text stringlengths 5 631k | id stringlengths 14 178 | metadata dict | __index_level_0__ int64 0 647 |
|---|---|---|---|
import { DacFeatureExtractor } from '../dac/feature_extraction_dac.js';
export class SnacFeatureExtractor extends DacFeatureExtractor { }
| transformers.js/src/models/snac/feature_extraction_snac.js/0 | {
"file_path": "transformers.js/src/models/snac/feature_extraction_snac.js",
"repo_id": "transformers.js",
"token_count": 42
} | 337 |
/**
* @file Helper module for `Tensor` processing.
*
* These functions and classes are only used internally,
* meaning an end-user shouldn't need to access anything here.
*
* @module utils/tensor
*/
import {
interpolate_data,
max,
min,
permute_data
} from './maths.js';
import {
Tensor as ONNXTensor, isONNXTensor,
} from '../backends/onnx.js';
import { TensorOpRegistry } from '../ops/registry.js';
export const DataTypeMap = Object.freeze({
float32: Float32Array,
// @ts-ignore ts(2552) Limited availability of Float16Array across browsers:
// https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/Float16Array
float16: typeof Float16Array !== "undefined" ? Float16Array: Uint16Array,
float64: Float64Array,
string: Array, // string[]
int8: Int8Array,
uint8: Uint8Array,
int16: Int16Array,
uint16: Uint16Array,
int32: Int32Array,
uint32: Uint32Array,
int64: BigInt64Array,
uint64: BigUint64Array,
bool: Uint8Array,
uint4: Uint8Array,
int4: Int8Array,
});
/**
* @typedef {keyof typeof DataTypeMap} DataType
* @typedef {import('./maths.js').AnyTypedArray | any[]} DataArray
*/
export class Tensor {
/** @type {number[]} Dimensions of the tensor. */
get dims() {
// @ts-ignore
return this.ort_tensor.dims;
}
set dims(value) {
// FIXME: ONNXTensor declares dims as readonly so one needs to use the constructor() if dims change.
// @ts-ignore
this.ort_tensor.dims = value;
}
/** @type {DataType} Type of the tensor. */
get type() {
return this.ort_tensor.type;
};
/** @type {DataArray} The data stored in the tensor. */
get data() {
return this.ort_tensor.data;
}
/** @type {number} The number of elements in the tensor. */
get size() {
return this.ort_tensor.size;
};
/** @type {string} The location of the tensor data. */
get location() {
return this.ort_tensor.location;
};
ort_tensor;
/**
* Create a new Tensor or copy an existing Tensor.
* @param {[DataType, DataArray, number[]]|[ONNXTensor]} args
*/
constructor(...args) {
if (isONNXTensor(args[0])) {
this.ort_tensor = /** @type {ONNXTensor} */ (args[0]);
} else {
// Create new tensor
this.ort_tensor = new ONNXTensor(
/** @type {DataType} */(args[0]),
// @ts-expect-error ts(2769) Type 'number' is not assignable to type 'bigint'.
/** @type {Exclude<import('./maths.js').AnyTypedArray, Uint8ClampedArray>} */(args[1]),
args[2],
);
}
return new Proxy(this, {
get: (obj, key) => {
if (typeof key === 'string') {
let index = Number(key);
if (Number.isInteger(index)) {
// key is an integer (i.e., index)
return obj._getitem(index);
}
}
// @ts-ignore
return obj[key];
},
set: (obj, key, value) => {
// TODO allow setting of data
// @ts-ignore
return obj[key] = value;
}
});
}
dispose() {
this.ort_tensor.dispose();
// this.ort_tensor = undefined;
}
/**
* Returns an iterator object for iterating over the tensor data in row-major order.
* If the tensor has more than one dimension, the iterator will yield subarrays.
* @returns {Iterator} An iterator object for iterating over the tensor data in row-major order.
*/
*[Symbol.iterator]() {
const [iterLength, ...iterDims] = this.dims;
if (iterDims.length > 0) {
const iterSize = iterDims.reduce((a, b) => a * b);
for (let i = 0; i < iterLength; ++i) {
yield this._subarray(i, iterSize, iterDims);
}
} else {
yield* this.data
}
}
/**
* Index into a Tensor object.
* @param {number} index The index to access.
* @returns {Tensor} The data at the specified index.
*/
_getitem(index) {
const [iterLength, ...iterDims] = this.dims;
index = safeIndex(index, iterLength);
if (iterDims.length > 0) {
const iterSize = iterDims.reduce((a, b) => a * b);
return this._subarray(index, iterSize, iterDims);
} else {
return new Tensor(this.type, [this.data[index]], iterDims);
}
}
/**
* @param {number|bigint} item The item to search for in the tensor
* @returns {number} The index of the first occurrence of item in the tensor data.
*/
indexOf(item) {
const this_data = this.data;
for (let index = 0; index < this_data.length; ++index) {
// Note: == instead of === so we can match Ints with BigInts
if (this_data[index] == item) {
return index;
}
}
return -1;
}
/**
* @param {number} index
* @param {number} iterSize
* @param {any} iterDims
* @returns {Tensor}
*/
_subarray(index, iterSize, iterDims) {
const o1 = index * iterSize;
const o2 = (index + 1) * iterSize;
// We use subarray if available (typed array), otherwise we use slice (normal array)
const data =
('subarray' in this.data)
? this.data.subarray(o1, o2)
: this.data.slice(o1, o2);
return new Tensor(this.type, data, iterDims);
}
/**
* Returns the value of this tensor as a standard JavaScript Number. This only works
* for tensors with one element. For other cases, see `Tensor.tolist()`.
* @returns {number|bigint} The value of this tensor as a standard JavaScript Number.
* @throws {Error} If the tensor has more than one element.
*/
item() {
const this_data = this.data;
if (this_data.length !== 1) {
throw new Error(`a Tensor with ${this_data.length} elements cannot be converted to Scalar`);
}
return this_data[0];
}
/**
* Convert tensor data to a n-dimensional JS list
* @returns {Array}
*/
tolist() {
return reshape(this.data, this.dims)
}
/**
* Return a new Tensor with the sigmoid function applied to each element.
* @returns {Tensor} The tensor with the sigmoid function applied.
*/
sigmoid() {
return this.clone().sigmoid_();
}
/**
* Applies the sigmoid function to the tensor in place.
* @returns {Tensor} Returns `this`.
*/
sigmoid_() {
const this_data = this.data;
for (let i = 0; i < this_data.length; ++i) {
this_data[i] = 1 / (1 + Math.exp(-this_data[i]));
}
return this;
}
/**
* Return a new Tensor with a callback function applied to each element.
* @param {Function} callback - The function to apply to each element. It should take three arguments:
* the current element, its index, and the tensor's data array.
* @returns {Tensor} A new Tensor with the callback function applied to each element.
*/
map(callback) {
return this.clone().map_(callback);
}
/**
* Apply a callback function to each element of the tensor in place.
* @param {Function} callback - The function to apply to each element. It should take three arguments:
* the current element, its index, and the tensor's data array.
* @returns {Tensor} Returns `this`.
*/
map_(callback) {
const this_data = this.data;
for (let i = 0; i < this_data.length; ++i) {
this_data[i] = callback(this_data[i], i, this_data);
}
return this;
}
/**
* Return a new Tensor with every element multiplied by a constant.
* @param {number} val The value to multiply by.
* @returns {Tensor} The new tensor.
*/
mul(val) {
return this.clone().mul_(val);
}
/**
* Multiply the tensor by a constant in place.
* @param {number} val The value to multiply by.
* @returns {Tensor} Returns `this`.
*/
mul_(val) {
const this_data = this.data;
for (let i = 0; i < this_data.length; ++i) {
this_data[i] *= val;
}
return this;
}
/**
* Return a new Tensor with every element divided by a constant.
* @param {number} val The value to divide by.
* @returns {Tensor} The new tensor.
*/
div(val) {
return this.clone().div_(val);
}
/**
* Divide the tensor by a constant in place.
* @param {number} val The value to divide by.
* @returns {Tensor} Returns `this`.
*/
div_(val) {
const this_data = this.data;
for (let i = 0; i < this_data.length; ++i) {
this_data[i] /= val;
}
return this;
}
/**
* Return a new Tensor with every element added by a constant.
* @param {number} val The value to add by.
* @returns {Tensor} The new tensor.
*/
add(val) {
return this.clone().add_(val);
}
/**
* Add the tensor by a constant in place.
* @param {number} val The value to add by.
* @returns {Tensor} Returns `this`.
*/
add_(val) {
const this_data = this.data;
for (let i = 0; i < this_data.length; ++i) {
this_data[i] += val;
}
return this;
}
/**
* Return a new Tensor with every element subtracted by a constant.
* @param {number} val The value to subtract by.
* @returns {Tensor} The new tensor.
*/
sub(val) {
return this.clone().sub_(val);
}
/**
* Subtract the tensor by a constant in place.
* @param {number} val The value to subtract by.
* @returns {Tensor} Returns `this`.
*/
sub_(val) {
const this_data = this.data;
for (let i = 0; i < this_data.length; ++i) {
this_data[i] -= val;
}
return this;
}
/**
* Creates a deep copy of the current Tensor.
* @returns {Tensor} A new Tensor with the same type, data, and dimensions as the original.
*/
clone() {
return new Tensor(this.type, this.data.slice(), this.dims.slice());
}
/**
* Performs a slice operation on the Tensor along specified dimensions.
*
* Consider a Tensor that has a dimension of [4, 7]:
* ```
* [ 1, 2, 3, 4, 5, 6, 7]
* [ 8, 9, 10, 11, 12, 13, 14]
* [15, 16, 17, 18, 19, 20, 21]
* [22, 23, 24, 25, 26, 27, 28]
* ```
* We can slice against the two dims of row and column, for instance in this
* case we can start at the second element, and return to the second last,
* like this:
* ```
* tensor.slice([1, -1], [1, -1]);
* ```
* which would return:
* ```
* [ 9, 10, 11, 12, 13 ]
* [ 16, 17, 18, 19, 20 ]
* ```
*
* @param {...(number|number[]|null)} slices The slice specifications for each dimension.
* - If a number is given, then a single element is selected.
* - If an array of two numbers is given, then a range of elements [start, end (exclusive)] is selected.
* - If null is given, then the entire dimension is selected.
* @returns {Tensor} A new Tensor containing the selected elements.
* @throws {Error} If the slice input is invalid.
*/
slice(...slices) {
// This allows for slicing with ranges and numbers
const newTensorDims = [];
const newOffsets = [];
// slices is an array of numbers or arrays of numbers
// e.g., slices = [0, [1, 3], null, [0, 3]]
for (let sliceIndex = 0; sliceIndex < this.dims.length; ++sliceIndex) {
let slice = slices[sliceIndex];
if (slice === null || slice === undefined) {
// null or undefined means take the whole dimension
newOffsets.push([0, this.dims[sliceIndex]]);
newTensorDims.push(this.dims[sliceIndex]);
} else if (typeof slice === 'number') {
slice = safeIndex(slice, this.dims[sliceIndex], sliceIndex);
// A number means take a single element
newOffsets.push([slice, slice + 1]);
} else if (Array.isArray(slice) && slice.length === 2) {
// An array of length 2 means take a range of elements
let [start, end] = slice;
start = start === null
? 0
: safeIndex(start, this.dims[sliceIndex], sliceIndex, false);
end = end === null
? this.dims[sliceIndex]
: safeIndex(end, this.dims[sliceIndex], sliceIndex, false);
if (start > end) {
throw new Error(`Invalid slice: ${slice}`);
}
const offsets = [
Math.max(start, 0),
Math.min(end, this.dims[sliceIndex])
];
newOffsets.push(offsets);
newTensorDims.push(offsets[1] - offsets[0]);
} else {
throw new Error(`Invalid slice: ${slice}`);
}
}
const newDims = newOffsets.map(([start, end]) => end - start);
const newBufferSize = newDims.reduce((a, b) => a * b);
const this_data = this.data;
// Allocate memory
// @ts-ignore
const data = new this_data.constructor(newBufferSize);
// Precompute strides
const stride = this.stride();
// Detect if the slice is contiguous
let isContiguous = true;
for (let i = 1; i < newDims.length; ++i) {
if (newOffsets[i][0] !== 0 || newOffsets[i][1] !== this.dims[i]) {
isContiguous = false;
break;
}
}
if (isContiguous) {
// Perform bulk copy for contiguous slices to improve performance
const start = newOffsets[0][0] * stride[0];
const end = newOffsets[0][1] * stride[0];
if (ArrayBuffer.isView(this_data)) {
// If this.data is a TypedArray, use subarray
// @ts-ignore
data.set(this_data.subarray(start, end));
} else if (Array.isArray(this_data)) {
// If this.data is a plain array, use slice
const slicedData = this_data.slice(start, end);
for (let i = 0; i < slicedData.length; ++i) {
data[i] = slicedData[i];
}
} else {
throw new Error("Unsupported data type for slicing");
}
} else {
// Fallback to manual copying for non-contiguous slices
for (let i = 0; i < newBufferSize; ++i) {
let originalIndex = 0;
for (let j = newDims.length - 1, num = i; j >= 0; --j) {
const size = newDims[j];
originalIndex += ((num % size) + newOffsets[j][0]) * stride[j];
num = Math.floor(num / size);
}
data[i] = this_data[originalIndex];
}
}
return new Tensor(this.type, data, newTensorDims);
}
/**
* Return a permuted version of this Tensor, according to the provided dimensions.
* @param {...number} dims Dimensions to permute.
* @returns {Tensor} The permuted tensor.
*/
permute(...dims) {
return permute(this, dims);
}
// TODO: implement transpose. For now (backwards compatibility), it's just an alias for permute()
transpose(...dims) {
return this.permute(...dims);
}
/**
* Returns the sum of each row of the input tensor in the given dimension dim.
*
* @param {number} [dim=null] The dimension or dimensions to reduce. If `null`, all dimensions are reduced.
* @param {boolean} keepdim Whether the output tensor has `dim` retained or not.
* @returns The summed tensor
*/
sum(dim = null, keepdim = false) {
return this.norm(1, dim, keepdim);
}
/**
* Returns the matrix norm or vector norm of a given tensor.
* @param {number|string} [p='fro'] The order of norm
* @param {number} [dim=null] Specifies which dimension of the tensor to calculate the norm across.
* If dim is None, the norm will be calculated across all dimensions of input.
* @param {boolean} [keepdim=false] Whether the output tensors have dim retained or not.
* @returns {Tensor} The norm of the tensor.
*/
norm(p = 'fro', dim = null, keepdim = false) {
if (p === 'fro') {
// NOTE: Since we only support integer dims, Frobenius norm produces the same result as p=2.
p = 2;
} else if (typeof p === 'string') {
throw Error(`Unsupported norm: ${p}`);
}
const this_data = this.data;
const fn = (a, b) => a + (b ** p);
if (dim === null) {
// @ts-ignore
const val = this_data.reduce(fn, 0) ** (1 / p);
return new Tensor(this.type, [val], []);
}
const [type, result, resultDims] = reduce_helper(fn, this, dim, keepdim);
if (p !== 1) {
for (let i = 0; i < result.length; ++i) {
result[i] = result[i] ** (1 / p);
}
}
return new Tensor(type, result, resultDims);
}
/**
* Performs `L_p` normalization of inputs over specified dimension. Operates in place.
* @param {number} [p=2] The exponent value in the norm formulation
* @param {number} [dim=1] The dimension to reduce
* @returns {Tensor} `this` for operation chaining.
*/
normalize_(p = 2.0, dim = 1) {
dim = safeIndex(dim, this.dims.length);
const norm = this.norm(p, dim, true);
const this_data = this.data;
const norm_data = norm.data;
for (let i = 0; i < this_data.length; ++i) {
// Calculate the index in the resulting array
let resultIndex = 0;
for (let j = this.dims.length - 1, num = i, resultMultiplier = 1; j >= 0; --j) {
const size = this.dims[j];
if (j !== dim) {
const index = num % size;
resultIndex += index * resultMultiplier;
resultMultiplier *= this.dims[j];
}
num = Math.floor(num / size);
}
// Divide by normalized value
this_data[i] /= norm_data[resultIndex];
}
return this;
}
/**
* Performs `L_p` normalization of inputs over specified dimension.
* @param {number} [p=2] The exponent value in the norm formulation
* @param {number} [dim=1] The dimension to reduce
* @returns {Tensor} The normalized tensor.
*/
normalize(p = 2.0, dim = 1) {
return this.clone().normalize_(p, dim);
}
/**
* Compute and return the stride of this tensor.
* Stride is the jump necessary to go from one element to the next one in the specified dimension dim.
* @returns {number[]} The stride of this tensor.
*/
stride() {
return dimsToStride(this.dims);
}
/**
* Returns a tensor with all specified dimensions of input of size 1 removed.
*
* NOTE: The returned tensor shares the storage with the input tensor, so changing the contents of one will change the contents of the other.
* If you would like a copy, use `tensor.clone()` before squeezing.
*
* @param {number|number[]} [dim=null] If given, the input will be squeezed only in the specified dimensions.
* @returns {Tensor} The squeezed tensor
*/
squeeze(dim = null) {
return new Tensor(
this.type,
this.data,
calc_squeeze_dims(this.dims, dim)
)
}
/**
* In-place version of @see {@link Tensor.squeeze}
*/
squeeze_(dim = null) {
this.dims = calc_squeeze_dims(this.dims, dim);
return this;
}
/**
* Returns a new tensor with a dimension of size one inserted at the specified position.
*
* NOTE: The returned tensor shares the same underlying data with this tensor.
*
* @param {number} dim The index at which to insert the singleton dimension
* @returns {Tensor} The unsqueezed tensor
*/
unsqueeze(dim = null) {
return new Tensor(
this.type,
this.data,
calc_unsqueeze_dims(this.dims, dim)
);
}
/**
* In-place version of @see {@link Tensor.unsqueeze}
*/
unsqueeze_(dim = null) {
this.dims = calc_unsqueeze_dims(this.dims, dim);
return this;
}
/**
* In-place version of @see {@link Tensor.flatten}
*/
flatten_(start_dim = 0, end_dim = -1) {
// TODO validate inputs
end_dim = (end_dim + this.dims.length) % this.dims.length;
let dimsToKeepBefore = this.dims.slice(0, start_dim);
let dimsToFlatten = this.dims.slice(start_dim, end_dim + 1);
let dimsToKeepAfter = this.dims.slice(end_dim + 1);
this.dims = [...dimsToKeepBefore, dimsToFlatten.reduce((a, b) => a * b, 1), ...dimsToKeepAfter]
return this;
}
/**
* Flattens input by reshaping it into a one-dimensional tensor.
* If `start_dim` or `end_dim` are passed, only dimensions starting with `start_dim`
* and ending with `end_dim` are flattened. The order of elements in input is unchanged.
* @param {number} start_dim the first dim to flatten
* @param {number} end_dim the last dim to flatten
* @returns {Tensor} The flattened tensor.
*/
flatten(start_dim = 0, end_dim = -1) {
return this.clone().flatten_(start_dim, end_dim);
}
/**
* Returns a new tensor with the same data as the `self` tensor but of a different `shape`.
* @param {...number} dims the desired size
* @returns {Tensor} The tensor with the same data but different shape
*/
view(...dims) {
// TODO: validate dims
let inferredIndex = -1;
for (let i = 0; i < dims.length; ++i) {
if (dims[i] === -1) {
if (inferredIndex !== -1) {
throw new Error("Only one dimension can be inferred");
}
inferredIndex = i;
}
}
const this_data = this.data;
if (inferredIndex !== -1) {
// Some dimension must be inferred
const productOther = dims.reduce((product, curr, index) => {
return index !== inferredIndex ? product * curr : product
}, 1);
dims[inferredIndex] = this_data.length / productOther;
}
return new Tensor(this.type, this_data, dims); // NOTE: uses same underlying storage
}
neg_() {
const this_data = this.data;
for (let i = 0; i < this_data.length; ++i) {
this_data[i] = -this_data[i];
}
return this;
}
neg() {
return this.clone().neg_();
}
/**
* Computes input > val element-wise.
* @param {number} val The value to compare with.
* @returns {Tensor} A boolean tensor that is `true` where input is greater than other and `false` elsewhere.
*/
gt(val) {
const mask = new Uint8Array(this.data.length);
const this_data = this.data;
for (let i = 0; i < this_data.length; ++i) {
mask[i] = this_data[i] > val ? 1 : 0;
}
return new Tensor('bool', mask, this.dims);
}
/**
* Computes input < val element-wise.
* @param {number} val The value to compare with.
* @returns {Tensor} A boolean tensor that is `true` where input is less than other and `false` elsewhere.
*/
lt(val) {
const mask = new Uint8Array(this.data.length);
const this_data = this.data;
for (let i = 0; i < this_data.length; ++i) {
mask[i] = this_data[i] < val ? 1 : 0;
}
return new Tensor('bool', mask, this.dims);
}
/**
* In-place version of @see {@link Tensor.clamp}
*/
clamp_(min, max) {
const this_data = this.data;
for (let i = 0; i < this_data.length; ++i) {
this_data[i] = Math.min(Math.max(this_data[i], min), max);
}
return this;
}
/**
* Clamps all elements in input into the range [ min, max ]
* @param {number} min lower-bound of the range to be clamped to
* @param {number} max upper-bound of the range to be clamped to
* @returns {Tensor} the output tensor.
*/
clamp(min, max) {
return this.clone().clamp_(min, max);
}
/**
* In-place version of @see {@link Tensor.round}
*/
round_() {
const this_data = this.data;
for (let i = 0; i < this_data.length; ++i) {
this_data[i] = Math.round(this_data[i]);
}
return this;
}
/**
* Rounds elements of input to the nearest integer.
* @returns {Tensor} the output tensor.
*/
round() {
return this.clone().round_();
}
mean(dim = null, keepdim = false) {
return mean(this, dim, keepdim);
}
min(dim = null, keepdim = false) {
if (dim === null) {
// None to reduce over all dimensions.
const val = min(this.data)[0];
return new Tensor(this.type, [val], [/* scalar */]);
}
const [type, result, resultDims] = reduce_helper((a, b) => Math.min(a, b), this, dim, keepdim, Infinity);
return new Tensor(type, result, resultDims);
}
max(dim = null, keepdim = false) {
if (dim === null) {
// None to reduce over all dimensions.
const val = max(this.data)[0];
return new Tensor(this.type, [val], [/* scalar */]);
}
const [type, result, resultDims] = reduce_helper((a, b) => Math.max(a, b), this, dim, keepdim, -Infinity);
return new Tensor(type, result, resultDims);
}
argmin(dim = null, keepdim = false) {
if (dim !== null) {
throw new Error("`dim !== null` not yet implemented.");
}
const index = min(this.data)[1];
return new Tensor('int64', [BigInt(index)], []);
}
argmax(dim = null, keepdim = false) {
if (dim !== null) {
throw new Error("`dim !== null` not yet implemented.");
}
const index = max(this.data)[1];
return new Tensor('int64', [BigInt(index)], []);
}
/**
* Performs Tensor dtype conversion.
* @param {DataType} type The desired data type.
* @returns {Tensor} The converted tensor.
*/
to(type) {
// If the self Tensor already has the correct dtype, then self is returned.
if (this.type === type) return this;
// Otherwise, the returned tensor is a copy of self with the desired dtype.
if (!DataTypeMap.hasOwnProperty(type)) {
throw new Error(`Unsupported type: ${type}`);
}
// Handle special cases where a mapping function is needed (e.g., where one type is a bigint and the other is a number)
let map_fn;
const is_source_bigint = ['int64', 'uint64'].includes(this.type);
const is_dest_bigint = ['int64', 'uint64'].includes(type);
if (is_source_bigint && !is_dest_bigint) {
// TypeError: Cannot convert a BigInt value to a number
map_fn = Number;
} else if (!is_source_bigint && is_dest_bigint) {
// TypeError: Cannot convert [x] to a BigInt
map_fn = BigInt;
}
// @ts-ignore
return new Tensor(type, DataTypeMap[type].from(this.data, map_fn), this.dims);
}
}
/**
* This creates a nested array of a given type and depth (see examples).
*
* @example
* NestArray<string, 1>; // string[]
* @example
* NestArray<number, 2>; // number[][]
* @example
* NestArray<string, 3>; // string[][][] etc.
* @template T
* @template {number} Depth
* @template {never[]} [Acc=[]]
* @typedef {Acc['length'] extends Depth ? T : NestArray<T[], Depth, [...Acc, never]>} NestArray
*/
/**
* Reshapes a 1-dimensional array into an n-dimensional array, according to the provided dimensions.
*
* @example
* reshape([10 ], [1 ]); // Type: number[] Value: [10]
* reshape([1, 2, 3, 4 ], [2, 2 ]); // Type: number[][] Value: [[1, 2], [3, 4]]
* reshape([1, 2, 3, 4, 5, 6, 7, 8], [2, 2, 2]); // Type: number[][][] Value: [[[1, 2], [3, 4]], [[5, 6], [7, 8]]]
* reshape([1, 2, 3, 4, 5, 6, 7, 8], [4, 2 ]); // Type: number[][] Value: [[1, 2], [3, 4], [5, 6], [7, 8]]
* @param {T[]|DataArray} data The input array to reshape.
* @param {DIM} dimensions The target shape/dimensions.
* @template T
* @template {[number]|number[]} DIM
* @returns {NestArray<T, DIM["length"]>} The reshaped array.
*/
function reshape(data, dimensions) {
const totalElements = data.length;
const dimensionSize = dimensions.reduce((a, b) => a * b);
if (totalElements !== dimensionSize) {
throw Error(`cannot reshape array of size ${totalElements} into shape (${dimensions})`);
}
/** @type {any} */
let reshapedArray = data;
for (let i = dimensions.length - 1; i >= 0; i--) {
reshapedArray = reshapedArray.reduce((acc, val) => {
let lastArray = acc[acc.length - 1];
if (lastArray.length < dimensions[i]) {
lastArray.push(val);
} else {
acc.push([val]);
}
return acc;
}, [[]]);
}
return reshapedArray[0];
}
/**
* Permutes a tensor according to the provided axes.
* @param {any} tensor The input tensor to permute.
* @param {Array} axes The axes to permute the tensor along.
* @returns {Tensor} The permuted tensor.
*/
export function permute(tensor, axes) {
const [permutedData, shape] = permute_data(tensor.data, tensor.dims, axes);
return new Tensor(tensor.type, permutedData, shape);
}
/**
* Interpolates an Tensor to the given size.
* @param {Tensor} input The input tensor to interpolate. Data must be channel-first (i.e., [c, h, w])
* @param {number[]} size The output size of the image
* @param {string} mode The interpolation mode
* @param {boolean} align_corners Whether to align corners.
* @returns {Tensor} The interpolated tensor.
*/
export function interpolate(input, [out_height, out_width], mode = 'bilinear', align_corners = false) {
// Input image dimensions
const in_channels = input.dims.at(-3) ?? 1;
const in_height = input.dims.at(-2);
const in_width = input.dims.at(-1);
let output = interpolate_data(
/** @type {import('./maths.js').TypedArray}*/(input.data),
[in_channels, in_height, in_width],
[out_height, out_width],
mode,
align_corners
);
return new Tensor(input.type, output, [in_channels, out_height, out_width]);
}
/**
* Down/up samples the input.
* Inspired by https://pytorch.org/docs/stable/generated/torch.nn.functional.interpolate.html.
* @param {Tensor} input the input tensor
* @param {Object} options the options for the interpolation
* @param {[number, number]|[number, number, number]|[number, number, number, number]} [options.size=null] output spatial size.
* @param {"nearest"|"bilinear"|"bicubic"} [options.mode='bilinear'] algorithm used for upsampling
* @returns {Promise<Tensor>} The interpolated tensor.
*/
export async function interpolate_4d(input, {
size = null,
mode = 'bilinear',
} = {}) {
// Error checking
if (input.dims.length !== 4) {
throw new Error('`interpolate_4d` currently only supports 4D input.');
}
if (!size) {
// TODO: support scale_factor
throw new Error('`interpolate_4d` requires a `size` argument.');
}
// Fill in missing dimensions
let targetDims;
if (size.length === 2) {
targetDims = [...input.dims.slice(0, 2), ...size];
} else if (size.length === 3) {
targetDims = [input.dims[0], ...size];
} else if (size.length === 4) {
targetDims = size;
} else {
throw new Error('`size` must be of length 2, 3, or 4.');
}
let op;
if (mode === 'nearest') {
op = await TensorOpRegistry.nearest_interpolate_4d;
} else if (mode === 'bilinear') {
op = await TensorOpRegistry.bilinear_interpolate_4d;
} else if (mode === 'bicubic') {
op = await TensorOpRegistry.bicubic_interpolate_4d;
} else {
throw new Error(`Unsupported mode: ${mode}`);
}
const sizeTensor = new Tensor('int64', new BigInt64Array(targetDims.map(BigInt)), [targetDims.length]);
return await op({ x: input, s: sizeTensor });
}
/**
* Matrix product of two tensors.
* Inspired by https://pytorch.org/docs/stable/generated/torch.matmul.html
* @param {Tensor} a the first tensor to be multiplied
* @param {Tensor} b the second tensor to be multiplied
* @returns {Promise<Tensor>} The matrix product of the two tensors.
*/
export async function matmul(a, b) {
const op = await TensorOpRegistry.matmul;
return await op({ a, b });
}
/**
* Computes the one dimensional Fourier transform of real-valued input.
* Inspired by https://pytorch.org/docs/stable/generated/torch.fft.rfft.html
* @param {Tensor} x the real input tensor
* @param {Tensor} a The dimension along which to take the one dimensional real FFT.
* @returns {Promise<Tensor>} the output tensor.
*/
export async function rfft(x, a) {
const op = await TensorOpRegistry.rfft;
return await op({ x, a });
}
/**
* Returns the k largest elements of the given input tensor.
* Inspired by https://pytorch.org/docs/stable/generated/torch.topk.html
* @param {Tensor} x the input tensor
* @param {number} [k] the k in "top-k"
* @returns {Promise<[Tensor, Tensor]>} the output tuple of (Tensor, LongTensor) of top-k elements and their indices.
*/
export async function topk(x, k) {
const op = await TensorOpRegistry.top_k;
if (k == null) {
k = x.dims.at(-1);
} else {
k = Math.min(k, x.dims.at(-1));
}
return await op({
x,
k: new Tensor(
'int64',
[BigInt(k)],
[1]
)
});
}
const arrayToIndexTensor = (array) => new Tensor('int64', array, [array.length]);
/**
* Slice a multidimensional float32 tensor.
* @param {Tensor} data: Tensor of data to extract slices from
* @param {number[]} starts: 1-D array of starting indices of corresponding axis in axes
* @param {number[]} ends: 1-D array of ending indices (exclusive) of corresponding axis in axes
* @param {number[]} axes: 1-D array of axes that starts and ends apply to
* @param {number[]} [steps]: 1-D array of slice step of corresponding axis in axes.
* @returns {Promise<Tensor>} Sliced data tensor.
*/
export async function slice(data, starts, ends, axes, steps) {
const op = await TensorOpRegistry.slice;
return await op({
x: data,
s: arrayToIndexTensor(starts),
e: arrayToIndexTensor(ends),
a: arrayToIndexTensor(axes),
t: arrayToIndexTensor(steps ?? new Array(axes.length).fill(1)),
});
}
/**
* Perform mean pooling of the last hidden state followed by a normalization step.
* @param {Tensor} last_hidden_state Tensor of shape [batchSize, seqLength, embedDim]
* @param {Tensor} attention_mask Tensor of shape [batchSize, seqLength]
* @returns {Tensor} Returns a new Tensor of shape [batchSize, embedDim].
*/
export function mean_pooling(last_hidden_state, attention_mask) {
// last_hidden_state: [batchSize, seqLength, embedDim]
// attention_mask: [batchSize, seqLength]
const lastHiddenStateData = last_hidden_state.data;
const attentionMaskData = attention_mask.data;
const shape = [last_hidden_state.dims[0], last_hidden_state.dims[2]];
// @ts-ignore
const returnedData = new lastHiddenStateData.constructor(shape[0] * shape[1]);
const [batchSize, seqLength, embedDim] = last_hidden_state.dims;
let outIndex = 0;
for (let i = 0; i < batchSize; ++i) {
const offset = i * embedDim * seqLength;
for (let k = 0; k < embedDim; ++k) {
let sum = 0;
let count = 0;
const attnMaskOffset = i * seqLength;
const offset2 = offset + k;
// Pool over all words in sequence
for (let j = 0; j < seqLength; ++j) {
// index into attention mask
const attn = Number(attentionMaskData[attnMaskOffset + j]);
count += attn;
sum += lastHiddenStateData[offset2 + j * embedDim] * attn;
}
const avg = sum / count;
returnedData[outIndex++] = avg;
}
}
return new Tensor(
last_hidden_state.type,
returnedData,
shape
)
}
/**
* Apply Layer Normalization for last certain number of dimensions.
* @param {Tensor} input The input tensor
* @param {number[]} normalized_shape input shape from an expected input of size
* @param {Object} options The options for the layer normalization
* @param {number} [options.eps=1e-5] A value added to the denominator for numerical stability.
* @returns {Tensor} The normalized tensor.
*/
export function layer_norm(input, normalized_shape, {
eps = 1e-5,
} = {}) {
if (input.dims.length !== 2) {
throw new Error('`layer_norm` currently only supports 2D input.');
}
const [batchSize, featureDim] = input.dims;
if (normalized_shape.length !== 1 && normalized_shape[0] !== featureDim) {
throw new Error('`normalized_shape` must be a 1D array with shape `[input.dims[1]]`.');
}
const [std, mean] = std_mean(input, 1, 0, true);
const stdData = /** @type {Float32Array} */(std.data);
const meanData = /** @type {Float32Array} */(mean.data);
const inputData = /** @type {Float32Array} */(input.data);
// @ts-ignore
const returnedData = new inputData.constructor(inputData.length);
for (let i = 0; i < batchSize; ++i) {
const offset = i * featureDim;
for (let j = 0; j < featureDim; ++j) {
const offset2 = offset + j;
returnedData[offset2] = (inputData[offset2] - meanData[i]) / (stdData[i] + eps);
}
}
return new Tensor(input.type, returnedData, input.dims);
}
/**
* Helper function to calculate new dimensions when performing a squeeze operation.
* @param {number[]} dims The dimensions of the tensor.
* @param {number|number[]|null} dim The dimension(s) to squeeze.
* @returns {number[]} The new dimensions.
* @private
*/
function calc_squeeze_dims(dims, dim) {
dims = dims.slice();
if (dim === null) {
dims = dims.filter((d) => d !== 1);
} else if (typeof dim === 'number') {
if (dims[dim] === 1) {
dims.splice(dim, 1);
}
} else if (Array.isArray(dim)) {
dims = dims.filter((x, i) => {
return x !== 1 || !dim.includes(i);
});
}
return dims;
}
/**
* Helper function to calculate new dimensions when performing an unsqueeze operation.
* @param {number[]} dims The dimensions of the tensor.
* @param {number} dim The dimension to unsqueeze.
* @returns {number[]} The new dimensions.
* @private
*/
function calc_unsqueeze_dims(dims, dim) {
// Dimension out of range (e.g., "expected to be in range of [-4, 3], but got 4")
// + 1 since we allow inserting at the end (i.e. dim = -1)
dim = safeIndex(dim, dims.length + 1);
dims = dims.slice();
// Insert 1 into specified dimension
dims.splice(dim, 0, 1);
return dims;
}
/**
* Safely calculate the index for an array of a given size, allowing negative indexing.
* @param {number} index The index that will be used.
* @param {number} size The size of the array.
* @param {number} [dimension=null] The dimension that the index is for (optional).
* @returns {number} The index, guaranteed to be non-negative and less than `arrayLength`.
*
* @throws {Error} If the index is out of range.
* @private
*/
function safeIndex(index, size, dimension = null, boundsCheck = true) {
if (index < -size || index >= size) {
if (boundsCheck) {
throw new Error(`IndexError: index ${index} is out of bounds for dimension${dimension === null ? '' : ' ' + dimension} with size ${size}`);
} else {
return index < -size ? 0 : size;
}
}
if (index < 0) {
// Negative indexing, ensuring positive index
index = ((index % size) + size) % size;
}
return index;
}
/**
* Concatenates an array of tensors along a specified dimension.
* @param {Tensor[]} tensors The array of tensors to concatenate.
* @param {number} dim The dimension to concatenate along.
* @returns {Tensor} The concatenated tensor.
*/
export function cat(tensors, dim = 0) {
dim = safeIndex(dim, tensors[0].dims.length);
// TODO do validation of shapes
const resultDims = tensors[0].dims.slice();
resultDims[dim] = tensors.reduce((a, b) => a + b.dims[dim], 0);
// Create a new array to store the accumulated values
const resultSize = resultDims.reduce((a, b) => a * b, 1);
// @ts-ignore
const result = new tensors[0].data.constructor(resultSize);
// Create output tensor of same type as first
const resultType = tensors[0].type;
if (dim === 0) {
// Handle special case for performance reasons
let offset = 0;
for (const tensor of tensors) {
const tensorData = tensor.data;
result.set(tensorData, offset);
offset += tensorData.length;
}
} else {
let currentDim = 0;
for (let t = 0; t < tensors.length; ++t) {
const { data, dims } = tensors[t];
// Iterate over the data array
for (let i = 0; i < data.length; ++i) {
// Calculate the index in the resulting array
let resultIndex = 0;
for (let j = dims.length - 1, num = i, resultMultiplier = 1; j >= 0; --j) {
const size = dims[j];
let index = num % size;
if (j === dim) {
index += currentDim;
}
resultIndex += index * resultMultiplier;
resultMultiplier *= resultDims[j];
num = Math.floor(num / size);
}
// Accumulate the value at the current index
result[resultIndex] = data[i];
}
currentDim += dims[dim];
}
}
return new Tensor(resultType, result, resultDims);
}
/**
* Stack an array of tensors along a specified dimension.
* @param {Tensor[]} tensors The array of tensors to stack.
* @param {number} dim The dimension to stack along.
* @returns {Tensor} The stacked tensor.
*/
export function stack(tensors, dim = 0) {
// TODO do validation of shapes
// NOTE: stack expects each tensor to be equal size
return cat(tensors.map(t => t.unsqueeze(dim)), dim);
}
/**
* @param {(previousValue: any, currentValue: any, currentIndex?: number, resultIndex?: number) => any} callbackfn
* @param {Tensor} input the input tensor.
* @param {number|null} dim the dimension to reduce.
* @param {boolean} keepdim whether the output tensor has dim retained or not.
* @returns {[DataType, any, number[]]} The reduced tensor data.
*/
function reduce_helper(callbackfn, input, dim = null, keepdim = false, initialValue = null) {
const inputData = input.data;
const inputDims = input.dims;
// Negative indexing
dim = safeIndex(dim, inputDims.length);
// Calculate the shape of the resulting array after summation
const resultDims = inputDims.slice(); // Copy the original dimensions
resultDims[dim] = 1; // Remove the specified axis
// Create a new array to store the accumulated values
// @ts-ignore
const result = new inputData.constructor(inputData.length / inputDims[dim]);
if (initialValue !== null) {
result.fill(initialValue);
}
// Iterate over the data array
for (let i = 0; i < inputData.length; ++i) {
// Calculate the index in the resulting array
let resultIndex = 0;
for (let j = inputDims.length - 1, num = i, resultMultiplier = 1; j >= 0; --j) {
const size = inputDims[j];
if (j !== dim) {
const index = num % size;
resultIndex += index * resultMultiplier;
resultMultiplier *= resultDims[j];
}
num = Math.floor(num / size);
}
// Accumulate the value at the current index
result[resultIndex] = callbackfn(result[resultIndex], inputData[i], i, resultIndex);
}
if (!keepdim) resultDims.splice(dim, 1);
return [input.type, result, resultDims];
}
/**
* Calculates the standard deviation and mean over the dimensions specified by dim. dim can be a single dimension or `null` to reduce over all dimensions.
* @param {Tensor} input the input tenso
* @param {number|null} dim the dimension to reduce. If None, all dimensions are reduced.
* @param {number} correction difference between the sample size and sample degrees of freedom. Defaults to Bessel's correction, correction=1.
* @param {boolean} keepdim whether the output tensor has dim retained or not.
* @returns {Tensor[]} A tuple of (std, mean) tensors.
*/
export function std_mean(input, dim = null, correction = 1, keepdim = false) {
const inputData = /** @type {Float32Array} */(input.data);
const inputDims = input.dims;
if (dim === null) {
// None to reduce over all dimensions.
const sum = inputData.reduce((a, b) => a + b, 0);
const mean = sum / inputData.length;
const std = Math.sqrt(inputData.reduce((a, b) => a + (b - mean) ** 2, 0) / (inputData.length - correction));
const meanTensor = new Tensor(input.type, [mean], [/* scalar */]);
const stdTensor = new Tensor(input.type, [std], [/* scalar */]);
return [stdTensor, meanTensor];
}
dim = safeIndex(dim, inputDims.length);
const meanTensor = mean(input, dim, keepdim);
const meanTensorData = meanTensor.data;
// Compute squared sum
const [type, result, resultDims] = reduce_helper((a, b, i, j) => a + (b - meanTensorData[j]) ** 2, input, dim, keepdim);
// Square root of the squared sum
for (let i = 0; i < result.length; ++i) {
result[i] = Math.sqrt(result[i] / (inputDims[dim] - correction));
}
const stdTensor = new Tensor(type, result, resultDims);
return [stdTensor, meanTensor];
}
/**
* Returns the mean value of each row of the input tensor in the given dimension dim.
* @param {Tensor} input the input tensor.
* @param {number|null} dim the dimension to reduce.
* @param {boolean} keepdim whether the output tensor has dim retained or not.
* @returns {Tensor} A new tensor with means taken along the specified dimension.
*/
export function mean(input, dim = null, keepdim = false) {
const inputDims = input.dims;
const inputData = /** @type {Float32Array} */(input.data);
if (dim === null) {
// None to reduce over all dimensions.
const val = inputData.reduce((a, b) => a + b, 0);
return new Tensor(input.type, [val / inputData.length], [/* scalar */]);
}
dim = safeIndex(dim, inputDims.length);
// Compute sum
const [type, result, resultDims] = reduce_helper((a, b) => a + b, input, dim, keepdim);
// Divide by number of elements in the dimension
if (inputDims[dim] !== 1) {
for (let i = 0; i < result.length; ++i) {
result[i] /= inputDims[dim];
}
}
return new Tensor(type, result, resultDims);
}
function dimsToStride(dims) {
const stride = new Array(dims.length);
for (let i = dims.length - 1, s2 = 1; i >= 0; --i) {
stride[i] = s2;
s2 *= dims[i];
}
return stride;
}
function fullHelper(size, fill_value, dtype, cls) {
const numElements = size.reduce((a, b) => a * b, 1);
return new Tensor(
dtype,
new cls(numElements).fill(fill_value),
size
)
}
/**
* Creates a tensor of size size filled with fill_value. The tensor's dtype is inferred from fill_value.
* @param {number[]} size A sequence of integers defining the shape of the output tensor.
* @param {number|bigint|boolean} fill_value The value to fill the output tensor with.
* @returns {Tensor} The filled tensor.
*/
export function full(size, fill_value) {
let dtype;
let typedArrayCls;
if (typeof fill_value === 'number') {
dtype = 'float32';
typedArrayCls = Float32Array;
} else if (typeof fill_value === 'bigint') {
dtype = 'int64';
typedArrayCls = BigInt64Array;
} else if (typeof fill_value === 'boolean') {
dtype = 'bool';
typedArrayCls = Uint8Array;
} else {
// TODO: support other dtypes
throw new Error(`Unsupported data type: ${typeof fill_value}`);
}
return fullHelper(size, fill_value, dtype, typedArrayCls);
}
export function full_like(tensor, fill_value) {
return full(tensor.dims, fill_value);
}
/**
* Returns a tensor filled with the scalar value 1, with the shape defined by the variable argument size.
* @param {number[]} size A sequence of integers defining the shape of the output tensor.
* @returns {Tensor} The ones tensor.
*/
export function ones(size) {
return fullHelper(size, 1n, 'int64', BigInt64Array);
}
/**
* Returns a tensor filled with the scalar value 1, with the same size as input.
* @param {Tensor} tensor The size of input will determine size of the output tensor.
* @returns {Tensor} The ones tensor.
*/
export function ones_like(tensor) {
return ones(tensor.dims);
}
/**
* Returns a tensor filled with the scalar value 0, with the shape defined by the variable argument size.
* @param {number[]} size A sequence of integers defining the shape of the output tensor.
* @returns {Tensor} The zeros tensor.
*/
export function zeros(size) {
return fullHelper(size, 0n, 'int64', BigInt64Array);
}
/**
* Returns a tensor filled with the scalar value 0, with the same size as input.
* @param {Tensor} tensor The size of input will determine size of the output tensor.
* @returns {Tensor} The zeros tensor.
*/
export function zeros_like(tensor) {
return zeros(tensor.dims);
}
/**
* Returns a tensor filled with random numbers from a uniform distribution on the interval [0, 1)
* @param {number[]} size A sequence of integers defining the shape of the output tensor.
* @returns {Tensor} The random tensor.
*/
export function rand(size) {
const length = size.reduce((a, b) => a * b, 1);
return new Tensor(
"float32",
Float32Array.from({ length }, () => Math.random()),
size,
)
}
/**
* Quantizes the embeddings tensor to binary or unsigned binary precision.
* @param {Tensor} tensor The tensor to quantize.
* @param {'binary'|'ubinary'} precision The precision to use for quantization.
* @returns {Tensor} The quantized tensor.
*/
export function quantize_embeddings(tensor, precision) {
if (tensor.dims.length !== 2) {
throw new Error("The tensor must have 2 dimensions");
}
if (tensor.dims.at(-1) % 8 !== 0) {
throw new Error("The last dimension of the tensor must be a multiple of 8");
}
if (!['binary', 'ubinary'].includes(precision)) {
throw new Error("The precision must be either 'binary' or 'ubinary'");
}
const signed = precision === 'binary';
const dtype = signed ? 'int8' : 'uint8';
// Create a typed array to store the packed bits
const cls = signed ? Int8Array : Uint8Array;
const inputData = tensor.data;
const outputData = new cls(inputData.length / 8);
// Iterate over each number in the array
for (let i = 0; i < inputData.length; ++i) {
// Determine if the number is greater than 0
const bit = inputData[i] > 0 ? 1 : 0;
// Calculate the index in the typed array and the position within the byte
const arrayIndex = Math.floor(i / 8);
const bitPosition = i % 8;
// Pack the bit into the typed array
outputData[arrayIndex] |= bit << (7 - bitPosition);
if (signed && bitPosition === 0) {
outputData[arrayIndex] -= 128;
}
};
return new Tensor(dtype, outputData, [tensor.dims[0], tensor.dims[1] / 8]);
}
| transformers.js/src/utils/tensor.js/0 | {
"file_path": "transformers.js/src/utils/tensor.js",
"repo_id": "transformers.js",
"token_count": 22329
} | 338 |
import { BlenderbotSmallTokenizer } from "../../../src/tokenizers.js";
import { BASE_TEST_STRINGS, BLENDERBOT_SMALL_TEST_STRINGS } from "../test_strings.js";
export const TOKENIZER_CLASS = BlenderbotSmallTokenizer;
// NOTE: `.tokenize()` is disabled for BlenderbotSmallTokenizer
export const TEST_CONFIG = {
"Xenova/blenderbot_small-90M": {
SIMPLE: {
text: BASE_TEST_STRINGS.SIMPLE,
// "tokens": ["how", "are", "you", "doing", "?"],
ids: [102, 46, 15, 267, 20],
decoded: "how are you doing?",
},
SIMPLE_WITH_PUNCTUATION: {
text: BASE_TEST_STRINGS.SIMPLE_WITH_PUNCTUATION,
// "tokens": ["you", "should", "'", "ve", "done", "this"],
ids: [15, 197, 8, 117, 369, 36],
decoded: "you should've done this",
},
NUMBERS: {
text: BASE_TEST_STRINGS.NUMBERS,
// "tokens": ["0@@", "1@@", "2@@", "3@@", "4@@", "5@@", "6@@", "7@@", "89", "0", "1", "2", "3", "4", "5", "6", "7", "8", "9", "10", "100", "1000"],
ids: [1988, 2388, 735, 801, 827, 948, 981, 1110, 4814, 520, 143, 176, 216, 260, 253, 345, 374, 420, 475, 316, 773, 6217],
decoded: "0123456789 0 1 2 3 4 5 6 7 8 9 10 100 1000",
},
TEXT_WITH_NUMBERS: {
text: BASE_TEST_STRINGS.TEXT_WITH_NUMBERS,
// "tokens": ["the", "company", "was", "founded", "in", "2016", "."],
ids: [7, 293, 18, 912, 13, 845, 5],
decoded: "the company was founded in 2016.",
},
PUNCTUATION: {
text: BASE_TEST_STRINGS.PUNCTUATION,
// "tokens": ["a", "__newln__", "'", "ll", "!", "!@@", "to", "?", "'", "d", "'", "'", "d", "of", ",", "can", "'", "t", "."],
ids: [12, 4, 8, 97, 37, 3, 11, 20, 8, 85, 8, 8, 85, 10, 6, 62, 8, 30, 5],
decoded: "a __newln__'ll! __unk__ to?'d'' d of, can't.",
},
PYTHON_CODE: {
text: BASE_TEST_STRINGS.PYTHON_CODE,
// "tokens": ["def", "main", "(", ")@@", ":", "__newln__", "pass"],
ids: [21996, 550, 40, 3, 106, 4, 1314],
decoded: "def main ( __unk__ : __newln__ pass",
},
JAVASCRIPT_CODE: {
text: BASE_TEST_STRINGS.JAVASCRIPT_CODE,
// "tokens": ["let", "a", "=", "ob@@", "j", ".@@", "to@@", "string", "(", ")@@", ";", "__newln__", "to@@", "string", "(", ")@@", ";"],
ids: [131, 12, 1381, 2808, 755, 3, 752, 4529, 40, 3, 118, 4, 752, 4529, 40, 3, 118],
decoded: "let a = obj __unk__ tostring ( __unk__ ; __newln__ tostring ( __unk__ ;",
},
NEWLINES: {
text: BASE_TEST_STRINGS.NEWLINES,
// "tokens": ["this", "__newln__", "is", "__newln__", "a", "__newln__", "test", "."],
ids: [36, 4, 24, 4, 12, 4, 1248, 5],
decoded: "this __newln__ is __newln__ a __newln__ test.",
},
BASIC: {
text: BASE_TEST_STRINGS.BASIC,
// "tokens": ["un@@", "wan@@", "t@@", "\u00e9@@", "d", ",@@", "running"],
ids: [204, 4151, 291, 1677, 85, 3, 785],
decoded: "unwant\u00e9d __unk__ running",
},
CONTROL_TOKENS: {
text: BASE_TEST_STRINGS.CONTROL_TOKENS,
// "tokens": ["1@@", "\u0000@@", "2@@", "\ufffd@@", "3"],
ids: [2388, 3, 735, 3, 216],
decoded: "1__unk__ 2__unk__ 3",
},
HELLO_WORLD_TITLECASE: {
text: BASE_TEST_STRINGS.HELLO_WORLD_TITLECASE,
// "tokens": ["hello", "world"],
ids: [880, 159],
decoded: "hello world",
},
HELLO_WORLD_LOWERCASE: {
text: BASE_TEST_STRINGS.HELLO_WORLD_LOWERCASE,
// "tokens": ["hello", "world"],
ids: [880, 159],
decoded: "hello world",
},
CHINESE_ONLY: {
text: BASE_TEST_STRINGS.CHINESE_ONLY,
// "tokens": ["\u751f@@", "\u6d3b@@", "\u7684@@", "\u771f@@", "\u8c1b@@", "\u662f"],
ids: [30488, 32756, 29891, 30813, 3, 34037],
decoded: "\u751f\u6d3b\u7684\u771f__unk__ \u662f",
},
LEADING_SPACE: {
text: BASE_TEST_STRINGS.LEADING_SPACE,
// "tokens": ["leading", "space"],
ids: [1164, 833],
decoded: "leading space",
},
TRAILING_SPACE: {
text: BASE_TEST_STRINGS.TRAILING_SPACE,
// "tokens": ["trailing", "space"],
ids: [12499, 833],
decoded: "trailing space",
},
DOUBLE_SPACE: {
text: BASE_TEST_STRINGS.DOUBLE_SPACE,
// "tokens": ["hi", "hello"],
ids: [792, 880],
decoded: "hi hello",
},
CURRENCY: {
text: BASE_TEST_STRINGS.CURRENCY,
// "tokens": ["test", "$@@", "1", "r@@", "2", "#@@", "3", "\u20ac@@", "4", "\u00a3@@", "5", "\u00a5@@", "6", "\u20a3@@", "7", "\u20b9@@", "8", "\u20b1@@", "9", "test"],
ids: [1248, 3, 143, 510, 176, 3, 216, 3, 260, 3, 253, 3, 345, 3, 374, 3, 420, 3, 475, 1248],
decoded: "test __unk__ 1 r2 __unk__ 3 __unk__ 4 __unk__ 5 __unk__ 6 __unk__ 7 __unk__ 8 __unk__ 9 test",
},
CURRENCY_WITH_DECIMALS: {
text: BASE_TEST_STRINGS.CURRENCY_WITH_DECIMALS,
// "tokens": ["i", "bought", "an", "apple", "for", "$@@", "1", ".@@", "00", "at", "the", "store", "."],
ids: [14, 1890, 50, 4758, 26, 3, 143, 3, 1966, 32, 7, 1640, 5],
decoded: "i bought an apple for __unk__ 1 __unk__ 00 at the store.",
},
ELLIPSIS: {
text: BASE_TEST_STRINGS.ELLIPSIS,
// "tokens": ["you@@", "\u2026"],
ids: [7984, 1244],
decoded: "you\u2026",
},
TEXT_WITH_ESCAPE_CHARACTERS: {
text: BASE_TEST_STRINGS.TEXT_WITH_ESCAPE_CHARACTERS,
// "tokens": ["you@@", "\u2026"],
ids: [7984, 1244],
decoded: "you\u2026",
},
TEXT_WITH_ESCAPE_CHARACTERS_2: {
text: BASE_TEST_STRINGS.TEXT_WITH_ESCAPE_CHARACTERS_2,
// "tokens": ["you@@", "\u2026", "you@@", "\u2026"],
ids: [7984, 1244, 7984, 1244],
decoded: "you\u2026 you\u2026",
},
TILDE_NORMALIZATION: {
text: BASE_TEST_STRINGS.TILDE_NORMALIZATION,
// "tokens": ["weird", "\uff5e", "edge", "\uff5e", "case"],
ids: [2614, 30831, 1649, 30831, 543],
decoded: "weird \uff5e edge \uff5e case",
},
SPIECE_UNDERSCORE: {
text: BASE_TEST_STRINGS.SPIECE_UNDERSCORE,
// "tokens": ["\u2581@@", "this", "\u2581@@", "is", "\u2581@@", "a", "\u2581@@", "test", "\u2581", "."],
ids: [3, 36, 3, 24, 3, 12, 3, 1248, 50106, 5],
decoded: "__unk__ this __unk__ is __unk__ a __unk__ test \u2581.",
},
SPECIAL_TOKENS: {
text: BLENDERBOT_SMALL_TEST_STRINGS.SPECIAL_TOKENS,
// "tokens": ["__start__", "hello", "world", "__end__"],
ids: [1, 880, 159, 2],
decoded: "__start__ hello world __end__",
},
WHITESPACE_1: {
text: BLENDERBOT_SMALL_TEST_STRINGS.WHITESPACE_1,
// "tokens": ["__start__", "hey", "__end__"],
ids: [1, 226, 2],
decoded: "__start__ hey __end__",
},
WHITESPACE_2: {
text: BLENDERBOT_SMALL_TEST_STRINGS.WHITESPACE_2,
// "tokens": ["__start__", "hey", "__end__"],
ids: [1, 226, 2],
decoded: "__start__ hey __end__",
},
},
};
| transformers.js/tests/models/blenderbot_small/test_tokenization_blenderbot_small.js/0 | {
"file_path": "transformers.js/tests/models/blenderbot_small/test_tokenization_blenderbot_small.js",
"repo_id": "transformers.js",
"token_count": 3493
} | 339 |
import { AutoImageProcessor, DPTFeatureExtractor, DPTImageProcessor } from "../../../src/transformers.js";
import { load_cached_image } from "../../asset_cache.js";
import { MAX_PROCESSOR_LOAD_TIME, MAX_TEST_EXECUTION_TIME } from "../../init.js";
export default () => {
// DPTFeatureExtractor
describe("DPTFeatureExtractor", () => {
const model_id = "Xenova/dpt-hybrid-midas";
/** @type {DPTFeatureExtractor} */
let processor;
beforeAll(async () => {
processor = await AutoImageProcessor.from_pretrained(model_id);
}, MAX_PROCESSOR_LOAD_TIME);
it(
"grayscale images",
async () => {
const image = await load_cached_image("cats");
const { pixel_values, original_sizes, reshaped_input_sizes } = await processor(image);
expect(pixel_values.dims).toEqual([1, 3, 384, 384]);
expect(pixel_values.mean().item()).toBeCloseTo(0.0372855559389454, 6);
expect(original_sizes).toEqual([[480, 640]]);
expect(reshaped_input_sizes).toEqual([[384, 384]]);
},
MAX_TEST_EXECUTION_TIME,
);
});
// DPTImageProcessor
// - tests ensure_multiple_of
// - tests keep_aspect_ratio
// - tests bankers rounding
describe("DPTImageProcessor", () => {
const model_id = "Xenova/depth-anything-small-hf";
/** @type {DPTImageProcessor} */
let processor;
beforeAll(async () => {
processor = await AutoImageProcessor.from_pretrained(model_id);
}, MAX_PROCESSOR_LOAD_TIME);
it(
"ensure_multiple_of w/ normal rounding",
async () => {
const image = await load_cached_image("cats");
const { pixel_values, original_sizes, reshaped_input_sizes } = await processor(image);
expect(pixel_values.dims).toEqual([1, 3, 518, 686]);
expect(pixel_values.mean().item()).toBeCloseTo(0.30337387323379517, 3);
expect(original_sizes).toEqual([[480, 640]]);
expect(reshaped_input_sizes).toEqual([[518, 686]]);
},
MAX_TEST_EXECUTION_TIME,
);
it(
"ensure_multiple_of w/ bankers rounding",
async () => {
const image = await load_cached_image("checkerboard_64x32");
const { pixel_values, original_sizes, reshaped_input_sizes } = await processor(image);
// NOTE: without bankers rounding, this would be [1, 3, 266, 518]
expect(pixel_values.dims).toEqual([1, 3, 252, 518]);
expect(pixel_values.mean().item()).toBeCloseTo(0.2267402559518814, 1);
expect(original_sizes).toEqual([[32, 64]]);
expect(reshaped_input_sizes).toEqual([[252, 518]]);
},
MAX_TEST_EXECUTION_TIME,
);
});
};
| transformers.js/tests/models/dpt/test_image_processing_dpt.js/0 | {
"file_path": "transformers.js/tests/models/dpt/test_image_processing_dpt.js",
"repo_id": "transformers.js",
"token_count": 1111
} | 340 |
import { MarianTokenizer, MarianMTModel } from "../../../src/transformers.js";
import { MAX_MODEL_LOAD_TIME, MAX_TEST_EXECUTION_TIME, MAX_MODEL_DISPOSE_TIME, DEFAULT_MODEL_OPTIONS } from "../../init.js";
export default () => {
describe("MarianMTModel", () => {
const model_id = "onnx-community/tiny-random-MarianMTModel";
/** @type {MarianMTModel} */
let model;
/** @type {MarianTokenizer} */
let tokenizer;
beforeAll(async () => {
model = await MarianMTModel.from_pretrained(model_id, DEFAULT_MODEL_OPTIONS);
tokenizer = await MarianTokenizer.from_pretrained(model_id);
}, MAX_MODEL_LOAD_TIME);
it(
"batch_size=1",
async () => {
const inputs = tokenizer("hello");
const outputs = await model.generate({
...inputs,
max_length: 10,
});
expect(outputs.tolist()).toEqual([[3n, 40672n, 8358n, 32810n, 32810n, 32810n, 32810n, 35687n, 33073n, 6870n]]);
},
MAX_TEST_EXECUTION_TIME,
);
it(
"batch_size>1",
async () => {
const inputs = tokenizer(["hello", "hello world"], { padding: true });
const outputs = await model.generate({
...inputs,
max_length: 10,
});
expect(outputs.tolist()).toEqual([
[3n, 40672n, 8358n, 32810n, 32810n, 32810n, 32810n, 35687n, 33073n, 6870n],
[3n, 40672n, 8358n, 32810n, 32810n, 32810n, 32810n, 35687n, 33073n, 6870n],
]);
},
MAX_TEST_EXECUTION_TIME,
);
afterAll(async () => {
await model?.dispose();
}, MAX_MODEL_DISPOSE_TIME);
});
};
| transformers.js/tests/models/marian/test_modeling_marian.js/0 | {
"file_path": "transformers.js/tests/models/marian/test_modeling_marian.js",
"repo_id": "transformers.js",
"token_count": 753
} | 341 |
import { AutoImageProcessor, Swin2SRImageProcessor } from "../../../src/transformers.js";
import { load_cached_image } from "../../asset_cache.js";
import { MAX_PROCESSOR_LOAD_TIME, MAX_TEST_EXECUTION_TIME } from "../../init.js";
export default () => {
// Swin2SRImageProcessor
// - tests when padding is a number (do_pad=true, pad_size=8)
describe("Swin2SRImageProcessor", () => {
const model_id = "Xenova/swin2SR-classical-sr-x2-64";
/** @type {Swin2SRImageProcessor} */
let processor;
beforeAll(async () => {
processor = await AutoImageProcessor.from_pretrained(model_id);
}, MAX_PROCESSOR_LOAD_TIME);
it(
"Pad to multiple of 8 (3x3 -> 8x8)",
async () => {
const image = await load_cached_image("pattern_3x3");
const { pixel_values } = await processor(image);
expect(pixel_values.dims).toEqual([1, 3, 8, 8]);
expect(pixel_values.mean().item()).toBeCloseTo(0.5458333368102709, 6);
},
MAX_TEST_EXECUTION_TIME,
);
it(
"Do not pad if already a multiple of 8 (8x8 -> 8x8)",
async () => {
const image = await load_cached_image("checkerboard_8x8");
const { pixel_values } = await processor(image);
expect(pixel_values.dims).toEqual([1, 3, 8, 8]);
expect(pixel_values.mean().item()).toBeCloseTo(0.5, 6);
},
MAX_TEST_EXECUTION_TIME,
);
});
};
| transformers.js/tests/models/swin2sr/test_image_processing_swin2sr.js/0 | {
"file_path": "transformers.js/tests/models/swin2sr/test_image_processing_swin2sr.js",
"repo_id": "transformers.js",
"token_count": 599
} | 342 |
import { pipeline, Text2TextGenerationPipeline } from "../../src/transformers.js";
import { MAX_MODEL_LOAD_TIME, MAX_TEST_EXECUTION_TIME, MAX_MODEL_DISPOSE_TIME, DEFAULT_MODEL_OPTIONS } from "../init.js";
const PIPELINE_ID = "text2text-generation";
export default () => {
describe("Text to Text Generation", () => {
const model_id = "hf-internal-testing/tiny-random-T5ForConditionalGeneration";
/** @type {Text2TextGenerationPipeline} */
let pipe;
beforeAll(async () => {
pipe = await pipeline(PIPELINE_ID, model_id, DEFAULT_MODEL_OPTIONS);
}, MAX_MODEL_LOAD_TIME);
it("should be an instance of Text2TextGenerationPipeline", () => {
expect(pipe).toBeInstanceOf(Text2TextGenerationPipeline);
});
describe("batch_size=1", () => {
it(
"default",
async () => {
const text = "This is a test.";
const output = await pipe(text, {
max_new_tokens: 5,
});
const target = [{ generated_text: "" }];
expect(output).toEqual(target);
},
MAX_TEST_EXECUTION_TIME,
);
});
afterAll(async () => {
await pipe.dispose();
}, MAX_MODEL_DISPOSE_TIME);
});
};
| transformers.js/tests/pipelines/test_pipelines_text2text_generation.js/0 | {
"file_path": "transformers.js/tests/pipelines/test_pipelines_text2text_generation.js",
"repo_id": "transformers.js",
"token_count": 520
} | 343 |
import { RawImage, rand } from "../../src/transformers.js";
import { load_cached_image } from "../asset_cache.js";
const TEST_IMAGES = {
rgba: new RawImage(new Uint8ClampedArray([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23]), 2, 3, 4),
rgb: new RawImage(new Uint8ClampedArray([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17]), 2, 3, 3),
la: new RawImage(new Uint8ClampedArray([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11]), 2, 3, 2),
l: new RawImage(new Uint8ClampedArray([0, 1, 2, 3, 4, 5]), 2, 3, 1),
};
describe("Image utilities", () => {
describe("Padding", () => {
it("should pad image", async () => {
/** @type {RawImage} */
const padded_image = await load_cached_image("blue_image")
.then((image) => image.resize(224, 224))
.then((image) => image.pad([128, 128, 128, 128]));
expect(padded_image.size).toEqual([480, 480]);
const avg = padded_image.data.reduce((acc, val) => acc + val, 0) / padded_image.data.length;
expect(avg).toBeCloseTo((224 * 224 * 255) / (3 * 480 * 480), 6);
});
});
describe("Tensor to Image", () => {
it("should create an image from a tensor (CHW)", () => {
const tensor_chw = rand([3, 128, 256]).mul_(255).to("uint8");
const image = RawImage.fromTensor(tensor_chw);
expect(image.size).toEqual([256, 128]);
});
it("should create an image from a tensor (HWC)", () => {
const tensor_hwc = rand([128, 256, 3]).mul_(255).to("uint8");
const image = RawImage.fromTensor(tensor_hwc, "HWC");
expect(image.size).toEqual([256, 128]);
});
});
describe("Channel conversions", () => {
it("should convert RGBA to L (grayscale)", async () => {
const grayscale = TEST_IMAGES.rgba.clone().grayscale();
expect(grayscale.size).toEqual(TEST_IMAGES.rgba.size);
expect(grayscale.channels).toEqual(1);
});
it("should convert RGB to L (grayscale)", async () => {
const grayscale = TEST_IMAGES.rgb.clone().grayscale();
expect(grayscale.size).toEqual(TEST_IMAGES.rgb.size);
expect(grayscale.channels).toEqual(1);
});
it("should convert L to RGB", async () => {
const rgb = TEST_IMAGES.l.clone().rgb();
expect(rgb.size).toEqual(TEST_IMAGES.l.size);
expect(rgb.channels).toEqual(3);
});
it("should convert L to RGBA", async () => {
const rgba = TEST_IMAGES.l.clone().rgba();
expect(rgba.size).toEqual(TEST_IMAGES.l.size);
expect(rgba.channels).toEqual(4);
});
it("should convert RGB to RGBA", async () => {
const rgba = TEST_IMAGES.rgb.clone().rgba();
expect(rgba.size).toEqual(TEST_IMAGES.rgb.size);
expect(rgba.channels).toEqual(4);
});
it("should convert RGBA to RGB", async () => {
const rgb = TEST_IMAGES.rgba.clone().rgb();
expect(rgb.size).toEqual(TEST_IMAGES.rgba.size);
expect(rgb.channels).toEqual(3);
});
});
describe("putAlpha", () => {
it("should add alpha to RGB image", async () => {
const rgba = TEST_IMAGES.rgb.clone().putAlpha(TEST_IMAGES.l);
expect(rgba.size).toEqual(TEST_IMAGES.rgb.size);
expect(rgba.channels).toEqual(4);
});
it("should add alpha to RGBA image", async () => {
const rgba = TEST_IMAGES.rgba.clone().putAlpha(TEST_IMAGES.l);
expect(rgba.size).toEqual(TEST_IMAGES.rgb.size);
expect(rgba.channels).toEqual(4);
});
});
});
| transformers.js/tests/utils/image.test.js/0 | {
"file_path": "transformers.js/tests/utils/image.test.js",
"repo_id": "transformers.js",
"token_count": 1496
} | 344 |
# Benchmarks
You might want to add new benchmarks.
You will need to define a python function named `run_benchmark` in your python file and the file must be located in this `benchmark/` directory.
The expected function signature is the following:
```py
def run_benchmark(logger: Logger, branch: str, commit_id: str, commit_msg: str, num_tokens_to_generate=100):
```
## Writing metrics to the database
`MetricsRecorder` is thread-safe, in the sense of the python [`Thread`](https://docs.python.org/3/library/threading.html#threading.Thread). This means you can start a background thread to do the readings on the device measurements while not blocking the main thread to execute the model measurements.
cf [`llama.py`](./llama.py) to see an example of this in practice.
```py
from benchmarks_entrypoint import MetricsRecorder
import psycopg2
def run_benchmark(logger: Logger, branch: str, commit_id: str, commit_msg: str, num_tokens_to_generate=100):
metrics_recorder = MetricsRecorder(psycopg2.connect("dbname=metrics"), logger, branch, commit_id, commit_msg)
benchmark_id = metrics_recorder.initialise_benchmark({"gpu_name": gpu_name, "model_id": model_id})
# To collect device measurements
metrics_recorder.collect_device_measurements(
benchmark_id, cpu_util, mem_megabytes, gpu_util, gpu_mem_megabytes
)
# To collect your model measurements
metrics_recorder.collect_model_measurements(
benchmark_id,
{
"model_load_time": model_load_time,
"first_eager_forward_pass_time_secs": first_eager_fwd_pass_time,
"second_eager_forward_pass_time_secs": second_eager_fwd_pass_time,
"first_eager_generate_time_secs": first_eager_generate_time,
"second_eager_generate_time_secs": second_eager_generate_time,
"time_to_first_token_secs": time_to_first_token,
"time_to_second_token_secs": time_to_second_token,
"time_to_third_token_secs": time_to_third_token,
"time_to_next_token_mean_secs": mean_time_to_next_token,
"first_compile_generate_time_secs": first_compile_generate_time,
"second_compile_generate_time_secs": second_compile_generate_time,
"third_compile_generate_time_secs": third_compile_generate_time,
"fourth_compile_generate_time_secs": fourth_compile_generate_time,
},
)
```
| transformers/benchmark/README.md/0 | {
"file_path": "transformers/benchmark/README.md",
"repo_id": "transformers",
"token_count": 936
} | 345 |
FROM python:3.9-slim
ENV PYTHONDONTWRITEBYTECODE=1
ARG REF=main
USER root
RUN apt-get update && apt-get install -y libsndfile1-dev espeak-ng time git
RUN apt-get install -y g++ cmake
ENV UV_PYTHON=/usr/local/bin/python
RUN pip --no-cache-dir install uv
RUN uv pip install --no-cache-dir -U pip setuptools albumentations seqeval
RUN uv pip install --upgrade --no-cache-dir "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[tf-cpu,sklearn,testing,sentencepiece,tf-speech,vision]"
RUN uv pip install --no-cache-dir "protobuf==3.20.3"
RUN uv pip uninstall transformers
RUN apt-get clean && rm -rf /var/lib/apt/lists/*
| transformers/docker/examples-tf.dockerfile/0 | {
"file_path": "transformers/docker/examples-tf.dockerfile",
"repo_id": "transformers",
"token_count": 250
} | 346 |
FROM rocm/dev-ubuntu-22.04:6.2.4
LABEL maintainer="Hugging Face"
ARG DEBIAN_FRONTEND=noninteractive
ARG PYTORCH='2.6.0'
ARG TORCH_VISION='0.21.0'
ARG TORCH_AUDIO='2.6.0'
ARG ROCM='6.2.4'
RUN apt update && \
apt install -y --no-install-recommends \
libaio-dev \
git \
# These are required to build deepspeed.
python3-dev \
python-is-python3 \
rocrand-dev \
rocthrust-dev \
rocblas-dev \
hipsolver-dev \
hipsparse-dev \
hipblas-dev \
hipblaslt-dev && \
apt clean && \
rm -rf /var/lib/apt/lists/*
RUN python3 -m pip install --no-cache-dir --upgrade pip ninja "pydantic>=2.0.0"
RUN python3 -m pip uninstall -y apex torch torchvision torchaudio
RUN python3 -m pip install torch==$PYTORCH torchvision==$TORCH_VISION torchaudio==$TORCH_AUDIO --index-url https://download.pytorch.org/whl/rocm$ROCM --no-cache-dir
# Pre-build DeepSpeed, so it's be ready for testing (to avoid timeout)
RUN DS_BUILD_CPU_ADAM=1 DS_BUILD_FUSED_ADAM=1 python3 -m pip install deepspeed --global-option="build_ext" --global-option="-j8" --no-cache-dir -v --disable-pip-version-check 2>&1
ARG REF=main
WORKDIR /
# Invalidate docker cache from here if new commit is available.
ADD https://api.github.com/repos/huggingface/transformers/git/refs/heads/main version.json
RUN git clone https://github.com/huggingface/transformers && cd transformers && git checkout $REF
RUN python3 -m pip install --no-cache-dir ./transformers[accelerate,testing,sentencepiece,sklearn]
# When installing in editable mode, `transformers` is not recognized as a package.
# this line must be added in order for python to be aware of transformers.
RUN cd transformers && python3 setup.py develop
RUN python3 -c "from deepspeed.launcher.runner import main"
# Remove nvml as it is not compatible with ROCm
RUN python3 -m pip uninstall py3nvml pynvml nvidia-ml-py apex -y
# `kernels` may causes many failing tests
RUN python3 -m pip uninstall -y kernels
| transformers/docker/transformers-pytorch-deepspeed-amd-gpu/Dockerfile/0 | {
"file_path": "transformers/docker/transformers-pytorch-deepspeed-amd-gpu/Dockerfile",
"repo_id": "transformers",
"token_count": 735
} | 347 |
# التدريب الموزع باستخدام 🤗 Accelerate
مع تزايد حجم النماذج اللغوية، برز التوازي كأحد الاستراتيجيات لتدريب نماذج أكبر على أجهزة محدودة وتسريع عملية التدريب بمقدار كبير. أنشأنا في Hugging Face، قمنا بإنشاء مكتبة [ Accelerate](https://huggingface.co/docs/accelerate) لمساعدة المستخدمين على تدريب أي نموذج من Transformers بسهولة على أي نوع من الإعدادات الموزعة، سواء كان ذلك على عدة وحدات معالجة رسومات (GPUs) على جهاز واحد أو على عدة وحدات معالجة رسومات موزعة على عدة أجهزة. في هذا الدليل، تعلم كيفية تخصيص حلقة تدريب PyTorch الأصلية لتمكين التدريب في بيئة موزعة.
## الإعداد
ابدأ بتثبيت 🤗 Accelerate:
```bash
pip install accelerate
```
ثم قم باستيراد وإنشاء كائن [`~accelerate.Accelerator`]. سيقوم [`~accelerate.Accelerator`] تلقائيًا باكتشاف نوع الإعداد الموزع الخاص بك وتهيئة جميع المكونات اللازمة للتدريب. لن تحتاج إلى وضع نموذجك على جهاز بشكل معين.
```py
>>> from accelerate import Accelerator
>>> accelerator = Accelerator()
```
## الاستعداد للتسريع
الخطوة التالية هي تمرير جميع كائنات التدريب ذات الصلة إلى دالة الإعداد [`~accelerate.Accelerator.prepare`]. ويشمل ذلك DataLoaders للتدريب والتقييم، ونموذجًا ومُحَسِّنً المعاملات (optimizer):
```py
>>> train_dataloader, eval_dataloader, model, optimizer = accelerator.prepare(
... train_dataloader, eval_dataloader, model, optimizer
... )
```
## الخلفي Backward
الإضافة الأخيرة هي استبدال الدالة المعتادة `loss.backward()` في حلقة التدريب الخاصة بك بدالة [`~accelerate.Accelerator.backward`] في 🤗 Accelerate:
```py
>>> for epoch in range(num_epochs):
... for batch in train_dataloader:
... outputs = model(**batch)
... loss = outputs.loss
... accelerator.backward(loss)
... optimizer.step()
... lr_scheduler.step()
... optimizer.zero_grad()
... progress_bar.update(1)
```
كما يمكنك أن ترى في الكود التالي، فأنت بحاجة فقط إلى إضافة أربعة أسطر من الكود إلى حلقة التدريب الخاصة بك لتمكين التدريب الموزع!
```diff
+ from accelerate import Accelerator
from transformers import AdamW, AutoModelForSequenceClassification, get_scheduler
+ accelerator = Accelerator()
model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=2)
optimizer = AdamW(model.parameters(), lr=3e-5)
- device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
- model.to(device)
+ train_dataloader, eval_dataloader, model, optimizer = accelerator.prepare(
+ train_dataloader, eval_dataloader, model, optimizer
+ )
num_epochs = 3
num_training_steps = num_epochs * len(train_dataloader)
lr_scheduler = get_scheduler(
"linear",
optimizer=optimizer,
num_warmup_steps=0,
num_training_steps=num_training_steps
)
progress_bar = tqdm(range(num_training_steps))
model.train()
for epoch in range(num_epochs):
for batch in train_dataloader:
- batch = {k: v.to(device) for k, v in batch.items()}
outputs = model(**batch)
loss = outputs.loss
- loss.backward()
+ accelerator.backward(loss)
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
```
## تدريب
بمجرد إضافة أسطر الكود ذات الصلة، قم بتشغيل التدريب الخاص بك في أحد النصوص أو الدفاتر مثل Colaboratory.
### التدريب باستخدام نص برمجي
إذا كنت تشغل التدريب الخاص بك من نص برمجي، فقم بتشغيل الأمر التالي لإنشاء وحفظ ملف تكوين:
```bash
accelerate config
```
ثم قم بتشغيل التدريب الخاص بك باستخدام:
```bash
accelerate launch train.py
```
### التدريب باستخدام دفتر ملاحظات
يمكن أيضًا تشغيل 🤗 Accelerate في دفاتر إذا كنت تخطط لاستخدام وحدات معالجة الرسوميات (TPUs) في Colaboratory. قم بتغليف كل الكود المسؤول عن التدريب في دالة، ومررها إلى [`~accelerate.notebook_launcher`]:
```py
>>> from accelerate import notebook_launcher
>>> notebook_launcher(training_function)
```
للحصول على مزيد من المعلومات حول 🤗 Accelerate وميزاته الغنية، يرجى الرجوع إلى [الوثائق](https://huggingface.co/docs/accelerate). | transformers/docs/source/ar/accelerate.md/0 | {
"file_path": "transformers/docs/source/ar/accelerate.md",
"repo_id": "transformers",
"token_count": 2534
} | 348 |
# تحسين نماذج اللغة الكبيرة من حيث السرعة والذاكرة
[[open-in-colab]]
تحقق نماذج اللغة الكبيرة (LLMs) مثل GPT3/4، [Falcon](https://huggingface.co/tiiuae/falcon-40b)، و [Llama](https://huggingface.co/meta-llama/Llama-2-70b-hf) تقدمًا سريعًا في قدرتها على معالجة المهام التي تركز على الإنسان، مما يجعلها أدوات أساسية في الصناعات القائمة على المعرفة الحديثة.
لا يزال نشر هذه النماذج في المهام الواقعية يمثل تحديًا، ومع ذلك:
- لكي تظهر نماذج اللغة الكبيرة قدرات فهم وتوليد النصوص قريبة من قدرات الإنسان، فإنها تتطلب حاليًا إلى تكوينها من مليارات المعلمات (انظر [كابلان وآخرون](https://huggingface.co/papers/2001.08361)، [وي وآخرون](https://huggingface.co/papers/2206.07682)). وهذا بدوره يزيد من متطلبات الذاكرة للاستدلال.
- في العديد من المهام الواقعية، تحتاج نماذج اللغة الكبيرة إلى معلومات سياقية شاملة. يتطلب ذلك قدرة النموذج على إدارة تسلسلات إدخال طويلة للغاية أثناء الاستدلال.
يكمن جوهر صعوبة هذه التحديات في تعزيز القدرات الحسابية والذاكرة لنماذج اللغة الكبيرة، خاصة عند التعامل مع تسلسلات الإدخال الضخمة.
في هذا الدليل، سنستعرض التقنيات الفعالة لتُحسِّن من كفاءة نشر نماذج اللغة الكبيرة:
1. سنتناول تقنية "دقة أقل" التي أثبتت الأبحاث فعاليتها في تحقيق مزايا حسابية دون التأثير بشكل ملحوظ على أداء النموذج عن طريق العمل بدقة رقمية أقل [8 بت و4 بت](/main_classes/quantization).
2. **اFlash Attention:** إن Flash Attention وهي نسخة مُعدَّلة من خوارزمية الانتباه التي لا توفر فقط نهجًا أكثر كفاءة في استخدام الذاكرة، ولكنها تحقق أيضًا كفاءة متزايدة بسبب الاستخدام الأمثل لذاكرة GPU.
3. **الابتكارات المعمارية:** حيث تم اقتراح هياكل متخصصة تسمح باستدلال أكثر فعالية نظرًا لأن نماذج اللغة الكبيرة يتم نشرها دائمًا بنفس الطريقة أثناء عملية الاستدلال، أي توليد النص التنبؤي التلقائي مع سياق الإدخال الطويل، فقد تم اقتراح بنيات نموذج متخصصة تسمح بالاستدلال الأكثر كفاءة. أهم تقدم في بنيات النماذج هنا هو [عذر](https://huggingface.co/papers/2108.12409)، [الترميز الدوار](https://huggingface.co/papers/2104.09864)، [الاهتمام متعدد الاستعلامات (MQA)](https://huggingface.co/papers/1911.02150) و [مجموعة الانتباه بالاستعلام (GQA)](https://huggingface.co/papers/2305.13245).
على مدار هذا الدليل، سنقدم تحليلًا للتوليد التنبؤي التلقائي من منظور المُوتِّرات. نتعمق في مزايا وعيوب استخدام دقة أقل، ونقدم استكشافًا شاملاً لخوارزميات الانتباه الأحدث، ونناقش بنيات نماذج نماذج اللغة الكبيرة المحسنة. سندعم الشرح بأمثلة عملية تُبرِز كل تحسين على حدة.
## 1. دقة أقل
يمكن فهم متطلبات ذاكرة نماذج اللغة الكبيرة بشكل أفضل من خلال النظر إلى نموذج اللغة الكبيرة على أنها مجموعة من المصفوفات والمتجهات الوزنية، ومدخلات النص على أنها تسلسل من المتجهات. فيما يلي، سيتم استخدام تعريف "الأوزان" للإشارة إلى جميع مصفوفات الأوزان والمتجهات في النموذج.
في وقت كتابة هذا الدليل، تتكون نماذج اللغة الكبيرة من مليارات المعلمات على الأقل.كل معلمة يتم تمثيلها برقم عشري مثل 4.5689 `` والذي يتم تخزينه عادةً بتنسيق [float32](https://en.wikipedia.org/wiki/Single-precision_floating-point_format)، [bfloat16](https://en.wikipedia.org/wiki/Bfloat16_floating-point_format)، أو [float16](https://en.wikipedia.org/wiki/Half-precision_floating-point_format) . يسمح لنا هذا بحساب متطلبات الذاكرة لتحميل نموذج اللغة الكبيرة في الذاكرة بسهولة:
> *يتطلب تحميل أوزان نموذج به X مليار معلمة حوالي 4 * X جيجابايت من ذاكرة الفيديو العشوائية (VRAM) بدقة float32*
ومع ذلك، نادرًا ما يتم تدريب النماذج في الوقت الحالي بدقة float32 الكاملة، ولكن عادةً ما تكون بدقة bfloat16 أو بشكل أقل في تنسيق float16. لذلك، تصبح القاعدة الإرشادية كما يلي:
> *يتطلب تحميل أوزان نموذج به X مليار معلمة حوالي 2 * X جيجابايت من ذاكرة الفيديو العشوائية (VRAM) بدقة bfloat16/float16*
بالنسبة لمدخلات النصوص القصيرة (أقل من 1024 رمزًا)، فإن متطلبات الذاكرة للاستدلال تهيمن عليها إلى حد كبير متطلبات الذاكرة لتحميل الأوزان. لذلك، دعنا نفترض، في الوقت الحالي، أن متطلبات الذاكرة للاستدلال تساوي متطلبات الذاكرة لتحميل النموذج في ذاكرة VRAM لوحدة معالجة الرسومات GPU..
ولإعطاء بعض الأمثلة على مقدار ذاكرة الفيديو العشوائية (VRAM) التي يتطلبها تحميل نموذج بتنسيق bfloat16 تقريبًا:
- **GPT3** يتطلب 2 \* 175 جيجا بايت = **350 جيجا بايت** VRAM
- [**بلوم**](https://huggingface.co/bigscience/bloom) يتطلب 2 \* 176 جيجا بايت = **352 جيجا بايت** VRAM
- [**Llama-2-70b**](https://huggingface.co/meta-llama/Llama-2-70b-hf) يتطلب 2 \* 70 جيجا بايت = **140 جيجا بايت** VRAM
- [**Falcon-40b**](https://huggingface.co/tiiuae/falcon-40b) يتطلب 2 \* 40 جيجا بايت = **80 جيجا بايت** VRAM
- [**MPT-30b**](https://huggingface.co/mosaicml/mpt-30b) يتطلب 2 \* 30 جيجا بايت = **60 جيجا بايت** VRAM
- [**bigcode/starcoder**](https://huggingface.co/bigcode/starcoder) يتطلب 2 \* 15.5 = **31 جيجا بايت** VRAM
عند كتابة هذا الدليل، أكبر شريحة لوحدة معالجة الرسومات المتوفّرة هي A100 و H100 التي توفر 80 جيجابايت من ذاكرة الفيديو العشوائية (VRAM). تتطلب معظم النماذج المدرجة أعلاه أكثر من 80 جيجابايت فقط لتحميلها، وبالتالي فهي تتطلب بالضرورة [التوازي للموتّرات](https://huggingface.co/docs/transformers/perf_train_gpu_many#tensor-parallelism) و/أو [لتوازي الخطي](https://huggingface.co/docs/transformers/perf_train_gpu_many#naive-model-parallelism-vertical-and-pipeline-parallelism).
🤗 لا يدعم Transformers موازاة التنسور خارج الصندوق لأنه يتطلب كتابة هيكلة النموذج بطريقة محددة. إذا كنت مهتمًا بكتابة نماذج بطريقة صديقة لموازاة التنسور، فلا تتردد في إلقاء نظرة على [مكتبة الاستدلال بتوليد النص](https://github.com/huggingface/text-generation-inference/tree/main/server/text_generation_server/models/custom_modeling).
بدعم موازاة قنوات المعالجة البسيطة خارج الصندوق. للقيام بذلك، قم بتحميل النموذج باستخدام `device="auto"` والذي سيقوم تلقائيًا بوضع الطبقات المختلفة على وحدات معالجة الرسومات (GPU) المتاحة كما هو موضح [هنا](https://huggingface.co/docs/accelerate/v0.22.0/en/concept_guides/big_model_inference).
لاحظ، مع ذلك، أنه في حين أن موازاة قنوات المعالجة البسيطة فعالة للغاية، إلا أنها لا تعالج مشكلات عدم نشاط وحدة معالجة الرسومات (GPU). لهذا، تكون موازاة قنوات المعالجة المتقدمة مطلوبة كما هو موضح [هنا](https://huggingface.co/docs/transformers/en/perf_train_gpu_many#naive-model-parallelism-vertical-and-pipeline-parallelism).
إذا كان لديك حق الوصول إلى عقدة 8 x 80 جيجابايت A100، فيمكنك تحميل BLOOM كما يلي
```bash
!pip install transformers accelerate bitsandbytes optimum
```
```python
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("bigscience/bloom", device_map="auto", pad_token_id=0)
```
من خلال استخدام `device_map="auto"` سيتم توزيع طبقات الاهتمام بالتساوي عبر جميع وحدات معالجة الرسومات (GPU) المتاحة.
في هذا الدليل، سنستخدم [bigcode/octocoder](https://huggingface.co/bigcode/octocoder) لأنه يمكن تشغيله على شريحة جهاز GPU A100 ذات 40 جيجا بايت. لاحظ أن جميع تحسينات الذاكرة والسرعة التي سنطبقها من الآن فصاعدًا تنطبق بالتساوي على النماذج التي تتطلب موازاة النماذج أو المصفوفات.
نظرًا لأن النموذج مُحمَّل بدقة bfloat16، فباستخدام قاعدتنا الإرشادية أعلاه، نتوقع أن تكون متطلبات الذاكرة لتشغيل الاستدلال باستخدام `bigcode/octocoder` حوالي 31 جيجا بايت من ذاكرة الفيديو العشوائية (VRAM). دعنا نجرب.
نقوم أولاً بتحميل النموذج والمجزىء اللغوي ثم نقوم بتمرير كلاهما إلى كائن [قنوات المعالجة](https://huggingface.co/docs/transformers/main_classes/pipelines) في Transformers.
```python
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
import torch
model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", dtype=torch.bfloat16, device_map="auto", pad_token_id=0)
tokenizer = AutoTokenizer.from_pretrained("bigcode/octocoder")
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
```
```python
prompt = "Question: Please write a function in Python that transforms bytes to Giga bytes.\n\nAnswer:"
result = pipe(prompt, max_new_tokens=60)[0]["generated_text"][len(prompt):]
result
```
**الإخراج**:
```
Here is a Python function that transforms bytes to Giga bytes:\n\n```python\ndef bytes_to_giga_bytes(bytes):\n return bytes / 1024 / 1024 / 1024\n```\n\nThis function takes a single
```
رائع، يمكننا الآن استخدام النتيجة مباشرة لتحويل البايت إلى جيجا بايت.
```python
def bytes_to_giga_bytes(bytes):
return bytes / 1024 / 1024 / 1024
```
دعونا نستدعي [`torch.cuda.max_memory_allocated`](https://pytorch.org/docs/stable/generated/torch.cuda.max_memory_allocated.html) لقياس ذروة تخصيص ذاكرة وحدة معالجة الرسومات (GPU).
```python
bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
```
**الإخراج**:
```bash
29.0260648727417
```
قريب بما يكفي من حسابنا التقريبي! يمكننا أن نرى أن الرقم غير صحيح تمامًا لأن الانتقال من البايت إلى الكيلوبايت يتطلب الضرب في 1024 بدلاً من 1000. لذلك يمكن أيضًا فهم صيغة التقريب على أنها حساب "بحد أقصى X جيجا بايت".
لاحظ أنه إذا حاولنا تشغيل النموذج بدقة float32 الكاملة، فستكون هناك حاجة إلى 64 جيجا بايت من ذاكرة الفيديو العشوائية (VRAM).
> يتم تدريب جميع النماذج تقريبًا بتنسيق bfloat16 في الوقت الحالي، ولا يوجد سبب لتشغيل النموذج بدقة float32 الكاملة إذا [كانت وحدة معالجة الرسومات (GPU) الخاصة بك تدعم bfloat16](https://discuss.pytorch.org/t/bfloat16-native-support/117155/5). لن توفر دقة float32 نتائج استدلال أفضل من الدقة التي تم استخدامها لتدريب النموذج.
إذا لم تكن متأكدًا من تنسيق تخزين أوزان النموذج على Hub، فيمكنك دائمًا الاطلاع على تهيئة نقطة التفتيش في `"dtype"`، على سبيل المثال [هنا](https://huggingface.co/meta-llama/Llama-2-7b-hf/blob/6fdf2e60f86ff2481f2241aaee459f85b5b0bbb9/config.json#L21). يوصى بتعيين النموذج إلى نفس نوع الدقة كما هو مكتوب في التهيئة عند التحميل باستخدام `from_pretrained(..., dtype=...)` إلا إذا كان النوع الأصلي هو float32، وفي هذه الحالة يمكن استخدام `float16` أو `bfloat16` للاستدلال.
دعونا نحدد وظيفة `flush(...)` لتحرير جميع الذاكرة المخصصة بحيث يمكننا قياس ذروة ذاكرة وحدة معالجة الرسومات (GPU) المخصصة بدقة.
```python
del pipe
del model
import gc
import torch
def flush():
gc.collect()
torch.cuda.empty_cache()
torch.cuda.reset_peak_memory_stats()
```
دعونا نستدعيه الآن للتجربة التالية.
```python
flush()
```
في الإصدار الأخير من مكتبة Accelerate، يمكنك أيضًا استخدام طريقة مساعدة تسمى `release_memory()`
```python
from accelerate.utils import release_memory
# ...
release_memory(model)
```
```python
from accelerate.utils import release_memory
# ...
release_memory(model)
```
والآن ماذا لو لم يكن لدى وحدة معالجة الرسومات (GPU) لديك 32 جيجا بايت من ذاكرة الفيديو العشوائية (VRAM)؟ لقد وجد أن أوزان النماذج يمكن تحويلها إلى 8 بتات أو 4 بتات دون خسارة كبيرة في الأداء (انظر [Dettmers et al.](https://huggingface.co/papers/2208.07339)).
يمكن تحويل النموذج إلى 3 بتات أو 2 بتات مع فقدان مقبول في الأداء كما هو موضح في ورقة [GPTQ](https://huggingface.co/papers/2210.17323) 🤯.
دون الدخول في الكثير من التفاصيل، تهدف مخططات التكميم إلى تخفيض دقة الأوزان مع محاولة الحفاظ على دقة نتائج النموذج كما هي (*أي* أقرب ما يمكن إلى bfloat16).
لاحظ أن التكميم يعمل بشكل خاص جيدًا لتوليد النص حيث كل ما نهتم به هو اختيار *مجموعة الرموز الأكثر احتمالًا التالية* ولا نهتم حقًا بالقيم الدقيقة لتوزيع الرمز التالي *logit*.
كل ما يهم هو أن توزيع الرمز التالي *logit* يظل كما هو تقريبًا بحيث تعطي عملية `argmax` أو `topk` نفس النتائج.
هناك عدة تقنيات للتكميم، والتي لن نناقشها بالتفصيل هنا، ولكن بشكل عام، تعمل جميع تقنيات التكميم كما يلي:
- 1. تكميم جميع الأوزان إلى الدقة المستهدفة
- 2. تحميل الأوزان المحولة، ومرر تسلسل المدخلات من المتجهات بتنسيق bfloat16
- 3. قم بتحويل الأوزان ديناميكيًا إلى bfloat1 لإجراء الحسابات مع متجهات المدخلات بدقة `bfloat16`
باختصار، هذا يعني أن مضروبات *مصفوفة المدخلات-الوزن*، حيث \\( X \\) هي المدخلات، \\( W \\) هي مصفوفة وزن و \\( Y \\) هي الناتج:
$$ Y = X * W $$
تتغير إلى
$$ Y = X * \text{dequantize}(W) $$
لكل عملية ضرب المصفوفات. يتم تنفيذ إلغاء التكميم وإعادة التكميم بشكل متسلسل لجميع مصفوفات الأوزان أثناء مرور المدخلات عبر رسم الشبكة.
لذلك، غالبًا ما لا يتم تقليل وقت الاستدلال عند استخدام الأوزان المكممة، ولكن بدلاً من ذلك يزيد.
كفى نظرية، دعنا نجرب! لتكميم الأوزان باستخدام المحولات، تحتاج إلى التأكد من تثبيت مكتبة [`bitsandbytes`](https://github.com/TimDettmers/bitsandbytes).
```bash
!pip install bitsandbytes
```
يمكننا بعد ذلك تحميل النماذج في تكميم 8 بت ببساطة عن طريق إضافة علامة `load_in_8bit=True` إلى `from_pretrained`.
```python
model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", load_in_8bit=True, pad_token_id=0)
```
الآن، دعنا نعيد تشغيل مثالنا ونقيس استخدام الذاكرة.
```python
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
result = pipe(prompt, max_new_tokens=60)[0]["generated_text"][len(prompt):]
result
```
**الإخراج**:
```
Here is a Python function that transforms bytes to Giga bytes:\n\n```python\ndef bytes_to_giga_bytes(bytes):\n return bytes / 1024 / 1024 / 1024\n```\n\nThis function takes a single
```
جميل، نحصل على نفس النتيجة كما في السابق، لذلك لا يوجد فقدان في الدقة! دعنا نلقي نظرة على مقدار الذاكرة المستخدمة هذه المرة.
```python
bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
```
**الإخراج**:
```
15.219234466552734
```
أقل بكثير! لقد انخفضنا إلى ما يزيد قليلاً عن 15 جيجابايت، وبالتالي يمكننا تشغيل هذا النموذج على وحدات معالجة الرسومات للمستهلك مثل 4090.
نرى مكسبًا لطيفًا جدًا في كفاءة الذاكرة ولا يوجد تقريبًا أي تدهور في ناتج النموذج. ومع ذلك، يمكننا أيضًا ملاحظة تباطؤ طفيف أثناء الاستدلال.
نحذف النماذج ونفرغ الذاكرة مرة أخرى.
```python
del model
del pipe
```
```python
flush()
```
دعنا نرى ما هو استهلاك ذاكرة GPU الذروة الذي يوفره تكميم 4 بت. يمكن تكميم النموذج إلى 4 بت باستخدام نفس واجهة برمجة التطبيقات كما في السابق - هذه المرة عن طريق تمرير `load_in_4bit=True` بدلاً من `load_in_8bit=True`.
```python
model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", load_in_4bit=True, pad_token_id=0)
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
result = pipe(prompt, max_new_tokens=60)[0]["generated_text"][len(prompt):]
result
```
**الإخراج**:
```
Here is a Python function that transforms bytes to Giga bytes:\n\n```\ndef bytes_to_gigabytes(bytes):\n return bytes / 1024 / 1024 / 1024\n```\n\nThis function takes a single argument
```
نحن نرى تقريبًا نفس نص الإخراج كما في السابق - فقط `python` مفقود قبل مقطع الكود. دعنا نرى مقدار الذاكرة المطلوبة.
```python
bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
```
**الإخراج**:
```
9.543574333190918
```
فقط 9.5 جيجابايت! هذا ليس كثيرًا بالفعل لنموذج يزيد عدد معاملاته عن 15 مليار.
على الرغم من أننا نرى تدهورًا بسيطًا جدًا في الدقة لنموذجنا هنا، إلا أن تكميم 4 بت يمكن أن يؤدي في الممارسة العملية غالبًا إلى نتائج مختلفة مقارنة بتكميم 8 بت أو الاستدلال الكامل `bfloat16`. الأمر متروك للمستخدم لتجربته.
لاحظ أيضًا أن الاستدلال هنا كان أبطأ قليلاً مقارنة بتكميم 8 بت والذي يرجع إلى طريقة التكميم الأكثر عدوانية المستخدمة لتكميم 4 بت مما يؤدي إلى \\( \text{quantize} \\) و \\( \text{dequantize} \\) يستغرق وقتًا أطول أثناء الاستدلال.
```python
del model
del pipe
```
```python
flush()
```
بشكل عام، رأينا أن تشغيل OctoCoder بدقة 8 بت قلل من ذاكرة GPU VRAM المطلوبة من 32G GPU VRAM إلى 15 جيجابايت فقط، وتشغيل النموذج بدقة 4 بت يقلل من ذاكرة GPU VRAM المطلوبة إلى ما يزيد قليلاً عن 9 جيجابايت.
يسمح تكميم 4 بت بتشغيل النموذج على وحدات معالجة الرسومات مثل RTX3090 و V100 و T4 والتي يمكن الوصول إليها بسهولة لمعظم الأشخاص.
لمزيد من المعلومات حول التكميم ولمعرفة كيف يمكن تكميم النماذج لطلب ذاكرة GPU VRAM أقل حتى من 4 بت، نوصي بالاطلاع على تنفيذ [`AutoGPTQ`](https://huggingface.co/docs/transformers/main/en/main_classes/quantization#autogptq-integration%60).
> كاستنتاج، من المهم تذكر أن تكميم النموذج يتداول كفاءة الذاكرة المحسنة مقابل الدقة وفي بعض الحالات وقت الاستدلال.
إذا لم تكن ذاكرة GPU قيدًا لحالتك الاستخدام، فغالبًا لا توجد حاجة للنظر في التكميم. ومع ذلك، لا يمكن للعديد من وحدات معالجة الرسومات ببساطة تشغيل نماذج اللغة الكبيرة بدون طرق التكميم وفي هذه الحالة، تعد مخططات التكميم 4 بت و 8 بت أدوات مفيدة للغاية.
لمزيد من المعلومات حول الاستخدام التفصيلي، نوصي بشدة بإلقاء نظرة على [وثائق تكميم المحولات](https://huggingface.co/docs/transformers/main_classes/quantization#general-usage).
بعد ذلك، دعنا نلقي نظرة على كيفية تحسين الكفاءة الحسابية وكفاءة الذاكرة باستخدام خوارزميات أفضل وبنية نموذج محسنة.
## 2. الانتباه السريع
تتشارك نماذج اللغة الكبيرة (LLMs) الأعلى أداءً اليوم تقريبًا نفس البنية الأساسية التي تتكون من طبقات التغذية الأمامية، وطبقات التنشيط، وطبقات التطبيع الطبقي، والأهم من ذلك، طبقات الانتباه الذاتي.
تعد طبقات الانتباه الذاتي مركزية لنماذج اللغة الكبيرة (LLMs) حيث تمكن النموذج من فهم العلاقات السياقية بين رموز المدخلات.
ومع ذلك، فإن استهلاك ذاكرة GPU الذروة لطبقات الانتباه الذاتي ينمو بشكل *رباعي* في كل من التعقيد الحسابي وتعقيد الذاكرة مع عدد رموز المدخلات (والذي يُطلق عليه أيضًا *طول التسلسل*) الذي نسميه في ما يلي بـ \\( N \\) .
على الرغم من أن هذا غير ملحوظ حقًا للتسلسلات الأقصر (حتى 1000 رمز إدخال)، إلا أنه يصبح مشكلة خطيرة للتسلسلات الأطول (حوالي 16000 رمز إدخال).
دعنا نلقي نظرة أقرب. الصيغة لحساب الناتج \\( \mathbf{O} \\) لطبقة الانتباه الذاتي لإدخال \\( \mathbf{X} \\) بطول \\( N \\) هي:
$$ \textbf{O} = \text{Attn}(\mathbf{X}) = \mathbf{V} \times \text{Softmax}(\mathbf{QK}^T) \text{ with } \mathbf{Q} = \mathbf{W}_q \mathbf{X}, \mathbf{V} = \mathbf{W}_v \mathbf{X}, \mathbf{K} = \mathbf{W}_k \mathbf{X} $$
يعد \\( \mathbf{X} = (\mathbf{x}_1, ... \mathbf{x}_{N}) \\) بالتالي تسلسل الإدخال إلى طبقة الاهتمام. وستتكون كل من الإسقاطات \\( \mathbf{Q} \\) و \\( \mathbf{K} \\) من \\( N \\) من المتجهات مما يؤدي إلى أن يكون حجم \\( \mathbf{QK}^T \\) هو \\( N^2 \\).
عادة ما يكون لدى LLMs العديد من رؤوس الاهتمام، وبالتالي يتم إجراء العديد من حسابات الاهتمام الذاتي بالتوازي.
وبافتراض أن LLM لديها 40 رأس اهتمام وتعمل بدقة bfloat16، يمكننا حساب متطلبات الذاكرة لتخزين مصفوفات \\( \mathbf{QK^T} \\) لتكون \\( 40 * 2 * N^2 \\) بايت. بالنسبة لـ \\( N=1000 \\)، لا يلزم سوى حوالي 50 ميجابايت من VRAM، ولكن بالنسبة لـ \\( N=16000 \\) سنحتاج إلى 19 جيجابايت من VRAM، وبالنسبة لـ \\( N=100,000 \\) سنحتاج إلى ما يقرب من 1 تيرابايت فقط لتخزين مصفوفات \\( \mathbf{QK}^T \\).
باختصار، سرعان ما يصبح خوارزمية الانتباه الذاتي الافتراضية مكلفة للغاية من حيث الذاكرة بالنسبة لسياقات الإدخال الكبيرة.
مع تحسن LLMs في فهم النص وتوليد النص، يتم تطبيقها على مهام متزايدة التعقيد. في حين أن النماذج كانت تتعامل سابقًا مع ترجمة أو تلخيص بضع جمل، فإنها الآن تدير صفحات كاملة، مما يتطلب القدرة على معالجة أطوال إدخال واسعة.
كيف يمكننا التخلص من متطلبات الذاكرة الباهظة للتطويلات المدخلة الكبيرة؟ نحن بحاجة إلى طريقة جديدة لحساب آلية الاهتمام الذاتي التي تتخلص من مصفوفة \\( QK^T \\). [طريقه داو وآخرون.](https://huggingface.co/papers/2205.14135) طوروا بالضبط مثل هذا الخوارزمية الجديدة وأطلقوا عليها اسم **Flash Attention**.
باختصار، يكسر الاهتمام الفلاشي حساب \\( \mathbf{V} \times \operatorname{Softmax}(\mathbf{QK}^T\\)) ويحسب بدلاً من ذلك قطعًا أصغر من الإخراج عن طريق التكرار عبر العديد من خطوات حساب Softmax:
$$ \textbf{O}_i \leftarrow s^a_{ij} * \textbf{O}_i + s^b_{ij} * \mathbf{V}_{j} \times \operatorname{Softmax}(\mathbf{QK}^T_{i,j}) \text{ for multiple } i, j \text{ iterations } $$
مع \\( s^a_{ij} \\) و \\( s^b_{ij} \\) كونها بعض إحصائيات التطبيع softmax التي يجب إعادة حسابها لكل \\( i \\) و \\( j \\).
يرجى ملاحظة أن Flash Attention بالكامل أكثر تعقيدًا إلى حد ما ويتم تبسيطه بشكل كبير هنا حيث أن التعمق كثيرًا يخرج عن نطاق هذا الدليل. القارئ مدعو لإلقاء نظرة على ورقة Flash Attention المكتوبة جيدًا [1] لمزيد من التفاصيل.
الفكرة الرئيسية هنا هي:
> من خلال تتبع إحصائيات التطبيع softmax واستخدام بعض الرياضيات الذكية، يعطي Flash Attention **مخرجات متطابقة رقميًا** مقارنة بطبقة الاهتمام الذاتي الافتراضية بتكلفة ذاكرة لا تزيد خطيًا مع \\( N \\).
عند النظر إلى الصيغة، قد يقول المرء بديهيًا أن الاهتمام الفلاشي يجب أن يكون أبطأ بكثير مقارنة بصيغة الاهتمام الافتراضية حيث يلزم إجراء المزيد من الحسابات. في الواقع، يتطلب Flash Attention المزيد من عمليات الفاصلة العائمة مقارنة بالاهتمام العادي حيث يجب إعادة حساب إحصائيات التطبيع softmax باستمرار (راجع [الورقة](https://huggingface.co/papers/2205.14135) لمزيد من التفاصيل إذا كنت مهتمًا)
> ومع ذلك، فإن الاهتمام الفلاشي أسرع بكثير في الاستدلال مقارنة بالاهتمام الافتراضي الذي يأتي من قدرته على تقليل الطلبات على ذاكرة GPU الأبطأ ذات النطاق الترددي العالي (VRAM)، والتركيز بدلاً من ذلك على ذاكرة SRAM الأسرع الموجودة على الشريحة.
من الناحية الأساسية، يتأكد Flash Attention من إمكانية إجراء جميع عمليات الكتابة والقراءة الوسيطة باستخدام ذاكرة SRAM السريعة الموجودة على الشريحة بدلاً من الاضطرار إلى الوصول إلى ذاكرة VRAM الأبطأ لحساب متجه الإخراج \\( \mathbf{O} \\).
من الناحية العملية، لا يوجد حاليًا أي سبب **عدم** استخدام الاهتمام الفلاشي إذا كان متاحًا. الخوارزمية تعطي نفس المخرجات رياضيا، وأسرع وأكثر كفاءة في استخدام الذاكرة.
لنلقِ نظرة على مثال عملي.
يحصل نموذج OctoCoder الخاص بنا الآن على موجه إدخال أطول بشكل كبير يتضمن ما يسمى *موجه النظام*. تُستخدم موجهات النظام لتوجيه LLM إلى مساعد أفضل مصمم لمهام المستخدمين.
فيما يلي، نستخدم موجه النظام الذي سيجعل OctoCoder مساعد ترميز أفضل.
```python
system_prompt = """Below are a series of dialogues between various people and an AI technical assistant.
The assistant tries to be helpful, polite, honest, sophisticated, emotionally aware, and humble but knowledgeable.
The assistant is happy to help with code questions and will do their best to understand exactly what is needed.
It also tries to avoid giving false or misleading information, and it caveats when it isn't entirely sure about the right answer.
That said, the assistant is practical really does its best, and doesn't let caution get too much in the way of being useful.
The Starcoder models are a series of 15.5B parameter models trained on 80+ programming languages from The Stack (v1.2) (excluding opt-out requests).
The model uses Multi Query Attention, was trained using the Fill-in-the-Middle objective, and with 8,192 tokens context window for a trillion tokens of heavily deduplicated data.
-----
Question: Write a function that takes two lists and returns a list that has alternating elements from each input list.
Answer: Sure. Here is a function that does that.
def alternating(list1, list2):
results = []
for i in range(len(list1)):
results.append(list1[i])
results.append(list2[i])
return results
Question: Can you write some test cases for this function?
Answer: Sure, here are some tests.
assert alternating([10, 20, 30], [1, 2, 3]) == [10, 1, 20, 2, 30, 3]
assert alternating([True, False], [4, 5]) == [True, 4, False, 5]
assert alternating([], []) == []
Question: Modify the function so that it returns all input elements when the lists have uneven length. The elements from the longer list should be at the end.
Answer: Here is the modified function.
def alternating(list1, list2):
results = []
for i in range(min(len(list1), len(list2))):
results.append(list1[i])
results.append(list2[i])
if len(list1) > len(list2):
results.extend(list1[i+1:])
else:
results.extend(list2[i+1:])
return results
-----
"""
```
لأغراض التوضيح، سنكرر موجه النظام عشر مرات بحيث يكون طول الإدخال طويلاً بما يكفي لملاحظة وفورات ذاكرة Flash Attention.
نضيف موجه النص الأصلي "سؤال: يرجى كتابة وظيفة في Python تقوم بتحويل البايتات إلى جيجا بايت.
```python
long_prompt = 10 * system_prompt + prompt
```
نقوم بتنفيذ نموذجنا مرة أخرى بدقة bfloat16.
```python
model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("bigcode/octocoder")
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
```
دعنا الآن نقوم بتشغيل النموذج تمامًا مثلما كان من قبل *بدون اهتمام فلاشي* وقياس متطلبات ذاكرة GPU وقت الذروة ووقت الاستدلال.
```python
import time
start_time = time.time()
result = pipe(long_prompt, max_new_tokens=60)[0]["generated_text"][len(long_prompt):]
print(f"Generated in {time.time() - start_time} seconds.")
result
```
**الإخراج**:
```
تم التوليد في 10.96854019165039 ثانية.
بالتأكيد. إليك وظيفة للقيام بذلك.
def bytes_to_giga(bytes):
return bytes / 1024 / 1024 / 1024
الإجابة: بالتأكيد. إليك وظيفة للقيام بذلك.
ديف
```
نحصل على نفس الإخراج كما كان من قبل، ولكن هذه المرة، يقوم النموذج بتكرار الإجابة عدة مرات حتى يتم قطعها عند 60 رمزًا. ليس من المستغرب أننا كررنا موجه النظام عشر مرات لأغراض التوضيح وبالتالي قمنا بتشغيل النموذج لتكرار نفسه.
**ملاحظة** لا ينبغي تكرار موجه النظام عشر مرات في التطبيقات الواقعية - مرة واحدة كافية!
دعنا نقيس متطلبات ذاكرة GPU وقت الذروة.
```python
bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
```
**الإخراج**:
```
37.668193340301514
```
كما نرى، فإن متطلبات ذاكرة GPU وقت الذروة أعلى بكثير مما كانت عليه في البداية، وهو ما يرجع إلى حد كبير إلى تسلسل الإدخال الأطول. أيضًا، يستغرق التوليد أكثر من دقيقة بقليل الآن.
نستدعي `flush()` لتحرير ذاكرة GPU لتجربتنا التالية.
```python
flush()
```
لمقارنة، دعونا نقوم بتشغيل نفس الدالة، ولكن تمكين الاهتمام فلاش بدلا من ذلك.
للقيام بذلك، نقوم بتحويل النموذج إلى [BetterTransformer](Https://huggingface.co/docs/optimum/bettertransformer/overview) ومن خلال القيام بذلك تمكين PyTorch's [SDPA self-attention](Https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention) والتي بدورها قادرة على استخدام الاهتمام فلاش.
```python
model.to_bettertransformer()
```
الآن نقوم بتشغيل نفس مقتطف التعليمات البرمجية بالضبط كما كان من قبل وتحت الغطاء سوف تستخدم المحولات الاهتمام فلاش.
```py
start_time = time.time()
with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
result = pipe(long_prompt, max_new_tokens=60)[0]["generated_text"][len(long_prompt):]
print(f"Generated in {time.time() - start_time} seconds.")
result
```
**الإخراج**:
```
تم التوليد في 3.0211617946624756 ثانية.
بالتأكيد. إليك وظيفة للقيام بذلك.
def bytes_to_giga(bytes):
return bytes / 1024 / 1024 / 1024
الإجابة: بالتأكيد. إليك وظيفة للقيام بذلك.
ديف
```
نحصل على نفس النتيجة بالضبط كما كان من قبل، ولكن يمكننا ملاحظة تسريع كبير بفضل الاهتمام فلاش.
دعنا نقيس استهلاك الذاكرة لآخر مرة.
```python
bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
```
**الإخراج**:
```
32.617331981658936
```
ونحن تقريبا مرة أخرى إلى ذاكرة GPU الذروة الأصلية لدينا 29GB.
يمكننا أن نلاحظ أننا نستخدم فقط حوالي 100 ميجابايت إضافية من ذاكرة GPU عند تمرير تسلسل إدخال طويل جدًا مع الاهتمام فلاش مقارنة بتمرير تسلسل إدخال قصير كما فعلنا في البداية.
```py
flush()
```
لمزيد من المعلومات حول كيفية استخدام Flash Attention، يرجى الاطلاع على [صفحة doc هذه](Https://huggingface.co/docs/transformers/en/perf_infer_gpu_one#flashattention-2).
## 3. الابتكارات المعمارية
حتى الآن، نظرنا في تحسين الكفاءة الحسابية والذاكرة من خلال:
- صب الأوزان في تنسيق دقة أقل
- استبدال خوارزمية الاهتمام الذاتي بإصدار أكثر كفاءة من حيث الذاكرة والحساب
دعونا الآن نلقي نظرة على كيفية تغيير بنية LLM بحيث تكون أكثر فعالية وكفاءة للمهام التي تتطلب مدخلات نصية طويلة، على سبيل المثال:
- استرجاع الأسئلة المعززة،
- تلخيص،
- الدردشة
لاحظ أن "الدردشة" لا تتطلب من LLM التعامل مع مدخلات نصية طويلة فحسب، بل تتطلب أيضًا أن يكون LLM قادرًا على التعامل بكفاءة مع الحوار ذهابًا وإيابًا بين المستخدم والمساعد (مثل ChatGPT).
بمجرد تدريبها، يصبح من الصعب تغيير بنية LLM الأساسية، لذلك من المهم مراعاة مهام LLM مسبقًا وتحسين بنية النموذج وفقًا لذلك.
هناك مكونان مهمان لبنية النموذج يصبحان بسرعة عنق زجاجة للذاكرة و/أو الأداء لتسلسلات الإدخال الكبيرة.
- الترميزات الموضعية
- ذاكرة التخزين المؤقت للقيمة الرئيسية
دعنا نلقي نظرة على كل مكون بمزيد من التفاصيل
### 3.1 تحسين الترميزات الموضعية لـ LLMs
يضع الاهتمام الذاتي كل رمز في علاقة مع رموز أخرى.
كمثال، يمكن أن تبدو مصفوفة \\( \operatorname{Softmax}(\mathbf{QK}^T) \\) لتسلسل الإدخال النصي *"مرحبًا"، "أنا"، "أحب"، "أنت"* كما يلي:
يتم منح كل رمز كلمة كتلة احتمال يتم من خلالها الاهتمام بجميع رموز الكلمات الأخرى، وبالتالي يتم وضعها في علاقة مع جميع رموز الكلمات الأخرى. على سبيل المثال، تحضر كلمة *"الحب"* كلمة *"مرحبًا"* بنسبة 5%، و *"أنا"* بنسبة 30%، ونفسها بنسبة 65%.
سيواجه LLM القائم على الاهتمام الذاتي، ولكن بدون الترميزات الموضعية، صعوبات كبيرة في فهم مواضع نصوص الإدخال بالنسبة لبعضها البعض.
ويرجع ذلك إلى أن درجة الاحتمال التي يحسبها \\( \mathbf{QK}^T \\) تربط كل رمز كلمة بكل رمز كلمة أخرى في حسابات \\( O (1) \\) بغض النظر عن مسافة الموضع النسبي بينهما.
لذلك، بالنسبة إلى LLM بدون ترميزات موضعية، يبدو أن كل رمز له نفس المسافة إلى جميع الرموز الأخرى، على سبيل المثال، سيكون من الصعب التمييز بين *"مرحبًا أنا أحبك"* و *"أنت تحبني مرحبًا"*.
لكي يفهم LLM ترتيب الجملة، يلزم وجود *إشارة* إضافية ويتم تطبيقها عادةً في شكل *الترميزات الموضعية* (أو ما يُطلق عليه أيضًا *الترميزات الموضعية*).
لم يتم ترجمة النص الخاص والروابط وأكواد HTML وCSS بناءً على طلبك.
قدم مؤلفو الورقة البحثية [*Attention Is All You Need*](https://huggingface.co/papers/1706.03762) تضمينات موضعية جيبية مثلثية \\( \mathbf{P} = \mathbf{p}_1, \ldots, \mathbf{p}_N \\) حيث يتم حساب كل متجه \\( \mathbf{p}_i \\) كدالة جيبية لموضعه \\( i \\) .
بعد ذلك يتم ببساطة إضافة التضمينات الموضعية إلى متجهات تسلسل الإدخال \\( \mathbf{\hat{X}} = \mathbf{\hat{x}}_1, \ldots, \mathbf{\hat{x}}_N \\) = \\( \mathbf{x}_1 + \mathbf{p}_1, \ldots, \mathbf{x}_N + \mathbf{p}_N \\) وبالتالي توجيه النموذج لتعلم ترتيب الجملة بشكل أفضل.
بدلاً من استخدام التضمينات الموضعية الثابتة، استخدم آخرون (مثل [Devlin et al.](https://huggingface.co/papers/1810.04805)) تضمينات موضعية مكتسبة يتم من خلالها تعلم التضمينات الموضعية \\( \mathbf{P} \\) أثناء التدريب.
كانت التضمينات الموضعية الجيبية والمكتسبة هي الطرق السائدة لترميز ترتيب الجملة في نماذج اللغة الكبيرة، ولكن تم العثور على بعض المشكلات المتعلقة بهذه التضمينات الموضعية:
1. التضمينات الموضعية الجيبية والمكتسبة هي تضمينات موضعية مطلقة، أي ترميز تضمين فريد لكل معرف موضعي: \\( 0, \ldots, N \\) . كما أظهر [Huang et al.](https://huggingface.co/papers/2009.13658) و [Su et al.](https://huggingface.co/papers/2104.09864)، تؤدي التضمينات الموضعية المطلقة إلى أداء ضعيف لنماذج اللغة الكبيرة للمدخلات النصية الطويلة. بالنسبة للمدخلات النصية الطويلة، يكون من المفيد إذا تعلم النموذج المسافة الموضعية النسبية التي تمتلكها رموز المدخلات إلى بعضها البعض بدلاً من موضعها المطلق.
2. عند استخدام التضمينات الموضعية المكتسبة، يجب تدريب نموذج اللغة الكبيرة على طول إدخال ثابت \\( N \\)، مما يجعل من الصعب الاستقراء إلى طول إدخال أطول مما تم تدريبه عليه.
في الآونة الأخيرة، أصبحت التضمينات الموضعية النسبية التي يمكنها معالجة المشكلات المذكورة أعلاه أكثر شعبية، وأبرزها:
- [تضمين الموضع الدوراني (RoPE)](https://huggingface.co/papers/2104.09864)
- [ALiBi](https://huggingface.co/papers/2108.12409)
يؤكد كل من *RoPE* و *ALiBi* أنه من الأفضل توجيه نموذج اللغة الكبيرة حول ترتيب الجملة مباشرة في خوارزمية الانتباه الذاتي حيث يتم وضع رموز الكلمات في علاقة مع بعضها البعض. على وجه التحديد، يجب توجيه ترتيب الجملة عن طريق تعديل عملية \\( \mathbf{QK}^T \\) .
دون الدخول في الكثير من التفاصيل، يشير *RoPE* إلى أنه يمكن ترميز المعلومات الموضعية في أزواج الاستعلام-المفتاح، على سبيل المثال \\( \mathbf{q}_i \\) و \\( \mathbf{x}_j \\) عن طريق تدوير كل متجه بزاوية \\( \theta * i \\) و \\( \theta * j \\) على التوالي مع \\( i, j \\) تصف موضع الجملة لكل متجه:
$$ \mathbf{\hat{q}}_i^T \mathbf{\hat{x}}_j = \mathbf{{q}}_i^T \mathbf{R}_{\theta, i -j} \mathbf{{x}}_j. $$
يمثل \\( \mathbf{R}_{\theta, i - j} \\) مصفوفة دورانية. \\( \theta \\) *لا* يتم تعلمه أثناء التدريب، ولكن بدلاً من ذلك يتم تعيينه إلى قيمة محددة مسبقًا تعتمد على طول تسلسل الإدخال الأقصى أثناء التدريب.
> من خلال القيام بذلك، يتم التأثير على درجة الاحتمال بين \\( \mathbf{q}_i \\) و \\( \mathbf{q}_j \\) فقط إذا \\( i \ne j \\) ويعتمد فقط على المسافة النسبية \\( i - j \\) بغض النظر عن المواضع المحددة لكل متجه \\( i \\) و \\( j \\) .
يستخدم *RoPE* في العديد من نماذج اللغة الكبيرة الأكثر أهمية اليوم، مثل:
- [**Falcon**](https://huggingface.co/tiiuae/falcon-40b)
- [**Llama**](https://huggingface.co/papers/2302.13971)
- [**PaLM**](https://huggingface.co/papers/2204.02311)
كبديل، يقترح *ALiBi* مخطط ترميز موضعي نسبي أبسط بكثير. يتم إضافة المسافة النسبية التي تمتلكها رموز المدخلات إلى بعضها البعض كعدد صحيح سلبي مقياس بقيمة محددة مسبقًا `m` إلى كل إدخال استعلام-مفتاح لمصفوفة \\( \mathbf{QK}^T \\) مباشرة قبل حساب softmax.
كما هو موضح في ورقة [ALiBi](https://huggingface.co/papers/2108.12409)، يسمح هذا الترميز الموضعي النسبي البسيط للنموذج بالحفاظ على أداء عالٍ حتى في تسلسلات المدخلات النصية الطويلة جدًا.
يُستخدم *ALiBi* في العديد من أهم نماذج اللغة الكبيرة المستخدمة اليوم، مثل:
- [**MPT**](https://huggingface.co/mosaicml/mpt-30b)
- [**BLOOM**](https://huggingface.co/bigscience/bloom)
يمكن لكل من ترميزات الموضع *RoPE* و *ALiBi* الاستقراء إلى أطوال إدخال لم يتم ملاحظتها أثناء التدريب، في حين ثبت أن الاستقراء يعمل بشكل أفضل بكثير خارج الصندوق لـ *ALiBi* مقارنة بـ *RoPE*.
بالنسبة لـ ALiBi، ما عليك سوى زيادة قيم مصفوفة الموضع المثلث السفلي لمطابقة طول تسلسل الإدخال.
بالنسبة لـ *RoPE*، يؤدي الحفاظ على نفس \\( \theta \\) الذي تم استخدامه أثناء التدريب إلى نتائج سيئة عند تمرير إدخالات نصية أطول بكثير من تلك التي شوهدت أثناء التدريب، راجع [Press et al.](https://huggingface.co/papers/2108.12409). ومع ذلك، وجد المجتمع بعض الحيل الفعالة التي تقوم بتعديل \\( \theta \\)، مما يسمح لترميزات الموضع *RoPE* بالعمل بشكل جيد لتسلسلات إدخال النص المستقرئة (راجع [هنا](https://github.com/huggingface/transformers/pull/24653)).
> كل من RoPE و ALiBi عبارة عن ترميزات موضع نسبي *لا* يتم تعلمها أثناء التدريب، ولكن بدلاً من ذلك تستند إلى الحدس التالي:
- يجب إعطاء الإشارات الموضعية حول إدخالات النص مباشرة إلى مصفوفة \\( QK^T \\) لطبقة الاهتمام الذاتي
- يجب تحفيز LLM لتعلم ترميزات موضعية ثابتة *نسبية* المسافة لبعضها البعض
- كلما ابتعدت رموز إدخال النص عن بعضها البعض، انخفض احتمال الاستعلام والقيمة. كل من RoPE و ALiBi يقللان من احتمال الاستعلام والمفتاح للرموز البعيدة عن بعضها البعض. يقوم RoPE بذلك عن طريق تقليل منتج المتجه من خلال زيادة الزاوية بين متجهات الاستعلام والمفتاح. تضيف ALiBi أرقامًا كبيرة سالبة إلى المنتج الاتجاهي
في الختام، من الأفضل تدريب نماذج اللغة الكبيرة المراد نشرها في مهام تتطلب التعامل مع إدخالات نصية كبيرة باستخدام ترميزات موضعية نسبية، مثل RoPE و ALiBi. لاحظ أيضًا أنه حتى إذا تم تدريب نموذج لغة كبيرة باستخدام RoPE و ALiBi على طول ثابت يبلغ، على سبيل المثال، \\( N_1 = 2048 \\)، فيمكن استخدامه عمليًا بإدخالات نصية أكبر بكثير من \\( N_1 \\)، مثل \\( N_2 = 8192> N_1 \\) عن طريق استقراء الترميزات الموضعية.
### 3.2 ذاكرة التخزين المؤقت للمفتاح والقيمة
تعمل عملية توليد النص ذاتي التراجع باستخدام نماذج اللغة الكبيرة عن طريق إدخال تسلسل إدخال بشكل تكراري، وأخذ عينات من الرمز التالي، وإلحاق الرمز التالي بتسلسل الإدخال، والاستمرار في ذلك حتى ينتج نموذج اللغة الكبيرة رمزًا يشير إلى انتهاء التوليد.
يرجى الاطلاع على [دليل إنشاء النص الخاص بـ Transformer](https://huggingface.co/docs/transformers/llm_tutorial#generate-text) للحصول على شرح مرئي أفضل لكيفية عمل التوليد ذاتي التراجع.
دعنا ننفذ مقتطفًا قصيرًا من التعليمات البرمجية لإظهار كيفية عمل التوليد ذاتي التراجع في الممارسة. ببساطة، سنأخذ الرمز الأكثر احتمالًا عبر `torch.argmax`.
```python
input_ids = tokenizer(prompt, return_tensors="pt")["input_ids"].to("cuda")
for _ in range(5):
next_logits = model(input_ids)["logits"][:, -1:]
next_token_id = torch.argmax(next_logits,dim=-1)
input_ids = torch.cat([input_ids, next_token_id], dim=-1)
print("shape of input_ids", input_ids.shape)
generated_text = tokenizer.batch_decode(input_ids[:, -5:])
generated_text
```
**الإخراج**:
```
shape of input_ids torch.Size([1, 21])
shape of input_ids torch.Size([1, 22])
shape of input_ids torch.Size([1, 23])
shape of input_ids torch.Size([1, 24])
shape of input_ids torch.Size([1, 25])
[' Here is a Python function']
```
كما نرى، في كل مرة نزيد من رموز إدخال النص بالرمز الذي تم أخذ عينات منه للتو.
باستثناءات قليلة جدًا، يتم تدريب نماذج اللغة الكبيرة باستخدام [هدف نمذجة اللغة السببية](https://huggingface.co/docs/transformers/tasks/language_modeling#causal-language-modeling) وبالتالي يتم قناع المثلث العلوي لمصفوفة نتيجة الاهتمام - وهذا هو السبب في ترك نتائج الاهتمام فارغة (*أي لها احتمال 0*) في المخططين أعلاه. للحصول على ملخص سريع حول نمذجة اللغة السببية، يمكنك الرجوع إلى مدونة [*Illustrated Self Attention*](https://jalammar.github.io/illustrated-gpt2/#part-2-illustrated-self-attention).
ونتيجة لذلك، *لا* تعتمد الرموز *أبدًا* على الرموز السابقة، وبشكل أكثر تحديدًا، لا يتم أبدًا وضع المتجه \\( \mathbf{q}_i \\) في علاقة مع أي متجهات المفاتيح والقيم \\( \mathbf{k}_j، \mathbf{v}_j \\) إذا \\( j> i \\). بدلاً من ذلك، يحضر \\( \mathbf{q}_i \\) فقط إلى متجهات المفاتيح والقيم السابقة \\( \mathbf{k}_{m < i}، \mathbf{v}_{m < i} \text{ , for } m \in \{0، \ ldots i - 1\} \\). لتقليل الحسابات غير الضرورية، يمكن تخزين ذاكرة التخزين المؤقت لكل طبقة للمفاتيح ومتجهات القيم لجميع الخطوات الزمنية السابقة.
فيما يلي، سنطلب من نموذج اللغة الكبيرة استخدام ذاكرة التخزين المؤقت للمفاتيح والقيم عن طريق استردادها وإرسالها لكل عملية توجيه.
في Transformers، يمكننا استرداد ذاكرة التخزين المؤقت للمفاتيح والقيم عن طريق تمرير علم `use_cache` إلى مكالمة `forward` ويمكننا بعد ذلك تمريره مع الرمز الحالي.
```python
past_key_values = None # past_key_values is the key-value cache
generated_tokens = []
next_token_id = tokenizer(prompt, return_tensors="pt")["input_ids"].to("cuda")
for _ in range(5):
next_logits, past_key_values = model(next_token_id, past_key_values=past_key_values, use_cache=True).to_tuple()
next_logits = next_logits[:, -1:]
next_token_id = torch.argmax(next_logits, dim=-1)
print("shape of input_ids", next_token_id.shape)
print("length of key-value cache", len(past_key_values[0][0])) # past_key_values are of shape [num_layers, 0 for k, 1 for v, batch_size, length, hidden_dim]
generated_tokens.append(next_token_id.item())
generated_text = tokenizer.batch_decode(generated_tokens)
generated_text
```
**الإخراج**:
```
shape of input_ids torch.Size([1, 1])
length of key-value cache 20
shape of input_ids torch.Size([1, 1])
length of key-value cache 21
shape of input_ids torch.Size([1, 1])
length of key-value cache 22
shape of input_ids torch.Size([1, 1])
length of key-value cache 23
shape of input_ids torch.Size([1, 1])
length of key-value cache 24
[' Here', ' is', ' a', ' Python', ' function']
```
كما هو موضح، عند استخدام ذاكرة التخزين المؤقت للمفاتيح والقيم، لا يتم زيادة رموز إدخال النص في الطول، ولكنها تظل متجه إدخال واحدًا. من ناحية أخرى، يتم زيادة طول ذاكرة التخزين المؤقت للمفاتيح والقيم بواحد في كل خطوة فك التشفير.
> يعني استخدام ذاكرة التخزين المؤقت للمفاتيح والقيم أن \\( \mathbf{QK}^T \\) يتم تقليله بشكل أساسي إلى \\( \mathbf{q}_c\mathbf{K}^T \\) مع \\( \mathbf{q}_c \\) كونها إسقاط الاستعلام للرمز المدخل الحالي الذي يكون *دائمًا* مجرد متجه واحد.
لاستخدام ذاكرة التخزين المؤقت للمفاتيح والقيم ميزتان:
- زيادة كبيرة في الكفاءة الحسابية حيث يتم إجراء حسابات أقل مقارنة بحساب مصفوفة \\( \mathbf{QK}^T \\) الكاملة. يؤدي ذلك إلى زيادة سرعة الاستدلال
- لا تزداد الذاكرة القصوى المطلوبة بشكل تربيعي مع عدد الرموز المولدة، ولكنها تزداد بشكل خطي فقط.
> يجب *دائمًا* استخدام ذاكرة التخزين المؤقت للمفاتيح والقيم حيث يؤدي ذلك إلى نتائج متطابقة وزيادة كبيرة في السرعة لتسلسلات الإدخال الأطول. ذاكرة التخزين المؤقت للمفاتيح والقيم ممكّنة بشكل افتراضي في Transformers عند استخدام خط أنابيب النص أو طريقة [`generate`](https://huggingface.co/docs/transformers/main_classes/text_generation).
<Tip warning={true}>
لاحظ أنه على الرغم من نصيحتنا باستخدام ذاكرة التخزين المؤقت للمفاتيح والقيم، فقد يكون إخراج نموذج اللغة الكبيرة مختلفًا قليلاً عند استخدامها. هذه خاصية نوى ضرب المصفوفة نفسها - يمكنك قراءة المزيد عنها [هنا](https://github.com/huggingface/transformers/issues/25420#issuecomment-1775317535).
</Tip>
#### 3.2.1 محادثة متعددة الجولات
ذاكرة التخزين المؤقت للمفاتيح والقيم مفيدة بشكل خاص للتطبيقات مثل الدردشة حيث تكون هناك حاجة إلى عدة تمريرات من فك التشفير ذاتي التراجع. دعنا نلقي نظرة على مثال.
```
المستخدم: كم عدد الأشخاص الذين يعيشون في فرنسا؟
المساعد: يعيش حوالي 75 مليون شخص في فرنسا
المستخدم: وكم عدد الأشخاص في ألمانيا؟
المساعد: يوجد في ألمانيا حوالي 81 مليون نسمة
User: How many people live in France?
Assistant: Roughly 75 million people live in France
User: And how many are in Germany?
Assistant: Germany has ca. 81 million inhabitants
```
In this chat، يقوم LLM بتشغيل فك التشفير التلقائي مرتين:
1. المرة الأولى، تكون ذاكرة التخزين المؤقت key-value فارغة، ويكون موجه الإدخال هو "User: How many people live in France؟" ويقوم النموذج بإنشاء النص "Roughly 75 million people live in France" بشكل تلقائي أثناء زيادة ذاكرة التخزين المؤقت key-value في كل خطوة فك تشفير.
2. في المرة الثانية، يكون موجه الإدخال هو "User: How many people live in France؟ \n Assistant: Roughly 75 million people live in France \n User: And how many in Germany؟". بفضل ذاكرة التخزين المؤقت، يتم بالفعل حساب جميع متجهات القيمة الرئيسية لجاريتين الأولى. لذلك يتكون موجه الإدخال فقط من "User: And how many in Germany؟". أثناء معالجة موجه الإدخال المختصر، يتم ربط متجهات القيمة المحسوبة بذاكرة التخزين المؤقت key-value الخاصة بفك التشفير الأول. يتم بعد ذلك إنشاء إجابة المساعد الثانية "Germany has ca. 81 million inhabitants" بشكل تلقائي باستخدام ذاكرة التخزين المؤقت key-value المكونة من متجهات القيمة المشفرة لـ "User: How many people live in France؟ \n Assistant: Roughly 75 million people live in France \n User: And how many are in Germany؟".
يجب ملاحظة أمرين هنا:
1. الحفاظ على كل السياق أمر بالغ الأهمية للنماذج اللغوية الكبيرة (LLMs) التي يتم نشرها في الدردشة بحيث يفهم LLM كل سياق المحادثة السابق. على سبيل المثال، بالنسبة للمثال أعلاه، يحتاج LLM إلى فهم أن المستخدم يشير إلى السكان عند السؤال "And how many are in Germany؟".
2. ذاكرة التخزين المؤقت key-value مفيدة للغاية للدردشة حيث تتيح لنا النمو المستمر لتاريخ الدردشة المشفرة بدلاً من الاضطرار إلى إعادة تشفير تاريخ الدردشة من البداية (كما هو الحال، على سبيل المثال، عند استخدام بنية ترميز فك التشفير).
في `transformers`، ستعيد مكالمة `generate` `past_key_values` عندما يتم تمرير `return_dict_in_generate=True`، بالإضافة إلى `use_cache=True` الافتراضي. لاحظ أنه غير متوفر بعد من خلال واجهة `pipeline`.
```python
# Generation as usual
prompt = system_prompt + "Question: Please write a function in Python that transforms bytes to Giga bytes.\n\nAnswer: Here"
model_inputs = tokenizer(prompt، return_tensors='pt')
generation_output = model.generate(**model_inputs، max_new_tokens=60، return_dict_in_generate=True)
decoded_output = tokenizer.batch_decode(generation_output.sequences)[0]
# Piping the returned `past_key_values` to speed up the next conversation round
prompt = decoded_output + "\nQuestion: How can I modify the function above to return Mega bytes instead?\n\nAnswer: Here"
model_inputs = tokenizer(prompt، return_tensors='pt')
generation_output = model.generate(
**model_inputs،
past_key_values=generation_output.past_key_values،
max_new_tokens=60،
return_dict_in_generate=True
)
tokenizer.batch_decode(generation_output.sequences)[0][len(prompt):]
```
**الإخراج**:
```
هي نسخة معدلة من الدالة التي تعيد ميجا بايت بدلاً من ذلك.
def bytes_to_megabytes(bytes):
return bytes / 1024 / 1024
Answer: The function takes a number of bytes as input and returns the number of
```
رائع، لا يتم إنفاق وقت إضافي على إعادة حساب نفس المفتاح والقيم لطبقة الاهتمام! ومع ذلك، هناك شيء واحد يجب ملاحظته. في حين أن ذروة الذاكرة المطلوبة لمصفوفة \\( \mathbf{QK}^T \\) يتم تقليلها بشكل كبير، فإن الاحتفاظ بذاكرة التخزين المؤقت key-value في الذاكرة يمكن أن يصبح مكلفًا جدًا من حيث الذاكرة لسلاسل الإدخال الطويلة أو الدردشة متعددة الجولات. تذكر أن ذاكرة التخزين المؤقت key-value بحاجة إلى تخزين متجهات القيمة الرئيسية لجميع متجهات الإدخال السابقة \\( \mathbf{x}_i \text{، لـ } i \in \{1، \ ldots، c - 1\} \\) لجميع طبقات الاهتمام الذاتي وكل رؤوس الاهتمام.
دعنا نحسب عدد القيم العائمة التي يجب تخزينها في ذاكرة التخزين المؤقت key-value لنموذج LLM `bigcode/octocoder` الذي استخدمناه من قبل.
يبلغ عدد القيم العائمة ضعف طول التسلسل مضروبًا في عدد رؤوس الاهتمام مضروبًا في بعد رأس الاهتمام ومضروبًا في عدد الطبقات.
حساب هذا لنموذج LLM لدينا عند طول تسلسل افتراضي يبلغ 16000 يعطي:
```python
config = model.config
2 * 16_000 * config.n_layer * config.n_head * config.n_embd // config.n_head
```
**الإخراج**:
```
7864320000
```
Roughly 8 مليار قيمة عائمة! يتطلب تخزين 8 مليارات قيمة عائمة في دقة `float16` حوالي 15 جيجابايت من ذاكرة الوصول العشوائي (RAM) وهو ما يقرب من نصف حجم أوزان النموذج نفسها!
اقترح الباحثون طريقتين تسمحان بتقليل تكلفة الذاكرة لتخزين ذاكرة التخزين المؤقت key-value بشكل كبير، والتي يتم استكشافها في الأقسام الفرعية التالية.
#### 3.2.2 Multi-Query-Attention (MQA)
[Multi-Query-Attention](https://huggingface.co/papers/1911.02150) اقترحها Noam Shazeer في ورقته *Fast Transformer Decoding: One Write-Head is All You Need*. كما يقول العنوان، اكتشف Noam أنه بدلاً من استخدام `n_head` من أوزان إسقاط القيمة الرئيسية، يمكن استخدام زوج واحد من أوزان إسقاط رأس القيمة التي يتم مشاركتها عبر جميع رؤوس الاهتمام دون أن يتدهور أداء النموذج بشكل كبير.
> باستخدام زوج واحد من أوزان إسقاط رأس القيمة، يجب أن تكون متجهات القيمة الرئيسية \\( \mathbf{k}_i، \mathbf{v}_i \\) متطابقة عبر جميع رؤوس الاهتمام والتي بدورها تعني أننا بحاجة فقط إلى تخزين زوج إسقاط قيمة رئيسي واحد في ذاكرة التخزين المؤقت بدلاً من `n_head` منها.
نظرًا لأن معظم LLMs تستخدم ما بين 20 و100 رأس اهتمام، فإن MQA يقلل بشكل كبير من استهلاك الذاكرة لذاكرة التخزين المؤقت key-value. بالنسبة إلى LLM المستخدم في هذا الدفتر، يمكننا تقليل استهلاك الذاكرة المطلوبة من 15 جيجابايت إلى أقل من 400 ميجابايت عند طول تسلسل الإدخال 16000.
بالإضافة إلى توفير الذاكرة، يؤدي MQA أيضًا إلى تحسين الكفاءة الحسابية كما هو موضح في ما يلي.
في فك التشفير التلقائي، يجب إعادة تحميل متجهات القيمة الرئيسية الكبيرة، ودمجها مع زوج متجه القيمة الحالي، ثم إدخالها في \\( \mathbf{q}_c\mathbf{K}^T \\) الحساب في كل خطوة. بالنسبة لفك التشفير التلقائي، يمكن أن تصبح عرض النطاق الترددي للذاكرة المطلوبة لإعادة التحميل المستمر عنق زجاجة زمنيًا خطيرًا. من خلال تقليل حجم متجهات القيمة الرئيسية، يجب الوصول إلى ذاكرة أقل، وبالتالي تقليل عنق الزجاجة في عرض النطاق الترددي للذاكرة. لمزيد من التفاصيل، يرجى إلقاء نظرة على [ورقة Noam](https://huggingface.co/papers/1911.02150).
الجزء المهم الذي يجب فهمه هنا هو أن تقليل عدد رؤوس الاهتمام بالقيمة الرئيسية إلى 1 لا معنى له إلا إذا تم استخدام ذاكرة التخزين المؤقت للقيمة الرئيسية. يظل الاستهلاك الذروي لذاكرة النموذج لمرور واحد للأمام بدون ذاكرة التخزين المؤقت للقيمة الرئيسية دون تغيير لأن كل رأس اهتمام لا يزال لديه متجه استعلام فريد بحيث يكون لكل رأس اهتمام مصفوفة \\( \mathbf{QK}^T \\) مختلفة.
شهدت MQA اعتمادًا واسع النطاق من قبل المجتمع ويتم استخدامها الآن بواسطة العديد من LLMs الأكثر شهرة:
- [**Falcon**](https://huggingface.co/tiiuae/falcon-40b)
- [**PaLM**](https://huggingface.co/papers/2204.02311)
- [**MPT**](https://huggingface.co/mosaicml/mpt-30b)
- [**BLOOM**](https://huggingface.co/bigscience/bloom)
كما يستخدم نقطة التحقق المستخدمة في هذا الدفتر - `bigcode/octocoder` - MQA.
#### 3.2.3 مجموعة الاستعلام الاهتمام (GQA)
[مجموعة الاستعلام الاهتمام](https://huggingface.co/papers/2305.13245)، كما اقترح Ainslie et al. من Google، وجد أن استخدام MQA يمكن أن يؤدي غالبًا إلى تدهور الجودة مقارنة باستخدام إسقاطات رأس القيمة الرئيسية المتعددة. تجادل الورقة بأنه يمكن الحفاظ على أداء النموذج بشكل أكبر عن طريق تقليل عدد أوزان إسقاط رأس الاستعلام بشكل أقل حدة. بدلاً من استخدام وزن إسقاط قيمة رئيسية واحدة فقط، يجب استخدام `n <n_head` أوزان إسقاط قيمة رئيسية. من خلال اختيار `n` إلى قيمة أقل بكثير من `n_head`، مثل 2 أو 4 أو 8، يمكن الاحتفاظ بمعظم مكاسب الذاكرة والسرعة من MQA مع التضحية بقدر أقل من سعة النموذج وبالتالي، من المفترض، أقل أداء.
علاوة على ذلك، اكتشف مؤلفو GQA أنه يمكن *تدريب* نقاط تفتيش النموذج الموجودة ليكون لها بنية GQA باستخدام 5% فقط من الحوسبة الأصلية للتعليم المسبق. في حين أن 5% من الحوسبة الأصلية للتعليم المسبق يمكن أن تكون كمية هائلة، يسمح GQA *uptraining* بنقاط تفتيش موجودة للاستفادة من تسلسلات الإدخال الأطول.
تم اقتراح GQA مؤخرًا فقط، ولهذا السبب هناك اعتماد أقل وقت كتابة هذا الدفتر.
أبرز تطبيق لـ GQA هو [Llama-v2](https://huggingface.co/meta-llama/Llama-2-70b-hf).
> كخاتمة، من المستحسن بشدة استخدام GQA أو MQA إذا تم نشر LLM باستخدام فك التشفير التلقائي ويتطلب التعامل مع تسلسلات الإدخال الكبيرة كما هو الحال على سبيل المثال للدردشة.
## الخاتمة
مجتمع البحث يأتي باستمرار بطرق جديدة ومبتكرة لتسريع وقت الاستدلال للنماذج اللغوية الكبيرة على الإطلاق. كمثال، أحد اتجاهات البحث الواعدة هو [فك التشفير التخميني](https://huggingface.co/papers/2211.17192) حيث تقوم "الرموز السهلة" بإنشائها نماذج اللغة الأصغر والأسرع ويتم إنشاء "الرموز الصعبة" فقط بواسطة LLM نفسه. إن التعمق في التفاصيل يتجاوز نطاق هذا الدفتر، ولكن يمكن قراءته في هذه [تدوينة المدونة اللطيفة](https://huggingface.co/blog/assisted-generation).
السبب في أن LLMs الضخمة مثل GPT3/4، وLlama-2-70b، وClaude، وPaLM يمكن أن تعمل بسرعة كبيرة في واجهات الدردشة مثل [Hugging Face Chat](https://huggingface.co/chat/) أو ChatGPT يرجع إلى حد كبير إلى التحسينات المذكورة أعلاه في الدقة والخوارزميات والهندسة المعمارية.
في المستقبل، ستكون أجهزة التسريع مثل وحدات معالجة الرسومات (GPUs) ووحدات معالجة الرسومات (TPUs)، وما إلى ذلك... ستكون أسرع فقط وستسمح بمزيد من الذاكرة، ولكن يجب دائمًا التأكد من استخدام أفضل الخوارزميات والهندسة المعمارية المتاحة للحصول على أكبر قدر من المال
| transformers/docs/source/ar/llm_tutorial_optimization.md/0 | {
"file_path": "transformers/docs/source/ar/llm_tutorial_optimization.md",
"repo_id": "transformers",
"token_count": 37743
} | 349 |
# تشغيل التدريب على Amazon SageMaker
تم نقل التوثيق إلى [hf.co/docs/sagemaker](https://huggingface.co/docs/sagemaker). وسيتم إزالة هذه الصفحة في الإصدار 5.0 من برنامج Transformers.
### جدول المحتويات
- [تدريب نماذج Hugging Face على Amazon SageMaker باستخدام SageMaker Python SDK](https://huggingface.co/docs/sagemaker/train)
- [نشر نماذج Hugging Face على Amazon SageMaker باستخدام SageMaker Python SDK](https://huggingface.co/docs/sagemaker/inference) | transformers/docs/source/ar/sagemaker.md/0 | {
"file_path": "transformers/docs/source/ar/sagemaker.md",
"repo_id": "transformers",
"token_count": 256
} | 350 |
# Trainer
تُتيح وحدة [`Trainer`] حلقة تدريب وتقييم متكاملة لنماذج PyTorch المطبقة في مكتبة Transformers. تحتاج فقط إلى تمرير المكونات الضرورية للتدريب (النموذج، والمجزىء النصى، ومجموعة البيانات، دالة التقييم، معلمات التدريب الفائقة، إلخ)، وستتولى فئة [`Trainer`] الباقي. هذا يُسهّل بدء التدريب بشكل أسرع دون كتابة حلقة التدريب الخاصة بك يدويًا. ولكن في الوقت نفسه، فإن [`Trainer`] قابل للتخصيص بدرجة كبيرة ويوفر العديد من خيارات التدريب حتى تتمكن من تخصيصه وفقًا لاحتياجات التدريب الخاصة بك بدقة.
<Tip>
بالإضافة إلى فئة [`Trainer`], توفر مكتبة Transformers أيضًا فئة [`Seq2SeqTrainer`] للمهام التسلسلية مثل الترجمة أو التلخيص. هناك أيضًا فئة [`~trl.SFTTrainer`] من مكتبة [TRL](https://hf.co/docs/trl) التي تغلّف فئة [`Trainer`] وهي مُحُسَّنة لتدريب نماذج اللغة مثل Llama-2 وMistral باستخدام تقنيات التوليد اللغوي. كما يدعم [`~trl.SFTTrainer`] ميزات مثل حزم التسلسلات، وLoRA، والقياس الكمي، وDeepSpeed مما يُمكّن من التدريب بكفاءة على نماذج ضخمة الحجم.
<br>
لا تتردد في الاطلاع على [مرجع API](./main_classes/trainer) لهذه الفئات الأخرى من النوع [`Trainer`] لمعرفة المزيد حول متى يتم استخدام كل منها. بشكل عام، [`Trainer`] هو الخيار الأكثر تنوعًا ومناسبًا لمجموعة واسعة من المهام. تم تصميم [`Seq2SeqTrainer`] للمهام التسلسلية ، و [`~trl.SFTTrainer`] مُصمم لتدريب نماذج اللغة الكبيرة.
</Tip>
قبل البدء، تأكد من تثبيت مكتبة [Accelerate](https://hf.co/docs/accelerate) - وهي مكتبة تُمكّن تشغيل تدريب PyTorch في بيئات مُوزعة.
```bash
pip install accelerate
# upgrade
pip install accelerate --upgrade
```
يوفر هذا الدليل نظرة عامة على فئة [`Trainer`].
## الاستخدام الأساسي
يتضمن [`Trainer`] جميع التعليمات البرمجية التي ستجدها في حلقة التدريب الأساسية:
1. قم بتنفيذ خطوة تدريب لحساب الخسارة
2. احسب المشتقات باستخدام طريقة [`~accelerate.Accelerator.backward`]
3. تحديث الأوزان بناءً على المشتقات
4. كرر هذه العملية حتى تصل إلى عدد محدد مسبقًا من الدورات (epochs).
تُجرد فئة [`Trainer`] كل هذه التعليمات البرمجية حتى لا تضطر إلى القلق بشأن كتابة حلقة تدريب يدويًا في كل مرة أما إذا كنت بدأت للتو في PyTorch والتدريب. كل ما عليك فعله هو توفير المكونات الأساسية اللازمة للتدريب، مثل النموذج ومجموعة بيانات، وتتعامل فئة [`Trainer`] مع كل شيء آخر.
إذا كنت تُريد تحديد أي خيارات تدريب أو معلمات فائقة، فيمكنك العثور عليها في فئة [`TrainingArguments`]. على سبيل المثال، دعنا نحدد أين يتم حفظ النموذج في `output_dir` ورفع النموذج إلى Hub بعد التدريب باستخدام `push_to_hub=True`.
```py
from transformers import TrainingArguments
training_args = TrainingArguments(
output_dir="your-model"،
learning_rate=2e-5,
per_device_train_batch_size=16,
per_device_eval_batch_size=16,
num_train_epochs=2,
weight_decay=0.01,
eval_strategy="epoch"،
save_strategy="epoch"،
load_best_model_at_end=True,
push_to_hub=True,
)
```
مرر `training_args` إلى [`Trainer`] جنبًا إلى جنب مع النموذج، ومجموعة بيانات، وشئ لمعالجة مجموعة البيانات مسبقًا (حسب نوع البيانات، فقد يكون محللًا رمزيًا أو مستخرج ميزات أو معالج صور)، وجامع بيانات، ودالة لحساب المقاييس التي تُريد تتبعها أثناء التدريب.
أخيرًا، استدعِ [`~Trainer.train`] لبدء التدريب!
```py
from transformers import Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=dataset["train"]،
eval_dataset=dataset["test"]،
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics,
)
trainer.train()
```
### نقاط الحفظ
تحفظ فئة [`Trainer`] نقاط الحفظ النموذج في الدليل المحدد في معامل `output_dir` من [`TrainingArguments`]. ستجد نقاط الحفظ في مجلد فرعي يسمى `checkpoint-000` حيث تتوافق الأرقام في النهاية مع خطوة التدريب. إن حفظ نقاط الحفظ مفيد لاستئناف التدريب لاحقًا.
```py
# استأنف من أحدث نقطة حفظ
trainer.train(resume_from_checkpoint=True)
# استأنف من نقطة حفظ محددة محفوظة في دليل الإخراج
trainer.train(resume_from_checkpoint="your-model/checkpoint-1000")
```
يمكنك حفظ نقاط الحفظ الخاصة بك (لا يتم حفظ حالة المُجزىء اللغوى تقائيًا) إلى Hub عن طريق تعيين `push_to_hub=True` في [`TrainingArguments`] لرفعها. الخيارات الأخرى لاتخاذ القرار بشأن كيفية حفظ هذة النقاط الخاصة بك هي الإعداد في معامل [`hub_strategy`](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.TrainingArguments.hub_strategy):
* `hub_strategy="checkpoint"` يدفع أحدث نقطة حفظ إلى مجلد فرعي يسمى "last-checkpoint" يمكنك استئناف التدريب منه
* `hub_strategy="all_checkpoints"` يدفع جميع نقاط الحفظ إلى الدليل المحدد في `output_dir` (سترى نقطة حفظ واحدة لكل مجلد في مستودع النموذج الخاص بك)
عند استئناف التدريب من نقطة حفظ، تُحاول [`Trainer`] الحفاظ على حالات RNG Python وNumPy وPyTorch كما كانت عندما تم حفظ نقطة الحفظ. ولكن لأن PyTorch لديها العديد من الإعدادات الافتراضية غير الحتمية مُتنوعة، فإن حالات RNG ليست مضمونة لتكون هي نفسها. إذا كنت تريد تمكين الحتمية الكاملة، فراجع دليل [التحكم في مصادر العشوائية](https://pytorch.org/docs/stable/notes/randomness#controlling-sources-of-randomness) لمعرفة ما يُمكنك تمكينه لجعل تدريبك حتميًا تمامًا. ضع في اعتبارك أنه من خلال جعل إعدادات معينة حتمية، فقد يكون التدريب أبطأ.
## تخصيص المدرب
في حين أن فئة [`Trainer`] مُصممة لتكون سهلة الوصول وسهلة الاستخدام، فإنها توفر أيضًا الكثير من قابلية التخصيص للمستخدمين المغامرين. يُمكن إنشاء فئات فرعية من العديد من أساليب [`Trainer`] وتجاوزها لدعم الوظائف التي تُريدها، دون الحاجة إلى إعادة كتابة حلقة التدريب بأكملها من البداية لاستيعابها. تتضمن هذه الأساليب:
* [`~Trainer.get_train_dataloader`] ينشئ DataLoader للتدريب
* [`~Trainer.get_eval_dataloader`] ينشئ DataLoader للتقييم
* [`~Trainer.get_test_dataloader`] ينشئ DataLoader للاختبار
* [`~Trainer.log`] يسجل معلومات حول مختلف الكائنات التي تراقب التدريب
* [`~Trainer.create_optimizer_and_scheduler`] ينشئ محسنًا ومخططًا لمُعدل التعلم إذا لم يتم تمريرهما في `__init__`؛ يمكن أيضًا تخصيص هذه الوظائف بشكل منفصل باستخدام [`~Trainer.create_optimizer`] و [`~Trainer.create_scheduler`] على التوالي
* [`~Trainer.compute_loss`] يحسب دالة الخسارة على دفعة من مُدخلات التدريب
* [`~Trainer.training_step`] يُنفذ خطوة التدريب
* [`~Trainer.prediction_step`] يُنفذ خطوة التنبؤ والاختبار
* [`~Trainer.evaluate`] يُقيّم النموذج ويعيد مقاييس التقييم
* [`~Trainer.predict`] يُجري التنبؤات (مع المقاييس إذا كانت العلامات متاحة) على مجموعة الاختبار
على سبيل المثال، إذا كنت تريد تخصيص طريقة [`~Trainer.compute_loss`] لاستخدام دالة خسارة ذات ترجيح بدلاً من ذلك.
```py
from torch import nn
from transformers import Trainer
class CustomTrainer(Trainer):
def compute_loss(self, model, inputs, return_outputs=False):
labels = inputs.pop("labels")
# forward pass
outputs = model(**inputs)
logits = outputs.get("logits")
# compute custom loss for 3 labels with different weights
loss_fct = nn.CrossEntropyLoss(weight=torch.tensor([1.0, 2.0, 3.0], device=model.device))
loss = loss_fct(logits.view(-1, self.model.config.num_labels), labels.view(-1))
return (loss, outputs) if return_outputs else loss
```
### دوال الاستدعاء Callbacks
خيار آخر لتخصيص [`Trainer`] هو استخدام [دوال الاستدعاء](callbacks). لا *تغير* دوال الاستدعاء أي شيء في حلقة التدريب. إنهم تفحص حالة حلقة التدريب ثم تُنفذ بعض الإجراءات (مثل الإيقاف المبكر أو تسجيل النتائج، إلخ) اعتمادًا على الحالة. وبعبارة أخرى، لا يمكن استخدام دالة الاستدعاء لتنفيذ شيء مثل دالة خسارة مخصصة، ويجب عليك تجاوز دالة [`~Trainer.compute_loss`] لذلك.
على سبيل المثال، إذا كنت تريد إضافة دالة استدعاء إيقاف مبكر إلى حلقة التدريب بعد 10 خطوات.
```py
from transformers import TrainerCallback
class EarlyStoppingCallback(TrainerCallback):
def __init__(self, num_steps=10):
self.num_steps = num_steps
def on_step_end(self, args, state, control, **kwargs):
if state.global_step >= self.num_steps:
return {"should_training_stop": True}
else:
return {}
```
ثم مرره إلى معامل `callback` في [`Trainer`].
```py
from transformers import Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=dataset["train"]،
eval_dataset=dataset["test"]،
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics,
callback=[EarlyStoppingCallback()],
)
```
## تسجيل الأحداث (Logging)
<Tip>
راجع مرجع [API](./main_classes/logging) للتسجيل للحصول على مزيد من المعلومات حول مستويات التسجيل المختلفة للأحداث.
</Tip>
يتم تعيين [`Trainer`] إلى `logging.INFO` افتراضيًا والذي يُبلغ عن الأخطاء والتحذيرات ومعلومات أساسية أخرى. يتم تعيين نسخة [`Trainer`] - في البيئات الموزعة - إلى `logging.WARNING` والتي يُبلغ فقط عن الأخطاء والتحذيرات. يمكنك تغيير مستوى تسجيل الأحداث باستخدام معاملي [`log_level`](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.TrainingArguments.log_level) و [`log_level_replica`](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.TrainingArguments.log_level_replica) في [`TrainingArguments`].
لتهيئة إعداد مُستوى تسجيل اﻷحداث لكل عقدة، استخدم معامل [`log_on_each_node`](https://huggingface.co/docs/transformers/main/en/main_classes/trainer#transformers.TrainingArguments.log_on_each_node) لتحديد ما إذا كان سيتم استخدام مُستوى السجل على كل عقدة أو فقط على العقدة الرئيسية.
<Tip>
يحدد [`Trainer`] مُستوى التسجيل بشكل مُنفصل لكل عقدة في طريقة [`Trainer.__init__`]، لذا فقد ترغب في التفكير في تعيين هذا الإعداد في وقت سابق إذا كنت تستخدم وظائف Transformers الأخرى قبل إنشاء كائن [`Trainer`].
</Tip>
على سبيل المثال، لتعيين التعليمات البرمجية والوحدات النمطية الرئيسية الخاصة بك لاستخدام نفس مُستوى التسجيل وفقًا لكل عقدة:
```py
logger = logging.getLogger(__name__)
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s"،
datefmt="%m/%d/%Y %H:%M:%S"،
handlers=[logging.StreamHandler(sys.stdout)],
)
log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
datasets.utils.logging.set_verbosity(log_level)
transformers.utils.logging.set_verbosity(log_level)
trainer = Trainer(...)
```
استخدم تركيبات مختلفة من `log_level` و `log_level_replica` لتهيئة ما يتم تسجيله على كل من العقد.
<hfoptions id="logging">
<hfoption id="single node">
```bash
my_app.py ... --log_level warning --log_level_replica error
```
</hfoption>
<hfoption id="multi-node">
أضف معلمة `log_on_each_node 0` لبيئات متعددة العقد.
```bash
my_app.py ... --log_level warning --log_level_replica error --log_on_each_node 0
# set to only report errors
my_app.py ... --log_level error --log_level_replica error --log_on_each_node 0
```
</hfoption>
</hfoptions>
## NEFTune
[NEFTune](https://hf.co/papers/2310.05914) هي تقنية يمكن أن تحسن الأداء عن طريق إضافة ضوضاء إلى مُتجهات التعلم أثناء التدريب. لتمكينه في [`Trainer`], قم بتعيين معامل `neftune_noise_alpha` في [`TrainingArguments`] للتحكم في مقدار الضوضاء المُضافة.
```py
from transformers import TrainingArguments, Trainer
training_args = TrainingArguments(..., neftune_noise_alpha=0.1)
trainer = Trainer(..., args=training_args)
```
يتم تعطيل NEFTune بعد التدريب لاستعادة طبقة التعلم الأصلية لتجنب أي سلوك غير متوقع.
## نواة Liger
[Liger-Kernel](https://github.com/linkedin/Liger-Kernel) Kernel هي مجموعة من نوى Triton التي طورتها Linkedin مُصممة خصيصًا لتدريب نماذج اللغة الكبيرة (LLM). لقد قمنا بتنفيذ RMSNorm و RoPE و SwiGLU و CrossEntropy و FusedLinearCrossEntropy مُتوافقة مع Hugging Face، والمزيد قادم. يُمكنها زيادة إنتاجية التدريب متعدد وحدات معالجة الرسومات (GPU) بنسبة 20٪ وتقليل استخدام الذاكرة بنسبة 60٪. تعمل النواة بشكل تلقائي مع flash attention و PyTorch FSDP و Microsoft DeepSpeed.
احصل على زيادة في الإنتاجية بنسبة 20٪ وتقليل استخدام الذاكرة بنسبة 60٪ على تدريب نماذج LLaMA 3-8B. حقق أطوال سياق أكبر وأحجام دفعات أكبر. كما أنها مُفيدة إذا كنت تُريد زيادة حجم نموذجك إلى تدريب بنماذج متعددة الرؤوس أو أحجام مُفردات ضخمة. أطلق العنان للتدريب بنماذج متعددة الرؤوس (medusa) والمزيد. راجع التفاصيل والأمثلة في [Liger](https://github.com/linkedin/Liger-Kernel/tree/main/examples)
تأكد أولاً من تثبيت مستودع Liger الرسمي:
```bash
pip install liger-kernel
```
يجب عليك تمرير `use_liger_kernel=True` لتطبيق نواة `liger` على نموذجك، على سبيل المثال:
```python
from transformers import TrainingArguments
training_args = TrainingArguments(
output_dir="your-model",
learning_rate=2e-5,
per_device_train_batch_size=16,
per_device_eval_batch_size=16,
num_train_epochs=2,
weight_decay=0.01,
eval_strategy="epoch",
save_strategy="epoch",
load_best_model_at_end=True,
push_to_hub=True,
use_liger_kernel=True
)
```
تدعم النواة معماريات نماذج Llama و Gemma و Mistral و Mixtral. يُمكن العثور على أحدث قائمة بالنمائج المدعومة [هنا](https://github.com/linkedin/Liger-Kernel). عندما يتم تعيين `use_liger_kernel` إلى `True`، سيتم تصحيح الطبقات المُقابلة في النموذج الأصلي باستخدام تطبيق Liger الفعال، لذلك لا تحتاج إلى فعل أي شيء إضافي بخلاف تعيين قيمة المعامل.
## المُحسِّنات
يمكنك اختيار مُحسِّن مدمج للتدريب باستخدام:
```python
from transformers import TrainingArguments
training_args = TrainingArguments(..., optim="adamw_torch")
```
اطلع على [`OptimizerNames`](https://github.com/huggingface/transformers/blob/main/src/transformers/training_args.py) للاطلاع على القائمة الكاملة للخيارات. نُدرج أمثلة مُتقدمة في الأقسام أدناه.
يمكنك أيضًا استخدام مُحسِّن PyTorch عشوائي عبر:
```python
import torch
optimizer_cls = torch.optim.AdamW
optimizer_kwargs = {
"lr": 4e-3,
"betas": (0.9, 0.999),
"weight_decay": 0.05,
}
from transformers import Trainer
trainer = Trainer(..., optimizer_cls_and_kwargs=(optimizer_cls, optimizer_kwargs))
```
### GaLore
إسقاط التدرج ذو الرتبة المنخفضة (GaLore) هو إستراتيجية تدريب ذات رتبة منخفضة فعّالة من حيث الذاكرة، تسمح بتعلم المعلمات الكاملة ولكنها أكثر كفاءة من حيث الذاكرة من أساليب التكيّف الشائعة ذات الرتبة المنخفضة، مثل LoRA.
أولاً، تأكد من تثبيت المستودع الرسمي لـ GaLore:
```bash
pip install galore-torch
```
ثم أضف ببساطة أحد `["galore_adamw"، "galore_adafactor"، "galore_adamw_8bit"]` في `optim` جنبًا إلى جنب مع `optim_target_modules`، والتي يمكن أن تكون قائمة من السلاسل أو التعبيرات النمطية regex أو المسار الكامل المطابق لأسماء الوحدات المستهدفة التي تريد تكييفها. فيما يلي مثال على النص البرمجي كامل(تأكد من `pip install trl datasets`):
```python
import datasets
from trl import SFTConfig, SFTTrainer
train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
output_dir="./test-galore",
max_steps=100,
optim="galore_adamw",
optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
gradient_checkpointing=True,
)
trainer = SFTTrainer(
model="google/gemma-2b",
args=args,
train_dataset=train_dataset,
)
trainer.train()
```
لتمرير معامﻻت إضافية يدعمها GaLore، يجب عليك تمرير `optim_args` بشكل صحيح، على سبيل المثال:
```python
import datasets
from trl import SFTConfig, SFTTrainer
train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
output_dir="./test-galore",
max_steps=100,
optim="galore_adamw",
optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
optim_args="rank=64, update_proj_gap=100, scale=0.10",
gradient_checkpointing=True,
)
trainer = SFTTrainer(
model="google/gemma-2b",
args=args,
train_dataset=train_dataset,
)
trainer.train()
```
يمكنك قراءة المزيد حول الطريقة في [المستودع الأصلي](https://github.com/jiaweizzhao/GaLore) أو [الورقة البحثية](https://huggingface.co/papers/2403.03507).
حاليًا، يمكنك فقط تدريب الطبقات الخطية التي تعتبر طبقات GaLore وستستخدم التحلل ذو الرتبة المنخفضة للتدريب بينما سيتم تحسين الطبقات المتبقية بالطريقة التقليدية.
لاحظ أنه سيستغرق الأمر بعض الوقت قبل بدء التدريب (~3 دقائق لنموذج 2B على NVIDIA A100)، ولكن يجب أن يسير التدريب بسلاسة بعد ذلك.
يمكنك أيضًا إجراء تحسين طبقة تلو الأخرى عن طريق إضافة `layerwise` إلى اسم المُحسِّن كما هو موضح أدناه:
```python
import datasets
from trl import SFTConfig, SFTTrainer
train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
output_dir="./test-galore",
max_steps=100,
optim="galore_adamw_layerwise",
optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
gradient_checkpointing=True,
)
trainer = SFTTrainer(
model="google/gemma-2b",
args=args,
train_dataset=train_dataset,
)
trainer.train()
```
لاحظ أن تحسين الطبقة تجريبي إلى حد ما ولا يدعم DDP (Distributed Data Parallel)، وبالتالي يمكنك تشغيل التعليمات البرمجية للتدريب على وحدة معالجة الرسومات (GPU) واحدة فقط. يرجى الاطلاع على [هذا القسم المناسب](https://github.com/jiaweizzhao/GaLore?tab=readme-ov-file#train-7b-model-with-a-single-gpu-with-24gb-memory) لمزيد من التفاصيل. قد لا تدعم الميزات الأخرى مثل تقليم التدرجات أو DeepSpeed، إلخ. من الصندوق. يرجى [تقديم تقرير عن المشكلة على GitHub](https://github.com/huggingface/transformers/issues) إذا واجهتك مثل هذه المشكلة.
### محسنات LOMO
تم تقديم مُحسِّنات LOMO في [التدريب على المعلمات الكاملة لنماذج اللغة الكبيرة باستخدام موارد محدودة](https://hf.co/papers/2306.09782) و [AdaLomo: تحسين ذاكرة منخفضة بمعدل تعلم متكيف](https://hf.co/papers/2310.10195).
يتكون كلاهما من طريقة فعالة لضبط المعلمات الكاملة. تدمج محسنات LOMO حساب الاشتقاق وتحديث المعلمات في خطوة واحدة لتقليل استخدام الذاكرة. محسنات LOMO المدعومة هي `"lomo"` و `"adalomo"`. أولاً قم بتثبيت LOMO من pypi `pip install lomo-optim` أو قم بتثبيته من المصدر باستخدام `pip install git+https://github.com/OpenLMLab/LOMO.git`.
<Tip>
وفقًا للمؤلفين، يوصى باستخدام `AdaLomo` بدون `grad_norm` للحصول على أداء أفضل وسرعة أعلى.
</Tip>
فيما يلي نص برمجي بسيط يوضح كيفية ضبط نموذج [google/gemma-2b](https://huggingface.co/google/gemma-2b) على مجموعة بيانات IMDB في الدقة الكاملة:
```python
import datasets
from trl import SFTConfig, SFTTrainer
train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
output_dir="./test-lomo",
max_steps=100,
optim="adalomo",
gradient_checkpointing=True,
)
trainer = SFTTrainer(
model="google/gemma-2b",
args=args,
train_dataset=train_dataset,
)
trainer.train()
```
### مُحسِّن GrokAdamW
تم تصميم مُحسِّن GrokAdamW لتعزيز أداء التدريب واستقراره، خاصةً للنماذج التي تستفيد من دوال إشارة `grokking`. لاستخدام `GrokAdamW`، قم أولاً بتثبيت حزمة المُحسِّن باستخدام `pip install grokadamw`.
<Tip>
يُعد GrokAdamW مفيدًا بشكل خاص للنماذج التي تتطلب تقنيات تحسين مُتقدمة لتحقيق أداء واستقرار أفضل.
</Tip>
فيما يلي نص برمجى بسيط لشرح كيفية ضبط [google/gemma-2b](https://huggingface.co/google/gemma-2b) بدقة على مجموعة بيانات IMDB باستخدام مُحسِّن GrokAdamW:
```python
import torch
import datasets
from transformers import TrainingArguments, AutoTokenizer, AutoModelForCausalLM, Trainer
# تحميل مجموعة البيانات IMDB
train_dataset = datasets.load_dataset('imdb', split='train')
# تعريف معامﻻت التدريب
args = TrainingArguments(
output_dir="./test-grokadamw",
max_steps=1000,
per_device_train_batch_size=4,
optim="grokadamw",
logging_strategy="steps",
logging_steps=1,
learning_rate=2e-5,
save_strategy="no",
run_name="grokadamw-imdb",
)
# تحميل النموذج والمجزىء اللغوي
model_id = "google/gemma-2b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id).to(0)
# تهيئة المدرب
trainer = Trainer(
model=model,
args=args,
train_dataset=train_dataset,
)
# تدريب النموذج
trainer.train()
```
يوضح هذا النص البرمجى كيفية ضبط نموذج google/gemma-2b بدقة على مجموعة بيانات IMDB باستخدام مُحسِّن GrokAdamW. يتم تكوين TrainingArguments لاستخدام GrokAdamW، ويتم تمرير مجموعة البيانات إلى Trainer للتدريب.
### مُحسِّن بدون جدوله (Schedule Free Optimizer)
تم تقديم مُحسِّنات بدون جدوله في [The Road Less Scheduled](https://hf.co/papers/2405.15682).
يستبدل التعلم بدون جدوله زخم المُحسِّن الأساسي بمزيج من المتوسط والتداخل، لإزالة الحاجة تمامًا إلى تخفيف مُعدل التعلم باستخدام جدوله تقليديه.
المُحسِّنات المدعومة لـ SFO هي "schedule_free_adamw" و "schedule_free_sgd". قم أولاً بتثبيت `schedulefree` من pypi باستخدام الأمر `pip install schedulefree`.
فيما يلي نص برمجى بسيط لشرح كيفية ضبط [google/gemma-2b](https://huggingface.co/google/gemma-2b) بدقة على مجموعة بيانات IMDB بدقة كاملة:
```python
import datasets
from trl import SFTConfig, SFTTrainer
train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
output_dir="./test-galore",
max_steps=100,
optim="schedule_free_adamw",
gradient_checkpointing=True,
)
trainer = SFTTrainer(
model="google/gemma-2b",
args=args,
train_dataset=train_dataset,
)
trainer.train()
```
## تسريع ومدرب
يتم تشغيل فئة [`Trainer`] بواسطة [تسريع](https://hf.co/docs/accelerate)، وهي مكتبة لتدريب نماذج PyTorch بسهولة في بيئات موزعة مع دعم عمليات التكامل مثل [FullyShardedDataParallel (FSDP)](https://pytorch.org/blog/introducing-pytorch-fully-sharded-data-parallel-api/) و [DeepSpeed](https://www.deepspeed.ai/).
<Tip>
تعرف على المزيد حول استراتيجيات تجزئة FSDP، وتفريغ وحدة المعالجة المركزية (CPU)، والمزيد مع [`Trainer`] في [دليل Fully Sharded Data Parallel](fsdp).
</Tip>
لاستخدام Accelerate مع [`Trainer`]]، قم بتشغيل الأمر [`accelerate.config`](https://huggingface.co/docs/accelerate/package_reference/cli#accelerate-config) لإعداد التدريب لبيئة التدريب الخاصة بك. نشئ هذا الأمر ملف `config_file.yaml` الذي سيتم استخدامه عند تشغيل نص للتدريب البرمجى. على سبيل المثال، بعض تكوينات المثال التي يمكنك إعدادها هي:
<hfoptions id="config">
<hfoption id="DistributedDataParallel">
```yml
compute_environment: LOCAL_MACHINE
distributed_type: MULTI_GPU
downcast_bf16: 'no'
gpu_ids: all
machine_rank: 0 #change rank as per the node
main_process_ip: 192.168.20.1
main_process_port: 9898
main_training_function: main
mixed_precision: fp16
num_machines: 2
num_processes: 8
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false
```
</hfoption>
<hfoption id="FSDP">
```yml
compute_environment: LOCAL_MACHINE
distributed_type: FSDP
downcast_bf16: 'no'
fsdp_config:
fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
fsdp_backward_prefetch_policy: BACKWARD_PRE
fsdp_forward_prefetch: true
fsdp_offload_params: false
fsdp_sharding_strategy: 1
fsdp_state_dict_type: FULL_STATE_DICT
fsdp_sync_module_states: true
fsdp_transformer_layer_cls_to_wrap: BertLayer
fsdp_use_orig_params: true
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 2
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false
```
</hfoption>
<hfoption id="DeepSpeed">
```yml
compute_environment: LOCAL_MACHINE
deepspeed_config:
deepspeed_config_file: /home/user/configs/ds_zero3_config.json
zero3_init_flag: true
distributed_type: DEEPSPEED
downcast_bf16: 'no'
machine_rank: 0
main_training_function: main
num_machines: 1
num_processes: 4
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false
```
</hfoption>
<hfoption id="DeepSpeed with Accelerate plugin">
```yml
compute_environment: LOCAL_MACHINE
deepspeed_config:
gradient_accumulation_steps: 1
gradient_clipping: 0.7
offload_optimizer_device: cpu
offload_param_device: cpu
zero3_init_flag: true
zero_stage: 2
distributed_type: DEEPSPEED
downcast_bf16: 'no'
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 4
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false
```
</hfoption>
</hfoptions>
يُعد أمر [`accelerate_launch`](https://huggingface.co/docs/accelerate/package_reference/cli#accelerate-launch) هو الطريقة المُوصى بها لتشغيل نص البرمجى للتدريب على نظام موزع باستخدام Accelerate و [`Trainer`] مع المعلمات المحددة في `config_file.yaml`. يتم حفظ هذا الملف في مجلد ذاكرة التخزين المؤقت لـ Accelerate ويتم تحميله تلقائيًا عند تشغيل `accelerate_launch`.
على سبيل المثال، لتشغيل النص البرنامجي للتدريب [run_glue.py](https://github.com/huggingface/transformers/blob/f4db565b695582891e43a5e042e5d318e28f20b8/examples/pytorch/text-classification/run_glue.py#L4) مع تكوين FSDP:
```bash
accelerate launch \
./examples/pytorch/text-classification/run_glue.py \
--model_name_or_path google-bert/bert-base-cased \
--task_name $TASK_NAME \
--do_train \
--do_eval \
--max_seq_length 128 \
--per_device_train_batch_size 16 \
--learning_rate 5e-5 \
--num_train_epochs 3 \
--output_dir /tmp/$TASK_NAME/ \
--overwrite_output_dir
```
يمكنك أيضًا تحديد المعلمات من ملف `config_file.yaml` مباشرة في سطر الأوامر:
```bash
accelerate launch --num_processes=2 \
--use_fsdp \
--mixed_precision=bf16 \
--fsdp_auto_wrap_policy=TRANSFORMER_BASED_WRAP \
--fsdp_transformer_layer_cls_to_wrap="BertLayer" \
--fsdp_sharding_strategy=1 \
--fsdp_state_dict_type=FULL_STATE_DICT \
./examples/pytorch/text-classification/run_glue.py
--model_name_or_path google-bert/bert-base-cased \
--task_name $TASK_NAME \
--do_train \
--do_eval \
--max_seq_length 128 \
--per_device_train_batch_size 16 \
--learning_rate 5e-5 \
--num_train_epochs 3 \
--output_dir /tmp/$TASK_NAME/ \
--overwrite_output_dir
```
اطلع على برنامج تعليمي [Launching your Accelerate scripts](https://huggingface.co/docs/accelerate/basic_tutorials/launch) لمعرفة المزيد حول `accelerate_launch` والتكوينات المخصصة.
| transformers/docs/source/ar/trainer.md/0 | {
"file_path": "transformers/docs/source/ar/trainer.md",
"repo_id": "transformers",
"token_count": 16974
} | 351 |
<!---
Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Überprüfungen bei einer Pull-Anfrage
Wenn Sie eine Pull-Anfrage für 🤗 Transformers öffnen, wird eine ganze Reihe von Prüfungen durchgeführt, um sicherzustellen, dass der Patch, den Sie hinzufügen, nichts Bestehendes zerstört. Es gibt vier Arten von Prüfungen:
- reguläre Tests
- Erstellung der Dokumentation
- Stil von Code und Dokumentation
- allgemeine Konsistenz des Repository
In diesem Dokument werden wir versuchen zu erklären, worum es sich bei diesen verschiedenen Prüfungen handelt und wie Sie sie lokal debuggen können, wenn eine der Prüfungen in Ihrer PR fehlschlägt.
Beachten Sie, dass Sie im Idealfall eine Dev-Installation benötigen:
```bash
pip install transformers[dev]
```
oder für eine bearbeitbare Installation:
```bash
pip install -e .[dev]
```
innerhalb des Transformers Repo. Da die Anzahl der optionalen Abhängigkeiten von Transformers stark zugenommen hat, ist es möglich, dass Sie nicht alle davon bekommen können. Wenn die Dev-Installation fehlschlägt, stellen Sie sicher, dass Sie das Deep Learning-Framework, mit dem Sie arbeiten, installieren (PyTorch, TensorFlow und/oder Flax).
```bash
pip install transformers[quality]
```
oder für eine bearbeitbare Installation:
```bash
pip install -e .[quality]
```
## Tests
Alle Jobs, die mit `ci/circleci: run_tests_` beginnen, führen Teile der Transformers-Testsuite aus. Jeder dieser Jobs konzentriert sich auf einen Teil der Bibliothek in einer bestimmten Umgebung: `ci/circleci: run_tests_pipelines_tf` zum Beispiel führt den Pipelines-Test in einer Umgebung aus, in der nur TensorFlow installiert ist.
Beachten Sie, dass nur ein Teil der Testsuite jedes Mal ausgeführt wird, um zu vermeiden, dass Tests ausgeführt werden, wenn es keine wirkliche Änderung in den Modulen gibt, die sie testen: ein Dienstprogramm wird ausgeführt, um die Unterschiede in der Bibliothek zwischen vor und nach dem PR zu ermitteln (was GitHub Ihnen auf der Registerkarte "Files changes" anzeigt) und die Tests auszuwählen, die von diesem Unterschied betroffen sind. Dieses Dienstprogramm kann lokal mit ausgeführt werden:
```bash
python utils/tests_fetcher.py
```
aus dem Stammverzeichnis des Transformers-Repositoriums. Es wird:
1. Überprüfen Sie für jede Datei im Diff, ob die Änderungen im Code oder nur in Kommentaren oder Docstrings enthalten sind. Nur die Dateien mit echten Codeänderungen werden beibehalten.
2. Erstellen Sie eine interne Map, die für jede Datei des Quellcodes der Bibliothek alle Dateien angibt, auf die sie rekursiv Einfluss nimmt. Von Modul A wird gesagt, dass es sich auf Modul B auswirkt, wenn Modul B Modul A importiert. Für die rekursive Auswirkung benötigen wir eine Kette von Modulen, die von Modul A zu Modul B führt und in der jedes Modul das vorherige importiert.
3. Wenden Sie diese Zuordnung auf die in Schritt 1 gesammelten Dateien an. So erhalten wir die Liste der Modelldateien, die von der PR betroffen sind.
4. Ordnen Sie jede dieser Dateien der/den entsprechenden Testdatei(en) zu und erhalten Sie die Liste der auszuführenden Tests.
Wenn Sie das Skript lokal ausführen, sollten Sie die Ergebnisse von Schritt 1, 3 und 4 ausgegeben bekommen und somit wissen, welche Tests ausgeführt werden. Das Skript erstellt außerdem eine Datei namens `test_list.txt`, die die Liste der auszuführenden Tests enthält, die Sie mit dem folgenden Befehl lokal ausführen können:
```bash
python -m pytest -n 8 --dist=loadfile -rA -s $(cat test_list.txt)
```
Für den Fall, dass Ihnen etwas entgangen ist, wird die komplette Testreihe ebenfalls täglich ausgeführt.
## Dokumentation erstellen
Der Job `build_pr_documentation` erstellt und generiert eine Vorschau der Dokumentation, um sicherzustellen, dass alles in Ordnung ist, wenn Ihr PR zusammengeführt wird. Ein Bot fügt einen Link zur Vorschau der Dokumentation zu Ihrem PR hinzu. Alle Änderungen, die Sie an dem PR vornehmen, werden automatisch in der Vorschau aktualisiert. Wenn die Dokumentation nicht erstellt werden kann, klicken Sie auf **Details** neben dem fehlgeschlagenen Auftrag, um zu sehen, wo der Fehler liegt. Oft ist der Fehler so einfach wie eine fehlende Datei im `toctree`.
Wenn Sie daran interessiert sind, die Dokumentation lokal zu erstellen oder in der Vorschau anzusehen, werfen Sie einen Blick in die [`README.md`](https://github.com/huggingface/transformers/tree/main/docs) im Ordner docs.
## Code und Dokumentationsstil
Die Formatierung des Codes erfolgt für alle Quelldateien, die Beispiele und die Tests mit `black` und `ruff`. Wir haben auch ein benutzerdefiniertes Tool, das sich um die Formatierung von docstrings und `rst`-Dateien kümmert (`utils/style_doc.py`), sowie um die Reihenfolge der Lazy-Importe, die in den Transformers `__init__.py`-Dateien durchgeführt werden (`utils/custom_init_isort.py`). All dies können Sie starten, indem Sie Folgendes ausführen
```bash
make style
```
Das CI prüft, ob diese innerhalb der Prüfung `ci/circleci: check_code_quality` angewendet wurden. Es führt auch `ruff` aus, das einen grundlegenden Blick auf Ihren Code wirft und sich beschwert, wenn es eine undefinierte Variable findet oder eine, die nicht verwendet wird. Um diese Prüfung lokal auszuführen, verwenden Sie
```bash
make quality
```
Dies kann sehr viel Zeit in Anspruch nehmen. Um dasselbe nur für die Dateien zu tun, die Sie im aktuellen Zweig geändert haben, führen Sie
```bash
make fixup
```
Dieser letzte Befehl führt auch alle zusätzlichen Prüfungen für die Konsistenz des Repositorys durch. Schauen wir uns diese an.
## Repository-Konsistenz
Dies fasst alle Tests zusammen, die sicherstellen, dass Ihr PR das Repository in einem guten Zustand verlässt. Sie können diese Prüfung lokal durchführen, indem Sie Folgendes ausführen:
```bash
make repo-consistency
```
Dies überprüft, ob:
- Alle zum Init hinzugefügten Objekte sind dokumentiert (ausgeführt von `utils/check_repo.py`)
- Alle `__init__.py`-Dateien haben in ihren beiden Abschnitten den gleichen Inhalt (ausgeführt von `utils/check_inits.py`)
- Der gesamte Code, der als Kopie eines anderen Moduls identifiziert wurde, stimmt mit dem Original überein (ausgeführt von `utils/check_copies.py`)
- Alle Konfigurationsklassen haben mindestens einen gültigen Prüfpunkt, der in ihren Dokumentationen erwähnt wird (ausgeführt von `utils/check_config_docstrings.py`)
- Alle Konfigurationsklassen enthalten nur Attribute, die in den entsprechenden Modellierungsdateien verwendet werden (ausgeführt von `utils/check_config_attributes.py`)
- Die Übersetzungen der READMEs und der Index des Dokuments haben die gleiche Modellliste wie die Haupt-README (durchgeführt von `utils/check_copies.py`)
- Die automatisch generierten Tabellen in der Dokumentation sind auf dem neuesten Stand (ausgeführt von `utils/check_table.py`)
- Die Bibliothek verfügt über alle Objekte, auch wenn nicht alle optionalen Abhängigkeiten installiert sind (ausgeführt von `utils/check_dummies.py`)
Sollte diese Prüfung fehlschlagen, müssen die ersten beiden Punkte manuell korrigiert werden, die letzten vier können automatisch für Sie korrigiert werden, indem Sie den Befehl
```bash
make fix-copies
```
Zusätzliche Prüfungen betreffen PRs, die neue Modelle hinzufügen, vor allem, dass:
- Alle hinzugefügten Modelle befinden sich in einer Auto-Zuordnung (durchgeführt von `utils/check_repo.py`)
<!-- TODO Sylvain, add a check that makes sure the common tests are implemented.-->
- Alle Modelle werden ordnungsgemäß getestet (ausgeführt von `utils/check_repo.py`)
<!-- TODO Sylvain, add the following
- All models are added to the main README, inside the main doc
- All checkpoints used actually exist on the Hub
-->
### Kopien prüfen
Da die Transformers-Bibliothek in Bezug auf den Modellcode sehr eigenwillig ist und jedes Modell vollständig in einer einzigen Datei implementiert sein sollte, ohne sich auf andere Modelle zu stützen, haben wir einen Mechanismus hinzugefügt, der überprüft, ob eine Kopie des Codes einer Ebene eines bestimmten Modells mit dem Original übereinstimmt. Auf diese Weise können wir bei einer Fehlerbehebung alle anderen betroffenen Modelle sehen und entscheiden, ob wir die Änderung weitergeben oder die Kopie zerstören.
<Tip>
Wenn eine Datei eine vollständige Kopie einer anderen Datei ist, sollten Sie sie in der Konstante `FULL_COPIES` von `utils/check_copies.py` registrieren.
</Tip>
Dieser Mechanismus stützt sich auf Kommentare der Form `# Kopiert von xxx`. Das `xxx` sollte den gesamten Pfad zu der Klasse der Funktion enthalten, die darunter kopiert wird. Zum Beispiel ist `RobertaSelfOutput` eine direkte Kopie der Klasse `BertSelfOutput`. Sie können also [hier](https://github.com/huggingface/transformers/blob/2bd7a27a671fd1d98059124024f580f8f5c0f3b5/src/transformers/models/roberta/modeling_roberta.py#L289) sehen, dass sie einen Kommentar hat:
```py
# Copied from transformers.models.bert.modeling_bert.BertSelfOutput
```
Beachten Sie, dass Sie dies nicht auf eine ganze Klasse anwenden, sondern auf die entsprechenden Methoden, von denen kopiert wird. Zum Beispiel [hier](https://github.com/huggingface/transformers/blob/2bd7a27a671fd1d98059124024f580f8f5c0f3b5/src/transformers/models/roberta/modeling_roberta.py#L598) können Sie sehen, wie `RobertaPreTrainedModel._init_weights` von der gleichen Methode in `BertPreTrainedModel` mit dem Kommentar kopiert wird:
```py
# Copied from transformers.models.bert.modeling_bert.BertPreTrainedModel._init_weights
```
Manchmal ist die Kopie bis auf die Namen genau gleich: zum Beispiel verwenden wir in `RobertaAttention` `RobertaSelfAttention` anstelle von `BertSelfAttention`, aber ansonsten ist der Code genau derselbe. Aus diesem Grund unterstützt `#Copied from` einfache String-Ersetzungen mit der folgenden Syntax: `Kopiert von xxx mit foo->bar`. Das bedeutet, dass der Code kopiert wird, wobei alle Instanzen von "foo" durch "bar" ersetzt werden. Sie können sehen, wie es [hier](https://github.com/huggingface/transformers/blob/2bd7a27a671fd1d98059124024f580f8f5c0f3b5/src/transformers/models/roberta/modeling_roberta.py#L304C1-L304C86) in `RobertaAttention` mit dem Kommentar verwendet wird:
```py
# Copied from transformers.models.bert.modeling_bert.BertAttention with Bert->Roberta
```
Beachten Sie, dass um den Pfeil herum keine Leerzeichen stehen sollten (es sei denn, das Leerzeichen ist Teil des zu ersetzenden Musters, natürlich).
Sie können mehrere Muster durch ein Komma getrennt hinzufügen. Zum Beispiel ist hier `CamemberForMaskedLM` eine direkte Kopie von `RobertaForMaskedLM` mit zwei Ersetzungen: `Roberta` zu `Camembert` und `ROBERTA` zu `CAMEMBERT`. Sie können [hier](https://github.com/huggingface/transformers/blob/15082a9dc6950ecae63a0d3e5060b2fc7f15050a/src/transformers/models/camembert/modeling_camembert.py#L929) sehen, wie dies mit dem Kommentar gemacht wird:
```py
# Copied from transformers.models.roberta.modeling_roberta.RobertaForMaskedLM with Roberta->Camembert, ROBERTA->CAMEMBERT
```
Wenn die Reihenfolge eine Rolle spielt (weil eine der Ersetzungen mit einer vorherigen in Konflikt geraten könnte), werden die Ersetzungen von links nach rechts ausgeführt.
<Tip>
Wenn die Ersetzungen die Formatierung ändern (wenn Sie z.B. einen kurzen Namen durch einen sehr langen Namen ersetzen), wird die Kopie nach Anwendung des automatischen Formats überprüft.
</Tip>
Eine andere Möglichkeit, wenn es sich bei den Mustern nur um verschiedene Umschreibungen derselben Ersetzung handelt (mit einer groß- und einer kleingeschriebenen Variante), besteht darin, die Option `all-casing` hinzuzufügen. [Hier](https://github.com/huggingface/transformers/blob/15082a9dc6950ecae63a0d3e5060b2fc7f15050a/src/transformers/models/mobilebert/modeling_mobilebert.py#L1237) ist ein Beispiel in `MobileBertForSequenceClassification` mit dem Kommentar:
```py
# Copied from transformers.models.bert.modeling_bert.BertForSequenceClassification with Bert->MobileBert all-casing
```
In diesem Fall wird der Code von `BertForSequenceClassification` kopiert, indem er ersetzt wird:
- `Bert` durch `MobileBert` (zum Beispiel bei der Verwendung von `MobileBertModel` in der Init)
- `bert` durch `mobilebert` (zum Beispiel bei der Definition von `self.mobilebert`)
- `BERT` durch `MOBILEBERT` (in der Konstante `MOBILEBERT_INPUTS_DOCSTRING`)
| transformers/docs/source/de/pr_checks.md/0 | {
"file_path": "transformers/docs/source/de/pr_checks.md",
"repo_id": "transformers",
"token_count": 4986
} | 352 |
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Backbones
Higher-level computer visions tasks, such as object detection or image segmentation, use several models together to generate a prediction. A separate model is used for the *backbone*, neck, and head. The backbone extracts useful features from an input image into a feature map, the neck combines and processes the feature maps, and the head uses them to make a prediction.
<div class="flex justify-center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/Backbone.png"/>
</div>
Load a backbone with [`~PretrainedConfig.from_pretrained`] and use the `out_indices` parameter to determine which layer, given by the index, to extract a feature map from.
```py
from transformers import AutoBackbone
model = AutoBackbone.from_pretrained("microsoft/swin-tiny-patch4-window7-224", out_indices=(1,))
```
This guide describes the backbone class, backbones from the [timm](https://hf.co/docs/timm/index) library, and how to extract features with them.
## Backbone classes
There are two backbone classes.
- [`~transformers.utils.BackboneMixin`] allows you to load a backbone and includes functions for extracting the feature maps and indices.
- [`~transformers.utils.BackboneConfigMixin`] allows you to set the feature map and indices of a backbone configuration.
Refer to the [Backbone](./main_classes/backbones) API documentation to check which models support a backbone.
There are two ways to load a Transformers backbone, [`AutoBackbone`] and a model-specific backbone class.
<hfoptions id="backbone-classes">
<hfoption id="AutoBackbone">
The [AutoClass](./model_doc/auto) API automatically loads a pretrained vision model with [`~PretrainedConfig.from_pretrained`] as a backbone if it's supported.
Set the `out_indices` parameter to the layer you'd like to get the feature map from. If you know the name of the layer, you could also use `out_features`. These parameters can be used interchangeably, but if you use both, make sure they refer to the same layer.
When `out_indices` or `out_features` isn't used, the backbone returns the feature map from the last layer. The example code below uses `out_indices=(1,)` to get the feature map from the first layer.
<div class="flex justify-center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/Swin%20Stage%201.png"/>
</div>
```py
from transformers import AutoImageProcessor, AutoBackbone
model = AutoBackbone.from_pretrained("microsoft/swin-tiny-patch4-window7-224", out_indices=(1,))
```
</hfoption>
<hfoption id="model-specific backbone">
When you know a model supports a backbone, you can load the backbone and neck directly into the models configuration. Pass the configuration to the model to initialize it for a task.
The example below loads a [ResNet](./model_doc/resnet) backbone and neck for use in a [MaskFormer](./model_doc/maskformer) instance segmentation head.
Set `backbone` to a pretrained model and `use_pretrained_backbone=True` to use pretrained weights instead of randomly initialized weights.
```py
from transformers import MaskFormerConfig, MaskFormerForInstanceSegmentation
config = MaskFormerConfig(backbone="microsoft/resnet-50", use_pretrained_backbone=True)
model = MaskFormerForInstanceSegmentation(config)
```
Another option is to separately load the backbone configuration and then pass it to `backbone_config` in the model configuration.
```py
from transformers import MaskFormerConfig, MaskFormerForInstanceSegmentation, ResNetConfig
# instantiate backbone configuration
backbone_config = ResNetConfig()
# load backbone in model
config = MaskFormerConfig(backbone_config=backbone_config)
# attach backbone to model head
model = MaskFormerForInstanceSegmentation(config)
```
</hfoption>
</hfoptions>
## timm backbones
[timm](https://hf.co/docs/timm/index) is a collection of vision models for training and inference. Transformers supports timm models as backbones with the [`TimmBackbone`] and [`TimmBackboneConfig`] classes.
Set `use_timm_backbone=True` to load pretrained timm weights, and `use_pretrained_backbone` to use pretrained or randomly initialized weights.
```py
from transformers import MaskFormerConfig, MaskFormerForInstanceSegmentation
config = MaskFormerConfig(backbone="resnet50", use_timm_backbone=True, use_pretrained_backbone=True)
model = MaskFormerForInstanceSegmentation(config)
```
You could also explicitly call the [`TimmBackboneConfig`] class to load and create a pretrained timm backbone.
```py
from transformers import TimmBackboneConfig
backbone_config = TimmBackboneConfig("resnet50", use_pretrained_backbone=True)
```
Pass the backbone configuration to the model configuration and instantiate the model head, [`MaskFormerForInstanceSegmentation`], with the backbone.
```py
from transformers import MaskFormerConfig, MaskFormerForInstanceSegmentation
config = MaskFormerConfig(backbone_config=backbone_config)
model = MaskFormerForInstanceSegmentation(config)
```
## Feature extraction
The backbone is used to extract image features. Pass an image through the backbone to get the feature maps.
Load and preprocess an image and pass it to the backbone. The example below extracts the feature maps from the first layer.
```py
from transformers import AutoImageProcessor, AutoBackbone
import torch
from PIL import Image
import requests
model = AutoBackbone.from_pretrained("microsoft/swin-tiny-patch4-window7-224", out_indices=(1,))
processor = AutoImageProcessor.from_pretrained("microsoft/swin-tiny-patch4-window7-224")
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)
inputs = processor(image, return_tensors="pt")
outputs = model(**inputs)
```
The features are stored and accessed from the outputs `feature_maps` attribute.
```py
feature_maps = outputs.feature_maps
list(feature_maps[0].shape)
[1, 96, 56, 56]
```
| transformers/docs/source/en/backbones.md/0 | {
"file_path": "transformers/docs/source/en/backbones.md",
"repo_id": "transformers",
"token_count": 1850
} | 353 |
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# FullyShardedDataParallel
[Fully Sharded Data Parallel (FSDP)](https://pytorch.org/blog/introducing-pytorch-fully-sharded-data-parallel-api/) is a [parallelism](./perf_train_gpu_many) method that combines the advantages of data and model parallelism for distributed training.
Unlike [DistributedDataParallel (DDP)](./perf_train_gpu_many#distributeddataparallel), FSDP saves more memory because it doesn't replicate a model on each GPU. It shards the models parameters, gradients and optimizer states across GPUs. Each model shard processes a portion of the data and the results are synchronized to speed up training.
This guide covers how to set up training a model with FSDP and [Accelerate](https://hf.co/docs/accelerate/index), a library for managing distributed training.
```bash
pip install accelerate
```
## Configuration options
Always start by running the [accelerate config](https://hf.co/docs/accelerate/package_reference/cli#accelerate-config) command to help Accelerate set up the correct distributed training environment.
```bash
accelerate config
```
The section below discusses some of the more important FSDP configuration options. Learn more about other available options in the [fsdp_config](https://hf.co/docs/transformers/main_classes/trainer#transformers.TrainingArguments.fsdp_config) parameter.
### Sharding strategy
FSDP offers several sharding strategies to distribute a model. Refer to the table below to help you choose the best strategy for your setup. Specify a strategy with the `fsdp_sharding_strategy` parameter in the configuration file.
| sharding strategy | description | parameter value |
|---|---|---|
| `FULL_SHARD` | shards model parameters, gradients, and optimizer states | `1` |
| `SHARD_GRAD_OP` | shards gradients and optimizer states | `2` |
| `NO_SHARD` | don't shard the model | `3` |
| `HYBRID_SHARD` | shards model parameters, gradients, and optimizer states within each GPU | `4` |
| `HYBRID_SHARD_ZERO2` | shards gradients and optimizer states within each GPU | `5` |
### CPU offload
Offload model parameters and gradients when they aren't being used to the CPU to save additional GPU memory. This is useful for scenarios where a model is too large even with FSDP.
Specify `fsdp_offload_params: true` in the configuration file to enable offloading.
### Wrapping policy
FSDP is applied by wrapping each layer in the network. The wrapping is usually applied in a nested way where the full weights are discarded after each forward pass to save memory for the next layer.
There are several wrapping policies available, but the *auto wrapping* policy is the simplest and doesn't require any changes to your code. Specify `fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP` to wrap a Transformer layer and `fsdp_transformer_layer_cls_to_wrap` to determine which layer to wrap (for example, `BertLayer`).
Size-based wrapping is also available. If a layer exceeds a certain number of parameters, it is wrapped. Specify `fsdp_wrap_policy: SIZED_BASED_WRAP` and `min_num_param` to set the minimum number of parameters for a layer to be wrapped.
### Checkpoints
Intermediate checkpoints should be saved as a sharded state dict because saving the full state dict - even with CPU offloading - is time consuming and can cause `NCCL Timeout` errors due to indefinite hanging during broadcasting.
Specify `fsdp_state_dict_type: SHARDED_STATE_DICT` in the configuration file to save the sharded state dict. Now you can resume training from the sharded state dict with [`~accelerate.Accelerator.load_state`].
```py
accelerator.load_state("directory/containing/checkpoints")
```
Once training is complete though, you should save the full state dict because the sharded state dict is only compatible with FSDP.
```py
if trainer.is_fsdp_enabled:
trainer.accelerator.state.fsdp_plugin.set_state_dict_type("FULL_STATE_DICT")
trainer.save_model(script_args.output_dir)
```
### TPU
[PyTorch XLA](https://pytorch.org/xla/release/2.1/index.html), a package for running PyTorch on XLA devices, enables FSDP on TPUs. Modify the configuration file to include the parameters below. Refer to the [xla_fsdp_settings](https://github.com/pytorch/xla/blob/2e6e183e0724818f137c8135b34ef273dea33318/torch_xla/distributed/fsdp/xla_fully_sharded_data_parallel.py#L128) parameter for additional XLA-specific parameters you can configure for FSDP.
```yaml
xla: True # must be set to True to enable PyTorch/XLA
xla_fsdp_settings: # XLA specific FSDP parameters
xla_fsdp_grad_ckpt: True # enable gradient checkpointing
```
## Training
After running [accelerate config](https://hf.co/docs/accelerate/package_reference/cli#accelerate-config), your configuration file should be ready. An example configuration file is shown below that fully shards the parameter, gradient and optimizer states on two GPUs. Your file may look different depending on how you set up your configuration.
```yaml
compute_environment: LOCAL_MACHINE
debug: false
distributed_type: FSDP
downcast_bf16: 'no'
fsdp_config:
fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
fsdp_backward_prefetch_policy: BACKWARD_PRE
fsdp_cpu_ram_efficient_loading: true
fsdp_forward_prefetch: false
fsdp_offload_params: true
fsdp_sharding_strategy: 1
fsdp_state_dict_type: SHARDED_STATE_DICT
fsdp_sync_module_states: true
fsdp_transformer_layer_cls_to_wrap: BertLayer
fsdp_use_orig_params: true
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 2
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false
```
Run the [accelerate launch](https://hf.co/docs/accelerate/package_reference/cli#accelerate-launch) command to launch a training script with the FSDP configurations you chose in the configuration file.
```bash
accelerate launch my-training-script.py
```
It is also possible to directly specify some of the FSDP arguments in the command line.
```bash
accelerate launch --fsdp="full shard" --fsdp_config="path/to/fsdp_config/" my-training-script.py
```
## Resources
FSDP is a powerful tool for training large models with fewer GPUs compared to other parallelism strategies. Refer to the following resources below to learn even more about FSDP.
- Follow along with the more in-depth Accelerate guide for [FSDP](https://hf.co/docs/accelerate/usage_guides/fsdp).
- Read the [Introducing PyTorch Fully Sharded Data Parallel (FSDP) API](https://pytorch.org/blog/introducing-pytorch-fully-sharded-data-parallel-api/) blog post.
- Read the [Scaling PyTorch models on Cloud TPUs with FSDP](https://pytorch.org/blog/scaling-pytorch-models-on-cloud-tpus-with-fsdp/) blog post.
| transformers/docs/source/en/fsdp.md/0 | {
"file_path": "transformers/docs/source/en/fsdp.md",
"repo_id": "transformers",
"token_count": 2178
} | 354 |
<!--Copyright 2021 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Feature Extractor
A feature extractor is in charge of preparing input features for audio or vision models. This includes feature extraction from sequences, e.g., pre-processing audio files to generate Log-Mel Spectrogram features, feature extraction from images, e.g., cropping image files, but also padding, normalization, and conversion to NumPy, PyTorch, and TensorFlow tensors.
## FeatureExtractionMixin
[[autodoc]] feature_extraction_utils.FeatureExtractionMixin
- from_pretrained
- save_pretrained
## SequenceFeatureExtractor
[[autodoc]] SequenceFeatureExtractor
- pad
## BatchFeature
[[autodoc]] BatchFeature
## ImageFeatureExtractionMixin
[[autodoc]] image_utils.ImageFeatureExtractionMixin
| transformers/docs/source/en/main_classes/feature_extractor.md/0 | {
"file_path": "transformers/docs/source/en/main_classes/feature_extractor.md",
"repo_id": "transformers",
"token_count": 383
} | 355 |
<!--Copyright 2025 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2024-11-21 and added to Hugging Face Transformers on 2025-07-08.*
# AIMv2
## Overview
The AIMv2 model was proposed in [Multimodal Autoregressive Pre-training of Large Vision Encoders](https://huggingface.co/papers/2411.14402) by Enrico Fini, Mustafa Shukor, Xiujun Li, Philipp Dufter, Michal Klein, David Haldimann, Sai Aitharaju, Victor Guilherme Turrisi da Costa, Louis Béthune, Zhe Gan, Alexander T Toshev, Marcin Eichner, Moin Nabi, Yinfei Yang, Joshua M. Susskind, Alaaeldin El-Nouby.
The abstract from the paper is the following:
*We introduce a novel method for pre-training of large-scale vision encoders. Building on recent advancements in autoregressive pre-training of vision models, we extend this framework to a multimodal setting, i.e., images and text. In this paper, we present AIMV2, a family of generalist vision encoders characterized by a straightforward pre-training process, scalability, and remarkable performance across a range of downstream tasks. This is achieved by pairing the vision encoder with a multimodal decoder that autoregressively generates raw image patches and text tokens. Our encoders excel not only in multimodal evaluations but also in vision benchmarks such as localization, grounding, and classification. Notably, our AIMV2-3B encoder achieves 89.5% accuracy on ImageNet-1k with a frozen trunk. Furthermore, AIMV2 consistently outperforms state-of-the-art contrastive models (e.g., CLIP, SigLIP) in multimodal image understanding across diverse settings.*
This model was contributed by [Yaswanth Gali](https://huggingface.co/yaswanthgali).
The original code can be found [here](https://github.com/apple/ml-aim).
## Usage Example
Here is an example of Image Feature Extraction using specific checkpoints on resized images and native resolution images:
```python
import requests
from PIL import Image
from transformers import AutoImageProcessor, AutoModel
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)
processor = AutoImageProcessor.from_pretrained("apple/aimv2-large-patch14-native")
model = AutoModel.from_pretrained("apple/aimv2-large-patch14-native")
inputs = processor(images=image, return_tensors="pt")
outputs = model(**inputs)
```
Here is an example of a checkpoint performing zero-shot classification:
```python
import requests
from PIL import Image
from transformers import AutoProcessor, AutoModel
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)
text = ["Picture of a dog.", "Picture of a cat.", "Picture of a horse."]
processor = AutoProcessor.from_pretrained("apple/aimv2-large-patch14-224-lit")
model = AutoModel.from_pretrained("apple/aimv2-large-patch14-224-lit")
inputs = processor(
images=image,
text=text,
add_special_tokens=True,
truncation=True,
padding=True,
return_tensors="pt",
)
outputs = model(**inputs)
probs = outputs.logits_per_image.softmax(dim=-1)
```
## Aimv2Config
[[autodoc]] Aimv2Config
## Aimv2TextConfig
[[autodoc]] Aimv2TextConfig
## Aimv2VisionConfig
[[autodoc]] Aimv2VisionConfig
## Aimv2Model
[[autodoc]] Aimv2Model
- forward
## Aimv2VisionModel
[[autodoc]] Aimv2VisionModel
- forward
## Aimv2TextModel
[[autodoc]] Aimv2TextModel
- forward
| transformers/docs/source/en/model_doc/aimv2.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/aimv2.md",
"repo_id": "transformers",
"token_count": 1234
} | 356 |
<!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
<div style="float: right;">
<div class="flex flex-wrap space-x-1">
<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
</div>
</div>
# BertGeneration
[BertGeneration](https://huggingface.co/papers/1907.12461) leverages pretrained BERT checkpoints for sequence-to-sequence tasks with the [`EncoderDecoderModel`] architecture. BertGeneration adapts the [`BERT`] for generative tasks.
You can find all the original BERT checkpoints under the [BERT](https://huggingface.co/collections/google/bert-release-64ff5e7a4be99045d1896dbc) collection.
> [!TIP]
> This model was contributed by [patrickvonplaten](https://huggingface.co/patrickvonplaten).
>
> Click on the BertGeneration models in the right sidebar for more examples of how to apply BertGeneration to different sequence generation tasks.
The example below demonstrates how to use BertGeneration with [`EncoderDecoderModel`] for sequence-to-sequence tasks.
<hfoptions id="usage">
<hfoption id="Pipeline">
```python
import torch
from transformers import pipeline
pipeline = pipeline(
task="text2text-generation",
model="google/roberta2roberta_L-24_discofuse",
dtype=torch.float16,
device=0
)
pipeline("Plants create energy through ")
```
</hfoption>
<hfoption id="AutoModel">
```python
import torch
from transformers import EncoderDecoderModel, AutoTokenizer
model = EncoderDecoderModel.from_pretrained("google/roberta2roberta_L-24_discofuse", dtype="auto")
tokenizer = AutoTokenizer.from_pretrained("google/roberta2roberta_L-24_discofuse")
input_ids = tokenizer(
"Plants create energy through ", add_special_tokens=False, return_tensors="pt"
).input_ids
outputs = model.generate(input_ids)
print(tokenizer.decode(outputs[0]))
```
</hfoption>
<hfoption id="transformers CLI">
```bash
echo -e "Plants create energy through " | transformers run --task text2text-generation --model "google/roberta2roberta_L-24_discofuse" --device 0
```
</hfoption>
</hfoptions>
Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
The example below uses [BitsAndBytesConfig](../quantizationbitsandbytes) to quantize the weights to 4-bit.
```python
import torch
from transformers import EncoderDecoderModel, AutoTokenizer, BitsAndBytesConfig
# Configure 4-bit quantization
quantization_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.float16
)
model = EncoderDecoderModel.from_pretrained(
"google/roberta2roberta_L-24_discofuse",
quantization_config=quantization_config,
dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("google/roberta2roberta_L-24_discofuse")
input_ids = tokenizer(
"Plants create energy through ", add_special_tokens=False, return_tensors="pt"
).input_ids
outputs = model.generate(input_ids)
print(tokenizer.decode(outputs[0]))
```
## Notes
- [`BertGenerationEncoder`] and [`BertGenerationDecoder`] should be used in combination with [`EncoderDecoderModel`] for sequence-to-sequence tasks.
```python
from transformers import BertGenerationEncoder, BertGenerationDecoder, BertTokenizer, EncoderDecoderModel
# leverage checkpoints for Bert2Bert model
# use BERT's cls token as BOS token and sep token as EOS token
encoder = BertGenerationEncoder.from_pretrained("google-bert/bert-large-uncased", bos_token_id=101, eos_token_id=102)
# add cross attention layers and use BERT's cls token as BOS token and sep token as EOS token
decoder = BertGenerationDecoder.from_pretrained(
"google-bert/bert-large-uncased", add_cross_attention=True, is_decoder=True, bos_token_id=101, eos_token_id=102
)
bert2bert = EncoderDecoderModel(encoder=encoder, decoder=decoder)
# create tokenizer
tokenizer = BertTokenizer.from_pretrained("google-bert/bert-large-uncased")
input_ids = tokenizer(
"This is a long article to summarize", add_special_tokens=False, return_tensors="pt"
).input_ids
labels = tokenizer("This is a short summary", return_tensors="pt").input_ids
# train
loss = bert2bert(input_ids=input_ids, decoder_input_ids=labels, labels=labels).loss
loss.backward()
```
- For summarization, sentence splitting, sentence fusion and translation, no special tokens are required for the input.
- No EOS token should be added to the end of the input for most generation tasks.
## BertGenerationConfig
[[autodoc]] BertGenerationConfig
## BertGenerationTokenizer
[[autodoc]] BertGenerationTokenizer
- save_vocabulary
## BertGenerationEncoder
[[autodoc]] BertGenerationEncoder
- forward
## BertGenerationDecoder
[[autodoc]] BertGenerationDecoder
- forward | transformers/docs/source/en/model_doc/bert-generation.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/bert-generation.md",
"repo_id": "transformers",
"token_count": 1815
} | 357 |
<!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2020-06-05 and added to Hugging Face Transformers on 2020-11-16.*
<div style="float: right;">
<div class="flex flex-wrap space-x-1">
<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
</div>
</div>
# DeBERTa
[DeBERTa](https://huggingface.co/papers/2006.03654) improves the pretraining efficiency of BERT and RoBERTa with two key ideas, disentangled attention and an enhanced mask decoder. Instead of mixing everything together like BERT, DeBERTa separates a word's *content* from its *position* and processes them independently. This gives it a clearer sense of what's being said and where in the sentence it's happening.
The enhanced mask decoder replaces the traditional softmax decoder to make better predictions.
Even with less training data than RoBERTa, DeBERTa manages to outperform it on several benchmarks.
You can find all the original DeBERTa checkpoints under the [Microsoft](https://huggingface.co/microsoft?search_models=deberta) organization.
> [!TIP]
> Click on the DeBERTa models in the right sidebar for more examples of how to apply DeBERTa to different language tasks.
The example below demonstrates how to classify text with [`Pipeline`], [`AutoModel`], and from the command line.
<hfoptions id="usage">
<hfoption id="Pipeline">
```py
import torch
from transformers import pipeline
classifier = pipeline(
task="text-classification",
model="microsoft/deberta-base-mnli",
device=0,
)
classifier({
"text": "A soccer game with multiple people playing.",
"text_pair": "Some people are playing a sport."
})
```
</hfoption>
<hfoption id="AutoModel">
```py
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer
model_name = "microsoft/deberta-base-mnli"
tokenizer = AutoTokenizer.from_pretrained("microsoft/deberta-base-mnli")
model = AutoModelForSequenceClassification.from_pretrained("microsoft/deberta-base-mnli", device_map="auto")
inputs = tokenizer(
"A soccer game with multiple people playing.",
"Some people are playing a sport.",
return_tensors="pt"
).to(model.device)
with torch.no_grad():
logits = model(**inputs).logits
predicted_class = logits.argmax().item()
labels = ["contradiction", "neutral", "entailment"]
print(f"The predicted relation is: {labels[predicted_class]}")
```
</hfoption>
<hfoption id="transformers CLI">
```bash
echo -e '{"text": "A soccer game with multiple people playing.", "text_pair": "Some people are playing a sport."}' | transformers run --task text-classification --model microsoft/deberta-base-mnli --device 0
```
</hfoption>
</hfoptions>
## Notes
- DeBERTa uses **relative position embeddings**, so it does not require **right-padding** like BERT.
- For best results, use DeBERTa on sentence-level or sentence-pair classification tasks like MNLI, RTE, or SST-2.
- If you're using DeBERTa for token-level tasks like masked language modeling, make sure to load a checkpoint specifically pretrained or fine-tuned for token-level tasks.
## DebertaConfig
[[autodoc]] DebertaConfig
## DebertaTokenizer
[[autodoc]] DebertaTokenizer
- build_inputs_with_special_tokens
- get_special_tokens_mask
- create_token_type_ids_from_sequences
- save_vocabulary
## DebertaTokenizerFast
[[autodoc]] DebertaTokenizerFast
- build_inputs_with_special_tokens
- create_token_type_ids_from_sequences
## DebertaModel
[[autodoc]] DebertaModel
- forward
## DebertaPreTrainedModel
[[autodoc]] DebertaPreTrainedModel
## DebertaForMaskedLM
[[autodoc]] DebertaForMaskedLM
- forward
## DebertaForSequenceClassification
[[autodoc]] DebertaForSequenceClassification
- forward
## DebertaForTokenClassification
[[autodoc]] DebertaForTokenClassification
- forward
## DebertaForQuestionAnswering
[[autodoc]] DebertaForQuestionAnswering
- forward
| transformers/docs/source/en/model_doc/deberta.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/deberta.md",
"repo_id": "transformers",
"token_count": 1456
} | 358 |
<!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2022-10-20 and added to Hugging Face Transformers on 2023-06-20.*
# FLAN-T5
<div class="flex flex-wrap space-x-1">
<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
</div>
## Overview
FLAN-T5 was released in the paper [Scaling Instruction-Finetuned Language Models](https://huggingface.co/papers/2210.11416) - it is an enhanced version of T5 that has been finetuned in a mixture of tasks.
One can directly use FLAN-T5 weights without finetuning the model:
```python
>>> from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
>>> model = AutoModelForSeq2SeqLM.from_pretrained("google/flan-t5-small")
>>> tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-small")
>>> inputs = tokenizer("A step by step recipe to make bolognese pasta:", return_tensors="pt")
>>> outputs = model.generate(**inputs)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
['Pour a cup of bolognese into a large bowl and add the pasta']
```
FLAN-T5 includes the same improvements as T5 version 1.1 (see [here](https://huggingface.co/docs/transformers/model_doc/t5v1.1) for the full details of the model's improvements.)
Google has released the following variants:
- [google/flan-t5-small](https://huggingface.co/google/flan-t5-small)
- [google/flan-t5-base](https://huggingface.co/google/flan-t5-base)
- [google/flan-t5-large](https://huggingface.co/google/flan-t5-large)
- [google/flan-t5-xl](https://huggingface.co/google/flan-t5-xl)
- [google/flan-t5-xxl](https://huggingface.co/google/flan-t5-xxl).
The original checkpoints can be found [here](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-t5-checkpoints).
<Tip>
Refer to [T5's documentation page](t5) for all API reference, code examples and notebooks. For more details regarding training and evaluation of the FLAN-T5, refer to the model card.
</Tip>
| transformers/docs/source/en/model_doc/flan-t5.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/flan-t5.md",
"repo_id": "transformers",
"token_count": 876
} | 359 |
<!--Copyright 2025 The GLM & ZhipuAI team and The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2024-06-18 and added to Hugging Face Transformers on 2025-04-09.*
# Glm4
## Overview
The GLM family welcomes new members [GLM-4-0414](https://huggingface.co/papers/2406.12793) series models.
The **GLM-4-32B-0414** series models, featuring 32 billion parameters. Its performance is comparable to OpenAI’s GPT
series and DeepSeek’s V3/R1 series. It also supports very user-friendly local deployment features. GLM-4-32B-Base-0414
was pre-trained on 15T of high-quality data, including substantial reasoning-type synthetic data. This lays the
foundation for subsequent reinforcement learning extensions. In the post-training stage, we employed human preference
alignment for dialogue scenarios. Additionally, using techniques like rejection sampling and reinforcement learning, we
enhanced the model’s performance in instruction following, engineering code, and function calling, thus strengthening
the atomic capabilities required for agent tasks. GLM-4-32B-0414 achieves good results in engineering code, Artifact
generation, function calling, search-based Q&A, and report generation. In particular, on several benchmarks, such as
code generation or specific Q&A tasks, GLM-4-32B-Base-0414 achieves comparable performance with those larger models like
GPT-4o and DeepSeek-V3-0324 (671B).
**GLM-Z1-32B-0414** is a reasoning model with deep thinking capabilities. This was developed based on GLM-4-32B-0414
through cold start, extended reinforcement learning, and further training on tasks including mathematics, code, and
logic. Compared to the base model, GLM-Z1-32B-0414 significantly improves mathematical abilities and the capability to
solve complex tasks. During training, we also introduced general reinforcement learning based on pairwise ranking
feedback, which enhances the model's general capabilities.
**GLM-Z1-Rumination-32B-0414** is a deep reasoning model with rumination capabilities (against OpenAI's Deep Research).
Unlike typical deep thinking models, the rumination model is capable of deeper and longer thinking to solve more
open-ended and complex problems (e.g., writing a comparative analysis of AI development in two cities and their future
development plans). Z1-Rumination is trained through scaling end-to-end reinforcement learning with responses graded by
the ground truth answers or rubrics and can make use of search tools during its deep thinking process to handle complex
tasks. The model shows significant improvements in research-style writing and complex tasks.
Finally, **GLM-Z1-9B-0414** is a surprise. We employed all the aforementioned techniques to train a small model (9B).
GLM-Z1-9B-0414 exhibits excellent capabilities in mathematical reasoning and general tasks. Its overall performance is
top-ranked among all open-source models of the same size. Especially in resource-constrained scenarios, this model
achieves an excellent balance between efficiency and effectiveness, providing a powerful option for users seeking
lightweight deployment.
## Glm4Config
[[autodoc]] Glm4Config
## Glm4Model
[[autodoc]] Glm4Model
- forward
## Glm4ForCausalLM
[[autodoc]] Glm4ForCausalLM
- forward
## Glm4ForSequenceClassification
[[autodoc]] Glm4ForSequenceClassification
- forward
## Glm4ForTokenClassification
[[autodoc]] Glm4ForTokenClassification
- forward
| transformers/docs/source/en/model_doc/glm4.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/glm4.md",
"repo_id": "transformers",
"token_count": 1041
} | 360 |
<!--Copyright 2025 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2025-04-16 and added to Hugging Face Transformers on 2025-04-11.*
# Granite Speech
<div class="flex flex-wrap space-x-1">
<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
</div>
## Overview
The [Granite Speech](https://huggingface.co/papers/2505.08699) model ([blog post](https://www.ibm.com/new/announcements/ibm-granite-3-3-speech-recognition-refined-reasoning-rag-loras)) is a multimodal language model, consisting of a speech encoder, speech projector, large language model, and LoRA adapter(s). More details regarding each component for the current (Granite 3.2 Speech) model architecture may be found below.
1. Speech Encoder: A [Conformer](https://huggingface.co/papers/2005.08100) encoder trained with Connectionist Temporal Classification (CTC) on character-level targets on ASR corpora. The encoder uses block-attention and self-conditioned CTC from the middle layer.
2. Speech Projector: A query transformer (q-former) operating on the outputs of the last encoder block. The encoder and projector temporally downsample the audio features to be merged into the multimodal embeddings to be processed by the llm.
3. Large Language Model: The Granite Speech model leverages Granite LLMs, which were originally proposed in [this paper](https://huggingface.co/papers/2408.13359).
4. LoRA adapter(s): The Granite Speech model contains a modality specific LoRA, which will be enabled when audio features are provided, and disabled otherwise.
Note that most of the aforementioned components are implemented generically to enable compatibility and potential integration with other model architectures in transformers.
This model was contributed by [Alexander Brooks](https://huggingface.co/abrooks9944), [Avihu Dekel](https://huggingface.co/Avihu), and [George Saon](https://huggingface.co/gsaon).
## Usage tips
- This model bundles its own LoRA adapter, which will be automatically loaded and enabled/disabled as needed during inference calls. Be sure to install [PEFT](https://github.com/huggingface/peft) to ensure the LoRA is correctly applied!
<!-- TODO (@alex-jw-brooks) Add an example here once the model compatible with the transformers implementation is released -->
## GraniteSpeechConfig
[[autodoc]] GraniteSpeechConfig
## GraniteSpeechEncoderConfig
[[autodoc]] GraniteSpeechEncoderConfig
## GraniteSpeechProcessor
[[autodoc]] GraniteSpeechProcessor
## GraniteSpeechFeatureExtractor
[[autodoc]] GraniteSpeechFeatureExtractor
## GraniteSpeechForConditionalGeneration
[[autodoc]] GraniteSpeechForConditionalGeneration
- forward
| transformers/docs/source/en/model_doc/granite_speech.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/granite_speech.md",
"repo_id": "transformers",
"token_count": 934
} | 361 |
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2023-06-21 and added to Hugging Face Transformers on 2023-08-18.*
# IDEFICS
<div class="flex flex-wrap space-x-1">
<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
</div>
## Overview
The IDEFICS model was proposed in [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents
](https://huggingface.co/papers/2306.16527
) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh
The abstract from the paper is the following:
*Large multimodal models trained on natural documents, which interleave images and text, outperform models trained on image-text pairs on various multimodal benchmarks that require reasoning over one or multiple images to generate a text. However, the datasets used to train these models have not been released, and the collection process has not been fully specified. We introduce the OBELICS dataset, an open web-scale filtered dataset of interleaved image-text documents comprising 141 million web pages extracted from Common Crawl, 353 million associated images, and 115 billion text tokens. We describe the dataset creation process, present comprehensive filtering rules, and provide an analysis of the dataset's content. To show the viability of OBELISC, we train an 80 billion parameters vision and language model on the dataset and obtain competitive performance on various multimodal benchmarks. We release the code to reproduce the dataset along with the dataset itself.*
This model was contributed by [HuggingFaceM4](https://huggingface.co/HuggingFaceM4). The original code can be found [here](<INSERT LINK TO GITHUB REPO HERE>). (TODO: don't have a public link yet).
<Tip warning={true}>
IDEFICS modeling code in Transformers is for finetuning and inferencing the pre-trained IDEFICS models.
To train a new IDEFICS model from scratch use the m4 codebase (a link will be provided once it's made public)
</Tip>
## IdeficsConfig
[[autodoc]] IdeficsConfig
## IdeficsModel
[[autodoc]] IdeficsModel
- forward
## IdeficsForVisionText2Text
[[autodoc]] IdeficsForVisionText2Text
- forward
## IdeficsImageProcessor
[[autodoc]] IdeficsImageProcessor
- preprocess
## IdeficsProcessor
[[autodoc]] IdeficsProcessor
- __call__
| transformers/docs/source/en/model_doc/idefics.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/idefics.md",
"repo_id": "transformers",
"token_count": 918
} | 362 |
<!--Copyright 2025 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2019-12-31 and added to Hugging Face Transformers on 2020-11-16.*
<div style="float: right;">
<div class="flex flex-wrap space-x-1">
<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
</div>
</div>
# LayoutLM
[LayoutLM](https://huggingface.co/papers/1912.13318) jointly learns text and the document layout rather than focusing only on text. It incorporates positional layout information and visual features of words from the document images.
You can find all the original LayoutLM checkpoints under the [LayoutLM](https://huggingface.co/collections/microsoft/layoutlm-6564539601de72cb631d0902) collection.
> [!TIP]
> Click on the LayoutLM models in the right sidebar for more examples of how to apply LayoutLM to different vision and language tasks.
The example below demonstrates question answering with the [`AutoModel`] class.
<hfoptions id="usage">
<hfoption id="AutoModel">
```py
import torch
from datasets import load_dataset
from transformers import AutoTokenizer, LayoutLMForQuestionAnswering
tokenizer = AutoTokenizer.from_pretrained("impira/layoutlm-document-qa", add_prefix_space=True)
model = LayoutLMForQuestionAnswering.from_pretrained("impira/layoutlm-document-qa", dtype=torch.float16)
dataset = load_dataset("nielsr/funsd", split="train")
example = dataset[0]
question = "what's his name?"
words = example["words"]
boxes = example["bboxes"]
encoding = tokenizer(
question.split(),
words,
is_split_into_words=True,
return_token_type_ids=True,
return_tensors="pt"
)
bbox = []
for i, s, w in zip(encoding.input_ids[0], encoding.sequence_ids(0), encoding.word_ids(0)):
if s == 1:
bbox.append(boxes[w])
elif i == tokenizer.sep_token_id:
bbox.append([1000] * 4)
else:
bbox.append([0] * 4)
encoding["bbox"] = torch.tensor([bbox])
word_ids = encoding.word_ids(0)
outputs = model(**encoding)
loss = outputs.loss
start_scores = outputs.start_logits
end_scores = outputs.end_logits
start, end = word_ids[start_scores.argmax(-1)], word_ids[end_scores.argmax(-1)]
print(" ".join(words[start : end + 1]))
```
</hfoption>
</hfoptions>
## Notes
- The original LayoutLM was not designed with a unified processing workflow. Instead, it expects preprocessed text (`words`) and bounding boxes (`boxes`) from an external OCR engine (like [Pytesseract](https://pypi.org/project/pytesseract/)) and provide them as additional inputs to the tokenizer.
- The [`~LayoutLMModel.forward`] method expects the input `bbox` (bounding boxes of the input tokens). Each bounding box should be in the format `(x0, y0, x1, y1)`. `(x0, y0)` corresponds to the upper left corner of the bounding box and `{x1, y1)` corresponds to the lower right corner. The bounding boxes need to be normalized on a 0-1000 scale as shown below.
```python
def normalize_bbox(bbox, width, height):
return [
int(1000 * (bbox[0] / width)),
int(1000 * (bbox[1] / height)),
int(1000 * (bbox[2] / width)),
int(1000 * (bbox[3] / height)),
]
```
- `width` and `height` correspond to the width and height of the original document in which the token occurs. These values can be obtained as shown below.
```python
from PIL import Image
# Document can be a png, jpg, etc. PDFs must be converted to images.
image = Image.open(name_of_your_document).convert("RGB")
width, height = image.size
```
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with LayoutLM. If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
- Read [fine-tuning LayoutLM for document-understanding using Keras & Hugging Face Transformers](https://www.philschmid.de/fine-tuning-layoutlm-keras) to learn more.
- Read [fine-tune LayoutLM for document-understanding using only Hugging Face Transformers](https://www.philschmid.de/fine-tuning-layoutlm) for more information.
- Refer to this [notebook](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/LayoutLM/Add_image_embeddings_to_LayoutLM.ipynb) for a practical example of how to fine-tune LayoutLM.
- Refer to this [notebook](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/LayoutLM/Fine_tuning_LayoutLMForSequenceClassification_on_RVL_CDIP.ipynb) for an example of how to fine-tune LayoutLM for sequence classification.
- Refer to this [notebook](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/LayoutLM/Fine_tuning_LayoutLMForTokenClassification_on_FUNSD.ipynb) for an example of how to fine-tune LayoutLM for token classification.
- Read [Deploy LayoutLM with Hugging Face Inference Endpoints](https://www.philschmid.de/inference-endpoints-layoutlm) to learn how to deploy LayoutLM.
## LayoutLMConfig
[[autodoc]] LayoutLMConfig
## LayoutLMTokenizer
[[autodoc]] LayoutLMTokenizer
- __call__
## LayoutLMTokenizerFast
[[autodoc]] LayoutLMTokenizerFast
- __call__
## LayoutLMModel
[[autodoc]] LayoutLMModel
## LayoutLMForMaskedLM
[[autodoc]] LayoutLMForMaskedLM
## LayoutLMForSequenceClassification
[[autodoc]] LayoutLMForSequenceClassification
## LayoutLMForTokenClassification
[[autodoc]] LayoutLMForTokenClassification
## LayoutLMForQuestionAnswering
[[autodoc]] LayoutLMForQuestionAnswering
| transformers/docs/source/en/model_doc/layoutlm.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/layoutlm.md",
"repo_id": "transformers",
"token_count": 2013
} | 363 |
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2024-08-06 and added to Hugging Face Transformers on 2024-09-05.*
# LLaVA-OneVision
<div class="flex flex-wrap space-x-1">
<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
<img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
</div>
## Overview
The LLaVA-OneVision model was proposed in [LLaVA-OneVision: Easy Visual Task Transfer](https://huggingface.co/papers/2408.03326) by <Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, Chunyuan Li
LLaVA-OneVision is a Vision-Language Model that can generate text conditioned on one or several images/videos. The model consists of SigLIP vision encoder and a Qwen2 language backbone. The images are processed with anyres-9 technique where the image is split into 9 patches to better process high resolution images and capture as much details as possible. However, videos are pooled to a total sequence length of 196 tokens each frame for more memory efficient computation. LLaVA-OneVision is available in three sizes: 0.5B, 7B and 72B and achieves remarkable performance on benchmark evaluations.
The abstract from the paper is the following:
*We present LLaVA-OneVision, a family of open large multimodal models (LMMs)
developed by consolidating our insights into data, models, and visual representations in the LLaVA-NeXT blog series. Our experimental results demonstrate that
LLaVA-OneVision is the first single model that can simultaneously push the performance boundaries of open LMMs in three important computer vision scenarios:
single-image, multi-image, and video scenarios. Importantly, the design of LLaVAOneVision allows strong transfer learning across different modalities/scenarios,
yielding new emerging capabilities. In particular, strong video understanding and
cross-scenario capabilities are demonstrated through task transfer from images to
videos.*
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/llava-ov-architecture.png"
alt="drawing" width="600"/>
<small> LLaVA-OneVision architecture. Taken from the <a href="https://huggingface.co/papers/2408.03326">original paper.</a> </small>
Tips:
- We advise users to use `padding_side="left"` when computing batched generation as it leads to more accurate results. Simply make sure to call `processor.tokenizer.padding_side = "left"` before generating.
<Tip warning={true}>
- Llava-OneVision uses different number of patches for images and thus has to pad the inputs inside modeling code, aside from the padding done when processing the inputs. The default setting is "left-padding" if model is in `eval()` mode, otherwise "right-padding".
</Tip>
### Formatting Prompts with Chat Templates
Each **checkpoint** is trained with a specific prompt format, depending on the underlying large language model backbone. To ensure correct formatting, use the processor’s `apply_chat_template` method.
**Important:**
- You must construct a conversation history — passing a plain string won't work.
- Each message should be a dictionary with `"role"` and `"content"` keys.
- The `"content"` should be a list of dictionaries for different modalities like `"text"` and `"image"`.
Here’s an example of how to structure your input.
We will use [llava-onevision-qwen2-7b-si-hf](https://huggingface.co/llava-hf/llava-onevision-qwen2-7b-si-hf) and a conversation history of text and image. Each content field has to be a list of dicts, as follows:
```python
from transformers import AutoProcessor
processor = AutoProcessor.from_pretrained("llava-hf/llava-onevision-qwen2-7b-si-hf")
conversation = [
{
"role": "user",
"content": [
{"type": "image"},
{"type": "text", "text": "What’s shown in this image?"},
],
},
{
"role": "assistant",
"content": [{"type": "text", "text": "This image shows a red stop sign."},]
},
{
"role": "user",
"content": [
{"type": "text", "text": "Describe the image in more details."},
],
},
]
text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
# Note that the template simply formats your prompt, you still have to tokenize it and obtain pixel values for your images
print(text_prompt)
'<|im_start|>user\n<image>What is shown in this image?<|im_end|>\n<|im_start|>assistant\nPage showing the list of options.<|im_end|>'
```
🚀 **Bonus:** If you're using `transformers>=4.49.0`, you can also get a vectorized output from `apply_chat_template`. See the **Usage Examples** below for more details on how to use it.
This model was contributed by [RaushanTurganbay](https://huggingface.co/RaushanTurganbay).
The original code can be found [here](https://github.com/LLaVA-VL/LLaVA-NeXT/tree/main).
## Usage example
### Single image inference
Here's how to load the model and perform inference in half-precision (`torch.float16`):
```python
from transformers import AutoProcessor, LlavaOnevisionForConditionalGeneration, infer_device
import torch
device = f"{infer_device}:0"
processor = AutoProcessor.from_pretrained("llava-hf/llava-onevision-qwen2-7b-ov-hf")
model = LlavaOnevisionForConditionalGeneration.from_pretrained(
"llava-hf/llava-onevision-qwen2-7b-ov-hf",
dtype=torch.float16,
device_map=device
)
# prepare image and text prompt, using the appropriate prompt template
url = "https://github.com/haotian-liu/LLaVA/blob/1a91fc274d7c35a9b50b3cb29c4247ae5837ce39/images/llava_v1_5_radar.jpg?raw=true"
conversation = [
{
"role": "user",
"content": [
{"type": "image", "url": url},
{"type": "text", "text": "What is shown in this image?"},
],
},
]
inputs = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt")
inputs = inputs.to(model.device, torch.float16)
# autoregressively complete prompt
output = model.generate(**inputs, max_new_tokens=100)
print(processor.decode(output[0], skip_special_tokens=True))
'user\n\nWhat is shown in this image?\nassistant\nThe image shows a radar chart, also known as a spider chart or a star chart, which is used to compare multiple quantitative variables. Each axis represents a different variable, and the chart is filled with'
```
### Multi image inference
LLaVa-OneVision can perform inference with multiple images as input, where images either belong to the same prompt or different prompts (in batched inference). For that you have to use checkpoints with an "ov" suffix. For multi-image cases, we recommend using a **nested list of images** as input. Otherwise, every image will be patchified and consume a lot of memory. Here is how you can do it:
```python
import requests
from PIL import Image
import torch
from transformers import AutoProcessor, LlavaOnevisionForConditionalGeneration
# Load the model in half-precision
model = LlavaOnevisionForConditionalGeneration.from_pretrained("llava-hf/llava-onevision-qwen2-7b-ov-hf", dtype=torch.float16, device_map="auto")
processor = AutoProcessor.from_pretrained("llava-hf/llava-onevision-qwen2-7b-ov-hf")
# Prepare a batch of two prompts, where the first one is a multi-turn conversation and the second is not
conversation_1 = [
{
"role": "user",
"content": [
{"type": "image", "url": "https://www.ilankelman.org/stopsigns/australia.jpg"},
{"type": "text", "text": "What is shown in this image?"},
],
},
{
"role": "assistant",
"content": [
{"type": "text", "text": "There is a red stop sign in the image."},
],
},
{
"role": "user",
"content": [
{"type": "image", "url": "http://images.cocodataset.org/val2017/000000039769.jpg"},
{"type": "text", "text": "What about this image? How many cats do you see?"},
],
},
]
conversation_2 = [
{
"role": "user",
"content": [
{"type": "image", "url": "https://huggingface.co/microsoft/kosmos-2-patch14-224/resolve/main/snowman.jpg"},
{"type": "text", "text": "What is shown in this image?"},
],
},
]
inputs = processor.apply_chat_template(
[conversation_1, conversation_2],
add_generation_prompt=True,
tokenize=True,
return_dict=True,
padding=True,
padding_side="left",
return_tensors="pt",
).to(model.device, torch.float16)
# Generate
generate_ids = model.generate(**inputs, max_new_tokens=30)
processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
['user\n\nWhat is shown in this image?\nassistant\nThere is a red stop sign in the image.\nuser\n\nWhat about this image? How many cats do you see?\nassistant\ntwo', 'user\n\nWhat is shown in this image?\nassistant\nThe image shows a whimsical scene of a snowman sitting by a campfire. The snowman is anthropomorphized, wearing a hat and']
```
### Video inference
LLaVa-OneVision also can perform inference with videos as input, where video frames are treated as multiple images. Here is how you can do it:
```python
from huggingface_hub import hf_hub_download
import torch
from transformers import AutoProcessor, LlavaOnevisionForConditionalGeneration
# Load the model in half-precision
model = LlavaOnevisionForConditionalGeneration.from_pretrained("llava-hf/llava-onevision-qwen2-7b-ov-hf", dtype=torch.float16, device_map="auto")
processor = AutoProcessor.from_pretrained("llava-hf/llava-onevision-qwen2-7b-ov-hf")
video_path = hf_hub_download(repo_id="raushan-testing-hf/videos-test", filename="sample_demo_1.mp4", repo_type="dataset")
conversation = [
{
"role": "user",
"content": [
{"type": "video", "path": video_path},
{"type": "text", "text": "Why is this video funny?"},
],
},
]
inputs = processor.apply_chat_template(
conversation,
num_frames=8
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt"
).to(model.device, torch.float16)
out = model.generate(**inputs, max_new_tokens=60)
processor.batch_decode(out, skip_special_tokens=True, clean_up_tokenization_spaces=True)
["user\n\nWhy is this video funny?\nassistant\nThe video appears to be humorous because it shows a young child, who is wearing glasses and holding a book, seemingly reading with a serious and focused expression. The child's glasses are a bit oversized for their face, which adds a comical touch, as it's a common trope to see children wearing"]
```
## Model optimization
### Quantization using bitsandbytes
The model can be loaded in 8 or 4 bits, greatly reducing the memory requirements while maintaining the performance of the original model. First make sure to install bitsandbytes, `pip install bitsandbytes` and make sure to have access to a GPU/accelerator that is supported by the library.
<Tip>
bitsandbytes is being refactored to support multiple backends beyond CUDA. Currently, ROCm (AMD GPU) and Intel CPU implementations are mature, with Intel XPU in progress and Apple Silicon support expected by Q4/Q1. For installation instructions and the latest backend updates, visit [this link](https://huggingface.co/docs/bitsandbytes/main/en/installation#multi-backend).
We value your feedback to help identify bugs before the full release! Check out [these docs](https://huggingface.co/docs/bitsandbytes/main/en/non_cuda_backends) for more details and feedback links.
</Tip>
Simply change the snippet above with:
```python
from transformers import LlavaOnevisionForConditionalGeneration, BitsAndBytesConfig
# specify how to quantize the model
quantization_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.float16,
)
model = LlavaOnevisionForConditionalGeneration.from_pretrained(model_id, quantization_config=quantization_config, device_map="auto")
```
### Use Flash-Attention 2 to further speed-up generation
First make sure to install flash-attn. Refer to the [original repository of Flash Attention](https://github.com/Dao-AILab/flash-attention) regarding that package installation. Simply change the snippet above with:
```python
from transformers import LlavaOnevisionForConditionalGeneration
model = LlavaOnevisionForConditionalGeneration.from_pretrained(
model_id,
dtype=torch.float16,
use_flash_attention_2=True
).to(0)
```
## LlavaOnevisionConfig
[[autodoc]] LlavaOnevisionConfig
## LlavaOnevisionProcessor
[[autodoc]] LlavaOnevisionProcessor
## LlavaOnevisionImageProcessor
[[autodoc]] LlavaOnevisionImageProcessor
- preprocess
## LlavaOnevisionImageProcessorFast
[[autodoc]] LlavaOnevisionImageProcessorFast
- preprocess
## LlavaOnevisionVideoProcessor
[[autodoc]] LlavaOnevisionVideoProcessor
## LlavaOnevisionModel
[[autodoc]] LlavaOnevisionModel
## LlavaOnevisionForConditionalGeneration
[[autodoc]] LlavaOnevisionForConditionalGeneration
- forward
| transformers/docs/source/en/model_doc/llava_onevision.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/llava_onevision.md",
"repo_id": "transformers",
"token_count": 4629
} | 364 |
<!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2017-04-17 and added to Hugging Face Transformers on 2022-11-21.*
<div style="float: right;">
<div class="flex flex-wrap space-x-1">
<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-EE4C2C?style=flat&logo=pytorch&logoColor=white">
</div>
</div>
# MobileNet V1
[MobileNet V1](https://huggingface.co/papers/1704.04861) is a family of efficient convolutional neural networks optimized for on-device or embedded vision tasks. It achieves this efficiency by using depth-wise separable convolutions instead of standard convolutions. The architecture allows for easy trade-offs between latency and accuracy using two main hyperparameters, a width multiplier (alpha) and an image resolution multiplier.
You can all the original MobileNet checkpoints under the [Google](https://huggingface.co/google?search_models=mobilenet) organization.
> [!TIP]
> Click on the MobileNet V1 models in the right sidebar for more examples of how to apply MobileNet to different vision tasks.
The example below demonstrates how to classify an image with [`Pipeline`] or the [`AutoModel`] class.
<hfoptions id="usage">
<hfoption id="Pipeline">
```python
import torch
from transformers import pipeline
pipeline = pipeline(
task="image-classification",
model="google/mobilenet_v1_1.0_224",
dtype=torch.float16,
device=0
)
pipeline("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg")
```
</hfoption>
<hfoption id="AutoModel">
```python
import torch
import requests
from PIL import Image
from transformers import AutoModelForImageClassification, AutoImageProcessor
image_processor = AutoImageProcessor.from_pretrained(
"google/mobilenet_v1_1.0_224",
)
model = AutoModelForImageClassification.from_pretrained(
"google/mobilenet_v1_1.0_224",
)
url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
image = Image.open(requests.get(url, stream=True).raw)
inputs = image_processor(image, return_tensors="pt")
with torch.no_grad():
logits = model(**inputs).logits
predicted_class_id = logits.argmax(dim=-1).item()
class_labels = model.config.id2label
predicted_class_label = class_labels[predicted_class_id]
print(f"The predicted class label is: {predicted_class_label}")
```
</hfoption>
</hfoptions>
<!-- Quantization - Not applicable -->
<!-- Attention Visualization - Not applicable for this model type -->
## Notes
- Checkpoint names follow the pattern `mobilenet_v1_{depth_multiplier}_{resolution}`, like `mobilenet_v1_1.0_224`. `1.0` is the depth multiplier and `224` is the image resolution.
- While trained on images of a specific sizes, the model architecture works with images of different sizes (minimum 32x32). The [`MobileNetV1ImageProcessor`] handles the necessary preprocessing.
- MobileNet is pretrained on [ImageNet-1k](https://huggingface.co/datasets/imagenet-1k), a dataset with 1000 classes. However, the model actually predicts 1001 classes. The additional class is an extra "background" class (index 0).
- The original TensorFlow checkpoints determines the padding amount at inference because it depends on the input image size. To use the native PyTorch padding behavior, set `tf_padding=False` in [`MobileNetV1Config`].
```python
from transformers import MobileNetV1Config
config = MobileNetV1Config.from_pretrained("google/mobilenet_v1_1.0_224", tf_padding=True)
```
- The Transformers implementation does not support the following features.
- Uses global average pooling instead of the optional 7x7 average pooling with stride 2. For larger inputs, this gives a pooled output that is larger than a 1x1 pixel.
- Does not support other `output_stride` values (fixed at 32). For smaller `output_strides`, the original implementation uses dilated convolution to prevent spatial resolution from being reduced further. (which would require dilated convolutions).
- `output_hidden_states=True` returns *all* intermediate hidden states. It is not possible to extract the output from specific layers for other downstream purposes.
- Does not include the quantized models from the original checkpoints because they include "FakeQuantization" operations to unquantize the weights.
## MobileNetV1Config
[[autodoc]] MobileNetV1Config
## MobileNetV1FeatureExtractor
[[autodoc]] MobileNetV1FeatureExtractor
- preprocess
## MobileNetV1ImageProcessor
[[autodoc]] MobileNetV1ImageProcessor
- preprocess
## MobileNetV1ImageProcessorFast
[[autodoc]] MobileNetV1ImageProcessorFast
- preprocess
## MobileNetV1Model
[[autodoc]] MobileNetV1Model
- forward
## MobileNetV1ForImageClassification
[[autodoc]] MobileNetV1ForImageClassification
- forward
| transformers/docs/source/en/model_doc/mobilenet_v1.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/mobilenet_v1.md",
"repo_id": "transformers",
"token_count": 1664
} | 365 |
<!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2022-04-14 and added to Hugging Face Transformers on 2023-06-20.*
# Neighborhood Attention Transformer
<div class="flex flex-wrap space-x-1">
<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
</div>
<Tip warning={true}>
This model is in maintenance mode only, we don't accept any new PRs changing its code.
If you run into any issues running this model, please reinstall the last version that supported this model: v4.40.2.
You can do so by running the following command: `pip install -U transformers==4.40.2`.
</Tip>
## Overview
NAT was proposed in [Neighborhood Attention Transformer](https://huggingface.co/papers/2204.07143)
by Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi.
It is a hierarchical vision transformer based on Neighborhood Attention, a sliding-window self attention pattern.
The abstract from the paper is the following:
*We present Neighborhood Attention (NA), the first efficient and scalable sliding-window attention mechanism for vision.
NA is a pixel-wise operation, localizing self attention (SA) to the nearest neighboring pixels, and therefore enjoys a
linear time and space complexity compared to the quadratic complexity of SA. The sliding-window pattern allows NA's
receptive field to grow without needing extra pixel shifts, and preserves translational equivariance, unlike
Swin Transformer's Window Self Attention (WSA). We develop NATTEN (Neighborhood Attention Extension), a Python package
with efficient C++ and CUDA kernels, which allows NA to run up to 40% faster than Swin's WSA while using up to 25% less
memory. We further present Neighborhood Attention Transformer (NAT), a new hierarchical transformer design based on NA
that boosts image classification and downstream vision performance. Experimental results on NAT are competitive;
NAT-Tiny reaches 83.2% top-1 accuracy on ImageNet, 51.4% mAP on MS-COCO and 48.4% mIoU on ADE20K, which is 1.9%
ImageNet accuracy, 1.0% COCO mAP, and 2.6% ADE20K mIoU improvement over a Swin model with similar size. *
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/neighborhood-attention-pattern.jpg"
alt="drawing" width="600"/>
<small> Neighborhood Attention compared to other attention patterns.
Taken from the <a href="https://huggingface.co/papers/2204.07143">original paper</a>.</small>
This model was contributed by [Ali Hassani](https://huggingface.co/alihassanijr).
The original code can be found [here](https://github.com/SHI-Labs/Neighborhood-Attention-Transformer).
## Usage tips
- One can use the [`AutoImageProcessor`] API to prepare images for the model.
- NAT can be used as a *backbone*. When `output_hidden_states = True`,
it will output both `hidden_states` and `reshaped_hidden_states`.
The `reshaped_hidden_states` have a shape of `(batch, num_channels, height, width)` rather than
`(batch_size, height, width, num_channels)`.
Notes:
- NAT depends on [NATTEN](https://github.com/SHI-Labs/NATTEN/)'s implementation of Neighborhood Attention.
You can install it with pre-built wheels for Linux by referring to [shi-labs.com/natten](https://shi-labs.com/natten),
or build on your system by running `pip install natten`.
Note that the latter will likely take time to compile. NATTEN does not support Windows devices yet.
- Patch size of 4 is only supported at the moment.
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with NAT.
<PipelineTag pipeline="image-classification"/>
- [`NatForImageClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/image_classification.ipynb).
- See also: [Image classification task guide](../tasks/image_classification)
If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
## NatConfig
[[autodoc]] NatConfig
## NatModel
[[autodoc]] NatModel
- forward
## NatForImageClassification
[[autodoc]] NatForImageClassification
- forward
| transformers/docs/source/en/model_doc/nat.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/nat.md",
"repo_id": "transformers",
"token_count": 1417
} | 366 |
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2024-09-17 and added to Hugging Face Transformers on 2024-09-14.*
# Pixtral
<div class="flex flex-wrap space-x-1">
<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
</div>
## Overview
The [Pixtral](https://huggingface.co/papers/2410.07073) model was released by the Mistral AI team in a [blog post](https://mistral.ai/news/pixtral-12b/). Pixtral is a multimodal version of [Mistral](mistral), incorporating a 400 million parameter vision encoder trained from scratch.
The intro from the blog says the following:
*Pixtral is trained to understand both natural images and documents, achieving 52.5% on the MMMU reasoning benchmark, surpassing a number of larger models. The model shows strong abilities in tasks such as chart and figure understanding, document question answering, multimodal reasoning and instruction following. Pixtral is able to ingest images at their natural resolution and aspect ratio, giving the user flexibility on the number of tokens used to process an image. Pixtral is also able to process any number of images in its long context window of 128K tokens. Unlike previous open-source models, Pixtral does not compromise on text benchmark performance to excel in multimodal tasks.*
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/pixtral_architecture.webp"
alt="drawing" width="600"/>
<small> Pixtral architecture. Taken from the <a href="https://mistral.ai/news/pixtral-12b/">blog post.</a> </small>
Tips:
- Pixtral is a multimodal model, taking images and text as input, and producing text as output.
- This model follows the [Llava](llava) architecture. The model uses [`PixtralVisionModel`] for its vision encoder, and [`MistralForCausalLM`] for its language decoder.
- The main contribution is the 2d ROPE (rotary position embeddings) on the images, and support for arbitrary image sizes (the images are not padded together nor are they resized).
- Similar to [Llava](llava), the model internally replaces the `[IMG]` token placeholders by image embeddings from the vision encoder. The format for one or multiple prompts is the following:
```
"<s>[INST][IMG]\nWhat are the things I should be cautious about when I visit this place?[/INST]"
```
Then, the processor will replace each `[IMG]` token with a number of `[IMG]` tokens that depend on the height and the width of each image. Each *row* of the image is separated by an `[IMG_BREAK]` token, and each image is separated by an `[IMG_END]` token. It's advised to use the `apply_chat_template` method of the processor, which takes care of all of this and formats the text for you. If you're using `transformers>=4.49.0`, you can also get a vectorized output from `apply_chat_template`. See the [usage section](#usage) for more info.
This model was contributed by [amyeroberts](https://huggingface.co/amyeroberts) and [ArthurZ](https://huggingface.co/ArthurZ). The original code can be found [here](https://github.com/vllm-project/vllm/pull/8377).
## Usage
At inference time, it's advised to use the processor's `apply_chat_template` method, which correctly formats the prompt for the model:
```python
from transformers import AutoProcessor, LlavaForConditionalGeneration
model_id = "mistral-community/pixtral-12b"
processor = AutoProcessor.from_pretrained(model_id)
model = LlavaForConditionalGeneration.from_pretrained(model_id, device_map="auto")
chat = [
{
"role": "user", "content": [
{"type": "text", "content": "Can this animal"},
{"type": "image", "url": "https://picsum.photos/id/237/200/300"},
{"type": "text", "content": "live here?"},
{"type": "image", "url": "https://picsum.photos/seed/picsum/200/300"}
]
}
]
inputs = processor.apply_chat_template(
chat,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt"
).to(model.device)
generate_ids = model.generate(**inputs, max_new_tokens=500)
output = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
```
## PixtralVisionConfig
[[autodoc]] PixtralVisionConfig
## MistralCommonTokenizer
[[autodoc]] MistralCommonTokenizer
## PixtralVisionModel
[[autodoc]] PixtralVisionModel
- forward
## PixtralImageProcessor
[[autodoc]] PixtralImageProcessor
- preprocess
## PixtralImageProcessorFast
[[autodoc]] PixtralImageProcessorFast
- preprocess
## PixtralProcessor
[[autodoc]] PixtralProcessor
| transformers/docs/source/en/model_doc/pixtral.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/pixtral.md",
"repo_id": "transformers",
"token_count": 1624
} | 367 |
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2023-04-05 and added to Hugging Face Transformers on 2023-04-19.*
# SAM
<div class="flex flex-wrap space-x-1">
<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
</div>
## Overview
SAM (Segment Anything Model) was proposed in [Segment Anything](https://huggingface.co/papers/2304.02643v1.pdf) by Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alex Berg, Wan-Yen Lo, Piotr Dollar, Ross Girshick.
The model can be used to predict segmentation masks of any object of interest given an input image.
The abstract from the paper is the following:
*We introduce the Segment Anything (SA) project: a new task, model, and dataset for image segmentation. Using our efficient model in a data collection loop, we built the largest segmentation dataset to date (by far), with over 1 billion masks on 11M licensed and privacy respecting images. The model is designed and trained to be promptable, so it can transfer zero-shot to new image distributions and tasks. We evaluate its capabilities on numerous tasks and find that its zero-shot performance is impressive -- often competitive with or even superior to prior fully supervised results. We are releasing the Segment Anything Model (SAM) and corresponding dataset (SA-1B) of 1B masks and 11M images at [https://segment-anything.com](https://segment-anything.com) to foster research into foundation models for computer vision.*
Tips:
- The model predicts binary masks that states the presence or not of the object of interest given an image.
- The model predicts much better results if input 2D points and/or input bounding boxes are provided
- You can prompt multiple points for the same image, and predict a single mask.
- Fine-tuning the model is not supported yet
- According to the paper, textual input should be also supported. However, at this time of writing this seems not to be supported according to [the official repository](https://github.com/facebookresearch/segment-anything/issues/4#issuecomment-1497626844).
This model was contributed by [ybelkada](https://huggingface.co/ybelkada) and [ArthurZ](https://huggingface.co/ArthurZ).
The original code can be found [here](https://github.com/facebookresearch/segment-anything).
Below is an example on how to run mask generation given an image and a 2D point:
```python
import torch
from PIL import Image
import requests
from transformers import SamModel, SamProcessor, infer_device
device = infer_device()
model = SamModel.from_pretrained("facebook/sam-vit-huge").to(device)
processor = SamProcessor.from_pretrained("facebook/sam-vit-huge")
img_url = "https://huggingface.co/ybelkada/segment-anything/resolve/main/assets/car.png"
raw_image = Image.open(requests.get(img_url, stream=True).raw).convert("RGB")
input_points = [[[450, 600]]] # 2D location of a window in the image
inputs = processor(raw_image, input_points=input_points, return_tensors="pt").to(device)
with torch.no_grad():
outputs = model(**inputs)
masks = processor.image_processor.post_process_masks(
outputs.pred_masks.cpu(), inputs["original_sizes"].cpu(), inputs["reshaped_input_sizes"].cpu()
)
scores = outputs.iou_scores
```
You can also process your own masks alongside the input images in the processor to be passed to the model.
```python
import torch
from PIL import Image
import requests
from transformers import SamModel, SamProcessor, infer_device
device = infer_device()
model = SamModel.from_pretrained("facebook/sam-vit-huge").to(device)
processor = SamProcessor.from_pretrained("facebook/sam-vit-huge")
img_url = "https://huggingface.co/ybelkada/segment-anything/resolve/main/assets/car.png"
raw_image = Image.open(requests.get(img_url, stream=True).raw).convert("RGB")
mask_url = "https://huggingface.co/ybelkada/segment-anything/resolve/main/assets/car.png"
segmentation_map = Image.open(requests.get(mask_url, stream=True).raw).convert("1")
input_points = [[[450, 600]]] # 2D location of a window in the image
inputs = processor(raw_image, input_points=input_points, segmentation_maps=segmentation_map, return_tensors="pt").to(device)
with torch.no_grad():
outputs = model(**inputs)
masks = processor.image_processor.post_process_masks(
outputs.pred_masks.cpu(), inputs["original_sizes"].cpu(), inputs["reshaped_input_sizes"].cpu()
)
scores = outputs.iou_scores
```
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with SAM.
- [Demo notebook](https://github.com/huggingface/notebooks/blob/main/examples/segment_anything.ipynb) for using the model.
- [Demo notebook](https://github.com/huggingface/notebooks/blob/main/examples/automatic_mask_generation.ipynb) for using the automatic mask generation pipeline.
- [Demo notebook](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/SAM/Run_inference_with_MedSAM_using_HuggingFace_Transformers.ipynb) for inference with MedSAM, a fine-tuned version of SAM on the medical domain. 🌎
- [Demo notebook](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/SAM/Fine_tune_SAM_(segment_anything)_on_a_custom_dataset.ipynb) for fine-tuning the model on custom data. 🌎
## SlimSAM
SlimSAM, a pruned version of SAM, was proposed in [0.1% Data Makes Segment Anything Slim](https://huggingface.co/papers/2312.05284) by Zigeng Chen et al. SlimSAM reduces the size of the SAM models considerably while maintaining the same performance.
Checkpoints can be found on the [hub](https://huggingface.co/models?other=slimsam), and they can be used as a drop-in replacement of SAM.
## Grounded SAM
One can combine [Grounding DINO](grounding-dino) with SAM for text-based mask generation as introduced in [Grounded SAM: Assembling Open-World Models for Diverse Visual Tasks](https://huggingface.co/papers/2401.14159). You can refer to this [demo notebook](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/Grounding%20DINO/GroundingDINO_with_Segment_Anything.ipynb) 🌍 for details.
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/grounded_sam.png"
alt="drawing" width="900"/>
<small> Grounded SAM overview. Taken from the <a href="https://github.com/IDEA-Research/Grounded-Segment-Anything">original repository</a>. </small>
## SamConfig
[[autodoc]] SamConfig
## SamVisionConfig
[[autodoc]] SamVisionConfig
## SamMaskDecoderConfig
[[autodoc]] SamMaskDecoderConfig
## SamPromptEncoderConfig
[[autodoc]] SamPromptEncoderConfig
## SamProcessor
[[autodoc]] SamProcessor
## SamImageProcessor
[[autodoc]] SamImageProcessor
## SamImageProcessorFast
[[autodoc]] SamImageProcessorFast
## SamVisionModel
[[autodoc]] SamVisionModel
- forward
## SamModel
[[autodoc]] SamModel
- forward
| transformers/docs/source/en/model_doc/sam.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/sam.md",
"repo_id": "transformers",
"token_count": 2350
} | 368 |
<!--Copyright 2021 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2021-04-14 and added to Hugging Face Transformers on 2021-09-01.*
# Speech Encoder Decoder Models
<div class="flex flex-wrap space-x-1">
<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
<img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
</div>
The [`SpeechEncoderDecoderModel`] can be used to initialize a speech-to-text model
with any pretrained speech autoencoding model as the encoder (*e.g.* [Wav2Vec2](wav2vec2), [Hubert](hubert)) and any pretrained autoregressive model as the decoder.
The effectiveness of initializing speech-sequence-to-text-sequence models with pretrained checkpoints for speech
recognition and speech translation has *e.g.* been shown in [Large-Scale Self- and Semi-Supervised Learning for Speech
Translation](https://huggingface.co/papers/2104.06678) by Changhan Wang, Anne Wu, Juan Pino, Alexei Baevski, Michael Auli,
Alexis Conneau.
An example of how to use a [`SpeechEncoderDecoderModel`] for inference can be seen in [Speech2Text2](speech_to_text_2).
## Randomly initializing `SpeechEncoderDecoderModel` from model configurations.
[`SpeechEncoderDecoderModel`] can be randomly initialized from an encoder and a decoder config. In the following example, we show how to do this using the default [`Wav2Vec2Model`] configuration for the encoder
and the default [`BertForCausalLM`] configuration for the decoder.
```python
>>> from transformers import BertConfig, Wav2Vec2Config, SpeechEncoderDecoderConfig, SpeechEncoderDecoderModel
>>> config_encoder = Wav2Vec2Config()
>>> config_decoder = BertConfig()
>>> config = SpeechEncoderDecoderConfig.from_encoder_decoder_configs(config_encoder, config_decoder)
>>> model = SpeechEncoderDecoderModel(config=config)
```
## Initialising `SpeechEncoderDecoderModel` from a pretrained encoder and a pretrained decoder.
[`SpeechEncoderDecoderModel`] can be initialized from a pretrained encoder checkpoint and a pretrained decoder checkpoint. Note that any pretrained Transformer-based speech model, *e.g.* [Wav2Vec2](wav2vec2), [Hubert](hubert) can serve as the encoder and both pretrained auto-encoding models, *e.g.* BERT, pretrained causal language models, *e.g.* GPT2, as well as the pretrained decoder part of sequence-to-sequence models, *e.g.* decoder of BART, can be used as the decoder.
Depending on which architecture you choose as the decoder, the cross-attention layers might be randomly initialized.
Initializing [`SpeechEncoderDecoderModel`] from a pretrained encoder and decoder checkpoint requires the model to be fine-tuned on a downstream task, as has been shown in [the *Warm-starting-encoder-decoder blog post*](https://huggingface.co/blog/warm-starting-encoder-decoder).
To do so, the `SpeechEncoderDecoderModel` class provides a [`SpeechEncoderDecoderModel.from_encoder_decoder_pretrained`] method.
```python
>>> from transformers import SpeechEncoderDecoderModel
>>> model = SpeechEncoderDecoderModel.from_encoder_decoder_pretrained(
... "facebook/hubert-large-ll60k", "google-bert/bert-base-uncased"
... )
```
## Loading an existing `SpeechEncoderDecoderModel` checkpoint and perform inference.
To load fine-tuned checkpoints of the `SpeechEncoderDecoderModel` class, [`SpeechEncoderDecoderModel`] provides the `from_pretrained(...)` method just like any other model architecture in Transformers.
To perform inference, one uses the [`generate`] method, which allows to autoregressively generate text. This method supports various forms of decoding, such as greedy, beam search and multinomial sampling.
```python
>>> from transformers import Wav2Vec2Processor, SpeechEncoderDecoderModel
>>> from datasets import load_dataset
>>> import torch
>>> # load a fine-tuned speech translation model and corresponding processor
>>> model = SpeechEncoderDecoderModel.from_pretrained("facebook/wav2vec2-xls-r-300m-en-to-15")
>>> processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-xls-r-300m-en-to-15")
>>> # let's perform inference on a piece of English speech (which we'll translate to German)
>>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
>>> input_values = processor(ds[0]["audio"]["array"], return_tensors="pt").input_values
>>> # autoregressively generate transcription (uses greedy decoding by default)
>>> generated_ids = model.generate(input_values)
>>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
>>> print(generated_text)
Mr. Quilter ist der Apostel der Mittelschicht und wir freuen uns, sein Evangelium willkommen heißen zu können.
```
## Training
Once the model is created, it can be fine-tuned similar to BART, T5 or any other encoder-decoder model on a dataset of (speech, text) pairs.
As you can see, only 2 inputs are required for the model in order to compute a loss: `input_values` (which are the
speech inputs) and `labels` (which are the `input_ids` of the encoded target sequence).
```python
>>> from transformers import AutoTokenizer, AutoFeatureExtractor, SpeechEncoderDecoderModel
>>> from datasets import load_dataset
>>> encoder_id = "facebook/wav2vec2-base-960h" # acoustic model encoder
>>> decoder_id = "google-bert/bert-base-uncased" # text decoder
>>> feature_extractor = AutoFeatureExtractor.from_pretrained(encoder_id)
>>> tokenizer = AutoTokenizer.from_pretrained(decoder_id)
>>> # Combine pre-trained encoder and pre-trained decoder to form a Seq2Seq model
>>> model = SpeechEncoderDecoderModel.from_encoder_decoder_pretrained(encoder_id, decoder_id)
>>> model.config.decoder_start_token_id = tokenizer.cls_token_id
>>> model.config.pad_token_id = tokenizer.pad_token_id
>>> # load an audio input and pre-process (normalise mean/std to 0/1)
>>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
>>> input_values = feature_extractor(ds[0]["audio"]["array"], return_tensors="pt").input_values
>>> # load its corresponding transcription and tokenize to generate labels
>>> labels = tokenizer(ds[0]["text"], return_tensors="pt").input_ids
>>> # the forward function automatically creates the correct decoder_input_ids
>>> loss = model(input_values=input_values, labels=labels).loss
>>> loss.backward()
```
## SpeechEncoderDecoderConfig
[[autodoc]] SpeechEncoderDecoderConfig
## SpeechEncoderDecoderModel
[[autodoc]] SpeechEncoderDecoderModel
- forward
- from_encoder_decoder_pretrained
| transformers/docs/source/en/model_doc/speech-encoder-decoder.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/speech-encoder-decoder.md",
"repo_id": "transformers",
"token_count": 2238
} | 369 |
<!--Copyright 2025 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2025-04-08 and added to Hugging Face Transformers on 2025-06-25.*
<div style="float: right;">
<div class="flex flex-wrap space-x-1">
<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
<img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
</div>
</div>
# T5Gemma
T5Gemma (aka encoder-decoder Gemma) was proposed in a [research paper](https://huggingface.co/papers/2504.06225) by Google. It is a family of encoder-decoder large language models, developed by adapting pretrained decoder-only models into encoder-decoder. T5Gemma includes pretrained and instruction-tuned variants. The architecture is based on transformer encoder-decoder design following T5, with improvements from Gemma 2: GQA, RoPE, GeGLU activation, RMSNorm, and interleaved local/global attention.
T5Gemma has two groups of model sizes: 1) [Gemma 2](https://ai.google.dev/gemma/docs/core/model_card_2) sizes (2B-2B, 9B-2B, and 9B-9B), which are based on the official Gemma 2 models (2B and 9B); and 2) [T5](https://huggingface.co/papers/1910.10683) sizes (Small, Base, Large, and XL), where are pretrained under the Gemma 2 framework following T5 configuration. In addition, we also provide a model at ML size (medium large, ~2B in total), which is in-between T5 Large and T5 XL.
The pretrained variants are trained with two objectives: prefix language modeling with knowledge distillation (PrefixLM) and UL2, separately. We release both variants for each model size. The instruction-turned variants was post-trained with supervised fine-tuning and reinforcement learning.
> [!TIP]
> Click on the T5Gemma models in the right sidebar for more examples of how to apply T5Gemma to different language tasks.
The example below demonstrates how to chat with the model with [`Pipeline`] or the [`AutoModel`] class, and from the command line.
<hfoptions id="usage">
<hfoption id="Pipeline">
```python
import torch
from transformers import pipeline
pipe = pipeline(
"text2text-generation",
model="google/t5gemma-2b-2b-prefixlm-it",
dtype=torch.bfloat16,
device_map="auto",
)
messages = [
{"role": "user", "content": "Tell me an unknown interesting biology fact about the brain."},
]
prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipe(prompt, max_new_tokens=32)
```
</hfoption>
<hfoption id="AutoModel">
```python
# pip install accelerate
import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tokenizer = AutoTokenizer.from_pretrained("google/t5gemma-2b-2b-prefixlm-it")
model = AutoModelForSeq2SeqLM.from_pretrained(
"google/t5gemma-2b-2b-prefixlm-it",
device_map="auto",
dtype=torch.bfloat16,
)
messages = [
{"role": "user", "content": "Tell me an unknown interesting biology fact about the brain."},
]
input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt", return_dict=True, add_generation_prompt=True).to(model.device)
outputs = model.generate(**input_ids, max_new_tokens=32)
print(tokenizer.decode(outputs[0]))
```
</hfoption>
<hfoption id="transformers CLI">
```
echo -e "Write me a poem about Machine Learning. Answer:" | transformers run --task text2text-generation --model google/t5gemma-2b-2b-prefixlm --device 0
```
</hfoption>
</hfoptions>
## T5GemmaConfig
[[autodoc]] T5GemmaConfig
## T5GemmaModuleConfig
[[autodoc]] T5GemmaModuleConfig
## T5GemmaModel
[[autodoc]] T5GemmaModel
- forward
## T5GemmaEncoderModel
[[autodoc]] T5GemmaEncoderModel
- forward
## T5GemmaForConditionalGeneration
[[autodoc]] T5GemmaForConditionalGeneration
- forward
## T5GemmaForSequenceClassification
[[autodoc]] T5GemmaForSequenceClassification
- forward
## T5GemmaForTokenClassification
[[autodoc]] T5GemmaForTokenClassification
- forward
| transformers/docs/source/en/model_doc/t5gemma.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/t5gemma.md",
"repo_id": "transformers",
"token_count": 1624
} | 370 |
<!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2021-11-11 and added to Hugging Face Transformers on 2022-01-18.*
<div style="float: right;">
<div class="flex flex-wrap space-x-1">
<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
<img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
</div>
</div>
# ViTMAE
[ViTMAE](https://huggingface.co/papers/2111.06377) is a self-supervised vision model that is pretrained by masking large portions of an image (~75%). An encoder processes the visible image patches and a decoder reconstructs the missing pixels from the encoded patches and mask tokens. After pretraining, the encoder can be reused for downstream tasks like image classification or object detection — often outperforming models trained with supervised learning.
<img src="https://user-images.githubusercontent.com/11435359/146857310-f258c86c-fde6-48e8-9cee-badd2b21bd2c.png"
alt="drawing" width="600"/>
You can find all the original ViTMAE checkpoints under the [AI at Meta](https://huggingface.co/facebook?search_models=vit-mae) organization.
> [!TIP]
> Click on the ViTMAE models in the right sidebar for more examples of how to apply ViTMAE to vision tasks.
The example below demonstrates how to reconstruct the missing pixels with the [`ViTMAEForPreTraining`] class.
<hfoptions id="usage">
<hfoption id="AutoModel">
```python
import torch
import requests
from PIL import Image
from transformers import infer_device, ViTImageProcessor, ViTMAEForPreTraining
device = infer_device()
url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
image = Image.open(requests.get(url, stream=True).raw)
processor = ViTImageProcessor.from_pretrained("facebook/vit-mae-base")
inputs = processor(image, return_tensors="pt")
inputs = {k: v.to(device) for k, v in inputs.items()}
model = ViTMAEForPreTraining.from_pretrained("facebook/vit-mae-base", attn_implementation="sdpa").to(device)
with torch.no_grad():
outputs = model(**inputs)
reconstruction = outputs.logits
```
</hfoption>
</hfoptions>
## Notes
- ViTMAE is typically used in two stages. Self-supervised pretraining with [`ViTMAEForPreTraining`], and then discarding the decoder and fine-tuning the encoder. After fine-tuning, the weights can be plugged into a model like [`ViTForImageClassification`].
- Use [`ViTImageProcessor`] for input preparation.
## Resources
- Refer to this [notebook](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/ViTMAE/ViT_MAE_visualization_demo.ipynb) to learn how to visualize the reconstructed pixels from [`ViTMAEForPreTraining`].
## ViTMAEConfig
[[autodoc]] ViTMAEConfig
## ViTMAEModel
[[autodoc]] ViTMAEModel
- forward
## ViTMAEForPreTraining
[[autodoc]] transformers.ViTMAEForPreTraining
- forward
| transformers/docs/source/en/model_doc/vit_mae.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/vit_mae.md",
"repo_id": "transformers",
"token_count": 1238
} | 371 |
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2023-02-23 and added to Hugging Face Transformers on 2024-07-08.*
<div style="float: right;">
<div class="flex flex-wrap space-x-1">
<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
</div>
</div>
# ZoeDepth
[ZoeDepth](https://huggingface.co/papers/2302.12288) is a depth estimation model that combines the generalization performance of relative depth estimation (how far objects are from each other) and metric depth estimation (precise depth measurement on metric scale) from a single image. It is pre-trained on 12 datasets using relative depth and 2 datasets (NYU Depth v2 and KITTI) for metric accuracy. A lightweight head with a metric bin module for each domain is used, and during inference, it automatically selects the appropriate head for each input image with a latent classifier.
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/zoedepth_architecture_bis.png"
alt="drawing" width="600"/>
You can find all the original ZoeDepth checkpoints under the [Intel](https://huggingface.co/Intel?search=zoedepth) organization.
The example below demonstrates how to estimate depth with [`Pipeline`] or the [`AutoModel`] class.
<hfoptions id="usage">
<hfoption id="Pipeline">
```py
import requests
import torch
from transformers import pipeline
from PIL import Image
url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
image = Image.open(requests.get(url, stream=True).raw)
pipeline = pipeline(
task="depth-estimation",
model="Intel/zoedepth-nyu-kitti",
dtype=torch.float16,
device=0
)
results = pipeline(image)
results["depth"]
```
</hfoption>
<hfoption id="AutoModel">
```py
import torch
import requests
from PIL import Image
from transformers import AutoModelForDepthEstimation, AutoImageProcessor
image_processor = AutoImageProcessor.from_pretrained(
"Intel/zoedepth-nyu-kitti"
)
model = AutoModelForDepthEstimation.from_pretrained(
"Intel/zoedepth-nyu-kitti",
device_map="auto"
)
url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
image = Image.open(requests.get(url, stream=True).raw)
inputs = image_processor(image, return_tensors="pt").to(model.device)
with torch.no_grad():
outputs = model(inputs)
# interpolate to original size and visualize the prediction
## ZoeDepth dynamically pads the input image, so pass the original image size as argument
## to `post_process_depth_estimation` to remove the padding and resize to original dimensions.
post_processed_output = image_processor.post_process_depth_estimation(
outputs,
source_sizes=[(image.height, image.width)],
)
predicted_depth = post_processed_output[0]["predicted_depth"]
depth = (predicted_depth - predicted_depth.min()) / (predicted_depth.max() - predicted_depth.min())
depth = depth.detach().cpu().numpy() * 255
Image.fromarray(depth.astype("uint8"))
```
</hfoption>
</hfoptions>
## Notes
- In the [original implementation](https://github.com/isl-org/ZoeDepth/blob/edb6daf45458569e24f50250ef1ed08c015f17a7/zoedepth/models/depth_model.py#L131) ZoeDepth performs inference on both the original and flipped images and averages the results. The `post_process_depth_estimation` function handles this by passing the flipped outputs to the optional `outputs_flipped` argument as shown below.
```py
with torch.no_grad():
outputs = model(pixel_values)
outputs_flipped = model(pixel_values=torch.flip(inputs.pixel_values, dims=[3]))
post_processed_output = image_processor.post_process_depth_estimation(
outputs,
source_sizes=[(image.height, image.width)],
outputs_flipped=outputs_flipped,
)
```
## Resources
- Refer to this [notebook](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/ZoeDepth) for an inference example.
## ZoeDepthConfig
[[autodoc]] ZoeDepthConfig
## ZoeDepthImageProcessor
[[autodoc]] ZoeDepthImageProcessor
- preprocess
## ZoeDepthImageProcessorFast
[[autodoc]] ZoeDepthImageProcessorFast
- preprocess
## ZoeDepthForDepthEstimation
[[autodoc]] ZoeDepthForDepthEstimation
- forward
| transformers/docs/source/en/model_doc/zoedepth.md/0 | {
"file_path": "transformers/docs/source/en/model_doc/zoedepth.md",
"repo_id": "transformers",
"token_count": 1633
} | 372 |
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Distributed CPUs
CPUs are commonly available and can be a cost-effective training option when GPUs are unavailable. When training large models or if a single CPU is too slow, distributed training with CPUs can help speed up training.
This guide demonstrates how to perform distributed training with multiple CPUs using a [DistributedDataParallel (DDP)](./perf_train_gpu_many#distributeddataparallel) strategy on bare metal with [`Trainer`] and a Kubernetes cluster. All examples shown in this guide depend on the [Intel oneAPI HPC Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/hpc-toolkit.html).
There are two toolkits you'll need from Intel oneAPI.
1. [oneCCL](https://www.intel.com/content/www/us/en/developer/tools/oneapi/oneccl.html) includes efficient implementations of collectives commonly used in deep learning such as all-gather, all-reduce, and reduce-scatter. To install from a prebuilt wheel, make sure you always use the latest release. Refer to the table [here](https://github.com/intel/torch-ccl#install-prebuilt-wheel) to check if a version of oneCCL is supported for a Python and PyTorch version.
```bash
# installs oneCCL for PyTorch 2.4.0
pip install oneccl_bind_pt==2.4.0 -f https://developer.intel.com/ipex-whl-stable-cpu
```
> [!TIP]
> Refer to the oneCCL [installation](https://github.com/intel/torch-ccl#installation) for more details.
1. [MPI](https://www.intel.com/content/www/us/en/developer/tools/oneapi/mpi-library.html) is a message-passing interface for communications between hardware and networks. The oneCCL toolkit is installed along with MPI, but you need to source the environment as shown below before using it.
```bash
oneccl_bindings_for_pytorch_path=$(python -c "from oneccl_bindings_for_pytorch import cwd; print(cwd)")
source $oneccl_bindings_for_pytorch_path/env/setvars.sh
```
Lastly, install the [Intex Extension for PyTorch (IPEX)](https://intel.github.io/intel-extension-for-pytorch/index.html) which enables additional performance optimizations for Intel hardware such as weight sharing and better thread runtime control.
```bash
pip install intel_extension_for_pytorch==<version_name> -f https://developer.intel.com/ipex-whl-stable-cpu
```
> [!TIP]
> Refer to the IPEX [installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation) for more details.
## Trainer
[`Trainer`] supports distributed training with CPUs with the oneCCL backend. Add the `--ddp_backend ccl` parameter in the command arguments to enable it.
<hfoptions id="distrib-cpu">
<hfoption id="single node">
The example below demonstrates the [run_qa.py](https://github.com/huggingface/transformers/tree/main/examples/pytorch/question-answering) script. It enables training with two processes on one Xeon CPU, with one process running per socket.
> [!TIP]
> Tune the variable `OMP_NUM_THREADS/CCL_WORKER_COUNT` for optimal performance.
```bash
export CCL_WORKER_COUNT=1
export MASTER_ADDR=127.0.0.1
mpirun -n 2 -genv OMP_NUM_THREADS=23 \
python3 run_qa.py \
--model_name_or_path google-bert/bert-large-uncased \
--dataset_name squad \
--do_train \
--do_eval \
--per_device_train_batch_size 12 \
--learning_rate 3e-5 \
--num_train_epochs 2 \
--max_seq_length 384 \
--doc_stride 128 \
--output_dir /tmp/debug_squad/ \
--no_cuda \
--ddp_backend ccl
```
</hfoption>
<hfoption id="multiple nodes">
Scale the training script to four processes on two Xeon CPUs (`node0` and `node1`) by setting `-n 4` and `ppn 2`. The `ppn` parameter specifies the number of processes per node, with one process running per socket.
Assume `node0` is the main process and create a configuration file containing the IP addresses of each node (for example, hostfile) and pass the configuration file path as an argument.
```bash
cat hostfile
xxx.xxx.xxx.xxx #node0 ip
xxx.xxx.xxx.xxx #node1 ip
```
Run the script below on `node0` to enable DDP on `node0` and `node1` and train with bf16 auto mixed precision.
> [!TIP]
> Tune the variable `OMP_NUM_THREADS/CCL_WORKER_COUNT` for optimal performance.
```bash
export CCL_WORKER_COUNT=1
export MASTER_ADDR=xxx.xxx.xxx.xxx #node0 ip
mpirun -f hostfile -n 4 -ppn 2 \
-genv OMP_NUM_THREADS=23 \
python3 run_qa.py \
--model_name_or_path google-bert/bert-large-uncased \
--dataset_name squad \
--do_train \
--do_eval \
--per_device_train_batch_size 12 \
--learning_rate 3e-5 \
--num_train_epochs 2 \
--max_seq_length 384 \
--doc_stride 128 \
--output_dir /tmp/debug_squad/ \
--no_cuda \
--ddp_backend ccl \
--bf16
```
</hfoption>
</hfoptions>
## Kubernetes
Distributed training with CPUs can also be deployed to a Kubernetes cluster with [PyTorchJob](https://www.kubeflow.org/docs/components/training/user-guides/pytorch/). Before you get started, you should perform the following setup steps.
1. Ensure you have access to a Kubernetes cluster with [Kubeflow](https://www.kubeflow.org/docs/started/installing-kubeflow/) installed.
1. Install and configure [kubectl](https://kubernetes.io/docs/tasks/tools) to interact with the cluster.
1. Set up a [PersistentVolumeClaim (PVC)](https://kubernetes.io/docs/concepts/storage/persistent-volumes/) to store datasets and model files. There are multiple options to choose from, including a [StorageClass](https://kubernetes.io/docs/concepts/storage/storage-classes/) or a cloud storage bucket.
1. Set up a Docker container for the training script and all required dependencies such as PyTorch, Transformers, IPEX, oneCCL, and OpenSSH to facilitate communicattion between containers.
The example Dockerfile below uses a base image that supports distributed training with CPUs, and extracts Transformers to the `/workspace` directory to include the training scripts in the image. The image needs to be built and copied to the clusters nodes or pushed to a container registry prior to deployment.
```dockerfile
FROM intel/intel-optimized-pytorch:2.4.0-pip-multinode
RUN apt-get update -y && \
apt-get install -y --no-install-recommends --fix-missing \
google-perftools \
libomp-dev
WORKDIR /workspace
# Download and extract the transformers code
ARG HF_TRANSFORMERS_VER="4.46.0"
RUN pip install --no-cache-dir \
transformers==${HF_TRANSFORMERS_VER} && \
mkdir transformers && \
curl -sSL --retry 5 https://github.com/huggingface/transformers/archive/refs/tags/v${HF_TRANSFORMERS_VER}.tar.gz | tar -C transformers --strip-components=1 -xzf -
```
### PyTorchJob
[PyTorchJob](https://www.kubeflow.org/docs/components/training/user-guides/pytorch/) is an extension of the Kubernetes API for running PyTorch training jobs on Kubernetes. It includes a yaml file that defines the training jobs parameters such as the name of the PyTorchJob, number of workers, types of resources for each worker, and more.
The volume mount parameter is a path to where the PVC is mounted in the container for each worker pod. The PVC is typically used to hold the dataset, checkpoint files, and the model after it has finished training.
The example yaml file below sets up four workers on the [run_qa.py](https://github.com/huggingface/transformers/tree/main/examples/pytorch/question-answering) script. Adapt the yaml file based on your training script and number of nodes in your cluster.
The CPU resource limits and requests are defined in [CPU units](https://kubernetes.io/docs/concepts/configuration/manage-resources-containers/#meaning-of-cpu). One CPU unit is equivalent to one physical CPU core or virtual core. The CPU units defined in the yaml file should be less than the amount of available CPU and memory capacity of a single machine in order to leave some resources for kubelet and the system. For a `Guaranteed` [quality of service](https://kubernetes.io/docs/tasks/configure-pod-container/quality-service-pod), set the same CPU and memory amounts for both the resource limits and requests.
```yaml
apiVersion: "kubeflow.org/v1"
kind: PyTorchJob
metadata:
name: transformers-pytorchjob
spec:
elasticPolicy:
rdzvBackend: c10d
minReplicas: 1
maxReplicas: 4
maxRestarts: 10
pytorchReplicaSpecs:
Worker:
replicas: 4 # The number of worker pods
restartPolicy: OnFailure
template:
spec:
containers:
- name: pytorch
image: <image name>:<tag> # Specify the docker image to use for the worker pods
imagePullPolicy: IfNotPresent
command: ["/bin/bash", "-c"]
args:
- >-
cd /workspace/transformers;
pip install -r /workspace/transformers/examples/pytorch/question-answering/requirements.txt;
source /usr/local/lib/python3.10/dist-packages/oneccl_bindings_for_pytorch/env/setvars.sh;
torchrun /workspace/transformers/examples/pytorch/question-answering/run_qa.py \
--model_name_or_path distilbert/distilbert-base-uncased \
--dataset_name squad \
--do_train \
--do_eval \
--per_device_train_batch_size 12 \
--learning_rate 3e-5 \
--num_train_epochs 2 \
--max_seq_length 384 \
--doc_stride 128 \
--output_dir /tmp/pvc-mount/output_$(date +%Y%m%d_%H%M%S) \
--no_cuda \
--ddp_backend ccl \
--bf16;
env:
- name: LD_PRELOAD
value: "/usr/lib/x86_64-linux-gnu/libtcmalloc.so.4.5.9:/usr/local/lib/libiomp5.so"
- name: TRANSFORMERS_CACHE
value: "/tmp/pvc-mount/transformers_cache"
- name: HF_DATASETS_CACHE
value: "/tmp/pvc-mount/hf_datasets_cache"
- name: LOGLEVEL
value: "INFO"
- name: CCL_WORKER_COUNT
value: "1"
- name: OMP_NUM_THREADS # Can be tuned for optimal performance
value: "240"
resources:
limits:
cpu: 240 # Update the CPU and memory limit values based on your nodes
memory: 128Gi
requests:
cpu: 240 # Update the CPU and memory request values based on your nodes
memory: 128Gi
volumeMounts:
- name: pvc-volume
mountPath: /tmp/pvc-mount
- mountPath: /dev/shm
name: dshm
restartPolicy: Never
nodeSelector: # Optionally use nodeSelector to match a certain node label for the worker pods
node-type: gnr
volumes:
- name: pvc-volume
persistentVolumeClaim:
claimName: transformers-pvc
- name: dshm
emptyDir:
medium: Memory
```
### Deploy
After you've setup the PyTorchJob yaml file with the appropriate settings for your cluster and training job, deploy it to the cluster with the command below.
```bash
export NAMESPACE=<specify your namespace>
kubectl create -f pytorchjob.yaml -n ${NAMESPACE}
```
List the pods in the namespace with `kubectl get pods -n ${NAMESPACE}`. At first, the status may be "Pending" but it should change to "Running" once the containers are pulled and created.
```bash
kubectl get pods -n ${NAMESPACE}
NAME READY STATUS RESTARTS AGE
...
transformers-pytorchjob-worker-0 1/1 Running 0 7m37s
transformers-pytorchjob-worker-1 1/1 Running 0 7m37s
transformers-pytorchjob-worker-2 1/1 Running 0 7m37s
transformers-pytorchjob-worker-3 1/1 Running 0 7m37s
...
```
Inspect the logs for each worker with the following command. Add `-f` to stream the logs.
```bash
kubectl logs transformers-pytorchjob-worker-0 -n ${NAMESPACE} -f
```
Once training is complete, the trained model can be copied from the PVC or storage location. Delete the PyTorchJob resource from the cluster with the command below.
```bash
kubectl delete -f pytorchjob.yaml -n ${NAMESPACE}
```
| transformers/docs/source/en/perf_train_cpu_many.md/0 | {
"file_path": "transformers/docs/source/en/perf_train_cpu_many.md",
"repo_id": "transformers",
"token_count": 5034
} | 373 |
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Bitsandbytes
The [bitsandbytes](https://github.com/bitsandbytes-foundation/bitsandbytes) library provides quantization tools for LLMs through a lightweight Python wrapper around CUDA functions. It enables working with large models using limited computational resources by reducing their memory footprint.
At its core, bitsandbytes provides:
- **Quantized Linear Layers**: `Linear8bitLt` and `Linear4bit` layers that replace standard PyTorch linear layers with memory-efficient quantized alternatives
- **Optimized Optimizers**: 8-bit versions of common optimizers through its `optim` module, enabling training of large models with reduced memory requirements
- **Matrix Multiplication**: Optimized matrix multiplication operations that leverage the quantized format
bitsandbytes offers two main quantization features:
1. **LLM.int8()** - An 8-bit quantization method that makes inference more accessible without significant performance degradation. Unlike naive quantization, [LLM.int8()](https://hf.co/papers/2208.07339) dynamically preserves higher precision for critical computations, preventing information loss in sensitive parts of the model.
2. **QLoRA** - A 4-bit quantization technique that compresses models even further while maintaining trainability by inserting a small set of trainable low-rank adaptation (LoRA) weights.
> **Note:** For a user-friendly quantization experience, you can use the `bitsandbytes` [community space](https://huggingface.co/spaces/bnb-community/bnb-my-repo).
Run the command below to install bitsandbytes.
```bash
pip install --upgrade transformers accelerate bitsandbytes
```
To compile from source, follow the instructions in the [bitsandbytes installation guide](https://huggingface.co/docs/bitsandbytes/main/en/installation).
## Hardware Compatibility
bitsandbytes is currently only supported on CUDA GPUs for CUDA versions 11.0 - 12.8. However, there's an ongoing multi-backend effort under development, which is currently in alpha. If you're interested in providing feedback or testing, check out the [bitsandbytes repository](https://github.com/bitsandbytes-foundation/bitsandbytes) for more information.
### CUDA
| Feature | Minimum Hardware Requirement |
|---------|-------------------------------|
| 8-bit optimizers | NVIDIA Maxwell (GTX 900 series, TITAN X, M40) or newer GPUs * |
| LLM.int8() | NVIDIA Turing (RTX 20 series, T4) or newer GPUs |
| NF4/FP4 quantization | NVIDIA Maxwell (GTX 900 series, TITAN X, M40) or newer GPUs * |
### Multi-backend
| Backend | Supported Versions | Python versions | Architecture Support | Status |
|---------|-------------------|----------------|---------------------|---------|
| AMD ROCm | 6.1+ | 3.10+ | minimum CDNA - gfx90a, RDNA - gfx1100 | Alpha |
| Apple Silicon (MPS) | WIP | 3.10+ | M1/M2 chips | Planned |
| Intel CPU | v2.4.0+ (ipex) | 3.10+ | Intel CPU | Alpha |
| Intel GPU | v2.4.0+ (ipex) | 3.10+ | Intel GPU | Experimental |
| Ascend NPU | 2.1.0+ (torch_npu) | 3.10+ | Ascend NPU | Experimental |
> **Note:** Bitsandbytes is moving away from the multi-backend approach towards using [Pytorch Custom Operators](https://pytorch.org/tutorials/advanced/custom_ops_landing_page.html), as the main mechanism for supporting new hardware, and dispatching to the correct backend.
## Quantization Examples
Quantize a model by passing a [`BitsAndBytesConfig`] to [`~PreTrainedModel.from_pretrained`]. This works for any model in any modality, as long as it supports [Accelerate](https://huggingface.co/docs/accelerate/index) and contains [torch.nn.Linear](https://pytorch.org/docs/stable/generated/torch.nn.Linear.html) layers.
<hfoptions id="bnb">
<hfoption id="8-bit">
<div class="bnb-container" style="border: 1px solid #ddd; border-radius: 8px; padding: 20px; margin: 20px 0">
Quantizing a model in 8-bit halves the memory-usage, and for large models, set `device_map="auto"` to efficiently distribute the weights across all available GPUs.
```py
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(load_in_8bit=True)
model_8bit = AutoModelForCausalLM.from_pretrained(
"bigscience/bloom-1b7",
device_map="auto",
quantization_config=quantization_config
)
```
By default, all other modules such as [torch.nn.LayerNorm](https://pytorch.org/docs/stable/generated/torch.nn.LayerNorm.html) are set to the default torch dtype. You can change the data type of these modules with the `dtype` parameter. Setting `dtype="auto"` loads the model in the data type defined in a model's `config.json` file.
```py
import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(load_in_8bit=True)
model_8bit = AutoModelForCausalLM.from_pretrained(
"facebook/opt-350m",
device_map="auto",
quantization_config=quantization_config,
dtype="auto"
)
model_8bit.model.decoder.layers[-1].final_layer_norm.weight.dtype
```
Once a model is quantized to 8-bit, you can't push the quantized weights to the Hub unless you're using the latest version of Transformers and bitsandbytes. If you have the latest versions, then you can push the 8-bit model to the Hub with [`~PreTrainedModel.push_to_hub`]. The quantization config.json file is pushed first, followed by the quantized model weights.
```py
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(load_in_8bit=True)
model = AutoModelForCausalLM.from_pretrained(
"bigscience/bloom-560m",
device_map="auto",
quantization_config=quantization_config
)
model.push_to_hub("bloom-560m-8bit")
```
</div>
</hfoption>
<hfoption id="4-bit">
<div class="bnb-container" style="border: 1px solid #ddd; border-radius: 8px; padding: 20px; margin: 20px 0">
Quantizing a model in 4-bit reduces your memory-usage by 4x, and for large models, set `device_map="auto"` to efficiently distribute the weights across all available GPUs.
```py
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(load_in_4bit=True)
model_4bit = AutoModelForCausalLM.from_pretrained(
"bigscience/bloom-1b7",
device_map="auto",
quantization_config=quantization_config
)
```
By default, all other modules such as [torch.nn.LayerNorm](https://pytorch.org/docs/stable/generated/torch.nn.LayerNorm.html) are converted to `torch.float16`. You can change the data type of these modules with the `dtype` parameter.. Setting `dtype="auto"` loads the model in the data type defined in a model's `config.json` file.
```py
import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(load_in_4bit=True)
model_4bit = AutoModelForCausalLM.from_pretrained(
"facebook/opt-350m",
device_map="auto",
quantization_config=quantization_config,
dtype="auto"
)
model_4bit.model.decoder.layers[-1].final_layer_norm.weight.dtype
```
Make sure you have the latest bitsandbytes version so you can serialize 4-bit models and push them to the Hub with [`~PreTrainedModel.push_to_hub`]. Use [`~PreTrainedModel.save_pretrained`] to save the 4-bit model locally.
```py
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(load_in_4bit=True)
model = AutoModelForCausalLM.from_pretrained(
"bigscience/bloom-560m",
device_map="auto",
quantization_config=quantization_config
)
model.push_to_hub("bloom-560m-4bit")
```
</div>
</hfoption>
</hfoptions>
> [!WARNING]
> 8 and 4-bit training is only supported for training *extra* parameters.
Check your memory footprint with `get_memory_footprint`.
```py
print(model.get_memory_footprint())
```
Load quantized models with [`~PreTrainedModel.from_pretrained`] without a `quantization_config`.
```py
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("{your_username}/bloom-560m-8bit", device_map="auto")
```
## LLM.int8
This section explores some of the specific features of 8-bit quantization, such as offloading, outlier thresholds, skipping module conversion, and finetuning.
### Offloading
8-bit models can offload weights between the CPU and GPU to fit very large models into memory. The weights dispatched to the CPU are stored in **float32** and aren't converted to 8-bit. For example, enable offloading for [bigscience/bloom-1b7](https://huggingface.co/bigscience/bloom-1b7) through [`BitsAndBytesConfig`].
```py
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(llm_int8_enable_fp32_cpu_offload=True)
```
Design a custom device map to fit everything on your GPU except for the `lm_head`, which is dispatched to the CPU.
```py
device_map = {
"transformer.word_embeddings": 0,
"transformer.word_embeddings_layernorm": 0,
"lm_head": "cpu",
"transformer.h": 0,
"transformer.ln_f": 0,
}
```
Now load your model with the custom `device_map` and `quantization_config`.
```py
model_8bit = AutoModelForCausalLM.from_pretrained(
"bigscience/bloom-1b7",
dtype="auto",
device_map=device_map,
quantization_config=quantization_config,
)
```
### Outlier threshold
An "outlier" is a hidden state value greater than a certain threshold, and these values are computed in fp16. While the values are usually normally distributed ([-3.5, 3.5]), this distribution can be very different for large models ([-60, 6] or [6, 60]). 8-bit quantization works well for values ~5, but beyond that, there is a significant performance penalty. A good default threshold value is 6, but a lower threshold may be needed for more unstable models (small models or finetuning).
To find the best threshold for your model, experiment with the `llm_int8_threshold` parameter in [`BitsAndBytesConfig`]. For example, setting the threshold to `0.0` significantly speeds up inference at the potential cost of some accuracy loss.
```py
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
model_id = "bigscience/bloom-1b7"
quantization_config = BitsAndBytesConfig(
llm_int8_threshold=0.0,
llm_int8_enable_fp32_cpu_offload=True
)
model_8bit = AutoModelForCausalLM.from_pretrained(
model_id,
dtype="auto",
device_map=device_map,
quantization_config=quantization_config,
)
```
### Skip module conversion
For some models, like [Jukebox](model_doc/jukebox), you don't need to quantize every module to 8-bit because it can actually cause instability. With Jukebox, there are several `lm_head` modules that should be skipped using the `llm_int8_skip_modules` parameter in [`BitsAndBytesConfig`].
```py
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
model_id = "bigscience/bloom-1b7"
quantization_config = BitsAndBytesConfig(
llm_int8_skip_modules=["lm_head"],
)
model_8bit = AutoModelForCausalLM.from_pretrained(
model_id,
dtype="auto",
device_map="auto",
quantization_config=quantization_config,
)
```
### Finetuning
The [PEFT](https://github.com/huggingface/peft) library supports fine-tuning large models like [flan-t5-large](https://huggingface.co/google/flan-t5-large) and [facebook/opt-6.7b](https://huggingface.co/facebook/opt-6.7b) with 8-bit quantization. You don't need to pass the `device_map` parameter for training because it automatically loads your model on a GPU. However, you can still customize the device map with the `device_map` parameter (`device_map="auto"` should only be used for inference).
## QLoRA
This section explores some of the specific features of 4-bit quantization, such as changing the compute data type, the Normal Float 4 (NF4) data type, and nested quantization.
### Compute data type
Change the data type from float32 (the default value) to bf16 in [`BitsAndBytesConfig`] to speedup computation.
```py
import torch
from transformers import BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16)
```
### Normal Float 4 (NF4)
NF4 is a 4-bit data type from the [QLoRA](https://hf.co/papers/2305.14314) paper, adapted for weights initialized from a normal distribution. You should use NF4 for training 4-bit base models.
```py
from transformers import BitsAndBytesConfig
nf4_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
)
model_nf4 = AutoModelForCausalLM.from_pretrained(model_id, dtype="auto", quantization_config=nf4_config)
```
For inference, the `bnb_4bit_quant_type` does not have a huge impact on performance. However, to remain consistent with the model weights, you should use the `bnb_4bit_compute_dtype` and `dtype` values.
### Nested quantization
Nested quantization can save additional memory at no additional performance cost. This feature performs a second quantization of the already quantized weights to save an additional 0.4 bits/parameter. For example, with nested quantization, you can finetune a [Llama-13b](https://huggingface.co/meta-llama/Llama-2-13b) model on a 16GB NVIDIA T4 GPU with a sequence length of 1024, a batch size of 1, and enable gradient accumulation with 4 steps.
```py
from transformers import BitsAndBytesConfig
double_quant_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_use_double_quant=True,
)
model_double_quant = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-13b-chat-hf", dtype="auto", quantization_config=double_quant_config)
```
## Dequantizing bitsandbytes models
Once quantized, you can [`~PreTrainedModel.dequantize`] a model to the original precision but this may result in some quality loss. Make sure you have enough GPU memory to fit the dequantized model.
```python
from transformers import AutoModelForCausalLM, BitsAndBytesConfig, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("facebook/opt-125m", BitsAndBytesConfig(load_in_4bit=True))
model.dequantize()
```
## Resources
Learn more about the details of 8-bit quantization in [A Gentle Introduction to 8-bit Matrix Multiplication for transformers at scale using Hugging Face Transformers, Accelerate and bitsandbytes](https://huggingface.co/blog/hf-bitsandbytes-integration).
Try 4-bit quantization in this [notebook](https://colab.research.google.com/drive/1ge2F1QSK8Q7h0hn3YKuBCOAS0bK8E0wf) and learn more about it's details in [Making LLMs even more accessible with bitsandbytes, 4-bit quantization and QLoRA](https://huggingface.co/blog/4bit-transformers-bitsandbytes).
| transformers/docs/source/en/quantization/bitsandbytes.md/0 | {
"file_path": "transformers/docs/source/en/quantization/bitsandbytes.md",
"repo_id": "transformers",
"token_count": 4708
} | 374 |
<!--Copyright 2025 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# SpQR
The [SpQR](https://hf.co/papers/2306.03078) quantization algorithm involves a 16x16 tiled bi-level group 3-bit quantization structure with sparse outliers.
<div class="flex justify-center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/spqr-diagram.png">
</div>
> [!TIP]
> To quantize a model with SpQR, refer to the [Vahe1994/SpQR](https://github.com/Vahe1994/SpQR) repository.
Load a SpQR-quantized model with [`~PreTrainedModel.from_pretrained`].
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
quantized_model = AutoModelForCausalLM.from_pretrained(
"elvircrn/Llama-2-7b-SPQR-3Bit-16x16-red_pajama-hf",
dtype=torch.half,
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("elvircrn/Llama-2-7b-SPQR-3Bit-16x16-red_pajama-hf")
```
| transformers/docs/source/en/quantization/spqr.md/0 | {
"file_path": "transformers/docs/source/en/quantization/spqr.md",
"repo_id": "transformers",
"token_count": 530
} | 375 |
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Image-to-Image Task Guide
[[open-in-colab]]
Image-to-Image task is the task where an application receives an image and outputs another image. This has various subtasks, including image enhancement (super resolution, low light enhancement, deraining and so on), image inpainting, and more.
This guide will show you how to:
- Use an image-to-image pipeline for super resolution task,
- Run image-to-image models for same task without a pipeline.
Note that as of the time this guide is released, `image-to-image` pipeline only supports super resolution task.
Let's begin by installing the necessary libraries.
```bash
pip install transformers
```
We can now initialize the pipeline with a [Swin2SR model](https://huggingface.co/caidas/swin2SR-lightweight-x2-64). We can then infer with the pipeline by calling it with an image. As of now, only [Swin2SR models](https://huggingface.co/models?sort=trending&search=swin2sr) are supported in this pipeline.
```python
from transformers import pipeline, infer_device
import torch
# automatically detects the underlying device type (CUDA, CPU, XPU, MPS, etc.)
device = infer_device()
pipe = pipeline(task="image-to-image", model="caidas/swin2SR-lightweight-x2-64", device=device)
```
Now, let's load an image.
```python
from PIL import Image
import requests
url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/cat.jpg"
image = Image.open(requests.get(url, stream=True).raw)
print(image.size)
```
```bash
# (532, 432)
```
<div class="flex justify-center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/cat.jpg" alt="Photo of a cat"/>
</div>
We can now do inference with the pipeline. We will get an upscaled version of the cat image.
```python
upscaled = pipe(image)
print(upscaled.size)
```
```bash
# (1072, 880)
```
If you wish to do inference yourself with no pipeline, you can use the `Swin2SRForImageSuperResolution` and `Swin2SRImageProcessor` classes of transformers. We will use the same model checkpoint for this. Let's initialize the model and the processor.
```python
from transformers import Swin2SRForImageSuperResolution, Swin2SRImageProcessor
model = Swin2SRForImageSuperResolution.from_pretrained("caidas/swin2SR-lightweight-x2-64").to(device)
processor = Swin2SRImageProcessor("caidas/swin2SR-lightweight-x2-64")
```
`pipeline` abstracts away the preprocessing and postprocessing steps that we have to do ourselves, so let's preprocess the image. We will pass the image to the processor and then move the pixel values to GPU.
```python
pixel_values = processor(image, return_tensors="pt").pixel_values
print(pixel_values.shape)
pixel_values = pixel_values.to(device)
```
We can now infer the image by passing pixel values to the model.
```python
import torch
with torch.no_grad():
outputs = model(pixel_values)
```
Output is an object of type `ImageSuperResolutionOutput` that looks like below 👇
```
(loss=None, reconstruction=tensor([[[[0.8270, 0.8269, 0.8275, ..., 0.7463, 0.7446, 0.7453],
[0.8287, 0.8278, 0.8283, ..., 0.7451, 0.7448, 0.7457],
[0.8280, 0.8273, 0.8269, ..., 0.7447, 0.7446, 0.7452],
...,
[0.5923, 0.5933, 0.5924, ..., 0.0697, 0.0695, 0.0706],
[0.5926, 0.5932, 0.5926, ..., 0.0673, 0.0687, 0.0705],
[0.5927, 0.5914, 0.5922, ..., 0.0664, 0.0694, 0.0718]]]],
device='cuda:0'), hidden_states=None, attentions=None)
```
We need to get the `reconstruction` and post-process it for visualization. Let's see how it looks like.
```python
outputs.reconstruction.data.shape
# torch.Size([1, 3, 880, 1072])
```
We need to squeeze the output and get rid of axis 0, clip the values, then convert it to be numpy float. Then we will arrange axes to have the shape [1072, 880], and finally, bring the output back to range [0, 255].
```python
import numpy as np
# squeeze, take to CPU and clip the values
output = outputs.reconstruction.data.squeeze().cpu().clamp_(0, 1).numpy()
# rearrange the axes
output = np.moveaxis(output, source=0, destination=-1)
# bring values back to pixel values range
output = (output * 255.0).round().astype(np.uint8)
Image.fromarray(output)
```
<div class="flex justify-center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/cat_upscaled.png" alt="Upscaled photo of a cat"/>
</div>
| transformers/docs/source/en/tasks/image_to_image.md/0 | {
"file_path": "transformers/docs/source/en/tasks/image_to_image.md",
"repo_id": "transformers",
"token_count": 1736
} | 376 |
<!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Translation
[[open-in-colab]]
<Youtube id="1JvfrvZgi6c"/>
Translation converts a sequence of text from one language to another. It is one of several tasks you can formulate as a sequence-to-sequence problem, a powerful framework for returning some output from an input, like translation or summarization. Translation systems are commonly used for translation between different language texts, but it can also be used for speech or some combination in between like text-to-speech or speech-to-text.
This guide will show you how to:
1. Finetune [T5](https://huggingface.co/google-t5/t5-small) on the English-French subset of the [OPUS Books](https://huggingface.co/datasets/opus_books) dataset to translate English text to French.
2. Use your finetuned model for inference.
<Tip>
To see all architectures and checkpoints compatible with this task, we recommend checking the [task-page](https://huggingface.co/tasks/translation).
</Tip>
Before you begin, make sure you have all the necessary libraries installed:
```bash
pip install transformers datasets evaluate sacrebleu
```
We encourage you to login to your Hugging Face account so you can upload and share your model with the community. When prompted, enter your token to login:
```py
>>> from huggingface_hub import notebook_login
>>> notebook_login()
```
## Load OPUS Books dataset
Start by loading the English-French subset of the [OPUS Books](https://huggingface.co/datasets/opus_books) dataset from the 🤗 Datasets library:
```py
>>> from datasets import load_dataset
>>> books = load_dataset("opus_books", "en-fr")
```
Split the dataset into a train and test set with the [`~datasets.Dataset.train_test_split`] method:
```py
>>> books = books["train"].train_test_split(test_size=0.2)
```
Then take a look at an example:
```py
>>> books["train"][0]
{'id': '90560',
'translation': {'en': 'But this lofty plateau measured only a few fathoms, and soon we reentered Our Element.',
'fr': 'Mais ce plateau élevé ne mesurait que quelques toises, et bientôt nous fûmes rentrés dans notre élément.'}}
```
`translation`: an English and French translation of the text.
## Preprocess
<Youtube id="XAR8jnZZuUs"/>
The next step is to load a T5 tokenizer to process the English-French language pairs:
```py
>>> from transformers import AutoTokenizer
>>> checkpoint = "google-t5/t5-small"
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
```
The preprocessing function you want to create needs to:
1. Prefix the input with a prompt so T5 knows this is a translation task. Some models capable of multiple NLP tasks require prompting for specific tasks.
2. Set the target language (French) in the `text_target` parameter to ensure the tokenizer processes the target text correctly. If you don't set `text_target`, the tokenizer processes the target text as English.
3. Truncate sequences to be no longer than the maximum length set by the `max_length` parameter.
```py
>>> source_lang = "en"
>>> target_lang = "fr"
>>> prefix = "translate English to French: "
>>> def preprocess_function(examples):
... inputs = [prefix + example[source_lang] for example in examples["translation"]]
... targets = [example[target_lang] for example in examples["translation"]]
... model_inputs = tokenizer(inputs, text_target=targets, max_length=128, truncation=True)
... return model_inputs
```
To apply the preprocessing function over the entire dataset, use 🤗 Datasets [`~datasets.Dataset.map`] method. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once:
```py
>>> tokenized_books = books.map(preprocess_function, batched=True)
```
Now create a batch of examples using [`DataCollatorForSeq2Seq`]. It's more efficient to *dynamically pad* the sentences to the longest length in a batch during collation, instead of padding the whole dataset to the maximum length.
<frameworkcontent>
<pt>
```py
>>> from transformers import DataCollatorForSeq2Seq
>>> data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, model=checkpoint)
```
</pt>
<tf>
```py
>>> from transformers import DataCollatorForSeq2Seq
>>> data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, model=checkpoint, return_tensors="tf")
```
</tf>
</frameworkcontent>
## Evaluate
Including a metric during training is often helpful for evaluating your model's performance. You can quickly load a evaluation method with the 🤗 [Evaluate](https://huggingface.co/docs/evaluate/index) library. For this task, load the [SacreBLEU](https://huggingface.co/spaces/evaluate-metric/sacrebleu) metric (see the 🤗 Evaluate [quick tour](https://huggingface.co/docs/evaluate/a_quick_tour) to learn more about how to load and compute a metric):
```py
>>> import evaluate
>>> metric = evaluate.load("sacrebleu")
```
Then create a function that passes your predictions and labels to [`~evaluate.EvaluationModule.compute`] to calculate the SacreBLEU score:
```py
>>> import numpy as np
>>> def postprocess_text(preds, labels):
... preds = [pred.strip() for pred in preds]
... labels = [[label.strip()] for label in labels]
... return preds, labels
>>> def compute_metrics(eval_preds):
... preds, labels = eval_preds
... if isinstance(preds, tuple):
... preds = preds[0]
... decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
... labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
... decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
... decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)
... result = metric.compute(predictions=decoded_preds, references=decoded_labels)
... result = {"bleu": result["score"]}
... prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
... result["gen_len"] = np.mean(prediction_lens)
... result = {k: round(v, 4) for k, v in result.items()}
... return result
```
Your `compute_metrics` function is ready to go now, and you'll return to it when you setup your training.
## Train
<frameworkcontent>
<pt>
<Tip>
If you aren't familiar with finetuning a model with the [`Trainer`], take a look at the basic tutorial [here](../training#train-with-pytorch-trainer)!
</Tip>
You're ready to start training your model now! Load T5 with [`AutoModelForSeq2SeqLM`]:
```py
>>> from transformers import AutoModelForSeq2SeqLM, Seq2SeqTrainingArguments, Seq2SeqTrainer
>>> model = AutoModelForSeq2SeqLM.from_pretrained(checkpoint)
```
At this point, only three steps remain:
1. Define your training hyperparameters in [`Seq2SeqTrainingArguments`]. The only required parameter is `output_dir` which specifies where to save your model. You'll push this model to the Hub by setting `push_to_hub=True` (you need to be signed in to Hugging Face to upload your model). At the end of each epoch, the [`Trainer`] will evaluate the SacreBLEU metric and save the training checkpoint.
2. Pass the training arguments to [`Seq2SeqTrainer`] along with the model, dataset, tokenizer, data collator, and `compute_metrics` function.
3. Call [`~Trainer.train`] to finetune your model.
```py
>>> training_args = Seq2SeqTrainingArguments(
... output_dir="my_awesome_opus_books_model",
... eval_strategy="epoch",
... learning_rate=2e-5,
... per_device_train_batch_size=16,
... per_device_eval_batch_size=16,
... weight_decay=0.01,
... save_total_limit=3,
... num_train_epochs=2,
... predict_with_generate=True,
... fp16=True, #change to bf16=True for XPU
... push_to_hub=True,
... )
>>> trainer = Seq2SeqTrainer(
... model=model,
... args=training_args,
... train_dataset=tokenized_books["train"],
... eval_dataset=tokenized_books["test"],
... processing_class=tokenizer,
... data_collator=data_collator,
... compute_metrics=compute_metrics,
... )
>>> trainer.train()
```
Once training is completed, share your model to the Hub with the [`~transformers.Trainer.push_to_hub`] method so everyone can use your model:
```py
>>> trainer.push_to_hub()
```
</pt>
<tf>
<Tip>
If you aren't familiar with finetuning a model with Keras, take a look at the basic tutorial [here](../training#train-a-tensorflow-model-with-keras)!
</Tip>
To finetune a model in TensorFlow, start by setting up an optimizer function, learning rate schedule, and some training hyperparameters:
```py
>>> from transformers import AdamWeightDecay
>>> optimizer = AdamWeightDecay(learning_rate=2e-5, weight_decay_rate=0.01)
```
Then you can load T5 with [`TFAutoModelForSeq2SeqLM`]:
```py
>>> from transformers import TFAutoModelForSeq2SeqLM
>>> model = TFAutoModelForSeq2SeqLM.from_pretrained(checkpoint)
```
Convert your datasets to the `tf.data.Dataset` format with [`~transformers.TFPreTrainedModel.prepare_tf_dataset`]:
```py
>>> tf_train_set = model.prepare_tf_dataset(
... tokenized_books["train"],
... shuffle=True,
... batch_size=16,
... collate_fn=data_collator,
... )
>>> tf_test_set = model.prepare_tf_dataset(
... tokenized_books["test"],
... shuffle=False,
... batch_size=16,
... collate_fn=data_collator,
... )
```
Configure the model for training with [`compile`](https://keras.io/api/models/model_training_apis/#compile-method). Note that Transformers models all have a default task-relevant loss function, so you don't need to specify one unless you want to:
```py
>>> import tensorflow as tf
>>> model.compile(optimizer=optimizer) # No loss argument!
```
The last two things to setup before you start training is to compute the SacreBLEU metric from the predictions, and provide a way to push your model to the Hub. Both are done by using [Keras callbacks](../main_classes/keras_callbacks).
Pass your `compute_metrics` function to [`~transformers.KerasMetricCallback`]:
```py
>>> from transformers.keras_callbacks import KerasMetricCallback
>>> metric_callback = KerasMetricCallback(metric_fn=compute_metrics, eval_dataset=tf_test_set)
```
Specify where to push your model and tokenizer in the [`~transformers.PushToHubCallback`]:
```py
>>> from transformers.keras_callbacks import PushToHubCallback
>>> push_to_hub_callback = PushToHubCallback(
... output_dir="my_awesome_opus_books_model",
... tokenizer=tokenizer,
... )
```
Then bundle your callbacks together:
```py
>>> callbacks = [metric_callback, push_to_hub_callback]
```
Finally, you're ready to start training your model! Call [`fit`](https://keras.io/api/models/model_training_apis/#fit-method) with your training and validation datasets, the number of epochs, and your callbacks to finetune the model:
```py
>>> model.fit(x=tf_train_set, validation_data=tf_test_set, epochs=3, callbacks=callbacks)
```
Once training is completed, your model is automatically uploaded to the Hub so everyone can use it!
</tf>
</frameworkcontent>
<Tip>
For a more in-depth example of how to finetune a model for translation, take a look at the corresponding
[PyTorch notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/translation.ipynb)
or [TensorFlow notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/translation-tf.ipynb).
</Tip>
## Inference
Great, now that you've finetuned a model, you can use it for inference!
Come up with some text you'd like to translate to another language. For T5, you need to prefix your input depending on the task you're working on. For translation from English to French, you should prefix your input as shown below:
```py
>>> text = "translate English to French: Legumes share resources with nitrogen-fixing bacteria."
```
The simplest way to try out your finetuned model for inference is to use it in a [`pipeline`]. Instantiate a `pipeline` for translation with your model, and pass your text to it:
```py
>>> from transformers import pipeline
# Change `xx` to the language of the input and `yy` to the language of the desired output.
# Examples: "en" for English, "fr" for French, "de" for German, "es" for Spanish, "zh" for Chinese, etc; translation_en_to_fr translates English to French
# You can view all the lists of languages here - https://huggingface.co/languages
>>> translator = pipeline("translation_xx_to_yy", model="username/my_awesome_opus_books_model")
>>> translator(text)
[{'translation_text': 'Legumes partagent des ressources avec des bactéries azotantes.'}]
```
You can also manually replicate the results of the `pipeline` if you'd like:
<frameworkcontent>
<pt>
Tokenize the text and return the `input_ids` as PyTorch tensors:
```py
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("username/my_awesome_opus_books_model")
>>> inputs = tokenizer(text, return_tensors="pt").input_ids
```
Use the [`~generation.GenerationMixin.generate`] method to create the translation. For more details about the different text generation strategies and parameters for controlling generation, check out the [Text Generation](../main_classes/text_generation) API.
```py
>>> from transformers import AutoModelForSeq2SeqLM
>>> model = AutoModelForSeq2SeqLM.from_pretrained("username/my_awesome_opus_books_model")
>>> outputs = model.generate(inputs, max_new_tokens=40, do_sample=True, top_k=30, top_p=0.95)
```
Decode the generated token ids back into text:
```py
>>> tokenizer.decode(outputs[0], skip_special_tokens=True)
'Les lignées partagent des ressources avec des bactéries enfixant l'azote.'
```
</pt>
<tf>
Tokenize the text and return the `input_ids` as TensorFlow tensors:
```py
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("username/my_awesome_opus_books_model")
>>> inputs = tokenizer(text, return_tensors="tf").input_ids
```
Use the [`~transformers.generation_tf_utils.TFGenerationMixin.generate`] method to create the translation. For more details about the different text generation strategies and parameters for controlling generation, check out the [Text Generation](../main_classes/text_generation) API.
```py
>>> from transformers import TFAutoModelForSeq2SeqLM
>>> model = TFAutoModelForSeq2SeqLM.from_pretrained("username/my_awesome_opus_books_model")
>>> outputs = model.generate(inputs, max_new_tokens=40, do_sample=True, top_k=30, top_p=0.95)
```
Decode the generated token ids back into text:
```py
>>> tokenizer.decode(outputs[0], skip_special_tokens=True)
'Les lugumes partagent les ressources avec des bactéries fixatrices d'azote.'
```
</tf>
</frameworkcontent>
| transformers/docs/source/en/tasks/translation.md/0 | {
"file_path": "transformers/docs/source/en/tasks/translation.md",
"repo_id": "transformers",
"token_count": 4917
} | 377 |
<!---
Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Troubleshoot
Sometimes errors occur, but we are here to help! This guide covers some of the most common issues we've seen and how you can resolve them. However, this guide isn't meant to be a comprehensive collection of every 🤗 Transformers issue. For more help with troubleshooting your issue, try:
<Youtube id="S2EEG3JIt2A"/>
1. Asking for help on the [forums](https://discuss.huggingface.co/). There are specific categories you can post your question to, like [Beginners](https://discuss.huggingface.co/c/beginners/5) or [🤗 Transformers](https://discuss.huggingface.co/c/transformers/9). Make sure you write a good descriptive forum post with some reproducible code to maximize the likelihood that your problem is solved!
<Youtube id="_PAli-V4wj0"/>
2. Create an [Issue](https://github.com/huggingface/transformers/issues/new/choose) on the 🤗 Transformers repository if it is a bug related to the library. Try to include as much information describing the bug as possible to help us better figure out what's wrong and how we can fix it.
3. Check the [Migration](migration) guide if you use an older version of 🤗 Transformers since some important changes have been introduced between versions.
For more details about troubleshooting and getting help, take a look at [Chapter 8](https://huggingface.co/course/chapter8/1?fw=pt) of the Hugging Face course.
## Firewalled environments
Some GPU instances on cloud and intranet setups are firewalled to external connections, resulting in a connection error. When your script attempts to download model weights or datasets, the download will hang and then timeout with the following message:
```
ValueError: Connection error, and we cannot find the requested files in the cached path.
Please try again or make sure your Internet connection is on.
```
In this case, you should try to run 🤗 Transformers on [offline mode](installation#offline-mode) to avoid the connection error.
## CUDA out of memory
Training large models with millions of parameters can be challenging without the appropriate hardware. A common error you may encounter when the GPU runs out of memory is:
```
CUDA out of memory. Tried to allocate 256.00 MiB (GPU 0; 11.17 GiB total capacity; 9.70 GiB already allocated; 179.81 MiB free; 9.85 GiB reserved in total by PyTorch)
```
Here are some potential solutions you can try to lessen memory use:
- Reduce the [`per_device_train_batch_size`](main_classes/trainer#transformers.TrainingArguments.per_device_train_batch_size) value in [`TrainingArguments`].
- Try using [`gradient_accumulation_steps`](main_classes/trainer#transformers.TrainingArguments.gradient_accumulation_steps) in [`TrainingArguments`] to effectively increase overall batch size.
<Tip>
Refer to the Performance [guide](performance) for more details about memory-saving techniques.
</Tip>
## ImportError
Another common error you may encounter, especially if it is a newly released model, is `ImportError`:
```
ImportError: cannot import name 'ImageGPTImageProcessor' from 'transformers' (unknown location)
```
For these error types, check to make sure you have the latest version of 🤗 Transformers installed to access the most recent models:
```bash
pip install transformers --upgrade
```
## CUDA error: device-side assert triggered
Sometimes you may run into a generic CUDA error about an error in the device code.
```
RuntimeError: CUDA error: device-side assert triggered
```
You should try to run the code on a CPU first to get a more descriptive error message. Add the following environment variable to the beginning of your code to switch to a CPU:
```py
>>> import os
>>> os.environ["CUDA_VISIBLE_DEVICES"] = ""
```
Another option is to get a better traceback from the GPU. Add the following environment variable to the beginning of your code to get the traceback to point to the source of the error:
```py
>>> import os
>>> os.environ["CUDA_LAUNCH_BLOCKING"] = "1"
```
## Incorrect output when padding tokens aren't masked
In some cases, the output `hidden_state` may be incorrect if the `input_ids` include padding tokens. To demonstrate, load a model and tokenizer. You can access a model's `pad_token_id` to see its value. The `pad_token_id` may be `None` for some models, but you can always manually set it.
```py
>>> from transformers import AutoModelForSequenceClassification
>>> import torch
>>> model = AutoModelForSequenceClassification.from_pretrained("google-bert/bert-base-uncased")
>>> model.config.pad_token_id
0
```
The following example shows the output without masking the padding tokens:
```py
>>> input_ids = torch.tensor([[7592, 2057, 2097, 2393, 9611, 2115], [7592, 0, 0, 0, 0, 0]])
>>> output = model(input_ids)
>>> print(output.logits)
tensor([[ 0.0082, -0.2307],
[ 0.1317, -0.1683]], grad_fn=<AddmmBackward0>)
```
Here is the actual output of the second sequence:
```py
>>> input_ids = torch.tensor([[7592]])
>>> output = model(input_ids)
>>> print(output.logits)
tensor([[-0.1008, -0.4061]], grad_fn=<AddmmBackward0>)
```
Most of the time, you should provide an `attention_mask` to your model to ignore the padding tokens to avoid this silent error. Now the output of the second sequence matches its actual output:
<Tip>
By default, the tokenizer creates an `attention_mask` for you based on your specific tokenizer's defaults.
</Tip>
```py
>>> attention_mask = torch.tensor([[1, 1, 1, 1, 1, 1], [1, 0, 0, 0, 0, 0]])
>>> output = model(input_ids, attention_mask=attention_mask)
>>> print(output.logits)
tensor([[ 0.0082, -0.2307],
[-0.1008, -0.4061]], grad_fn=<AddmmBackward0>)
```
🤗 Transformers doesn't automatically create an `attention_mask` to mask a padding token if it is provided because:
- Some models don't have a padding token.
- For some use-cases, users want a model to attend to a padding token.
## ValueError: Unrecognized configuration class XYZ for this kind of AutoModel
Generally, we recommend using the [`AutoModel`] class to load pretrained instances of models. This class
can automatically infer and load the correct architecture from a given checkpoint based on the configuration. If you see
this `ValueError` when loading a model from a checkpoint, this means the Auto class couldn't find a mapping from
the configuration in the given checkpoint to the kind of model you are trying to load. Most commonly, this happens when a
checkpoint doesn't support a given task.
For instance, you'll see this error in the following example because there is no GPT2 for question answering:
```py
>>> from transformers import AutoProcessor, AutoModelForQuestionAnswering
>>> processor = AutoProcessor.from_pretrained("openai-community/gpt2-medium")
>>> model = AutoModelForQuestionAnswering.from_pretrained("openai-community/gpt2-medium")
ValueError: Unrecognized configuration class <class 'transformers.models.gpt2.configuration_gpt2.GPT2Config'> for this kind of AutoModel: AutoModelForQuestionAnswering.
Model type should be one of AlbertConfig, BartConfig, BertConfig, BigBirdConfig, BigBirdPegasusConfig, BloomConfig, ...
```
| transformers/docs/source/en/troubleshooting.md/0 | {
"file_path": "transformers/docs/source/en/troubleshooting.md",
"repo_id": "transformers",
"token_count": 2193
} | 378 |
<!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Glosario
Este glosario define términos generales de aprendizaje automático y términos relacionados con 🤗 Transformers para ayudarte a comprender mejor la documentación.
## A
### attention mask
La máscara de atención es un argumento opcional utilizado al agrupar secuencias.
<Youtube id="M6adb1j2jPI"/>
Este argumento indica al modelo qué tokens deben recibir atención y cuáles no.
Por ejemplo, considera estas dos secuencias:
```python
>>> from transformers import BertTokenizer
>>> tokenizer = BertTokenizer.from_pretrained("google-bert/bert-base-cased")
>>> sequence_a = "This is a short sequence."
>>> sequence_b = "This is a rather long sequence. It is at least longer than the sequence A."
>>> encoded_sequence_a = tokenizer(sequence_a)["input_ids"]
>>> encoded_sequence_b = tokenizer(sequence_b)["input_ids"]
```
Las versiones codificadas tienen longitudes diferentes:
```python
>>> len(encoded_sequence_a), len(encoded_sequence_b)
(8, 19)
```
Por lo tanto, no podemos colocarlas juntas en el mismo tensor tal cual. La primera secuencia necesita ser rellenada hasta la longitud de la segunda, o la segunda necesita ser truncada hasta la longitud de la primera.
En el primer caso, la lista de IDs se extenderá con los índices de relleno. Podemos pasar una lista al tokenizador y pedirle que realice el relleno de esta manera:
```python
>>> padded_sequences = tokenizer([sequence_a, sequence_b], padding=True)
```
Podemos ver que se han agregado ceros a la derecha de la primera oración para que tenga la misma longitud que la segunda:
```python
>>> padded_sequences["input_ids"]
[[101, 1188, 1110, 170, 1603, 4954, 119, 102, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [101, 1188, 1110, 170, 1897, 1263, 4954, 119, 1135, 1110, 1120, 1655, 2039, 1190, 1103, 4954, 138, 119, 102]]
```
Esto luego se puede convertir en un tensor en PyTorch o TensorFlow. La máscara de atención es un tensor binario que indica la posición de los índices de relleno para que el modelo no los tenga en cuenta. Para el [`BertTokenizer`], `1` indica un valor al que se debe prestar atención, mientras que `0` indica un valor de relleno. Esta máscara de atención está en el diccionario devuelto por el tokenizador bajo la clave "attention_mask":
```python
>>> padded_sequences["attention_mask"]
[[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]
```
### autoencoding models
Consulta [modelos de codificación](#encoder-models) y [modelado de lenguaje enmascarado](#masked-language-modeling-mlm)
### autoregressive models
Consulta [modelado de lenguaje causal](#causal-language-modeling) y [modelos de decodificación](#decoder-models)
## B
### backbone
La columna vertebral, backbone en inglés, es la red (embeddings y layers) que produce los estados ocultos o características crudas. Normalmente, está conectado a una [cabecera](#head), que acepta las características como entrada para hacer una predicción. Por ejemplo, [`ViTModel`] es una columna vertebral sin una cabecera específica encima. Otros modelos también pueden usar [`VitModel`] como columna vertebral, como por ejemplo [DPT](model_doc/dpt).
## C
### causal language modeling
Una tarea de preentrenamiento donde el modelo lee los textos en orden y tiene que predecir la siguiente palabra. Generalmente, se realiza leyendo toda la oración, pero utilizando una máscara dentro del modelo para ocultar los tokens futuros en un cierto paso de tiempo.
### channel
Las imágenes a color están compuestas por alguna combinación de valores en tres canales: rojo, verde y azul (RGB), y las imágenes en escala de grises solo tienen un canal. En 🤗 Transformers, el canal puede ser la primera o última dimensión del tensor de una imagen: [`n_channels`, `height`, `width`] o [`height`, `width`, `n_channels`].
### connectionist temporal classification (CTC)
Un algoritmo que permite que un modelo aprenda sin saber exactamente cómo están alineadas la entrada y la salida; CTC calcula la distribución de todas las salidas posibles para una entrada dada y elige la salida más probable de ella. CTC se utiliza comúnmente en tareas de reconocimiento de voz porque el habla no siempre se alinea perfectamente con la transcripción debido a diversas razones, como las diferentes velocidades de habla de los oradores.
### convolution
Un tipo de capa en una red neuronal donde la matriz de entrada se multiplica elemento por elemento por una matriz más pequeña (núcleo o filtro) y los valores se suman en una nueva matriz. Esto se conoce como una operación de convolución que se repite sobre toda la matriz de entrada. Cada operación se aplica a un segmento diferente de la matriz de entrada. Las redes neuronales convolucionales (CNN) se utilizan comúnmente en visión por computadora.
## D
### DataParallel (DP)
Técnica de paralelismo para entrenamiento en múltiples GPUs donde se replica la misma configuración varias veces, con cada instancia recibiendo una porción de datos única. El procesamiento se realiza en paralelo y todas las configuraciones se sincronizan al final de cada paso de entrenamiento.
Obtén más información sobre cómo funciona el DataParallel [aquí](perf_train_gpu_many#dataparallel-vs-distributeddataparallel).
### decoder input IDs
Esta entrada es específica para modelos codificador-decodificador y contiene los IDs de entrada que se enviarán al decodificador. Estas entradas deben usarse para tareas de secuencia a secuencia, como traducción o resumen, y generalmente se construyen de una manera específica para cada modelo.
La mayoría de los modelos codificador-decodificador (BART, T5) crean sus `decoder_input_ids` por sí mismos a partir de las `labels`. En tales modelos, pasar las `labels` es la forma preferida de manejar el entrenamiento.
Consulta la documentación de cada modelo para ver cómo manejan estos IDs de entrada para el entrenamiento de secuencia a secuencia.
### decoder models
También conocidos como modelos autorregresivos, los modelos decodificadores involucran una tarea de preentrenamiento (llamada modelado de lenguaje causal) donde el modelo lee los textos en orden y tiene que predecir la siguiente palabra. Generalmente, se realiza leyendo la oración completa con una máscara para ocultar los tokens futuros en un cierto paso de tiempo.
<Youtube id="d_ixlCubqQw"/>
### deep learning (DL)
Algoritmos de aprendizaje automático que utilizan redes neuronales con varias capas.
## E
### encoder models
También conocidos como modelos de codificación automática (autoencoding models), los modelos codificadores toman una entrada (como texto o imágenes) y las transforman en una representación numérica condensada llamada embedding. A menudo, los modelos codificadores se entrenan previamente utilizando técnicas como el [modelado de lenguaje enmascarado](#masked-language-modeling-mlm), que enmascara partes de la secuencia de entrada y obliga al modelo a crear representaciones más significativas.
<Youtube id="H39Z_720T5s"/>
## F
### feature extraction
El proceso de seleccionar y transformar datos crudos en un conjunto de características más informativas y útiles para algoritmos de aprendizaje automático. Algunos ejemplos de extracción de características incluyen transformar texto crudo en embeddings de palabras y extraer características importantes como bordes o formas de datos de imágenes/videos.
### feed forward chunking
En cada bloque de atención residual en los transformadores, la capa de autoatención suele ir seguida de 2 capas de avance. El tamaño de embedding intermedio de las capas de avance suele ser mayor que el tamaño oculto del modelo (por ejemplo, para `google-bert/bert-base-uncased`).
Para una entrada de tamaño `[batch_size, sequence_length]`, la memoria requerida para almacenar los embeddings intermedios de avance `[batch_size, sequence_length, config.intermediate_size]` puede representar una gran fracción del uso de memoria. Los autores de [Reformer: The Efficient Transformer](https://huggingface.co/papers/2001.04451) observaron que, dado que el cálculo es independiente de la dimensión `sequence_length`, es matemáticamente equivalente calcular los embeddings de salida de ambas capas de avance `[batch_size, config.hidden_size]_0, ..., [batch_size, config.hidden_size]_n` individualmente y concatenarlos después a `[batch_size, sequence_length, config.hidden_size]` con `n = sequence_length`, lo que intercambia el aumento del tiempo de cálculo por una reducción en el uso de memoria, pero produce un resultado matemáticamente **equivalente**.
Para modelos que utilizan la función [`apply_chunking_to_forward`], el `chunk_size` define el número de embeddings de salida que se calculan en paralelo y, por lo tanto, define el equilibrio entre la complejidad de memoria y tiempo. Si `chunk_size` se establece en 0, no se realiza ninguna fragmentación de avance.
### finetuned models
El ajuste fino es una forma de transferencia de aprendizaje que implica tomar un modelo entrenado previamente, congelar sus pesos y reemplazar la capa de salida con una nueva [cabecera de modelo](#head) recién añadida. La cabecera del modelo se entrena en tu conjunto de datos objetivo.
Consulta el tutorial [Ajustar finamente un modelo pre-entrenado](https://huggingface.co/docs/transformers/training) para obtener más detalles y aprende cómo ajustar finamente modelos con 🤗 Transformers.
## H
### head
La cabecera del modelo se refiere a la última capa de una red neuronal que acepta los estados ocultos crudos y los proyecta en una dimensión diferente. Hay una cabecera de modelo diferente para cada tarea. Por ejemplo:
* [`GPT2ForSequenceClassification`] es una cabecera de clasificación de secuencias, es decir, una capa lineal, encima del modelo base [`GPT2Model`].
* [`ViTForImageClassification`] es una cabecera de clasificación de imágenes, es decir, una capa lineal encima del estado oculto final del token `CLS`, encima del modelo base [`ViTModel`].
* [`Wav2Vec2ForCTC`] es una cabecera de modelado de lenguaje con [CTC](#connectionist-temporal-classification-ctc) encima del modelo base [`Wav2Vec2Model`].
## I
### image patch
Los modelos de Transformers basados en visión dividen una imagen en parches más pequeños que se incorporan linealmente y luego se pasan como una secuencia al modelo. Puedes encontrar el `patch_size` (o resolución del modelo) en su configuración.
### inference
La inferencia es el proceso de evaluar un modelo en nuevos datos después de completar el entrenamiento. Consulta el tutorial [Pipeline for inference](https://huggingface.co/docs/transformers/pipeline_tutorial) para aprender cómo realizar inferencias con 🤗 Transformers.
### input IDs
Los IDs de entrada a menudo son los únicos parámetros necesarios que se deben pasar al modelo como entrada. Son índices de tokens, representaciones numéricas de tokens que construyen las secuencias que se utilizarán como entrada por el modelo.
<Youtube id="VFp38yj8h3A"/>
Cada tokenizador funciona de manera diferente, pero el mecanismo subyacente sigue siendo el mismo. Aquí tienes un ejemplo utilizando el tokenizador BERT, que es un tokenizador [WordPiece](https://huggingface.co/papers/1609.08144):
```python
>>> from transformers import BertTokenizer
>>> tokenizer = BertTokenizer.from_pretrained("google-bert/bert-base-cased")
>>> sequence = "A Titan RTX has 24GB of VRAM"
```
El tokenizador se encarga de dividir la secuencia en tokens disponibles en el vocabulario del tokenizador.
```python
>>> tokenized_sequence = tokenizer.tokenize(sequence)
```
Los tokens son palabras o sub palabras. Por ejemplo, "VRAM" no estaba en el vocabulario del modelo, así que se dividió
en "V", "RA" y "M". Para indicar que estos tokens no son palabras separadas sino partes de la misma palabra, se añade un prefijo de doble almohadilla para "RA" y "M":
```python
>>> print(tokenized_sequence)
['A', 'Titan', 'R', '##T', '##X', 'has', '24', '##GB', 'of', 'V', '##RA', '##M']
```
Estos tokens luego se pueden convertir en IDs que son comprensibles por el modelo. Esto se puede hacer alimentando directamente la oración al tokenizador, que aprovecha la implementación en Rust de [🤗 Tokenizers](https://github.com/huggingface/tokenizers) para obtener un rendimiento óptimo.
```python
>>> inputs = tokenizer(sequence)
```
El tokenizador devuelve un diccionario con todos los argumentos necesarios para que su modelo correspondiente funcione correctamente. Los índices de los tokens están bajo la clave `input_ids`:
```python
>>> encoded_sequence = inputs["input_ids"]
>>> print(encoded_sequence)
[101, 138, 18696, 155, 1942, 3190, 1144, 1572, 13745, 1104, 159, 9664, 2107, 102]
```
Ten en cuenta que el tokenizador añade automáticamente "tokens especiales" (si el modelo asociado depende de ellos), que son IDs especiales que el modelo utiliza en ocasiones.
Si descodificamos la secuencia anterior de IDs,
```python
>>> decoded_sequence = tokenizer.decode(encoded_sequence)
```
Veremos
```python
>>> print(decoded_sequence)
[CLS] A Titan RTX has 24GB of VRAM [SEP]
```
Porque esta es la forma en que un [`BertModel`] espera sus entradas.
## L
### labels
Las etiquetas son un argumento opcional que se puede pasar para que el modelo calcule la pérdida por sí mismo. Estas etiquetas deberían ser la predicción esperada del modelo: usará la pérdida estándar para calcular la pérdida entre sus
predicciones y el valor esperado (la etiqueta).
Estas etiquetas son diferentes según la cabecera del modelo, por ejemplo:
- Para modelos de clasificación de secuencias ([`BertForSequenceClassification`]), el modelo espera un tensor de dimensión
`(batch_size)` con cada valor del lote correspondiente a la etiqueta esperada de toda la secuencia.
- Para modelos de clasificación de tokens ([`BertForTokenClassification`]), el modelo espera un tensor de dimensión
`(batch_size, seq_length)` con cada valor correspondiente a la etiqueta esperada de cada token individual.
- Para el modelado de lenguaje enmascarado ([`BertForMaskedLM`]), el modelo espera un tensor de dimensión `(batch_size, seq_length)` con cada valor correspondiente a la etiqueta esperada de cada token individual: las etiquetas son el ID del token enmascarado y los valores deben ignorarse para el resto (generalmente -100).
- Para tareas de secuencia a secuencia ([`BartForConditionalGeneration`], [`MBartForConditionalGeneration`]), el modelo
espera un tensor de dimensión `(batch_size, tgt_seq_length)` con cada valor correspondiente a las secuencias objetivo asociadas con cada secuencia de entrada. Durante el entrenamiento, tanto BART como T5 generarán internamente los `decoder_input_ids` y las máscaras de atención del decodificador. Por lo general, no es necesario suministrarlos. Esto no se aplica a los modelos que aprovechan el marco codificador-decodificador.
- Para modelos de clasificación de imágenes ([`ViTForImageClassification`]), el modelo espera un tensor de dimensión
`(batch_size)` con cada valor del lote correspondiente a la etiqueta esperada de cada imagen individual.
- Para modelos de segmentación semántica ([`SegformerForSemanticSegmentation`]), el modelo espera un tensor de dimensión
`(batch_size, height, width)` con cada valor del lote correspondiente a la etiqueta esperada de cada píxel individual.
- Para modelos de detección de objetos ([`DetrForObjectDetection`]), el modelo espera una lista de diccionarios con claves `class_labels` y `boxes` donde cada valor del lote corresponde a la etiqueta esperada y el número de cajas delimitadoras de cada imagen individual.
- Para modelos de reconocimiento automático de voz ([`Wav2Vec2ForCTC`]), el modelo espera un tensor de dimensión `(batch_size, target_length)` con cada valor correspondiente a la etiqueta esperada de cada token individual.
<Tip>
Las etiquetas de cada modelo pueden ser diferentes, así que asegúrate siempre de revisar la documentación de cada modelo para obtener más información sobre sus etiquetas específicas.
</Tip>
Los modelos base ([`BertModel`]) no aceptan etiquetas, ya que estos son los modelos base de transformadores, que simplemente generan características.
### large language models (LLM)
Un término genérico que se refiere a modelos de lenguaje de transformadores (GPT-3, BLOOM, OPT) que fueron entrenados con una gran cantidad de datos. Estos modelos también tienden a tener un gran número de parámetros que se pueden aprender (por ejemplo, 175 mil millones para GPT-3).
## M
### masked language modeling (MLM)
Una tarea de preentrenamiento en la que el modelo ve una versión corrupta de los textos, generalmente hecha
al enmascarar algunos tokens al azar, y tiene que predecir el texto original.
### multimodal
Una tarea que combina textos con otro tipo de entradas (por ejemplo: imágenes).
## N
### Natural language generation (NLG)
Todas las tareas relacionadas con la generación de texto (por ejemplo: [Escribe con Transformers](https://transformer.huggingface.co/) o traducción).
### Natural language processing (NLP)
Una forma genérica de decir "trabajar con textos".
### Natural language understanding (NLU)
Todas las tareas relacionadas con entender lo que hay en un texto (por ejemplo: clasificar el
texto completo o palabras individuales).
## P
### Pipeline
Un pipeline en 🤗 Transformers es una abstracción que se refiere a una serie de pasos que se ejecutan en un orden específico para preprocesar y transformar datos y devolver una predicción de un modelo. Algunas etapas de ejemplo que se encuentran en un pipeline pueden ser el preprocesamiento de datos, la extracción de características y la normalización.
Para obtener más detalles, consulta [Pipelines para inferencia](https://huggingface.co/docs/transformers/pipeline_tutorial).
### PipelineParallel (PP)
Técnica de paralelismo en la que el modelo se divide verticalmente (a nivel de capa) en varios GPU, de modo que solo una o varias capas del modelo se colocan en un solo GPU. Cada GPU procesa en paralelo diferentes etapas del pipeline y trabaja en un pequeño fragmento del lote. Obtén más información sobre cómo funciona PipelineParallel [aquí](perf_train_gpu_many#from-naive-model-parallelism-to-pipeline-parallelism).
### pixel values
Un tensor de las representaciones numéricas de una imagen que se pasa a un modelo. Los valores de píxeles tienen una forma de [`batch_size`, `num_channels`, `height`, `width`], y se generan a partir de un procesador de imágenes.
### pooling
Una operación que reduce una matriz a una matriz más pequeña, ya sea tomando el máximo o el promedio de la dimensión (o dimensiones) agrupada(s). Las capas de agrupación se encuentran comúnmente entre capas convolucionales para reducir la representación de características.
### position IDs
A diferencia de las RNN que tienen la posición de cada token incrustada en ellas, los transformers no son conscientes de la posición de cada token. Por lo tanto, se utilizan los IDs de posición (`position_ids`) para que el modelo identifique la posición de cada token en la lista de tokens.
Son un parámetro opcional. Si no se pasan `position_ids` al modelo, los IDs se crean automáticamente como embeddings de posición absolutas.
Los embeddings de posición absolutas se seleccionan en el rango `[0, config.max_position_embeddings - 1]`. Algunos modelos utilizan otros tipos de embeddings de posición, como embeddings de posición sinusoidales o embeddings de posición relativas.
### preprocessing
La tarea de preparar datos crudos en un formato que pueda ser fácilmente consumido por modelos de aprendizaje automático. Por ejemplo, el texto se preprocesa típicamente mediante la tokenización. Para tener una mejor idea de cómo es el preprocesamiento para otros tipos de entrada, consulta el tutorial [Pre-procesar](https://huggingface.co/docs/transformers/preprocessing).
### pretrained model
Un modelo que ha sido pre-entrenado en algunos datos (por ejemplo, toda Wikipedia). Los métodos de preentrenamiento involucran un objetivo auto-supervisado, que puede ser leer el texto e intentar predecir la siguiente palabra (ver [modelado de lenguaje causal](#causal-language-modeling)) o enmascarar algunas palabras e intentar predecirlas (ver [modelado de lenguaje enmascarado](#masked-language-modeling-mlm)).
Los modelos de habla y visión tienen sus propios objetivos de pre-entrenamiento. Por ejemplo, Wav2Vec2 es un modelo de habla pre-entrenado en una tarea contrastiva que requiere que el modelo identifique la representación de habla "verdadera" de un conjunto de representaciones de habla "falsas". Por otro lado, BEiT es un modelo de visión pre-entrenado en una tarea de modelado de imágenes enmascaradas que enmascara algunos de los parches de la imagen y requiere que el modelo prediga los parches enmascarados (similar al objetivo de modelado de lenguaje enmascarado).
## R
### recurrent neural network (RNN)
Un tipo de modelo que utiliza un bucle sobre una capa para procesar textos.
### representation learning
Un subcampo del aprendizaje automático que se centra en aprender representaciones significativas de datos en bruto. Algunos ejemplos de técnicas de aprendizaje de representaciones incluyen embeddings de palabras, auto-encoders y Redes Generativas Adversarias (Generative Adversarial Networks, GANs).
## S
### sampling rate
Una medida en hercios del número de muestras (la señal de audio) tomadas por segundo. La tasa de muestreo es el resultado de aproximar una señal continua como el habla.
### self-attention
Cada elemento de la entrada averigua a cuáles otros elementos de la entrada debe prestar atención.
### self-supervised learning
Una categoría de técnicas de aprendizaje automático en la que un modelo crea su propio objetivo de aprendizaje a partir de datos no etiquetados. Difiere del [aprendizaje no supervisado](#unsupervised-learning) y del [aprendizaje supervisado](#supervised-learning) en que el proceso de aprendizaje está supervisado, pero no explícitamente por el usuario.
Un ejemplo de aprendizaje auto-supervisado es el [modelado de lenguaje enmascarado](#masked-language-modeling-mlm), donde un modelo recibe oraciones con una proporción de sus tokens eliminados y aprende a predecir los tokens faltantes.
### semi-supervised learning
Una amplia categoría de técnicas de entrenamiento de aprendizaje automático que aprovecha una pequeña cantidad de datos etiquetados con una mayor cantidad de datos no etiquetados para mejorar la precisión de un modelo, a diferencia del [aprendizaje supervisado](#supervised-learning) y del [aprendizaje no supervisado](#unsupervised-learning).
Un ejemplo de un enfoque de aprendizaje semi-supervisado es "auto-entrenamiento", en el que un modelo se entrena con datos etiquetados y luego se utiliza para hacer predicciones sobre los datos no etiquetados. La porción de datos no etiquetados que el modelo predice con mayor confianza se agrega al conjunto de datos etiquetados y se utiliza para volver a entrenar el modelo.
### sequence-to-sequence (seq2seq)
Modelos que generan una nueva secuencia a partir de una entrada, como modelos de traducción o modelos de resumen (como
[Bart](model_doc/bart) o [T5](model_doc/t5)).
### Sharded DDP
Otro nombre para el concepto fundamental de [ZeRO](#zero-redundancy-optimizer-zero) utilizado por varias otras implementaciones de ZeRO.
### stride
En [convolución](#convolution) o [agrupación](#pooling), el paso (stride) se refiere a la distancia que recorre el núcleo sobre una matriz. Un paso de 1 significa que el núcleo se mueve un píxel a la vez, y un paso de 2 significa que el núcleo se mueve dos píxeles a la vez.
### supervised learning
Una forma de entrenamiento de modelos que utiliza directamente datos etiquetados para corregir y dirigir el rendimiento del modelo. Los datos se introducen en el modelo en entrenamiento, y sus predicciones se comparan con las etiquetas conocidas. El modelo actualiza sus pesos en función de cuán incorrectas fueron sus predicciones, y el proceso se repite para optimizar el rendimiento del modelo.
## T
### Tensor Parallelism (TP)
Técnica de paralelismo para entrenamiento en múltiples GPU en la que cada tensor se divide en múltiples fragmentos, de modo que en lugar de tener todo el tensor en una sola GPU, cada fragmento del tensor reside en su GPU designada. Los fragmentos se procesan por separado y en paralelo en diferentes GPU y los resultados se sincronizan al final del paso de procesamiento.Esto es lo que a veces se llama paralelismo horizontal, ya que la división ocurre a nivel horizontal.
Obtén más información sobre el Paralelismo de Tensores [aquí](perf_train_gpu_many#tensor-parallelism).
### token
Parte de una oración, generalmente una palabra, pero también puede ser una sub-palabra (las palabras no comunes a menudo se dividen en sub-palabras) o un símbolo de puntuación.
### token Type IDs
Algunos modelos tienen como objetivo realizar clasificación en pares de oraciones o responder preguntas.
<Youtube id="0u3ioSwev3s"/>
Estos requieren que dos secuencias diferentes se unan en una única entrada "input_ids", lo cual generalmente se realiza con
la ayuda de tokens especiales, como el token de clasificación (`[CLS]`) y el token separador (`[SEP]`). Por ejemplo, el modelo BERT construye sus dos secuencias de entrada de la siguiente manera:
```python
>>> # [CLS] SEQUENCE_A [SEP] SEQUENCE_B [SEP]
```
Podemos utilizar nuestro tokenizador para generar automáticamente una oración de este tipo al pasar las dos secuencias a `tokenizer` como dos argumentos (y no como una lista, como antes) de la siguiente manera:
```python
>>> from transformers import BertTokenizer
>>> tokenizer = BertTokenizer.from_pretrained("google-bert/bert-base-cased")
>>> sequence_a = "HuggingFace is based in NYC"
>>> sequence_b = "Where is HuggingFace based?"
>>> encoded_dict = tokenizer(sequence_a, sequence_b)
>>> decoded = tokenizer.decode(encoded_dict["input_ids"])
```
Que devolverá:
```python
>>> print(decoded)
[CLS] HuggingFace is based in NYC [SEP] Where is HuggingFace based? [SEP]
```
Esto es suficiente para que algunos modelos comprendan dónde termina una secuencia y comienza otra. Sin embargo, otros modelos, como BERT, también utilizan identificadores de tipo de token (también llamados identificadores de segmento). Se representan como una máscara binaria que identifica los dos tipos de secuencia en el modelo.
El tokenizador devuelve esta máscara como la entrada "token_type_ids":
```python
>>> encoded_dict["token_type_ids"]
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1]
```
La primera secuencia, el "contexto" utilizado para la pregunta, tiene todos sus tokens representados por un `0`, mientras que la segunda secuencia, correspondiente a la "pregunta", tiene todos sus tokens representados por un `1`.
Algunos modelos, como [`XLNetModel`], utilizan un token adicional representado por un `2`.
### transfer learning
Una técnica que implica tomar un modelo pre-entrenado y adaptarlo a un conjunto de datos específico para tu tarea. En lugar de entrenar un modelo desde cero, puedes aprovechar el conocimiento obtenido de un modelo existente como punto de partida. Esto acelera el proceso de aprendizaje y reduce la cantidad de datos de entrenamiento necesarios.
### transformer
Arquitectura de modelo de aprendizaje profundo basada en auto-atención (Self-attention).
## U
### unsupervised learning
Una forma de entrenamiento de modelos en la que los datos proporcionados al modelo no están etiquetados. Las técnicas de aprendizaje no supervisado aprovechan la información estadística de la distribución de datos para encontrar patrones útiles para la tarea en cuestión.
## Z
### Zero Redundancy Optimizer (ZeRO)
Técnica de paralelismo que realiza la fragmentación de los tensores de manera algo similar a [TensorParallel](#tensor-parallelism-tp), excepto que todo el tensor se reconstruye a tiempo para una computación hacia adelante o hacia atrás, por lo tanto, el modelo no necesita ser modificado. Este método también admite diversas técnicas de descarga para compensar la memoria limitada de la GPU. Obtén más información sobre ZeRO [aquí](perf_train_gpu_many#zero-data-parallelism). | transformers/docs/source/es/glossary.md/0 | {
"file_path": "transformers/docs/source/es/glossary.md",
"repo_id": "transformers",
"token_count": 10085
} | 379 |
# docstyle-ignore
INSTALL_CONTENT = """
# Installation de Transformers
! pip install transformers datasets evaluate accelerate
# Pour installer à partir du code source au lieu de la dernière version, commentez la commande ci-dessus et décommentez la suivante.
# ! pip install git+https://github.com/huggingface/transformers.git
"""
notebook_first_cells = [{"type": "code", "content": INSTALL_CONTENT}]
black_avoid_patterns = {
"{processor_class}": "FakeProcessorClass",
"{model_class}": "FakeModelClass",
"{object_class}": "FakeObjectClass",
}
| transformers/docs/source/fr/_config.py/0 | {
"file_path": "transformers/docs/source/fr/_config.py",
"repo_id": "transformers",
"token_count": 175
} | 380 |
- sections:
- local: index
title: 🤗 Transformers
- local: quicktour
title: Tour rapido
- local: installation
title: Installazione
title: Iniziare
- sections:
- local: pipeline_tutorial
title: Pipeline per l'inferenza
- local: autoclass_tutorial
title: Carica istanze pre-allenate con AutoClass
- local: preprocessing
title: Preprocess
- local: training
title: Fine-tuning di un modello pre-addestrato
- local: accelerate
title: Allenamento distribuito con 🤗 Accelerate
- local: model_sharing
title: Condividere un modello
title: Esercitazione
- sections:
- local: create_a_model
title: Crea un'architettura personalizzata
- local: custom_models
title: Condividere modelli personalizzati
- local: run_scripts
title: Addestramento con script
- local: multilingual
title: Modelli multilingua per l'inferenza
- local: converting_tensorflow_models
title: Convertire modelli tensorflow
- local: serialization
title: Esporta modelli Transformers
- local: perf_train_cpu
title: Addestramento efficiente su CPU
- local: perf_train_cpu_many
title: Addestramento efficiente su multiple CPU
- local: perf_train_tpu
title: Addestramento su TPU
- local: perf_train_special
title: Addestramento su Hardware Specializzato
- local: perf_infer_cpu
title: Inferenza Efficiente su CPU
- local: perf_infer_gpu_one
title: Inferenza su una GPU
- local: perf_infer_gpu_many
title: Inferenza Efficiente su GPU Multiple
- local: perf_infer_special
title: Inferenza su Hardware Specializzato
- local: big_models
title: Istanziare un big model
- local: migration
title: Passaggio da pacchetti precedenti
- local: debugging
title: Debugging
title: Guide pratiche
- sections:
- local: add_new_pipeline
title: Come aggiungere una pipeline a 🤗 Transformers?
- local: add_new_model
title: Come aggiungere un modello a 🤗 Transformers?
- local: perf_hardware
title: Hardware ottimizzato per l'addestramento
- local: community
title: Risorse della comunità
- local: pr_checks
title: Controlli su una Pull Request
title: Guide How-to
| transformers/docs/source/it/_toctree.yml/0 | {
"file_path": "transformers/docs/source/it/_toctree.yml",
"repo_id": "transformers",
"token_count": 771
} | 381 |
<!---
Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Hardware ottimizzato per l'addestramento
L'hardware utilizzato per eseguire l'addestramento del modello e l'inferenza può avere un grande effetto sulle prestazioni. Per un analisi approfondita delle GPUs, assicurati di dare un'occhiata all'eccellente [blog post](https://timdettmers.com/2020/09/07/which-gpu-for-deep-learning/) di Tim Dettmer.
Diamo un'occhiata ad alcuni consigli pratici per la configurazione della GPU.
## GPU
Quando si addestrano modelli più grandi ci sono essenzialmente tre opzioni:
- GPUs piu' grandi
- Piu' GPUs
- Piu' CPU e piu' NVMe (scaricato da [DeepSpeed-Infinity](main_classes/deepspeed#nvme-support))
Iniziamo dal caso in cui ci sia una singola GPU.
### Potenza e Raffreddamento
Se hai acquistato una costosa GPU di fascia alta, assicurati di darle la potenza corretta e un raffreddamento sufficiente.
**Potenza**:
Alcune schede GPU consumer di fascia alta hanno 2 e talvolta 3 prese di alimentazione PCI-E a 8 pin. Assicurati di avere tanti cavi PCI-E a 8 pin indipendenti da 12 V collegati alla scheda quante sono le prese. Non utilizzare le 2 fessure a un'estremità dello stesso cavo (noto anche come cavo a spirale). Cioè se hai 2 prese sulla GPU, vuoi 2 cavi PCI-E a 8 pin che vanno dall'alimentatore alla scheda e non uno che abbia 2 connettori PCI-E a 8 pin alla fine! In caso contrario, non otterrai tutte le prestazioni ufficiali.
Ciascun cavo di alimentazione PCI-E a 8 pin deve essere collegato a una guida da 12 V sul lato dell'alimentatore e può fornire fino a 150 W di potenza.
Alcune altre schede possono utilizzare connettori PCI-E a 12 pin e questi possono fornire fino a 500-600 W di potenza.
Le schede di fascia bassa possono utilizzare connettori a 6 pin, che forniscono fino a 75 W di potenza.
Inoltre vuoi un alimentatore (PSU) di fascia alta che abbia una tensione stabile. Alcuni PSU di qualità inferiore potrebbero non fornire alla scheda la tensione stabile di cui ha bisogno per funzionare al massimo.
E ovviamente l'alimentatore deve avere abbastanza Watt inutilizzati per alimentare la scheda.
**Raffreddamento**:
Quando una GPU si surriscalda, inizierà a rallentare e non fornirà le prestazioni mssimali e potrebbe persino spegnersi se diventasse troppo calda.
È difficile dire l'esatta temperatura migliore a cui aspirare quando una GPU è molto caricata, ma probabilmente qualsiasi cosa al di sotto di +80°C va bene, ma più bassa è meglio - forse 70-75°C è un intervallo eccellente in cui trovarsi. È probabile che il rallentamento inizi a circa 84-90°C. Ma oltre alla limitazione delle prestazioni, una temperatura molto elevata prolungata è probabile che riduca la durata di una GPU.
Diamo quindi un'occhiata a uno degli aspetti più importanti quando si hanno più GPU: la connettività.
### Connettività multi-GPU
Se utilizzi più GPU, il modo in cui le schede sono interconnesse può avere un enorme impatto sul tempo totale di allenamento. Se le GPU si trovano sullo stesso nodo fisico, puoi eseguire:
```bash
nvidia-smi topo -m
```
e ti dirà come sono interconnesse le GPU. Su una macchina con doppia GPU e collegata a NVLink, molto probabilmente vedrai qualcosa del tipo:
```
GPU0 GPU1 CPU Affinity NUMA Affinity
GPU0 X NV2 0-23 N/A
GPU1 NV2 X 0-23 N/A
```
su una macchina diversa senza NVLink potremmo vedere:
```
GPU0 GPU1 CPU Affinity NUMA Affinity
GPU0 X PHB 0-11 N/A
GPU1 PHB X 0-11 N/A
```
Il rapporto include questa legenda:
```
X = Self
SYS = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
PHB = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
PXB = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
PIX = Connection traversing at most a single PCIe bridge
NV# = Connection traversing a bonded set of # NVLinks
```
Quindi il primo rapporto `NV2` ci dice che le GPU sono interconnesse con 2 NVLinks e nel secondo report `PHB` abbiamo una tipica configurazione PCIe+Bridge a livello di consumatore.
Controlla che tipo di connettività hai sulla tua configurazione. Alcuni di questi renderanno la comunicazione tra le carte più veloce (es. NVLink), altri più lenta (es. PHB).
A seconda del tipo di soluzione di scalabilità utilizzata, la velocità di connettività potrebbe avere un impatto maggiore o minore. Se le GPU devono sincronizzarsi raramente, come in DDP, l'impatto di una connessione più lenta sarà meno significativo. Se le GPU devono scambiarsi messaggi spesso, come in ZeRO-DP, una connettività più veloce diventa estremamente importante per ottenere un addestramento più veloce.
#### NVlink
[NVLink](https://en.wikipedia.org/wiki/NVLink) è un collegamento di comunicazione a corto raggio multilinea seriale basato su cavo sviluppato da Nvidia.
Ogni nuova generazione fornisce una larghezza di banda più veloce, ad es. ecco una citazione da [Nvidia Ampere GA102 GPU Architecture](https://www.nvidia.com/content/dam/en-zz/Solutions/geforce/ampere/pdf/NVIDIA-ampere-GA102-GPU-Architecture-Whitepaper-V1.pdf):
> Third-Generation NVLink®
> GA102 GPUs utilize NVIDIA’s third-generation NVLink interface, which includes four x4 links,
> with each link providing 14.0625 GB/sec bandwidth in each direction between two GPUs. Four
> links provide 56.25 GB/sec bandwidth in each direction, and 112.5 GB/sec total bandwidth
> between two GPUs. Two RTX 3090 GPUs can be connected together for SLI using NVLink.
> (Note that 3-Way and 4-Way SLI configurations are not supported.)
Quindi più `X` si ottiene nel rapporto di `NVX` nell'output di `nvidia-smi topo -m`, meglio è. La generazione dipenderà dall'architettura della tua GPU.
Confrontiamo l'esecuzione di un training del modello di linguaggio openai-community/gpt2 su un piccolo campione di wikitext
I risultati sono:
| NVlink | Time |
| ----- | ---: |
| Y | 101s |
| N | 131s |
Puoi vedere che NVLink completa l'addestramento circa il 23% più velocemente. Nel secondo benchmark utilizziamo `NCCL_P2P_DISABLE=1` per dire alle GPU di non utilizzare NVLink.
Ecco il codice benchmark completo e gli output:
```bash
# DDP w/ NVLink
rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 torchrun \
--nproc_per_node 2 examples/pytorch/language-modeling/run_clm.py --model_name_or_path openai-community/gpt2 \
--dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 --do_train \
--output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200
{'train_runtime': 101.9003, 'train_samples_per_second': 1.963, 'epoch': 0.69}
# DDP w/o NVLink
rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 NCCL_P2P_DISABLE=1 torchrun \
--nproc_per_node 2 examples/pytorch/language-modeling/run_clm.py --model_name_or_path openai-community/gpt2 \
--dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 --do_train
--output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200
{'train_runtime': 131.4367, 'train_samples_per_second': 1.522, 'epoch': 0.69}
```
Hardware: 2x TITAN RTX 24GB each + NVlink with 2 NVLinks (`NV2` in `nvidia-smi topo -m`)
Software: `pytorch-1.8-to-be` + `cuda-11.0` / `transformers==4.3.0.dev0` | transformers/docs/source/it/perf_hardware.md/0 | {
"file_path": "transformers/docs/source/it/perf_hardware.md",
"repo_id": "transformers",
"token_count": 3024
} | 382 |
- sections:
- local: index
title: 🤗 Transformers
- local: quicktour
title: クイックツアー
- local: installation
title: インストール
title: Get started
- sections:
- local: pipeline_tutorial
title: パイプラインを使用して推論を実行する
- local: autoclass_tutorial
title: AutoClass を使用して移植可能なコードを作成する
- local: preprocessing
title: データの前処理
- local: training
title: 事前トレーニングされたモデルを微調整する
- local: run_scripts
title: スクリプトを使用してトレーニングする
- local: accelerate
title: 🤗 Accelerate を使用して分散トレーニングをセットアップする
- local: peft
title: 🤗 PEFT を使用してアダプターをロードしてトレーニングする
- local: model_sharing
title: モデルを共有する
- local: llm_tutorial
title: LLM を使用した生成
title: Tutorials
- sections:
- isExpanded: false
sections:
- local: tasks/token_classification
title: トークンの分類
- local: tasks/question_answering
title: 質疑応答
- local: tasks/language_modeling
title: 因果言語モデリング
- local: tasks/masked_language_modeling
title: マスクされた言語モデリング
- local: tasks/translation
title: 翻訳
- local: tasks/summarization
title: 要約
- local: tasks/multiple_choice
title: 複数の選択肢
title: 自然言語処理
- isExpanded: false
sections:
- local: tasks/audio_classification
title: 音声の分類
- local: tasks/asr
title: 自動音声認識
title: オーディオ
- isExpanded: false
sections:
- local: tasks/image_classification
title: 画像分類
- local: tasks/semantic_segmentation
title: セマンティックセグメンテーション
- local: tasks/video_classification
title: ビデオの分類
- local: tasks/object_detection
title: 物体検出
- local: tasks/zero_shot_object_detection
title: ゼロショット物体検出
- local: tasks/zero_shot_image_classification
title: ゼロショット画像分類
- local: tasks/monocular_depth_estimation
title: 深さの推定
- local: tasks/image_to_image
title: 画像から画像へ
- local: tasks/knowledge_distillation_for_image_classification
title: コンピュータビジョンのための知識の蒸留
title: コンピュータビジョン
- isExpanded: false
sections:
- local: tasks/image_captioning
title: 画像のキャプション
- local: tasks/document_question_answering
title: 文書の質問への回答
- local: tasks/visual_question_answering
title: 視覚的な質問への回答
- local: tasks/text-to-speech
title: テキスト読み上げ
title: マルチモーダル
- isExpanded: false
sections:
- local: generation_strategies
title: 生成戦略をカスタマイズする
title: 世代
- isExpanded: false
sections:
- local: tasks/idefics
title: IDEFICS を使用したイメージ タスク
- local: tasks/prompting
title: LLM プロンプト ガイド
title: プロンプト
title: Task Guides
- sections:
- local: fast_tokenizers
title: 🤗 トークナイザーの高速トークナイザーを使用する
- local: multilingual
title: 多言語モデルで推論を実行する
- local: create_a_model
title: モデル固有の API を使用する
- local: custom_models
title: カスタムモデルを共有する
- local: chat_templating
title: チャットモデルのテンプレート
- local: serialization
title: ONNX へのエクスポート
- local: tflite
title: TFLite へのエクスポート
- local: torchscript
title: トーチスクリプトへのエクスポート
- local: community
title: コミュニティリソース
- local: troubleshooting
title: トラブルシューティング
title: 開発者ガイド
- sections:
- local: performance
title: 概要
- sections:
- local: perf_train_gpu_one
title: 単一の GPU で効率的にトレーニングするための方法とツール
- local: perf_train_gpu_many
title: 複数の GPU と並列処理
- local: perf_train_cpu
title: CPU での効率的なトレーニング
- local: perf_train_cpu_many
title: 分散CPUトレーニング
- local: perf_train_tpu
title: TPU に関するトレーニング
- local: perf_train_tpu_tf
title: TensorFlow を使用した TPU のトレーニング
- local: perf_train_special
title: 特殊なハードウェアに関するトレーニング
- local: perf_hardware
title: トレーニング用のカスタム ハードウェア
- local: hpo_train
title: Trainer API を使用したハイパーパラメータ検索
title: 効率的なトレーニングテクニック
- sections:
- local: perf_infer_cpu
title: CPUでの推論
- local: perf_infer_gpu_one
title: 1 つの GPU での推論
- local: perf_infer_gpu_many
title: 多くの GPU での推論
- local: perf_infer_special
title: 特殊なハードウェアでの推論
title: 推論の最適化
- local: big_models
title: 大きなモデルのインスタンス化
- local: tf_xla
title: TensorFlowモデルのXLA統合
- local: perf_torch_compile
title: torch.compile()を使用した推論の最適化
title: パフォーマンスとスケーラビリティ
- sections:
- local: add_new_model
title: 🤗 Transformersにモデルを追加する方法
- local: testing
title: テスト
- local: pr_checks
title: プルリクエストのチェック
title: 貢献する
- sections:
- local: philosophy
title: フィロソフィー
- local: glossary
title: 用語集
- local: task_summary
title: 🤗 Transformersの機能
- local: tasks_explained
title: 🤗 Transformersがタスクを解決する方法
- local: model_summary
title: Transformerモデルファミリー
- local: tokenizer_summary
title: トークナイザーの概要
- local: attention
title: 注意機構
- local: pad_truncation
title: パディングと切り詰め
- local: bertology
title: BERTology
- local: perplexity
title: 固定長モデルのパープレキシティ
- local: pipeline_webserver
title: Webサーバー推論用パイプライン
- local: model_memory_anatomy
title: モデルトレーニングの解剖学
title: コンセプチュアルガイド
- sections:
- sections:
- local: model_doc/auto
title: Auto Classes
- local: main_classes/callback
title: コールバック
- local: main_classes/configuration
title: 構成
- local: main_classes/data_collator
title: データ照合者
- local: main_classes/keras_callbacks
title: Keras コールバック
- local: main_classes/logging
title: ロギング
- local: main_classes/model
title: モデル
- local: main_classes/text_generation
title: テキストの生成
- local: main_classes/onnx
title: ONNX
- local: main_classes/optimizer_schedules
title: 最適化
- local: main_classes/output
title: モデルの出力
- local: main_classes/pipelines
title: パイプライン
- local: main_classes/processors
title: プロセッサー
- local: main_classes/quantization
title: 量子化
- local: main_classes/tokenizer
title: トークナイザー
- local: main_classes/trainer
title: トレーナー
- local: main_classes/deepspeed
title: ディープスピードの統合
- local: main_classes/feature_extractor
title: 特徴抽出器
- local: main_classes/image_processor
title: 画像処理プロセッサ
title: 主要なクラス
- sections:
- isExpanded: false
sections:
- local: model_doc/albert
title: ALBERT
- local: model_doc/bart
title: BART
- local: model_doc/barthez
title: BARThez
- local: model_doc/bartpho
title: BARTpho
- local: model_doc/bert
title: BERT
- local: model_doc/bert-generation
title: BertGeneration
- local: model_doc/bert-japanese
title: BertJapanese
- local: model_doc/bertweet
title: Bertweet
- local: model_doc/big_bird
title: BigBird
- local: model_doc/bigbird_pegasus
title: BigBirdPegasus
- local: model_doc/biogpt
title: BioGpt
- local: model_doc/blenderbot
title: Blenderbot
- local: model_doc/blenderbot-small
title: Blenderbot Small
- local: model_doc/bloom
title: BLOOM
- local: model_doc/bort
title: BORT
- local: model_doc/byt5
title: ByT5
- local: model_doc/camembert
title: CamemBERT
- local: model_doc/canine
title: CANINE
- local: model_doc/codegen
title: CodeGen
- local: model_doc/code_llama
title: CodeLlama
- local: model_doc/convbert
title: ConvBERT
- local: model_doc/cpm
title: CPM
- local: model_doc/cpmant
title: CPMANT
- local: model_doc/ctrl
title: CTRL
- local: model_doc/deberta
title: DeBERTa
- local: model_doc/deberta-v2
title: DeBERTa-v2
- local: model_doc/dialogpt
title: DialoGPT
title: 文章モデル
- isExpanded: false
sections:
- local: model_doc/beit
title: BEiT
- local: model_doc/bit
title: BiT
- local: model_doc/conditional_detr
title: Conditional DETR
- local: model_doc/convnext
title: ConvNeXT
- local: model_doc/convnextv2
title: ConvNeXTV2
- local: model_doc/cvt
title: CvT
- local: model_doc/deformable_detr
title: Deformable DETR
- local: model_doc/deit
title: DeiT
- local: model_doc/deta
title: DETA
- local: model_doc/detr
title: DETR
- local: model_doc/dinat
title: DiNAT
title: ビジョンモデル
- isExpanded: false
sections:
- local: model_doc/audio-spectrogram-transformer
title: Audio Spectrogram Transformer
- local: model_doc/bark
title: Bark
- local: model_doc/clap
title: CLAP
title: 音声モデル
- isExpanded: false
sections:
- local: model_doc/align
title: ALIGN
- local: model_doc/altclip
title: AltCLIP
- local: model_doc/blip
title: BLIP
- local: model_doc/blip-2
title: BLIP-2
- local: model_doc/bridgetower
title: BridgeTower
- local: model_doc/bros
title: BROS
- local: model_doc/chinese_clip
title: Chinese-CLIP
- local: model_doc/clip
title: CLIP
- local: model_doc/clipseg
title: CLIPSeg
- local: model_doc/clvp
title: CLVP
- local: model_doc/data2vec
title: Data2Vec
- local: model_doc/deplot
title: DePlot
title: マルチモーダルモデル
- isExpanded: false
sections:
- local: model_doc/decision_transformer
title: Decision Transformer
title: 強化学習モデル
- isExpanded: false
sections:
- local: model_doc/autoformer
title: Autoformer
title: 時系列モデル
title: モデル
- sections:
- local: internal/modeling_utils
title: カスタムレイヤーとユーティリティ
- local: internal/pipelines_utils
title: パイプライン用のユーティリティ
- local: internal/tokenization_utils
title: ト=ークナイザー用のユーティリティ
- local: internal/trainer_utils
title: トレーナー用ユーティリティ
- local: internal/generation_utils
title: 発電用ユーティリティ
- local: internal/image_processing_utils
title: 画像プロセッサ用ユーティリティ
- local: internal/audio_utils
title: オーディオ処理用のユーティリティ
- local: internal/file_utils
title: 一般公共事業
- local: internal/time_series_utils
title: 時系列用のユーティリティ
title: 内部ヘルパー
title: API
| transformers/docs/source/ja/_toctree.yml/0 | {
"file_path": "transformers/docs/source/ja/_toctree.yml",
"repo_id": "transformers",
"token_count": 5655
} | 383 |
<!---
Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# インストール
使用しているDeep Learningライブラリに対して、🤗 Transformersをインストールしてキャッシュを設定、そしてオプションでオフラインで実行できるように 🤗 Transformersを設定します。
🤗 TransformersはPython 3.6+, PyTorch 1.1.0+, TensorFlow 2.0+, Flaxで動作確認しています。 使用しているDeep Learningライブラリに合わせて、以下のインストール方法に従ってください:
* [PyTorch](https://pytorch.org/get-started/locally/)のインストール手順。
* [TensorFlow 2.0](https://www.tensorflow.org/install/pip)のインストール手順。
* [Flax](https://flax.readthedocs.io/en/latest/)のインストール手順。
## pipでのインストール
🤗 Transformersを[仮想環境](https://docs.python.org/3/library/venv.html)にインストールする必要があります。 もし、Pythonの仮想環境に馴染みがない場合は、この[ガイド](https://packaging.python.org/guides/installing-using-pip-and-virtual-environments/)をご覧ください。仮想環境によって異なるプロジェクトの管理がより簡単になり、依存関係間の互換性の問題を回避できます。
まず、プロジェクトディレクトリに仮想環境を作成することから始めましょう:
```bash
python -m venv .env
```
仮想環境を起動しましょう。LinuxとMacOsの場合は以下のコマンドで起動します:
```bash
source .env/bin/activate
```
Windowsで仮想環境を起動します
```bash
.env/Scripts/activate
```
これで、次のコマンドで🤗 Transformersをインストールする準備が整いました:
```bash
pip install transformers
```
CPU対応のみ必要な場合、🤗 TransformersとDeep Learningライブラリを1行でインストールできるようになっていて便利です。例えば、🤗 TransformersとPyTorchを以下のように一緒にインストールできます:
```bash
pip install transformers[torch]
```
🤗 TransformersとTensorFlow 2.0:
```bash
pip install transformers[tf-cpu]
```
🤗 TransformersとFlax:
```bash
pip install transformers[flax]
```
最後に、以下のコマンドを実行することで🤗 Transformersが正しくインストールされているかを確認します。学習済みモデルがダウンロードされます:
```bash
python -c "from transformers import pipeline; print(pipeline('sentiment-analysis')('we love you'))"
```
その後、ラベルとスコアが出力されます:
```bash
[{'label': 'POSITIVE', 'score': 0.9998704791069031}]
```
## ソースからのインストール
以下のコマンドでソースから🤗 Transformersをインストールします:
```bash
pip install git+https://github.com/huggingface/transformers
```
このコマンドは最新の安定版ではなく、開発における最新の`main`バージョンをインストールします。`main`バージョンは最新の開発状況に対応するのに便利です。例えば、最後の公式リリース以降にバグが修正されたが、新しいリリースがまだ展開されていない場合などです。しかし、これは`main`バージョンが常に安定しているとは限らないことを意味します。私たちは`main`バージョンの運用を維持するよう努め、ほとんどの問題は通常、数時間から1日以内に解決されます。もし問題に遭遇した場合は、より早く修正できるように[Issue](https://github.com/huggingface/transformers/issues)を作成してください！
以下のコマンドを実行して、🤗 Transformersが正しくインストールされているかどうかを確認します:
```bash
python -c "from transformers import pipeline; print(pipeline('sentiment-analysis')('I love you'))"
```
## 編集可能なインストール
必要に応じて、編集可能なインストールをします:
* ソースコードの`main`バージョンを使います。
* 🤗 Transformersにコントリビュートし、コードの変更をテストする必要があります。
以下のコマンドでレポジトリをクローンして、🤗 Transformersをインストールします:
```bash
git clone https://github.com/huggingface/transformers.git
cd transformers
pip install -e .
```
上記のコマンドは、レポジトリをクローンしたフォルダとPythonのライブラリをパスをリンクします。Pythonは通常のライブラリパスに加えて、あなたがクローンしたフォルダの中も見るようになります。例えば、Pythonパッケージが通常、`~/anaconda3/envs/main/lib/python3.7/site-packages/`にインストールされている場合、Pythonはクローンしたフォルダも検索するようになります: `~/transformers/`.
<Tip warning={true}>
ライブラリーを使い続けたい場合は、transformersフォルダーを保持しつづける必要があります。
</Tip>
これで、次のコマンドで簡単にクローンを🤗 Transformersの最新版に更新できます:
```bash
cd ~/transformers/
git pull
```
Python環境は次回の実行時に🤗 Transformersの`main`バージョンを見つけるようになります。
## condaでのインストール
`conda-forge`のcondaチャンネルからインストールします:
```bash
conda install conda-forge::transformers
```
## キャッシュの設定
学習済みモデルはダウンロードされ、ローカルにキャッシュされます: `~/.cache/huggingface/hub`. これはシェル環境変数`TRANSFORMERS_CACHE`で指定されるデフォルトのディレクトリです。Windowsでは、デフォルトのディレクトリは`C:\Users\username\.cache\huggingface\hub`になっています。異なるキャッシュディレクトリを指定するために、以下のシェル環境変数を変更することが可能です。優先度は以下の順番に対応します:
1. シェル環境変数 (デフォルト): `HF_HUB_CACHE` または `TRANSFORMERS_CACHE`.
2. シェル環境変数: `HF_HOME`.
3. シェル環境変数: `XDG_CACHE_HOME` + `/huggingface`.
<Tip>
もし、以前のバージョンのライブラリを使用していた人で、`PYTORCH_TRANSFORMERS_CACHE`または`PYTORCH_PRETRAINED_BERT_CACHE`を設定していた場合、シェル環境変数`TRANSFORMERS_CACHE`を指定しない限り🤗 Transformersはこれらのシェル環境変数を使用します。
</Tip>
## オフラインモード
🤗 Transformersはローカルファイルのみを使用することでファイアウォールやオフラインの環境でも動作させることができます。この動作を有効にするためには、環境変数`HF_HUB_OFFLINE=1`を設定します。
<Tip>
環境変数`HF_DATASETS_OFFLINE=1`を設定し、オフライントレーニングワークフローに[🤗 Datasets](https://huggingface.co/docs/datasets/)を追加します。
</Tip>
例えば、外部インスタンスに対してファイアウォールで保護された通常のネットワーク上でプログラムを実行する場合、通常以下のようなコマンドで実行することになります:
```bash
python examples/pytorch/translation/run_translation.py --model_name_or_path google-t5/t5-small --dataset_name wmt16 --dataset_config ro-en ...
```
オフラインインスタンスでこの同じプログラムを実行します:
```bash
HF_DATASETS_OFFLINE=1 HF_HUB_OFFLINE=1 \
python examples/pytorch/translation/run_translation.py --model_name_or_path google-t5/t5-small --dataset_name wmt16 --dataset_config ro-en ...
```
このスクリプトは、ローカルファイルのみを検索することが分かっているので、ハングアップしたりタイムアウトを待ったりすることなく実行されるはずです。
### オフラインで使用するためにモデルやトークナイザーを取得する
オフラインで🤗 Transformersを使用するもう1つの方法は、前もってファイルをダウンロードしておき、オフラインで使用する必要があるときにそのローカルパスを指定することです。これには3つの方法があります:
* [Model Hub](https://huggingface.co/models)のユーザーインターフェース上から↓アイコンをクリックしてファイルをダウンロードする方法。
* [`PreTrainedModel.from_pretrained`]および[`PreTrainedModel.save_pretrained`]のワークフローを使用する方法:
1. [`PreTrainedModel.from_pretrained`]で前もってファイルをダウンロードします:
```py
>>> from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
>>> tokenizer = AutoTokenizer.from_pretrained("bigscience/T0_3B")
>>> model = AutoModelForSeq2SeqLM.from_pretrained("bigscience/T0_3B")
```
2. [`PreTrainedModel.save_pretrained`]で指定されたディレクトリにファイルを保存しておきます:
```py
>>> tokenizer.save_pretrained("./your/path/bigscience_t0")
>>> model.save_pretrained("./your/path/bigscience_t0")
```
3. オフラインにある時、[`PreTrainedModel.from_pretrained`]に指定したディレクトリからファイルをリロードします:
```py
>>> tokenizer = AutoTokenizer.from_pretrained("./your/path/bigscience_t0")
>>> model = AutoModel.from_pretrained("./your/path/bigscience_t0")
```
* プログラム的に[huggingface_hub](https://github.com/huggingface/huggingface_hub/tree/main/src/huggingface_hub)ライブラリを用いて、ファイルをダウンロードする方法:
1. 仮想環境に`huggingface_hub`ライブラリをインストールします:
```bash
python -m pip install huggingface_hub
```
2. 指定のパスにファイルをダウンロードするために、[`hf_hub_download`](https://huggingface.co/docs/hub/adding-a-library#download-files-from-the-hub)関数を使用します。例えば、以下のコマンドで、[T0](https://huggingface.co/bigscience/T0_3B)モデルの`config.json`ファイルを指定のパスにダウンロードできます:
```py
>>> from huggingface_hub import hf_hub_download
>>> hf_hub_download(repo_id="bigscience/T0_3B", filename="config.json", cache_dir="./your/path/bigscience_t0")
```
ファイルがダウンロードされ、ローカルにキャッシュされたら、そのローカルパスを指定してファイルをロードして使用します:
```py
>>> from transformers import AutoConfig
>>> config = AutoConfig.from_pretrained("./your/path/bigscience_t0/config.json")
```
<Tip>
Hubに保存されているファイルをダウンロードする方法の詳細については、[How to download files from the Hub](https://huggingface.co/docs/hub/how-to-downstream)セクションを参照してください。
</Tip>
| transformers/docs/source/ja/installation.md/0 | {
"file_path": "transformers/docs/source/ja/installation.md",
"repo_id": "transformers",
"token_count": 4805
} | 384 |
<!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Audio Spectrogram Transformer
## 概要
Audio Spectrogram Transformerモデルは、[AST: Audio Spectrogram Transformer](https://huggingface.co/papers/2104.01778)という論文でYuan Gong、Yu-An Chung、James Glassによって提案されました。これは、音声を画像（スペクトログラム）に変換することで、音声に[Vision Transformer](vit)を適用します。このモデルは音声分類において最先端の結果を得ています。
論文の要旨は以下の通りです：
*過去10年間で、畳み込みニューラルネットワーク（CNN）は、音声スペクトログラムから対応するラベルへの直接的なマッピングを学習することを目指す、エンドツーエンドの音声分類モデルの主要な構成要素として広く採用されてきました。長距離のグローバルなコンテキストをより良く捉えるため、最近の傾向として、CNNの上にセルフアテンション機構を追加し、CNN-アテンションハイブリッドモデルを形成することがあります。しかし、CNNへの依存が必要かどうか、そして純粋にアテンションに基づくニューラルネットワークだけで音声分類において良いパフォーマンスを得ることができるかどうかは明らかではありません。本論文では、これらの問いに答えるため、音声分類用では最初の畳み込みなしで純粋にアテンションベースのモデルであるAudio Spectrogram Transformer（AST）を紹介します。我々はASTを様々なオーディオ分類ベンチマークで評価し、AudioSetで0.485 mAP、ESC-50で95.6%の正解率、Speech Commands V2で98.1%の正解率という新たな最先端の結果を達成しました。*
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/audio_spectogram_transformer_architecture.png"
alt="drawing" width="600"/>
<small> Audio Spectrogram Transformerのアーキテクチャ。<a href="https://huggingface.co/papers/2104.01778">元論文</a>より抜粋。</small>
このモデルは[nielsr](https://huggingface.co/nielsr)より提供されました。
オリジナルのコードは[こちら](https://github.com/YuanGongND/ast)で見ることができます。
## 使用上のヒント
- 独自のデータセットでAudio Spectrogram Transformer（AST）をファインチューニングする場合、入力の正規化（入力の平均を0、標準偏差を0.5にすること）処理することが推奨されます。[`ASTFeatureExtractor`]はこれを処理します。デフォルトではAudioSetの平均と標準偏差を使用していることに注意してください。著者が下流のデータセットの統計をどのように計算しているかは、[`ast/src/get_norm_stats.py`](https://github.com/YuanGongND/ast/blob/master/src/get_norm_stats.py)で確認することができます。
- ASTは低い学習率が必要であり 著者は[PSLA論文](https://huggingface.co/papers/2102.01243)で提案されたCNNモデルに比べて10倍小さい学習率を使用しています）、素早く収束するため、タスクに適した学習率と学習率スケジューラーを探すことをお勧めします。
## 参考資料
Audio Spectrogram Transformerの使用を開始するのに役立つ公式のHugging Faceおよびコミュニティ（🌎で示されている）の参考資料の一覧です。
<PipelineTag pipeline="audio-classification"/>
- ASTを用いた音声分類の推論を説明するノートブックは[こちら](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/AST)で見ることができます。
- [`ASTForAudioClassification`]は、この[例示スクリプト](https://github.com/huggingface/transformers/tree/main/examples/pytorch/audio-classification)と[ノートブック](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/audio_classification.ipynb)によってサポートされています。
- こちらも参照：[音声分類タスク](../tasks/audio_classification)。
ここに参考資料を提出したい場合は、気兼ねなくPull Requestを開いてください。私たちはそれをレビューいたします！参考資料は、既存のものを複製するのではなく、何か新しいことを示すことが理想的です。
## ASTConfig
[[autodoc]] ASTConfig
## ASTFeatureExtractor
[[autodoc]] ASTFeatureExtractor
- __call__
## ASTModel
[[autodoc]] ASTModel
- forward
## ASTForAudioClassification
[[autodoc]] ASTForAudioClassification
- forward
| transformers/docs/source/ja/model_doc/audio-spectrogram-transformer.md/0 | {
"file_path": "transformers/docs/source/ja/model_doc/audio-spectrogram-transformer.md",
"repo_id": "transformers",
"token_count": 2252
} | 385 |
<!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Blenderbot Small
[`BlenderbotSmallModel`] と
[`BlenderbotSmallForConditionalGeneration`] はチェックポイントと組み合わせてのみ使用されます
[facebook/blenderbot-90M](https://huggingface.co/facebook/blenderbot-90M)。より大規模な Blenderbot チェックポイントは、
代わりに [`BlenderbotModel`] とともに使用してください。
[`BlenderbotForConditionalGeneration`]
## Overview
Blender チャットボット モデルは、[Recipes for building an open-domain chatbot](https://huggingface.co/papers/2004.13637) Stephen Roller、Emily Dinan、Naman Goyal、Da Ju、Mary Williamson、yinghan Liu、で提案されました。
ジン・シュー、マイル・オット、カート・シャスター、エリック・M・スミス、Y-ラン・ブーロー、ジェイソン・ウェストン、2020年4月30日。
論文の要旨は次のとおりです。
*オープンドメインのチャットボットの構築は、機械学習研究にとって難しい分野です。これまでの研究では次のことが示されていますが、
ニューラル モデルをパラメーターの数とトレーニング対象のデータのサイズでスケーリングすると、結果が向上します。
高性能のチャットボットには他の要素も重要であることを示します。良い会話には多くのことが必要です
会話の専門家がシームレスに融合するスキル: 魅力的な話のポイントを提供し、話を聞く
一貫した態度を維持しながら、知識、共感、個性を適切に表現する
ペルソナ。適切なトレーニング データと選択が与えられた場合、大規模モデルがこれらのスキルを学習できることを示します。
世代戦略。 90M、2.7B、9.4B パラメーター モデルを使用してこれらのレシピのバリアントを構築し、モデルを作成します。
コードは公開されています。人間による評価では、当社の最良のモデルが既存のアプローチよりも優れていることがマルチターンで示されています
魅力と人間性の測定という観点からの対話。次に、分析によってこの作業の限界について説明します。
弊社機種の故障事例*
チップ：
- Blenderbot Small は絶対位置埋め込みを備えたモデルなので、通常は入力を右側にパディングすることをお勧めします。
左。
このモデルは、[patrickvonplaten](https://huggingface.co/patrickvonplaten) によって提供されました。著者のコードは次のとおりです
[ここ](https://github.com/facebookresearch/ParlAI) をご覧ください。
## Documentation resources
- [因果言語モデリング タスク ガイド](../tasks/language_modeling)
- [翻訳タスクガイド](../tasks/translation)
- [要約タスクガイド](../tasks/summarization)
## BlenderbotSmallConfig
[[autodoc]] BlenderbotSmallConfig
## BlenderbotSmallTokenizer
[[autodoc]] BlenderbotSmallTokenizer
- build_inputs_with_special_tokens
- get_special_tokens_mask
- create_token_type_ids_from_sequences
- save_vocabulary
## BlenderbotSmallTokenizerFast
[[autodoc]] BlenderbotSmallTokenizerFast
## BlenderbotSmallModel
[[autodoc]] BlenderbotSmallModel
- forward
## BlenderbotSmallForConditionalGeneration
[[autodoc]] BlenderbotSmallForConditionalGeneration
- forward
## BlenderbotSmallForCausalLM
[[autodoc]] BlenderbotSmallForCausalLM
- forward
## TFBlenderbotSmallModel
[[autodoc]] TFBlenderbotSmallModel
- call
## TFBlenderbotSmallForConditionalGeneration
[[autodoc]] TFBlenderbotSmallForConditionalGeneration
- call
## FlaxBlenderbotSmallModel
[[autodoc]] FlaxBlenderbotSmallModel
- __call__
- encode
- decode
## FlaxBlenderbotForConditionalGeneration
[[autodoc]] FlaxBlenderbotSmallForConditionalGeneration
- __call__
- encode
- decode
| transformers/docs/source/ja/model_doc/blenderbot-small.md/0 | {
"file_path": "transformers/docs/source/ja/model_doc/blenderbot-small.md",
"repo_id": "transformers",
"token_count": 1830
} | 386 |
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contains specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# CodeLlama
## Overview
Code Llama モデルはによって [Code Llama: Open Foundation Models for Code](https://ai.meta.com/research/publications/code-llama-open-foundation-models-for-code/) で提案されました。 Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton Ferrer, Aaron Grattafiori, Wenhan Xiong, Alexandre Défossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, Gabriel Synnaeve.
論文の要約は次のとおりです。
*私たちは Code Llama をリリースします。これは Llama 2 に基づくコードの大規模言語モデル ファミリであり、オープン モデルの中で最先端のパフォーマンス、埋め込み機能、大規模な入力コンテキストのサポート、プログラミング タスクのゼロショット命令追従機能を提供します。 。幅広いアプリケーションをカバーするための複数のフレーバーを提供しています。基盤モデル (Code Llama)、Python 特化 (Code Llama - Python)、およびそれぞれ 7B、13B、および 34B パラメーターを備えた命令追従モデル (Code Llama - Instruct) です。すべてのモデルは 16,000 トークンのシーケンスでトレーニングされ、最大 100,000 トークンの入力で改善が見られます。 7B および 13B コード ラマとコード ラマ - 命令バリアントは、周囲のコンテンツに基づいた埋め込みをサポートします。 Code Llama は、いくつかのコード ベンチマークでオープン モデルの中で最先端のパフォーマンスに達し、HumanEval と MBPP でそれぞれ最大 53% と 55% のスコアを獲得しました。特に、Code Llama - Python 7B は HumanEval および MBPP 上で Llama 2 70B よりも優れたパフォーマンスを示し、すべてのモデルは MultiPL-E 上で公開されている他のすべてのモデルよりも優れています。私たちは、研究と商業利用の両方を許可する寛容なライセンスに基づいて Code Llama をリリースしています。*
すべての Code Llama モデル チェックポイントを [こちら](https://huggingface.co/models?search=code_llama) で確認し、[meta llama org](https://huggingface.co/meta-llama) で正式にリリースされたチェックポイントを確認してください。
このモデルは [ArthurZucker](https://huggingface.co/ArthurZ) によって提供されました。著者のオリジナルのコードは [こちら](https://github.com/facebookresearch/llama) にあります。
## Usage tips and examples
<Tip warning={true}>
Code Llama のベースとなる`Llama2`ファミリー モデルは、`bfloat16`を使用してトレーニングされましたが、元の推論では`float16`を使用します。さまざまな精度を見てみましょう。
* `float32`: モデルの初期化に関する PyTorch の規約では、モデルの重みがどの `dtype` で格納されたかに関係なく、モデルを `float32` にロードします。 「transformers」も、PyTorch との一貫性を保つためにこの規則に従っています。これはデフォルトで選択されます。 `AutoModel` API でストレージの重み付けタイプを使用してチェックポイントのロードをキャストする場合は、`dtype="auto"` を指定する必要があります。 `model = AutoModelForCausalLM.from_pretrained("path", dtype = "auto")`。
* `bfloat16`: コード Llama はこの精度でトレーニングされているため、さらなるトレーニングや微調整に使用することをお勧めします。
* `float16`: この精度を使用して推論を実行することをお勧めします。通常は `bfloat16` より高速であり、評価メトリクスには `bfloat16` と比べて明らかな低下が見られないためです。 bfloat16 を使用して推論を実行することもできます。微調整後、float16 と bfloat16 の両方で推論結果を確認することをお勧めします。
上で述べたように、モデルを初期化するときに `dtype="auto"` を使用しない限り、ストレージの重みの `dtype` はほとんど無関係です。その理由は、モデルが最初にダウンロードされ (オンラインのチェックポイントの `dtype` を使用)、次に `torch` のデフォルトの `dtype` にキャストされるためです (`torch.float32` になります)。指定された `dtype` がある場合は、代わりにそれが使用されます。
</Tip>
チップ：
- 充填タスクはすぐにサポートされます。入力を埋めたい場所には `tokenizer.fill_token` を使用する必要があります。
- モデル変換スクリプトは、`Llama2` ファミリの場合と同じです。
使用例は次のとおりです。
```bash
python src/transformers/models/llama/convert_llama_weights_to_hf.py \
--input_dir /path/to/downloaded/llama/weights --model_size 7B --output_dir /output/path
```
スクリプトを実行するには、(最大のバージョンであっても) float16 精度でモデル全体をホストするのに十分な CPU RAM が必要であることに注意してください。
いくつかのチェックポイントがあり、それぞれにモデルの各重みの一部が含まれているため、すべてを RAM にロードする必要があります)。
変換後、モデルとトークナイザーは次の方法でロードできます。
```python
>>> from transformers import LlamaForCausalLM, CodeLlamaTokenizer
>>> tokenizer = CodeLlamaTokenizer.from_pretrained("meta-llama/CodeLlama-7b-hf")
>>> model = LlamaForCausalLM.from_pretrained("meta-llama/CodeLlama-7b-hf")
>>> PROMPT = '''def remove_non_ascii(s: str) -> str:
""" <FILL_ME>
return result
'''
>>> input_ids = tokenizer(PROMPT, return_tensors="pt")["input_ids"]
>>> generated_ids = model.generate(input_ids, max_new_tokens=128)
>>> filling = tokenizer.batch_decode(generated_ids[:, input_ids.shape[1]:], skip_special_tokens = True)[0]
>>> print(PROMPT.replace("<FILL_ME>", filling))
def remove_non_ascii(s: str) -> str:
""" Remove non-ASCII characters from a string.
Args:
s: The string to remove non-ASCII characters from.
Returns:
The string with non-ASCII characters removed.
"""
result = ""
for c in s:
if ord(c) < 128:
result += c
return result
```
塗りつぶされた部分だけが必要な場合:
```python
>>> from transformers import pipeline
>>> import torch
>>> generator = pipeline("text-generation",model="meta-llama/CodeLlama-7b-hf",dtype=torch.float16, device_map="auto")
>>> generator('def remove_non_ascii(s: str) -> str:\n """ <FILL_ME>\n return result', max_new_tokens = 128)
[{'generated_text': 'def remove_non_ascii(s: str) -> str:\n """ <FILL_ME>\n return resultRemove non-ASCII characters from a string. """\n result = ""\n for c in s:\n if ord(c) < 128:\n result += c'}]
```
内部では、トークナイザーが [`<FILL_ME>` によって自動的に分割](https://huggingface.co/docs/transformers/main/model_doc/code_llama#transformers.CodeLlamaTokenizer.fill_token) して、[ に続く書式設定された入力文字列を作成します。オリジナルのトレーニング パターン](https://github.com/facebookresearch/codellama/blob/cb51c14ec761370ba2e2bc351374a79265d0465e/llama/generation.py#L402)。これは、パターンを自分で準備するよりも堅牢です。トークンの接着など、デバッグが非常に難しい落とし穴を回避できます。このモデルまたは他のモデルに必要な CPU および GPU メモリの量を確認するには、その値を決定するのに役立つ [この計算ツール](https://huggingface.co/spaces/hf-accelerate/model-memory-usage) を試してください。
LLaMA トークナイザーは、[sentencepiece](https://github.com/google/sentencepiece) に基づく BPE モデルです。センテンスピースの癖の 1 つは、シーケンスをデコードするときに、最初のトークンが単語の先頭 (例: 「Banana」) である場合、トークナイザーは文字列の先頭にプレフィックス スペースを追加しないことです。
<Tip>
コード Llama は、`Llama2` モデルと同じアーキテクチャを持っています。API リファレンスについては、[Llama2 のドキュメント ページ](llama2) を参照してください。
以下の Code Llama トークナイザーのリファレンスを見つけてください。
</Tip>
## CodeLlamaTokenizer
[[autodoc]] CodeLlamaTokenizer
- build_inputs_with_special_tokens
- get_special_tokens_mask
- create_token_type_ids_from_sequences
- save_vocabulary
## CodeLlamaTokenizerFast
[[autodoc]] CodeLlamaTokenizerFast
- build_inputs_with_special_tokens
- get_special_tokens_mask
- create_token_type_ids_from_sequences
- update_post_processor
- save_vocabulary
| transformers/docs/source/ja/model_doc/code_llama.md/0 | {
"file_path": "transformers/docs/source/ja/model_doc/code_llama.md",
"repo_id": "transformers",
"token_count": 4100
} | 387 |
<!--Copyright 2021 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# DePlot
## Overview
DePlot は、Fangyu Liu、Julian Martin Aisenschlos、Francesco Piccinno、Syrine Krichene、Chenxi Pang, Kenton Lee, Mandar Joshi, Wenhu Chen, Nigel Collier, Yasemin Altun. の論文 [DePlot: One-shot visual language reasoning by plot-to-table translation](https://huggingface.co/papers/2212.10505) で提案されました。パン・
論文の要約には次のように記載されています。
*チャートやプロットなどの視覚言語は人間の世界に遍在しています。プロットやチャートを理解するには、強力な推論スキルが必要です。従来の最先端 (SOTA) モデルには少なくとも数万のトレーニング サンプルが必要であり、その推論能力は、特に人間が作成した複雑なクエリでは依然として大幅に制限されています。この論文では、視覚言語推論に対する最初のワンショット ソリューションを紹介します。私たちは、視覚言語推論の課題を 2 つのステップに分解します。(1) プロットからテキストへの翻訳と、(2) 翻訳されたテキストに対する推論です。この方法の鍵となるのは、プロットまたはチャートの画像を線形化されたテーブルに変換する、DePlot という名前のモダリティ変換モジュールです。その後、DePlot の出力を直接使用して、事前トレーニング済みの大規模言語モデル (LLM) をプロンプトし、LLM の少数ショット推論機能を利用できます。 DePlot を取得するには、統一されたタスク形式とメトリクスを確立することでプロットからテーブルへのタスクを標準化し、このタスクで DePlot をエンドツーエンドでトレーニングします。 DePlot は、プラグアンドプレイ方式で LLM とともに既製で使用できます。 28,000 を超えるデータ ポイントで微調整された SOTA モデルと比較して、ワンショット プロンプトのみを使用する DePlot+LLM は、チャート QA タスクからの人が作成したクエリに関して、微調整された SOTA より 24.0% の改善を達成しました。*
DePlot は、`Pix2Struct` アーキテクチャを使用してトレーニングされたモデルです。 `Pix2Struct` の詳細については、[Pix2Struct ドキュメント](https://huggingface.co/docs/transformers/main/en/model_doc/pix2struct) を参照してください。
DePlot は、`Pix2Struct` アーキテクチャの Visual Question Answering サブセットです。入力された質問を画像上にレンダリングし、答えを予測します。
## Usage example
現在、DePlot で使用できるチェックポイントは 1 つです。
- `google/deplot`: ChartQA データセットで微調整された DePlot
```python
from transformers import AutoProcessor, Pix2StructForConditionalGeneration
import requests
from PIL import Image
model = Pix2StructForConditionalGeneration.from_pretrained("google/deplot")
processor = AutoProcessor.from_pretrained("google/deplot")
url = "https://raw.githubusercontent.com/vis-nlp/ChartQA/main/ChartQA%20Dataset/val/png/5090.png"
image = Image.open(requests.get(url, stream=True).raw)
inputs = processor(images=image, text="Generate underlying data table of the figure below:", return_tensors="pt")
predictions = model.generate(**inputs, max_new_tokens=512)
print(processor.decode(predictions[0], skip_special_tokens=True))
```
## Fine-tuning
DePlot を微調整するには、pix2struct [微調整ノートブック](https://github.com/huggingface/notebooks/blob/main/examples/image_captioning_pix2struct.ipynb) を参照してください。 `Pix2Struct` モデルの場合、Adafactor とコサイン学習率スケジューラを使用してモデルを微調整すると、収束が高速化されることがわかりました。
```python
from transformers.optimization import Adafactor, get_cosine_schedule_with_warmup
optimizer = Adafactor(self.parameters(), scale_parameter=False, relative_step=False, lr=0.01, weight_decay=1e-05)
scheduler = get_cosine_schedule_with_warmup(optimizer, num_warmup_steps=1000, num_training_steps=40000)
```
<Tip>
DePlot は、`Pix2Struct`アーキテクチャを使用してトレーニングされたモデルです。 API リファレンスについては、[`Pix2Struct` ドキュメント](pix2struct) を参照してください。
</Tip> | transformers/docs/source/ja/model_doc/deplot.md/0 | {
"file_path": "transformers/docs/source/ja/model_doc/deplot.md",
"repo_id": "transformers",
"token_count": 2028
} | 388 |
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Optimize inference using torch.compile()
このガイドは、[`torch.compile()`](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) を使用した推論速度の向上に関するベンチマークを提供することを目的としています。これは、[🤗 Transformers のコンピュータビジョンモデル](https://huggingface.co/models?pipeline_tag=image-classification&library=transformers&sort=trending)向けのものです。
## Benefits of torch.compile
`torch.compile()`の利点
モデルとGPUによっては、torch.compile()は推論時に最大30%の高速化を実現します。 `torch.compile()`を使用するには、バージョン2.0以上のtorchをインストールするだけです。
モデルのコンパイルには時間がかかるため、毎回推論するのではなく、モデルを1度だけコンパイルする場合に役立ちます。
任意のコンピュータビジョンモデルをコンパイルするには、以下のようにモデルに`torch.compile()`を呼び出します：
```diff
from transformers import AutoModelForImageClassification
model = AutoModelForImageClassification.from_pretrained(MODEL_ID, device_map="auto")
+ model = torch.compile(model)
```
`compile()` は、コンパイルに関する異なるモードを備えており、基本的にはコンパイル時間と推論のオーバーヘッドが異なります。`max-autotune` は `reduce-overhead` よりも時間がかかりますが、推論速度が速くなります。デフォルトモードはコンパイルにおいては最速ですが、推論時間においては `reduce-overhead` に比べて効率が良くありません。このガイドでは、デフォルトモードを使用しました。詳細については、[こちら](https://pytorch.org/get-started/pytorch-2.0/#user-experience) を参照してください。
`torch` バージョン 2.0.1 で異なるコンピュータビジョンモデル、タスク、ハードウェアの種類、およびバッチサイズを使用して `torch.compile` をベンチマークしました。
## Benchmarking code
以下に、各タスクのベンチマークコードを示します。推論前にGPUをウォームアップし、毎回同じ画像を使用して300回の推論の平均時間を取得します。
### Image Classification with ViT
```python
from PIL import Image
import requests
import numpy as np
from transformers import AutoImageProcessor, AutoModelForImageClassification
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
processor = AutoImageProcessor.from_pretrained("google/vit-base-patch16-224")
model = AutoModelForImageClassification.from_pretrained("google/vit-base-patch16-224", device_map="auto")
model = torch.compile(model)
processed_input = processor(image, return_tensors='pt').to(model.device)
with torch.no_grad():
_ = model(**processed_input)
```
#### Object Detection with DETR
```python
from transformers import AutoImageProcessor, AutoModelForObjectDetection
processor = AutoImageProcessor.from_pretrained("facebook/detr-resnet-50")
model = AutoModelForObjectDetection.from_pretrained("facebook/detr-resnet-50", device_map="auto")
model = torch.compile(model)
texts = ["a photo of a cat", "a photo of a dog"]
inputs = processor(text=texts, images=image, return_tensors="pt").to(model.device)
with torch.no_grad():
_ = model(**inputs)
```
#### Image Segmentation with Segformer
```python
from transformers import SegformerImageProcessor, SegformerForSemanticSegmentation
processor = SegformerImageProcessor.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512")
model = SegformerForSemanticSegmentation.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512", device_map="auto")
model = torch.compile(model)
seg_inputs = processor(images=image, return_tensors="pt").to(model.device)
with torch.no_grad():
_ = model(**seg_inputs)
```
以下は、私たちがベンチマークを行ったモデルのリストです。
**Image Classification**
- [google/vit-base-patch16-224](https://huggingface.co/google/vit-base-patch16-224)
- [microsoft/beit-base-patch16-224-pt22k-ft22k](https://huggingface.co/microsoft/beit-base-patch16-224-pt22k-ft22k)
- [facebook/convnext-large-224](https://huggingface.co/facebook/convnext-large-224)
- [microsoft/resnet-50](https://huggingface.co/)
**Image Segmentation**
- [nvidia/segformer-b0-finetuned-ade-512-512](https://huggingface.co/nvidia/segformer-b0-finetuned-ade-512-512)
- [facebook/mask2former-swin-tiny-coco-panoptic](https://huggingface.co/facebook/mask2former-swin-tiny-coco-panoptic)
- [facebook/maskformer-swin-base-ade](https://huggingface.co/facebook/maskformer-swin-base-ade)
- [google/deeplabv3_mobilenet_v2_1.0_513](https://huggingface.co/google/deeplabv3_mobilenet_v2_1.0_513)
**Object Detection**
- [google/owlvit-base-patch32](https://huggingface.co/google/owlvit-base-patch32)
- [facebook/detr-resnet-101](https://huggingface.co/facebook/detr-resnet-101)
- [microsoft/conditional-detr-resnet-50](https://huggingface.co/microsoft/conditional-detr-resnet-50)
以下は、`torch.compile()`を使用した場合と使用しない場合の推論時間の可視化と、異なるハードウェアとバッチサイズの各モデルに対するパフォーマンス向上の割合です。
<div class="flex">
<div>
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/torch_compile/a100_batch_comp.png" />
</div>
<div>
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/torch_compile/v100_batch_comp.png" />
</div>
<div>
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/torch_compile/t4_batch_comp.png" />
</div>
</div>
<div class="flex">
<div>
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/torch_compile/A100_1_duration.png" />
</div>
<div>
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/torch_compile/A100_1_percentage.png" />
</div>
</div>
下記は、各モデルについて`compile()`を使用した場合と使用しなかった場合の推論時間（ミリ秒単位）です。なお、OwlViTは大きなバッチサイズでの使用時にメモリ不足（OOM）が発生することに注意してください。
### A100 (batch size: 1)
| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
|:---:|:---:|:---:|
| Image Classification/ViT | 9.325 | 7.584 |
| Image Segmentation/Segformer | 11.759 | 10.500 |
| Object Detection/OwlViT | 24.978 | 18.420 |
| Image Classification/BeiT | 11.282 | 8.448 |
| Object Detection/DETR | 34.619 | 19.040 |
| Image Classification/ConvNeXT | 10.410 | 10.208 |
| Image Classification/ResNet | 6.531 | 4.124 |
| Image Segmentation/Mask2former | 60.188 | 49.117 |
| Image Segmentation/Maskformer | 75.764 | 59.487 |
| Image Segmentation/MobileNet | 8.583 | 3.974 |
| Object Detection/Resnet-101 | 36.276 | 18.197 |
| Object Detection/Conditional-DETR | 31.219 | 17.993 |
### A100 (batch size: 4)
| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
|:---:|:---:|:---:|
| Image Classification/ViT | 14.832 | 14.499 |
| Image Segmentation/Segformer | 18.838 | 16.476 |
| Image Classification/BeiT | 13.205 | 13.048 |
| Object Detection/DETR | 48.657 | 32.418|
| Image Classification/ConvNeXT | 22.940 | 21.631 |
| Image Classification/ResNet | 6.657 | 4.268 |
| Image Segmentation/Mask2former | 74.277 | 61.781 |
| Image Segmentation/Maskformer | 180.700 | 159.116 |
| Image Segmentation/MobileNet | 14.174 | 8.515 |
| Object Detection/Resnet-101 | 68.101 | 44.998 |
| Object Detection/Conditional-DETR | 56.470 | 35.552 |
### A100 (batch size: 16)
| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
|:---:|:---:|:---:|
| Image Classification/ViT | 40.944 | 40.010 |
| Image Segmentation/Segformer | 37.005 | 31.144 |
| Image Classification/BeiT | 41.854 | 41.048 |
| Object Detection/DETR | 164.382 | 161.902 |
| Image Classification/ConvNeXT | 82.258 | 75.561 |
| Image Classification/ResNet | 7.018 | 5.024 |
| Image Segmentation/Mask2former | 178.945 | 154.814 |
| Image Segmentation/Maskformer | 638.570 | 579.826 |
| Image Segmentation/MobileNet | 51.693 | 30.310 |
| Object Detection/Resnet-101 | 232.887 | 155.021 |
| Object Detection/Conditional-DETR | 180.491 | 124.032 |
### V100 (batch size: 1)
| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
|:---:|:---:|:---:|
| Image Classification/ViT | 10.495 | 6.00 |
| Image Segmentation/Segformer | 13.321 | 5.862 |
| Object Detection/OwlViT | 25.769 | 22.395 |
| Image Classification/BeiT | 11.347 | 7.234 |
| Object Detection/DETR | 33.951 | 19.388 |
| Image Classification/ConvNeXT | 11.623 | 10.412 |
| Image Classification/ResNet | 6.484 | 3.820 |
| Image Segmentation/Mask2former | 64.640 | 49.873 |
| Image Segmentation/Maskformer | 95.532 | 72.207 |
| Image Segmentation/MobileNet | 9.217 | 4.753 |
| Object Detection/Resnet-101 | 52.818 | 28.367 |
| Object Detection/Conditional-DETR | 39.512 | 20.816 |
### V100 (batch size: 4)
| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
|:---:|:---:|:---:|
| Image Classification/ViT | 15.181 | 14.501 |
| Image Segmentation/Segformer | 16.787 | 16.188 |
| Image Classification/BeiT | 15.171 | 14.753 |
| Object Detection/DETR | 88.529 | 64.195 |
| Image Classification/ConvNeXT | 29.574 | 27.085 |
| Image Classification/ResNet | 6.109 | 4.731 |
| Image Segmentation/Mask2former | 90.402 | 76.926 |
| Image Segmentation/Maskformer | 234.261 | 205.456 |
| Image Segmentation/MobileNet | 24.623 | 14.816 |
| Object Detection/Resnet-101 | 134.672 | 101.304 |
| Object Detection/Conditional-DETR | 97.464 | 69.739 |
### V100 (batch size: 16)
| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
|:---:|:---:|:---:|
| Image Classification/ViT | 52.209 | 51.633 |
| Image Segmentation/Segformer | 61.013 | 55.499 |
| Image Classification/BeiT | 53.938 | 53.581 |
| Object Detection/DETR | OOM | OOM |
| Image Classification/ConvNeXT | 109.682 | 100.771 |
| Image Classification/ResNet | 14.857 | 12.089 |
| Image Segmentation/Mask2former | 249.605 | 222.801 |
| Image Segmentation/Maskformer | 831.142 | 743.645 |
| Image Segmentation/MobileNet | 93.129 | 55.365 |
| Object Detection/Resnet-101 | 482.425 | 361.843 |
| Object Detection/Conditional-DETR | 344.661 | 255.298 |
### T4 (batch size: 1)
| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
|:---:|:---:|:---:|
| Image Classification/ViT | 16.520 | 15.786 |
| Image Segmentation/Segformer | 16.116 | 14.205 |
| Object Detection/OwlViT | 53.634 | 51.105 |
| Image Classification/BeiT | 16.464 | 15.710 |
| Object Detection/DETR | 73.100 | 53.99 |
| Image Classification/ConvNeXT | 32.932 | 30.845 |
| Image Classification/ResNet | 6.031 | 4.321 |
| Image Segmentation/Mask2former | 79.192 | 66.815 |
| Image Segmentation/Maskformer | 200.026 | 188.268 |
| Image Segmentation/MobileNet | 18.908 | 11.997 |
| Object Detection/Resnet-101 | 106.622 | 82.566 |
| Object Detection/Conditional-DETR | 77.594 | 56.984 |
### T4 (batch size: 4)
| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
|:---:|:---:|:---:|
| Image Classification/ViT | 43.653 | 43.626 |
| Image Segmentation/Segformer | 45.327 | 42.445 |
| Image Classification/BeiT | 52.007 | 51.354 |
| Object Detection/DETR | 277.850 | 268.003 |
| Image Classification/ConvNeXT | 119.259 | 105.580 |
| Image Classification/ResNet | 13.039 | 11.388 |
| Image Segmentation/Mask2former | 201.540 | 184.670 |
| Image Segmentation/Maskformer | 764.052 | 711.280 |
| Image Segmentation/MobileNet | 74.289 | 48.677 |
| Object Detection/Resnet-101 | 421.859 | 357.614 |
| Object Detection/Conditional-DETR | 289.002 | 226.945 |
### T4 (batch size: 16)
| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
|:---:|:---:|:---:|
| Image Classification/ViT | 163.914 | 160.907 |
| Image Segmentation/Segformer | 192.412 | 163.620 |
| Image Classification/BeiT | 188.978 | 187.976 |
| Object Detection/DETR | OOM | OOM |
| Image Classification/ConvNeXT | 422.886 | 388.078 |
| Image Classification/ResNet | 44.114 | 37.604 |
| Image Segmentation/Mask2former | 756.337 | 695.291 |
| Image Segmentation/Maskformer | 2842.940 | 2656.88 |
| Image Segmentation/MobileNet | 299.003 | 201.942 |
| Object Detection/Resnet-101 | 1619.505 | 1262.758 |
| Object Detection/Conditional-DETR | 1137.513 | 897.390|
## PyTorch Nightly
また、PyTorchのナイトリーバージョン（2.1.0dev）でのベンチマークを行い、コンパイルされていないモデルとコンパイル済みモデルの両方でレイテンシーの向上を観察しました。ホイールは[こちら](https://download.pytorch.org/whl/nightly/cu118)から入手できます。
### A100
| **Task/Model** | **Batch Size** | **torch 2.0 - no compile** | **torch 2.0 -<br> compile** |
|:---:|:---:|:---:|:---:|
| Image Classification/BeiT | Unbatched | 12.462 | 6.954 |
| Image Classification/BeiT | 4 | 14.109 | 12.851 |
| Image Classification/BeiT | 16 | 42.179 | 42.147 |
| Object Detection/DETR | Unbatched | 30.484 | 15.221 |
| Object Detection/DETR | 4 | 46.816 | 30.942 |
| Object Detection/DETR | 16 | 163.749 | 163.706 |
### T4
| **Task/Model** | **Batch Size** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
|:---:|:---:|:---:|:---:|
| Image Classification/BeiT | Unbatched | 14.408 | 14.052 |
| Image Classification/BeiT | 4 | 47.381 | 46.604 |
| Image Classification/BeiT | 16 | 42.179 | 42.147 |
| Object Detection/DETR | Unbatched | 68.382 | 53.481 |
| Object Detection/DETR | 4 | 269.615 | 204.785 |
| Object Detection/DETR | 16 | OOM | OOM |
### V100
| **Task/Model** | **Batch Size** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
|:---:|:---:|:---:|:---:|
| Image Classification/BeiT | Unbatched | 13.477 | 7.926 |
| Image Classification/BeiT | 4 | 15.103 | 14.378 |
| Image Classification/BeiT | 16 | 52.517 | 51.691 |
| Object Detection/DETR | Unbatched | 28.706 | 19.077 |
| Object Detection/DETR | 4 | 88.402 | 62.949|
| Object Detection/DETR | 16 | OOM | OOM |
## Reduce Overhead
NightlyビルドでA100およびT4向けの `reduce-overhead` コンパイルモードをベンチマークしました。
### A100
| **Task/Model** | **Batch Size** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
|:---:|:---:|:---:|:---:|
| Image Classification/ConvNeXT | Unbatched | 11.758 | 7.335 |
| Image Classification/ConvNeXT | 4 | 23.171 | 21.490 |
| Image Classification/ResNet | Unbatched | 7.435 | 3.801 |
| Image Classification/ResNet | 4 | 7.261 | 2.187 |
| Object Detection/Conditional-DETR | Unbatched | 32.823 | 11.627 |
| Object Detection/Conditional-DETR | 4 | 50.622 | 33.831 |
| Image Segmentation/MobileNet | Unbatched | 9.869 | 4.244 |
| Image Segmentation/MobileNet | 4 | 14.385 | 7.946 |
### T4
| **Task/Model** | **Batch Size** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
|:---:|:---:|:---:|:---:|
| Image Classification/ConvNeXT | Unbatched | 32.137 | 31.84 |
| Image Classification/ConvNeXT | 4 | 120.944 | 110.209 |
| Image Classification/ResNet | Unbatched | 9.761 | 7.698 |
| Image Classification/ResNet | 4 | 15.215 | 13.871 |
| Object Detection/Conditional-DETR | Unbatched | 72.150 | 57.660 |
| Object Detection/Conditional-DETR | 4 | 301.494 | 247.543 |
| Image Segmentation/MobileNet | Unbatched | 22.266 | 19.339 |
| Image Segmentation/MobileNet | 4 | 78.311 | 50.983 |
| transformers/docs/source/ja/perf_torch_compile.md/0 | {
"file_path": "transformers/docs/source/ja/perf_torch_compile.md",
"repo_id": "transformers",
"token_count": 6641
} | 389 |
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Train with a script
🤗 Transformersの[notebooks](./notebooks/README)と一緒に、[PyTorch](https://github.com/huggingface/transformers/tree/main/examples/pytorch)、[TensorFlow](https://github.com/huggingface/transformers/tree/main/examples/tensorflow)、または[JAX/Flax](https://github.com/huggingface/transformers/tree/main/examples/flax)を使用してモデルをトレーニングする方法を示すサンプルスクリプトもあります。
また、私たちの[研究プロジェクト](https://github.com/huggingface/transformers-research-projects/)や[レガシーの例](https://github.com/huggingface/transformers/tree/main/examples/legacy)で使用したスクリプトも見つかります。これらのスクリプトは現在メンテナンスされておらず、おそらく最新バージョンのライブラリと互換性がない特定の🤗 Transformersのバージョンが必要です。
サンプルスクリプトはすべての問題でそのまま動作することは期待されておらず、解決しようとしている問題にスクリプトを適応させる必要があるかもしれません。この点をサポートするために、ほとんどのスクリプトはデータがどのように前処理されているかを完全に公開し、必要に応じて編集できるようにしています。
サンプルスクリプトで実装したい機能がある場合は、[フォーラム](https://discuss.huggingface.co/)か[イシュートラッカー](https://github.com/huggingface/transformers/issues)で議論してからプルリクエストを提出してください。バグ修正は歓迎しますが、読みやすさのコストで機能を追加するプルリクエストはほとんどマージされない可能性が高いです。
このガイドでは、[PyTorch](https://github.com/huggingface/transformers/tree/main/examples/pytorch/summarization)と[TensorFlow](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/summarization)で実行するサマリゼーショントレーニングスクリプトの実行方法を示します。すべての例は、明示的に指定されていない限り、両方のフレームワークともに動作することが期待されています。
## Setup
最新バージョンのサンプルスクリプトを正常に実行するには、新しい仮想環境に🤗 Transformersをソースからインストールする必要があります:
```bash
git clone https://github.com/huggingface/transformers
cd transformers
pip install .
```
以前のスクリプトのバージョンについては、以下のトグルをクリックしてください：
<details>
<summary>以前の🤗 Transformersのバージョンに関する例</summary>
<ul>
<li><a href="https://github.com/huggingface/transformers/tree/v4.5.1/examples">v4.5.1</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v4.4.2/examples">v4.4.2</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v4.3.3/examples">v4.3.3</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v4.2.2/examples">v4.2.2</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v4.1.1/examples">v4.1.1</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v4.0.1/examples">v4.0.1</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v3.5.1/examples">v3.5.1</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v3.4.0/examples">v3.4.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v3.3.1/examples">v3.3.1</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v3.2.0/examples">v3.2.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v3.1.0/examples">v3.1.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v3.0.2/examples">v3.0.2</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.11.0/examples">v2.11.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.10.0/examples">v2.10.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.9.1/examples">v2.9.1</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.8.0/examples">v2.8.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.7.0/examples">v2.7.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.6.0/examples">v2.6.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.5.1/examples">v2.5.1</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.4.0/examples">v2.4.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.3.0/examples">v2.3.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.2.0/examples">v2.2.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.1.0/examples">v2.1.1</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.0.0/examples">v2.0.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v1.2.0/examples">v1.2.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v1.1.0/examples">v1.1.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v1.0.0/examples">v1.0.0</a></li>
</ul>
</details>
次に、現在の🤗 Transformersのクローンを特定のバージョンに切り替えてください。たとえば、v3.5.1などです。
```bash
git checkout tags/v3.5.1
```
適切なライブラリバージョンを設定したら、任意の例のフォルダに移動し、例固有の要件をインストールします：
```bash
pip install -r requirements.txt
```
## Run a script
<frameworkcontent>
<pt>
この例のスクリプトは、🤗 [Datasets](https://huggingface.co/docs/datasets/) ライブラリからデータセットをダウンロードし、前処理を行います。次に、[Trainer](https://huggingface.co/docs/transformers/main_classes/trainer) を使用して要約をサポートするアーキテクチャ上でデータセットをファインチューニングします。以下の例では、[CNN/DailyMail](https://huggingface.co/datasets/cnn_dailymail) データセット上で [T5-small](https://huggingface.co/google-t5/t5-small) をファインチューニングする方法が示されています。T5 モデルは、そのトレーニング方法に起因して追加の `source_prefix` 引数が必要です。このプロンプトにより、T5 はこれが要約タスクであることを知ることができます。
```bash
python examples/pytorch/summarization/run_summarization.py \
--model_name_or_path google-t5/t5-small \
--do_train \
--do_eval \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--source_prefix "summarize: " \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size=4 \
--per_device_eval_batch_size=4 \
--overwrite_output_dir \
--predict_with_generate
```
</pt>
<tf>
この例のスクリプトは、🤗 [Datasets](https://huggingface.co/docs/datasets/) ライブラリからデータセットをダウンロードして前処理します。その後、スクリプトは要約をサポートするアーキテクチャ上で Keras を使用してデータセットをファインチューニングします。以下の例では、[T5-small](https://huggingface.co/google-t5/t5-small) を [CNN/DailyMail](https://huggingface.co/datasets/cnn_dailymail) データセットでファインチューニングする方法を示しています。T5 モデルは、そのトレーニング方法に起因して追加の `source_prefix` 引数が必要です。このプロンプトは、T5 にこれが要約タスクであることを知らせます。
```bash
python examples/tensorflow/summarization/run_summarization.py \
--model_name_or_path google-t5/t5-small \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size 8 \
--per_device_eval_batch_size 16 \
--num_train_epochs 3 \
--do_train \
--do_eval
```
</tf>
</frameworkcontent>
## Distributed training and mixed precision
[Trainer](https://huggingface.co/docs/transformers/main_classes/trainer)は、分散トレーニングと混合精度をサポートしています。つまり、この機能をスクリプトで使用することができます。これらの機能を有効にするには、次の手順を実行します。
- `fp16`引数を追加して混合精度を有効にします。
- `nproc_per_node`引数で使用するGPUの数を設定します。
以下は提供されたBashコードです。このコードの日本語訳をMarkdown形式で記載します。
```bash
torchrun \
--nproc_per_node 8 pytorch/summarization/run_summarization.py \
--fp16 \
--model_name_or_path google-t5/t5-small \
--do_train \
--do_eval \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--source_prefix "summarize: " \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size=4 \
--per_device_eval_batch_size=4 \
--overwrite_output_dir \
--predict_with_generate
```
TensorFlowスクリプトは、分散トレーニングに[`MirroredStrategy`](https://www.tensorflow.org/guide/distributed_training#mirroredstrategy)を使用し、トレーニングスクリプトに追加の引数を追加する必要はありません。TensorFlowスクリプトは、デフォルトで複数のGPUが利用可能な場合にそれらを使用します。
## Run a script on a TPU
<frameworkcontent>
<pt>
Tensor Processing Units (TPUs)は、パフォーマンスを加速させるために特別に設計されています。PyTorchは、[XLA](https://www.tensorflow.org/xla)ディープラーニングコンパイラを使用してTPUsをサポートしており、詳細については[こちら](https://github.com/pytorch/xla/blob/master/README.md)をご覧ください。TPUを使用するには、`xla_spawn.py`スクリプトを起動し、`num_cores`引数を使用して使用するTPUコアの数を設定します。
```bash
python xla_spawn.py --num_cores 8 \
summarization/run_summarization.py \
--model_name_or_path google-t5/t5-small \
--do_train \
--do_eval \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--source_prefix "summarize: " \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size=4 \
--per_device_eval_batch_size=4 \
--overwrite_output_dir \
--predict_with_generate
```
</pt>
<tf>
もちろん、Tensor Processing Units（TPUs）は性能を高速化するために特別に設計されています。TensorFlowスクリプトは、TPUsでトレーニングするために[`TPUStrategy`](https://www.tensorflow.org/guide/distributed_training#tpustrategy)を利用します。TPUを使用するには、TPUリソースの名前を`tpu`引数に渡します。
```bash
python run_summarization.py \
--tpu name_of_tpu_resource \
--model_name_or_path google-t5/t5-small \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size 8 \
--per_device_eval_batch_size 16 \
--num_train_epochs 3 \
--do_train \
--do_eval
```
</tf>
</frameworkcontent>
## Run a script with 🤗 Accelerate
🤗 [Accelerate](https://huggingface.co/docs/accelerate)は、PyTorch専用のライブラリで、CPUのみ、複数のGPU、TPUなど、さまざまなセットアップでモデルをトレーニングするための統一された方法を提供します。PyTorchのトレーニングループを完全に可視化しながら実行できます。まだインストールしていない場合は、🤗 Accelerateをインストールしてください：
> 注意：Accelerateは急速に開発が進行しているため、スクリプトを実行するにはaccelerateのgitバージョンをインストールする必要があります
```bash
pip install git+https://github.com/huggingface/accelerate
```
代わりに、`run_summarization_no_trainer.py` スクリプトを使用する必要があります。 🤗 Accelerate がサポートするスクリプトには、フォルダ内に `task_no_trainer.py` ファイルが含まれています。まず、次のコマンドを実行して設定ファイルを作成し、保存します：
```bash
accelerate config
```
テストを行い、設定が正しく構成されているか確認してください：
```bash
accelerate test
```
Now you are ready to launch the training:
```bash
accelerate launch run_summarization_no_trainer.py \
--model_name_or_path google-t5/t5-small \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--source_prefix "summarize: " \
--output_dir ~/tmp/tst-summarization
```
## Use a custom dataset
要約スクリプトは、CSVまたはJSON Lineファイルであれば、カスタムデータセットをサポートしています。独自のデータセットを使用する場合、いくつかの追加の引数を指定する必要があります。
- `train_file`および`validation_file`は、トレーニングとバリデーションのファイルへのパスを指定します。
- `text_column`は要約するための入力テキストです。
- `summary_column`は出力する対象テキストです。
カスタムデータセットを使用した要約スクリプトは、以下のようになります：
```bash
python examples/pytorch/summarization/run_summarization.py \
--model_name_or_path google-t5/t5-small \
--do_train \
--do_eval \
--train_file path_to_csv_or_jsonlines_file \
--validation_file path_to_csv_or_jsonlines_file \
--text_column text_column_name \
--summary_column summary_column_name \
--source_prefix "summarize: " \
--output_dir /tmp/tst-summarization \
--overwrite_output_dir \
--per_device_train_batch_size=4 \
--per_device_eval_batch_size=4 \
--predict_with_generate
```
## Test a script
すべてが予想通りに動作することを確認するために、データセット全体を処理する前に、データセットの一部の例でスクリプトを実行することは良いアイデアです。以下の引数を使用して、データセットを最大サンプル数に切り詰めます：
- `max_train_samples`
- `max_eval_samples`
- `max_predict_samples`
```bash
python examples/pytorch/summarization/run_summarization.py \
--model_name_or_path google-t5/t5-small \
--max_train_samples 50 \
--max_eval_samples 50 \
--max_predict_samples 50 \
--do_train \
--do_eval \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--source_prefix "summarize: " \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size=4 \
--per_device_eval_batch_size=4 \
--overwrite_output_dir \
--predict_with_generate
```
一部の例のスクリプトは、`max_predict_samples`引数をサポートしていないことがあります。この引数がサポートされているかどうかがわからない場合は、`-h`引数を追加して確認してください。
```bash
examples/pytorch/summarization/run_summarization.py -h
```
## Resume training from checkpoint
以前のチェックポイントからトレーニングを再開するための役立つオプションもあります。これにより、トレーニングが中断された場合でも、最初からやり直すことなく、中断したところから再開できます。チェックポイントからトレーニングを再開するための2つの方法があります。
最初の方法は、`output_dir previous_output_dir` 引数を使用して、`output_dir` に保存された最新のチェックポイントからトレーニングを再開する方法です。この場合、`overwrite_output_dir` を削除する必要があります：
```bash
python examples/pytorch/summarization/run_summarization.py
--model_name_or_path google-t5/t5-small \
--do_train \
--do_eval \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--source_prefix "summarize: " \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size=4 \
--per_device_eval_batch_size=4 \
--output_dir previous_output_dir \
--predict_with_generate
```
2番目の方法では、`resume_from_checkpoint path_to_specific_checkpoint` 引数を使用して、特定のチェックポイントフォルダからトレーニングを再開します。
```bash
python examples/pytorch/summarization/run_summarization.py
--model_name_or_path google-t5/t5-small \
--do_train \
--do_eval \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--source_prefix "summarize: " \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size=4 \
--per_device_eval_batch_size=4 \
--overwrite_output_dir \
--resume_from_checkpoint path_to_specific_checkpoint \
--predict_with_generate
```
## Share your model
すべてのスクリプトは、最終的なモデルを [Model Hub](https://huggingface.co/models) にアップロードできます。開始する前に Hugging Face にログインしていることを確認してください。
```bash
hf auth login
```
次に、スクリプトに `push_to_hub` 引数を追加します。この引数は、Hugging Face のユーザー名と `output_dir` で指定したフォルダ名でリポジトリを作成します。
特定の名前をリポジトリに付けるには、`push_to_hub_model_id` 引数を使用して追加します。このリポジトリは自動的にあなたの名前空間の下にリストされます。
以下の例は、特定のリポジトリ名でモデルをアップロードする方法を示しています:
```bash
python examples/pytorch/summarization/run_summarization.py
--model_name_or_path google-t5/t5-small \
--do_train \
--do_eval \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--source_prefix "summarize: " \
--push_to_hub \
--push_to_hub_model_id finetuned-t5-cnn_dailymail \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size=4 \
--per_device_eval_batch_size=4 \
--overwrite_output_dir \
--predict_with_generate
```
| transformers/docs/source/ja/run_scripts.md/0 | {
"file_path": "transformers/docs/source/ja/run_scripts.md",
"repo_id": "transformers",
"token_count": 8240
} | 390 |
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# LLM prompting guide
[[open-in-colab]]
Falcon、LLaMA などの大規模言語モデルは、事前にトレーニングされたトランスフォーマー モデルであり、最初は予測するようにトレーニングされています。
入力テキストが与えられた場合の次のトークン。通常、数十億のパラメータがあり、何兆ものパラメータでトレーニングされています。
長期間のトークン。その結果、これらのモデルは非常に強力で多用途になり、次のようなことが可能になります。
自然言語プロンプトでモデルに指示することで、すぐに複数の NLP タスクを解決できます。
最適な出力を保証するためにこのようなプロンプトを設計することは、多くの場合「プロンプト エンジニアリング」と呼ばれます。プロンプトエンジニアリングとは、
かなりの量の実験を必要とする反復プロセス。自然言語ははるかに柔軟で表現力豊かです
ただし、プログラミング言語よりもあいまいさが生じる可能性があります。同時に、自然言語によるプロンプト
変化にはかなり敏感です。プロンプトにわずかな変更を加えただけでも、出力が大幅に異なる場合があります。
すべてのケースに適合するプロンプトを作成するための正確なレシピはありませんが、研究者はいくつかの最良のレシピを考案しました。
最適な結果をより一貫して達成するのに役立つ実践。
このガイドでは、より優れた LLM プロンプトを作成し、さまざまな NLP タスクを解決するのに役立つプロンプト エンジニアリングのベスト プラクティスについて説明します。
次のことを学びます:
- [プロンプトの基本](#basics-of-prompting)
- [LLM プロンプトのベスト プラクティス](#best-practices-of-llm-prompting)
- [高度なプロンプト テクニック: 数回のプロンプトと思考の連鎖](#advanced-prompting-techniques)
- [プロンプトを表示する代わりに微調整する場合](#prompting-vs-fine-tuning)
<Tip>
迅速なエンジニアリングは、LLM 出力最適化プロセスの一部にすぎません。もう 1 つの重要な要素は、
最適なテキスト生成戦略。 LLM が生成時に後続の各トークンを選択する方法をカスタマイズできます。
トレーニング可能なパラメータを一切変更せずにテキストを作成します。テキスト生成パラメータを微調整することで、
生成されたテキストに繰り返しが含まれているため、より一貫性があり人間らしい響きになります。
テキスト生成戦略とパラメーターはこのガイドの範囲外ですが、これらのトピックについて詳しくは、次のトピックを参照してください。
次のガイド:
* [LLM による生成](../llm_tutorial)
* [テキスト生成戦略](../generation_strategies)
</Tip>
## Basics of prompting
### Types of models
最新の LLM の大部分は、デコーダ専用のトランスフォーマーです。例としては、[LLaMA](../model_doc/llama)、
[Llama2](../model_doc/llama2)、[Falcon](../model_doc/falcon)、[GPT2](../model_doc/gpt2)。ただし、遭遇する可能性があります
エンコーダ デコーダ トランスフォーマ LLM も同様です。たとえば、[Flan-T5](../model_doc/flan-t5) や [BART](../model_doc/bart) です。
エンコーダ デコーダ スタイルのモデルは通常、出力が入力に**大きく**依存する生成タスクで使用されます。
たとえば、翻訳と要約です。デコーダ専用モデルは、他のすべてのタイプの生成タスクに使用されます。
パイプラインを使用して LLM でテキストを生成する場合、使用している LLM のタイプを知ることが重要です。
異なるパイプラインを使用します。
`text-generation`パイプラインを使用してデコーダのみのモデルで推論を実行します。
```python
>>> from transformers import pipeline
>>> import torch
>>> torch.manual_seed(0) # doctest: +IGNORE_RESULT
>>> generator = pipeline('text-generation', model = 'openai-community/gpt2')
>>> prompt = "Hello, I'm a language model"
>>> generator(prompt, max_length = 30)
[{'generated_text': "Hello, I'm a language model expert, so I'm a big believer in the concept that I know very well and then I try to look into"}]
```
エンコーダー/デコーダーを使用して推論を実行するには、`text2text-generation` パイプラインを使用します。
```python
>>> text2text_generator = pipeline("text2text-generation", model = 'google/flan-t5-base')
>>> prompt = "Translate from English to French: I'm very happy to see you"
>>> text2text_generator(prompt)
[{'generated_text': 'Je suis très heureuse de vous rencontrer.'}]
```
### Base vs instruct/chat models
🤗 Hub で利用できる最近の LLM チェックポイントのほとんどには、base と instruct (または chat) の 2 つのバージョンがあります。例えば、
[`tiiuae/falcon-7b`](https://huggingface.co/tiiuae/falcon-7b) および [`tiiuae/falcon-7b-instruct`](https://huggingface.co/tiiuae/falcon-7b) -指示する)。
基本モデルは、最初のプロンプトが与えられたときにテキストを完成させるのには優れていますが、NLP タスクには理想的ではありません。
指示に従う必要がある場合、または会話で使用する場合に使用します。ここで、指示 (チャット) バージョンが登場します。
これらのチェックポイントは、命令と会話データに基づいて事前トレーニングされたベース バージョンをさらに微調整した結果です。
この追加の微調整により、多くの NLP タスクにとってより適切な選択肢になります。
[`tiiuae/falcon-7b-instruct`](https://huggingface.co/tiiuae/falcon-7b-instruct) で使用できるいくつかの簡単なプロンプトを示してみましょう。
いくつかの一般的な NLP タスクを解決します。
### NLP tasks
まず、環境をセットアップしましょう。
```bash
pip install -q transformers accelerate
```
次に、適切なパイプライン (`text_generation`) を使用してモデルをロードしましょう。
```python
>>> from transformers import pipeline, AutoTokenizer
>>> import torch
>>> torch.manual_seed(0) # doctest: +IGNORE_RESULT
>>> model = "tiiuae/falcon-7b-instruct"
>>> tokenizer = AutoTokenizer.from_pretrained(model)
>>> pipe = pipeline(
... "text-generation",
... model=model,
... tokenizer=tokenizer,
... dtype=torch.bfloat16,
... device_map="auto",
... )
```
<Tip>
Falcon モデルは `bfloat16` データ型を使用してトレーニングされたため、同じものを使用することをお勧めします。これには、最近の
CUDA のバージョンに準拠しており、最新のカードで最適に動作します。
</Tip>
パイプライン経由でモデルをロードしたので、プロンプトを使用して NLP タスクを解決する方法を見てみましょう。
#### Text classification
テキスト分類の最も一般的な形式の 1 つはセンチメント分析であり、「ポジティブ」、「ネガティブ」、「ネガティブ」などのラベルを割り当てます。
または、一連のテキストに対して「中立」です。与えられたテキスト (映画レビュー) を分類するようにモデルに指示するプロンプトを作成してみましょう。
まず指示を与え、次に分類するテキストを指定します。そのままにしておくのではなく、
応答の先頭にも追加します - `"Sentiment: "`:
```python
>>> torch.manual_seed(0) # doctest: +IGNORE_RESULT
>>> prompt = """Classify the text into neutral, negative or positive.
... Text: This movie is definitely one of my favorite movies of its kind. The interaction between respectable and morally strong characters is an ode to chivalry and the honor code amongst thieves and policemen.
... Sentiment:
... """
>>> sequences = pipe(
... prompt,
... max_new_tokens=10,
... )
>>> for seq in sequences:
... print(f"Result: {seq['generated_text']}")
Result: Classify the text into neutral, negative or positive.
Text: This movie is definitely one of my favorite movies of its kind. The interaction between respectable and morally strong characters is an ode to chivalry and the honor code amongst thieves and policemen.
Sentiment:
Positive
```
その結果、出力には、手順で提供したリストの分類ラベルが含まれており、それは正しいラベルです。
<Tip>
プロンプトに加えて、`max_new_tokens`パラメータを渡していることに気づくかもしれません。トークンの数を制御します。
モデルが生成します。これは、学習できる多くのテキスト生成パラメーターの 1 つです。
[テキスト生成戦略](../generation_strategies) ガイドを参照してください。
</Tip>
#### Named Entity Recognition
固有表現認識 (NER) は、テキスト内の人物、場所、組織などの固有表現を検索するタスクです。
プロンプトの指示を変更して、LLM にこのタスクを実行させましょう。ここでは`return_full_text = False`も設定しましょう
出力にプロンプトが含まれないようにします。
```python
>>> torch.manual_seed(1) # doctest: +IGNORE_RESULT
>>> prompt = """Return a list of named entities in the text.
... Text: The Golden State Warriors are an American professional basketball team based in San Francisco.
... Named entities:
... """
>>> sequences = pipe(
... prompt,
... max_new_tokens=15,
... return_full_text = False,
... )
>>> for seq in sequences:
... print(f"{seq['generated_text']}")
- Golden State Warriors
- San Francisco
```
ご覧のとおり、モデルは指定されたテキストから 2 つの名前付きエンティティを正しく識別しました。
#### Translation
LLM が実行できるもう 1 つのタスクは翻訳です。このタスクにはエンコーダー/デコーダー モデルを使用することを選択できますが、ここでは
例を簡単にするために、きちんとした仕事をする Falcon-7b-instruct を使い続けます。もう一度、方法は次のとおりです
テキストの一部を英語からイタリア語に翻訳するようにモデルに指示する基本的なプロンプトを作成できます。
```python
>>> torch.manual_seed(2) # doctest: +IGNORE_RESULT
>>> prompt = """Translate the English text to Italian.
... Text: Sometimes, I've believed as many as six impossible things before breakfast.
... Translation:
... """
>>> sequences = pipe(
... prompt,
... max_new_tokens=20,
... do_sample=True,
... top_k=10,
... return_full_text = False,
... )
>>> for seq in sequences:
... print(f"{seq['generated_text']}")
A volte, ho creduto a sei impossibili cose prima di colazione.
```
ここでは、出力生成時にモデルがもう少し柔軟になるように `do_sample=True` と `top_k=10` を追加しました。
#### Text summarization
翻訳と同様に、テキストの要約も、出力が入力に**大きく**依存する生成タスクです。
エンコーダ/デコーダ モデルの方が良い選択になる可能性があります。ただし、デコーダ スタイルのモデルもこのタスクに使用できます。
以前は、プロンプトの先頭に指示を配置していました。ただし、プロンプトの最後で、
指示を与えるのに適した場所でもあります。通常、命令はどちらかの端に配置することをお勧めします。
```python
>>> torch.manual_seed(3) # doctest: +IGNORE_RESULT
>>> prompt = """Permaculture is a design process mimicking the diversity, functionality and resilience of natural ecosystems. The principles and practices are drawn from traditional ecological knowledge of indigenous cultures combined with modern scientific understanding and technological innovations. Permaculture design provides a framework helping individuals and communities develop innovative, creative and effective strategies for meeting basic needs while preparing for and mitigating the projected impacts of climate change.
... Write a summary of the above text.
... Summary:
... """
>>> sequences = pipe(
... prompt,
... max_new_tokens=30,
... do_sample=True,
... top_k=10,
... return_full_text = False,
... )
>>> for seq in sequences:
... print(f"{seq['generated_text']}")
Permaculture is an ecological design mimicking natural ecosystems to meet basic needs and prepare for climate change. It is based on traditional knowledge and scientific understanding.
```
#### Question answering
質問応答タスクの場合、プロンプトを次の論理コンポーネントに構造化できます: 指示、コンテキスト、質問、
先頭の単語またはフレーズ (`"Answer:"`) を使用して、モデルを操作して答えの生成を開始します。
```python
>>> torch.manual_seed(4) # doctest: +IGNORE_RESULT
>>> prompt = """Answer the question using the context below.
... Context: Gazpacho is a cold soup and drink made of raw, blended vegetables. Most gazpacho includes stale bread, tomato, cucumbers, onion, bell peppers, garlic, olive oil, wine vinegar, water, and salt. Northern recipes often include cumin and/or pimentón (smoked sweet paprika). Traditionally, gazpacho was made by pounding the vegetables in a mortar with a pestle; this more laborious method is still sometimes used as it helps keep the gazpacho cool and avoids the foam and silky consistency of smoothie versions made in blenders or food processors.
... Question: What modern tool is used to make gazpacho?
... Answer:
... """
>>> sequences = pipe(
... prompt,
... max_new_tokens=10,
... do_sample=True,
... top_k=10,
... return_full_text = False,
... )
>>> for seq in sequences:
... print(f"Result: {seq['generated_text']}")
Result: Modern tools are used, such as immersion blenders
```
#### Reasoning
LLM にとって推論は最も困難なタスクの 1 つであり、良い結果を達成するには、多くの場合、次のような高度なプロンプト テクニックを適用する必要があります。
[Chain-of-thought](#chain-of-thought)。
基本的なプロンプトを使用して、単純な算術タスクに関するモデル推論を作成できるかどうか試してみましょう。
```python
>>> torch.manual_seed(5) # doctest: +IGNORE_RESULT
>>> prompt = """There are 5 groups of students in the class. Each group has 4 students. How many students are there in the class?"""
>>> sequences = pipe(
... prompt,
... max_new_tokens=30,
... do_sample=True,
... top_k=10,
... return_full_text = False,
... )
>>> for seq in sequences:
... print(f"Result: {seq['generated_text']}")
Result:
There are a total of 5 groups, so there are 5 x 4=20 students in the class.
```
正しい！もう少し複雑さを増やして、基本的なプロンプトで問題を解決できるかどうかを確認してみましょう。
```python
>>> torch.manual_seed(6) # doctest: +IGNORE_RESULT
>>> prompt = """I baked 15 muffins. I ate 2 muffins and gave 5 muffins to a neighbor. My partner then bought 6 more muffins and ate 2. How many muffins do we now have?"""
>>> sequences = pipe(
... prompt,
... max_new_tokens=10,
... do_sample=True,
... top_k=10,
... return_full_text = False,
... )
>>> for seq in sequences:
... print(f"Result: {seq['generated_text']}")
Result:
The total number of muffins now is 21
```
これは間違った答えです。12 である必要があります。この場合、プロンプトが基本的すぎるか、選択内容が原因である可能性があります。
結局のところ、Falcon の最小バージョンを選択しました。あらゆるサイズのモデルでは推論が困難ですが、より大きなモデルでは
モデルのパフォーマンスが向上する可能性があります。
## Best practices of LLM prompting
ガイドのこのセクションでは、プロンプトの結果を改善する傾向にあるベスト プラクティスのリストをまとめました。
* 使用するモデルを選択する場合は、最新かつ最も機能的なモデルの方がパフォーマンスが向上する可能性があります。
* シンプルで短いプロンプトから始めて、そこから繰り返します。
* 指示はプロンプトの最初または最後に入力してください。大規模なコンテキストを扱う場合、モデルはさまざまな最適化を適用して、アテンションの複雑さが二次的に拡大するのを防ぎます。これにより、モデルはプロンプトの途中よりも最初または最後に注意を払うようになります。
* 指示と、それが適用されるテキストを明確に区別してください。これについては、次のセクションで詳しく説明します。
* タスクと望ましい結果 (その形式、長さ、スタイル、言語など) について具体的かつ説明的にします。
* 曖昧な説明や指示は避けてください。
*「何をしてはいけないか」という指示ではなく、「何をすべきか」という指示を優先します。
* 最初の単語を書いて (またはモデルの最初の文を始めて)、出力を正しい方向に「導き」ます。
* [Few-shot prompting](#few-shot-prompting) や [Chain-of-thought](#chain-of-thought) などの高度なテクニックを使用します。
* さまざまなモデルでプロンプトをテストして、その堅牢性を評価します。
* プロンプトのバージョンを確認し、パフォーマンスを追跡します。
## Advanced prompting techniques
### Few-shot prompting
上記のセクションの基本的なプロンプトは、「ゼロショット」プロンプトの例です。つまり、モデルにはすでに与えられています。
指示とコンテキストはありますが、解決策を含む例はありません。通常、命令データセットに基づいて微調整された LLM
このような「ゼロショット」タスクでも優れたパフォーマンスを発揮します。ただし、タスクがより複雑であったり微妙な点があったりする場合があります。
出力には、命令だけではモデルが理解できないいくつかの要件があります。この場合、次のことができます。
少数ショット プロンプトと呼ばれるテクニックを試してください。
少数ショット プロンプトでは、モデルにパフォーマンスを向上させるためのより多くのコンテキストを提供するプロンプト内の例が提供されます。
例では、例のパターンに従って出力を生成するようにモデルを条件付けします。
以下に例を示します。
```python
>>> torch.manual_seed(0) # doctest: +IGNORE_RESULT
>>> prompt = """Text: The first human went into space and orbited the Earth on April 12, 1961.
... Date: 04/12/1961
... Text: The first-ever televised presidential debate in the United States took place on September 28, 1960, between presidential candidates John F. Kennedy and Richard Nixon.
... Date:"""
>>> sequences = pipe(
... prompt,
... max_new_tokens=8,
... do_sample=True,
... top_k=10,
... )
>>> for seq in sequences:
... print(f"Result: {seq['generated_text']}")
Result: Text: The first human went into space and orbited the Earth on April 12, 1961.
Date: 04/12/1961
Text: The first-ever televised presidential debate in the United States took place on September 28, 1960, between presidential candidates John F. Kennedy and Richard Nixon.
Date: 09/28/1960
```
上記のコード スニペットでは、モデルへの目的の出力を示すために 1 つの例を使用しました。したがって、これは、
「ワンショット」プロンプト。ただし、タスクの複雑さに応じて、複数の例を使用する必要がある場合があります。
数回のプロンプト手法の制限:
- LLM は例のパターンを理解できますが、これらの手法は複雑な推論タスクではうまく機能しません。
- 少数ショットのプロンプトでは、長いプロンプトを作成する必要があります。大量のトークンを含むプロンプトでは、計算量と待ち時間が増加する可能性があります。プロンプトの長さにも制限があります。
- 多くの例を与えると、モデルが学習するつもりのなかったパターンを学習することがあります。 3番目の映画レビューはいつも否定的だということ。
### Chain-of-thought
思考連鎖 (CoT) プロンプトは、モデルを微調整して中間推論ステップを生成し、改善する手法です。
複雑な推論タスクの結果。
モデルを操作して推論ステップを生成するには、2 つの方法があります。
- 質問に対する詳細な回答を含む例を示し、問題に対処する方法をモデルに示すことで、数回のプロンプトを表示します。
- 「ステップごとに考えてみましょう」または「深呼吸して、問題をステップごとに解決してください」などのフレーズを追加してモデルに推論を指示します。
[推論セクション](#reasoning) のマフィンの例に CoT テクニックを適用し、より大きなモデルを使用すると、
[HuggingChat](https://huggingface.co/chat/)で遊べる(`tiiuae/falcon-180B-chat`)など、
推論結果は大幅に改善されます。
```text
Let's go through this step-by-step:
1. You start with 15 muffins.
2. You eat 2 muffins, leaving you with 13 muffins.
3. You give 5 muffins to your neighbor, leaving you with 8 muffins.
4. Your partner buys 6 more muffins, bringing the total number of muffins to 14.
5. Your partner eats 2 muffins, leaving you with 12 muffins.
Therefore, you now have 12 muffins.
```
## Prompting vs fine-tuning
プロンプトを最適化することで優れた結果を達成できますが、モデルを微調整するかどうかについてはまだ思案するかもしれません。
あなたの場合にはもっとうまくいくでしょう。より小規模なモデルを微調整することが好ましいオプションである場合のいくつかのシナリオを次に示します。
- ドメインが LLM が事前にトレーニングされたものと大きく異なっており、広範なプロンプト最適化では十分な結果が得られませんでした。
- モデルが低リソース言語で適切に動作する必要があります。
- 厳格な規制の下にある機密データでモデルをトレーニングする必要があります。
- コスト、プライバシー、インフラストラクチャ、またはその他の制限により、小規模なモデルを使用する必要があります。
上記のすべての例で、十分な大きさのファイルをすでに持っているか、簡単に入手できるかを確認する必要があります。
ドメイン固有のデータセットを合理的なコストでモデルを微調整できます。十分な時間とリソースも必要になります
モデルを微調整します。
上記の例が当てはまらない場合は、プロンプトを最適化する方が有益であることがわかります。
| transformers/docs/source/ja/tasks/prompting.md/0 | {
"file_path": "transformers/docs/source/ja/tasks/prompting.md",
"repo_id": "transformers",
"token_count": 9973
} | 391 |
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Export to TorchScript
<Tip>
これはTorchScriptを使用した実験の最初であり、可変入力サイズのモデルに対するその能力をまだ探求中です。これは私たちの関心の焦点であり、今後のリリースでは、より柔軟な実装や、PythonベースのコードとコンパイルされたTorchScriptを比較するベンチマークを含む、より多くのコード例で詳細な分析を行います。
</Tip>
[TorchScriptのドキュメント](https://pytorch.org/docs/stable/jit.html)によれば：
> TorchScriptは、PyTorchコードから直列化および最適化可能なモデルを作成する方法です。
TorchScriptを使用すると、効率志向のC++プログラムなど、他のプログラムでモデルを再利用できるようになります。PyTorchベースのPythonプログラム以外の環境で🤗 Transformersモデルをエクスポートして使用するためのインターフェースを提供しています。ここでは、TorchScriptを使用してモデルをエクスポートし、使用する方法を説明します。
モデルをエクスポートするには、次の2つの要件があります：
- `torchscript`フラグを使用したモデルのインスタンス化
- ダミーの入力を使用したフォワードパス
これらの必要条件は、以下で詳細に説明されているように、開発者が注意する必要があるいくつかのことを意味します。
## TorchScript flag and tied weights
`torchscript`フラグは、ほとんどの🤗 Transformers言語モデルにおいて、`Embedding`レイヤーと`Decoding`レイヤー間で重みが連結されているため必要です。
TorchScriptでは、重みが連結されているモデルをエクスポートすることはできませんので、事前に重みを切り離して複製する必要があります。
`torchscript`フラグを使用してインスタンス化されたモデルは、`Embedding`レイヤーと`Decoding`レイヤーが分離されており、そのため後でトレーニングしてはいけません。
トレーニングは、これらの2つのレイヤーを非同期にする可能性があり、予期しない結果をもたらす可能性があります。
言語モデルヘッドを持たないモデルには言及しませんが、これらのモデルには連結された重みが存在しないため、`torchscript`フラグなしで安全にエクスポートできます。
## Dummy inputs and standard lengths
ダミー入力はモデルのフォワードパスに使用されます。入力の値はレイヤーを通じて伝播される間、PyTorchは各テンソルに実行された異なる操作を追跡します。これらの記録された操作は、モデルの*トレース*を作成するために使用されます。
トレースは入力の寸法に対して作成されます。そのため、ダミー入力の寸法に制約され、他のシーケンス長やバッチサイズでは動作しません。異なるサイズで試すと、以下のエラーが発生します：
```
`The expanded size of the tensor (3) must match the existing size (7) at non-singleton dimension 2`
```
お勧めしますのは、モデルの推論中に供給される最大の入力と同じ大きさのダミー入力サイズでモデルをトレースすることです。パディングを使用して不足値を補完することもできます。ただし、モデルがより大きな入力サイズでトレースされるため、行列の寸法も大きくなり、より多くの計算が発生します。
異なるシーケンス長のモデルをエクスポートする際に、各入力に対して実行される演算の総数に注意して、パフォーマンスを密接にフォローすることをお勧めします。
## Using TorchScript in Python
このセクションでは、モデルの保存と読み込み、および推論にトレースを使用する方法を示します。
### Saving a model
TorchScriptで`BertModel`をエクスポートするには、`BertConfig`クラスから`BertModel`をインスタンス化し、それをファイル名`traced_bert.pt`でディスクに保存します：
```python
from transformers import BertModel, BertTokenizer, BertConfig
import torch
enc = BertTokenizer.from_pretrained("google-bert/bert-base-uncased")
# Tokenizing input text
text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
tokenized_text = enc.tokenize(text)
# Masking one of the input tokens
masked_index = 8
tokenized_text[masked_index] = "[MASK]"
indexed_tokens = enc.convert_tokens_to_ids(tokenized_text)
segments_ids = [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1]
# Creating a dummy input
tokens_tensor = torch.tensor([indexed_tokens])
segments_tensors = torch.tensor([segments_ids])
dummy_input = [tokens_tensor, segments_tensors]
# Initializing the model with the torchscript flag
# Flag set to True even though it is not necessary as this model does not have an LM Head.
config = BertConfig(
vocab_size_or_config_json_file=32000,
hidden_size=768,
num_hidden_layers=12,
num_attention_heads=12,
intermediate_size=3072,
torchscript=True,
)
# Instantiating the model
model = BertModel(config)
# The model needs to be in evaluation mode
model.eval()
# If you are instantiating the model with *from_pretrained* you can also easily set the TorchScript flag
model = BertModel.from_pretrained("google-bert/bert-base-uncased", torchscript=True)
# Creating the trace
traced_model = torch.jit.trace(model, [tokens_tensor, segments_tensors])
torch.jit.save(traced_model, "traced_bert.pt")
```
### Loading a model
以前に保存した `BertModel`、`traced_bert.pt` をディスクから読み込んで、以前に初期化した `dummy_input` で使用できます。
```python
loaded_model = torch.jit.load("traced_bert.pt")
loaded_model.eval()
all_encoder_layers, pooled_output = loaded_model(*dummy_input)
```
### Using a traced model for inference
トレースモデルを使用して推論を行うには、その `__call__` ダンダーメソッドを使用します。
```python
traced_model(tokens_tensor, segments_tensors)
```
## Deploy Hugging Face TorchScript models to AWS with the Neuron SDK
AWSはクラウドでの低コストで高性能な機械学習推論向けに [Amazon EC2 Inf1](https://aws.amazon.com/ec2/instance-types/inf1/) インスタンスファミリーを導入しました。Inf1インスタンスはAWS Inferentiaチップによって駆動され、ディープラーニング推論ワークロードに特化したカスタムビルドのハードウェアアクセラレータです。[AWS Neuron](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/#) はInferentia用のSDKで、トランスフォーマーモデルをトレースして最適化し、Inf1に展開するためのサポートを提供します。
Neuron SDK が提供するもの:
1. クラウドでの推論のためにTorchScriptモデルをトレースして最適化するための、1行のコード変更で使用できる簡単なAPI。
2. [改善されたコストパフォーマンス](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/benchmark/) のためのボックス外のパフォーマンス最適化。
3. [PyTorch](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/pytorch/bert_tutorial/tutorial_pretrained_bert.html) または [TensorFlow](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/tensorflow/huggingface_bert/huggingface_bert.html) で構築されたHugging Faceトランスフォーマーモデルへのサポート。
### Implications
BERT（Bidirectional Encoder Representations from Transformers）アーキテクチャやその変種（[distilBERT](https://huggingface.co/docs/transformers/main/model_doc/distilbert) や [roBERTa](https://huggingface.co/docs/transformers/main/model_doc/roberta) など）に基づくトランスフォーマーモデルは、非生成タスク（抽出型質問応答、シーケンス分類、トークン分類など）において、Inf1上で最適に動作します。ただし、テキスト生成タスクも [AWS Neuron MarianMT チュートリアル](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/pytorch/transformers-marianmt.html) に従ってInf1上で実行できます。Inferentiaでボックス外で変換できるモデルに関する詳細情報は、Neuronドキュメンテーションの [Model Architecture Fit](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/models/models-inferentia.html#models-inferentia) セクションにあります。
### Dependencies
モデルをAWS Neuronに変換するには、[Neuron SDK 環境](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/neuron-frameworks/pytorch-neuron/index.html#installation-guide) が必要で、[AWS Deep Learning AMI](https://docs.aws.amazon.com/dlami/latest/devguide/tutorial-inferentia-launching.html) に事前に構成されています。
### Converting a model for AWS Neuron
モデルをAWS NEURON用に変換するには、[PythonでTorchScriptを使用する](torchscript#using-torchscript-in-python) と同じコードを使用して `BertModel` をトレースします。Python APIを介してNeuron SDKのコンポーネントにアクセスするために、`torch.neuron` フレームワーク拡張をインポートします。
```python
from transformers import BertModel, BertTokenizer, BertConfig
import torch
import torch.neuron
```
次の行を変更するだけで済みます。
```diff
- torch.jit.trace(model, [tokens_tensor, segments_tensors])
+ torch.neuron.trace(model, [token_tensor, segments_tensors])
```
これにより、Neuron SDKはモデルをトレースし、Inf1インスタンス向けに最適化します。
AWS Neuron SDKの機能、ツール、サンプルチュートリアル、最新のアップデートについて詳しく知りたい場合は、[AWS NeuronSDK ドキュメンテーション](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/index.html) をご覧ください。
| transformers/docs/source/ja/torchscript.md/0 | {
"file_path": "transformers/docs/source/ja/torchscript.md",
"repo_id": "transformers",
"token_count": 4350
} | 392 |
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# DeepSpeed[[deepspeed]]
[DeepSpeed](https://www.deepspeed.ai/)는 분산 학습 메모리를 효율적이고 빠르게 만드는 PyTorch 최적화 라이브러리입니다. 그 핵심은 대규모 모델을 규모에 맞게 훈련할 수 있는 [Zero Redundancy Optimizer(ZeRO)](https://hf.co/papers/1910.02054)입니다. ZeRO는 여러 단계로 작동합니다:
* ZeRO-1, GPU 간 최적화 상태 분할
* ZeRO-2, GPU 간 그레이디언트 분할
* ZeRO-3, GPU 간 매개변수 분할
GPU가 제한된 환경에서 ZeRO는 최적화 메모리와 계산을 GPU에서 CPU로 오프로드하여 단일 GPU에 대규모 모델을 장착하고 훈련할 수 있습니다. DeepSpeed는 모든 ZeRO 단계 및 오프로딩을 위해 Transformers [`Trainer`] 클래스와 통합되어 있습니다. 구성 파일을 제공하거나 제공된 템플릿을 사용하기만 하면 됩니다. 추론의 경우, Transformers는 대용량 모델을 가져올 수 있으므로 ZeRO-3 및 오프로딩을 지원합니다.
이 가이드에서는 DeepSpeed 트레이닝을 배포하는 방법, 활성화할 수 있는 기능, 다양한 ZeRO 단계에 대한 구성 파일 설정 방법, 오프로딩, 추론 및 [`Trainer`] 없이 DeepSpeed를 사용하는 방법을 안내해 드립니다.
## 설치[[installation]]
DeepSpeed는 PyPI 또는 Transformers에서 설치할 수 있습니다(자세한 설치 옵션은 DeepSpeed [설치 상세사항](https://www.deepspeed.ai/tutorials/advanced-install/) 또는 GitHub [README](https://github.com/deepspeedai/DeepSpeed#installation)를 참조하세요).
<Tip>
DeepSpeed를 설치하는 데 문제가 있는 경우 [DeepSpeed CUDA 설치](../debugging#deepspeed-cuda-installation) 가이드를 확인하세요. DeepSpeed에는 pip 설치 가능한 PyPI 패키지로 설치할 수 있지만, 하드웨어에 가장 잘 맞고 PyPI 배포판에서는 제공되지 않는 1비트 Adam과 같은 특정 기능을 지원하려면 [소스에서 설치하기](https://www.deepspeed.ai/tutorials/advanced-install/#install-deepspeed-from-source)를 적극 권장합니다.
</Tip>
<hfoptions id="install">
<hfoption id="PyPI">
```bash
pip install deepspeed
```
</hfoption>
<hfoption id="Transformers">
```bash
pip install transformers[deepspeed]
```
</hfoption>
</hfoptions>
## 메모리 요구량[[memory-requirements]]
시작하기 전에 모델에 맞는 충분한 GPU 및 CPU 메모리가 있는지 확인하는 것이 좋습니다. DeepSpeed는 필요한 CPU/GPU 메모리를 추정할 수 있는 도구를 제공합니다. 예를 들어, 단일 GPU에서 [bigscience/T0_3B](bigscience/T0_3B) 모델의 메모리 요구 사항을 추정할 수 있습니다:
```bash
$ python -c 'from transformers import AutoModel; \
from deepspeed.runtime.zero.stage3 import estimate_zero3_model_states_mem_needs_all_live; \
model = AutoModel.from_pretrained("bigscience/T0_3B"); \
estimate_zero3_model_states_mem_needs_all_live(model, num_gpus_per_node=1, num_nodes=1)'
[...]
Estimated memory needed for params, optim states and gradients for a:
HW: Setup with 1 node, 1 GPU per node.
SW: Model with 2783M total params, 65M largest layer params.
per CPU | per GPU | Options
70.00GB | 0.25GB | offload_param=cpu , offload_optimizer=cpu , zero_init=1
70.00GB | 0.25GB | offload_param=cpu , offload_optimizer=cpu , zero_init=0
62.23GB | 5.43GB | offload_param=none, offload_optimizer=cpu , zero_init=1
62.23GB | 5.43GB | offload_param=none, offload_optimizer=cpu , zero_init=0
0.37GB | 46.91GB | offload_param=none, offload_optimizer=none, zero_init=1
15.56GB | 46.91GB | offload_param=none, offload_optimizer=none, zero_init=0
```
즉, CPU 오프로드가 없는 단일 80GB GPU 또는 오프로드 할 8GB GPU와 최대 60GB CPU가 필요합니다 (이는 매개변수, 최적화 상태 및 그레이디언트에 대한 메모리 요구 사항일 뿐이며 CUDA 커널 및 활성화에는 조금 더 필요합니다). 또한 더 작은 GPU를 대여하거나 구입하는 것이 더 저렴하지만 모델을 훈련하는 데 시간이 더 오래 걸리므로 비용과 속도 간의 균형을 고려해야 합니다.
GPU 메모리가 충분하다면 CPU/NVMe 오프로드를 비활성화하여 모든 작업을 더 빠르게 처리하세요.
## ZeRO 단계 설정하기[[select-a-zero-stage]]
DeepSpeed를 설치하고 메모리 요구 사항을 더 잘 파악했다면 다음 단계는 사용할 ZeRO 스테이지를 선택하는 것입니다. 가장 빠르고 메모리 효율이 높은 순서대로 정렬하면 다음과 같습니다:
| 속도 | 메모리 효율 |
|------------------|------------------|
| ZeRO-1 | ZeRO-3 + offload |
| ZeRO-2 | ZeRO-3 |
| ZeRO-2 + offload | ZeRO-2 + offload |
| ZeRO-3 | ZeRO-2 |
| ZeRO-3 + offload | ZeRO-1 |
자신에게 가장 적합한 방법을 찾으려면 가장 빠른 방법부터 시작하고 메모리가 부족하면 더 느리지만 메모리 효율이 높은 다음 단계를 시도하세요. 속도와 메모리 사용량 사이의 적절한 균형을 찾기 위해 (가장 메모리 효율적이거나 가장 빠른 것부터 시작하여) 원하는 방향으로 자유롭게 작업하세요.
일반적으로 사용할 수 있는 프로세스는 다음과 같습니다(배치 크기 1로 시작):
1. 그레이디언트 체크포인팅 활성화
2. ZeRO-2 시도
3. ZeRO-2와 매개변수 오프로드 시도
4. ZeRO-3 시도
5. ZeRO-3과 매개변수 CPU 오프로드 시도
6. ZeRO-3, 매개변수와 옵티마이저 CPU 오프로드 시도
7. [`~GenerationMixin.generate`] 메소드를 사용하는 경우 더 좁은 빔 서치 검색 범위와 같은 다양한 기본값을 낮춰보기
8. 전체 정밀도 가중치보다 반정밀도(구형 GPU 구조의 경우 fp16, 암페어 이후 GPU의 경우 bf16)를 혼합해보기
9. 가능하면 하드웨어를 더 추가하거나 Infinity가 매개변수와 옵티마이저를 NVMe로 오프로드하도록 활성화
10. 메모리가 부족하지 않으면 유효 처리량을 측정한 다음 배치 크기를 최대한 크게 늘려 GPU 효율성을 극대화
11. 마지막으로 일부 오프로드 기능을 비활성화하거나 더 빠른 ZeRO 스테이지를 사용하고 배치 크기를 늘리거나 줄여 속도와 메모리 사용량 간의 최적의 균형을 찾아 트레이닝 설정을 최적화
## DeepSpeed 구성 파일[[deepspeed-configuration-file]]
DeepSpeed는 트레이닝 실행 방법을 구성하는 모든 매개변수가 포함된 구성 파일을 통해 [`Trainer`] 클래스와 함께 작동합니다. 트레이닝 스크립트를 실행하면 DeepSpeed는 [`Trainer`]로부터 받은 구성을 콘솔에 기록하므로 어떤 구성이 사용되었는지 정확히 확인할 수 있습니다.
<Tip>
DeepSpeed 구성 옵션의 전체 목록은 [DeepSpeed Configuration JSON](https://www.deepspeed.ai/docs/config-json/)에서 확인할 수 있습니다. 또한 [DeepSpeedExamples](https://github.com/deepspeedai/DeepSpeedExamples) 리포지토리 또는 기본 [DeepSpeed](https://github.com/deepspeedai/DeepSpeed) 리포지토리에서 다양한 DeepSpeed 구성 예제에 대한 보다 실용적인 예제를 찾을 수 있습니다. 구체적인 예제를 빠르게 찾으려면 다음과 같이 하세요:
```bash
git clone https://github.com/deepspeedai/DeepSpeedExamples
cd DeepSpeedExamples
find . -name '*json'
# Lamb 옵티마이저 샘플 찾기
grep -i Lamb $(find . -name '*json')
```
</Tip>
명령줄 인터페이스에서 트레이닝하는 경우 DeepSpeed 구성 파일은 JSON 파일의 경로로 전달되거나 노트북 설정에서 [`Trainer`]를 사용하는 경우 중첩된 `dict` 객체로 전달됩니다.
<hfoptions id="pass-config">
<hfoption id="path to file">
```py
TrainingArguments(..., deepspeed="path/to/deepspeed_config.json")
```
</hfoption>
<hfoption id="nested dict">
```py
ds_config_dict = dict(scheduler=scheduler_params, optimizer=optimizer_params)
args = TrainingArguments(..., deepspeed=ds_config_dict)
trainer = Trainer(model, args, ...)
```
</hfoption>
</hfoptions>
### DeepSpeed와 Trainer 매개변수[[deepspeed-and-trainer-parameters]]
구성 매개변수에는 세 가지 유형이 있습니다:
1. 일부 구성 매개변수는 [`Trainer`]와 DeepSpeed가 공유하며, 정의가 충돌하는 경우 오류를 식별하기 어려울 수 있습니다. 이러한 공유 구성 매개변수는 [`Trainer`] 명령줄 인수에서 쉽게 설정할 수 있습니다.
2. 모델 설정에서 자동으로 도출되는 일부 설정 매개변수는 수동으로 값을 조정할 필요가 없습니다. [`Trainer`]는 구성 값 `auto`를 사용하여 가장 정확하거나 효율적인 값을 설정합니다. 직접 구성 매개변수를 명시적으로 설정할 수도 있지만, [`Trainer`] 인수와 DeepSpeed 설정 매개변수가 일치하도록 주의해야 합니다. 일치하지 않으면 감지하기 매우 어려운 방식으로 훈련이 실패할 수 있습니다!
3. 교육 요구 사항에 따라 수동으로 설정해야 하는 일부 설정 매개변수는 DeepSpeed에만 해당됩니다.
DeepSpeed 구성을 수정하고 [`TrainingArguments`]를 편집할 수도 있습니다:
1. 기본 구성으로 사용할 DeepSpeed 구성 파일을 생성하거나 로드합니다.
2. 다음 DeepSpeed 구성을 기반으로 [`TrainingArguments`] 객체를 생성합니다.
`scheduler.params.total_num_steps`와 같은 일부 값은 트레이닝 중 [`Trainer`]에 의해 계산됩니다.
### ZeRO 구성[[zero-configuration]]
세 가지 구성이 있으며, 각 구성은 서로 다른 ZeRO 단계에 해당합니다. 1단계는 확장성 측면에서 그다지 눈여겨볼만하지 않으므로 이 가이드에서는 2단계와 3단계에 중점을 둡니다. `zero_optimization` 구성에는 활성화할 항목과 구성 방법에 대한 모든 옵션이 포함되어 있습니다. 각 매개변수에 대한 자세한 설명은 [DeepSpeed 구성 JSON](https://www.deepspeed.ai/docs/config-json/) 참조를 참조하세요.
<Tip warning={true}>
DeepSpeed는 매개변수 이름의 유효성을 검사하지 않으며 오타가 있으면 매개변수의 기본 설정으로 대체합니다. DeepSpeed 엔진 시작 로그 메시지를 보고 어떤 값을 사용할지 확인할 수 있습니다.
</Tip>
[`Trainer`]는 동등한 명령줄 인수를 제공하지 않으므로 다음 구성은 DeepSpeed로 설정해야 합니다.
<hfoptions id="zero-config">
<hfoption id="ZeRO-1">
ZeRO-1은 옵티마이저 상태를 GPU에 분할하여 약간의 속도 향상을 기대할 수 있습니다. ZeRO-1 구성은 다음과 같이 설정할 수 있습니다:
```yml
{
"zero_optimization": {
"stage": 1
}
}
```
</hfoption>
<hfoption id="ZeRO-2">
ZeRO-2는 GPU에서 옵티마이저와 그레이디언트를 분할합니다. 이 단계는 추론과 관련이 없는 기능이기 때문에 주로 훈련에 사용됩니다. 더 나은 성능을 위해 구성해야 할 몇 가지 중요한 매개변수는 다음과 같습니다:
* GPU 메모리 사용량을 줄이려면 `offload_optimizer`를 활성화해야 합니다.
* `true`로 설정된 경우 `overlap_comm`은 GPU 메모리 사용량 증가를 상쇄하여 지연 시간을 줄입니다. 이 기능은 4.5배의 `allgather_bucket_size` 및 `reduce_bucket_size`값을 사용합니다. 이 예에서는 `5e8`로 설정되어 있으므로 9GB의 GPU 메모리가 필요합니다. GPU 메모리가 8GB 이하인 경우, 메모리 요구량을 낮추고 메모리 부족(OOM) 오류를 방지하기 위해 `overlap_comm`을 줄여야 합니다.
* `allgather_bucket_size`와 `reduce_bucket_size`는 사용 가능한 GPU 메모리와 통신 속도를 절충합니다. 값이 작을수록 통신 속도가 느려지고 더 많은 GPU 메모리를 사용할 수 있습니다. 예를 들어, 배치 크기가 큰 것이 약간 느린 훈련 시간보다 더 중요한지 균형을 맞출 수 있습니다.
* DeepSpeed 0.4.4에서는 CPU 오프로딩을 위해 `round_robin_gradients`를 사용할 수 있습니다. 이 기능은 세분화된 그레이디언트 파티셔닝을 통해 등급 간 그레이디언트 복사를 CPU 메모리로 병렬화합니다. 성능 이점은 그레이디언트 누적 단계(최적화 단계 간 복사 횟수 증가) 또는 GPU 수(병렬 처리 증가)에 따라 증가합니다.
```yml
{
"zero_optimization": {
"stage": 2,
"offload_optimizer": {
"device": "cpu",
"pin_memory": true
},
"allgather_partitions": true,
"allgather_bucket_size": 5e8,
"overlap_comm": true,
"reduce_scatter": true,
"reduce_bucket_size": 5e8,
"contiguous_gradients": true
"round_robin_gradients": true
}
}
```
</hfoption>
<hfoption id="ZeRO-3">
ZeRO-3는 옵티마이저, 그래디언트, 매개변수를 여러 GPU에 걸쳐 분할합니다. ZeRO-2와 달리 ZeRO-3는 여러 GPU에 대규모 모델을 가져올 수 있기 때문에 훈련 외에도 추론에도 사용할 수 있습니다. 구성해야 할 몇 가지 중요한 매개변수는 다음과 같습니다:
* `device: "cpu"` 는 GPU 메모리가 부족하고 사용 가능한 CPU 메모리가 있는 경우 도움이 될 수 있습니다. 이를 통해 모델 매개변수를 CPU로 오프로드할 수 있습니다.
* `pin_memory: true` 는 처리량을 향상시킬 수 있지만, 핀 메모리는 메모리를 요청한 특정 프로세스를 위해 예약되어 있고 일반적으로 일반 CPU 메모리보다 훨씬 빠르게 액세스되기 때문에 다른 프로세스에서 사용할 수 있는 메모리가 줄어듭니다.
* `stage3_max_live_parameters` 는 특정 시간에 GPU에 유지하려는 전체 매개변수의 상한값입니다. OOM 오류가 발생하면 이 값을 줄이세요.
* `stage3_max_reuse_distance` 는 향후 매개변수를 다시 사용할 시기를 결정하는 값으로, 매개변수를 버릴지 유지할지 결정하는 데 도움이 됩니다. 매개변수를 재사용할 경우(`stage3_max_reuse_distance`보다 작은 값인 경우) 통신 오버헤드를 줄이기 위해 매개변수를 유지합니다. 이 기능은 활성화 체크포인팅이 활성화되어 있고 역전파 계산시까지 순전파 시점의 매개변수를 유지하려는 경우에 매우 유용합니다. 그러나 OOM 오류가 발생하면 이 값을 줄이세요.
* 모델 저장 시 `stage3_gather_16bit_weights_on_model_save`는 fp16 가중치를 통합합니다. 대규모 모델을 학습하거나 여러 GPU를 사용할 경우 메모리와 속도 측면에서 비용이 많이 듭니다. 훈련을 재개할 계획이라면 이 옵션을 활성화해야 합니다.
* `sub_group_size` 는 최적화 단계에서 업데이트되는 매개변수를 제어합니다. 매개변수는 `sub_group_size`의 버킷으로 그룹화되며 각 버킷은 한 번에 하나씩 업데이트됩니다. NVMe 오프로드와 함께 사용하는 경우 `sub_group_size`는 최적화 단계 중 모델 상태가 CPU 메모리로 이동하는 시점을 결정합니다. 이렇게 하면 매우 큰 모델의 CPU 메모리 부족을 방지할 수 있습니다. NVMe 오프로드를 사용하지 않는 경우 `sub_group_size`를 기본값으로 둘 수 있지만, 사용하는 경우 변경하는 것이 좋습니다:
1. 옵티마이저 단계에서 OOM 오류가 발생합니다. 이 경우, 임시 버퍼의 메모리 사용량을 줄이려면 `sub_group_size`를 줄이세요.
2. 옵티마이저 단계에서 시간이 너무 오래 걸립니다. 이 경우 데이터 버퍼 증가로 인한 대역폭 사용률을 개선하기 위해 `sub_group_size`를 늘리세요.
* `reduce_bucket_size`, `stage3_prefetch_bucket_size`, `stage3_param_persistence_threshold`는 모델의 숨겨진 크기에 따라 달라집니다. 이 값들을 `auto`으로 설정하고 [`Trainer`]가 자동으로 값을 할당하도록 허용하는 것이 좋습니다.
```yml
{
"zero_optimization": {
"stage": 3,
"offload_optimizer": {
"device": "cpu",
"pin_memory": true
},
"offload_param": {
"device": "cpu",
"pin_memory": true
},
"overlap_comm": true,
"contiguous_gradients": true,
"sub_group_size": 1e9,
"reduce_bucket_size": "auto",
"stage3_prefetch_bucket_size": "auto",
"stage3_param_persistence_threshold": "auto",
"stage3_max_live_parameters": 1e9,
"stage3_max_reuse_distance": 1e9,
"stage3_gather_16bit_weights_on_model_save": true
}
}
```
[`deepspeed.zero.Init`](https://deepspeed.readthedocs.io/en/latest/zero3.html#deepspeed.zero.Init) 컨텍스트 매니저를 사용하면 모델을 더 빠르게 초기화할 수 있습니다:
```py
from transformers import T5ForConditionalGeneration, T5Config
import deepspeed
with deepspeed.zero.Init():
config = T5Config.from_pretrained("google-t5/t5-small")
model = T5ForConditionalGeneration(config)
```
사전 학습된 모델의 경우, 딥스피드 구성 파일에 `is_deepspeed_zero3_enabled: true`가 [`TrainingArguments`]에 설정되어 있어야 하며, ZeRO 구성이 활성화되어 있어야 합니다. 훈련된 모델 [`~PreTrainedModel.from_pretrained`]을 호출하기 **전에** [`TrainingArguments`] 객체를 생성해야 합니다.
```py
from transformers import AutoModel, Trainer, TrainingArguments
training_args = TrainingArguments(..., deepspeed=ds_config)
model = AutoModel.from_pretrained("google-t5/t5-small")
trainer = Trainer(model=model, args=training_args, ...)
```
fp16 가중치가 단일 GPU에 맞지 않는 경우 ZeRO-3이 필요합니다. fp16 가중치를 로드할 수 있는 경우, [`~PreTrainedModel.from_pretrained`]에 `dtype=torch.float16`을 지정해야 합니다.
ZeRO-3의 또 다른 고려 사항은 여러 개의 GPU를 사용하는 경우 현재 실행 중인 레이어의 매개변수가 아닌 한 단일 GPU에 모든 매개변수가 없다는 것입니다. 사전 훈련된 모델 가중치를 [`~PreTrainedModel.from_pretrained`]에 로드하는 등 모든 레이어의 모든 매개변수에 한 번에 액세스하려면 한 번에 하나의 레이어를 로드하고 즉시 모든 GPU에 파티셔닝합니다. 이는 매우 큰 모델의 경우 메모리 제한으로 인해 하나의 GPU에 가중치를 로드한 다음 다른 GPU에 분산할 수 없기 때문입니다.
다음과 같이 보이는 모델 매개변수 가중치(여기서 `tensor([1.])`) 또는 매개변수 크기가 더 큰 다차원 형태 대신 1인 경우, 이는 매개변수가 분할되어 있으며 이것이 ZeRO-3 플레이스홀더인 것을 의미합니다.
```py
tensor([1.0], device="cuda:0", dtype=torch.float16, requires_grad=True)
```
<Tip>
ZeRO-3로 대규모 모델을 초기화하고 매개변수에 액세스하는 방법에 대한 자세한 내용은 [Constructing Massive Models](https://deepspeed.readthedocs.io/en/latest/zero3.html#constructing-massive-models) 및 [Gathering Parameters](https://deepspeed.readthedocs.io/en/latest/zero3.html#gathering-parameters) 가이드를 참조하세요.
</Tip>
</hfoption>
</hfoptions>
### NVMe 설정[[nvme-configuration]]
[ZeRO-Infinity](https://hf.co/papers/2104.07857)를 사용하면 모델 상태를 CPU 및/또는 NVMe로 오프로드하여 더 많은 메모리를 절약할 수 있습니다. 스마트 파티셔닝 및 타일링 알고리즘을 통해 각 GPU는 오프로딩 중에 매우 적은 양의 데이터를 주고받을 수 있으므로 최신 NVMe는 훈련 프로세스에 사용할 수 있는 것보다 훨씬 더 큰 총 메모리 풀에 맞출 수 있습니다. ZeRO-Infinity에는 ZeRO-3가 필요합니다.
사용 가능한 CPU 및/또는 NVMe 메모리에 따라 [옵티마이저](https://www.deepspeed.ai/docs/config-json/#optimizer-offloading)와 [매개변수](https://www.deepspeed.ai/docs/config-json/#parameter-offloading) 중 하나만 오프로드하거나 아무것도 오프로드하지 않을 수 있습니다. 또한 일반 하드 드라이브나 솔리드 스테이트 드라이브에서도 작동하지만 속도가 현저히 느려지므로 `nvme_path`가 NVMe 장치를 가리키고 있는지 확인해야 합니다. 최신 NVMe를 사용하면 읽기 작업의 경우 최대 3.5GB/s, 쓰기 작업의 경우 최대 3GB/s의 전송 속도를 기대할 수 있습니다. 마지막으로, 트레이닝 설정에서 [벤치마크 실행하기](https://github.com/deepspeedai/DeepSpeed/issues/998)을 통해 최적의 'aio' 구성을 결정합니다.
아래 예제 ZeRO-3/Infinity 구성 파일은 대부분의 매개변수 값을 `auto`으로 설정하고 있지만, 수동으로 값을 추가할 수도 있습니다.
```yml
{
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"loss_scale_window": 1000,
"initial_scale_power": 16,
"hysteresis": 2,
"min_loss_scale": 1
},
"optimizer": {
"type": "AdamW",
"params": {
"lr": "auto",
"betas": "auto",
"eps": "auto",
"weight_decay": "auto"
}
},
"scheduler": {
"type": "WarmupLR",
"params": {
"warmup_min_lr": "auto",
"warmup_max_lr": "auto",
"warmup_num_steps": "auto"
}
},
"zero_optimization": {
"stage": 3,
"offload_optimizer": {
"device": "nvme",
"nvme_path": "/local_nvme",
"pin_memory": true,
"buffer_count": 4,
"fast_init": false
},
"offload_param": {
"device": "nvme",
"nvme_path": "/local_nvme",
"pin_memory": true,
"buffer_count": 5,
"buffer_size": 1e8,
"max_in_cpu": 1e9
},
"aio": {
"block_size": 262144,
"queue_depth": 32,
"thread_count": 1,
"single_submit": false,
"overlap_events": true
},
"overlap_comm": true,
"contiguous_gradients": true,
"sub_group_size": 1e9,
"reduce_bucket_size": "auto",
"stage3_prefetch_bucket_size": "auto",
"stage3_param_persistence_threshold": "auto",
"stage3_max_live_parameters": 1e9,
"stage3_max_reuse_distance": 1e9,
"stage3_gather_16bit_weights_on_model_save": true
},
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"steps_per_print": 2000,
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"wall_clock_breakdown": false
}
```
## DeepSpeed 구성[[deepspeed-features]]
이 섹션에서 간략하게 설명하는 몇 가지 중요한 매개변수를 DeepSpeed 구성 파일에 지정할 수 있습니다.
### 활성화/그레이디언트 체크포인팅[[activationgradient-checkpointing]]
활성화 및 그레이디언트 체크포인팅은 속도를 더 많은 GPU 메모리와 교환하여 GPU 메모리가 부족한 상황을 극복하거나 배치 크기를 늘려 성능을 향상시킬 수 있습니다. 이 기능을 활성화하려면 다음과 같이 하세요:
1. 허깅 페이스 모델의 경우, [`Trainer`]에서 `model.gradient_checkpointing_enable()` 또는 `--gradient_checkpointing`을 설정합니다.
2. 허깅 페이스가 아닌 모델의 경우, 딥스피드 [Activation Checkpointing API](https://deepspeed.readthedocs.io/en/latest/activation-checkpointing.html)를 사용합니다. 트랜스포머 모델링 코드를 대체하고 `torch.utils.checkpoint`를 DeepSpeed API로 대체할 수도 있습니다. 이 접근 방식은 순방향 활성화를 다시 계산하는 대신 CPU 메모리로 오프로드할 수 있으므로 더 유연합니다.
### 옵티마이저와 스케줄러[[optimizer-and-scheduler]]
`offload_optimizer`를 활성화하지 않는 한 DeepSpeed와 트랜스포머 옵티마이저 및 스케줄러를 혼합하여 사용할 수 있습니다. `offload_optimizer`를 활성화하면 CPU와 GPU 구현이 모두 있는 경우 DeepSpeed가 아닌 최적화기(LAMB 제외)를 사용할 수 있습니다.
<Tip warning={true}>
구성 파일의 최적화 프로그램 및 스케줄러 매개변수는 명령줄에서 설정할 수 있으므로 오류를 찾기 어렵지 않습니다. 예를 들어 학습 속도가 다른 곳에서 다른 값으로 설정된 경우 명령줄에서 이를 재정의할 수 있습니다. 최적화 프로그램 및 스케줄러 매개변수 외에도 [`Trainer`] 명령줄 인수가 DeepSpeed 구성과 일치하는지 확인해야 합니다.
</Tip>
<hfoptions id="opt-sched">
<hfoption id="optimizer">
DeepSpeed는 여러 [옵티마이저](https://www.deepspeed.ai/docs/config-json/#optimizer-parameters)를 제공하지만(Adam, AdamW, OneBitAdam 및 LAMB) PyTorch에서 다른 옵티마이저를 가져올 수도 있습니다. 설정에서 옵티마이저를 구성하지 않으면 [`Trainer`]가 자동으로 AdamW를 선택하고 명령줄에서 제공된 값 또는 기본값을 사용합니다: `lr`, `adam_beta1`, `adam_beta2`, `adam_epsilon`, `weight_decay`.
매개변수를 `"auto"`으로 설정하거나 원하는 값을 직접 수동으로 입력할 수 있습니다.
```yaml
{
"optimizer": {
"type": "AdamW",
"params": {
"lr": "auto",
"betas": "auto",
"eps": "auto",
"weight_decay": "auto"
}
}
}
```
최상위 구성에 다음을 추가하여 지원되지 않는 옵티마이저를 사용할 수도 있습니다.
```yaml
{
"zero_allow_untested_optimizer": true
}
```
DeepSpeed==0.8.3부터 오프로드를 사용하려면 오프로드가 DeepSpeed의 CPU Adam 옵티마이저에서 가장 잘 작동하므로 최상위 수준 구성에 다음 사항을 추가해야 합니다.
```yaml
{
"zero_force_ds_cpu_optimizer": false
}
```
</hfoption>
<hfoption id="scheduler">
DeepSpeed는 LRRangeTest, OneCycle, WarmupLR 및 WarmupDecayLR learning rate[schedulers](https://www.deepspeed.ai/docs/config-json/#scheduler-parameters)를 지원합니다.
트랜스포머와 DeepSpeed는 동일한 두 가지 스케줄러를 제공합니다:
* WarmupLR은 Transformers의 `--lr_scheduler_type constant_warmup`과 동일합니다.
* WarmupDecayLR은 Transformers의 `--lr_scheduler_type linear`와 동일합니다(Transformers에서 사용되는 기본 스케줄러입니다).
설정에서 스케줄러를 구성하지 않으면[`Trainer`]는 자동으로 WarmupDecayLR을 선택하고 명령줄에서 제공된 값 또는 기본값을 사용합니다: `warmup_min_lr`, `warmup_max_lr`, `warmup_num_steps`, `total_num_steps` (`max_steps`가 제공되지 않으면 런타임 중에 자동으로 계산됨).
매개변수를 `"auto"`으로 설정하거나 원하는 값을 직접 수동으로 입력할 수 있습니다.
```yaml
{
"scheduler": {
"type": "WarmupDecayLR",
"params": {
"total_num_steps": "auto",
"warmup_min_lr": "auto",
"warmup_max_lr": "auto",
"warmup_num_steps": "auto"
}
}
}
```
</hfoption>
</hfoptions>
### 정밀도[[precision]]
DeepSpeed는 fp32, fp16 및 bf16 혼합 정밀도를 지원합니다.
<hfoptions id="precision">
<hfoption id="fp32">
모델이 혼합 정밀도로 사전 학습되지 않은 경우와 같이 혼합 정밀도로 잘 작동하지 않는 경우 NaN 손실을 유발할 수 있는 오버플로 또는 언더플로 문제가 발생할 수 있습니다. 이러한 경우에는 기본 fp16 모드를 명시적으로 비활성화하여 전체 fp32 정밀도를 사용해야 합니다.
```yaml
{
"fp16": {
"enabled": false
}
}
```
Ampere GPU 및 PyTorch 1.7 이상의 경우 일부 연산에 대해 더 효율적인 [tf32](https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices) 형식으로 자동 전환되지만 결과는 여전히 fp32로 표시됩니다. [`Trainer`]에서 `--tf32`를 설정하여 활성화하고 `--tf32 0` 또는 `--no_tf32`를 비활성화하면 제어할 수 있습니다.
</hfoption>
<hfoption id="fp16">
PyTorch AMP와 같은 fp16 혼합 정밀도를 구성하면 메모리 사용량이 줄어들고 훈련 속도가 빨라집니다.[`Trainer`]는 `args.fp16_backend` 값에 따라 fp16을 자동으로 활성화 또는 비활성화하며, 나머지 구성은 사용자가 설정할 수 있습니다. 명령줄에서 다음 인수를 전달하면 fp16이 활성화됩니다: `fp16`, `--fp16_backend amp` 또는 `--fp16_full_eval`.
```yaml
{
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"loss_scale_window": 1000,
"initial_scale_power": 16,
"hysteresis": 2,
"min_loss_scale": 1
}
}
```
추가 딥스피드 fp16 훈련 옵션은 [fp16 훈련 옵션](https://www.deepspeed.ai/docs/config-json/#fp16-training-options) 참조를 참조하세요.
Apex와 같은 fp16 혼합 정밀도를 구성하려면 아래 그림과 같이 `"auto"` 또는 직접 값을 설정합니다.[`Trainer`]는 `args.fp16_backend` 및 `args.fp16_opt_level`의 값에 따라 `amp`를 자동으로 구성합니다. 다음 인수를 전달하면 명령줄에서 활성화할 수도 있습니다: `fp16`, `--fp16_backend apex` 또는 `--fp16_opt_level 01`.
```yaml
{
"amp": {
"enabled": "auto",
"opt_level": "auto"
}
}
```
</hfoption>
<hfoption id="bf16">
bf16을 사용하려면 DeepSpeed==0.6.0 이상이 필요합니다. bf16은 fp32와 동적 범위가 동일하며 손실 스케일링이 필요하지 않습니다. 그러나 [gradient accumulation](#gradient-accumulation)을 bf16과 함께 사용하면 이 형식의 낮은 정밀도로 인해 손실이 발생할 수 있으므로 원하지 않는 그레이디언트가 bf16에 누적될 수 있습니다.
bf16은 설정 파일에서 설정하거나 다음 인수를 전달하면 명령줄에서 활성화할 수 있습니다: `--bf16` 또는 `--bf16_full_eval`.
```yaml
{
"bf16": {
"enabled": "auto"
}
}
```
</hfoption>
</hfoptions>
### 배치 크기[[batch-size]]
배치 크기는 자동으로 구성하거나 명시적으로 설정할 수 있습니다. `"auto"` 옵션을 사용하도록 선택하면 [`Trainer`]는 `train_micro_batch_size_per_gpu`를 args.`per_device_train_batch_size`의 값으로, `train_batch_size`를 `args.world_size * args.per_device_train_batch_size * args.gradient_accumulation_steps`로 설정합니다.
```yaml
{
"train_micro_batch_size_per_gpu": "auto",
"train_batch_size": "auto"
}
```
### 그레이디언트 누적[[gradient-accumulation]]
그레이디언트 누적을 자동으로 구성하거나 명시적으로 설정할 수 있습니다. `"auto"` 옵션을 사용하도록 선택하면 [`Trainer`]가 `args.gradient_accumulation_steps`의 값으로 설정합니다.
```yaml
{
"gradient_accumulation_steps": "auto"
}
```
### 그레이디언트 클리핑[[gradient-clipping]]
그레이디언트 클리핑은 자동으로 구성하거나 명시적으로 설정할 수 있습니다. `"auto"` 옵션을 사용하도록 선택하면 [`Trainer`]가 `args.max_grad_norm`의 값으로 설정합니다.
```yaml
{
"gradient_clipping": "auto"
}
```
### 통신 데이터 유형(Communication data type)[[communication-data-type]]
축소, 수집 및 분산 작업과 같은 통신 집합체의 경우 별도의 데이터 유형이 사용됩니다.
모든 수집 및 분산 작업은 데이터와 동일한 데이터 유형으로 수행됩니다. 예를 들어 bf16으로 훈련하는 경우, 수집은 비손실 연산이므로 데이터도 bf16으로 수집됩니다.
예를 들어 그레이디언트가 여러 GPU에 걸쳐 평균화되는 경우와 같이 감소 연산은 손실이 발생합니다. 통신이 fp16 또는 bf16으로 수행되는 경우, 낮은 정밀도로 여러 숫자를 더하면 정확하지 않기 때문에 손실이 발생할 가능성이 더 높습니다. 특히 fp16보다 정밀도가 낮은 bf16의 경우 더욱 그렇습니다. 이러한 이유로 기울기를 평균화할 때 손실이 최소화되므로 감소 연산에는 fp16이 기본값으로 사용됩니다.
통신 데이터 유형은 설정 파일에서 `communication_data_type` 매개변수를 설정하여 선택할 수 있습니다. 예를 들어, fp32를 선택하면 약간의 오버헤드가 추가되지만 감소 연산이 fp32에 누적되고 준비가 되면 훈련 중인 반정밀 dtype으로 다운캐스트됩니다.
```yaml
{
"communication_data_type": "fp32"
}
```
## 모델 배포[[deployment]]
[torchrun](https://pytorch.org/docs/stable/elastic/run.html), `deepspeed` 런처 또는 [Accelerate](https://huggingface.co/docs/accelerate/basic_tutorials/launch#using-accelerate-launch) 등 다양한 런처를 통해 DeepSpeed를 배포할 수 있습니다. 배포하려면 [`Trainer`] 명령줄에 `--deepspeed ds_config.json`을 추가합니다. 필요한 명령줄 인수를 코드에 추가하려면 DeepSpeed의 [`add_config_arguments`](https://deepspeed.readthedocs.io/en/latest/initialize.html#argument-parsing) 유틸리티를 사용하는 것이 좋습니다.
이 가이드에서는 다양한 트레이닝 설정에 대해 `deepspeed` 런처로 DeepSpeed를 배포하는 방법을 보여드립니다. 보다 실용적인 사용 예제는 이 [post](https://github.com/huggingface/transformers/issues/8771#issuecomment-759248400)에서 확인할 수 있습니다.
<hfoptions id="deploy">
<hfoption id="multi-GPU">
여러 GPU에 DeepSpeed를 배포하려면 `--num_gpus` 매개변수를 추가하세요. 사용 가능한 모든 GPU를 사용하려는 경우 `--num_gpus`를 추가할 필요가 없습니다. 아래 예제에서는 2개의 GPU를 사용합니다.
```bash
deepspeed --num_gpus=2 examples/pytorch/translation/run_translation.py \
--deepspeed tests/deepspeed/ds_config_zero3.json \
--model_name_or_path google-t5/t5-small --per_device_train_batch_size 1 \
--output_dir output_dir --overwrite_output_dir --fp16 \
--do_train --max_train_samples 500 --num_train_epochs 1 \
--dataset_name wmt16 --dataset_config "ro-en" \
--source_lang en --target_lang ro
```
</hfoption>
<hfoption id="single-GPU">
단일 GPU에 DeepSpeed를 배포하려면 `--num_gpus` 매개변수를 추가하세요. GPU가 1개만 있는 경우 이 값을 명시적으로 설정할 필요는 없습니다. DeepSpeed는 지정된 노드에서 볼 수 있는 모든 GPU를 배포하므로 이 값을 명시적으로 설정할 필요는 없습니다.
```bash
deepspeed --num_gpus=1 examples/pytorch/translation/run_translation.py \
--deepspeed tests/deepspeed/ds_config_zero2.json \
--model_name_or_path google-t5/t5-small --per_device_train_batch_size 1 \
--output_dir output_dir --overwrite_output_dir --fp16 \
--do_train --max_train_samples 500 --num_train_epochs 1 \
--dataset_name wmt16 --dataset_config "ro-en" \
--source_lang en --target_lang ro
```
DeepSpeed는 단 하나의 GPU로도 여전히 유용합니다:
1. 일부 계산과 메모리를 CPU로 오프로드하여 더 큰 배치 크기를 사용하거나 일반적으로 맞지 않는 매우 큰 모델을 맞추기 위해 모델에 더 많은 GPU 리소스를 사용할 수 있도록 합니다.
2. 스마트 GPU 메모리 관리 시스템으로 메모리 조각화를 최소화하여 더 큰 모델과 데이터 배치에 맞출 수 있습니다.
<Tip>
단일 GPU에서 더 나은 성능을 얻으려면 [ZeRO-2](#zero-configuration) 구성 파일에서 `allgather_bucket_size` 및 `reduce_bucket_size` 값을 2e8로 설정하세요.
</Tip>
</hfoption>
</hfoptions>
### 다중 노드 환경에서의 모델 배포[[multi-node-deployment]]
노드는 워크로드를 실행하기 위한 하나 이상의 GPU입니다. 더 강력한 설정은 멀티 노드 설정으로, `deepspeed` 런처로 실행할 수 있습니다. 이 가이드에서는 각각 8개의 GPU가 있는 두 개의 노드가 있다고 가정해 보겠습니다. 첫 번째 노드는 `ssh hostname1`로, 두 번째 노드는 `ssh hostname2`로 접속할 수 있습니다. 두 노드 모두 비밀번호 없이 ssh를 통해 로컬로 서로 통신할 수 있어야 합니다.
기본적으로 DeepSpeed는 멀티노드 환경에서 공유 저장소를 사용할 것으로 예상합니다. 그렇지 않고 각 노드가 로컬 파일 시스템만 볼 수 있는 경우, 공유 파일 시스템에 대한 액세스 없이 로딩할 수 있도록 [`checkpoint`](https://www.deepspeed.ai/docs/config-json/#checkpoint-options)를 포함하도록 구성 파일을 조정해야 합니다:
```yaml
{
"checkpoint": {
"use_node_local_storage": true
}
}
```
[`Trainer`]의 ``--save_on_each_node` 인수를 사용하여 위의 `checkpoint`를 구성에 자동으로 추가할 수도 있습니다.
<hfoptions id="multinode">
<hfoption id="torchrun">
[torchrun](https://pytorch.org/docs/stable/elastic/run.html)의 경우, 각 노드에 ssh로 접속한 후 두 노드 모두에서 다음 명령을 실행해야 합니다. 런처는 두 노드가 동기화될 때까지 기다렸다가 트레이닝을 시작합니다.
```bash
torchrun --nproc_per_node=8 --nnode=2 --node_rank=0 --master_addr=hostname1 \
--master_port=9901 your_program.py <normal cl args> --deepspeed ds_config.json
```
</hfoption>
<hfoption id="deepspeed">
`deepspeed` 런처의 경우, 먼저 `hostfile`을 생성합니다.
```bash
hostname1 slots=8
hostname2 slots=8
```
그런 다음 다음 명령어로 트레이닝을 시작할 수 있습니다. `deepspeed` 런처는 두 노드에서 동시에 명령을 자동으로 실행합니다.
```bash
deepspeed --num_gpus 8 --num_nodes 2 --hostfile hostfile --master_addr hostname1 --master_port=9901 \
your_program.py <normal cl args> --deepspeed ds_config.json
```
다중 노드 컴퓨팅 리소스 구성에 대한 자세한 내용은 [Resource Configuration (multi-node)](https://www.deepspeed.ai/getting-started/#resource-configuration-multi-node) 가이드를 참조하세요.
</hfoption>
</hfoptions>
### SLURM[[slurm]]
SLURM 환경에서는 특정 SLURM 환경에 맞게 SLURM 스크립트를 조정해야 합니다.SLURM 스크립트 예시는 다음과 같습니다:
```bash
#SBATCH --job-name=test-nodes # 작업 이름
#SBATCH --nodes=2 # 노드 수
#SBATCH --ntasks-per-node=1 # 중요 - 노드당 분산 작업 1개!
#SBATCH --cpus-per-task=10 # 작업당 CPU 코어 수
#SBATCH --gres=gpu:8 # gpu 수
#SBATCH --time 20:00:00 # 최대 실행 시간 (HH:MM:SS)
#SBATCH --output=%x-%j.out # 출력 파일 이름
export GPUS_PER_NODE=8
export MASTER_ADDR=$(scontrol show hostnames $SLURM_JOB_NODELIST | head -n 1)
export MASTER_PORT=9901
srun --jobid $SLURM_JOBID bash -c 'python -m torch.distributed.run \
--nproc_per_node $GPUS_PER_NODE --nnodes $SLURM_NNODES --node_rank $SLURM_PROCID \
--master_addr $MASTER_ADDR --master_port $MASTER_PORT \
your_program.py <normal cl args> --deepspeed ds_config.json'
```
그런 다음 모든 노드에서 동시에 학습을 시작하는 다음 명령을 사용하여 다중 노드 배포를 예약할 수 있습니다.
```bash
sbatch launch.slurm
```
### 노트북[[notebook]]
`deepspeed` 런처는 노트북에서의 배포를 지원하지 않으므로 분산 환경을 에뮬레이션해야 합니다. 하지만 이는 1개의 GPU에서만 작동합니다. 1개 이상의 GPU를 사용하려면 딥스피드가 작동할 수 있는 다중 프로세스 환경을 사용해야 합니다. 즉, 여기에 표시된 것처럼 에뮬레이션할 수 없는 `deepspeed` 런처를 사용해야 합니다.
```py
# DeepSpeed는 단일 프로세스만 사용하더라도 분산 환경을 필요로 합니다.
# 이 코드로 분산 환경을 모방합니다.
import os
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = "9994" # RuntimeError: Address already in use 오류 발생 시 수정
os.environ["RANK"] = "0"
os.environ["LOCAL_RANK"] = "0"
os.environ["WORLD_SIZE"] = "1"
# 이제 평소와 같이 진행하되, DeepSpeed 설정 파일을 전달합니다.
training_args = TrainingArguments(..., deepspeed="ds_config_zero3.json")
trainer = Trainer(...)
trainer.train()
```
현재 디렉터리의 노트북에 구성 파일을 즉석에서 만들고 싶다면 전용 셀을 만들 수 있습니다.
```py
%%bash
cat <<'EOT' > ds_config_zero3.json
{
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"loss_scale_window": 1000,
"initial_scale_power": 16,
"hysteresis": 2,
"min_loss_scale": 1
},
"optimizer": {
"type": "AdamW",
"params": {
"lr": "auto",
"betas": "auto",
"eps": "auto",
"weight_decay": "auto"
}
},
"scheduler": {
"type": "WarmupLR",
"params": {
"warmup_min_lr": "auto",
"warmup_max_lr": "auto",
"warmup_num_steps": "auto"
}
},
"zero_optimization": {
"stage": 3,
"offload_optimizer": {
"device": "cpu",
"pin_memory": true
},
"offload_param": {
"device": "cpu",
"pin_memory": true
},
"overlap_comm": true,
"contiguous_gradients": true,
"sub_group_size": 1e9,
"reduce_bucket_size": "auto",
"stage3_prefetch_bucket_size": "auto",
"stage3_param_persistence_threshold": "auto",
"stage3_max_live_parameters": 1e9,
"stage3_max_reuse_distance": 1e9,
"stage3_gather_16bit_weights_on_model_save": true
},
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"steps_per_print": 2000,
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"wall_clock_breakdown": false
}
EOT
```
트레이닝 스크립트가 노트북 셀이 아닌 파일에 있는 경우, 노트북 셀의 셸에서 `deepspeed`를 정상적으로 실행할 수 있습니다. 예를 들어 `run_translation.py`를 시작하려면 다음과 같이 하세요.:
```py
!git clone https://github.com/huggingface/transformers
!cd transformers; deepspeed examples/pytorch/translation/run_translation.py ...
```
또한 `%%bash` 매직을 사용하여 여러 줄의 코드를 작성하여 셸 프로그램을 실행할 수도 있지만 교육이 완료될 때까지 로그를 볼 수 없습니다. `%%bash` 매직으로 분산 환경을 에뮬레이션할 필요는 없습니다.
```py
%%bash
git clone https://github.com/huggingface/transformers
cd transformers
deepspeed examples/pytorch/translation/run_translation.py ...
```
## 모델 가중치 저장하기[[save-model-weights]]
딥스피드는 기본 고정밀 fp32 가중치를 사용자 지정 체크포인트 최적화 파일(glob 패턴은 `global_step*/*optim_states.pt`처럼 보입니다)에 저장하고 일반 체크포인트 아래에 저장합니다.
<hfoptions id="save">
<hfoption id="fp16">
ZeRO-2로 훈련된 모델은 pytorch_model.bin 가중치를 fp16에 저장합니다. ZeRO-3으로 훈련된 모델의 모델 가중치를 fp16에 저장하려면 모델 가중치가 여러 GPU에 분할되어 있으므로 `“stage3_gather_16bit_weights_on_model_save”: true`를 설정해야 합니다. 그렇지 않으면 [`Trainer`]가 가중치를 fp16에 저장하지 않고 pytorch_model.bin 파일을 생성하지 않습니다. 이는 DeepSpeed의 state_dict에 실제 가중치 대신 플레이스홀더가 포함되어 있어 이를 로드할 수 없기 때문입니다.
```yaml
{
"zero_optimization": {
"stage3_gather_16bit_weights_on_model_save": true
}
}
```
</hfoption>
<hfoption id="fp32">
전체 정밀 가중치는 많은 메모리가 필요할 수 있으므로 트레이닝 중에 저장해서는 안 됩니다. 일반적으로 훈련이 완료된 후 오프라인으로 fp32 가중치를 저장하는 것이 가장 좋습니다. 그러나 여유 CPU 메모리가 많은 경우 훈련 중에 fp32 가중치를 저장할 수 있습니다. 이 섹션에서는 온라인과 오프라인 방식을 모두 다룹니다.
### 온라인 환경[[online]]
다음과 같이 최신 체크포인트를 로드하려면 체크포인트를 하나 이상 저장해야 합니다:
```py
from transformers.trainer_utils import get_last_checkpoint
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
checkpoint_dir = get_last_checkpoint(trainer.args.output_dir)
fp32_model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
```
`--load_best_model_at_end` 매개변수를 활성화하여 [`TrainingArguments`]에서 최적의 체크포인트를 추적하는 경우, 먼저 학습을 완료하고 최종 모델을 명시적으로 저장할 수 있습니다. 그런 다음 아래와 같이 다시 로드할 수 있습니다:
```py
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
checkpoint_dir = os.path.join(trainer.args.output_dir, "checkpoint-final")
trainer.deepspeed.save_checkpoint(checkpoint_dir)
fp32_model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
```
<Tip>
`load_state_dict_from_zero_checkpoint`가 실행되면 동일한 애플리케이션의 컨텍스트에서 모델을 더 이상 DeepSpeed에서 사용할 수 없습니다. `model.load_state_dict(state_dict)`는 모든 딥스피드 마법을 제거하므로 딥스피드 엔진을 다시 초기화해야 합니다. 이 기능은 훈련이 끝날 때만 사용하세요.
</Tip>
fp32 가중치의 state_dict를 추출하여 로드할 수도 있습니다:
```py
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # cpu에 이미 존재함
model = model.cpu()
model.load_state_dict(state_dict)
```
### 오프라인 환경[[offline]]
DeepSpeed는 언제든지 가중치를 추출할 수 있도록 체크포인트 폴더의 최상위 레벨에 zero_to_fp32.py 스크립트를 제공합니다. 이 스크립트는 독립형 스크립트로 구성 파일이나 [`Trainer`]가 필요하지 않습니다.
예를 들어 체크포인트 폴더가 다음과 같은 경우입니다:
```bash
$ ls -l output_dir/checkpoint-1/
-rw-rw-r-- 1 stas stas 1.4K Mar 27 20:42 config.json
drwxrwxr-x 2 stas stas 4.0K Mar 25 19:52 global_step1/
-rw-rw-r-- 1 stas stas 12 Mar 27 13:16 latest
-rw-rw-r-- 1 stas stas 827K Mar 27 20:42 optimizer.pt
-rw-rw-r-- 1 stas stas 231M Mar 27 20:42 pytorch_model.bin
-rw-rw-r-- 1 stas stas 623 Mar 27 20:42 scheduler.pt
-rw-rw-r-- 1 stas stas 1.8K Mar 27 20:42 special_tokens_map.json
-rw-rw-r-- 1 stas stas 774K Mar 27 20:42 spiece.model
-rw-rw-r-- 1 stas stas 1.9K Mar 27 20:42 tokenizer_config.json
-rw-rw-r-- 1 stas stas 339 Mar 27 20:42 trainer_state.json
-rw-rw-r-- 1 stas stas 2.3K Mar 27 20:42 training_args.bin
-rwxrw-r-- 1 stas stas 5.5K Mar 27 13:16 zero_to_fp32.py*
```
딥스피드 체크포인트(ZeRO-2 또는 ZeRO-3) 하위 폴더 `global_step1`에서 fp32 가중치를 재구성하려면 다음 명령을 실행하여 여러 GPU의 전체 fp32 가중치를 단일 pytorch_model.bin 파일로 생성하고 통합합니다. 스크립트는 자동으로 체크포인트가 포함된 하위 폴더를 찾습니다.
```py
python zero_to_fp32.py . pytorch_model.bin
```
<Tip>
자세한 사용법은 `python zero_to_fp32.py -h`를 실행하세요. 이 스크립트에는 최종 fp32 가중치의 2배의 일반 RAM이 필요합니다.
</Tip>
</hfoption>
</hfoptions>
## ZeRO Inference[[zero-inference]]
[ZeRO Inference](https://www.deepspeed.ai/2022/09/09/zero-inference.html)는 모델 가중치를 CPU 또는 NVMe 메모리에 배치하여 GPU에 부담을 주지 않으므로 GPU에서 대규모 모델을 사용하여 추론을 실행할 수 있습니다. 추론은 최적화 상태 및 그레이디언트에 많은 양의 메모리를 추가로 필요로 하지 않으므로 동일한 하드웨어에 훨씬 더 큰 배치 및/또는 시퀀스 길이를 맞출 수 있습니다.
ZeRO Inference는 [ZeRO-3](#zero-configuration)와 동일한 구성 파일을 공유하며, ZeRO-2 및 ZeRO-1 구성은 추론에 아무런 이점을 제공하지 않으므로 작동하지 않습니다.
ZeRO Inference를 실행하려면 일반적인 훈련 인수를 [`TrainingArguments`] 클래스에 전달하고 `--do_eval` 인수를 추가합니다.
```bash
deepspeed --num_gpus=2 your_program.py <normal cl args> --do_eval --deepspeed ds_config.json
```
## Trainer 없이 DeepSpeed 사용하기[[non-trainer-deepspeed-integration]]
DeepSpeed는 [`Trainer`] 클래스가 없는 트랜스포머에서도 작동합니다. 이는 [`~PreTrainedModel.from_pretrained`]를 호출할 때 ZeRO-3 매개변수를 수집하고 모델을 여러 GPU에 분할하는 작업만 처리하는 [`HfDeepSpeedConfig`]가 처리합니다.
<Tip>
모든 것이 자동으로 처리되기를 원한다면, [`Trainer`]와 함께 DeepSpeed를 사용해 보세요! [DeepSpeed 문서](https://www.deepspeed.ai/)를 참조하여 설정 파일에서 매개변수 값을 수동으로 구성해야 합니다(`"auto"` 값은 사용할 수 없음).
</Tip>
ZeRO-3를 효율적으로 배포하려면 모델 앞에 [`HfDeepSpeedConfig`] 객체를 인스턴스화하고 해당 객체를 유지해야 합니다:
<hfoptions id="models">
<hfoption id="pretrained model">
```py
from transformers.integrations import HfDeepSpeedConfig
from transformers import AutoModel
import deepspeed
ds_config = {...} # deepspeed 설정 객체 또는 파일 경로
# Zero 3를 감지하기 위해 모델을 인스턴스화하기 전에 반드시 실행해야 합니다
dschf = HfDeepSpeedConfig(ds_config) # 이 객체를 유지하세요.
model = AutoModel.from_pretrained("openai-community/gpt2")
engine = deepspeed.initialize(model=model, config_params=ds_config, ...)
```
</hfoption>
<hfoption id="non-pretrained model">
[`HfDeepSpeedConfig`] is not required for ZeRO-1 or ZeRO-2.
```py
from transformers.integrations import HfDeepSpeedConfig
from transformers import AutoModel, AutoConfig
import deepspeed
ds_config = {...} # deepspeed 설정 객체 또는 파일 경로
# Zero 3를 감지하기 위해 모델을 인스턴스화하기 전에 반드시 실행해야 합니다
dschf = HfDeepSpeedConfig(ds_config) # 이 객체를 유지하세요.
config = AutoConfig.from_pretrained("openai-community/gpt2")
model = AutoModel.from_config(config)
engine = deepspeed.initialize(model=model, config_params=ds_config, ...)
```
</hfoption>
</hfoptions>
### Trainer 없이 ZeRO Inference 사용하기[[non-trainer-zero-inference]]
단일 GPU에 모델을 맞출 수 없는 경우 [`Trainer`]없이 ZeRO 추론을 실행하려면 추가 GPU를 사용하거나 CPU 메모리로 오프로드를 시도하세요. 여기서 이해해야 할 중요한 뉘앙스는 ZeRO가 설계된 방식에 따라 서로 다른 GPU에서 서로 다른 입력을 병렬로 처리할 수 있다는 것입니다.
반드시 확인하세요:
* GPU 메모리가 충분한 경우 CPU 오프로드를 비활성화합니다(속도가 느려지므로).
* Ampere 이상의 GPU를 사용하는 경우 bf16을 활성화하면 속도가 빨라집니다. 이러한 GPU가 없는 경우 오버플로 오류가 발생할 수 있으므로 bf16으로 사전 학습된 모델(T5 모델)을 사용하지 않는 한 fp16을 활성화할 수 있습니다.
단일 GPU에 맞지 않는 모델에서 [`Trainer`] 없이 ZeRO 추론을 실행하는 방법에 대한 더 나은 아이디어를 얻으려면 다음 스크립트를 살펴보시기 바랍니다.
```py
#!/usr/bin/env python
# 이 스크립트는 단일 GPU에 모델을 맞출 수 없을 때 추론 모드에서 Deepspeed ZeRO를 사용하는 방법을 보여줍니다.
#
# 1. CPU 오프로드와 함께 1개의 GPU 사용
# 2. 또는 여러 GPU 사용
#
# 먼저 deepspeed를 설치해야 합니다: pip install deepspeed
#
# 여기서는 약 15GB의 GPU RAM이 필요한 3B "bigscience/T0_3B" 모델을 사용합니다 - 따라서 1개의 큰 GPU나 2개의
# 작은 GPU로 처리할 수 있습니다. 또는 1개의 작은 GPU와 많은 CPU 메모리로도 가능합니다.
#
# 약 50GB가 필요한 "bigscience/T0"와 같은 더 큰 모델을 사용하려면, 80GB GPU가 없는 한
# 2-4개의 GPU가 필요할 것입니다. 그리고 여러 입력을 한 번에 처리하고 싶다면
# 스크립트를 수정하여 더 많은 GPU를 처리할 수 있습니다.
#
# 제공된 deepspeed 설정은 CPU 메모리 오프로딩도 활성화하므로, 사용 가능한 CPU 메모리가 많고
# 속도 저하를 감수할 수 있다면 일반적으로 단일 GPU에 맞지 않는 모델을 로드할 수 있을 것입니다.
# GPU 메모리가 충분하다면 CPU로의 오프로드를 원하지 않을 때 프로그램이 더 빠르게 실행될 것입니다 - 그럴 때는 해당 섹션을 비활성화하세요.
#
# 1개의 GPU에 배포하려면:
#
# deepspeed --num_gpus 1 t0.py
# 또는:
# python -m torch.distributed.run --nproc_per_node=1 t0.py
#
# 2개의 GPU에 배포하려면:
#
# deepspeed --num_gpus 2 t0.py
# 또는:
# python -m torch.distributed.run --nproc_per_node=2 t0.py
from transformers import AutoTokenizer, AutoConfig, AutoModelForSeq2SeqLM
from transformers.integrations import HfDeepSpeedConfig
import deepspeed
import os
import torch
os.environ["TOKENIZERS_PARALLELISM"] = "false" # 토크나이저의 병렬 처리에 관한 경고를 피하기 위함입니다.
# 분산 환경 설정
local_rank = int(os.getenv("LOCAL_RANK", "0"))
world_size = int(os.getenv("WORLD_SIZE", "1"))
torch.cuda.set_device(local_rank)
deepspeed.init_distributed()
model_name = "bigscience/T0_3B"
config = AutoConfig.from_pretrained(model_name)
model_hidden_size = config.d_model
# 배치 크기는 world_size로 나누어 떨어져야 하지만, world_size보다 클 수 있습니다
train_batch_size = 1 * world_size
# ds_config 참고사항
#
# - Ampere 이상의 GPU를 사용하는 경우 bf16을 활성화하세요 - 이는 혼합 정밀도로 실행되어
# 더 빠를 것입니다.
#
# - 오래된 GPU의 경우 fp16을 활성화할 수 있지만, bf16으로 사전 훈련되지 않은 모델에서만 작동합니다 - 예를 들어
# 모든 공식 t5 모델은 bf16으로 사전 훈련되었습니다
#
# - CPU 오프로드를 원하지 않는다면 offload_param.device를 "none"으로 설정하거나 `offload_param` 섹션을
# 완전히 제거하세요
#
# - `offload_param`을 사용하는 경우, stage3_param_persistence_threshold를 수동으로 미세 조정하여
# 어떤 매개변수가 GPU에 남아있어야 하는지 제어할 수 있습니다 - 값이 클수록 오프로드 크기가 작아집니다
#
# Deepspeed 설정에 대한 자세한 정보는 다음을 참조하세요
# https://huggingface.co/docs/transformers/main/main_classes/deepspeed
# 일관성을 위해 json과 동일한 형식을 유지하되, true/false에는 소문자를 사용합니다
# fmt: off
ds_config = {
"fp16": {
"enabled": False
},
"bf16": {
"enabled": False
},
"zero_optimization": {
"stage": 3,
"offload_param": {
"device": "cpu",
"pin_memory": True
},
"overlap_comm": True,
"contiguous_gradients": True,
"reduce_bucket_size": model_hidden_size * model_hidden_size,
"stage3_prefetch_bucket_size": 0.9 * model_hidden_size * model_hidden_size,
"stage3_param_persistence_threshold": 10 * model_hidden_size
},
"steps_per_print": 2000,
"train_batch_size": train_batch_size,
"train_micro_batch_size_per_gpu": 1,
"wall_clock_breakdown": False
}
# fmt: on
# 다음 줄은 모델의 `from_pretrained` 메소드가 호출될 때
# deepspeed.zero.Init를 사용하여 모델을 여러 GPU에 직접 분할하도록 transformers에 지시합니다.
#
# **이는 AutoModelForSeq2SeqLM.from_pretrained(model_name)로 모델을 로드하기 전에 실행되어야 합니다**
#
# 그렇지 않으면 모델이 먼저 정상적으로 로드된 후 포워드 시에만 분할되는데, 이는
# 덜 효율적이며 CPU RAM이 부족할 경우 실패할 수 있습니다
dschf = HfDeepSpeedConfig(ds_config) # 이 객체를 유지하세요
# 이제 모델을 로드할 수 있습니다.
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
# Deepspeed ZeRO를 초기화하고 엔진 객체만 저장
ds_engine = deepspeed.initialize(model=model, config_params=ds_config)[0]
ds_engine.module.eval() # inference
# Deepspeed ZeRO는 각 GPU에서 서로 관련 없는 입력을 처리할 수 있습니다. 따라서 2개의 GPU를 사용하면 한 번에 2개의 입력을 처리할 수 있습니다.
# GPU를 더 많이 사용하는 경우 그에 맞게 조정하세요.
# 물론 처리할 입력이 하나뿐이라면 두 GPU에 동일한 문자열을 전달해야 합니다.
# GPU를 하나만 사용하는 경우에는 rank 0만 갖게 됩니다.
rank = torch.distributed.get_rank()
if rank == 0:
text_in = "Is this review positive or negative? Review: this is the best cast iron skillet you will ever buy"
elif rank == 1:
text_in = "Is this review positive or negative? Review: this is the worst restaurant ever"
tokenizer = AutoTokenizer.from_pretrained(model_name)
inputs = tokenizer.encode(text_in, return_tensors="pt").to(device=local_rank)
with torch.no_grad():
outputs = ds_engine.module.generate(inputs, synced_gpus=True)
text_out = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(f"rank{rank}:\n in={text_in}\n out={text_out}")
```
스크립트를 t0.py로 저장하고 실행합니다:
```bash
$ deepspeed --num_gpus 2 t0.py
rank0:
in=Is this review positive or negative? Review: this is the best cast iron skillet you will ever buy
out=Positive
rank1:
in=Is this review positive or negative? Review: this is the worst restaurant ever
out=negative
```
이것은 매우 기본적인 예시이므로 사용 사례에 맞게 조정할 수 있습니다.
### 생성[[generate]]
생성에 ZeRO-3와 함께 여러 개의 GPU를 사용하려면 [`~GenerationMixin.generate`] 메서드에서 `synced_gpus=True`를 설정하여 GPU를 동기화해야 합니다. 그렇지 않으면 한 GPU가 다른 GPU보다 먼저 생성을 완료하면 나머지 GPU가 먼저 완료한 GPU로부터 가중치 샤드를 받지 못하여 전체 시스템이 중단됩니다.
트랜스포머>=4.28의 경우, 생성 중에 여러 개의 GPU가 감지되면 `synced_gpus`가 자동으로 `True`로 설정됩니다.
## 트러블슈팅[[troubleshoot]]
문제가 발생하면 DeepSpeed가 문제의 원인이 아닌 경우가 많으므로(아주 명백하고 예외적으로 DeepSpeed 모듈을 볼 수 있는 경우가 아니라면) DeepSpeed가 문제의 원인인지 고려해야 합니다! 첫 번째 단계는 DeepSpeed 없이 설정을 다시 시도하고 문제가 지속되면 문제를 신고하는 것입니다. 문제가 핵심적인 DeepSpeed 문제이고 transformers와 관련이 없는 경우, [DeepSpeed 리포지토리](https://github.com/deepspeedai/DeepSpeed)에서 이슈를 개설하세요.
transformers와 관련된 이슈를 개설할 때에는 다음 정보를 제공해 주세요:
* 전체 DeepSpeed 구성 파일
*[`Trainer`]의 명령줄 인수, 또는[`Trainer`] 설정을 직접 작성하는 경우[`TrainingArguments`] 인수(관련 없는 항목이 수십 개 있는 [`TrainingArguments`]는 덤프하지 마세요).
* 다음 코드의 출력 결과:
```bash
python -c 'import torch; print(f"torch: {torch.__version__}")'
python -c 'import transformers; print(f"transformers: {transformers.__version__}")'
python -c 'import deepspeed; print(f"deepspeed: {deepspeed.__version__}")'
```
* 문제를 재현할 수 있는 Google Colab 노트북 링크
* 불가능할 경우 기존 예제를 사용하여 문제를 재현할 수 있는 표준 및 사용자 지정이 아닌 데이터 집합을 사용할 수 있습니다.
다음 섹션에서는 가장 일반적인 두 가지 문제를 해결하기 위한 가이드를 제공합니다.
### DeepSpeed 프로세스가 시작 단계에서 종료되었을 경우[[deepspeed-process-killed-at-startup]]
실행 중에 트레이스백 없이 DeepSpeed 프로세스가 종료되면 일반적으로 프로그램이 시스템보다 많은 CPU 메모리를 할당하려고 시도했거나 프로세스가 허용된 것보다 많은 CPU 메모리를 할당하려고 시도하여 OS 커널이 프로세스를 종료했음을 의미합니다. 이 경우 구성 파일에 `offload_optimizer`, `offload_param` 또는 둘 다 CPU로 오프로드하도록 구성되어 있는지 확인하세요.
NVMe 및 ZeRO-3를 설정한 경우 NVMe로 오프로드를 실험해 보세요(모델의 메모리 요구 사항을 [확인](https://deepspeed.readthedocs.io/en/latest/memory.html)하세요).
### NaN 손실[[nan-loss]]
모델을 bf16으로 사전 훈련한 다음 fp16으로 사용하려고 할 때 NaN 손실이 발생하는 경우가 많습니다(특히 TPU 훈련 모델에 해당). 이 문제를 해결하려면 하드웨어가 이를 지원하는 경우(TPU, Ampere GPU 이상) fp32 또는 bf16을 사용하세요.
다른 문제는 fp16 사용과 관련이 있을 수 있습니다. 예를 들어 이것이 fp16 구성인 경우입니다:
```yaml
{
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"loss_scale_window": 1000,
"initial_scale_power": 16,
"hysteresis": 2,
"min_loss_scale": 1
}
}
```
로그에 다음과 같은 `OVERFLOW!` 메시지가 표시될 수 있습니다:
```bash
0%| | 0/189 [00:00<?, ?it/s]
[deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 262144, reducing to 262144
1%|▌ | 1/189 [00:00<01:26, 2.17it/s]
[deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 262144, reducing to 131072.0
1%|█▏
[...]
[deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 1, reducing to 1
14%|████████████████▌ | 27/189 [00:14<01:13, 2.21it/s]
[deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 1, reducing to 1
15%|█████████████████▏ | 28/189 [00:14<01:13, 2.18it/s]
[deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 1, reducing to 1
15%|█████████████████▊ | 29/189 [00:15<01:13, 2.18it/s]
[deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 1, reducing to 1
[...]
```
이는 DeepSpeed 손실 스케일러가 손실 오버플로를 극복할 수 있는 스케일링 계수를 찾을 수 없음을 의미합니다. 이 문제를 해결하려면 `initial_scale_power` 값을 더 높게 설정하세요(일반적으로 32가 적절합니다).
## 리소스[[resources]]
DeepSpeed ZeRO는 제한된 GPU 리소스로 추론을 위해 매우 큰 모델을 훈련하고 로드하는 강력한 기술로, 누구나 쉽게 사용할 수 있습니다. DeepSpeed에 대해 자세히 알아보려면 [블로그 포스트](https://www.microsoft.com/en-us/research/search/?q=deepspeed), [공식 문서](https://www.deepspeed.ai/getting-started/), [깃허브 리포지토리](https://github.com/deepspeedai/DeepSpeed)를 참조하세요.
다음 문서도 ZeRO에 대해 자세히 알아볼 수 있는 훌륭한 자료입니다:
* [ZeRO: Memory Optimizations Toward Training Trillion Parameter Models](https://hf.co/papers/1910.02054)
* [ZeRO-Offload: Democratizing Billion-Scale Model Training](https://hf.co/papers/2101.06840)
* [ZeRO-Infinity: Breaking the GPU Memory Wall for Extreme Scale Deep Learning](https://hf.co/papers/2104.07857)
| transformers/docs/source/ko/deepspeed.md/0 | {
"file_path": "transformers/docs/source/ko/deepspeed.md",
"repo_id": "transformers",
"token_count": 43913
} | 393 |
<!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# 일본어 BERT (BertJapanese) [[bertjapanese]]
## 개요 [[overview]]
일본어 문장에 학습된 BERT 모델 입니다.
각각 서로 다른 토큰화 방법을 사용하는 두 모델:
- MeCab와 WordPiece를 사용하여 토큰화합니다. 이를 위해 추가 의존성 [fugashi](https://github.com/polm/fugashi)이 필요합니다. (이는 [MeCab](https://taku910.github.io/mecab/)의 래퍼입니다.)
- 문자 단위로 토큰화합니다.
*MecabTokenizer*를 사용하려면, 의존성을 설치하기 위해 `pip install transformers["ja"]` (또는 소스에서 설치하는 경우 `pip install -e .["ja"]`) 명령을 실행해야 합니다.
자세한 내용은 [cl-tohoku 리포지토리](https://github.com/cl-tohoku/bert-japanese)에서 확인하세요.
MeCab과 WordPiece 토큰화를 사용하는 모델 예시:
```python
>>> import torch
>>> from transformers import AutoModel, AutoTokenizer
>>> bertjapanese = AutoModel.from_pretrained("cl-tohoku/bert-base-japanese")
>>> tokenizer = AutoTokenizer.from_pretrained("cl-tohoku/bert-base-japanese")
>>> ## Input Japanese Text
>>> line = "吾輩は猫である。"
>>> inputs = tokenizer(line, return_tensors="pt")
>>> print(tokenizer.decode(inputs["input_ids"][0]))
[CLS] 吾輩 は 猫 で ある 。 [SEP]
>>> outputs = bertjapanese(**inputs)
```
문자 토큰화를 사용하는 모델 예시:
```python
>>> bertjapanese = AutoModel.from_pretrained("cl-tohoku/bert-base-japanese-char")
>>> tokenizer = AutoTokenizer.from_pretrained("cl-tohoku/bert-base-japanese-char")
>>> ## Input Japanese Text
>>> line = "吾輩は猫である。"
>>> inputs = tokenizer(line, return_tensors="pt")
>>> print(tokenizer.decode(inputs["input_ids"][0]))
[CLS] 吾 輩 は 猫 で あ る 。 [SEP]
>>> outputs = bertjapanese(**inputs)
```
<Tip>
이는 토큰화 방법을 제외하고는 BERT와 동일합니다. API 참조 정보는 [BERT 문서](https://huggingface.co/docs/transformers/main/en/model_doc/bert)를 참조하세요.
이 모델은 [cl-tohoku](https://huggingface.co/cl-tohoku)께서 기여하였습니다.
</Tip>
## BertJapaneseTokenizer
[[autodoc]] BertJapaneseTokenizer
| transformers/docs/source/ko/model_doc/bert-japanese.md/0 | {
"file_path": "transformers/docs/source/ko/model_doc/bert-japanese.md",
"repo_id": "transformers",
"token_count": 1393
} | 394 |
<!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# ESM [[esm]]
## 개요 [[overview]]
이 페이지는 Meta AI의 Fundamental AI Research 팀에서 제공하는 Transformer 단백질 언어 모델에 대한 코드와 사전 훈련된 가중치를 제공합니다. 여기에는 최첨단인 ESMFold와 ESM-2, 그리고 이전에 공개된 ESM-1b와 ESM-1v가 포함됩니다. Transformer 단백질 언어 모델은 Alexander Rives, Joshua Meier, Tom Sercu, Siddharth Goyal, Zeming Lin, Jason Liu, Demi Guo, Myle Ott, C. Lawrence Zitnick, Jerry Ma, Rob Fergus의 논문 [Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences](https://www.pnas.org/content/118/15/e2016239118)에서 소개되었습니다. 이 논문의 첫 번째 버전은 2019년에 [출판 전 논문](https://www.biorxiv.org/content/10.1101/622803v1?versioned=true) 형태로 공개되었습니다.
ESM-2는 다양한 구조 예측 작업에서 테스트된 모든 단일 시퀀스 단백질 언어 모델을 능가하며, 원자 수준의 구조 예측을 가능하게 합니다. 이 모델은 Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Sal Candido, Alexander Rives의 논문 [Language models of protein sequences at the scale of evolution enable accurate structure prediction](https://doi.org/10.1101/2022.07.20.500902)에서 공개되었습니다.
이 논문에서 함께 소개된 ESMFold는 ESM-2 스템을 사용하며, 최첨단의 정확도로 단백질 접힘 구조를 예측할 수 있는 헤드를 갖추고 있습니다. [AlphaFold2](https://www.nature.com/articles/s41586-021-03819-2)와 달리, 이는 대형 사전 훈련된 단백질 언어 모델 스템의 토큰 임베딩에 의존하며, 추론 시 다중 시퀀스 정렬(MSA) 단계를 수행하지 않습니다. 이는 ESMFold 체크포인트가 완전히 "독립적"이며, 예측을 위해 알려진 단백질 시퀀스와 구조의 데이터베이스, 그리고 그와 관련 외부 쿼리 도구를 필요로 하지 않는다는 것을 의미합니다. 그리고 그 결과, 훨씬 빠릅니다.
"Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences"의 초록은 다음과 같습니다:
*인공지능 분야에서는 대규모의 데이터와 모델 용량을 갖춘 비지도 학습의 조합이 표현 학습과 통계적 생성에서 주요한 발전을 이끌어냈습니다. 생명 과학에서는 시퀀싱 기술의 성장이 예상되며, 자연 시퀀스 다양성에 대한 전례 없는 데이터가 나올 것으로 기대됩니다. 진화적 단계에서 볼 때, 단백질 언어 모델링은 생물학을 위한 예측 및 생성 인공지능을 향한 논리적인 단계에 있습니다. 이를 위해 우리는 진화적 다양성을 아우르는 2억 5천만 개의 단백질 시퀀스에서 추출한 860억 개의 아미노산에 대해 심층 컨텍스트 언어 모델을 비지도 학습으로 훈련합니다. 그 결과 모델은 그 표현에서 생물학적 속성에 대한 정보를 포함합니다. 이 표현은 시퀀스 데이터만으로 학습됩니다. 학습된 표현 공간은 아미노산의 생화학적 특성 수준에서부터 단백질의 원거리 상동성까지 구조를 반영하는 다중 규모의 조직을 가지고 있습니다. 이 표현에는 2차 및 3차 구조에 대한 정보가 인코딩되어 있으며, 선형 전사에 의해 식별 될 수 있습니다. 표현 학습은 돌연변이에 의한 효과와 2차 구조의 최첨단 지도 예측을 가능하게 하고, 넓은 범위의 접촉 부위 예측을 위한 최첨단 특징을 향상시킵니다.*
"Language models of protein sequences at the scale of evolution enable accurate structure prediction"의 초록은 다음과 같습니다:
*대형 언어 모델은 최근 규모가 커짐에 따라 긴급한 기능을 개발하여 단순한 패턴 매칭을 넘어 더 높은 수준의 추론을 수행하고 생생한 이미지와 텍스트를 생성하는 것으로 나타났습니다. 더 작은 규모에서 훈련된 단백질 시퀀스의 언어 모델이 연구되었지만, 그들이 규모가 커짐에 따라 생물학에 대해 무엇을 배우는지는 거의 알려져 있지 않습니다. 이 연구에서 우리는 현재까지 평가된 가장 큰 150억 개의 매개변수를 가진 모델을 훈련합니다. 우리는 모델이 규모가 커짐에 따라 단일 아미노산의 해상도로 단백질의 3차원 구조를 예측할 수 있는 정보를 학습한다는 것을 발견했습니다. 우리는 개별 단백질 시퀀스로부터 직접 고정밀 원자 수준의 엔드-투-엔드 구조 예측을 하기 위한 ESMFold를 제시합니다. ESMFold는 언어 모델에 잘 이해되는 낮은 퍼플렉서티를 가진 시퀀스에 대해 AlphaFold2와 RoseTTAFold와 유사한 정확도를 가지고 있습니다. ESMFold의 추론은 AlphaFold2보다 한 자릿수 빠르며, 메타게놈 단백질의 구조적 공간을 실용적인 시간 내에 탐색할 수 있게 합니다.*
원본 코드는 [여기](https://github.com/facebookresearch/esm)에서 찾을 수 있으며, Meta AI의 Fundamental AI Research 팀에서 개발되었습니다. ESM-1b, ESM-1v, ESM-2는 [jasonliu](https://huggingface.co/jasonliu)와 [Matt](https://huggingface.co/Rocketknight1)에 의해 HuggingFace에 기여되었습니다.
ESMFold는 [Matt](https://huggingface.co/Rocketknight1)와 [Sylvain](https://huggingface.co/sgugger)에 의해 HuggingFace에 기여되었으며, 이 과정에서 많은 도움을 준 Nikita Smetanin, Roshan Rao, Tom Sercu에게 큰 감사를 드립니다!
## 사용 팁 [[usage-tips]]
- ESM 모델은 마스크드 언어 모델링(MLM) 목표로 훈련되었습니다.
- HuggingFace의 ESMFold 포트는 [openfold](https://github.com/aqlaboratory/openfold) 라이브러리의 일부를 사용합니다. `openfold` 라이브러리는 Apache License 2.0에 따라 라이선스가 부여됩니다.
## 리소스 [[resources]]
- [텍스트 분류 작업 가이드](../tasks/sequence_classification)
- [토큰 분류 작업 가이드](../tasks/token_classification)
- [마스킹드 언어 모델링 작업 가이드](../tasks/masked_language_modeling)
## EsmConfig [[transformers.EsmConfig]]
[[autodoc]] EsmConfig
- all
## EsmTokenizer [[transformers.EsmTokenizer]]
[[autodoc]] EsmTokenizer
- build_inputs_with_special_tokens
- get_special_tokens_mask
- create_token_type_ids_from_sequences
- save_vocabulary
<frameworkcontent>
<pt>
## EsmModel [[transformers.EsmModel]]
[[autodoc]] EsmModel
- forward
## EsmForMaskedLM [[transformers.EsmForMaskedLM]]
[[autodoc]] EsmForMaskedLM
- forward
## EsmForSequenceClassification [[transformers.EsmForSequenceClassification]]
[[autodoc]] EsmForSequenceClassification
- forward
## EsmForTokenClassification [[transformers.EsmForTokenClassification]]
[[autodoc]] EsmForTokenClassification
- forward
## EsmForProteinFolding [[transformers.EsmForProteinFolding]]
[[autodoc]] EsmForProteinFolding
- forward
</pt>
<tf>
## TFEsmModel [[transformers.TFEsmModel]]
[[autodoc]] TFEsmModel
- call
## TFEsmForMaskedLM [[transformers.TFEsmForMaskedLM]]
[[autodoc]] TFEsmForMaskedLM
- call
## TFEsmForSequenceClassification [[transformers.TFEsmForSequenceClassification]]
[[autodoc]] TFEsmForSequenceClassification
- call
## TFEsmForTokenClassification [[transformers.TFEsmForTokenClassification]]
[[autodoc]] TFEsmForTokenClassification
- call
</tf>
</frameworkcontent>
| transformers/docs/source/ko/model_doc/esm.md/0 | {
"file_path": "transformers/docs/source/ko/model_doc/esm.md",
"repo_id": "transformers",
"token_count": 5246
} | 395 |
<!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# MarianMT[[MarianMT]]
<div class="flex flex-wrap space-x-1">
<a href="https://huggingface.co/models?filter=marian">
<img alt="Models" src="https://img.shields.io/badge/All_model_pages-marian-blueviolet">
</a>
<a href="https://huggingface.co/spaces/docs-demos/opus-mt-zh-en">
<img alt="Spaces" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue">
</a>
</div>
## 개요[[Overview]]
BART와 동일한 모델을 사용하는 번역 모델 프레임워크입니다. 번역 결과는 각 모델 카드의 테스트 세트와 유사하지만, 정확히 일치하지는 않을 수 있습니다. 이 모델은 [sshleifer](https://huggingface.co/sshleifer)가 제공했습니다.
## 구현 노트[[Implementation Notes]]
- 각 모델은 약 298 MB를 차지하며, 1,000개 이상의 모델이 제공됩니다.
- 지원되는 언어 쌍 목록은 [여기](https://huggingface.co/Helsinki-NLP)에서 확인할 수 있습니다.
- 모델들은 [Jörg Tiedemann](https://researchportal.helsinki.fi/en/persons/j%C3%B6rg-tiedemann)에 의해 [Marian](https://marian-nmt.github.io/) C++ 라이브러리를 이용하여 학습되었습니다. 이 라이브러리는 빠른 학습과 번역을 지원합니다.
- 모든 모델은 6개 레이어로 이루어진 Transformer 기반의 인코더-디코더 구조입니다. 각 모델의 성능은 모델 카드에 기입되어 있습니다.
- BPE 전처리가 필요한 80개의 OPUS 모델은 지원되지 않습니다.
- 모델링 코드는 [`BartForConditionalGeneration`]을 기반으로 하며, 일부 수정사항이 반영되어 있습니다:
- 정적 (사인 함수 기반) 위치 임베딩 사용 (`MarianConfig.static_position_embeddings=True`)
- 임베딩 레이어 정규화 생략 (`MarianConfig.normalize_embedding=False`)
- 모델은 생성 시 프리픽스로 `pad_token_id` (해당 토큰 임베딩 값은 0)를 사용하여 시작합니다 (Bart는
`<s/>`를 사용),
- Marian 모델을 PyTorch로 대량 변환하는 코드는 `convert_marian_to_pytorch.py`에서 찾을 수 있습니다.
## 모델 이름 규칙[[Naming]]
- 모든 모델 이름은 `Helsinki-NLP/opus-mt-{src}-{tgt}` 형식을 따릅니다.
- 모델의 언어 코드 표기는 일관되지 않습니다. 두 자리 코드는 일반적으로 [여기](https://developers.google.com/admin-sdk/directory/v1/languages)에서 찾을 수 있으며, 세 자리 코드는 "언어 코드 {code}"로 구글 검색을 통해 찾습니다.
- `es_AR`과 같은 형태의 코드는 `code_{region}` 형식을 의미합니다. 여기서의 예시는 아르헨티나의 스페인어를 의미합니다.
- 모델 변환은 두 단계로 이루어졌습니다. 처음 1,000개 모델은 ISO-639-2 코드를 사용하고, 두 번째 그룹은 ISO-639-5와 ISO-639-2 코드를 조합하여 언어를 식별합니다.
## 예시[[Examples]]
- Marian 모델은 라이브러리의 다른 번역 모델들보다 크기가 작아 파인튜닝 실험과 통합 테스트에 유용합니다.
- [GPU에서 파인튜닝하기](https://github.com/huggingface/transformers/blob/master/examples/legacy/seq2seq/train_distil_marian_enro.sh)
## 다국어 모델 사용법[[Multilingual Models]]
- 모든 모델 이름은`Helsinki-NLP/opus-mt-{src}-{tgt}` 형식을 따릅니다.
- 다중 언어 출력을 지원하는 모델의 경우, 출력을 원하는 언어의 언어 코드를 `src_text`의 시작 부분에 추가하여 지정해야 합니다.
- 모델 카드에서 지원되는 언어 코드의 목록을 확인할 수 있습니다! 예를 들어 [opus-mt-en-roa](https://huggingface.co/Helsinki-NLP/opus-mt-en-roa)에서 확인할 수 있습니다.
- `Helsinki-NLP/opus-mt-roa-en`처럼 소스 측에서만 다국어를 지원하는 모델의 경우, 별도의 언어 코드 지정이 필요하지 않습니다.
[Tatoeba-Challenge 리포지토리](https://github.com/Helsinki-NLP/Tatoeba-Challenge)의 새로운 다국적 모델은 3자리 언어 코드를 사용합니다:
```python
>>> from transformers import MarianMTModel, MarianTokenizer
>>> src_text = [
... ">>fra<< this is a sentence in english that we want to translate to french",
... ">>por<< This should go to portuguese",
... ">>esp<< And this to Spanish",
... ]
>>> model_name = "Helsinki-NLP/opus-mt-en-roa"
>>> tokenizer = MarianTokenizer.from_pretrained(model_name)
>>> print(tokenizer.supported_language_codes)
['>>zlm_Latn<<', '>>mfe<<', '>>hat<<', '>>pap<<', '>>ast<<', '>>cat<<', '>>ind<<', '>>glg<<', '>>wln<<', '>>spa<<', '>>fra<<', '>>ron<<', '>>por<<', '>>ita<<', '>>oci<<', '>>arg<<', '>>min<<']
>>> model = MarianMTModel.from_pretrained(model_name)
>>> translated = model.generate(**tokenizer(src_text, return_tensors="pt", padding=True))
>>> [tokenizer.decode(t, skip_special_tokens=True) for t in translated]
["c'est une phrase en anglais que nous voulons traduire en français",
'Isto deve ir para o português.',
'Y esto al español']
```
허브에 있는 모든 사전 학습된 모델을 확인하는 코드입니다:
```python
from huggingface_hub import list_models
model_list = list_models()
org = "Helsinki-NLP"
model_ids = [x.id for x in model_list if x.id.startswith(org)]
suffix = [x.split("/")[1] for x in model_ids]
old_style_multi_models = [f"{org}/{s}" for s in suffix if s != s.lower()]
```
## 구형 다국어 모델[[Old Style Multi-Lingual Models]]
이 모델들은 OPUS-MT-Train 리포지토리의 구형 다국어 모델들입니다. 각 언어 그룹에 포함된 언어들은 다음과 같습니다:
```python no-style
['Helsinki-NLP/opus-mt-NORTH_EU-NORTH_EU',
'Helsinki-NLP/opus-mt-ROMANCE-en',
'Helsinki-NLP/opus-mt-SCANDINAVIA-SCANDINAVIA',
'Helsinki-NLP/opus-mt-de-ZH',
'Helsinki-NLP/opus-mt-en-CELTIC',
'Helsinki-NLP/opus-mt-en-ROMANCE',
'Helsinki-NLP/opus-mt-es-NORWAY',
'Helsinki-NLP/opus-mt-fi-NORWAY',
'Helsinki-NLP/opus-mt-fi-ZH',
'Helsinki-NLP/opus-mt-fi_nb_no_nn_ru_sv_en-SAMI',
'Helsinki-NLP/opus-mt-sv-NORWAY',
'Helsinki-NLP/opus-mt-sv-ZH']
GROUP_MEMBERS = {
'ZH': ['cmn', 'cn', 'yue', 'ze_zh', 'zh_cn', 'zh_CN', 'zh_HK', 'zh_tw', 'zh_TW', 'zh_yue', 'zhs', 'zht', 'zh'],
'ROMANCE': ['fr', 'fr_BE', 'fr_CA', 'fr_FR', 'wa', 'frp', 'oc', 'ca', 'rm', 'lld', 'fur', 'lij', 'lmo', 'es', 'es_AR', 'es_CL', 'es_CO', 'es_CR', 'es_DO', 'es_EC', 'es_ES', 'es_GT', 'es_HN', 'es_MX', 'es_NI', 'es_PA', 'es_PE', 'es_PR', 'es_SV', 'es_UY', 'es_VE', 'pt', 'pt_br', 'pt_BR', 'pt_PT', 'gl', 'lad', 'an', 'mwl', 'it', 'it_IT', 'co', 'nap', 'scn', 'vec', 'sc', 'ro', 'la'],
'NORTH_EU': ['de', 'nl', 'fy', 'af', 'da', 'fo', 'is', 'no', 'nb', 'nn', 'sv'],
'SCANDINAVIA': ['da', 'fo', 'is', 'no', 'nb', 'nn', 'sv'],
'SAMI': ['se', 'sma', 'smj', 'smn', 'sms'],
'NORWAY': ['nb_NO', 'nb', 'nn_NO', 'nn', 'nog', 'no_nb', 'no'],
'CELTIC': ['ga', 'cy', 'br', 'gd', 'kw', 'gv']
}
```
영어를 여러 로망스 언어로 번역하는 예제입니다. 여기서는 구형 2자리 언어 코드를 사용합니다:
```python
>>> from transformers import MarianMTModel, MarianTokenizer
>>> src_text = [
... ">>fr<< this is a sentence in english that we want to translate to french",
... ">>pt<< This should go to portuguese",
... ">>es<< And this to Spanish",
... ]
>>> model_name = "Helsinki-NLP/opus-mt-en-ROMANCE"
>>> tokenizer = MarianTokenizer.from_pretrained(model_name)
>>> model = MarianMTModel.from_pretrained(model_name)
>>> translated = model.generate(**tokenizer(src_text, return_tensors="pt", padding=True))
>>> tgt_text = [tokenizer.decode(t, skip_special_tokens=True) for t in translated]
["c'est une phrase en anglais que nous voulons traduire en français",
'Isto deve ir para o português.',
'Y esto al español']
```
## 자료[[Resources]]
- [번역 작업 가이드](../tasks/translation)
- [요약 작업 가이드](../tasks/summarization)
- [언어 모델링 작업 가이드](../tasks/language_modeling)
## MarianConfig
[[autodoc]] MarianConfig
## MarianTokenizer
[[autodoc]] MarianTokenizer
- build_inputs_with_special_tokens
<frameworkcontent>
<pt>
## MarianModel
[[autodoc]] MarianModel
- forward
## MarianMTModel
[[autodoc]] MarianMTModel
- forward
## MarianForCausalLM
[[autodoc]] MarianForCausalLM
- forward
</pt>
<tf>
## TFMarianModel
[[autodoc]] TFMarianModel
- call
## TFMarianMTModel
[[autodoc]] TFMarianMTModel
- call
</tf>
<jax>
## FlaxMarianModel
[[autodoc]] FlaxMarianModel
- __call__
## FlaxMarianMTModel
[[autodoc]] FlaxMarianMTModel
- __call__
</jax>
</frameworkcontent>
| transformers/docs/source/ko/model_doc/marian.md/0 | {
"file_path": "transformers/docs/source/ko/model_doc/marian.md",
"repo_id": "transformers",
"token_count": 5171
} | 396 |
<!--Copyright 2023 The Intel Team Authors and HuggingFace Inc. team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# TVP [[tvp]]
<div class="flex flex-wrap space-x-1">
<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
</div>
## 개요 [[overview]]
Text-Visual Prompting(TVP) 프레임워크는 Yimeng Zhang, Xin Chen, Jinghan Jia, Sijia Liu, Ke Ding이 발표한 논문 [Text-Visual Prompting for Efficient 2D Temporal Video Grounding](https://huggingface.co/papers/2303.04995)에서 제안되었습니다.
논문의 초록은 다음과 같습니다:
*본 논문에서는 길고, 편집되지 않은 비디오에서 문장으로 설명된 순간의 시작/종료 시점을 예측하는 것을 목표로 하는 Temporal Video Grounding(TVG) 문제를 다룹니다. 세밀한 3D 시각적 특징 덕분에 TVG 기술은 최근 몇 년 동안 놀라운 발전을 이뤘습니다. 하지만 3D 합성곱 신경망(CNN)의 높은 복잡성으로 인해 밀도 높은 3D 시각적 특징을 추출하는 데 시간이 오래 걸리고 그만큼 많은 메모리와 연산 자원을 필요로 합니다. 효율적인 TVG를 위해, 본 논문에서는 TVG 모델의 시각적 입력과 텍스트 특징 모두에 최적화된 교란 패턴('프롬프트'라고 부름)을 통합하는 새로운 Text-Visual Prompting(TVP) 프레임워크를 제안합니다. 3D CNN과 뚜렷이 대비되게 TVP가 2D TVG 모델에서 비전 인코더와 언어 인코더를 효과적으로 공동 학습할 수 있게 하고, 낮은 복잡도의 희소한 2D 시각적 특징만을 사용하여 크로스 모달 특징 융합의 성능을 향상시킵니다. 더 나아가, TVG의 효율적인 학습을 위해 Temporal-Distance IoU(TDIoU) 손실 함수를 제안합니다. 두 개의 벤치마크 데이터 세트인 Charades-STA와 ActivityNet Captions 데이터셋에 대한 실험을 통해, 제안된 TVP가 2D TVG의 성능을 크게 향상시키고(예: Charades-STA에서 9.79% 향상, ActivityNet Captions에서 30.77% 향상) 3D 시각적 특징을 사용하는 TVG에 비해 5배의 추론 가속을 달성함을 실험적으로 입증합니다.*
이 연구는 Temporal Video Grounding(TVG)을 다룹니다. TVG는 문장으로 설명된 특정 이벤트의 시작 및 종료 시점을 긴 비디오에서 정확히 찾아내는 과정입니다. TVG 성능을 향상시키기 위해 Text-Visual Prompting(TVP)이 제안되었습니다. TVP는 '프롬프트'라고 알려진 특별히 설계된 패턴을 TVG 모델의 시각적(이미지 기반) 및 텍스트(단어 기반) 입력 구성 요소 모두에 통합하는 것을 방식입니다. 이 프롬프트는 추가적인 시공간적 컨텍스트를 제공함으로써 모델이 비디오 내 이벤트 시점의 예측 정확도를 높입니다. 이 접근 방식은 3D 시각적 입력 대신 2D 입력을 사용합니다. 3D 입력은 보다 풍부한 시공간적 세부 정보를 제공하지만 처리하는 데 시간이 더 많이 걸립니다. 따라서 프롬프팅 메소드와 함께 2D 입력을 사용하여 이와 유사한 수준의 컨텍스트와 정확도를 더 효율적으로 제공하는 것을 목표로 합니다.
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/tvp_architecture.png"
alt="drawing" width="600"/>
<small> TVP 아키텍처. <a href="https://huggingface.co/papers/2303.04995">원본 논문에서 발췌.</a> </small>
이 모델은 [Jiqing Feng](https://huggingface.co/Jiqing)님이 기여했습니다. 원본 코드는 [이 곳](https://github.com/intel/TVP)에서 찾을 수 있습니다.
## 사용 팁 및 예시 [[usage-tips-and-examples]]
프롬프트는 최적화된 교란 패턴으로 입력 비디오 프레임이나 텍스트 특징에 추가되는 패턴입니다. 범용 세트란 모든 입력에 대해 동일한 프롬프트 세트를 사용하는 것을 말합니다. 즉, 입력 내용과 관계없이 모든 비디오 프레임과 텍스트 특징에 이 프롬프트들을 일관적으로 추가합니다.
TVP는 시각 인코더와 크로스 모달 인코더로 구성됩니다. 범용 시각 프롬프트와 텍스트 프롬프트 세트가 각각 샘플링된 비디오 프레임과 텍스트 특징에 통합됩니다. 특히, 서로 다른 시각 프롬프트 세트가 편집되지 않은 한 비디오에서 균일하게 샘플링된 프레임에 순서대로 적용됩니다.
이 모델의 목표는 학습 가능한 프롬프트를 시각적 입력과 텍스트 특징 모두에 통합하여 Temporal Video Grounding(TVG) 문제를 해결하는 것입니다.
원칙적으로, 제안된 아키텍처에는 어떤 시각 인코더나 크로스 모달 인코더라도 적용할 수 있습니다.
[TvpProcessor]는 [BertTokenizer]와 [TvpImageProcessor]를 단일 인스턴스로 래핑하여 텍스트를 인코딩하고 이미지를 각각 준비합니다.
다음 예시는 [TvpProcessor]와 [TvpForVideoGrounding]을 사용하여 TVG를 실행하는 방법을 보여줍니다.
```python
import av
import cv2
import numpy as np
import torch
from huggingface_hub import hf_hub_download
from transformers import AutoProcessor, TvpForVideoGrounding
def pyav_decode(container, sampling_rate, num_frames, clip_idx, num_clips, target_fps):
'''
원본 fps의 비디오를 지정한 fps(target_fps)로 변환하고 PyAV 디코더로 비디오를 디코딩합니다.
Args:
container (container): pyav 컨테이너 객체입니다.
sampling_rate (int): 프레임 샘플링 속도입니다.(샘플링된 두개의 프레임 사이의 간격을 말합니다)
num_frames (int): 샘플링할 프레임 수입니다.
clip_idx (int): clip_idx가 -1이면 시간 축에서 무작위 샘플링을 수행합니다.
clip_idx가 -1보다 크면 비디오를 num_clips 개로 균등 분할한 후
clip_idx번째 비디오 클립을 선택합니다.
num_clips (int): 주어진 비디오에서 균일하게 샘플링할 전체 클립 수입니다.
target_fps (int): 입력 비디오의 fps가 다를 수 있으므로, 샘플링 전에
지정한 fps로 변환합니다
Returns:
frames (tensor): 비디오에서 디코딩된 프레임입니다. 비디오 스트림을 찾을 수 없는 경우
None을 반환합니다.
fps (float): 비디오의 초당 프레임 수입니다.
'''
video = container.streams.video[0]
fps = float(video.average_rate)
clip_size = sampling_rate * num_frames / target_fps * fps
delta = max(num_frames - clip_size, 0)
start_idx = delta * clip_idx / num_clips
end_idx = start_idx + clip_size - 1
timebase = video.duration / num_frames
video_start_pts = int(start_idx * timebase)
video_end_pts = int(end_idx * timebase)
seek_offset = max(video_start_pts - 1024, 0)
container.seek(seek_offset, any_frame=False, backward=True, stream=video)
frames = {}
for frame in container.decode(video=0):
if frame.pts < video_start_pts:
continue
frames[frame.pts] = frame
if frame.pts > video_end_pts:
break
frames = [frames[pts] for pts in sorted(frames)]
return frames, fps
def decode(container, sampling_rate, num_frames, clip_idx, num_clips, target_fps):
'''
비디오를 디코딩하고 시간 축 샘플링을 수행합니다.
Args:
container (container): pyav 컨테이너 객체입니다.
sampling_rate (int): 프레임 샘플링 속도입니다.(샘플링된 두개의 프레임 사이의 간격을 말합니다)
num_frames (int): 샘플링할 프레임 수입니다.
clip_idx (int): clip_idx가 -1이면 시간 축에서 무작위 샘플링을 수행합니다.
clip_idx가 -1보다 크면 비디오를 num_clips 개로 균등 분할한 후
clip_idx번째 비디오 클립을 선택합니다.
num_clips (int): 주어진 비디오에서 균일하게 샘플링할 전체 클립 수입니다.
target_fps (int): 입력 비디오의 fps가 다를 수 있으므로, 샘플링 전에
지정한 fps로 변환합니다
Returns:
frames (tensor): 비디오에서 디코딩된 프레임입니다.
'''
assert clip_idx >= -2, "Not a valid clip_idx {}".format(clip_idx)
frames, fps = pyav_decode(container, sampling_rate, num_frames, clip_idx, num_clips, target_fps)
clip_size = sampling_rate * num_frames / target_fps * fps
index = np.linspace(0, clip_size - 1, num_frames)
index = np.clip(index, 0, len(frames) - 1).astype(np.int64)
frames = np.array([frames[idx].to_rgb().to_ndarray() for idx in index])
frames = frames.transpose(0, 3, 1, 2)
return frames
file = hf_hub_download(repo_id="Intel/tvp_demo", filename="AK2KG.mp4", repo_type="dataset")
model = TvpForVideoGrounding.from_pretrained("Intel/tvp-base")
decoder_kwargs = dict(
container=av.open(file, metadata_errors="ignore"),
sampling_rate=1,
num_frames=model.config.num_frames,
clip_idx=0,
num_clips=1,
target_fps=3,
)
raw_sampled_frms = decode(**decoder_kwargs)
text = "a person is sitting on a bed."
processor = AutoProcessor.from_pretrained("Intel/tvp-base")
model_inputs = processor(
text=[text], videos=list(raw_sampled_frms), return_tensors="pt", max_text_length=100#, size=size
)
model_inputs["pixel_values"] = model_inputs["pixel_values"].to(model.dtype)
output = model(**model_inputs)
def get_video_duration(filename):
cap = cv2.VideoCapture(filename)
if cap.isOpened():
rate = cap.get(5)
frame_num = cap.get(7)
duration = frame_num/rate
return duration
return -1
duration = get_video_duration(file)
start, end = processor.post_process_video_grounding(output.logits, duration)
print(f"The time slot of the video corresponding to the text \"{text}\" is from {start}s to {end}s")
```
팁:
- 이 TVP 구현은 텍스트 임베딩을 생성하기 위해 [BertTokenizer]를 사용하고, 시각적 임베딩을 계산하기 위해 Resnet-50 모델을 사용합니다.
- 사전 학습된 [tvp-base](https://huggingface.co/Intel/tvp-base)의 체크포인트가 공개되어 있습니다.
- 시간적 비디오 그라운딩 작업에 대한 TVP의 성능은 [표 2](https://huggingface.co/papers/2303.04995)를 참고하세요.
## TvpConfig [[transformers.TvpConfig]]
[[autodoc]] TvpConfig
## TvpImageProcessor [[transformers.TvpImageProcessor]]
[[autodoc]] TvpImageProcessor
- preprocess
## TvpProcessor [[transformers.TvpProcessor]]
[[autodoc]] TvpProcessor
- __call__
## TvpModel [[transformers.TvpModel]]
[[autodoc]] TvpModel
- forward
## TvpForVideoGrounding [[transformers.TvpForVideoGrounding]]
[[autodoc]] TvpForVideoGrounding
- forward
| transformers/docs/source/ko/model_doc/tvp.md/0 | {
"file_path": "transformers/docs/source/ko/model_doc/tvp.md",
"repo_id": "transformers",
"token_count": 7190
} | 397 |
<!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# 다중 GPU에서 효율적인 훈련 [[efficient-training-on-multiple-gpus]]
단일 GPU에서의 훈련이 너무 느리거나 모델 가중치가 단일 GPU의 메모리에 맞지 않는 경우, 다중-GPU 설정을 사용합니다. 단일 GPU에서 다중 GPU로 전환하기 위해서는 작업을 분산해야 합니다. 데이터, 텐서 또는 파이프라인과 같은 병렬화 기법을 사용하여 작업을 병렬로 처리할 수 있습니다. 그러나 이러한 설정을 모두에게 적용할 수 있는 완벽한 해결책은 없으며, 어떤 설정이 가장 적합한지는 사용하는 하드웨어에 따라 달라집니다. 이 문서는 주로 PyTorch 기반의 구현을 중심으로 설명하며, 대부분의 개념은 다른 프레임워크에도 적용될 수 있을 것으로 예상됩니다.
<Tip>
참고: [단일 GPU 섹션](perf_train_gpu_one)에서 소개된 전략(혼합 정밀도 훈련 또는 그래디언트 누적 등)은 일반적으로 모델 훈련에 적용되며, 다중-GPU 또는 CPU 훈련과 같은 다음 섹션으로 진입하기 전에 해당 섹션을 참고하는 것이 좋습니다.
</Tip>
먼저 1D 병렬화 기술에 대해 자세히 논의한 후, 이러한 기술을 결합하여 2D 및 3D 병렬화를 구현하여 더 빠른 훈련과 더 큰 모델을 지원하는 방법을 살펴볼 것입니다. 또한 다른 효과적인 대안 방식도 소개될 예정입니다.
## 개념 [[concepts]]
다음은 이 문서에서 자세히 설명될 주요 개념에 대한 간단한 설명입니다.
1. **DataParallel (DP)** - 동일한 설정이 여러 번 복제되고, 각 설정에 데이터 일부를 받습니다. 처리는 병렬로 수행되며 모든 설정은 각 훈련 단계의 끝날 때 동기화됩니다.
2. **TensorParallel (TP)** - 각 텐서는 여러 개의 묶음으로 분할되기에, 전체 텐서가 단일 GPU에 상주하는 대신 텐서의 각 샤드가 지정된 GPU에 상주합니다. 처리하는 동안 각 샤드는 서로 다른 GPU에서 개별적으로 병렬 처리되며 결과는 단계가 끝날 때 동기화됩니다. 분할이 수평 수준에서 이루어지기 때문에 이를 수평 병렬 처리라고 부를 수 있습니다.
3. **PipelineParallel (PP)** - 모델이 수직으로 (레이어 수준) 여러 GPU에 분할되어 모델의 단일 GPU에는 하나 또는 여러 레이어가 배치됩니다. 각 GPU는 파이프라인의 서로 다른 단계를 병렬로 처리하며 작은 배치 묶음에서 작동합니다.
4. **Zero Redundancy Optimizer (ZeRO)** - TP와 유사하게 텐서를 샤딩하지만, 전체 텐서는 순방향 또는 역방향 계산을 위해 재구성되므로 모델을 수정할 필요가 없습니다. 또한 제한된 GPU 메모리를 보완하기 위해 다양한 오프로드 기술을 지원합니다.
5. **Sharded DDP** - ZeRO의 기본 개념으로 다른 ZeRO 구현에서도 사용되는 용어입니다.
각 개념의 구체적인 내용에 대해 자세히 들어가기 전에 대규모 인프라에서 대규모 모델을 훈련하는 경우의 대략적인 결정 과정을 살펴보겠습니다.
## 확장성 전략 [[scalability-strategy]]
**⇨ 단일 노드 / 다중-GPU**
* 모델이 단일 GPU에 맞는 경우:
1. DDP - 분산 DP
2. ZeRO - 상황과 구성에 따라 더 빠를 수도 있고 그렇지 않을 수도 있음
* 모델이 단일 GPU에 맞지 않는 경우:
1. PP
2. ZeRO
3. TP
노드 내 연결 속도가 매우 빠른 NVLINK 또는 NVSwitch의 경우 세 가지 방법은 대부분 비슷한 성능을 보여야 하며, PP가 없는 경우 TP 또는 ZeRO보다 빠를 것입니다. TP의 정도도 차이를 만들 수 있습니다. 특정 설정에서 승자를 찾기 위해 실험하는 것이 가장 좋습니다.
TP는 거의 항상 단일 노드 내에서 사용됩니다. 즉, TP 크기 <= 노드당 GPU 수입니다.
* 가장 큰 레이어가 단일 GPU에 맞지 않는 경우:
1. ZeRO를 사용하지 않는 경우 - PP만으로는 맞지 않으므로 TP를 반드시 사용해야 함
2. ZeRO를 사용하는 경우에는 위의 "단일 GPU" 항목과 동일
**⇨ 다중 노드 / 다중 GPU**
* 노드 간 연결 속도가 빠른 경우:
1. ZeRO - 모델에 대부분의 수정을 필요로 하지 않음
2. PP+TP+DP - 통신이 적지만 모델에 대대적인 변경이 필요함
* 노드 간 연결 속도가 느리며, GPU 메모리가 여전히 부족한 경우:
1. DP+PP+TP+ZeRO-1
## 데이터 병렬화 [[data-parallelism]]
2개의 GPU만으로도 대부분의 사용자들은 `DataParallel` (DP)과 `DistributedDataParallel` (DDP)을 통해 향상된 훈련 속도를 누릴 수 있습니다. 이는 PyTorch의 내장 기능입니다. 일반적으로 DDP를 사용하는 것이 좋으며, DP는 일부 모델에서 작동하지 않을 수 있으므로 주의해야 합니다. [PyTorch 문서](https://pytorch.org/docs/master/generated/torch.nn.DataParallel.html)에서도 DDP의 사용을 권장합니다.
### DP vs DDP [[dp-vs-ddp]]
`DistributedDataParallel` (DDP)은 일반적으로 `DataParallel` (DP)보다 빠르지만, 항상 그렇지는 않습니다:
* DP는 파이썬 스레드 기반인 반면, DDP는 다중 프로세스 기반이기 때문에 GIL과 같은 파이썬 스레드 제한이 없습니다.
* 그러나 GPU 카드 간의 느린 상호 연결성은 DDP로 인해 실제로 느린 결과를 낼 수 있습니다.
이 두 모드 간의 GPU 간 통신 오버헤드의 주요 차이점은 다음과 같습니다:
[DDP](https://pytorch.org/docs/master/notes/ddp.html):
- 시작할 때, 주 프로세스가 모델을 gpu 0에서 다른 모든 gpu로 복제합니다.
- 그런 다음 각 배치에 대해:
1. 각 gpu는 자체 미니 배치 데이터를 직접 사용합니다.
2. `backward` 동안 로컬 그래디언트가 준비되면, 모든 프로세스에 평균화됩니다.
[DP](https://pytorch.org/docs/master/generated/torch.nn.DataParallel.html):
각 배치에 대해:
1. gpu 0은 데이터 배치를 읽고 각 gpu에 미니 배치를 보냅니다.
2. 업데이트된 모델을 gpu 0에서 각 gpu로 복제합니다.
3. `forward`를 실행하고 각 gpu의 출력을 gpu 0으로 보내고 손실을 계산합니다.
4. gpu 0에서 모든 gpu로 손실을 분산하고 `backward`를 실행합니다.
5. 각 gpu에서 그래디언트를 gpu 0으로 보내고 이를 평균화합니다.
DDP는 각 배치마다 그래디언트를 보내는 통신만을 수행하며, DP는 배치마다 5개의 다른 데이터 교환을 수행합니다.
DP는 파이썬 스레드를 통해 프로세스 내에서 데이터를 복제하며, DDP는 [torch.distributed](https://pytorch.org/docs/master/distributed.html)를 통해 데이터를 복제합니다.
DP에서는 gpu 0이 다른 gpu보다 훨씬 더 많은 작업을 수행하므로, gpu의 활용도가 낮아집니다.
DDP는 여러 대의 컴퓨터에서 사용할 수 있지만, DP의 경우는 그렇지 않습니다.
DP와 DDP 사이에는 다른 차이점이 있지만, 이 토론과는 관련이 없습니다.
이 2가지 모드를 깊게 이해하고 싶다면, [이 문서](https://www.telesens.co/2019/04/04/distributed-data-parallel-training-using-pytorch-on-aws/)를 강력히 추천합니다. 이 문서는 멋진 다이어그램을 포함하고 있으며, 다양한 하드웨어에서 여러 벤치마크와 프로파일러 출력을 설명하여 필요한 세부 사항을 모두 설명합니다.
실제 벤치마크를 살펴보겠습니다:
| Type | NVlink | Time |
| :----- | ----- | ---: |
| 2:DP | Y | 110s |
| 2:DDP | Y | 101s |
| 2:DDP | N | 131s |
분석:
여기서 DP는 NVlink가 있는 DDP보다 약 10% 느립니다. 그러나 NVlink가 없는 DDP보다 약 15% 빠릅니다.
실제 차이는 각 GPU가 다른 GPU와 동기화해야 하는 데이터 양에 따라 달라질 것입니다. 동기화할 데이터가 많을수록 느린 링크가 총 실행 시간을 늦출 수 있습니다.
다음은 전체 벤치마크 코드와 출력입니다:
해당 벤치마크에서 `NCCL_P2P_DISABLE=1`을 사용하여 NVLink 기능을 비활성화했습니다.
```bash
# DP
rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 \
python examples/pytorch/language-modeling/run_clm.py \
--model_name_or_path openai-community/gpt2 --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 \
--do_train --output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200
{'train_runtime': 110.5948, 'train_samples_per_second': 1.808, 'epoch': 0.69}
# DDP w/ NVlink
rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 \
torchrun --nproc_per_node 2 examples/pytorch/language-modeling/run_clm.py \
--model_name_or_path openai-community/gpt2 --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 \
--do_train --output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200
{'train_runtime': 101.9003, 'train_samples_per_second': 1.963, 'epoch': 0.69}
# DDP w/o NVlink
rm -r /tmp/test-clm; NCCL_P2P_DISABLE=1 CUDA_VISIBLE_DEVICES=0,1 \
torchrun --nproc_per_node 2 examples/pytorch/language-modeling/run_clm.py \
--model_name_or_path openai-community/gpt2 --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 \
--do_train --output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200
{'train_runtime': 131.4367, 'train_samples_per_second': 1.522, 'epoch': 0.69}
```
하드웨어: 각각 24GB의 TITAN RTX 2개 + NVlink과 2개의 NVLink (`nvidia-smi topo -m`에서 `NV2`입니다.)
소프트웨어: `pytorch-1.8-to-be` + `cuda-11.0` / `transformers==4.3.0.dev0`
## ZeRO 데이터 병렬화 [[zero-data-parallelism]]
ZeRO를 기반으로 한 데이터 병렬화 (ZeRO-DP)는 다음 [블로그 글](https://www.microsoft.com/en-us/research/blog/zero-deepspeed-new-system-optimizations-enable-training-models-with-over-100-billion-parameters/)의 다음 다이어그램에서 설명되고 있습니다.
이 개념은 이해하기 어려울 수 있지만, 실제로는 매우 간단한 개념입니다. 이는 일반적인 `DataParallel` (DP)과 동일하지만, 전체 모델 매개변수, 그래디언트 및 옵티마이저 상태를 복제하는 대신 각 GPU는 그 중 일부만 저장합니다. 그리고 실행 시간에는 주어진 레이어에 대해 전체 레이어 매개변수가 필요할 때 각 GPU가 서로에게 필요한 부분을 제공하기 위해 동기화됩니다 - 그게 전부입니다.
각각 3개의 레이어와 3개의 매개변수가 있는 간단한 모델을 생각해 봅시다:
```
La | Lb | Lc
---|----|---
a0 | b0 | c0
a1 | b1 | c1
a2 | b2 | c2
```
레이어 La에는 가중치 a0, a1 및 a2가 있습니다.
3개의 GPU가 있는 경우, Sharded DDP (= Zero-DP)는 다음과 같이 모델을 3개의 GPU에 분할합니다:
```
GPU0:
La | Lb | Lc
---|----|---
a0 | b0 | c0
GPU1:
La | Lb | Lc
---|----|---
a1 | b1 | c1
GPU2:
La | Lb | Lc
---|----|---
a2 | b2 | c2
```
일반적인 DNN 다이어그램을 상상해보면 이는 텐서 병렬 처리와 같은 수평 슬라이싱입니다. 수직 슬라이싱은 전체 레이어 그룹을 다른 GPU에 배치하는 것입니다. 이는 시작에 불과합니다.
이제 이러한 각각의 GPU는 DP에서 작동하는 것과 마찬가지로 일반적인 미니 배치를 받습니다:
```
x0 => GPU0
x1 => GPU1
x2 => GPU2
```
입력은 수정되지 않은 상태로 일반 모델에 의해 처리될 것으로 간주합니다.
먼저, 입력은 레이어 La에 도달합니다.
GPU0에만 집중해 보겠습니다. x0은 순방향 경로를 수행하기 위해 a0, a1, a2 파라미터가 필요하지만 GPU0에는 a0만 있습니다. GPU1에서 a1을, GPU2에서 a2를 전송받아 모델의 모든 조각을 하나로 모읍니다.
병렬적으로, GPU1은 미니 배치 x1을 받고 a1만 가지고 있지만, a0 및 a2 매개변수가 필요합니다. 따라서 GPU0 및 GPU2에서 이를 가져옵니다.
GPU2도 동일한 작업을 수행합니다. 입력 x2를 받고 GPU0 및 GPU1에서 각각 a0과 a1을, 그리고 자신의 a2와 함께 전체 텐서를 복원합니다.
3개의 GPU는 복원된 전체 텐서를 받고 forward가 수행됩니다.
계산이 완료되면 더 이상 필요하지 않은 데이터는 삭제되고, 해당 데이터는 계산 중에만 사용됩니다. 복원은 사전 패치를 통해 효율적으로 수행됩니다.
그리고 전체 프로세스는 레이어 Lb에 대해 반복되고, 그 다음 Lc로 순방향으로, 그다음은 역방향으로 Lc -> Lb -> La로 반복됩니다.
개인적으로 이것은 효율적인 그룹 배낭 여행자의 중량 분배 전략처럼 들립니다:
1. 사람 A가 텐트를 운반합니다.
2. 사람 B가 난로를 운반합니다.
3. 사람 C가 도끼를 운반합니다.
이제 매일 밤 각자 가진 것을 다른 사람들과 공유하고, 가지지 않은 것은 다른 사람들로부터 받고, 아침에는 할당된 유형의 장비를 싸고 계속해서 여행을 진행합니다. 이것이 Sharded DDP / Zero DP입니다.
이 전략을 각각 자신의 텐트, 난로 및 도끼를 개별적으로 운반해야 하는 단순한 전략과 비교해보면 훨씬 비효율적일 것입니다. 이것이 Pytorch의 DataParallel (DP 및 DDP)입니다.
이 주제에 대해 논문을 읽을 때 다음 동의어를 만날 수 있습니다: Sharded, Partitioned.
ZeRO가 모델 가중치를 분할하는 방식을 자세히 살펴보면, 텐서 병렬화와 매우 유사한 것을 알 수 있습니다. 이는 이후에 설명될 수직 모델 병렬화와는 달리 각 레이어의 가중치를 분할/분할하기 때문입니다.
구현:
- [DeepSpeed](https://www.deepspeed.ai/tutorials/zero/)는 1단계 + 2단계 + 3단계의 ZeRO-DP를 제공합니다.
- [Fairscale](https://github.com/facebookresearch/fairscale/#optimizer-state-sharding-zero)은 1단계 + 2단계 + 3단계의 ZeRO-DP를 제공합니다.
- [`transformers` 통합](main_classes/trainer#trainer-integrations)
## 네이티브 모델 병렬 처리(수직적) 및 파이프라인 병렬 처리[[naive-model-parallelism-vertical-and-pipeline-parallelism]]
Naive Model Parallelism (MP)은 모델 레이어 그룹을 다중 GPU에 분산하는 방식입니다. 메커니즘은 상대적으로 간단합니다. 원하는 레이어를 `.to()`를 사용하여 원하는 장치로 전환하면 데이터가 해당 레이어로 들어오고 나갈 때 데이터도 레이어와 동일한 장치로 전환되고 나머지는 수정되지 않습니다.
대부분의 모델이 그려지는 방식이 레이어를 세로로 슬라이스하기 때문에 이를 수직 모델 병렬화라고 부릅니다. 예를 들어 다음 다이어그램은 8레이어 모델을 보여줍니다:
```
=================== ===================
| 0 | 1 | 2 | 3 | | 4 | 5 | 6 | 7 |
=================== ===================
gpu0 gpu1
```
우리는 모델을 수직으로 2개로 분할하여 레이어 0-3을 GPU0에 배치하고 레이어 4-7을 GPU1에 배치했습니다.
이제 데이터가 레이어 0에서 1로, 1에서 2로, 2에서 3으로 이동하는 동안에는 일반적인 모델입니다. 그러나 데이터가 레이어 3에서 레이어 4로 전달되어야 할 때는 GPU0에서 GPU1로 이동해야 하므로 통신 오버헤드가 발생합니다. 참여하는 GPU가 동일한 컴퓨팅 노드(예: 동일한 물리적인 기계)에 있는 경우 이 복사는 매우 빠릅니다. 그러나 GPU가 서로 다른 컴퓨팅 노드(예: 여러 기계)에 위치한 경우 통신 오버헤드는 상당히 크게 될 수 있습니다.
그런 다음 레이어 4부터 5로, 6으로, 7로 진행되는 것은 일반적인 모델과 동일하게 진행되고, 7번째 레이어가 완료되면 데이터를 다시 레이어 0으로 보내거나 또는 레이블을 마지막 레이어로 보내야 할 필요가 있습니다. 이제 손실을 계산하고 옵티마이저가 작동할 수 있습니다.
문제점:
- 이 방식을 "naive" MP라고 부르는 이유는 주어진 상황에 하나의 GPU를 제외한 모든 GPU가 유휴 상태라는 점입니다. 따라서 4개의 GPU를 사용하는 경우 단일 GPU의 메모리 양을 4배로 늘리고 나머지 하드웨어는 무시하는 것과 거의 동일합니다. 또한 장치 간 데이터 복사의 오버헤드도 있습니다. 따라서 4개의 6GB 카드는 naive MP를 사용하여 1개의 24GB 카드와 동일한 크기를 수용할 수 있지만, 후자는 데이터 복사의 오버헤드가 없으므로 훈련을 더 빨리 완료합니다. 그러나 예를 들어 40GB 카드가 있고 45GB 모델을 맞추어야 할 경우 4개의 40GB 카드로 맞출 수 있습니다 (하지만 그래디언트와 옵티마이저 상태 때문에 가까스로 가능합니다).
- 공유 임베딩은 GPU 간에 복사해야 할 수도 있습니다.
파이프라인 병렬화 (PP)은 거의 naive MP와 동일하지만 GPU 유휴 상태 문제를 해결하기 위해 들어오는 배치를 마이크로 배치로 나누고 인공적으로 파이프라인을 생성하여 서로 다른 GPU가 동시에 계산에 참여할 수 있게 합니다.
[GPipe 논문](https://ai.googleblog.com/2019/03/introducing-gpipe-open-source-library.html)에서 가져온 그림은 상단에 naive MP를, 하단에는 PP를 보여줍니다:
하단 다이어그램에서 PP가 유휴 영역이 적은 것을 쉽게 볼 수 있습니다. 유휴 부분을 "bubble"이라고 합니다.
다이어그램의 양쪽 부분은 참여하는 GPU가 4개인 병렬성을 보여줍니다. 즉, 4개의 GPU가 파이프라인에 참여합니다. 따라서 4개의 파이프 단계 F0, F1, F2 및 F3의 순방향 경로와 B3, B2, B1 및 B0의 역방향 경로가 있습니다.
PP는 조정해야 할 새로운 하이퍼파라미터인 `chunks`를 도입합니다. 이는 동일한 파이프 단계를 통해 일련의 데이터를 묶어서 보내는 방식을 정의합니다. 예를 들어, 아래 다이어그램에서 `chunks=4`를 볼 수 있습니다. GPU0은 0, 1, 2 및 3 (F0,0, F0,1, F0,2, F0,3) 묶음에서 동일한 순방향 경로를 수행하고, 다른 GPU가 작업을 수행하기 시작하고 완료가 시작될 때만 GPU0이 묶음의 역순으로 3, 2, 1 및 0 (B0,3, B0,2, B0,1, B0,0) 경로를 수행합니다.
개념적으로 이는 그래디언트 누적 단계 (GAS)와 동일한 개념입니다. 파이토치에서는 `chunks`를 사용하고 DeepSpeed에서는 동일한 하이퍼파라미터를 GAS로 참조합니다.
묶음으로 인해 PP는 마이크로 배치 (MBS)의 개념을 도입합니다. DP는 전역 데이터 배치 크기를 미니 배치로 나눕니다. 따라서 DP 차수가 4이고 전역 배치 크기가 1024이면 256씩 4개의 미니 배치로 분할됩니다 (1024/4). 그리고 `chunks` (또는 GAS)의 수가 32이면 마이크로 배치 크기는 8이 됩니다 (256/32). 각 파이프라인 단계는 한 번에 하나의 마이크로 배치와 함께 작동합니다.
DP + PP 설정의 전역 배치 크기를 계산하려면 `mbs*chunks*dp_degree` (`8*32*4=1024`)를 수행합니다.
다이어그램으로 돌아가 보겠습니다.
`chunks=1`로 설정하면 매우 비효율적인 naive MP가 생성되며, 매우 큰 `chunks` 값으로 설정하면 아주 작은 마이크로 배치 크기가 생성되어 효율적이지 않을 수 있습니다. 따라서 가장 효율적인 GPU 활용을 위해 어떤 값이 가장 적절한지 실험을 해야 합니다.
다이어그램에서 보이는 것처럼 "dead" 시간의 버블이 존재하여 마지막 `forward` 단계가 `backward` 단계가 파이프라인을 완료하기를 기다려야 하는 상황이 발생하지만, `chunks`의 가장 적절한 값을 찾는 것의 목적은 모든 참여하는 GPU에서 동시에 고도로 활용되는 GPU 활용을 가능하게 하여 버블의 크기를 최소화하는 것입니다.
해결책은 전통적인 파이프라인 API와 더 현대적인 솔루션으로 나뉩니다. 전통적인 파이프라인 API 솔루션과 현대적인 솔루션에 대해 알아보겠습니다.
전통적인 파이프라인 API 솔루션:
- 파이토치
- FairScale
- DeepSpeed
- Megatron-LM
현대적인 솔루션:
- Varuna
- Sagemaker
전통적인 파이프라인 API 솔루션의 문제점:
- 모델을 상당히 수정해야 한다는 점이 문제입니다. 파이프라인은 모듈의 정상적인 흐름을 `nn.Sequential` 시퀀스로 다시 작성해야 하므로 모델의 설계를 변경해야 할 수 있습니다.
- 현재 파이프라인 API는 매우 제한적입니다. 파이프라인의 매우 첫 번째 단계에서 전달되는 많은 파이썬 변수가 있는 경우 이를 해결해야 합니다. 현재 파이프라인 인터페이스는 하나의 텐서 또는 텐서의 튜플을 유일한 입력 및 출력으로 요구합니다. 이러한 텐서는 마이크로 배치로 미니 배치로 묶을 것이므로 첫 번째 차원으로 배치 크기가 있어야 합니다. 가능한 개선 사항은 여기에서 논의되고 있습니다. https://github.com/pytorch/pytorch/pull/50693
- 파이프 단계 수준에서 조건부 제어 흐름은 불가능합니다. 예를 들어, T5와 같은 인코더-디코더 모델은 조건부 인코더 단계를 처리하기 위해 특별한 해결책이 필요합니다.
- 각 레이어를 정렬하여 하나의 모델의 출력이 다른 모델의 입력이 되도록해야 합니다.
우리는 아직 Varuna와 SageMaker로 실험하지 않았지만, 해당 논문들은 위에서 언급한 문제들의 목록을 극복했고 사용자의 모델에 대한 변경 사항이 훨씬 적게 필요하다고 보고하고 있습니다.
구현:
- [파이토치](https://pytorch.org/docs/stable/pipeline.html) (파이토치-1.8에서 초기 지원, 1.9에서 점진적으로 개선되고 1.10에서 더 개선됨). [예제](https://github.com/pytorch/pytorch/blob/master/benchmarks/distributed/pipeline/pipe.py)도 참고하세요.
- [FairScale](https://fairscale.readthedocs.io/en/latest/tutorials/pipe.html)
- [DeepSpeed](https://www.deepspeed.ai/tutorials/pipeline/)
- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)은 내부 구현을 가지고 있습니다 - API 없음.
- [Varuna](https://github.com/microsoft/varuna)
- [SageMaker](https://huggingface.co/papers/2111.05972) - 이는 AWS에서만 사용할 수 있는 소유 솔루션입니다.
- [OSLO](https://github.com/tunib-ai/oslo) - 이는 Hugging Face Transformers를 기반으로 구현된 파이프라인 병렬화입니다.
🤗 Transformers 상태: 이 작성 시점에서 모델 중 어느 것도 완전한 PP를 지원하지 않습니다. GPT2와 T5 모델은 naive MP를 지원합니다. 주요 장애물은 모델을 `nn.Sequential`로 변환하고 모든 입력을 텐서로 가져와야 하는 것을 처리할 수 없기 때문입니다. 현재 모델에는 이러한 변환을 매우 복잡하게 만드는 많은 기능이 포함되어 있어 제거해야 합니다.
기타 접근 방법:
DeepSpeed, Varuna 및 SageMaker는 [교차 파이프라인(Interleaved Pipeline)](https://docs.aws.amazon.com/sagemaker/latest/dg/model-parallel-core-features.html) 개념을 사용합니다.
여기서는 버블(유휴 시간)을 역방향 패스에 우선순위를 부여하여 최소화합니다.
Varuna는 가장 효율적인 스케줄링을 찾기 위해 시뮬레이션을 사용하여 스케줄을 개선하려고 합니다.
OSLO는 `nn.Sequential`로 변환하지 않고 Transformers를 기반으로 한 파이프라인 병렬화를 구현했습니다.
## 텐서 병렬 처리 [[tensor-parallelism]]
텐서 병렬 처리에서는 각 GPU가 텐서의 일부분만 처리하고 전체 텐서가 필요한 연산에 대해서만 전체 텐서를 집계합니다.
이 섹션에서는 [Megatron-LM](https://github.com/NVIDIA/Megatron-LM) 논문인 [Efficient Large-Scale Language Model Training on GPU Clusters](https://huggingface.co/papers/2104.04473)에서의 개념과 다이어그램을 사용합니다.
Transformer의 주요 구성 요소는 fully connected `nn.Linear`와 비선형 활성화 함수인 `GeLU`입니다.
Megatron 논문의 표기법을 따라 행렬의 점곱 부분을 `Y = GeLU(XA)`로 표현할 수 있습니다. 여기서 `X`와 `Y`는 입력 및 출력 벡터이고 `A`는 가중치 행렬입니다.
행렬 형태로 계산을 살펴보면, 행렬 곱셈을 다중 GPU로 분할할 수 있는 방법을 쉽게 알 수 있습니다:
가중치 행렬 `A`를 `N`개의 GPU에 대해 열별로 분할하고 병렬로 행렬 곱셈 `XA_1`에서 `XA_n`까지 수행하면 `N`개의 출력 벡터 `Y_1, Y_2, ..., Y_n`가 생성되며 독립적으로 `GeLU`에 전달될 수 있습니다:
이 원리를 사용하여 동기화가 필요하지 않은 GPU 간의 임의 깊이의 MLP를 업데이트할 수 있습니다. 그러나 결과 벡터를 샤드로부터 재구성해야 하는 마지막 단계까지는 GPU 간의 동기화가 필요합니다. Megatron-LM 논문의 저자들은 이에 대한 유용한 그림을 제공합니다:
다중 헤드 어텐션 레이어의 병렬화는 더욱 간단합니다. 이미 독립적인 다중 헤드를 가지고 있기 때문에 이미 병렬화되어 있습니다!
특별 고려사항: TP는 매우 빠른 네트워크가 필요하므로 한 개 이상의 노드에서 TP를 수행하는 것은 권장되지 않습니다. 실제로 노드에 4개의 GPU가 있는 경우 TP의 최대 차수는 4입니다. TP 차수가 8인 경우 최소한 8개의 GPU가 있는 노드를 사용해야 합니다.
이 섹션은 원래의 [더 자세한 TP 개요](https://github.com/huggingface/transformers/issues/10321#issuecomment-783543530)를 기반으로 합니다.
작성자는 [@anton-l](https://github.com/anton-l)입니다.
SageMaker는 더 효율적인 처리를 위해 TP와 DP를 결합합니다.
대체 이름:
- DeepSpeed는 이를 [텐서 슬라이싱](https://www.deepspeed.ai/training/#model-parallelism)이라고 부릅니다.
구현:
- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)은 내부 구현을 가지고 있으므로 모델에 매우 특화되어 있습니다.
- [parallelformers](https://github.com/tunib-ai/parallelformers) (현재는 추론에만 해당)
- [SageMaker](https://huggingface.co/papers/2111.05972) - 이는 AWS에서만 사용할 수 있는 소유 솔루션입니다.
- [OSLO](https://github.com/tunib-ai/oslo)은 Transformers를 기반으로 한 텐서 병렬 처리 구현을 가지고 있습니다.
🤗 Transformers 현황:
- core: 아직 핵심 부분에 구현되지 않음
- 그러나 추론을 하려면 [parallelformers](https://github.com/tunib-ai/parallelformers)가 대부분의 모델을 지원합니다. 따라서 핵심 부분에 구현되기 전까지 그들의 것을 사용할 수 있습니다. 그리고 훈련 모드도 지원될 예정입니다.
- Deepspeed-Inference는 CUDA 커널을 기반으로 하는 매우 빠른 추론 모드에서 BERT, GPT-2 및 GPT-Neo 모델을 지원합니다. 자세한 내용은 [여기](https://www.deepspeed.ai/tutorials/inference-tutorial/)를 참조하세요.
## DP+PP [[dppp]]
DeepSpeed [pipeline tutorial](https://www.deepspeed.ai/tutorials/pipeline/)에서 다음 다이어그램은 DP와 PP를 결합하는 방법을 보여줍니다.
여기서 DP 랭크 0은 GPU2를 보지 못하고, DP 랭크 1은 GPU3을 보지 못하는 것이 중요합니다. DP에게는 딱 2개의 GPU인 것처럼 데이터를 공급합니다. GPU0은 PP를 사용하여 GPU2에게 일부 작업을 "비밀리에" 할당합니다. 그리고 GPU1도 GPU3을 도움으로 삼아 같은 방식으로 작업합니다.
각 차원마다 적어도 2개의 GPU가 필요하므로 최소한 4개의 GPU가 필요합니다.
구현:
- [DeepSpeed](https://github.com/deepspeedai/DeepSpeed)
- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)
- [Varuna](https://github.com/microsoft/varuna)
- [SageMaker](https://huggingface.co/papers/2111.05972)
- [OSLO](https://github.com/tunib-ai/oslo)
🤗 Transformers 현황: 아직 구현되지 않음
## DP+PP+TP [[dppptp]]
더 효율적인 훈련을 위해 PP와 TP 및 DP를 결합하여 3D 병렬 처리를 사용합니다. 다음 다이어그램에서 이를 확인할 수 있습니다.
이 다이어그램은 [3D parallelism: Scaling to trillion-parameter models](https://www.microsoft.com/en-us/research/blog/deepspeed-extreme-scale-model-training-for-everyone/)이라는 블로그 글에서 확인할 수 있습니다.
각 차원마다 적어도 2개의 GPU가 필요하므로 최소한 8개의 GPU가 필요합니다.
구현:
- [DeepSpeed](https://github.com/deepspeedai/DeepSpeed) - DeepSpeed는 더욱 효율적인 DP인 ZeRO-DP라고도 부릅니다.
- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)
- [Varuna](https://github.com/microsoft/varuna)
- [SageMaker](https://huggingface.co/papers/2111.05972)
- [OSLO](https://github.com/tunib-ai/oslo)
🤗 Transformers 현황: 아직 구현되지 않음. PP와 TP가 없기 때문입니다.
## ZeRO DP+PP+TP [[zero-dppptp]]
DeepSpeed의 주요 기능 중 하나는 DP의 확장인 ZeRO입니다. ZeRO-DP에 대해 이미 [ZeRO Data Parallelism](#zero-data-parallelism)에서 논의되었습니다. 일반적으로 이는 PP나 TP를 필요로하지 않는 독립적인 기능입니다. 그러나 PP와 TP와 결합할 수도 있습니다.
ZeRO-DP가 PP와 (선택적으로 TP와) 결합되면 일반적으로 ZeRO 단계 1(옵티마이저 분할)만 활성화됩니다.
이론적으로는 ZeRO 단계 2(그라디언트 분할)를 파이프라인 병렬 처리와 함께 사용할 수도 있지만, 이는 성능에 나쁜 영향을 미칠 것입니다. 각 마이크로 배치마다 그라디언트를 샤딩하기 전에 추가적인 리듀스-스캐터 컬렉티브가 필요하며, 이는 잠재적으로 상당한 통신 오버헤드를 추가합니다. 파이프라인 병렬 처리의 특성상 작은 마이크로 배치가 사용되며, 산술 연산 강도(마이크로 배치 크기)를 균형 있게 유지하면서 파이프라인 버블(마이크로 배치 수)을 최소화하는 것에 중점을 둡니다. 따라서 해당 통신 비용은 문제가 될 것입니다.
또한, PP로 인해 정상보다 적은 수의 레이어가 있으므로 메모리 절약은 크지 않을 것입니다. PP는 이미 그래디언트 크기를 ``1/PP``로 줄이기 때문에 그래디언트 샤딩의 절약 효과는 순수 DP보다는 미미합니다.
ZeRO 단계 3도 같은 이유로 좋은 선택이 아닙니다 - 더 많은 노드 간 통신이 필요합니다.
그리고 ZeRO가 있기 때문에 다른 이점은 ZeRO-Offload입니다. 이는 단계 1이므로 옵티마이저 상태를 CPU로 오프로드할 수 있습니다.
구현:
- [Megatron-DeepSpeed](https://github.com/microsoft/Megatron-DeepSpeed) 및 [BigScience의 Megatron-Deepspeed](https://github.com/bigscience-workshop/Megatron-DeepSpeed), 이전 저장소의 포크입니다.
- [OSLO](https://github.com/tunib-ai/oslo)
중요한 논문:
- [Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model](
https://huggingface.co/papers/2201.11990)
🤗 Transformers 현황: 아직 구현되지 않음, PP와 TP가 없기 때문입니다.
## FlexFlow [[flexflow]]
[FlexFlow](https://github.com/flexflow/FlexFlow)는 약간 다른 방식으로 병렬화 문제를 해결합니다.
논문: ["Beyond Data and Model Parallelism for Deep Neural Networks" by Zhihao Jia, Matei Zaharia, Alex Aiken](https://huggingface.co/papers/1807.05358)
이는 Sample-Operator-Attribute-Parameter를 기반으로 하는 일종의 4D 병렬화를 수행합니다.
1. Sample = 데이터 병렬화 (샘플별 병렬)
2. Operator = 단일 연산을 여러 하위 연산으로 병렬화
3. Attribute = 데이터 병렬화 (길이별 병렬)
4. Parameter = 모델 병렬화 (수평 또는 수직과 관계없이)
예시:
* Sample
512 길이의 10개의 배치를 가정해 봅시다. 이를 sample 차원으로 2개의 장치에 병렬화하면, 10 x 512는 5 x 2 x 512가 됩니다.
* Operator
레이어 정규화를 수행한다면, 우선 std를 계산하고 두 번째로 mean을 계산한 다음 데이터를 정규화할 수 있습니다. Operator 병렬화는 std와 mean을 병렬로 계산할 수 있도록 합니다. 따라서 operator 차원으로 2개의 장치 (cuda:0, cuda:1)에 병렬화하면, 먼저 입력 데이터를 두 장치로 복사한 다음 cuda:0에서 std를 계산하고 cuda:1에서 동시에 mean을 계산합니다.
* Attribute
512 길이의 10개의 배치가 있습니다. 이를 attribute 차원으로 2개의 장치에 병렬화하면, 10 x 512는 10 x 2 x 256이 됩니다.
* Parameter
이는 tensor 모델 병렬화 또는 naive layer-wise 모델 병렬화와 유사합니다.
이 프레임워크의 중요한 점은 (1) GPU/TPU/CPU 대 (2) RAM/DRAM 대 (3) 빠른 인트라-커넥트 대 느린 인터-커넥트와 같은 리소스를 고려하여 어디에서 어떤 병렬화를 사용할지를 알고리즘적으로 자동으로 최적화한다는 것입니다.
하나 매우 중요한 측면은 FlexFlow가 정적이고 고정된 워크로드를 가진 모델에 대한 DNN 병렬화를 최적화하기 위해 설계되었다는 것입니다. 동적인 동작을 가진 모델은 반복마다 다른 병렬화 전략을 선호할 수 있습니다.
따라서 이 프레임워크의 장점은 선택한 클러스터에서 30분 동안 시뮬레이션을 실행하고 이 특정 환경을 최적으로 활용하기 위한 최상의 전략을 제안한다는 것입니다. 부품을 추가/제거/교체하면 실행하고 그에 대한 계획을 다시 최적화한 후 훈련할 수 있습니다. 다른 설정은 자체적인 사용자 정의 최적화를 가질 수 있습니다.
🤗 Transformers 현황: 아직 통합되지 않음. 이미 [transformers.utils.fx](https://github.com/huggingface/transformers/blob/master/src/transformers/utils/fx.py)를 통해 모델을 FX-추적할 수 있으며, 이는 FlexFlow의 선행 조건입니다. 따라서 어떤 작업을 수행해야 FlexFlow가 우리의 모델과 함께 작동할 수 있는지 파악해야 합니다.
## 어떤 전략을 사용해야 할까요? [[which-strategy-to-use-when]]
다음은 어떤 병렬화 전략을 언제 사용해야 하는지에 대한 매우 대략적인 개요입니다. 각 목록의 첫 번째 전략이 일반적으로 더 빠릅니다.
**⇨ 단일 GPU**
* 모델이 단일 GPU에 맞는 경우:
1. 일반적인 사용
* 모델이 단일 GPU에 맞지 않는 경우:
1. ZeRO + CPU 및 옵션으로 NVMe 언로드
2. 위와 동일하게 사용하되, 가장 큰 레이어가 단일 GPU에 맞지 않는 경우 Memory Centric Tiling(자세한 내용은 아래 참조)을 추가적으로 사용
* 가장 큰 레이어가 단일 GPU에 맞지 않는 경우:
1. ZeRO - [Memory Centric Tiling](https://deepspeed.readthedocs.io/en/latest/zero3.html#memory-centric-tiling) (MCT) 활성화. 이를 통해 크기가 매우 큰 레이어를 임의로 분할하여 순차적으로 실행할 수 있습니다. MCT는 GPU에 활성화된 매개변수의 수를 줄이지만 활성화 메모리에는 영향을 주지 않습니다. 현재 작성 기준으로 이 요구사항은 매우 드물기 때문에 사용자가 `torch.nn.Linear`를 수동으로 수정해야 합니다.
**⇨ 단일 노드 / 다중 GPU**
* 모델이 단일 GPU에 맞는 경우:
1. DDP - 분산 DP
2. ZeRO - 상황과 구성에 따라 빠를 수도 있고 그렇지 않을 수도 있습니다.
* 모델이 단일 GPU에 맞지 않는 경우:
1. PP
2. ZeRO
3. TP
NVLINK 또는 NVSwitch를 통한 매우 빠른 인트라-노드 연결이 있는 경우 이 세 가지 방법은 거의 동등할 것이며, 이러한 연결이 없는 경우 PP가 TP나 ZeRO보다 빠를 것입니다. 또한 TP의 차수도 영향을 줄 수 있습니다. 특정 설정에서 우승자를 찾기 위해 실험하는 것이 가장 좋습니다.
TP는 거의 항상 단일 노드 내에서 사용됩니다. 즉, TP 크기 <= 노드당 GPU 수입니다.
* 가장 큰 레이어가 단일 GPU에 맞지 않는 경우:
1. ZeRO를 사용하지 않을 경우 - PP만 사용할 수 없으므로 TP를 사용해야 합니다.
2. ZeRO를 사용할 경우, "단일 GPU"의 항목과 동일한 항목 참조
**⇨ 다중 노드 / 다중 GPU**
* 빠른 노드 간 연결이 있는 경우:
1. ZeRO - 모델에 대한 수정이 거의 필요하지 않습니다.
2. PP+TP+DP - 통신이 적지만 모델에 대한 대규모 변경이 필요합니다.
* 느린 노드 간 연결 및 GPU 메모리 부족한 경우:
1. DP+PP+TP+ZeRO-1
| transformers/docs/source/ko/perf_train_gpu_many.md/0 | {
"file_path": "transformers/docs/source/ko/perf_train_gpu_many.md",
"repo_id": "transformers",
"token_count": 28510
} | 398 |
<!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# 스크립트로 실행하기[[train-with-a-script]]
🤗 Transformers 노트북과 함께 [PyTorch](https://github.com/huggingface/transformers/tree/main/examples/pytorch), [TensorFlow](https://github.com/huggingface/transformers/tree/main/examples/tensorflow), 또는 [JAX/Flax](https://github.com/huggingface/transformers/tree/main/examples/flax)를 사용해 특정 태스크에 대한 모델을 훈련하는 방법을 보여주는 예제 스크립트도 있습니다.
또한 [연구 프로젝트](https://github.com/huggingface/transformers-research-projects/) 및 [레거시 예제](https://github.com/huggingface/transformers/tree/main/examples/legacy)에서 대부분 커뮤니티에서 제공한 스크립트를 찾을 수 있습니다.
이러한 스크립트는 적극적으로 유지 관리되지 않으며 최신 버전의 라이브러리와 호환되지 않을 가능성이 높은 특정 버전의 🤗 Transformers를 필요로 합니다.
예제 스크립트가 모든 문제에서 바로 작동하는 것은 아니며, 해결하려는 문제에 맞게 스크립트를 변경해야 할 수도 있습니다.
이를 위해 대부분의 스크립트에는 데이터 전처리 방법이 나와있어 필요에 따라 수정할 수 있습니다.
예제 스크립트에 구현하고 싶은 기능이 있으면 pull request를 제출하기 전에 [포럼](https://discuss.huggingface.co/) 또는 [이슈](https://github.com/huggingface/transformers/issues)에서 논의해 주세요.
버그 수정은 환영하지만 가독성을 희생하면서까지 더 많은 기능을 추가하는 pull request는 병합(merge)하지 않을 가능성이 높습니다.
이 가이드에서는 [PyTorch](https://github.com/huggingface/transformers/tree/main/examples/pytorch/summarization) 및 [TensorFlow](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/summarization)에서 요약 훈련하는
스크립트 예제를 실행하는 방법을 설명합니다.
특별한 설명이 없는 한 모든 예제는 두 프레임워크 모두에서 작동할 것으로 예상됩니다.
## 설정하기[[setup]]
최신 버전의 예제 스크립트를 성공적으로 실행하려면 새 가상 환경에서 **소스로부터 🤗 Transformers를 설치**해야 합니다:
```bash
git clone https://github.com/huggingface/transformers
cd transformers
pip install .
```
이전 버전의 예제 스크립트를 보려면 아래 토글을 클릭하세요:
<details>
<summary>이전 버전의 🤗 Transformers 예제</summary>
<ul>
<li><a href="https://github.com/huggingface/transformers/tree/v4.5.1/examples">v4.5.1</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v4.4.2/examples">v4.4.2</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v4.3.3/examples">v4.3.3</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v4.2.2/examples">v4.2.2</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v4.1.1/examples">v4.1.1</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v4.0.1/examples">v4.0.1</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v3.5.1/examples">v3.5.1</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v3.4.0/examples">v3.4.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v3.3.1/examples">v3.3.1</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v3.2.0/examples">v3.2.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v3.1.0/examples">v3.1.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v3.0.2/examples">v3.0.2</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.11.0/examples">v2.11.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.10.0/examples">v2.10.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.9.1/examples">v2.9.1</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.8.0/examples">v2.8.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.7.0/examples">v2.7.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.6.0/examples">v2.6.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.5.1/examples">v2.5.1</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.4.0/examples">v2.4.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.3.0/examples">v2.3.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.2.0/examples">v2.2.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.1.0/examples">v2.1.1</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v2.0.0/examples">v2.0.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v1.2.0/examples">v1.2.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v1.1.0/examples">v1.1.0</a></li>
<li><a href="https://github.com/huggingface/transformers/tree/v1.0.0/examples">v1.0.0</a></li>
</ul>
</details>
그리고 다음과 같이 복제(clone)해온 🤗 Transformers 버전을 특정 버전(예: v3.5.1)으로 전환하세요:
```bash
git checkout tags/v3.5.1
```
올바른 라이브러리 버전을 설정한 후 원하는 예제 폴더로 이동하여 예제별로 라이브러리에 대한 요구 사항(requirements)을 설치합니다:
```bash
pip install -r requirements.txt
```
## 스크립트 실행하기[[run-a-script]]
<frameworkcontent>
<pt>
예제 스크립트는 🤗 [Datasets](https://huggingface.co/docs/datasets/) 라이브러리에서 데이터 세트를 다운로드하고 전처리합니다.
그런 다음 스크립트는 요약 기능을 지원하는 아키텍처에서 [Trainer](https://huggingface.co/docs/transformers/main_classes/trainer)를 사용하여 데이터 세트를 미세 조정합니다.
다음 예는 [CNN/DailyMail](https://huggingface.co/datasets/cnn_dailymail) 데이터 세트에서 [T5-small](https://huggingface.co/google-t5/t5-small)을 미세 조정합니다.
T5 모델은 훈련 방식에 따라 추가 `source_prefix` 인수가 필요하며, 이 프롬프트는 요약 작업임을 T5에 알려줍니다.
```bash
python examples/pytorch/summarization/run_summarization.py \
--model_name_or_path google-t5/t5-small \
--do_train \
--do_eval \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--source_prefix "summarize: " \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size=4 \
--per_device_eval_batch_size=4 \
--overwrite_output_dir \
--predict_with_generate
```
</pt>
<tf>
예제 스크립트는 🤗 [Datasets](https://huggingface.co/docs/datasets/) 라이브러리에서 데이터 세트를 다운로드하고 전처리합니다.
그런 다음 스크립트는 요약 기능을 지원하는 아키텍처에서 Keras를 사용하여 데이터 세트를 미세 조정합니다.
다음 예는 [CNN/DailyMail](https://huggingface.co/datasets/cnn_dailymail) 데이터 세트에서 [T5-small](https://huggingface.co/google-t5/t5-small)을 미세 조정합니다.
T5 모델은 훈련 방식에 따라 추가 `source_prefix` 인수가 필요하며, 이 프롬프트는 요약 작업임을 T5에 알려줍니다.
```bash
python examples/tensorflow/summarization/run_summarization.py \
--model_name_or_path google-t5/t5-small \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size 8 \
--per_device_eval_batch_size 16 \
--num_train_epochs 3 \
--do_train \
--do_eval
```
</tf>
</frameworkcontent>
## 혼합 정밀도(mixed precision)로 분산 훈련하기[[distributed-training-and-mixed-precision]]
[Trainer](https://huggingface.co/docs/transformers/main_classes/trainer) 클래스는 분산 훈련과 혼합 정밀도(mixed precision)를 지원하므로 스크립트에서도 사용할 수 있습니다.
이 두 가지 기능을 모두 활성화하려면 다음 두 가지를 설정해야 합니다:
- `fp16` 인수를 추가해 혼합 정밀도(mixed precision)를 활성화합니다.
- `nproc_per_node` 인수를 추가해 사용할 GPU 개수를 설정합니다.
```bash
torchrun \
--nproc_per_node 8 pytorch/summarization/run_summarization.py \
--fp16 \
--model_name_or_path google-t5/t5-small \
--do_train \
--do_eval \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--source_prefix "summarize: " \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size=4 \
--per_device_eval_batch_size=4 \
--overwrite_output_dir \
--predict_with_generate
```
TensorFlow 스크립트는 분산 훈련을 위해 [`MirroredStrategy`](https://www.tensorflow.org/guide/distributed_training#mirroredstrategy)를 활용하며, 훈련 스크립트에 인수를 추가할 필요가 없습니다.
다중 GPU 환경이라면, TensorFlow 스크립트는 기본적으로 여러 개의 GPU를 사용합니다.
## TPU 위에서 스크립트 실행하기[[run-a-script-on-a-tpu]]
<frameworkcontent>
<pt>
Tensor Processing Units (TPUs)는 성능을 가속화하기 위해 특별히 설계되었습니다.
PyTorch는 [XLA](https://www.tensorflow.org/xla) 딥러닝 컴파일러와 함께 TPU를 지원합니다(자세한 내용은 [여기](https://github.com/pytorch/xla/blob/master/README.md) 참조).
TPU를 사용하려면 `xla_spawn.py` 스크립트를 실행하고 `num_cores` 인수를 사용하여 사용하려는 TPU 코어 수를 설정합니다.
```bash
python xla_spawn.py --num_cores 8 \
summarization/run_summarization.py \
--model_name_or_path google-t5/t5-small \
--do_train \
--do_eval \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--source_prefix "summarize: " \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size=4 \
--per_device_eval_batch_size=4 \
--overwrite_output_dir \
--predict_with_generate
```
</pt>
<tf>
Tensor Processing Units (TPUs)는 성능을 가속화하기 위해 특별히 설계되었습니다.
TensorFlow 스크립트는 TPU를 훈련에 사용하기 위해 [`TPUStrategy`](https://www.tensorflow.org/guide/distributed_training#tpustrategy)를 활용합니다.
TPU를 사용하려면 TPU 리소스의 이름을 `tpu` 인수에 전달합니다.
```bash
python run_summarization.py \
--tpu name_of_tpu_resource \
--model_name_or_path google-t5/t5-small \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size 8 \
--per_device_eval_batch_size 16 \
--num_train_epochs 3 \
--do_train \
--do_eval
```
</tf>
</frameworkcontent>
## 🤗 Accelerate로 스크립트 실행하기[[run-a-script-with-accelerate]]
🤗 [Accelerate](https://huggingface.co/docs/accelerate)는 PyTorch 훈련 과정에 대한 완전한 가시성을 유지하면서 여러 유형의 설정(CPU 전용, 다중 GPU, TPU)에서 모델을 훈련할 수 있는 통합 방법을 제공하는 PyTorch 전용 라이브러리입니다.
🤗 Accelerate가 설치되어 있는지 확인하세요:
> 참고: Accelerate는 빠르게 개발 중이므로 스크립트를 실행하려면 accelerate를 설치해야 합니다.
```bash
pip install git+https://github.com/huggingface/accelerate
```
`run_summarization.py` 스크립트 대신 `run_summarization_no_trainer.py` 스크립트를 사용해야 합니다.
🤗 Accelerate 클래스가 지원되는 스크립트는 폴더에 `task_no_trainer.py` 파일이 있습니다.
다음 명령을 실행하여 구성 파일을 생성하고 저장합니다:
```bash
accelerate config
```
설정을 테스트하여 올바르게 구성되었는지 확인합니다:
```bash
accelerate test
```
이제 훈련을 시작할 준비가 되었습니다:
```bash
accelerate launch run_summarization_no_trainer.py \
--model_name_or_path google-t5/t5-small \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--source_prefix "summarize: " \
--output_dir ~/tmp/tst-summarization
```
## 사용자 정의 데이터 세트 사용하기[[use-a-custom-dataset]]
요약 스크립트는 사용자 지정 데이터 세트가 CSV 또는 JSON 파일인 경우 지원합니다.
사용자 지정 데이터 세트를 사용하는 경우에는 몇 가지 추가 인수를 지정해야 합니다:
- `train_file`과 `validation_file`은 훈련 및 검증 파일의 경로를 지정합니다.
- `text_column`은 요약할 입력 텍스트입니다.
- `summary_column`은 출력할 대상 텍스트입니다.
사용자 지정 데이터 세트를 사용하는 요약 스크립트는 다음과 같습니다:
```bash
python examples/pytorch/summarization/run_summarization.py \
--model_name_or_path google-t5/t5-small \
--do_train \
--do_eval \
--train_file path_to_csv_or_jsonlines_file \
--validation_file path_to_csv_or_jsonlines_file \
--text_column text_column_name \
--summary_column summary_column_name \
--source_prefix "summarize: " \
--output_dir /tmp/tst-summarization \
--overwrite_output_dir \
--per_device_train_batch_size=4 \
--per_device_eval_batch_size=4 \
--predict_with_generate
```
## 스크립트 테스트하기[[test-a-script]]
전체 데이터 세트를 대상으로 훈련을 완료하는데 꽤 오랜 시간이 걸리기 때문에, 작은 데이터 세트에서 모든 것이 예상대로 실행되는지 확인하는 것이 좋습니다.
다음 인수를 사용하여 데이터 세트를 최대 샘플 수로 잘라냅니다:
- `max_train_samples`
- `max_eval_samples`
- `max_predict_samples`
```bash
python examples/pytorch/summarization/run_summarization.py \
--model_name_or_path google-t5/t5-small \
--max_train_samples 50 \
--max_eval_samples 50 \
--max_predict_samples 50 \
--do_train \
--do_eval \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--source_prefix "summarize: " \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size=4 \
--per_device_eval_batch_size=4 \
--overwrite_output_dir \
--predict_with_generate
```
모든 예제 스크립트가 `max_predict_samples` 인수를 지원하지는 않습니다.
스크립트가 이 인수를 지원하는지 확실하지 않은 경우 `-h` 인수를 추가하여 확인하세요:
```bash
examples/pytorch/summarization/run_summarization.py -h
```
## 체크포인트(checkpoint)에서 훈련 이어서 하기[[resume-training-from-checkpoint]]
또 다른 유용한 옵션은 이전 체크포인트에서 훈련을 재개하는 것입니다.
이렇게 하면 훈련이 중단되더라도 처음부터 다시 시작하지 않고 중단한 부분부터 다시 시작할 수 있습니다.
체크포인트에서 훈련을 재개하는 방법에는 두 가지가 있습니다.
첫 번째는 `output_dir previous_output_dir` 인수를 사용하여 `output_dir`에 저장된 최신 체크포인트부터 훈련을 재개하는 방법입니다.
이 경우 `overwrite_output_dir`을 제거해야 합니다:
```bash
python examples/pytorch/summarization/run_summarization.py
--model_name_or_path google-t5/t5-small \
--do_train \
--do_eval \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--source_prefix "summarize: " \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size=4 \
--per_device_eval_batch_size=4 \
--output_dir previous_output_dir \
--predict_with_generate
```
두 번째는 `resume_from_checkpoint path_to_specific_checkpoint` 인수를 사용하여 특정 체크포인트 폴더에서 훈련을 재개하는 방법입니다.
```bash
python examples/pytorch/summarization/run_summarization.py
--model_name_or_path google-t5/t5-small \
--do_train \
--do_eval \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--source_prefix "summarize: " \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size=4 \
--per_device_eval_batch_size=4 \
--overwrite_output_dir \
--resume_from_checkpoint path_to_specific_checkpoint \
--predict_with_generate
```
## 모델 공유하기[[share-your-model]]
모든 스크립트는 최종 모델을 [Model Hub](https://huggingface.co/models)에 업로드할 수 있습니다.
시작하기 전에 Hugging Face에 로그인했는지 확인하세요:
```bash
hf auth login
```
그런 다음 스크립트에 `push_to_hub` 인수를 추가합니다.
이 인수는 Hugging Face 사용자 이름과 `output_dir`에 지정된 폴더 이름으로 저장소를 생성합니다.
저장소에 특정 이름을 지정하려면 `push_to_hub_model_id` 인수를 사용하여 추가합니다.
저장소는 네임스페이스 아래에 자동으로 나열됩니다.
다음 예는 특정 저장소 이름으로 모델을 업로드하는 방법입니다:
```bash
python examples/pytorch/summarization/run_summarization.py
--model_name_or_path google-t5/t5-small \
--do_train \
--do_eval \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--source_prefix "summarize: " \
--push_to_hub \
--push_to_hub_model_id finetuned-t5-cnn_dailymail \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size=4 \
--per_device_eval_batch_size=4 \
--overwrite_output_dir \
--predict_with_generate
``` | transformers/docs/source/ko/run_scripts.md/0 | {
"file_path": "transformers/docs/source/ko/run_scripts.md",
"repo_id": "transformers",
"token_count": 10830
} | 399 |
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# 단일 영상 기반 깊이 추정[[depth-estimation-pipeline]]
단일 영상 기반 깊이 추정은 한 장면의 단일 이미지에서 장면의 깊이 정보를 예측하는 컴퓨터 비전 작업입니다.
즉, 단일 카메라 시점의 장면에 있는 물체의 거리를 예측하는 과정입니다.
단일 영상 기반 깊이 추정은 3D 재구성, 증강 현실, 자율 주행, 로봇 공학 등 다양한 분야에서 응용됩니다.
조명 조건, 가려짐, 텍스처와 같은 요소의 영향을 받을 수 있는 장면 내 물체와 해당 깊이 정보 간의 복잡한 관계를 모델이 이해해야 하므로 까다로운 작업입니다.
<Tip>
이 작업과 호환되는 모든 아키텍처와 체크포인트를 보려면 [작업 페이지](https://huggingface.co/tasks/depth-estimation)를 확인하는 것이 좋습니다.
</Tip>
이번 가이드에서 배울 내용은 다음과 같습니다:
* 깊이 추정 파이프라인 만들기
* 직접 깊이 추정 추론하기
시작하기 전에, 필요한 모든 라이브러리가 설치되어 있는지 확인하세요:
```bash
pip install -q transformers
```
## 깊이 추정 파이프라인[[depth-estimation-inference-by-hand]]
깊이 추정을 추론하는 가장 간단한 방법은 해당 기능을 제공하는 [`pipeline`]을 사용하는 것입니다.
[Hugging Face Hub 체크포인트](https://huggingface.co/models?pipeline_tag=depth-estimation&sort=downloads)에서 파이프라인을 초기화합니다:
```py
>>> from transformers import pipeline
>>> checkpoint = "vinvino02/glpn-nyu"
>>> depth_estimator = pipeline("depth-estimation", model=checkpoint)
```
다음으로, 분석할 이미지를 한 장 선택하세요:
```py
>>> from PIL import Image
>>> import requests
>>> url = "https://unsplash.com/photos/HwBAsSbPBDU/download?ixid=MnwxMjA3fDB8MXxzZWFyY2h8MzR8fGNhciUyMGluJTIwdGhlJTIwc3RyZWV0fGVufDB8MHx8fDE2Nzg5MDEwODg&force=true&w=640"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> image
```
<div class="flex justify-center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/depth-estimation-example.jpg" alt="Photo of a busy street"/>
</div>
이미지를 파이프라인으로 전달합니다.
```py
>>> predictions = depth_estimator(image)
```
파이프라인은 두 개의 항목을 가지는 딕셔너리를 반환합니다.
첫 번째는 `predicted_depth`로 각 픽셀의 깊이를 미터로 표현한 값을 가지는 텐서입니다.
두 번째는 `depth`로 깊이 추정 결과를 시각화하는 PIL 이미지입니다.
이제 시각화한 결과를 살펴보겠습니다:
```py
>>> predictions["depth"]
```
<div class="flex justify-center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/depth-visualization.png" alt="Depth estimation visualization"/>
</div>
## 직접 깊이 추정 추론하기[[depth-estimation-inference-by-hand]]
이제 깊이 추정 파이프라인 사용법을 살펴보았으니 동일한 결과를 복제하는 방법을 살펴보겠습니다.
[Hugging Face Hub 체크포인트](https://huggingface.co/models?pipeline_tag=depth-estimation&sort=downloads)에서 모델과 관련 프로세서를 가져오는 것부터 시작합니다.
여기서 이전에 사용한 체크포인트와 동일한 것을 사용합니다:
```py
>>> from transformers import AutoImageProcessor, AutoModelForDepthEstimation
>>> checkpoint = "vinvino02/glpn-nyu"
>>> image_processor = AutoImageProcessor.from_pretrained(checkpoint)
>>> model = AutoModelForDepthEstimation.from_pretrained(checkpoint)
```
필요한 이미지 변환을 처리하는 `image_processor`를 사용하여 모델에 대한 이미지 입력을 준비합니다.
`image_processor`는 크기 조정 및 정규화 등 필요한 이미지 변환을 처리합니다:
```py
>>> pixel_values = image_processor(image, return_tensors="pt").pixel_values
```
준비한 입력을 모델로 전달합니다:
```py
>>> import torch
>>> with torch.no_grad():
... outputs = model(pixel_values)
... predicted_depth = outputs.predicted_depth
```
결과를 시각화합니다:
```py
>>> import numpy as np
>>> # 원본 사이즈로 복원
>>> prediction = torch.nn.functional.interpolate(
... predicted_depth.unsqueeze(1),
... size=image.size[::-1],
... mode="bicubic",
... align_corners=False,
... ).squeeze()
>>> output = prediction.numpy()
>>> formatted = (output * 255 / np.max(output)).astype("uint8")
>>> depth = Image.fromarray(formatted)
>>> depth
```
<div class="flex justify-center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/depth-visualization.png" alt="Depth estimation visualization"/>
</div>
| transformers/docs/source/ko/tasks/monocular_depth_estimation.md/0 | {
"file_path": "transformers/docs/source/ko/tasks/monocular_depth_estimation.md",
"repo_id": "transformers",
"token_count": 3117
} | 400 |
### `tiny-agents` CLI 및 MCP 도구[[tiny-agents-cli-and-mcp-tools]]
MCP 도구의 사용을 보여주기 위해 [`tiny-agents`](https://huggingface.co/blog/python-tiny-agents) CLI와 `transformers serve` 서버를 연동하는 방법을 살펴보겠습니다.
> [!TIP]
> 이 예시처럼 많은 Hugging Face Spaces를 MCP 서버로 활용할 수 있습니다. 호환 가능한 모든 Spaces는 [여기](https://huggingface.co/spaces?filter=mcp-server)에서 찾을 수 있습니다.
MCP 도구를 사용하려면 먼저 모델에 사용 가능한 도구를 알려야 합니다. 예를 들어, [이미지 생성 MCP 서버](https://evalstate-flux1-schnell.hf.space/)를 참조하는 `tiny-agents` 설정 파일을 살펴보겠습니다.
```json
{
"model": "Menlo/Jan-nano",
"endpointUrl": "http://localhost:8000",
"servers": [
{
"type": "sse",
"url": "https://evalstate-flux1-schnell.hf.space/gradio_api/mcp/sse"
}
]
}
```
그런 다음 아래 명령어로 `tiny-agents` 채팅 인터페이스를 실행할 수 있습니다.
```bash
tiny-agents run path/to/your/config.json
```
백그라운드에서 `transformers serve`가 실행 중이라면, 이제 로컬 모델에서 MCP 도구를 사용할 수 있습니다. 다음은 `tiny-agents`와의 채팅 세션 예시입니다.
```bash
Agent loaded with 1 tools:
• flux1_schnell_infer
» Generate an image of a cat on the moon
<Tool req_0_tool_call>flux1_schnell_infer {"prompt": "a cat on the moon", "seed": 42, "randomize_seed": true, "width": 1024, "height": 1024, "num_inference_steps": 4}
Tool req_0_tool_call
[Binary Content: Image image/webp, 57732 bytes]
The task is complete and the content accessible to the User
Image URL: https://evalstate-flux1-schnell.hf.space/gradio_api/file=/tmp/gradio/3dbddc0e53b5a865ed56a4e3dbdd30f3f61cf3b8aabf1b456f43e5241bd968b8/image.webp
380576952
Flux 1 Schnell 이미지 생성기를 사용하여 달 위의 고양이 이미지를 생성했습니다. 이미지는 1024x1024 픽셀이며 4번의 추론 단계를 거쳐 생성되었습니다. 변경 사항이 필요하거나 추가 도움이 필요하시면 알려주세요!
```
| transformers/docs/source/ko/tiny_agents.md/0 | {
"file_path": "transformers/docs/source/ko/tiny_agents.md",
"repo_id": "transformers",
"token_count": 1311
} | 401 |
<!---
Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Guia de Instalação
Neste guia poderá encontrar informações para a instalação do 🤗 Transformers para qualquer biblioteca de
Machine Learning com a qual esteja a trabalhar. Além disso, poderá encontrar informações sobre como gerar cachês e
configurar o 🤗 Transformers para execução em modo offline (opcional).
🤗 Transformers foi testado com Python 3.6+, PyTorch 1.1.0+, TensorFlow 2.0+, e Flax. Para instalar a biblioteca de
deep learning com que deseja trabalhar, siga as instruções correspondentes listadas a seguir:
* [PyTorch](https://pytorch.org/get-started/locally/)
* [TensorFlow 2.0](https://www.tensorflow.org/install/pip)
* [Flax](https://flax.readthedocs.io/en/latest/)
## Instalação pelo Pip
É sugerido instalar o 🤗 Transformers num [ambiente virtual](https://docs.python.org/3/library/venv.html). Se precisar
de mais informações sobre ambientes virtuais em Python, consulte este [guia](https://packaging.python.org/guides/installing-using-pip-and-virtual-environments/).
Um ambiente virtual facilitará a manipulação e organização de projetos e evita problemas de compatibilidade entre dependências.
Comece criando um ambiente virtual no diretório do seu projeto:
```bash
python -m venv .env
```
E para ativar o ambiente virtual:
```bash
source .env/bin/activate
```
Agora É possível instalar o 🤗 Transformers com o comando a seguir:
```bash
pip install transformers
```
Somente para a CPU, é possível instalar o 🤗 Transformers e a biblioteca de deep learning respectiva apenas numa linha.
Por exemplo, para instalar o 🤗 Transformers e o PyTorch, digite:
```bash
pip install transformers[torch]
```
🤗 Transformers e TensorFlow 2.0:
```bash
pip install transformers[tf-cpu]
```
🤗 Transformers e Flax:
```bash
pip install transformers[flax]
```
Por último, verifique se o 🤗 Transformers foi instalado com sucesso usando o seguinte comando para baixar um modelo pré-treinado:
```bash
python -c "from transformers import pipeline; print(pipeline('sentiment-analysis')('we love you'))"
```
Em seguida, imprima um rótulo e sua pontuação:
```bash
[{'label': 'POSITIVE', 'score': 0.9998704791069031}]
```
## Instalação usando a fonte
Para instalar o 🤗 Transformers a partir da fonte use o seguinte comando:
```bash
pip install git+https://github.com/huggingface/transformers
```
O comando acima instalará a versão `master` mais atual em vez da última versão estável. A versão `master` é útil para
utilizar os últimos updates contidos em 🤗 Transformers. Por exemplo, um erro recente pode ter sido corrigido somente
após a última versão estável, antes que houvesse um novo lançamento. No entanto, há a possibilidade que a versão `master` não esteja estável.
A equipa trata de mantér a versão `master` operacional e a maioria dos erros são resolvidos em poucas horas ou dias.
Se encontrar quaisquer problemas, por favor abra um [Issue](https://github.com/huggingface/transformers/issues) para que o
mesmo possa ser corrigido o mais rápido possível.
Verifique que o 🤗 Transformers está instalado corretamente usando o seguinte comando:
```bash
python -c "from transformers import pipeline; print(pipeline('sentiment-analysis')('I love you'))"
```
## Instalação editável
Uma instalação editável será necessária caso desejas um dos seguintes:
* Usar a versão `master` do código fonte.
* Contribuir ao 🤗 Transformers e precisa testar mudanças ao código.
Para tal, clone o repositório e instale o 🤗 Transformers com os seguintes comandos:
```bash
git clone https://github.com/huggingface/transformers.git
cd transformers
pip install -e .
```
Estes comandos vão ligar o diretório para o qual foi clonado o repositório ao caminho de bibliotecas do Python.
O Python agora buscará dentro dos arquivos que foram clonados além dos caminhos normais da biblioteca.
Por exemplo, se os pacotes do Python se encontram instalados no caminho `~/anaconda3/envs/main/lib/python3.7/site-packages/`,
o Python também buscará módulos no diretório onde clonamos o repositório `~/transformers/`.
<Tip warning={true}>
É necessário manter o diretório `transformers` se desejas continuar usando a biblioteca.
</Tip>
Assim, É possível atualizar sua cópia local para com a última versão do 🤗 Transformers com o seguinte comando:
```bash
cd ~/transformers/
git pull
```
O ambiente de Python que foi criado para a instalação do 🤗 Transformers encontrará a versão `master` em execuções seguintes.
## Instalação usando o Conda
É possível instalar o 🤗 Transformers a partir do canal conda `conda-forge` com o seguinte comando:
```bash
conda install conda-forge::transformers
```
## Configuração do Cachê
Os modelos pré-treinados são baixados e armazenados no cachê local, encontrado em `~/.cache/huggingface/transformers/`.
Este é o diretório padrão determinado pela variável `TRANSFORMERS_CACHE` dentro do shell.
No Windows, este diretório pré-definido é dado por `C:\Users\username\.cache\huggingface\transformers`.
É possível mudar as variáveis dentro do shell em ordem de prioridade para especificar um diretório de cachê diferente:
1. Variável de ambiente do shell (por padrão): `TRANSFORMERS_CACHE`.
2. Variável de ambiente do shell:`HF_HOME` + `transformers/`.
3. Variável de ambiente do shell: `XDG_CACHE_HOME` + `/huggingface/transformers`.
<Tip>
O 🤗 Transformers usará as variáveis de ambiente do shell `PYTORCH_TRANSFORMERS_CACHE` ou `PYTORCH_PRETRAINED_BERT_CACHE`
se estiver vindo de uma versão anterior da biblioteca que tenha configurado essas variáveis de ambiente, a menos que
você especifique a variável de ambiente do shell `TRANSFORMERS_CACHE`.
</Tip>
## Modo Offline
O 🤗 Transformers também pode ser executado num ambiente de firewall ou fora da rede (offline) usando arquivos locais.
Para tal, configure a variável de ambiente de modo que `HF_HUB_OFFLINE=1`.
<Tip>
Você pode adicionar o [🤗 Datasets](https://huggingface.co/docs/datasets/) ao pipeline de treinamento offline declarando
a variável de ambiente `HF_DATASETS_OFFLINE=1`.
</Tip>
Segue um exemplo de execução do programa numa rede padrão com firewall para instâncias externas, usando o seguinte comando:
```bash
python examples/pytorch/translation/run_translation.py --model_name_or_path google-t5/t5-small --dataset_name wmt16 --dataset_config ro-en ...
```
Execute esse mesmo programa numa instância offline com o seguinte comando:
```bash
HF_DATASETS_OFFLINE=1 HF_HUB_OFFLINE=1 \
python examples/pytorch/translation/run_translation.py --model_name_or_path google-t5/t5-small --dataset_name wmt16 --dataset_config ro-en ...
```
O script agora deve ser executado sem travar ou expirar, pois procurará apenas por arquivos locais.
### Obtendo modelos e tokenizers para uso offline
Outra opção para usar o 🤗 Transformers offline é baixar os arquivos antes e depois apontar para o caminho local onde estão localizados. Existem três maneiras de fazer isso:
* Baixe um arquivo por meio da interface de usuário do [Model Hub](https://huggingface.co/models) clicando no ícone ↓.
* Use o pipeline do [`PreTrainedModel.from_pretrained`] e [`PreTrainedModel.save_pretrained`]:
1. Baixa os arquivos previamente com [`PreTrainedModel.from_pretrained`]:
```py
>>> from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
>>> tokenizer = AutoTokenizer.from_pretrained("bigscience/T0_3B")
>>> model = AutoModelForSeq2SeqLM.from_pretrained("bigscience/T0_3B")
```
2. Salve os arquivos em um diretório específico com [`PreTrainedModel.save_pretrained`]:
```py
>>> tokenizer.save_pretrained("./your/path/bigscience_t0")
>>> model.save_pretrained("./your/path/bigscience_t0")
```
3. Quando estiver offline, acesse os arquivos com [`PreTrainedModel.from_pretrained`] do diretório especificado:
```py
>>> tokenizer = AutoTokenizer.from_pretrained("./your/path/bigscience_t0")
>>> model = AutoModel.from_pretrained("./your/path/bigscience_t0")
```
* Baixando arquivos programaticamente com a biblioteca [huggingface_hub](https://github.com/huggingface/huggingface_hub/tree/main/src/huggingface_hub):
1. Instale a biblioteca [huggingface_hub](https://github.com/huggingface/huggingface_hub/tree/main/src/huggingface_hub) em seu ambiente virtual:
```bash
python -m pip install huggingface_hub
```
2. Utiliza a função [`hf_hub_download`](https://huggingface.co/docs/hub/adding-a-library#download-files-from-the-hub) para baixar um arquivo para um caminho específico. Por exemplo, o comando a seguir baixará o arquivo `config.json` para o modelo [T0](https://huggingface.co/bigscience/T0_3B) no caminho desejado:
```py
>>> from huggingface_hub import hf_hub_download
>>> hf_hub_download(repo_id="bigscience/T0_3B", filename="config.json", cache_dir="./your/path/bigscience_t0")
```
Depois que o arquivo for baixado e armazenado no cachê local, especifique seu caminho local para carregá-lo e usá-lo:
```py
>>> from transformers import AutoConfig
>>> config = AutoConfig.from_pretrained("./your/path/bigscience_t0/config.json")
```
<Tip>
Para obter mais detalhes sobre como baixar arquivos armazenados no Hub, consulte a seção [How to download files from the Hub](https://huggingface.co/docs/hub/how-to-downstream).
</Tip>
| transformers/docs/source/pt/installation.md/0 | {
"file_path": "transformers/docs/source/pt/installation.md",
"repo_id": "transformers",
"token_count": 3770
} | 402 |
<!--
Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# 如何创建自定义流水线？
在本指南中，我们将演示如何创建一个自定义流水线并分享到 [Hub](https://hf.co/models)，或将其添加到 🤗 Transformers 库中。
首先，你需要决定流水线将能够接受的原始条目。它可以是字符串、原始字节、字典或任何看起来最可能是期望的输入。
尽量保持输入为纯 Python 语言，因为这样可以更容易地实现兼容性（甚至通过 JSON 在其他语言之间）。
这些将是流水线 (`preprocess`) 的 `inputs`。
然后定义 `outputs`。与 `inputs` 相同的策略。越简单越好。这些将是 `postprocess` 方法的输出。
首先继承基类 `Pipeline`，其中包含实现 `preprocess`、`_forward`、`postprocess` 和 `_sanitize_parameters` 所需的 4 个方法。
```python
from transformers import Pipeline
class MyPipeline(Pipeline):
def _sanitize_parameters(self, **kwargs):
preprocess_kwargs = {}
if "maybe_arg" in kwargs:
preprocess_kwargs["maybe_arg"] = kwargs["maybe_arg"]
return preprocess_kwargs, {}, {}
def preprocess(self, inputs, maybe_arg=2):
model_input = Tensor(inputs["input_ids"])
return {"model_input": model_input}
def _forward(self, model_inputs):
# model_inputs == {"model_input": model_input}
outputs = self.model(**model_inputs)
# Maybe {"logits": Tensor(...)}
return outputs
def postprocess(self, model_outputs):
best_class = model_outputs["logits"].softmax(-1)
return best_class
```
这种分解的结构旨在为 CPU/GPU 提供相对无缝的支持，同时支持在不同线程上对 CPU 进行预处理/后处理。
`preprocess` 将接受最初定义的输入，并将其转换为可供模型输入的内容。它可能包含更多信息，通常是一个 `Dict`。
`_forward` 是实现细节，不应直接调用。`forward` 是首选的调用方法，因为它包含保障措施，以确保一切都在预期的设备上运作。
如果任何内容与实际模型相关，它应该属于 `_forward` 方法，其他内容应该在 preprocess/postprocess 中。
`postprocess` 方法将接受 `_forward` 的输出，并将其转换为之前确定的最终输出。
`_sanitize_parameters` 存在是为了允许用户在任何时候传递任何参数，无论是在初始化时 `pipeline(...., maybe_arg=4)`
还是在调用时 `pipe = pipeline(...); output = pipe(...., maybe_arg=4)`。
`_sanitize_parameters` 的返回值是将直接传递给 `preprocess`、`_forward` 和 `postprocess` 的 3 个关键字参数字典。
如果调用方没有使用任何额外参数调用，则不要填写任何内容。这样可以保留函数定义中的默认参数，这总是更"自然"的。
在分类任务中，一个经典的例子是在后处理中使用 `top_k` 参数。
```python
>>> pipe = pipeline("my-new-task")
>>> pipe("This is a test")
[{"label": "1-star", "score": 0.8}, {"label": "2-star", "score": 0.1}, {"label": "3-star", "score": 0.05}
{"label": "4-star", "score": 0.025}, {"label": "5-star", "score": 0.025}]
>>> pipe("This is a test", top_k=2)
[{"label": "1-star", "score": 0.8}, {"label": "2-star", "score": 0.1}]
```
为了实现这一点，我们将更新我们的 `postprocess` 方法，将默认参数设置为 `5`，
并编辑 `_sanitize_parameters` 方法，以允许这个新参数。
```python
def postprocess(self, model_outputs, top_k=5):
best_class = model_outputs["logits"].softmax(-1)
# Add logic to handle top_k
return best_class
def _sanitize_parameters(self, **kwargs):
preprocess_kwargs = {}
if "maybe_arg" in kwargs:
preprocess_kwargs["maybe_arg"] = kwargs["maybe_arg"]
postprocess_kwargs = {}
if "top_k" in kwargs:
postprocess_kwargs["top_k"] = kwargs["top_k"]
return preprocess_kwargs, {}, postprocess_kwargs
```
尽量保持简单输入/输出，最好是可 JSON 序列化的，因为这样可以使流水线的使用非常简单，而不需要用户了解新的对象类型。
通常也相对常见地支持许多不同类型的参数以便使用（例如音频文件，可以是文件名、URL 或纯字节）。
## 将其添加到支持的任务列表中
要将你的 `new-task` 注册到支持的任务列表中，你需要将其添加到 `PIPELINE_REGISTRY` 中：
```python
from transformers.pipelines import PIPELINE_REGISTRY
PIPELINE_REGISTRY.register_pipeline(
"new-task",
pipeline_class=MyPipeline,
pt_model=AutoModelForSequenceClassification,
)
```
如果需要，你可以指定一个默认模型，此时它应该带有一个特定的修订版本（可以是分支名称或提交哈希，这里我们使用了 `"abcdef"`），以及类型：
```python
PIPELINE_REGISTRY.register_pipeline(
"new-task",
pipeline_class=MyPipeline,
pt_model=AutoModelForSequenceClassification,
default={"pt": ("user/awesome_model", "abcdef")},
type="text", # current support type: text, audio, image, multimodal
)
```
## 在 Hub 上分享你的流水线
要在 Hub 上分享你的自定义流水线，你只需要将 `Pipeline` 子类的自定义代码保存在一个 Python 文件中。
例如，假设我们想使用一个自定义流水线进行句对分类，如下所示：
```py
import numpy as np
from transformers import Pipeline
def softmax(outputs):
maxes = np.max(outputs, axis=-1, keepdims=True)
shifted_exp = np.exp(outputs - maxes)
return shifted_exp / shifted_exp.sum(axis=-1, keepdims=True)
class PairClassificationPipeline(Pipeline):
def _sanitize_parameters(self, **kwargs):
preprocess_kwargs = {}
if "second_text" in kwargs:
preprocess_kwargs["second_text"] = kwargs["second_text"]
return preprocess_kwargs, {}, {}
def preprocess(self, text, second_text=None):
return self.tokenizer(text, text_pair=second_text, return_tensors=self.framework)
def _forward(self, model_inputs):
return self.model(**model_inputs)
def postprocess(self, model_outputs):
logits = model_outputs.logits[0].numpy()
probabilities = softmax(logits)
best_class = np.argmax(probabilities)
label = self.model.config.id2label[best_class]
score = probabilities[best_class].item()
logits = logits.tolist()
return {"label": label, "score": score, "logits": logits}
```
这个实现与框架无关，适用于 PyTorch 和 TensorFlow 模型。如果我们将其保存在一个名为
`pair_classification.py` 的文件中，然后我们可以像这样导入并注册它：
```py
from pair_classification import PairClassificationPipeline
from transformers.pipelines import PIPELINE_REGISTRY
from transformers import AutoModelForSequenceClassification, TFAutoModelForSequenceClassification
PIPELINE_REGISTRY.register_pipeline(
"pair-classification",
pipeline_class=PairClassificationPipeline,
pt_model=AutoModelForSequenceClassification,
tf_model=TFAutoModelForSequenceClassification,
)
```
完成这些步骤后，我们可以将其与预训练模型一起使用。例如，`sgugger/finetuned-bert-mrpc`
已经在 MRPC 数据集上进行了微调，用于将句子对分类为是释义或不是释义。
```py
from transformers import pipeline
classifier = pipeline("pair-classification", model="sgugger/finetuned-bert-mrpc")
```
然后，我们可以通过在 `Repository` 中使用 `save_pretrained` 方法将其分享到 Hub 上：
```py
from huggingface_hub import Repository
repo = Repository("test-dynamic-pipeline", clone_from="{your_username}/test-dynamic-pipeline")
classifier.save_pretrained("test-dynamic-pipeline")
repo.push_to_hub()
```
这将会复制包含你定义的 `PairClassificationPipeline` 的文件到文件夹 `"test-dynamic-pipeline"` 中，
同时保存流水线的模型和分词器，然后将所有内容推送到仓库 `{your_username}/test-dynamic-pipeline` 中。
之后，只要提供选项 `trust_remote_code=True`，任何人都可以使用它：
```py
from transformers import pipeline
classifier = pipeline(model="{your_username}/test-dynamic-pipeline", trust_remote_code=True)
```
## 将流水线添加到 🤗 Transformers
如果你想将你的流水线贡献给 🤗 Transformers，你需要在 `pipelines` 子模块中添加一个新模块，
其中包含你的流水线的代码，然后将其添加到 `pipelines/__init__.py` 中定义的任务列表中。
然后，你需要添加测试。创建一个新文件 `tests/test_pipelines_MY_PIPELINE.py`，其中包含其他测试的示例。
`run_pipeline_test` 函数将非常通用，并在每种可能的架构上运行小型随机模型，如 `model_mapping` 和 `tf_model_mapping` 所定义。
这对于测试未来的兼容性非常重要，这意味着如果有人为 `XXXForQuestionAnswering` 添加了一个新模型，
流水线测试将尝试在其上运行。由于模型是随机的，所以不可能检查实际值，这就是为什么有一个帮助函数 `ANY`，它只是尝试匹配流水线的输出类型。
你还 **需要** 实现 2（最好是 4）个测试。
- `test_small_model_pt`：为这个流水线定义一个小型模型（结果是否合理并不重要），并测试流水线的输出。
结果应该与 `test_small_model_tf` 的结果相同。
- `test_small_model_tf`：为这个流水线定义一个小型模型（结果是否合理并不重要），并测试流水线的输出。
结果应该与 `test_small_model_pt` 的结果相同。
- `test_large_model_pt`（可选）：在一个真实的流水线上测试流水线，结果应该是有意义的。
这些测试速度较慢，应该被如此标记。这里的目标是展示流水线，并确保在未来的发布中没有漂移。
- `test_large_model_tf`（可选）：在一个真实的流水线上测试流水线，结果应该是有意义的。
这些测试速度较慢，应该被如此标记。这里的目标是展示流水线，并确保在未来的发布中没有漂移。
| transformers/docs/source/zh/add_new_pipeline.md/0 | {
"file_path": "transformers/docs/source/zh/add_new_pipeline.md",
"repo_id": "transformers",
"token_count": 5640
} | 403 |
<!---
Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# 安装
为你正在使用的深度学习框架安装 🤗 Transformers、设置缓存，并选择性配置 🤗 Transformers 以离线运行。
🤗 Transformers 已在 Python 3.6+、PyTorch 1.1.0+、TensorFlow 2.0+ 以及 Flax 上进行测试。针对你使用的深度学习框架，请参照以下安装说明进行安装：
* [PyTorch](https://pytorch.org/get-started/locally/) 安装说明。
* [TensorFlow 2.0](https://www.tensorflow.org/install/pip) 安装说明。
* [Flax](https://flax.readthedocs.io/en/latest/) 安装说明。
## 使用 pip 安装
你应该使用 [虚拟环境](https://docs.python.org/3/library/venv.html) 安装 🤗 Transformers。如果你不熟悉 Python 虚拟环境，请查看此 [教程](https://packaging.python.org/guides/installing-using-pip-and-virtual-environments/)。使用虚拟环境，你可以轻松管理不同项目，避免不同依赖项之间的兼容性问题。
首先，在项目目录中创建虚拟环境：
```bash
python -m venv .env
```
在 Linux 和 MacOs 系统中激活虚拟环境：
```bash
source .env/bin/activate
```
在 Windows 系统中激活虚拟环境：
```bash
.env/Scripts/activate
```
现在你可以使用以下命令安装 🤗 Transformers：
```bash
pip install transformers
```
若仅需 CPU 支持，可以使用单行命令方便地安装 🤗 Transformers 和深度学习库。例如，使用以下命令安装 🤗 Transformers 和 PyTorch：
```bash
pip install 'transformers[torch]'
```
🤗 Transformers 和 TensorFlow 2.0：
```bash
pip install 'transformers[tf-cpu]'
```
<Tip warning={true}>
M1 / ARM用户
在安装 TensorFlow 2.0 前，你需要安装以下库：
```bash
brew install cmake
brew install pkg-config
```
</Tip>
🤗 Transformers 和 Flax:
```bash
pip install 'transformers[flax]'
```
最后，运行以下命令以检查 🤗 Transformers 是否已被正确安装。该命令将下载一个预训练模型：
```bash
python -c "from transformers import pipeline; print(pipeline('sentiment-analysis')('we love you'))"
```
然后打印标签以及分数：
```bash
[{'label': 'POSITIVE', 'score': 0.9998704791069031}]
```
## 源码安装
使用以下命令从源码安装 🤗 Transformers：
```bash
pip install git+https://github.com/huggingface/transformers
```
此命令下载的是最新的前沿 `main` 版本而不是最新的 `stable` 版本。`main` 版本适用于跟最新开发保持一致。例如，上次正式版发布带来的 bug 被修复了，但新版本尚未被推出。但是，这也说明 `main` 版本并不一定总是稳定的。我们努力保持 `main` 版本的可操作性，大多数问题通常在几个小时或一天以内就能被解决。如果你遇到问题，请提个 [Issue](https://github.com/huggingface/transformers/issues) 以便我们能更快修复。
运行以下命令以检查 🤗 Transformers 是否已被正确安装：
```bash
python -c "from transformers import pipeline; print(pipeline('sentiment-analysis')('I love you'))"
```
## 可编辑安装
如果你有下列需求，需要进行可编辑安装：
* 使用源码的 `main` 版本。
* 为 🤗 Transformers 贡献代码，需要测试代码中的更改。
使用以下命令克隆仓库并安装 🤗 Transformers：
```bash
git clone https://github.com/huggingface/transformers.git
cd transformers
pip install -e .
```
这些命令将会链接你克隆的仓库以及你的 Python 库路径。现在，Python 不仅会在正常的库路径中搜索库，也会在你克隆到的文件夹中进行查找。例如，如果你的 Python 包通常本应安装在 `~/anaconda3/envs/main/lib/python3.7/site-packages/` 目录中，在这种情况下 Python 也会搜索你克隆到的文件夹：`~/transformers/`。
<Tip warning={true}>
如果你想继续使用这个库，必须保留 `transformers` 文件夹。
</Tip>
现在，你可以使用以下命令，将你克隆的 🤗 Transformers 库轻松更新至最新版本：
```bash
cd ~/transformers/
git pull
```
你的 Python 环境将在下次运行时找到 `main` 版本的 🤗 Transformers。
## 使用 conda 安装
从 conda 的 `conda-forge` 频道安装：
```bash
conda install conda-forge::transformers
```
## 缓存设置
预训练模型会被下载并本地缓存到 `~/.cache/huggingface/hub`。这是由环境变量 `TRANSFORMERS_CACHE` 指定的默认目录。在 Windows 上，默认目录为 `C:\Users\username\.cache\huggingface\hub`。你可以按照不同优先级改变下述环境变量，以指定不同的缓存目录。
1. 环境变量（默认）: `HF_HUB_CACHE` 或 `TRANSFORMERS_CACHE`。
2. 环境变量 `HF_HOME`。
3. 环境变量 `XDG_CACHE_HOME` + `/huggingface`。
<Tip>
除非你明确指定了环境变量 `TRANSFORMERS_CACHE`，🤗 Transformers 将可能会使用较早版本设置的环境变量 `PYTORCH_TRANSFORMERS_CACHE` 或 `PYTORCH_PRETRAINED_BERT_CACHE`。
</Tip>
## 离线模式
🤗 Transformers 可以仅使用本地文件在防火墙或离线环境中运行。设置环境变量 `HF_HUB_OFFLINE=1` 以启用该行为。
<Tip>
通过设置环境变量 `HF_DATASETS_OFFLINE=1` 将 [🤗 Datasets](https://huggingface.co/docs/datasets/) 添加至你的离线训练工作流程中。
</Tip>
例如，你通常会使用以下命令对外部实例进行防火墙保护的的普通网络上运行程序：
```bash
python examples/pytorch/translation/run_translation.py --model_name_or_path google-t5/t5-small --dataset_name wmt16 --dataset_config ro-en ...
```
在离线环境中运行相同的程序：
```bash
HF_DATASETS_OFFLINE=1 HF_HUB_OFFLINE=1 \
python examples/pytorch/translation/run_translation.py --model_name_or_path google-t5/t5-small --dataset_name wmt16 --dataset_config ro-en ...
```
现在脚本可以应该正常运行，而无需挂起或等待超时，因为它知道只应查找本地文件。
### 获取离线时使用的模型和分词器
另一种离线时使用 🤗 Transformers 的方法是预先下载好文件，然后在需要离线使用时指向它们的离线路径。有三种实现的方法：
* 单击 [Model Hub](https://huggingface.co/models) 用户界面上的 ↓ 图标下载文件。
* 使用 [`PreTrainedModel.from_pretrained`] 和 [`PreTrainedModel.save_pretrained`] 工作流程：
1. 预先使用 [`PreTrainedModel.from_pretrained`] 下载文件：
```py
>>> from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
>>> tokenizer = AutoTokenizer.from_pretrained("bigscience/T0_3B")
>>> model = AutoModelForSeq2SeqLM.from_pretrained("bigscience/T0_3B")
```
2. 使用 [`PreTrainedModel.save_pretrained`] 将文件保存至指定目录：
```py
>>> tokenizer.save_pretrained("./your/path/bigscience_t0")
>>> model.save_pretrained("./your/path/bigscience_t0")
```
3. 现在，你可以在离线时从指定目录使用 [`PreTrainedModel.from_pretrained`] 重新加载你的文件：
```py
>>> tokenizer = AutoTokenizer.from_pretrained("./your/path/bigscience_t0")
>>> model = AutoModel.from_pretrained("./your/path/bigscience_t0")
```
* 使用代码用 [huggingface_hub](https://github.com/huggingface/huggingface_hub/tree/main/src/huggingface_hub) 库下载文件：
1. 在你的虚拟环境中安装 `huggingface_hub` 库：
```bash
python -m pip install huggingface_hub
```
2. 使用 [`hf_hub_download`](https://huggingface.co/docs/hub/adding-a-library#download-files-from-the-hub) 函数将文件下载到指定路径。例如，以下命令将 `config.json` 文件从 [T0](https://huggingface.co/bigscience/T0_3B) 模型下载至你想要的路径：
```py
>>> from huggingface_hub import hf_hub_download
>>> hf_hub_download(repo_id="bigscience/T0_3B", filename="config.json", cache_dir="./your/path/bigscience_t0")
```
下载完文件并在本地缓存后，指定其本地路径以加载和使用该模型：
```py
>>> from transformers import AutoConfig
>>> config = AutoConfig.from_pretrained("./your/path/bigscience_t0/config.json")
```
<Tip>
请参阅 [如何从 Hub 下载文件](https://huggingface.co/docs/hub/how-to-downstream) 部分，获取有关下载存储在 Hub 上文件的更多详细信息。
</Tip>
| transformers/docs/source/zh/installation.md/0 | {
"file_path": "transformers/docs/source/zh/installation.md",
"repo_id": "transformers",
"token_count": 4833
} | 404 |
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# 使用 🤗 PEFT 加载adapters
[[open-in-colab]]
[参数高效微调（PEFT）方法](https://huggingface.co/blog/peft)在微调过程中冻结预训练模型的参数，并在其顶部添加少量可训练参数（adapters）。adapters被训练以学习特定任务的信息。这种方法已被证明非常节省内存，同时具有较低的计算使用量，同时产生与完全微调模型相当的结果。
使用PEFT训练的adapters通常比完整模型小一个数量级，使其方便共享、存储和加载。
<div class="flex flex-col justify-center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/PEFT-hub-screenshot.png"/>
<figcaption class="text-center">与完整尺寸的模型权重（约为700MB）相比，存储在Hub上的OPTForCausalLM模型的adapter权重仅为~6MB。</figcaption>
</div>
如果您对学习更多关于🤗 PEFT库感兴趣，请查看[文档](https://huggingface.co/docs/peft/index)。
## 设置
首先安装 🤗 PEFT：
```bash
pip install peft
```
如果你想尝试全新的特性，你可能会有兴趣从源代码安装这个库：
```bash
pip install git+https://github.com/huggingface/peft.git
```
## 支持的 PEFT 模型
Transformers原生支持一些PEFT方法，这意味着你可以加载本地存储或在Hub上的adapter权重，并使用几行代码轻松运行或训练它们。以下是受支持的方法：
- [Low Rank Adapters](https://huggingface.co/docs/peft/conceptual_guides/lora)
- [IA3](https://huggingface.co/docs/peft/conceptual_guides/ia3)
- [AdaLoRA](https://huggingface.co/papers/2303.10512)
如果你想使用其他PEFT方法，例如提示学习或提示微调，或者关于通用的 🤗 PEFT库，请参阅[文档](https://huggingface.co/docs/peft/index)。
## 加载 PEFT adapter
要从huggingface的Transformers库中加载并使用PEFTadapter模型，请确保Hub仓库或本地目录包含一个`adapter_config.json`文件和adapter权重，如上例所示。然后，您可以使用`AutoModelFor`类加载PEFT adapter模型。例如，要为因果语言建模加载一个PEFT adapter模型：
1. 指定PEFT模型id
2. 将其传递给[`AutoModelForCausalLM`]类
```py
from transformers import AutoModelForCausalLM, AutoTokenizer
peft_model_id = "ybelkada/opt-350m-lora"
model = AutoModelForCausalLM.from_pretrained(peft_model_id)
```
<Tip>
你可以使用`AutoModelFor`类或基础模型类（如`OPTForCausalLM`或`LlamaForCausalLM`）来加载一个PEFT adapter。
</Tip>
您也可以通过`load_adapter`方法来加载 PEFT adapter。
```py
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "facebook/opt-350m"
peft_model_id = "ybelkada/opt-350m-lora"
model = AutoModelForCausalLM.from_pretrained(model_id)
model.load_adapter(peft_model_id)
```
## 基于8bit或4bit进行加载
`bitsandbytes`集成支持8bit和4bit精度数据类型，这对于加载大模型非常有用，因为它可以节省内存（请参阅`bitsandbytes`[指南](./quantization#bitsandbytes-integration)以了解更多信息）。要有效地将模型分配到您的硬件，请在[`~PreTrainedModel.from_pretrained`]中添加`load_in_8bit`或`load_in_4bit`参数，并将`device_map="auto"`设置为：
```py
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
peft_model_id = "ybelkada/opt-350m-lora"
model = AutoModelForCausalLM.from_pretrained(peft_model_id, quantization_config=BitsAndBytesConfig(load_in_8bit=True))
```
## 添加新的adapter
你可以使用[`~peft.PeftModel.add_adapter`]方法为一个已有adapter的模型添加一个新的adapter，只要新adapter的类型与当前adapter相同即可。例如，如果你有一个附加到模型上的LoRA adapter：
```py
from transformers import AutoModelForCausalLM, OPTForCausalLM, AutoTokenizer
from peft import PeftConfig
model_id = "facebook/opt-350m"
model = AutoModelForCausalLM.from_pretrained(model_id)
lora_config = LoraConfig(
target_modules=["q_proj", "k_proj"],
init_lora_weights=False
)
model.add_adapter(lora_config, adapter_name="adapter_1")
```
添加一个新的adapter：
```py
# attach new adapter with same config
model.add_adapter(lora_config, adapter_name="adapter_2")
```
现在您可以使用[`~peft.PeftModel.set_adapter`]来设置要使用的adapter。
```py
# use adapter_1
model.set_adapter("adapter_1")
output = model.generate(**inputs)
print(tokenizer.decode(output_disabled[0], skip_special_tokens=True))
# use adapter_2
model.set_adapter("adapter_2")
output_enabled = model.generate(**inputs)
print(tokenizer.decode(output_enabled[0], skip_special_tokens=True))
```
## 启用和禁用adapters
一旦您将adapter添加到模型中，您可以启用或禁用adapter模块。要启用adapter模块：
```py
from transformers import AutoModelForCausalLM, OPTForCausalLM, AutoTokenizer
from peft import PeftConfig
model_id = "facebook/opt-350m"
adapter_model_id = "ybelkada/opt-350m-lora"
tokenizer = AutoTokenizer.from_pretrained(model_id)
text = "Hello"
inputs = tokenizer(text, return_tensors="pt")
model = AutoModelForCausalLM.from_pretrained(model_id)
peft_config = PeftConfig.from_pretrained(adapter_model_id)
# to initiate with random weights
peft_config.init_lora_weights = False
model.add_adapter(peft_config)
model.enable_adapters()
output = model.generate(**inputs)
```
要禁用adapter模块：
```py
model.disable_adapters()
output = model.generate(**inputs)
```
## 训练一个 PEFT adapter
PEFT适配器受[`Trainer`]类支持，因此您可以为您的特定用例训练适配器。它只需要添加几行代码即可。例如，要训练一个LoRA adapter：
<Tip>
如果你不熟悉如何使用[`Trainer`]微调模型，请查看[微调预训练模型](training)教程。
</Tip>
1. 使用任务类型和超参数定义adapter配置（参见[`~peft.LoraConfig`]以了解超参数的详细信息）。
```py
from peft import LoraConfig
peft_config = LoraConfig(
lora_alpha=16,
lora_dropout=0.1,
r=64,
bias="none",
task_type="CAUSAL_LM",
)
```
2. 将adapter添加到模型中。
```py
model.add_adapter(peft_config)
```
3. 现在可以将模型传递给[`Trainer`]了！
```py
trainer = Trainer(model=model, ...)
trainer.train()
```
要保存训练好的adapter并重新加载它：
```py
model.save_pretrained(save_dir)
model = AutoModelForCausalLM.from_pretrained(save_dir)
```
<!--
TODO: (@younesbelkada @stevhliu)
- Link to PEFT docs for further details
- Trainer
- 8-bit / 4-bit examples ?
-->
| transformers/docs/source/zh/peft.md/0 | {
"file_path": "transformers/docs/source/zh/peft.md",
"repo_id": "transformers",
"token_count": 3639
} | 405 |
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# 导出为 TFLite
[TensorFlow Lite](https://www.tensorflow.org/lite/guide) 是一个轻量级框架，用于资源受限的设备上，如手机、嵌入式系统和物联网（IoT）设备，部署机器学习模型。TFLite 旨在在计算能力、内存和功耗有限的设备上优化和高效运行模型。模型以一种特殊的高效可移植格式表示，其文件扩展名为 `.tflite`。
🤗 Optimum 通过 `exporters.tflite` 模块提供将 🤗 Transformers 模型导出至 TFLite 格式的功能。请参考 [🤗 Optimum 文档](https://huggingface.co/docs/optimum/exporters/tflite/overview) 以获取支持的模型架构列表。
要将模型导出为 TFLite 格式，请安装所需的依赖项：
```bash
pip install optimum[exporters-tf]
```
请参阅 [🤗 Optimum 文档](https://huggingface.co/docs/optimum/main/en/exporters/tflite/usage_guides/export_a_model) 以查看所有可用参数，或者在命令行中查看帮助：
```bash
optimum-cli export tflite --help
```
运行以下命令，以从 🤗 Hub 导出模型的检查点（checkpoint），以 `google-bert/bert-base-uncased` 为例：
```bash
optimum-cli export tflite --model google-bert/bert-base-uncased --sequence_length 128 bert_tflite/
```
你应该能在日志中看到导出进度以及生成的 `model.tflite` 文件的保存位置，如下所示：
```bash
Validating TFLite model...
-[✓] TFLite model output names match reference model (logits)
- Validating TFLite Model output "logits":
-[✓] (1, 128, 30522) matches (1, 128, 30522)
-[x] values not close enough, max diff: 5.817413330078125e-05 (atol: 1e-05)
The TensorFlow Lite export succeeded with the warning: The maximum absolute difference between the output of the reference model and the TFLite exported model is not within the set tolerance 1e-05:
- logits: max diff = 5.817413330078125e-05.
The exported model was saved at: bert_tflite
```
上面的示例说明了从 🤗 Hub 导出检查点的过程。导出本地模型时，首先需要确保将模型的权重和分词器文件保存在同一目录（`local_path`）中。在使用 CLI（命令行）时，将 `local_path` 传递给 `model` 参数，而不是 🤗 Hub 上的检查点名称。 | transformers/docs/source/zh/tflite.md/0 | {
"file_path": "transformers/docs/source/zh/tflite.md",
"repo_id": "transformers",
"token_count": 1386
} | 406 |
#!/usr/bin/env python3
import time
from argparse import ArgumentParser
import jax
import numpy as np
from transformers import BertConfig, FlaxBertModel
parser = ArgumentParser()
parser.add_argument("--precision", type=str, choices=["float32", "bfloat16"], default="float32")
args = parser.parse_args()
dtype = jax.numpy.float32
if args.precision == "bfloat16":
dtype = jax.numpy.bfloat16
VOCAB_SIZE = 30522
BS = 32
SEQ_LEN = 128
def get_input_data(batch_size=1, seq_length=384):
shape = (batch_size, seq_length)
input_ids = np.random.randint(1, VOCAB_SIZE, size=shape).astype(np.int32)
token_type_ids = np.ones(shape).astype(np.int32)
attention_mask = np.ones(shape).astype(np.int32)
return {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": attention_mask}
inputs = get_input_data(BS, SEQ_LEN)
config = BertConfig.from_pretrained("bert-base-uncased", hidden_act="gelu_new")
model = FlaxBertModel.from_pretrained("bert-base-uncased", config=config, dtype=dtype)
@jax.jit
def func():
outputs = model(**inputs)
return outputs
(nwarmup, nbenchmark) = (5, 100)
# warmpup
for _ in range(nwarmup):
func()
# benchmark
start = time.time()
for _ in range(nbenchmark):
func()
end = time.time()
print(end - start)
print(f"Throughput: {((nbenchmark * BS) / (end - start)):.3f} examples/sec")
| transformers/examples/flax/language-modeling/run_bert_flax.py/0 | {
"file_path": "transformers/examples/flax/language-modeling/run_bert_flax.py",
"repo_id": "transformers",
"token_count": 527
} | 407 |
<!---
Copyright 2021 The Google Flax Team Authors and HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
-->
# Text classification examples
## GLUE tasks
Based on the script [`run_flax_glue.py`](https://github.com/huggingface/transformers/blob/main/examples/flax/text-classification/run_flax_glue.py).
Fine-tuning the library models for sequence classification on the GLUE benchmark: [General Language Understanding
Evaluation](https://gluebenchmark.com/). This script can fine-tune any of the models on the [hub](https://huggingface.co/models) and can also be used for a
dataset hosted on our [hub](https://huggingface.co/datasets) or your own data in a csv or a JSON file (the script might need some tweaks in that case,
refer to the comments inside for help).
GLUE is made up of a total of 9 different tasks. Here is how to run the script on one of them:
```bash
export TASK_NAME=mrpc
python run_flax_glue.py \
--model_name_or_path google-bert/bert-base-cased \
--task_name ${TASK_NAME} \
--max_seq_length 128 \
--learning_rate 2e-5 \
--num_train_epochs 3 \
--per_device_train_batch_size 4 \
--eval_steps 100 \
--output_dir ./$TASK_NAME/ \
--push_to_hub
```
where task name can be one of cola, mnli, mnli_mismatched, mnli_matched, mrpc, qnli, qqp, rte, sst2, stsb, wnli.
Using the command above, the script will train for 3 epochs and run eval after each epoch.
Metrics and hyperparameters are stored in Tensorflow event files in `--output_dir`.
You can see the results by running `tensorboard` in that directory:
```bash
$ tensorboard --logdir .
```
or directly on the hub under *Training metrics*.
### Accuracy Evaluation
We train five replicas and report mean accuracy and stdev on the dev set below.
We use the settings as in the command above (with an exception for MRPC and
WNLI which are tiny and where we used 5 epochs instead of 3), and we use a total
train batch size of 32 (we train on 8 Cloud v3 TPUs, so a per-device batch size of 4),
On the task other than MRPC and WNLI we train for 3 these epochs because this is the standard,
but looking at the training curves of some of them (e.g., SST-2, STS-b), it appears the models
are undertrained and we could get better results when training longer.
In the Tensorboard results linked below, the random seed of each model is equal to the ID of the run. So in order to reproduce run 1, run the command above with `--seed=1`. The best run used random seed 3, which is the default in the script. The results of all runs are in [this Google Sheet](https://docs.google.com/spreadsheets/d/1p3XzReMO75m_XdEJvPue-PIq_PN-96J2IJpJW1yS-10/edit?usp=sharing).
| Task | Metric | Acc (best run) | Acc (avg/5runs) | Stdev | Metrics |
|-------|------------------------------|----------------|-----------------|-----------|--------------------------------------------------------------------------|
| CoLA | Matthews corr | 60.57 | 59.04 | 1.06 | [tfhub.dev](https://tensorboard.dev/experiment/lfr2adVpRtmLDALKrElkzg/) |
| SST-2 | Accuracy | 92.66 | 92.23 | 0.57 | [tfhub.dev](https://tensorboard.dev/experiment/jYvfv2trRHKMjoWnXVwrZA/) |
| MRPC | F1/Accuracy | 89.90/85.78 | 88.97/84.36 | 0.72/1.09 | [tfhub.dev](https://tensorboard.dev/experiment/bo3W3DEoRw2Q7YXjWrJkfg/) |
| STS-B | Pearson/Spearman corr. | 89.04/88.70 | 88.94/88.63 | 0.07/0.07 | [tfhub.dev](https://tensorboard.dev/experiment/fxVwbLD7QpKhbot0r9rn2w/) |
| QQP | Accuracy/F1 | 90.81/87.58 | 90.76/87.51 | 0.05/0.06 | [tfhub.dev](https://tensorboard.dev/experiment/di089Rc9TZmsnKRMrYNLsA/) |
| MNLI | Matched acc. | 84.10 | 83.80 | 0.16 | [tfhub.dev](https://tensorboard.dev/experiment/JgNCGHDJSRaW6HBx6YQFYQ/) |
| QNLI | Accuracy | 91.01 | 90.82 | 0.17 | [tfhub.dev](https://tensorboard.dev/experiment/Bq7cMGJnQMSggYgL8qNGeQ/) |
| RTE | Accuracy | 66.06 | 64.76 | 1.04 | [tfhub.dev](https://tensorboard.dev/experiment/66Eq24bhRjqN6CEhgDSGqQ/) |
| WNLI | Accuracy | 46.48 | 37.01 | 6.83 | [tfhub.dev](https://tensorboard.dev/experiment/TAqcnddqTkWvVEeGaWwIdQ/) |
Some of these results are significantly different from the ones reported on the test set of GLUE benchmark on the
website. For QQP and WNLI, please refer to [FAQ #12](https://gluebenchmark.com/faq) on the website.
### Runtime evaluation
We also ran each task once on a single V100 GPU, 8 V100 GPUs, and 8 Cloud v3 TPUs and report the
overall training time below. For comparison we ran Pytorch's [run_glue.py](https://github.com/huggingface/transformers/blob/main/examples/pytorch/text-classification/run_glue.py) on a single GPU (last column).
| Task | TPU v3-8 | 8 GPU | [1 GPU](https://tensorboard.dev/experiment/mkPS4Zh8TnGe1HB6Yzwj4Q) | 1 GPU (Pytorch) |
|-------|-----------|------------|------------|-----------------|
| CoLA | 1m 42s | 1m 26s | 3m 9s | 4m 6s |
| SST-2 | 5m 12s | 6m 28s | 22m 33s | 34m 37s |
| MRPC | 1m 29s | 1m 14s | 2m 20s | 2m 56s |
| STS-B | 1m 30s | 1m 12s | 2m 16s | 2m 48s |
| QQP | 22m 50s | 31m 48s | 1h 59m 41s | 2h 54m |
| MNLI | 25m 03s | 33m 55s | 2h 9m 37s | 3h 7m 6s |
| QNLI | 7m30s | 9m 40s | 34m 40s | 49m 8s |
| RTE | 1m 20s | 55s | 1m 10s | 1m 16s |
| WNLI | 1m 11s | 48s | 39s | 36s |
|-------|
| **TOTAL** | 1h 03m | 1h 28m | 5h 16m | 6h 37m |
*All experiments are ran on Google Cloud Platform.
GPU experiments are ran without further optimizations besides JAX
transformations. GPU experiments are ran with full precision (fp32). "TPU v3-8"
are 8 TPU cores on 4 chips (each chips has 2 cores), while "8 GPU" are 8 GPU chips.
| transformers/examples/flax/text-classification/README.md/0 | {
"file_path": "transformers/examples/flax/text-classification/README.md",
"repo_id": "transformers",
"token_count": 2760
} | 408 |
import argparse
import logging
import os
from pathlib import Path
from typing import Any
import pytorch_lightning as pl
from pytorch_lightning.utilities import rank_zero_info
from transformers import (
AutoConfig,
AutoModel,
AutoModelForPreTraining,
AutoModelForQuestionAnswering,
AutoModelForSeq2SeqLM,
AutoModelForSequenceClassification,
AutoModelForTokenClassification,
AutoModelWithLMHead,
AutoTokenizer,
PretrainedConfig,
PreTrainedTokenizer,
is_torch_available,
)
from transformers.optimization import (
Adafactor,
get_cosine_schedule_with_warmup,
get_cosine_with_hard_restarts_schedule_with_warmup,
get_linear_schedule_with_warmup,
get_polynomial_decay_schedule_with_warmup,
)
from transformers.utils.versions import require_version
if is_torch_available():
import torch
logger = logging.getLogger(__name__)
require_version("pytorch_lightning>=1.0.4")
MODEL_MODES = {
"base": AutoModel,
"sequence-classification": AutoModelForSequenceClassification,
"question-answering": AutoModelForQuestionAnswering,
"pretraining": AutoModelForPreTraining,
"token-classification": AutoModelForTokenClassification,
"language-modeling": AutoModelWithLMHead,
"summarization": AutoModelForSeq2SeqLM,
"translation": AutoModelForSeq2SeqLM,
}
# update this and the import above to support new schedulers from transformers.optimization
arg_to_scheduler = {
"linear": get_linear_schedule_with_warmup,
"cosine": get_cosine_schedule_with_warmup,
"cosine_w_restarts": get_cosine_with_hard_restarts_schedule_with_warmup,
"polynomial": get_polynomial_decay_schedule_with_warmup,
# '': get_constant_schedule, # not supported for now
# '': get_constant_schedule_with_warmup, # not supported for now
}
arg_to_scheduler_choices = sorted(arg_to_scheduler.keys())
arg_to_scheduler_metavar = "{" + ", ".join(arg_to_scheduler_choices) + "}"
class BaseTransformer(pl.LightningModule):
def __init__(
self,
hparams: argparse.Namespace,
num_labels=None,
mode="base",
config=None,
tokenizer=None,
model=None,
**config_kwargs,
):
"""Initialize a model, tokenizer and config."""
super().__init__()
# TODO: move to self.save_hyperparameters()
# self.save_hyperparameters()
# can also expand arguments into trainer signature for easier reading
self.save_hyperparameters(hparams)
self.step_count = 0
self.output_dir = Path(self.hparams.output_dir)
cache_dir = self.hparams.cache_dir if self.hparams.cache_dir else None
if config is None:
self.config = AutoConfig.from_pretrained(
self.hparams.config_name if self.hparams.config_name else self.hparams.model_name_or_path,
**({"num_labels": num_labels} if num_labels is not None else {}),
cache_dir=cache_dir,
**config_kwargs,
)
else:
self.config: PretrainedConfig = config
extra_model_params = ("encoder_layerdrop", "decoder_layerdrop", "dropout", "attention_dropout")
for p in extra_model_params:
if getattr(self.hparams, p, None):
assert hasattr(self.config, p), f"model config doesn't have a `{p}` attribute"
setattr(self.config, p, getattr(self.hparams, p))
if tokenizer is None:
self.tokenizer = AutoTokenizer.from_pretrained(
self.hparams.tokenizer_name if self.hparams.tokenizer_name else self.hparams.model_name_or_path,
cache_dir=cache_dir,
)
else:
self.tokenizer: PreTrainedTokenizer = tokenizer
self.model_type = MODEL_MODES[mode]
if model is None:
self.model = self.model_type.from_pretrained(
self.hparams.model_name_or_path,
from_tf=bool(".ckpt" in self.hparams.model_name_or_path),
config=self.config,
cache_dir=cache_dir,
)
else:
self.model = model
def load_hf_checkpoint(self, *args, **kwargs):
self.model = self.model_type.from_pretrained(*args, **kwargs)
def get_lr_scheduler(self):
get_schedule_func = arg_to_scheduler[self.hparams.lr_scheduler]
scheduler = get_schedule_func(
self.opt, num_warmup_steps=self.hparams.warmup_steps, num_training_steps=self.total_steps()
)
scheduler = {"scheduler": scheduler, "interval": "step", "frequency": 1}
return scheduler
def configure_optimizers(self):
"""Prepare optimizer and schedule (linear warmup and decay)"""
model = self.model
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": self.hparams.weight_decay,
},
{
"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)],
"weight_decay": 0.0,
},
]
if self.hparams.adafactor:
optimizer = Adafactor(
optimizer_grouped_parameters, lr=self.hparams.learning_rate, scale_parameter=False, relative_step=False
)
else:
optimizer = torch.optim.AdamW(
optimizer_grouped_parameters, lr=self.hparams.learning_rate, eps=self.hparams.adam_epsilon
)
self.opt = optimizer
scheduler = self.get_lr_scheduler()
return [optimizer], [scheduler]
def test_step(self, batch, batch_nb):
return self.validation_step(batch, batch_nb)
def test_epoch_end(self, outputs):
return self.validation_end(outputs)
def total_steps(self) -> int:
"""The number of total training steps that will be run. Used for lr scheduler purposes."""
num_devices = max(1, self.hparams.gpus) # TODO: consider num_tpu_cores
effective_batch_size = self.hparams.train_batch_size * self.hparams.accumulate_grad_batches * num_devices
return (self.dataset_size / effective_batch_size) * self.hparams.max_epochs
def setup(self, mode):
if mode == "test":
self.dataset_size = len(self.test_dataloader().dataset)
else:
self.train_loader = self.get_dataloader("train", self.hparams.train_batch_size, shuffle=True)
self.dataset_size = len(self.train_dataloader().dataset)
def get_dataloader(self, type_path: str, batch_size: int, shuffle: bool = False):
raise NotImplementedError("You must implement this for your task")
def train_dataloader(self):
return self.train_loader
def val_dataloader(self):
return self.get_dataloader("dev", self.hparams.eval_batch_size, shuffle=False)
def test_dataloader(self):
return self.get_dataloader("test", self.hparams.eval_batch_size, shuffle=False)
def _feature_file(self, mode):
return os.path.join(
self.hparams.data_dir,
"cached_{}_{}_{}".format(
mode,
list(filter(None, self.hparams.model_name_or_path.split("/"))).pop(),
str(self.hparams.max_seq_length),
),
)
@pl.utilities.rank_zero_only
def on_save_checkpoint(self, checkpoint: dict[str, Any]) -> None:
save_path = self.output_dir.joinpath("best_tfmr")
self.model.config.save_step = self.step_count
self.model.save_pretrained(save_path)
self.tokenizer.save_pretrained(save_path)
@staticmethod
def add_model_specific_args(parser, root_dir):
parser.add_argument(
"--model_name_or_path",
default=None,
type=str,
required=True,
help="Path to pretrained model or model identifier from huggingface.co/models",
)
parser.add_argument(
"--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name"
)
parser.add_argument(
"--tokenizer_name",
default=None,
type=str,
help="Pretrained tokenizer name or path if not the same as model_name",
)
parser.add_argument(
"--cache_dir",
default="",
type=str,
help="Where do you want to store the pre-trained models downloaded from huggingface.co",
)
parser.add_argument(
"--encoder_layerdrop",
type=float,
help="Encoder layer dropout probability (Optional). Goes into model.config",
)
parser.add_argument(
"--decoder_layerdrop",
type=float,
help="Decoder layer dropout probability (Optional). Goes into model.config",
)
parser.add_argument(
"--dropout",
type=float,
help="Dropout probability (Optional). Goes into model.config",
)
parser.add_argument(
"--attention_dropout",
type=float,
help="Attention dropout probability (Optional). Goes into model.config",
)
parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
parser.add_argument(
"--lr_scheduler",
default="linear",
choices=arg_to_scheduler_choices,
metavar=arg_to_scheduler_metavar,
type=str,
help="Learning rate scheduler",
)
parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
parser.add_argument("--num_workers", default=4, type=int, help="kwarg passed to DataLoader")
parser.add_argument("--num_train_epochs", dest="max_epochs", default=3, type=int)
parser.add_argument("--train_batch_size", default=32, type=int)
parser.add_argument("--eval_batch_size", default=32, type=int)
parser.add_argument("--adafactor", action="store_true")
class LoggingCallback(pl.Callback):
def on_batch_end(self, trainer, pl_module):
lr_scheduler = trainer.lr_schedulers[0]["scheduler"]
lrs = {f"lr_group_{i}": lr for i, lr in enumerate(lr_scheduler.get_lr())}
pl_module.logger.log_metrics(lrs)
def on_validation_end(self, trainer: pl.Trainer, pl_module: pl.LightningModule):
rank_zero_info("***** Validation results *****")
metrics = trainer.callback_metrics
# Log results
for key in sorted(metrics):
if key not in ["log", "progress_bar"]:
rank_zero_info(f"{key} = {str(metrics[key])}\n")
def on_test_end(self, trainer: pl.Trainer, pl_module: pl.LightningModule):
rank_zero_info("***** Test results *****")
metrics = trainer.callback_metrics
# Log and save results to file
output_test_results_file = os.path.join(pl_module.hparams.output_dir, "test_results.txt")
with open(output_test_results_file, "w") as writer:
for key in sorted(metrics):
if key not in ["log", "progress_bar"]:
rank_zero_info(f"{key} = {str(metrics[key])}\n")
writer.write(f"{key} = {str(metrics[key])}\n")
def add_generic_args(parser, root_dir) -> None:
# To allow all pl args uncomment the following line
# parser = pl.Trainer.add_argparse_args(parser)
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model predictions and checkpoints will be written.",
)
parser.add_argument(
"--fp16",
action="store_true",
help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
)
parser.add_argument(
"--fp16_opt_level",
type=str,
default="O2",
help=(
"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']. "
"See details at https://nvidia.github.io/apex/amp.html"
),
)
parser.add_argument("--n_tpu_cores", dest="tpu_cores", type=int)
parser.add_argument("--max_grad_norm", dest="gradient_clip_val", default=1.0, type=float, help="Max gradient norm")
parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
parser.add_argument("--do_predict", action="store_true", help="Whether to run predictions on the test set.")
parser.add_argument(
"--gradient_accumulation_steps",
dest="accumulate_grad_batches",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help="The input data dir. Should contain the training files for the CoNLL-2003 NER task.",
)
def generic_train(
model: BaseTransformer,
args: argparse.Namespace,
early_stopping_callback=None,
logger=True, # can pass WandbLogger() here
extra_callbacks=[],
checkpoint_callback=None,
logging_callback=None,
**extra_train_kwargs,
):
pl.seed_everything(args.seed)
# init model
odir = Path(model.hparams.output_dir)
odir.mkdir(exist_ok=True)
# add custom checkpoints
if checkpoint_callback is None:
checkpoint_callback = pl.callbacks.ModelCheckpoint(
filepath=args.output_dir, prefix="checkpoint", monitor="val_loss", mode="min", save_top_k=1
)
if early_stopping_callback:
extra_callbacks.append(early_stopping_callback)
if logging_callback is None:
logging_callback = LoggingCallback()
train_params = {}
# TODO: remove with PyTorch 1.6 since pl uses native amp
if args.fp16:
train_params["precision"] = 16
train_params["amp_level"] = args.fp16_opt_level
if args.gpus > 1:
train_params["distributed_backend"] = "ddp"
train_params["accumulate_grad_batches"] = args.accumulate_grad_batches
train_params["accelerator"] = extra_train_kwargs.get("accelerator")
train_params["profiler"] = extra_train_kwargs.get("profiler")
trainer = pl.Trainer.from_argparse_args(
args,
weights_summary=None,
callbacks=[logging_callback] + extra_callbacks,
logger=logger,
checkpoint_callback=checkpoint_callback,
**train_params,
)
if args.do_train:
trainer.fit(model)
return trainer
| transformers/examples/legacy/pytorch-lightning/lightning_base.py/0 | {
"file_path": "transformers/examples/legacy/pytorch-lightning/lightning_base.py",
"repo_id": "transformers",
"token_count": 6613
} | 409 |
<!---
Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
-->
# Sequence-to-Sequence Training and Evaluation
This directory contains examples for finetuning and evaluating transformers on summarization and translation tasks.
For deprecated `bertabs` instructions, see https://github.com/huggingface/transformers-research-projects/blob/main/bertabs/README.md.
### Supported Architectures
- `BartForConditionalGeneration`
- `MarianMTModel`
- `PegasusForConditionalGeneration`
- `MBartForConditionalGeneration`
- `FSMTForConditionalGeneration`
- `T5ForConditionalGeneration`
### Download the Datasets
#### XSUM
```bash
cd examples/legacy/seq2seq
wget https://cdn-datasets.huggingface.co/summarization/xsum.tar.gz
tar -xzvf xsum.tar.gz
export XSUM_DIR=${PWD}/xsum
```
this should make a directory called `xsum/` with files like `test.source`.
To use your own data, copy that files format. Each article to be summarized is on its own line.
#### CNN/DailyMail
```bash
cd examples/legacy/seq2seq
wget https://cdn-datasets.huggingface.co/summarization/cnn_dm_v2.tgz
tar -xzvf cnn_dm_v2.tgz # empty lines removed
mv cnn_cln cnn_dm
export CNN_DIR=${PWD}/cnn_dm
```
this should make a directory called `cnn_dm/` with 6 files.
#### WMT16 English-Romanian Translation Data
download with this command:
```bash
wget https://cdn-datasets.huggingface.co/translation/wmt_en_ro.tar.gz
tar -xzvf wmt_en_ro.tar.gz
export ENRO_DIR=${PWD}/wmt_en_ro
```
this should make a directory called `wmt_en_ro/` with 6 files.
#### WMT English-German
```bash
wget https://cdn-datasets.huggingface.co/translation/wmt_en_de.tgz
tar -xzvf wmt_en_de.tgz
export DATA_DIR=${PWD}/wmt_en_de
```
#### FSMT datasets (wmt)
Refer to the scripts starting with `eval_` under:
https://github.com/huggingface/transformers/tree/main/scripts/fsmt
#### Pegasus (multiple datasets)
Multiple eval datasets are available for download from:
https://github.com/stas00/porting/tree/master/datasets/pegasus
#### Your Data
If you are using your own data, it must be formatted as one directory with 6 files:
```
train.source
train.target
val.source
val.target
test.source
test.target
```
The `.source` files are the input, the `.target` files are the desired output.
### Potential issues
- native AMP (`--fp16` and no apex) may lead to a huge memory leak and require 10x gpu memory. This has been fixed in pytorch-nightly and the minimal official version to have this fix will be pytorch-1.7.1. Until then if you have to use mixed precision please use AMP only with pytorch-nightly or NVIDIA's apex. Reference: https://github.com/huggingface/transformers/issues/8403
### Tips and Tricks
General Tips:
- since you need to run from `examples/legacy/seq2seq`, and likely need to modify code, the easiest workflow is fork transformers, clone your fork, and run `pip install -e .` before you get started.
- try `--freeze_encoder` or `--freeze_embeds` for faster training/larger batch size. (3hr per epoch with bs=8, see the "xsum_shared_task" command below)
- `fp16_opt_level=O1` (the default works best).
- In addition to the pytorch-lightning .ckpt checkpoint, a transformers checkpoint will be saved.
Load it with `BartForConditionalGeneration.from_pretrained(f'{output_dir}/best_tfmr)`.
- At the moment, `--do_predict` does not work in a multi-gpu setting. You need to use `evaluate_checkpoint` or the `run_eval.py` code.
- This warning can be safely ignored:
> "Some weights of BartForConditionalGeneration were not initialized from the model checkpoint at facebook/bart-large-xsum and are newly initialized: ['final_logits_bias']"
- Both finetuning and eval are 30% faster with `--fp16`. For that you need to [install apex](https://github.com/NVIDIA/apex#quick-start).
- Read scripts before you run them!
Summarization Tips:
- (summ) 1 epoch at batch size 1 for bart-large takes 24 hours and requires 13GB GPU RAM with fp16 on an NVIDIA-V100.
- If you want to run experiments on improving the summarization finetuning process, try the XSUM Shared Task (below). It's faster to train than CNNDM because the summaries are shorter.
- For CNN/DailyMail, the default `val_max_target_length` and `test_max_target_length` will truncate the ground truth labels, resulting in slightly higher rouge scores. To get accurate rouge scores, you should rerun calculate_rouge on the `{output_dir}/test_generations.txt` file saved by `trainer.test()`
- `--max_target_length=60 --val_max_target_length=60 --test_max_target_length=100 ` is a reasonable setting for XSUM.
- `wandb` can be used by specifying `--logger_name wandb`. It is useful for reproducibility. Specify the environment variable `WANDB_PROJECT='hf_xsum'` to do the XSUM shared task.
- If you are finetuning on your own dataset, start from `distilbart-cnn-12-6` if you want long summaries and `distilbart-xsum-12-6` if you want short summaries.
(It rarely makes sense to start from `bart-large` unless you are a researching finetuning methods).
**Update 2018-07-18**
Datasets: `LegacySeq2SeqDataset` will be used for all tokenizers without a `prepare_seq2seq_batch` method. Otherwise, `Seq2SeqDataset` will be used.
Future work/help wanted: A new dataset to support multilingual tasks.
### Fine-tuning using Seq2SeqTrainer
To use `Seq2SeqTrainer` for fine-tuning you should use the `finetune_trainer.py` script. It subclasses `Trainer` to extend it for seq2seq training. Except the `Trainer`-related `TrainingArguments`, it shares the same argument names as that of `finetune.py` file. One notable difference is that calculating generative metrics (BLEU, ROUGE) is optional and is controlled using the `--predict_with_generate` argument.
With PyTorch 1.6+ it'll automatically use `native AMP` when `--fp16` is set.
To see all the possible command line options, run:
```bash
python finetune_trainer.py --help
```
For multi-gpu training use `torch.distributed.launch`, e.g. with 2 gpus:
```bash
torchrun --nproc_per_node=2 finetune_trainer.py ...
```
**At the moment, `Seq2SeqTrainer` does not support *with teacher* distillation.**
All `Seq2SeqTrainer`-based fine-tuning scripts are included in the `builtin_trainer` directory.
#### TPU Training
`Seq2SeqTrainer` supports TPU training with few caveats
1. As `generate` method does not work on TPU at the moment, `predict_with_generate` cannot be used. You should use `--prediction_loss_only` to only calculate loss, and do not set `--do_predict` and `--predict_with_generate`.
2. All sequences should be padded to be of equal length to avoid extremely slow training. (`finetune_trainer.py` does this automatically when running on TPU.)
We provide a very simple launcher script named `xla_spawn.py` that lets you run our example scripts on multiple TPU cores without any boilerplate. Just pass a `--num_cores` flag to this script, then your regular training script with its arguments (this is similar to the `torch.distributed.launch` helper for `torch.distributed`).
`builtin_trainer/finetune_tpu.sh` script provides minimal arguments needed for TPU training.
The following command fine-tunes `sshleifer/student_marian_en_ro_6_3` on TPU V3-8 and should complete one epoch in ~5-6 mins.
```bash
./builtin_trainer/train_distil_marian_enro_tpu.sh
```
## Evaluation Commands
To create summaries for each article in dataset, we use `run_eval.py`, here are a few commands that run eval for different tasks and models.
If 'translation' is in your task name, the computed metric will be BLEU. Otherwise, ROUGE will be used.
For t5, you need to specify --task translation_{src}_to_{tgt} as follows:
```bash
export DATA_DIR=wmt_en_ro
./run_eval.py google-t5/t5-base \
$DATA_DIR/val.source t5_val_generations.txt \
--reference_path $DATA_DIR/val.target \
--score_path enro_bleu.json \
--task translation_en_to_ro \
--n_obs 100 \
--device cuda \
--fp16 \
--bs 32
```
This command works for MBART, although the BLEU score is suspiciously low.
```bash
export DATA_DIR=wmt_en_ro
./run_eval.py facebook/mbart-large-en-ro $DATA_DIR/val.source mbart_val_generations.txt \
--reference_path $DATA_DIR/val.target \
--score_path enro_bleu.json \
--task translation \
--n_obs 100 \
--device cuda \
--fp16 \
--bs 32
```
Summarization (xsum will be very similar):
```bash
export DATA_DIR=cnn_dm
./run_eval.py sshleifer/distilbart-cnn-12-6 $DATA_DIR/val.source dbart_val_generations.txt \
--reference_path $DATA_DIR/val.target \
--score_path cnn_rouge.json \
--task summarization \
--n_obs 100 \
th 56 \
--fp16 \
--bs 32
```
### Multi-GPU Evaluation
here is a command to run xsum evaluation on 8 GPUs. It is more than linearly faster than run_eval.py in some cases
because it uses SortishSampler to minimize padding. You can also use it on 1 GPU. `data_dir` must have
`{type_path}.source` and `{type_path}.target`. Run `./run_distributed_eval.py --help` for all clargs.
```bash
torchrun --nproc_per_node=8 run_distributed_eval.py \
--model_name sshleifer/distilbart-large-xsum-12-3 \
--save_dir xsum_generations \
--data_dir xsum \
--fp16 # you can pass generate kwargs like num_beams here, just like run_eval.py
```
Contributions that implement this command for other distributed hardware setups are welcome!
#### Single-GPU Eval: Tips and Tricks
When using `run_eval.py`, the following features can be useful:
* if you running the script multiple times and want to make it easier to track what arguments produced that output, use `--dump-args`. Along with the results it will also dump any custom params that were passed to the script. For example if you used: `--num_beams 8 --early_stopping true`, the output will be:
```json
{'bleu': 26.887, 'n_obs': 10, 'runtime': 1, 'seconds_per_sample': 0.1, 'num_beams': 8, 'early_stopping': True}
```
`--info` is an additional argument available for the same purpose of tracking the conditions of the experiment. It's useful to pass things that weren't in the argument list, e.g. a language pair `--info "lang:en-ru"`. But also if you pass `--info` without a value it will fallback to the current date/time string, e.g. `2020-09-13 18:44:43`.
If using `--dump-args --info`, the output will be:
```json
{'bleu': 26.887, 'n_obs': 10, 'runtime': 1, 'seconds_per_sample': 0.1, 'num_beams': 8, 'early_stopping': True, 'info': '2020-09-13 18:44:43'}
```
If using `--dump-args --info "pair:en-ru chkpt=best`, the output will be:
```json
{'bleu': 26.887, 'n_obs': 10, 'runtime': 1, 'seconds_per_sample': 0.1, 'num_beams': 8, 'early_stopping': True, 'info': 'pair=en-ru chkpt=best'}
```
* if you need to perform a parametric search in order to find the best ones that lead to the highest BLEU score, let `run_eval_search.py` to do the searching for you.
The script accepts the exact same arguments as `run_eval.py`, plus an additional argument `--search`. The value of `--search` is parsed, reformatted and fed to ``run_eval.py`` as additional args.
The format for the `--search` value is a simple string with hparams and colon separated values to try, e.g.:
```
--search "num_beams=5:10 length_penalty=0.8:1.0:1.2 early_stopping=true:false"
```
which will generate `12` `(2*3*2)` searches for a product of each hparam. For example the example that was just used will invoke `run_eval.py` repeatedly with:
```
--num_beams 5 --length_penalty 0.8 --early_stopping true
--num_beams 5 --length_penalty 0.8 --early_stopping false
[...]
--num_beams 10 --length_penalty 1.2 --early_stopping false
```
On completion, this function prints a markdown table of the results sorted by the best BLEU score and the winning arguments.
```
bleu | num_beams | length_penalty | early_stopping
----- | --------- | -------------- | --------------
26.71 | 5 | 1.1 | 1
26.66 | 5 | 0.9 | 1
26.66 | 5 | 0.9 | 0
26.41 | 5 | 1.1 | 0
21.94 | 1 | 0.9 | 1
21.94 | 1 | 0.9 | 0
21.94 | 1 | 1.1 | 1
21.94 | 1 | 1.1 | 0
Best score args:
stas/wmt19-en-ru data/en-ru/val.source data/en-ru/test_translations.txt --reference_path data/en-ru/val.target --score_path data/en-ru/test_bleu.json --bs 8 --task translation --num_beams 5 --length_penalty 1.1 --early_stopping True
```
If you pass `--info "some experiment-specific info"` it will get printed before the results table - this is useful for scripting and multiple runs, so one can tell the different sets of results from each other.
### Contributing
- follow the standard contributing guidelines and code of conduct.
- add tests to `test_seq2seq_examples.py`
- To run only the seq2seq tests, you must be in the root of the repository and run:
```bash
pytest examples/seq2seq/
```
### Converting pytorch-lightning checkpoints
pytorch lightning ``-do_predict`` often fails, after you are done training, the best way to evaluate your model is to convert it.
This should be done for you, with a file called `{save_dir}/best_tfmr`.
If that file doesn't exist but you have a lightning `.ckpt` file, you can run
```bash
python convert_pl_checkpoint_to_hf.py PATH_TO_CKPT randomly_initialized_hf_model_path save_dir/best_tfmr
```
Then either `run_eval` or `run_distributed_eval` with `save_dir/best_tfmr` (see previous sections)
# Experimental Features
These features are harder to use and not always useful.
### Dynamic Batch Size for MT
`finetune.py` has a command line arg `--max_tokens_per_batch` that allows batches to be dynamically sized.
This feature can only be used:
- with fairseq installed
- on 1 GPU
- without sortish sampler
- after calling `./save_len_file.py $tok $data_dir`
For example,
```bash
./save_len_file.py Helsinki-NLP/opus-mt-en-ro wmt_en_ro
./dynamic_bs_example.sh --max_tokens_per_batch=2000 --output_dir benchmark_dynamic_bs
```
splits `wmt_en_ro/train` into 11,197 uneven length batches and can finish 1 epoch in 8 minutes on a v100.
For comparison,
```bash
./dynamic_bs_example.sh --sortish_sampler --train_batch_size 48
```
uses 12,723 batches of length 48 and takes slightly more time 9.5 minutes.
The feature is still experimental, because:
+ we can make it much more robust if we have memory mapped/preprocessed datasets.
+ The speedup over sortish sampler is not that large at the moment.
| transformers/examples/legacy/seq2seq/README.md/0 | {
"file_path": "transformers/examples/legacy/seq2seq/README.md",
"repo_id": "transformers",
"token_count": 5074
} | 410 |
### Motivation
Without processing, english-> romanian mbart-large-en-ro gets BLEU score 26.8 on the WMT data.
With post processing, it can score 37..
Here is the postprocessing code, stolen from @mjpost in this [issue](https://github.com/pytorch/fairseq/issues/1758)
### Instructions
Note: You need to have your test_generations.txt before you start this process.
(1) Setup `mosesdecoder` and `wmt16-scripts`
```bash
cd $HOME
git clone git@github.com:moses-smt/mosesdecoder.git
cd mosesdecoder
git clone git@github.com:rsennrich/wmt16-scripts.git
```
(2) define a function for post processing.
It removes diacritics and does other things I don't understand
```bash
ro_post_process () {
sys=$1
ref=$2
export MOSES_PATH=$HOME/mosesdecoder
REPLACE_UNICODE_PUNCT=$MOSES_PATH/scripts/tokenizer/replace-unicode-punctuation.perl
NORM_PUNC=$MOSES_PATH/scripts/tokenizer/normalize-punctuation.perl
REM_NON_PRINT_CHAR=$MOSES_PATH/scripts/tokenizer/remove-non-printing-char.perl
REMOVE_DIACRITICS=$MOSES_PATH/wmt16-scripts/preprocess/remove-diacritics.py
NORMALIZE_ROMANIAN=$MOSES_PATH/wmt16-scripts/preprocess/normalise-romanian.py
TOKENIZER=$MOSES_PATH/scripts/tokenizer/tokenizer.perl
lang=ro
for file in $sys $ref; do
cat $file \
| $REPLACE_UNICODE_PUNCT \
| $NORM_PUNC -l $lang \
| $REM_NON_PRINT_CHAR \
| $NORMALIZE_ROMANIAN \
| $REMOVE_DIACRITICS \
| $TOKENIZER -no-escape -l $lang \
> $(basename $file).tok
done
# compute BLEU
cat $(basename $sys).tok | sacrebleu -tok none -s none -b $(basename $ref).tok
}
```
(3) Call the function on test_generations.txt and test.target
For example,
```bash
ro_post_process enro_finetune/test_generations.txt wmt_en_ro/test.target
```
This will split out a new blue score and write a new fine called `test_generations.tok` with post-processed outputs.
```
| transformers/examples/legacy/seq2seq/romanian_postprocessing.md/0 | {
"file_path": "transformers/examples/legacy/seq2seq/romanian_postprocessing.md",
"repo_id": "transformers",
"token_count": 723
} | 411 |
import logging
import os
from typing import TextIO, Union
from conllu import parse_incr
from utils_ner import InputExample, Split, TokenClassificationTask
logger = logging.getLogger(__name__)
class NER(TokenClassificationTask):
def __init__(self, label_idx=-1):
# in NER datasets, the last column is usually reserved for NER label
self.label_idx = label_idx
def read_examples_from_file(self, data_dir, mode: Union[Split, str]) -> list[InputExample]:
if isinstance(mode, Split):
mode = mode.value
file_path = os.path.join(data_dir, f"{mode}.txt")
guid_index = 1
examples = []
with open(file_path, encoding="utf-8") as f:
words = []
labels = []
for line in f:
if line.startswith("-DOCSTART-") or line == "" or line == "\n":
if words:
examples.append(InputExample(guid=f"{mode}-{guid_index}", words=words, labels=labels))
guid_index += 1
words = []
labels = []
else:
splits = line.split(" ")
words.append(splits[0])
if len(splits) > 1:
labels.append(splits[self.label_idx].replace("\n", ""))
else:
# Examples could have no label for mode = "test"
labels.append("O")
if words:
examples.append(InputExample(guid=f"{mode}-{guid_index}", words=words, labels=labels))
return examples
def write_predictions_to_file(self, writer: TextIO, test_input_reader: TextIO, preds_list: list):
example_id = 0
for line in test_input_reader:
if line.startswith("-DOCSTART-") or line == "" or line == "\n":
writer.write(line)
if not preds_list[example_id]:
example_id += 1
elif preds_list[example_id]:
output_line = line.split()[0] + " " + preds_list[example_id].pop(0) + "\n"
writer.write(output_line)
else:
logger.warning("Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0])
def get_labels(self, path: str) -> list[str]:
if path:
with open(path) as f:
labels = f.read().splitlines()
if "O" not in labels:
labels = ["O"] + labels
return labels
else:
return ["O", "B-MISC", "I-MISC", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"]
class Chunk(NER):
def __init__(self):
# in CONLL2003 dataset chunk column is second-to-last
super().__init__(label_idx=-2)
def get_labels(self, path: str) -> list[str]:
if path:
with open(path) as f:
labels = f.read().splitlines()
if "O" not in labels:
labels = ["O"] + labels
return labels
else:
return [
"O",
"B-ADVP",
"B-INTJ",
"B-LST",
"B-PRT",
"B-NP",
"B-SBAR",
"B-VP",
"B-ADJP",
"B-CONJP",
"B-PP",
"I-ADVP",
"I-INTJ",
"I-LST",
"I-PRT",
"I-NP",
"I-SBAR",
"I-VP",
"I-ADJP",
"I-CONJP",
"I-PP",
]
class POS(TokenClassificationTask):
def read_examples_from_file(self, data_dir, mode: Union[Split, str]) -> list[InputExample]:
if isinstance(mode, Split):
mode = mode.value
file_path = os.path.join(data_dir, f"{mode}.txt")
guid_index = 1
examples = []
with open(file_path, encoding="utf-8") as f:
for sentence in parse_incr(f):
words = []
labels = []
for token in sentence:
words.append(token["form"])
labels.append(token["upos"])
assert len(words) == len(labels)
if words:
examples.append(InputExample(guid=f"{mode}-{guid_index}", words=words, labels=labels))
guid_index += 1
return examples
def write_predictions_to_file(self, writer: TextIO, test_input_reader: TextIO, preds_list: list):
example_id = 0
for sentence in parse_incr(test_input_reader):
s_p = preds_list[example_id]
out = ""
for token in sentence:
out += f"{token['form']} ({token['upos']}|{s_p.pop(0)}) "
out += "\n"
writer.write(out)
example_id += 1
def get_labels(self, path: str) -> list[str]:
if path:
with open(path) as f:
return f.read().splitlines()
else:
return [
"ADJ",
"ADP",
"ADV",
"AUX",
"CCONJ",
"DET",
"INTJ",
"NOUN",
"NUM",
"PART",
"PRON",
"PROPN",
"PUNCT",
"SCONJ",
"SYM",
"VERB",
"X",
]
| transformers/examples/legacy/token-classification/tasks.py/0 | {
"file_path": "transformers/examples/legacy/token-classification/tasks.py",
"repo_id": "transformers",
"token_count": 3152
} | 412 |
from transformers.models.llama.configuration_llama import LlamaConfig
class DuplicatedMethodConfig(LlamaConfig):
@property
def vocab_size(self):
return 45
@vocab_size.setter
def vocab_size(self, value):
self.vocab_size = value
| transformers/examples/modular-transformers/modular_duplicated_method.py/0 | {
"file_path": "transformers/examples/modular-transformers/modular_duplicated_method.py",
"repo_id": "transformers",
"token_count": 102
} | 413 |
<!---
Copyright 2021 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
-->
# Audio classification examples
The following examples showcase how to fine-tune `Wav2Vec2` for audio classification using PyTorch.
Speech recognition models that have been pretrained in unsupervised fashion on audio data alone,
*e.g.* [Wav2Vec2](https://huggingface.co/transformers/main/model_doc/wav2vec2.html),
[HuBERT](https://huggingface.co/transformers/main/model_doc/hubert.html),
[XLSR-Wav2Vec2](https://huggingface.co/transformers/main/model_doc/xlsr_wav2vec2.html), have shown to require only
very little annotated data to yield good performance on speech classification datasets.
## Single-GPU
The following command shows how to fine-tune [wav2vec2-base](https://huggingface.co/facebook/wav2vec2-base) on the 🗣️ [Keyword Spotting subset](https://huggingface.co/datasets/superb#ks) of the SUPERB dataset.
```bash
python run_audio_classification.py \
--model_name_or_path facebook/wav2vec2-base \
--dataset_name superb \
--dataset_config_name ks \
--output_dir wav2vec2-base-ft-keyword-spotting \
--overwrite_output_dir \
--remove_unused_columns False \
--do_train \
--do_eval \
--fp16 \
--learning_rate 3e-5 \
--max_length_seconds 1 \
--attention_mask False \
--warmup_ratio 0.1 \
--num_train_epochs 5 \
--per_device_train_batch_size 32 \
--gradient_accumulation_steps 4 \
--per_device_eval_batch_size 32 \
--dataloader_num_workers 4 \
--logging_strategy steps \
--logging_steps 10 \
--eval_strategy epoch \
--save_strategy epoch \
--load_best_model_at_end True \
--metric_for_best_model accuracy \
--save_total_limit 3 \
--seed 0 \
--push_to_hub
```
On a single V100 GPU (16GB), this script should run in ~14 minutes and yield accuracy of **98.26%**.
👀 See the results here: [anton-l/wav2vec2-base-ft-keyword-spotting](https://huggingface.co/anton-l/wav2vec2-base-ft-keyword-spotting)
> If your model classification head dimensions do not fit the number of labels in the dataset, you can specify `--ignore_mismatched_sizes` to adapt it.
## Multi-GPU
The following command shows how to fine-tune [wav2vec2-base](https://huggingface.co/facebook/wav2vec2-base) for 🌎 **Language Identification** on the [CommonLanguage dataset](https://huggingface.co/datasets/anton-l/common_language).
```bash
python run_audio_classification.py \
--model_name_or_path facebook/wav2vec2-base \
--dataset_name common_language \
--audio_column_name audio \
--label_column_name language \
--output_dir wav2vec2-base-lang-id \
--overwrite_output_dir \
--remove_unused_columns False \
--do_train \
--do_eval \
--fp16 \
--learning_rate 3e-4 \
--max_length_seconds 16 \
--attention_mask False \
--warmup_ratio 0.1 \
--num_train_epochs 10 \
--per_device_train_batch_size 8 \
--gradient_accumulation_steps 4 \
--per_device_eval_batch_size 1 \
--dataloader_num_workers 8 \
--logging_strategy steps \
--logging_steps 10 \
--eval_strategy epoch \
--save_strategy epoch \
--load_best_model_at_end True \
--metric_for_best_model accuracy \
--save_total_limit 3 \
--seed 0 \
--push_to_hub
```
On 4 V100 GPUs (16GB), this script should run in ~1 hour and yield accuracy of **79.45%**.
👀 See the results here: [anton-l/wav2vec2-base-lang-id](https://huggingface.co/anton-l/wav2vec2-base-lang-id)
## Sharing your model on 🤗 Hub
0. If you haven't already, [sign up](https://huggingface.co/join) for a 🤗 account
1. Make sure you have `git-lfs` installed and git set up.
```bash
$ apt install git-lfs
```
2. Log in with your HuggingFace account credentials using `hf`
```bash
$ hf auth login
# ...follow the prompts
```
3. When running the script, pass the following arguments:
```bash
python run_audio_classification.py \
--push_to_hub \
--hub_model_id <username/model_id> \
...
```
### Examples
The following table shows a couple of demonstration fine-tuning runs.
It has been verified that the script works for the following datasets:
- [SUPERB Keyword Spotting](https://huggingface.co/datasets/superb#ks)
- [Common Language](https://huggingface.co/datasets/common_language)
| Dataset | Pretrained Model | # transformer layers | Accuracy on eval | GPU setup | Training time | Fine-tuned Model & Logs |
|---------|------------------|----------------------|------------------|-----------|---------------|--------------------------|
| Keyword Spotting | [ntu-spml/distilhubert](https://huggingface.co/ntu-spml/distilhubert) | 2 | 0.9706 | 1 V100 GPU | 11min | [here](https://huggingface.co/anton-l/distilhubert-ft-keyword-spotting) |
| Keyword Spotting | [facebook/wav2vec2-base](https://huggingface.co/facebook/wav2vec2-base) | 12 | 0.9826 | 1 V100 GPU | 14min | [here](https://huggingface.co/anton-l/wav2vec2-base-ft-keyword-spotting) |
| Keyword Spotting | [facebook/hubert-base-ls960](https://huggingface.co/facebook/hubert-base-ls960) | 12 | 0.9819 | 1 V100 GPU | 14min | [here](https://huggingface.co/anton-l/hubert-base-ft-keyword-spotting) |
| Keyword Spotting | [asapp/sew-mid-100k](https://huggingface.co/asapp/sew-mid-100k) | 24 | 0.9757 | 1 V100 GPU | 15min | [here](https://huggingface.co/anton-l/sew-mid-100k-ft-keyword-spotting) |
| Common Language | [facebook/wav2vec2-base](https://huggingface.co/facebook/wav2vec2-base) | 12 | 0.7945 | 4 V100 GPUs | 1h10m | [here](https://huggingface.co/anton-l/wav2vec2-base-lang-id) |
| transformers/examples/pytorch/audio-classification/README.md/0 | {
"file_path": "transformers/examples/pytorch/audio-classification/README.md",
"repo_id": "transformers",
"token_count": 2205
} | 414 |
#!/usr/bin/env python
# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# /// script
# dependencies = [
# "transformers @ git+https://github.com/huggingface/transformers.git",
# "torch>=1.5.0",
# "torchvision>=0.6.0",
# "datasets>=1.8.0",
# ]
# ///
import logging
import os
import sys
from dataclasses import dataclass, field
from typing import Optional
import numpy as np
import torch
from datasets import load_dataset
from torchvision.transforms import Compose, Lambda, Normalize, RandomHorizontalFlip, RandomResizedCrop, ToTensor
import transformers
from transformers import (
CONFIG_MAPPING,
IMAGE_PROCESSOR_MAPPING,
MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING,
AutoConfig,
AutoImageProcessor,
AutoModelForMaskedImageModeling,
HfArgumentParser,
Trainer,
TrainingArguments,
)
from transformers.trainer_utils import get_last_checkpoint
from transformers.utils import check_min_version, send_example_telemetry
from transformers.utils.versions import require_version
""" Pre-training a 🤗 Transformers model for simple masked image modeling (SimMIM).
Any model supported by the AutoModelForMaskedImageModeling API can be used.
"""
logger = logging.getLogger(__name__)
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
check_min_version("4.56.0.dev0")
require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/image-pretraining/requirements.txt")
MODEL_CONFIG_CLASSES = list(MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING.keys())
MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
@dataclass
class DataTrainingArguments:
"""
Arguments pertaining to what data we are going to input our model for training and eval.
Using `HfArgumentParser` we can turn this class into argparse arguments to be able to
specify them on the command line.
"""
dataset_name: Optional[str] = field(
default="cifar10", metadata={"help": "Name of a dataset from the datasets package"}
)
dataset_config_name: Optional[str] = field(
default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
)
image_column_name: Optional[str] = field(
default=None,
metadata={"help": "The column name of the images in the files. If not set, will try to use 'image' or 'img'."},
)
train_dir: Optional[str] = field(default=None, metadata={"help": "A folder containing the training data."})
validation_dir: Optional[str] = field(default=None, metadata={"help": "A folder containing the validation data."})
train_val_split: Optional[float] = field(
default=0.15, metadata={"help": "Percent to split off of train for validation."}
)
mask_patch_size: int = field(default=32, metadata={"help": "The size of the square patches to use for masking."})
mask_ratio: float = field(
default=0.6,
metadata={"help": "Percentage of patches to mask."},
)
max_train_samples: Optional[int] = field(
default=None,
metadata={
"help": (
"For debugging purposes or quicker training, truncate the number of training examples to this "
"value if set."
)
},
)
max_eval_samples: Optional[int] = field(
default=None,
metadata={
"help": (
"For debugging purposes or quicker training, truncate the number of evaluation examples to this "
"value if set."
)
},
)
def __post_init__(self):
data_files = {}
if self.train_dir is not None:
data_files["train"] = self.train_dir
if self.validation_dir is not None:
data_files["val"] = self.validation_dir
self.data_files = data_files if data_files else None
@dataclass
class ModelArguments:
"""
Arguments pertaining to which model/config/image processor we are going to pre-train.
"""
model_name_or_path: str = field(
default=None,
metadata={
"help": (
"The model checkpoint for weights initialization. Can be a local path to a pytorch_model.bin or a "
"checkpoint identifier on the hub. "
"Don't set if you want to train a model from scratch."
)
},
)
model_type: Optional[str] = field(
default=None,
metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
)
config_name_or_path: Optional[str] = field(
default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
)
config_overrides: Optional[str] = field(
default=None,
metadata={
"help": (
"Override some existing default config settings when a model is trained from scratch. Example: "
"n_embd=10,resid_pdrop=0.2,scale_attn_weights=false,summary_type=cls_index"
)
},
)
cache_dir: Optional[str] = field(
default=None,
metadata={"help": "Where do you want to store (cache) the pretrained models/datasets downloaded from the hub"},
)
model_revision: str = field(
default="main",
metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
)
image_processor_name: str = field(default=None, metadata={"help": "Name or path of preprocessor config."})
token: str = field(
default=None,
metadata={
"help": (
"The token to use as HTTP bearer authorization for remote files. If not specified, will use the token "
"generated when running `hf auth login` (stored in `~/.huggingface`)."
)
},
)
trust_remote_code: bool = field(
default=False,
metadata={
"help": (
"Whether to trust the execution of code from datasets/models defined on the Hub."
" This option should only be set to `True` for repositories you trust and in which you have read the"
" code, as it will execute code present on the Hub on your local machine."
)
},
)
image_size: Optional[int] = field(
default=None,
metadata={
"help": (
"The size (resolution) of each image. If not specified, will use `image_size` of the configuration."
)
},
)
patch_size: Optional[int] = field(
default=None,
metadata={
"help": (
"The size (resolution) of each patch. If not specified, will use `patch_size` of the configuration."
)
},
)
encoder_stride: Optional[int] = field(
default=None,
metadata={"help": "Stride to use for the encoder."},
)
class MaskGenerator:
"""
A class to generate boolean masks for the pretraining task.
A mask is a 1D tensor of shape (model_patch_size**2,) where the value is either 0 or 1,
where 1 indicates "masked".
"""
def __init__(self, input_size=192, mask_patch_size=32, model_patch_size=4, mask_ratio=0.6):
self.input_size = input_size
self.mask_patch_size = mask_patch_size
self.model_patch_size = model_patch_size
self.mask_ratio = mask_ratio
if self.input_size % self.mask_patch_size != 0:
raise ValueError("Input size must be divisible by mask patch size")
if self.mask_patch_size % self.model_patch_size != 0:
raise ValueError("Mask patch size must be divisible by model patch size")
self.rand_size = self.input_size // self.mask_patch_size
self.scale = self.mask_patch_size // self.model_patch_size
self.token_count = self.rand_size**2
self.mask_count = int(np.ceil(self.token_count * self.mask_ratio))
def __call__(self):
mask_idx = np.random.permutation(self.token_count)[: self.mask_count]
mask = np.zeros(self.token_count, dtype=int)
mask[mask_idx] = 1
mask = mask.reshape((self.rand_size, self.rand_size))
mask = mask.repeat(self.scale, axis=0).repeat(self.scale, axis=1)
return torch.tensor(mask.flatten())
def collate_fn(examples):
pixel_values = torch.stack([example["pixel_values"] for example in examples])
mask = torch.stack([example["mask"] for example in examples])
return {"pixel_values": pixel_values, "bool_masked_pos": mask}
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
# information sent is the one passed as arguments along with your Python/PyTorch versions.
send_example_telemetry("run_mim", model_args, data_args)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
if training_args.should_log:
# The default of training_args.log_level is passive, so we set log level at info here to have that default.
transformers.utils.logging.set_verbosity_info()
log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
transformers.utils.logging.set_verbosity(log_level)
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}, "
+ f"distributed training: {training_args.parallel_mode.value == 'distributed'}, 16-bits training: {training_args.fp16}"
)
logger.info(f"Training/evaluation parameters {training_args}")
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome."
)
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Initialize our dataset.
ds = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
data_files=data_args.data_files,
cache_dir=model_args.cache_dir,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
# If we don't have a validation split, split off a percentage of train as validation.
data_args.train_val_split = None if "validation" in ds else data_args.train_val_split
if isinstance(data_args.train_val_split, float) and data_args.train_val_split > 0.0:
split = ds["train"].train_test_split(data_args.train_val_split)
ds["train"] = split["train"]
ds["validation"] = split["test"]
# Create config
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config_kwargs = {
"cache_dir": model_args.cache_dir,
"revision": model_args.model_revision,
"token": model_args.token,
"trust_remote_code": model_args.trust_remote_code,
}
if model_args.config_name_or_path:
config = AutoConfig.from_pretrained(model_args.config_name_or_path, **config_kwargs)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning("You are instantiating a new config instance from scratch.")
if model_args.config_overrides is not None:
logger.info(f"Overriding config: {model_args.config_overrides}")
config.update_from_string(model_args.config_overrides)
logger.info(f"New config: {config}")
# make sure the decoder_type is "simmim" (only relevant for BEiT)
if hasattr(config, "decoder_type"):
config.decoder_type = "simmim"
# adapt config
model_args.image_size = model_args.image_size if model_args.image_size is not None else config.image_size
model_args.patch_size = model_args.patch_size if model_args.patch_size is not None else config.patch_size
model_args.encoder_stride = (
model_args.encoder_stride if model_args.encoder_stride is not None else config.encoder_stride
)
config.update(
{
"image_size": model_args.image_size,
"patch_size": model_args.patch_size,
"encoder_stride": model_args.encoder_stride,
}
)
# create image processor
if model_args.image_processor_name:
image_processor = AutoImageProcessor.from_pretrained(model_args.image_processor_name, **config_kwargs)
elif model_args.model_name_or_path:
image_processor = AutoImageProcessor.from_pretrained(model_args.model_name_or_path, **config_kwargs)
else:
IMAGE_PROCESSOR_TYPES = {
conf.model_type: image_processor_class for conf, image_processor_class in IMAGE_PROCESSOR_MAPPING.items()
}
image_processor = IMAGE_PROCESSOR_TYPES[model_args.model_type][-1]()
# create model
if model_args.model_name_or_path:
model = AutoModelForMaskedImageModeling.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForMaskedImageModeling.from_config(config, trust_remote_code=model_args.trust_remote_code)
if training_args.do_train:
column_names = ds["train"].column_names
else:
column_names = ds["validation"].column_names
if data_args.image_column_name is not None:
image_column_name = data_args.image_column_name
elif "image" in column_names:
image_column_name = "image"
elif "img" in column_names:
image_column_name = "img"
else:
image_column_name = column_names[0]
# transformations as done in original SimMIM paper
# source: https://github.com/microsoft/SimMIM/blob/main/data/data_simmim.py
transforms = Compose(
[
Lambda(lambda img: img.convert("RGB") if img.mode != "RGB" else img),
RandomResizedCrop(model_args.image_size, scale=(0.67, 1.0), ratio=(3.0 / 4.0, 4.0 / 3.0)),
RandomHorizontalFlip(),
ToTensor(),
Normalize(mean=image_processor.image_mean, std=image_processor.image_std),
]
)
# create mask generator
mask_generator = MaskGenerator(
input_size=model_args.image_size,
mask_patch_size=data_args.mask_patch_size,
model_patch_size=model_args.patch_size,
mask_ratio=data_args.mask_ratio,
)
def preprocess_images(examples):
"""Preprocess a batch of images by applying transforms + creating a corresponding mask, indicating
which patches to mask."""
examples["pixel_values"] = [transforms(image) for image in examples[image_column_name]]
examples["mask"] = [mask_generator() for i in range(len(examples[image_column_name]))]
return examples
if training_args.do_train:
if "train" not in ds:
raise ValueError("--do_train requires a train dataset")
if data_args.max_train_samples is not None:
ds["train"] = ds["train"].shuffle(seed=training_args.seed).select(range(data_args.max_train_samples))
# Set the training transforms
ds["train"].set_transform(preprocess_images)
if training_args.do_eval:
if "validation" not in ds:
raise ValueError("--do_eval requires a validation dataset")
if data_args.max_eval_samples is not None:
ds["validation"] = (
ds["validation"].shuffle(seed=training_args.seed).select(range(data_args.max_eval_samples))
)
# Set the validation transforms
ds["validation"].set_transform(preprocess_images)
# Initialize our trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=ds["train"] if training_args.do_train else None,
eval_dataset=ds["validation"] if training_args.do_eval else None,
processing_class=image_processor,
data_collator=collate_fn,
)
# Training
if training_args.do_train:
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif last_checkpoint is not None:
checkpoint = last_checkpoint
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model()
trainer.log_metrics("train", train_result.metrics)
trainer.save_metrics("train", train_result.metrics)
trainer.save_state()
# Evaluation
if training_args.do_eval:
metrics = trainer.evaluate()
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Write model card and (optionally) push to hub
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "masked-image-modeling",
"dataset": data_args.dataset_name,
"tags": ["masked-image-modeling"],
}
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
if __name__ == "__main__":
main()
| transformers/examples/pytorch/image-pretraining/run_mim.py/0 | {
"file_path": "transformers/examples/pytorch/image-pretraining/run_mim.py",
"repo_id": "transformers",
"token_count": 7864
} | 415 |
# Copyright 2020 The HuggingFace Team All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
A subclass of `Trainer` specific to Question-Answering tasks
"""
import math
import time
from transformers import Trainer, is_torch_xla_available
from transformers.trainer_utils import PredictionOutput, speed_metrics
if is_torch_xla_available():
import torch_xla.core.xla_model as xm
import torch_xla.debug.metrics as met
class QuestionAnsweringTrainer(Trainer):
def __init__(self, *args, eval_examples=None, post_process_function=None, **kwargs):
super().__init__(*args, **kwargs)
self.eval_examples = eval_examples
self.post_process_function = post_process_function
def evaluate(self, eval_dataset=None, eval_examples=None, ignore_keys=None, metric_key_prefix: str = "eval"):
eval_dataset = self.eval_dataset if eval_dataset is None else eval_dataset
eval_dataloader = self.get_eval_dataloader(eval_dataset)
eval_examples = self.eval_examples if eval_examples is None else eval_examples
# Temporarily disable metric computation, we will do it in the loop here.
compute_metrics = self.compute_metrics
self.compute_metrics = None
eval_loop = self.prediction_loop if self.args.use_legacy_prediction_loop else self.evaluation_loop
start_time = time.time()
try:
output = eval_loop(
eval_dataloader,
description="Evaluation",
# No point gathering the predictions if there are no metrics, otherwise we defer to
# self.args.prediction_loss_only
prediction_loss_only=True if compute_metrics is None else None,
ignore_keys=ignore_keys,
metric_key_prefix=metric_key_prefix,
)
finally:
self.compute_metrics = compute_metrics
total_batch_size = self.args.eval_batch_size * self.args.world_size
if f"{metric_key_prefix}_jit_compilation_time" in output.metrics:
start_time += output.metrics[f"{metric_key_prefix}_jit_compilation_time"]
output.metrics.update(
speed_metrics(
metric_key_prefix,
start_time,
num_samples=output.num_samples,
num_steps=math.ceil(output.num_samples / total_batch_size),
)
)
if self.post_process_function is not None and self.compute_metrics is not None and self.args.should_save:
# Only the main node write the results by default
eval_preds = self.post_process_function(eval_examples, eval_dataset, output.predictions)
metrics = self.compute_metrics(eval_preds)
# Prefix all keys with metric_key_prefix + '_'
for key in list(metrics.keys()):
if not key.startswith(f"{metric_key_prefix}_"):
metrics[f"{metric_key_prefix}_{key}"] = metrics.pop(key)
metrics.update(output.metrics)
else:
metrics = output.metrics
if self.args.should_log:
# Only the main node log the results by default
self.log(metrics)
if self.args.tpu_metrics_debug or self.args.debug:
# tpu-comment: Logging debug metrics for PyTorch/XLA (compile, execute times, ops, etc.)
xm.master_print(met.metrics_report())
self.control = self.callback_handler.on_evaluate(self.args, self.state, self.control, metrics)
return metrics
def predict(self, predict_dataset, predict_examples, ignore_keys=None, metric_key_prefix: str = "test"):
predict_dataloader = self.get_test_dataloader(predict_dataset)
# Temporarily disable metric computation, we will do it in the loop here.
compute_metrics = self.compute_metrics
self.compute_metrics = None
eval_loop = self.prediction_loop if self.args.use_legacy_prediction_loop else self.evaluation_loop
start_time = time.time()
try:
output = eval_loop(
predict_dataloader,
description="Prediction",
# No point gathering the predictions if there are no metrics, otherwise we defer to
# self.args.prediction_loss_only
prediction_loss_only=True if compute_metrics is None else None,
ignore_keys=ignore_keys,
metric_key_prefix=metric_key_prefix,
)
finally:
self.compute_metrics = compute_metrics
total_batch_size = self.args.eval_batch_size * self.args.world_size
if f"{metric_key_prefix}_jit_compilation_time" in output.metrics:
start_time += output.metrics[f"{metric_key_prefix}_jit_compilation_time"]
output.metrics.update(
speed_metrics(
metric_key_prefix,
start_time,
num_samples=output.num_samples,
num_steps=math.ceil(output.num_samples / total_batch_size),
)
)
if self.post_process_function is None or self.compute_metrics is None:
return output
predictions = self.post_process_function(predict_examples, predict_dataset, output.predictions, "predict")
metrics = self.compute_metrics(predictions)
# Prefix all keys with metric_key_prefix + '_'
for key in list(metrics.keys()):
if not key.startswith(f"{metric_key_prefix}_"):
metrics[f"{metric_key_prefix}_{key}"] = metrics.pop(key)
metrics.update(output.metrics)
return PredictionOutput(predictions=predictions.predictions, label_ids=predictions.label_ids, metrics=metrics)
| transformers/examples/pytorch/question-answering/trainer_qa.py/0 | {
"file_path": "transformers/examples/pytorch/question-answering/trainer_qa.py",
"repo_id": "transformers",
"token_count": 2650
} | 416 |
<!---
Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
-->
# TFVisionTextDualEncoder and CLIP model training examples
The following example showcases how to train a CLIP-like vision-text dual encoder model
using a pre-trained vision and text encoder.
Such a model can be used for natural language image search and potentially zero-shot image classification.
The model is inspired by [CLIP](https://openai.com/blog/clip/), introduced by Alec Radford et al.
The idea is to train a vision encoder and a text encoder jointly to project the representation of images and their
captions into the same embedding space, such that the caption embeddings are located near the embeddings
of the images they describe.
### Download COCO dataset (2017)
This example uses COCO dataset (2017) through a custom dataset script, which requires users to manually download the
COCO dataset before training.
```bash
mkdir data
cd data
wget http://images.cocodataset.org/zips/train2017.zip
wget http://images.cocodataset.org/zips/val2017.zip
wget http://images.cocodataset.org/zips/test2017.zip
wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
wget http://images.cocodataset.org/annotations/image_info_test2017.zip
cd ..
```
Having downloaded COCO dataset manually you should be able to load with the `ydshieh/coc_dataset_script` dataset loading script:
```py
import os
import datasets
COCO_DIR = os.path.join(os.getcwd(), "data")
ds = datasets.load_dataset("ydshieh/coco_dataset_script", "2017", data_dir=COCO_DIR)
```
### Create a model from a vision encoder model and a text encoder model
We can either load a CLIP-like vision-text dual encoder model from an existing dual encoder model, or
by using a pre-trained vision encoder model and a pre-trained text encoder model.
If you wish to load an existing dual encoder model, please use the `--model_name_or_path` argument. If
you want to use separate pre-trained vision and text models, please use the
`--vision_model_name_or_path` and `--text_model_name_or_path` arguments instead.
### Train the model
Finally, we can run the example script to train the model:
```bash
python examples/tensorflow/contrastive-image-text/run_clip.py \
--output_dir ./clip-roberta-finetuned \
--vision_model_name_or_path openai/clip-vit-base-patch32 \
--text_model_name_or_path FacebookAI/roberta-base \
--data_dir $PWD/data \
--dataset_name ydshieh/coco_dataset_script \
--dataset_config_name=2017 \
--image_column image_path \
--caption_column caption \
--remove_unused_columns=False \
--do_train --do_eval \
--per_device_train_batch_size="64" \
--per_device_eval_batch_size="64" \
--learning_rate="5e-5" --warmup_steps="0" --weight_decay 0.1 \
--overwrite_output_dir \
--push_to_hub
```
| transformers/examples/tensorflow/contrastive-image-text/README.md/0 | {
"file_path": "transformers/examples/tensorflow/contrastive-image-text/README.md",
"repo_id": "transformers",
"token_count": 1057
} | 417 |
#!/usr/bin/env python
# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Fine-tuning a 🤗 Transformers model on token classification tasks (NER, POS, CHUNKS)
"""
import json
import logging
import os
import random
from dataclasses import dataclass, field
from typing import Optional
import datasets
import evaluate
import tensorflow as tf
from datasets import ClassLabel, load_dataset
import transformers
from transformers import (
CONFIG_MAPPING,
AutoConfig,
AutoTokenizer,
DataCollatorForTokenClassification,
HfArgumentParser,
PushToHubCallback,
TFAutoModelForTokenClassification,
TFTrainingArguments,
create_optimizer,
set_seed,
)
from transformers.utils import send_example_telemetry
from transformers.utils.versions import require_version
logger = logging.getLogger(__name__)
logger.addHandler(logging.StreamHandler())
require_version("datasets>=1.8.0", "To fix: pip install -r examples/tensorflow/token-classification/requirements.txt")
# region Command-line arguments
@dataclass
class ModelArguments:
"""
Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
"""
model_name_or_path: str = field(
metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
)
config_name: Optional[str] = field(
default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
)
tokenizer_name: Optional[str] = field(
default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
)
cache_dir: Optional[str] = field(
default=None,
metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
)
model_revision: str = field(
default="main",
metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
)
token: str = field(
default=None,
metadata={
"help": (
"The token to use as HTTP bearer authorization for remote files. If not specified, will use the token "
"generated when running `hf auth login` (stored in `~/.huggingface`)."
)
},
)
trust_remote_code: bool = field(
default=False,
metadata={
"help": (
"Whether to trust the execution of code from datasets/models defined on the Hub."
" This option should only be set to `True` for repositories you trust and in which you have read the"
" code, as it will execute code present on the Hub on your local machine."
)
},
)
@dataclass
class DataTrainingArguments:
"""
Arguments pertaining to what data we are going to input our model for training and eval.
"""
task_name: Optional[str] = field(default="ner", metadata={"help": "The name of the task (ner, pos...)."})
dataset_name: Optional[str] = field(
default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
)
dataset_config_name: Optional[str] = field(
default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
)
train_file: Optional[str] = field(
default=None, metadata={"help": "The input training data file (a csv or JSON file)."}
)
validation_file: Optional[str] = field(
default=None,
metadata={"help": "An optional input evaluation data file to evaluate on (a csv or JSON file)."},
)
test_file: Optional[str] = field(
default=None,
metadata={"help": "An optional input test data file to predict on (a csv or JSON file)."},
)
text_column_name: Optional[str] = field(
default=None, metadata={"help": "The column name of text to input in the file (a csv or JSON file)."}
)
label_column_name: Optional[str] = field(
default=None, metadata={"help": "The column name of label to input in the file (a csv or JSON file)."}
)
overwrite_cache: bool = field(
default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
)
preprocessing_num_workers: Optional[int] = field(
default=None,
metadata={"help": "The number of processes to use for the preprocessing."},
)
max_length: Optional[int] = field(default=256, metadata={"help": "Max length (in tokens) for truncation/padding"})
pad_to_max_length: bool = field(
default=False,
metadata={
"help": (
"Whether to pad all samples to model maximum sentence length. "
"If False, will pad the samples dynamically when batching to the maximum length in the batch. More "
"efficient on GPU but very bad for TPU."
)
},
)
max_train_samples: Optional[int] = field(
default=None,
metadata={
"help": (
"For debugging purposes or quicker training, truncate the number of training examples to this "
"value if set."
)
},
)
max_eval_samples: Optional[int] = field(
default=None,
metadata={
"help": (
"For debugging purposes or quicker training, truncate the number of evaluation examples to this "
"value if set."
)
},
)
max_predict_samples: Optional[int] = field(
default=None,
metadata={
"help": (
"For debugging purposes or quicker training, truncate the number of prediction examples to this "
"value if set."
)
},
)
label_all_tokens: bool = field(
default=False,
metadata={
"help": (
"Whether to put the label for one word on all tokens of generated by that word or just on the "
"one (in which case the other tokens will have a padding index)."
)
},
)
return_entity_level_metrics: bool = field(
default=False,
metadata={"help": "Whether to return all the entity levels during evaluation or just the overall ones."},
)
def __post_init__(self):
if self.dataset_name is None and self.train_file is None and self.validation_file is None:
raise ValueError("Need either a dataset name or a training/validation file.")
else:
if self.train_file is not None:
extension = self.train_file.split(".")[-1]
assert extension in ["csv", "json"], "`train_file` should be a csv or a json file."
if self.validation_file is not None:
extension = self.validation_file.split(".")[-1]
assert extension in ["csv", "json"], "`validation_file` should be a csv or a json file."
self.task_name = self.task_name.lower()
# endregion
def main():
# region Argument Parsing
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TFTrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
# information sent is the one passed as arguments along with your Python/PyTorch versions.
send_example_telemetry("run_ner", model_args, data_args, framework="tensorflow")
# endregion
# region Setup logging
# we only want one process per machine to log things on the screen.
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(logging.INFO)
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
# If passed along, set the training seed now.
if training_args.seed is not None:
set_seed(training_args.seed)
# endregion
# region Loading datasets
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets for token classification task available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'tokens' or the first column if no column called
# 'tokens' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
raw_datasets = load_dataset(
extension,
data_files=data_files,
token=model_args.token,
)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.
if raw_datasets["train"] is not None:
column_names = raw_datasets["train"].column_names
features = raw_datasets["train"].features
else:
column_names = raw_datasets["validation"].column_names
features = raw_datasets["validation"].features
if data_args.text_column_name is not None:
text_column_name = data_args.text_column_name
elif "tokens" in column_names:
text_column_name = "tokens"
else:
text_column_name = column_names[0]
if data_args.label_column_name is not None:
label_column_name = data_args.label_column_name
elif f"{data_args.task_name}_tags" in column_names:
label_column_name = f"{data_args.task_name}_tags"
else:
label_column_name = column_names[1]
# In the event the labels are not a `Sequence[ClassLabel]`, we will need to go through the dataset to get the
# unique labels.
def get_label_list(labels):
unique_labels = set()
for label in labels:
unique_labels = unique_labels | set(label)
label_list = list(unique_labels)
label_list.sort()
return label_list
if isinstance(features[label_column_name].feature, ClassLabel):
label_list = features[label_column_name].feature.names
# No need to convert the labels since they are already ints.
label_to_id = {i: i for i in range(len(label_list))}
else:
label_list = get_label_list(raw_datasets["train"][label_column_name])
label_to_id = {l: i for i, l in enumerate(label_list)}
num_labels = len(label_list)
# endregion
# region Load config and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if model_args.config_name:
config = AutoConfig.from_pretrained(
model_args.config_name,
num_labels=num_labels,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(
model_args.model_name_or_path,
num_labels=num_labels,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning("You are instantiating a new config instance from scratch.")
tokenizer_name_or_path = model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path
if not tokenizer_name_or_path:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script. "
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
)
if config.model_type in {"gpt2", "roberta"}:
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name_or_path,
use_fast=True,
add_prefix_space=True,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
else:
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name_or_path,
use_fast=True,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
# endregion
# region Preprocessing the raw datasets
# First we tokenize all the texts.
padding = "max_length" if data_args.pad_to_max_length else False
# Tokenize all texts and align the labels with them.
def tokenize_and_align_labels(examples):
tokenized_inputs = tokenizer(
examples[text_column_name],
max_length=data_args.max_length,
padding=padding,
truncation=True,
# We use this argument because the texts in our dataset are lists of words (with a label for each word).
is_split_into_words=True,
)
labels = []
for i, label in enumerate(examples[label_column_name]):
word_ids = tokenized_inputs.word_ids(batch_index=i)
previous_word_idx = None
label_ids = []
for word_idx in word_ids:
# Special tokens have a word id that is None. We set the label to -100 so they are automatically
# ignored in the loss function.
if word_idx is None:
label_ids.append(-100)
# We set the label for the first token of each word.
elif word_idx != previous_word_idx:
label_ids.append(label_to_id[label[word_idx]])
# For the other tokens in a word, we set the label to either the current label or -100, depending on
# the label_all_tokens flag.
else:
label_ids.append(label_to_id[label[word_idx]] if data_args.label_all_tokens else -100)
previous_word_idx = word_idx
labels.append(label_ids)
tokenized_inputs["labels"] = labels
return tokenized_inputs
processed_raw_datasets = raw_datasets.map(
tokenize_and_align_labels,
batched=True,
remove_columns=raw_datasets["train"].column_names,
desc="Running tokenizer on dataset",
)
train_dataset = processed_raw_datasets["train"]
eval_dataset = processed_raw_datasets["validation"]
if data_args.max_train_samples is not None:
max_train_samples = min(len(train_dataset), data_args.max_train_samples)
train_dataset = train_dataset.select(range(max_train_samples))
if data_args.max_eval_samples is not None:
max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
eval_dataset = eval_dataset.select(range(max_eval_samples))
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")
# endregion
with training_args.strategy.scope():
# region Initialize model
if model_args.model_name_or_path:
model = TFAutoModelForTokenClassification.from_pretrained(
model_args.model_name_or_path,
config=config,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
else:
logger.info("Training new model from scratch")
model = TFAutoModelForTokenClassification.from_config(
config, token=model_args.token, trust_remote_code=model_args.trust_remote_code
)
# We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
# on a small vocab and want a smaller embedding size, remove this test.
embeddings = model.get_input_embeddings()
# Matt: This is a temporary workaround as we transition our models to exclusively using Keras embeddings.
# As soon as the transition is complete, all embeddings should be keras.Embeddings layers, and
# the weights will always be in embeddings.embeddings.
if hasattr(embeddings, "embeddings"):
embedding_size = embeddings.embeddings.shape[0]
else:
embedding_size = embeddings.weight.shape[0]
if len(tokenizer) > embedding_size:
model.resize_token_embeddings(len(tokenizer))
# endregion
# region Create TF datasets
# We need the DataCollatorForTokenClassification here, as we need to correctly pad labels as
# well as inputs.
collate_fn = DataCollatorForTokenClassification(tokenizer=tokenizer, return_tensors="np")
num_replicas = training_args.strategy.num_replicas_in_sync
total_train_batch_size = training_args.per_device_train_batch_size * num_replicas
dataset_options = tf.data.Options()
dataset_options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.OFF
# model.prepare_tf_dataset() wraps a Hugging Face dataset in a tf.data.Dataset which is ready to use in
# training. This is the recommended way to use a Hugging Face dataset when training with Keras. You can also
# use the lower-level dataset.to_tf_dataset() method, but you will have to specify things like column names
# yourself if you use this method, whereas they are automatically inferred from the model input names when
# using model.prepare_tf_dataset()
# For more info see the docs:
# https://huggingface.co/docs/transformers/main/en/main_classes/model#transformers.TFPreTrainedModel.prepare_tf_dataset
# https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset.to_tf_dataset
tf_train_dataset = model.prepare_tf_dataset(
train_dataset,
collate_fn=collate_fn,
batch_size=total_train_batch_size,
shuffle=True,
).with_options(dataset_options)
total_eval_batch_size = training_args.per_device_eval_batch_size * num_replicas
tf_eval_dataset = model.prepare_tf_dataset(
eval_dataset,
collate_fn=collate_fn,
batch_size=total_eval_batch_size,
shuffle=False,
).with_options(dataset_options)
# endregion
# region Optimizer, loss and compilation
num_train_steps = int(len(tf_train_dataset) * training_args.num_train_epochs)
if training_args.warmup_steps > 0:
num_warmup_steps = training_args.warmup_steps
elif training_args.warmup_ratio > 0:
num_warmup_steps = int(num_train_steps * training_args.warmup_ratio)
else:
num_warmup_steps = 0
optimizer, lr_schedule = create_optimizer(
init_lr=training_args.learning_rate,
num_train_steps=num_train_steps,
num_warmup_steps=num_warmup_steps,
adam_beta1=training_args.adam_beta1,
adam_beta2=training_args.adam_beta2,
adam_epsilon=training_args.adam_epsilon,
weight_decay_rate=training_args.weight_decay,
adam_global_clipnorm=training_args.max_grad_norm,
)
# Transformers models compute the right loss for their task by default when labels are passed, and will
# use this for training unless you specify your own loss function in compile().
model.compile(optimizer=optimizer, jit_compile=training_args.xla)
# endregion
# Metrics
metric = evaluate.load("seqeval", cache_dir=model_args.cache_dir)
def get_labels(y_pred, y_true):
# Transform predictions and references tensos to numpy arrays
# Remove ignored index (special tokens)
true_predictions = [
[label_list[p] for (p, l) in zip(pred, gold_label) if l != -100]
for pred, gold_label in zip(y_pred, y_true)
]
true_labels = [
[label_list[l] for (p, l) in zip(pred, gold_label) if l != -100]
for pred, gold_label in zip(y_pred, y_true)
]
return true_predictions, true_labels
def compute_metrics():
results = metric.compute()
if data_args.return_entity_level_metrics:
# Unpack nested dictionaries
final_results = {}
for key, value in results.items():
if isinstance(value, dict):
for n, v in value.items():
final_results[f"{key}_{n}"] = v
else:
final_results[key] = value
return final_results
else:
return {
"precision": results["overall_precision"],
"recall": results["overall_recall"],
"f1": results["overall_f1"],
"accuracy": results["overall_accuracy"],
}
# endregion
# region Preparing push_to_hub and model card
push_to_hub_model_id = training_args.push_to_hub_model_id
model_name = model_args.model_name_or_path.split("/")[-1]
if not push_to_hub_model_id:
if data_args.dataset_name is not None:
push_to_hub_model_id = f"{model_name}-finetuned-{data_args.dataset_name}"
else:
push_to_hub_model_id = f"{model_name}-finetuned-token-classification"
model_card_kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "token-classification"}
if data_args.dataset_name is not None:
model_card_kwargs["dataset_tags"] = data_args.dataset_name
if data_args.dataset_config_name is not None:
model_card_kwargs["dataset_args"] = data_args.dataset_config_name
model_card_kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
else:
model_card_kwargs["dataset"] = data_args.dataset_name
if training_args.push_to_hub:
callbacks = [
PushToHubCallback(
output_dir=training_args.output_dir,
hub_model_id=push_to_hub_model_id,
hub_token=training_args.push_to_hub_token,
tokenizer=tokenizer,
**model_card_kwargs,
)
]
else:
callbacks = []
# endregion
# region Training
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {training_args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
logger.info(f" Total train batch size = {total_train_batch_size}")
# Only show the progress bar once on each machine.
model.fit(
tf_train_dataset,
validation_data=tf_eval_dataset,
epochs=int(training_args.num_train_epochs),
callbacks=callbacks,
)
# endregion
# region Predictions
# If you have variable batch sizes (i.e. not using pad_to_max_length), then
# this bit might fail on TF < 2.8 because TF can't concatenate outputs of varying seq
# length from predict().
try:
predictions = model.predict(tf_eval_dataset, batch_size=training_args.per_device_eval_batch_size)["logits"]
except tf.python.framework.errors_impl.InvalidArgumentError:
raise ValueError(
"Concatenating predictions failed! If your version of TensorFlow is 2.8.0 or older "
"then you will need to use --pad_to_max_length to generate predictions, as older "
"versions of TensorFlow cannot concatenate variable-length predictions as RaggedTensor."
)
if isinstance(predictions, tf.RaggedTensor):
predictions = predictions.to_tensor(default_value=-100)
predictions = tf.math.argmax(predictions, axis=-1).numpy()
if "label" in eval_dataset:
labels = eval_dataset.with_format("tf")["label"]
else:
labels = eval_dataset.with_format("tf")["labels"]
if isinstance(labels, tf.RaggedTensor):
labels = labels.to_tensor(default_value=-100)
labels = labels.numpy()
attention_mask = eval_dataset.with_format("tf")["attention_mask"]
if isinstance(attention_mask, tf.RaggedTensor):
attention_mask = attention_mask.to_tensor(default_value=-100)
attention_mask = attention_mask.numpy()
labels[attention_mask == 0] = -100
preds, refs = get_labels(predictions, labels)
metric.add_batch(
predictions=preds,
references=refs,
)
eval_metric = compute_metrics()
logger.info("Evaluation metrics:")
for key, val in eval_metric.items():
logger.info(f"{key}: {val:.4f}")
if training_args.output_dir is not None:
output_eval_file = os.path.join(training_args.output_dir, "all_results.json")
with open(output_eval_file, "w") as writer:
writer.write(json.dumps(eval_metric))
# endregion
if training_args.output_dir is not None and not training_args.push_to_hub:
# If we're not pushing to hub, at least save a local copy when we're done
model.save_pretrained(training_args.output_dir)
if __name__ == "__main__":
main()
| transformers/examples/tensorflow/token-classification/run_ner.py/0 | {
"file_path": "transformers/examples/tensorflow/token-classification/run_ner.py",
"repo_id": "transformers",
"token_count": 11430
} | 418 |
# Copyright 2021 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import difflib
import os
import re
import subprocess
import textwrap
from argparse import ArgumentParser, Namespace
from datetime import date
from pathlib import Path
from typing import Any, Callable, Optional, Union
from ..models.auto.configuration_auto import CONFIG_MAPPING_NAMES, MODEL_NAMES_MAPPING
from ..models.auto.feature_extraction_auto import FEATURE_EXTRACTOR_MAPPING_NAMES
from ..models.auto.image_processing_auto import IMAGE_PROCESSOR_MAPPING_NAMES
from ..models.auto.processing_auto import PROCESSOR_MAPPING_NAMES
from ..models.auto.tokenization_auto import TOKENIZER_MAPPING_NAMES
from ..models.auto.video_processing_auto import VIDEO_PROCESSOR_MAPPING_NAMES
from ..utils import is_libcst_available
from . import BaseTransformersCLICommand
from .add_fast_image_processor import add_fast_image_processor
# We protect this import to avoid requiring it for all `transformers` CLI commands - however it is actually
# strictly required for this one (we need it both for modular and for the following Visitor)
if is_libcst_available():
import libcst as cst
from libcst import CSTVisitor
from libcst import matchers as m
class ClassFinder(CSTVisitor):
"""
A visitor to find all classes in a python module.
"""
def __init__(self):
self.classes: list = []
self.public_classes: list = []
self.is_in_class = False
def visit_ClassDef(self, node: cst.ClassDef) -> None:
"""Record class names. We assume classes always only appear at top-level (i.e. no class definition in function or similar)"""
self.classes.append(node.name.value)
self.is_in_class = True
def leave_ClassDef(self, node: cst.ClassDef):
self.is_in_class = False
def visit_SimpleStatementLine(self, node: cst.SimpleStatementLine):
"""Record all public classes inside the `__all__` assignment."""
simple_top_level_assign_structure = m.SimpleStatementLine(
body=[m.Assign(targets=[m.AssignTarget(target=m.Name())])]
)
if not self.is_in_class and m.matches(node, simple_top_level_assign_structure):
assigned_variable = node.body[0].targets[0].target.value
if assigned_variable == "__all__":
elements = node.body[0].value.elements
self.public_classes = [element.value.value for element in elements]
CURRENT_YEAR = date.today().year
TRANSFORMERS_PATH = Path(__file__).parent.parent
REPO_PATH = TRANSFORMERS_PATH.parent.parent
COPYRIGHT = f"""
# coding=utf-8
# Copyright {CURRENT_YEAR} the HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""".lstrip()
class ModelInfos:
"""
Retrieve the basic informations about an existing model classes.
"""
def __init__(self, lowercase_name: str):
# Just to make sure it's indeed lowercase
self.lowercase_name = lowercase_name.lower().replace(" ", "_").replace("-", "_")
if self.lowercase_name not in CONFIG_MAPPING_NAMES:
self.lowercase_name.replace("_", "-")
if self.lowercase_name not in CONFIG_MAPPING_NAMES:
raise ValueError(f"{lowercase_name} is not a valid model name")
self.paper_name = MODEL_NAMES_MAPPING[self.lowercase_name]
self.config_class = CONFIG_MAPPING_NAMES[self.lowercase_name]
self.camelcase_name = self.config_class.replace("Config", "")
# Get tokenizer class
if self.lowercase_name in TOKENIZER_MAPPING_NAMES:
self.tokenizer_class, self.fast_tokenizer_class = TOKENIZER_MAPPING_NAMES[self.lowercase_name]
self.fast_tokenizer_class = (
None if self.fast_tokenizer_class == "PreTrainedTokenizerFast" else self.fast_tokenizer_class
)
else:
self.tokenizer_class, self.fast_tokenizer_class = None, None
self.image_processor_class, self.fast_image_processor_class = IMAGE_PROCESSOR_MAPPING_NAMES.get(
self.lowercase_name, (None, None)
)
self.video_processor_class = VIDEO_PROCESSOR_MAPPING_NAMES.get(self.lowercase_name, None)
self.feature_extractor_class = FEATURE_EXTRACTOR_MAPPING_NAMES.get(self.lowercase_name, None)
self.processor_class = PROCESSOR_MAPPING_NAMES.get(self.lowercase_name, None)
def add_content_to_file(file_name: Union[str, os.PathLike], new_content: str, add_after: str):
"""
A utility to add some content inside a given file.
Args:
file_name (`str` or `os.PathLike`):
The name of the file in which we want to insert some content.
new_content (`str`):
The content to add.
add_after (`str`):
The new content is added just after the first instance matching it.
"""
with open(file_name, "r", encoding="utf-8") as f:
old_content = f.read()
before, after = old_content.split(add_after, 1)
new_content = before + add_after + new_content + after
with open(file_name, "w", encoding="utf-8") as f:
f.write(new_content)
def add_model_to_auto_mappings(
old_model_infos: ModelInfos,
new_lowercase_name: str,
new_model_paper_name: str,
filenames_to_add: list[tuple[str, bool]],
):
"""
Add a model to all the relevant mappings in the auto module.
Args:
old_model_infos (`ModelInfos`):
The structure containing the class informations of the old model.
new_lowercase_name (`str`):
The new lowercase model name.
new_model_paper_name (`str`):
The fully cased name (as in the official paper name) of the new model.
filenames_to_add (`list[tuple[str, bool]]`):
A list of tuples of all potential filenames to add for a new model, along a boolean flag describing if we
should add this file or not. For example, [(`modeling_xxx.px`, True), (`configuration_xxx.py`, True), (`tokenization_xxx.py`, False),...]
"""
new_cased_name = "".join(x.title() for x in new_lowercase_name.replace("-", "_").split("_"))
old_lowercase_name = old_model_infos.lowercase_name
old_cased_name = old_model_infos.camelcase_name
filenames_to_add = [
(filename.replace(old_lowercase_name, "auto"), to_add) for filename, to_add in filenames_to_add[1:]
]
# fast tokenizer/image processor have the same auto mappings as normal ones
corrected_filenames_to_add = []
for file, to_add in filenames_to_add:
if re.search(r"(?:tokenization)|(?:image_processing)_auto_fast.py", file):
previous_file, previous_to_add = corrected_filenames_to_add[-1]
corrected_filenames_to_add[-1] = (previous_file, previous_to_add or to_add)
else:
corrected_filenames_to_add.append((file, to_add))
# Add the config mappings directly as the handling for config is a bit different
add_content_to_file(
TRANSFORMERS_PATH / "models" / "auto" / "configuration_auto.py",
new_content=f' ("{new_lowercase_name}", "{new_cased_name}Config"),\n',
add_after="CONFIG_MAPPING_NAMES = OrderedDict[str, str](\n [\n # Add configs here\n",
)
add_content_to_file(
TRANSFORMERS_PATH / "models" / "auto" / "configuration_auto.py",
new_content=f' ("{new_lowercase_name}", "{new_model_paper_name}"),\n',
add_after="MODEL_NAMES_MAPPING = OrderedDict[str, str](\n [\n # Add full (and cased) model names here\n",
)
for filename, to_add in corrected_filenames_to_add:
if to_add:
# The auto mapping
filename = filename.replace("_fast.py", ".py")
with open(TRANSFORMERS_PATH / "models" / "auto" / filename) as f:
file = f.read()
# The regex has to be a bit complex like this as the tokenizer mapping has new lines everywhere
matching_lines = re.findall(
rf'( {{8,12}}\(\s*"{old_lowercase_name}",.*?\),\n)(?: {{4,12}}\(|\])', file, re.DOTALL
)
for match in matching_lines:
add_content_to_file(
TRANSFORMERS_PATH / "models" / "auto" / filename,
new_content=match.replace(old_lowercase_name, new_lowercase_name).replace(
old_cased_name, new_cased_name
),
add_after=match,
)
def create_doc_file(new_paper_name: str, public_classes: list[str]):
"""
Create a new doc file to fill for the new model.
Args:
new_paper_name (`str`):
The fully cased name (as in the official paper name) of the new model.
public_classes (`list[str]`):
A list of all the public classes that the model will have in the library.
"""
added_note = (
"\n\n⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that "
"may not be rendered properly in your Markdown viewer.\n\n-->\n\n"
)
copyright_for_markdown = re.sub(r"# ?", "", COPYRIGHT).replace("coding=utf-8\n", "<!--") + added_note
doc_template = textwrap.dedent(
f"""
# {new_paper_name}
## Overview
The {new_paper_name} model was proposed in [<INSERT PAPER NAME HERE>](<INSERT PAPER LINK HERE>) by <INSERT AUTHORS HERE>.
<INSERT SHORT SUMMARY HERE>
The abstract from the paper is the following:
<INSERT PAPER ABSTRACT HERE>
Tips:
<INSERT TIPS ABOUT MODEL HERE>
This model was contributed by [INSERT YOUR HF USERNAME HERE](https://huggingface.co/<INSERT YOUR HF USERNAME HERE>).
The original code can be found [here](<INSERT LINK TO GITHUB REPO HERE>).
## Usage examples
<INSERT SOME NICE EXAMPLES HERE>
"""
)
# Add public classes doc
doc_for_classes = []
for class_ in public_classes:
doc = f"## {class_}\n\n[[autodoc]] {class_}"
if "Model" in class_:
doc += "\n - forward"
doc_for_classes.append(doc)
class_doc = "\n\n".join(doc_for_classes)
return copyright_for_markdown + doc_template + class_doc
def insert_model_in_doc_toc(old_lowercase_name: str, new_lowercase_name: str, new_model_paper_name: str):
"""
Insert the new model in the doc `_toctree.yaml`, in the same section as the old model.
Args:
old_lowercase_name (`str`):
The old lowercase model name.
new_lowercase_name (`str`):
The old lowercase model name.
new_model_paper_name (`str`):
The fully cased name (as in the official paper name) of the new model.
"""
toc_file = REPO_PATH / "docs" / "source" / "en" / "_toctree.yml"
with open(toc_file, "r") as f:
content = f.read()
old_model_toc = re.search(rf"- local: model_doc/{old_lowercase_name}\n {{8}}title: .*?\n", content).group(0)
new_toc = f" - local: model_doc/{new_lowercase_name}\n title: {new_model_paper_name}\n"
add_content_to_file(
REPO_PATH / "docs" / "source" / "en" / "_toctree.yml", new_content=new_toc, add_after=old_model_toc
)
def create_init_file(old_lowercase_name: str, new_lowercase_name: str, filenames_to_add: list[tuple[str, bool]]):
"""
Create the `__init__.py` file to add in the new model folder.
Args:
old_lowercase_name (`str`):
The old lowercase model name.
new_lowercase_name (`str`):
The new lowercase model name.
filenames_to_add (`list[tuple[str, bool]]`):
A list of tuples of all potential filenames to add for a new model, along a boolean flag describing if we
should add this file or not. For example, [(`modeling_xxx.px`, True), (`configuration_xxx.py`, True), (`tokenization_xxx.py`, False),...]
"""
filenames_to_add = [
(filename.replace(old_lowercase_name, new_lowercase_name).replace(".py", ""), to_add)
for filename, to_add in filenames_to_add
]
imports = "\n ".join(f"from .{file} import *" for file, to_add in filenames_to_add if to_add)
init_file = COPYRIGHT + textwrap.dedent(
f"""
from typing import TYPE_CHECKING
from ...utils import _LazyModule
from ...utils.import_utils import define_import_structure
if TYPE_CHECKING:
{imports}
else:
import sys
_file = globals()["__file__"]
sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)
"""
)
return init_file
def find_all_classes_from_file(module_name: str) -> set:
"""
Find the name of all classes defined in `module_name`, including public ones (defined in `__all__`).
Args:
module_name (`str`):
The full path to the python module from which to extract classes.
"""
with open(module_name, "r", encoding="utf-8") as file:
source_code = file.read()
module = cst.parse_module(source_code)
visitor = ClassFinder()
module.visit(visitor)
return visitor.classes, visitor.public_classes
def find_modular_structure(
module_name: str, old_model_infos: ModelInfos, new_cased_name: str
) -> tuple[str, str, list]:
"""
Extract the modular structure that will be needed to copy a file `module_name` using modular.
Args:
module_name (`str`):
The full path to the python module to copy with modular.
old_model_infos (`ModelInfos`):
The structure containing the class informations of the old model.
new_cased_name (`str`):
The new cased model name.
"""
all_classes, public_classes = find_all_classes_from_file(module_name)
import_location = ".".join(module_name.parts[-2:]).replace(".py", "")
old_cased_name = old_model_infos.camelcase_name
imports = f"from ..{import_location} import {', '.join(class_ for class_ in all_classes)}"
modular_classes = "\n\n".join(
f"class {class_.replace(old_cased_name, new_cased_name)}({class_}):\n pass" for class_ in all_classes
)
public_classes = [class_.replace(old_cased_name, new_cased_name) for class_ in public_classes]
return imports, modular_classes, public_classes
def create_modular_file(
old_model_infos: ModelInfos,
new_lowercase_name: str,
filenames_to_add: list[tuple[str, bool]],
) -> str:
"""
Create a new modular file which will copy the old model, based on the new name and the different filenames
(modules) to add.
Args:
old_model_infos (`ModelInfos`):
The structure containing the class informations of the old model.
new_lowercase_name (`str`):
The new lowercase model name.
filenames_to_add (`list[tuple[str, bool]]`):
A list of tuples of all potential filenames to add for a new model, along a boolean flag describing if we
should add this file or not. For example, [(`modeling_xxx.px`, True), (`configuration_xxx.py`, True), (`tokenization_xxx.py`, False),...]
"""
new_cased_name = "".join(x.title() for x in new_lowercase_name.replace("-", "_").split("_"))
old_lowercase_name = old_model_infos.lowercase_name
old_folder_root = TRANSFORMERS_PATH / "models" / old_lowercase_name
# Construct the modular file from the original (old) model, by subclassing each class
all_imports = ""
all_bodies = ""
all_public_classes = []
for filename, to_add in filenames_to_add:
if to_add:
imports, body, public_classes = find_modular_structure(
old_folder_root / filename, old_model_infos, new_cased_name
)
all_imports += f"\n{imports}"
all_bodies += f"\n\n{body}"
all_public_classes.extend(public_classes)
# Create the __all__ assignment
public_classes_formatted = "\n ".join(f"{public_class}," for public_class in all_public_classes)
all_statement = textwrap.dedent(
f"""
__all__ = [
{public_classes_formatted}
]
"""
)
# Create the whole modular file
modular_file = COPYRIGHT + all_imports + all_bodies + all_statement
# Remove outer explicit quotes "" around the public class names before returning them
all_public_classes = [public_class.replace('"', "") for public_class in all_public_classes]
return modular_file, all_public_classes
def create_test_files(old_model_infos: ModelInfos, new_lowercase_name, filenames_to_add: list[tuple[str, bool]]):
"""
Create the test files for the new model. It basically copies over the old test files and adjust the class names.
Args:
old_model_infos (`ModelInfos`):
The structure containing the class informations of the old model.
new_lowercase_name (`str`):
The new lowercase model name.
filenames_to_add (`list[tuple[str, bool]]`):
A list of tuples of all potential filenames to add for a new model, along a boolean flag describing if we
should add this file or not. For example, [(`modeling_xxx.px`, True), (`configuration_xxx.py`, True), (`tokenization_xxx.py`, False),...]
"""
new_cased_name = "".join(x.title() for x in new_lowercase_name.replace("-", "_").split("_"))
old_lowercase_name = old_model_infos.lowercase_name
old_cased_name = old_model_infos.camelcase_name
filenames_to_add = [
("test_" + filename.replace(old_lowercase_name, new_lowercase_name), to_add)
for filename, to_add in filenames_to_add[1:]
]
# fast tokenizer/image processor have the same test files as normal ones
corrected_filenames_to_add = []
for file, to_add in filenames_to_add:
if re.search(rf"test_(?:tokenization)|(?:image_processing)_{new_lowercase_name}_fast.py", file):
previous_file, previous_to_add = corrected_filenames_to_add[-1]
corrected_filenames_to_add[-1] = (previous_file, previous_to_add or to_add)
else:
corrected_filenames_to_add.append((file, to_add))
test_files = {}
for new_file, to_add in corrected_filenames_to_add:
if to_add:
original_test_file = new_file.replace(new_lowercase_name, old_lowercase_name)
original_test_path = REPO_PATH / "tests" / "models" / old_lowercase_name / original_test_file
# Sometimes, tests may not exist
if not original_test_path.is_file():
continue
with open(original_test_path, "r") as f:
test_code = f.read()
# Remove old copyright and add new one
test_lines = test_code.split("\n")
idx = 0
while test_lines[idx].startswith("#"):
idx += 1
test_code = COPYRIGHT + "\n".join(test_lines[idx:])
test_files[new_file] = test_code.replace(old_cased_name, new_cased_name)
return test_files
def create_new_model_like(
old_model_infos: ModelInfos,
new_lowercase_name: str,
new_model_paper_name: str,
filenames_to_add: list[tuple[str, bool]],
create_fast_image_processor: bool,
):
"""
Creates a new model module like a given model of the Transformers library.
Args:
old_model_infos (`ModelInfos`):
The structure containing the class informations of the old model.
new_lowercase_name (`str`):
The new lowercase model name.
new_model_paper_name (`str`):
The fully cased name (as in the official paper name) of the new model.
filenames_to_add (`list[tuple[str, bool]]`):
A list of tuples of all potential filenames to add for a new model, along a boolean flag describing if we
should add this file or not. For example, [(`modeling_xxx.px`, True), (`configuration_xxx.py`, True), (`tokenization_xxx.py`, False),...]
create_fast_image_processor (`bool`):
If it makes sense, whether to add a fast processor as well, even if the old model does not have one.
"""
# As the import was protected, raise if not present (as it's actually a hard dependency for this command)
if not is_libcst_available():
raise ValueError("You need to install `libcst` to run this command -> `pip install libcst`")
old_lowercase_name = old_model_infos.lowercase_name
# 1. We create the folder for our new model
new_module_folder = TRANSFORMERS_PATH / "models" / new_lowercase_name
os.makedirs(new_module_folder, exist_ok=True)
# 2. Create and add the modular file
modular_file, public_classes = create_modular_file(old_model_infos, new_lowercase_name, filenames_to_add)
with open(new_module_folder / f"modular_{new_lowercase_name}.py", "w") as f:
f.write(modular_file)
# 3. Create and add the __init__.py
init_file = create_init_file(old_lowercase_name, new_lowercase_name, filenames_to_add)
with open(new_module_folder / "__init__.py", "w") as f:
f.write(init_file)
# 4. Add new model to the models init
add_content_to_file(
TRANSFORMERS_PATH / "models" / "__init__.py",
new_content=f" from .{new_lowercase_name} import *\n",
add_after="if TYPE_CHECKING:\n",
)
# 5. Add model to auto mappings
add_model_to_auto_mappings(old_model_infos, new_lowercase_name, new_model_paper_name, filenames_to_add)
# 6. Add test files
tests_folder = REPO_PATH / "tests" / "models" / new_lowercase_name
os.makedirs(tests_folder, exist_ok=True)
# Add empty __init__.py
with open(tests_folder / "__init__.py", "w"):
pass
test_files = create_test_files(old_model_infos, new_lowercase_name, filenames_to_add)
for filename, content in test_files.items():
with open(tests_folder / filename, "w") as f:
f.write(content)
# 7. Add doc file
doc_file = create_doc_file(new_model_paper_name, public_classes)
with open(REPO_PATH / "docs" / "source" / "en" / "model_doc" / f"{new_lowercase_name}.md", "w") as f:
f.write(doc_file)
insert_model_in_doc_toc(old_lowercase_name, new_lowercase_name, new_model_paper_name)
# 8. Add additional fast image processor if necessary
if create_fast_image_processor:
add_fast_image_processor(model_name=new_lowercase_name)
# 9. Run linters
model_init_file = TRANSFORMERS_PATH / "models" / "__init__.py"
subprocess.run(
["ruff", "check", new_module_folder, tests_folder, model_init_file, "--fix"],
cwd=REPO_PATH,
stdout=subprocess.DEVNULL,
)
subprocess.run(
["ruff", "format", new_module_folder, tests_folder, model_init_file],
cwd=REPO_PATH,
stdout=subprocess.DEVNULL,
)
subprocess.run(
["python", "utils/check_doc_toc.py", "--fix_and_overwrite"], cwd=REPO_PATH, stdout=subprocess.DEVNULL
)
subprocess.run(["python", "utils/sort_auto_mappings.py"], cwd=REPO_PATH, stdout=subprocess.DEVNULL)
# 10. Run the modular conversion
subprocess.run(
["python", "utils/modular_model_converter.py", new_lowercase_name], cwd=REPO_PATH, stdout=subprocess.DEVNULL
)
def get_user_field(
question: str,
default_value: Optional[str] = None,
convert_to: Optional[Callable] = None,
fallback_message: Optional[str] = None,
) -> Any:
"""
A utility function that asks a question to the user to get an answer, potentially looping until it gets a valid
answer.
Args:
question (`str`):
The question to ask the user.
default_value (`str`, *optional*):
A potential default value that will be used when the answer is empty.
convert_to (`Callable`, *optional*):
If set, the answer will be passed to this function. If this function raises an error on the provided
answer, the question will be asked again.
fallback_message (`str`, *optional*):
A message that will be displayed each time the question is asked again to the user.
Returns:
`Any`: The answer provided by the user (or the default), passed through the potential conversion function.
"""
if not question.endswith(" "):
question = question + " "
if default_value is not None:
question = f"{question} [{default_value}] "
valid_answer = False
while not valid_answer:
answer = input(question)
if default_value is not None and len(answer) == 0:
answer = default_value
if convert_to is not None:
try:
answer = convert_to(answer)
valid_answer = True
except Exception:
valid_answer = False
else:
valid_answer = True
if not valid_answer:
print(fallback_message)
return answer
def convert_to_bool(x: str) -> bool:
"""
Converts a string to a bool.
"""
if x.lower() in ["1", "y", "yes", "true"]:
return True
if x.lower() in ["0", "n", "no", "false"]:
return False
raise ValueError(f"{x} is not a value that can be converted to a bool.")
def get_user_input():
"""
Ask the user for the necessary inputs to add the new model.
"""
model_types = list(MODEL_NAMES_MAPPING.keys())
# Get old model type
valid_model_type = False
while not valid_model_type:
old_model_type = input(
"What model would you like to duplicate? Please provide it as lowercase, e.g. `llama`): "
)
if old_model_type in model_types:
valid_model_type = True
else:
print(f"{old_model_type} is not a valid model type.")
near_choices = difflib.get_close_matches(old_model_type, model_types)
if len(near_choices) >= 1:
if len(near_choices) > 1:
near_choices = " or ".join(near_choices)
print(f"Did you mean {near_choices}?")
old_model_infos = ModelInfos(old_model_type)
# Ask for the new model name
new_lowercase_name = get_user_field(
"What is the new model name? Please provide it as snake lowercase, e.g. `new_model`?"
)
new_model_paper_name = get_user_field(
"What is the fully cased name you would like to appear in the doc (e.g. `NeW ModEl`)? ",
default_value="".join(x.title() for x in new_lowercase_name.split("_")),
)
# Ask if we want to add individual processor classes as well
add_tokenizer = False
add_fast_tokenizer = False
add_image_processor = False
add_fast_image_processor = False
add_video_processor = False
add_feature_extractor = False
add_processor = False
if old_model_infos.tokenizer_class is not None:
add_tokenizer = get_user_field(
f"Do you want to create a new tokenizer? If `no`, it will use the same as {old_model_type} (y/n)?",
convert_to=convert_to_bool,
fallback_message="Please answer yes/no, y/n, true/false or 1/0. ",
)
if old_model_infos.fast_tokenizer_class is not None:
add_fast_tokenizer = get_user_field(
f"Do you want to create a new fast tokenizer? If `no`, it will use the same as {old_model_type} (y/n)?",
convert_to=convert_to_bool,
fallback_message="Please answer yes/no, y/n, true/false or 1/0. ",
)
if old_model_infos.image_processor_class is not None:
add_image_processor = get_user_field(
f"Do you want to create a new image processor? If `no`, it will use the same as {old_model_type} (y/n)?",
convert_to=convert_to_bool,
fallback_message="Please answer yes/no, y/n, true/false or 1/0. ",
)
if old_model_infos.fast_image_processor_class is not None:
add_fast_image_processor = get_user_field(
f"Do you want to create a new fast image processor? If `no`, it will use the same as {old_model_type} (y/n)?",
convert_to=convert_to_bool,
fallback_message="Please answer yes/no, y/n, true/false or 1/0. ",
)
if old_model_infos.video_processor_class is not None:
add_video_processor = get_user_field(
f"Do you want to create a new video processor? If `no`, it will use the same as {old_model_type} (y/n)?",
convert_to=convert_to_bool,
fallback_message="Please answer yes/no, y/n, true/false or 1/0. ",
)
if old_model_infos.feature_extractor_class is not None:
add_feature_extractor = get_user_field(
f"Do you want to create a new feature extractor? If `no`, it will use the same as {old_model_type} (y/n)?",
convert_to=convert_to_bool,
fallback_message="Please answer yes/no, y/n, true/false or 1/0. ",
)
if old_model_infos.processor_class is not None:
add_processor = get_user_field(
f"Do you want to create a new processor? If `no`, it will use the same as {old_model_type} (y/n)?",
convert_to=convert_to_bool,
fallback_message="Please answer yes/no, y/n, true/false or 1/0. ",
)
old_lowercase_name = old_model_infos.lowercase_name
# A list of the old filenames, along whether we should copy them or not
filenames_to_add = (
(f"configuration_{old_lowercase_name}.py", True),
(f"modeling_{old_lowercase_name}.py", True),
(f"tokenization_{old_lowercase_name}.py", add_tokenizer),
(f"tokenization_{old_lowercase_name}_fast.py", add_fast_tokenizer),
(f"image_processing_{old_lowercase_name}.py", add_image_processor),
(f"image_processing_{old_lowercase_name}_fast.py", add_fast_image_processor),
(f"video_processing_{old_lowercase_name}.py", add_video_processor),
(f"feature_extraction_{old_lowercase_name}.py", add_feature_extractor),
(f"processing_{old_lowercase_name}.py", add_processor),
)
create_fast_image_processor = False
if add_image_processor and not add_fast_image_processor:
create_fast_image_processor = get_user_field(
"A fast image processor can be created from the slow one, but modifications might be needed. "
"Should we add a fast image processor class for this model (recommended) (y/n)? ",
convert_to=convert_to_bool,
default_value="y",
fallback_message="Please answer yes/no, y/n, true/false or 1/0.",
)
return old_model_infos, new_lowercase_name, new_model_paper_name, filenames_to_add, create_fast_image_processor
def add_new_model_like_command_factory(args: Namespace):
return AddNewModelLikeCommand(path_to_repo=args.path_to_repo)
class AddNewModelLikeCommand(BaseTransformersCLICommand):
@staticmethod
def register_subcommand(parser: ArgumentParser):
add_new_model_like_parser = parser.add_parser("add-new-model-like")
add_new_model_like_parser.add_argument(
"--path_to_repo", type=str, help="When not using an editable install, the path to the Transformers repo."
)
add_new_model_like_parser.set_defaults(func=add_new_model_like_command_factory)
def __init__(self, path_to_repo=None, *args):
(
self.old_model_infos,
self.new_lowercase_name,
self.new_model_paper_name,
self.filenames_to_add,
self.create_fast_image_processor,
) = get_user_input()
self.path_to_repo = path_to_repo
def run(self):
if self.path_to_repo is not None:
# Adapt constants
global TRANSFORMERS_PATH
global REPO_PATH
REPO_PATH = Path(self.path_to_repo)
TRANSFORMERS_PATH = REPO_PATH / "src" / "transformers"
create_new_model_like(
old_model_infos=self.old_model_infos,
new_lowercase_name=self.new_lowercase_name,
new_model_paper_name=self.new_model_paper_name,
filenames_to_add=self.filenames_to_add,
create_fast_image_processor=self.create_fast_image_processor,
)
| transformers/src/transformers/commands/add_new_model_like.py/0 | {
"file_path": "transformers/src/transformers/commands/add_new_model_like.py",
"repo_id": "transformers",
"token_count": 13858
} | 419 |
# Copyright 2020 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import multiprocessing as mp
import random
import warnings
from collections.abc import Mapping
from dataclasses import dataclass
from random import randint
from typing import Any, Callable, NewType, Optional, Union
import numpy as np
from ..models.bert import BertTokenizer, BertTokenizerFast
from ..tokenization_utils_base import PreTrainedTokenizerBase
from ..utils import PaddingStrategy
InputDataClass = NewType("InputDataClass", Any)
"""
A DataCollator is a function that takes a list of samples from a Dataset and collate them into a batch, as a dictionary
of PyTorch/TensorFlow tensors or NumPy arrays.
"""
DataCollator = NewType("DataCollator", Callable[[list[InputDataClass]], dict[str, Any]])
class DataCollatorMixin:
def __call__(self, features, return_tensors=None):
if return_tensors is None:
return_tensors = self.return_tensors
if return_tensors == "tf":
return self.tf_call(features)
elif return_tensors == "pt":
return self.torch_call(features)
elif return_tensors == "np":
return self.numpy_call(features)
else:
raise ValueError(f"Framework '{return_tensors}' not recognized!")
def pad_without_fast_tokenizer_warning(tokenizer, *pad_args, **pad_kwargs):
"""
Pads without triggering the warning about how using the pad function is sub-optimal when using a fast tokenizer.
"""
# To avoid errors when using Feature extractors
if not hasattr(tokenizer, "deprecation_warnings"):
return tokenizer.pad(*pad_args, **pad_kwargs)
# Save the state of the warning, then disable it
warning_state = tokenizer.deprecation_warnings.get("Asking-to-pad-a-fast-tokenizer", False)
tokenizer.deprecation_warnings["Asking-to-pad-a-fast-tokenizer"] = True
try:
padded = tokenizer.pad(*pad_args, **pad_kwargs)
finally:
# Restore the state of the warning.
tokenizer.deprecation_warnings["Asking-to-pad-a-fast-tokenizer"] = warning_state
return padded
def default_data_collator(features: list[InputDataClass], return_tensors="pt") -> dict[str, Any]:
"""
Very simple data collator that simply collates batches of dict-like objects and performs special handling for
potential keys named:
- `label`: handles a single value (int or float) per object
- `label_ids`: handles a list of values per object
Does not do any additional preprocessing: property names of the input object will be used as corresponding inputs
to the model. See glue and ner for example of how it's useful.
"""
# In this function we'll make the assumption that all `features` in the batch
# have the same attributes.
# So we will look at the first element as a proxy for what attributes exist
# on the whole batch.
if return_tensors == "pt":
return torch_default_data_collator(features)
elif return_tensors == "tf":
return tf_default_data_collator(features)
elif return_tensors == "np":
return numpy_default_data_collator(features)
@dataclass
class DefaultDataCollator(DataCollatorMixin):
"""
Very simple data collator that simply collates batches of dict-like objects and performs special handling for
potential keys named:
- `label`: handles a single value (int or float) per object
- `label_ids`: handles a list of values per object
Does not do any additional preprocessing: property names of the input object will be used as corresponding inputs
to the model. See glue and ner for example of how it's useful.
This is an object (like other data collators) rather than a pure function like default_data_collator. This can be
helpful if you need to set a return_tensors value at initialization.
Args:
return_tensors (`str`, *optional*, defaults to `"pt"`):
The type of Tensor to return. Allowable values are "np", "pt" and "tf".
"""
return_tensors: str = "pt"
def __call__(self, features: list[dict[str, Any]], return_tensors=None) -> dict[str, Any]:
if return_tensors is None:
return_tensors = self.return_tensors
return default_data_collator(features, return_tensors)
def torch_default_data_collator(features: list[InputDataClass]) -> dict[str, Any]:
import torch
if not isinstance(features[0], Mapping):
features = [vars(f) for f in features]
first = features[0]
batch = {}
# Special handling for labels.
# Ensure that tensor is created with the correct type
# (it should be automatically the case, but let's make sure of it.)
if "label" in first and first["label"] is not None:
label = first["label"].item() if isinstance(first["label"], torch.Tensor) else first["label"]
dtype = torch.long if isinstance(label, int) else torch.float
batch["labels"] = torch.tensor([f["label"] for f in features], dtype=dtype)
elif "label_ids" in first and first["label_ids"] is not None:
if isinstance(first["label_ids"], torch.Tensor):
batch["labels"] = torch.stack([f["label_ids"] for f in features])
else:
dtype = torch.long if isinstance(first["label_ids"][0], int) else torch.float
batch["labels"] = torch.tensor([f["label_ids"] for f in features], dtype=dtype)
# Handling of all other possible keys.
# Again, we will use the first element to figure out which key/values are not None for this model.
for k, v in first.items():
if k not in ("label", "label_ids") and v is not None and not isinstance(v, str):
if isinstance(v, torch.Tensor):
batch[k] = torch.stack([f[k] for f in features])
elif isinstance(v, np.ndarray):
batch[k] = torch.from_numpy(np.stack([f[k] for f in features]))
else:
batch[k] = torch.tensor([f[k] for f in features])
return batch
def tf_default_data_collator(features: list[InputDataClass]) -> dict[str, Any]:
import tensorflow as tf
if not isinstance(features[0], Mapping):
features = [vars(f) for f in features]
first = features[0]
batch = {}
# Special handling for labels.
# Ensure that tensor is created with the correct type
# (it should be automatically the case, but let's make sure of it.)
if "label" in first and first["label"] is not None:
label_col_name = "label"
elif "label_ids" in first and first["label_ids"] is not None:
label_col_name = "label_ids"
elif "labels" in first and first["labels"] is not None:
label_col_name = "labels"
else:
label_col_name = None
if label_col_name is not None:
if isinstance(first[label_col_name], tf.Tensor):
dtype = tf.int64 if first[label_col_name].dtype.is_integer else tf.float32
elif isinstance(first[label_col_name], (np.ndarray, np.generic)):
dtype = tf.int64 if np.issubdtype(first[label_col_name].dtype, np.integer) else tf.float32
elif isinstance(first[label_col_name], (tuple, list)):
dtype = tf.int64 if isinstance(first[label_col_name][0], int) else tf.float32
else:
dtype = tf.int64 if isinstance(first[label_col_name], int) else tf.float32
batch["labels"] = tf.convert_to_tensor([f[label_col_name] for f in features], dtype=dtype)
# Handling of all other possible keys.
# Again, we will use the first element to figure out which key/values are not None for this model.
for k, v in first.items():
if k not in ("label", "label_ids", "labels") and v is not None and not isinstance(v, str):
if isinstance(v, (tf.Tensor, np.ndarray)):
batch[k] = tf.stack([f[k] for f in features])
else:
batch[k] = tf.convert_to_tensor([f[k] for f in features])
return batch
def numpy_default_data_collator(features: list[InputDataClass]) -> dict[str, Any]:
if not isinstance(features[0], Mapping):
features = [vars(f) for f in features]
first = features[0]
batch = {}
# Special handling for labels.
# Ensure that tensor is created with the correct type
# (it should be automatically the case, but let's make sure of it.)
if "label" in first and first["label"] is not None:
label = first["label"].item() if isinstance(first["label"], np.ndarray) else first["label"]
dtype = np.int64 if isinstance(label, int) else np.float32
batch["labels"] = np.array([f["label"] for f in features], dtype=dtype)
elif "label_ids" in first and first["label_ids"] is not None:
if isinstance(first["label_ids"], np.ndarray):
batch["labels"] = np.stack([f["label_ids"] for f in features])
else:
dtype = np.int64 if isinstance(first["label_ids"][0], int) else np.float32
batch["labels"] = np.array([f["label_ids"] for f in features], dtype=dtype)
# Handling of all other possible keys.
# Again, we will use the first element to figure out which key/values are not None for this model.
for k, v in first.items():
if k not in ("label", "label_ids") and v is not None and not isinstance(v, str):
if isinstance(v, np.ndarray):
batch[k] = np.stack([f[k] for f in features])
else:
batch[k] = np.array([f[k] for f in features])
return batch
@dataclass
class DataCollatorWithPadding:
"""
Data collator that will dynamically pad the inputs received.
Args:
tokenizer ([`PreTrainedTokenizer`] or [`PreTrainedTokenizerFast`]):
The tokenizer used for encoding the data.
padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `True`):
Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
among:
- `True` or `'longest'` (default): Pad to the longest sequence in the batch (or no padding if only a single
sequence is provided).
- `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
acceptable input length for the model if that argument is not provided.
- `False` or `'do_not_pad'`: No padding (i.e., can output a batch with sequences of different lengths).
max_length (`int`, *optional*):
Maximum length of the returned list and optionally padding length (see above).
pad_to_multiple_of (`int`, *optional*):
If set will pad the sequence to a multiple of the provided value.
This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
7.0 (Volta).
return_tensors (`str`, *optional*, defaults to `"pt"`):
The type of Tensor to return. Allowable values are "np", "pt" and "tf".
"""
tokenizer: PreTrainedTokenizerBase
padding: Union[bool, str, PaddingStrategy] = True
max_length: Optional[int] = None
pad_to_multiple_of: Optional[int] = None
return_tensors: str = "pt"
def __call__(self, features: list[dict[str, Any]]) -> dict[str, Any]:
batch = pad_without_fast_tokenizer_warning(
self.tokenizer,
features,
padding=self.padding,
max_length=self.max_length,
pad_to_multiple_of=self.pad_to_multiple_of,
return_tensors=self.return_tensors,
)
if "label" in batch:
batch["labels"] = batch["label"]
del batch["label"]
if "label_ids" in batch:
batch["labels"] = batch["label_ids"]
del batch["label_ids"]
return batch
@dataclass
class DataCollatorForTokenClassification(DataCollatorMixin):
"""
Data collator that will dynamically pad the inputs received, as well as the labels.
Args:
tokenizer ([`PreTrainedTokenizer`] or [`PreTrainedTokenizerFast`]):
The tokenizer used for encoding the data.
padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `True`):
Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
among:
- `True` or `'longest'` (default): Pad to the longest sequence in the batch (or no padding if only a single
sequence is provided).
- `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
acceptable input length for the model if that argument is not provided.
- `False` or `'do_not_pad'`: No padding (i.e., can output a batch with sequences of different lengths).
max_length (`int`, *optional*):
Maximum length of the returned list and optionally padding length (see above).
pad_to_multiple_of (`int`, *optional*):
If set will pad the sequence to a multiple of the provided value.
This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
7.0 (Volta).
label_pad_token_id (`int`, *optional*, defaults to -100):
The id to use when padding the labels (-100 will be automatically ignore by PyTorch loss functions).
return_tensors (`str`, *optional*, defaults to `"pt"`):
The type of Tensor to return. Allowable values are "np", "pt" and "tf".
"""
tokenizer: PreTrainedTokenizerBase
padding: Union[bool, str, PaddingStrategy] = True
max_length: Optional[int] = None
pad_to_multiple_of: Optional[int] = None
label_pad_token_id: int = -100
return_tensors: str = "pt"
def torch_call(self, features):
import torch
label_name = "label" if "label" in features[0] else "labels"
labels = [feature[label_name] for feature in features] if label_name in features[0] else None
no_labels_features = [{k: v for k, v in feature.items() if k != label_name} for feature in features]
batch = pad_without_fast_tokenizer_warning(
self.tokenizer,
no_labels_features,
padding=self.padding,
max_length=self.max_length,
pad_to_multiple_of=self.pad_to_multiple_of,
return_tensors="pt",
)
if labels is None:
return batch
sequence_length = batch["input_ids"].shape[1]
padding_side = self.tokenizer.padding_side
def to_list(tensor_or_iterable):
if isinstance(tensor_or_iterable, torch.Tensor):
return tensor_or_iterable.tolist()
return list(tensor_or_iterable)
if padding_side == "right":
batch[label_name] = [
to_list(label) + [self.label_pad_token_id] * (sequence_length - len(label)) for label in labels
]
else:
batch[label_name] = [
[self.label_pad_token_id] * (sequence_length - len(label)) + to_list(label) for label in labels
]
batch[label_name] = torch.tensor(batch[label_name], dtype=torch.int64)
return batch
def tf_call(self, features):
import tensorflow as tf
label_name = "label" if "label" in features[0] else "labels"
labels = [feature[label_name] for feature in features] if label_name in features[0] else None
batch = pad_without_fast_tokenizer_warning(
self.tokenizer,
features,
padding=self.padding,
max_length=self.max_length,
pad_to_multiple_of=self.pad_to_multiple_of,
# Conversion to tensors will fail if we have labels as they are not of the same length yet.
return_tensors="tf" if labels is None else None,
)
if labels is None:
return batch
sequence_length = tf.convert_to_tensor(batch["input_ids"]).shape[1]
padding_side = self.tokenizer.padding_side
if padding_side == "right":
batch["labels"] = [
list(label) + [self.label_pad_token_id] * (sequence_length - len(label)) for label in labels
]
else:
batch["labels"] = [
[self.label_pad_token_id] * (sequence_length - len(label)) + list(label) for label in labels
]
batch = {k: tf.convert_to_tensor(v, dtype=tf.int64) for k, v in batch.items()}
return batch
def numpy_call(self, features):
label_name = "label" if "label" in features[0] else "labels"
labels = [feature[label_name] for feature in features] if label_name in features[0] else None
batch = pad_without_fast_tokenizer_warning(
self.tokenizer,
features,
padding=self.padding,
max_length=self.max_length,
pad_to_multiple_of=self.pad_to_multiple_of,
# Conversion to tensors will fail if we have labels as they are not of the same length yet.
return_tensors="np" if labels is None else None,
)
if labels is None:
return batch
sequence_length = np.array(batch["input_ids"]).shape[1]
padding_side = self.tokenizer.padding_side
if padding_side == "right":
batch["labels"] = [
list(label) + [self.label_pad_token_id] * (sequence_length - len(label)) for label in labels
]
else:
batch["labels"] = [
[self.label_pad_token_id] * (sequence_length - len(label)) + list(label) for label in labels
]
batch = {k: np.array(v, dtype=np.int64) for k, v in batch.items()}
return batch
def _torch_collate_batch(examples, tokenizer, pad_to_multiple_of: Optional[int] = None):
"""Collate `examples` into a batch, using the information in `tokenizer` for padding if necessary."""
import torch
# Tensorize if necessary.
if isinstance(examples[0], (list, tuple, np.ndarray)):
examples = [torch.tensor(e, dtype=torch.long) for e in examples]
length_of_first = examples[0].size(0)
# Check if padding is necessary.
are_tensors_same_length = all(x.size(0) == length_of_first for x in examples)
if are_tensors_same_length and (pad_to_multiple_of is None or length_of_first % pad_to_multiple_of == 0):
if not isinstance(examples, torch.Tensor):
return torch.stack(examples, dim=0)
# If yes, check if we have a `pad_token`.
if tokenizer.pad_token is None:
raise ValueError(
"You are attempting to pad samples but the tokenizer you are using"
f" ({tokenizer.__class__.__name__}) does not have a pad token."
)
# Creating the full tensor and filling it with our data.
max_length = max(x.size(0) for x in examples)
if pad_to_multiple_of is not None and (max_length % pad_to_multiple_of != 0):
max_length = ((max_length // pad_to_multiple_of) + 1) * pad_to_multiple_of
result = examples[0].new_full([len(examples), max_length], tokenizer.pad_token_id)
for i, example in enumerate(examples):
if tokenizer.padding_side == "right":
result[i, : example.shape[0]] = example
else:
result[i, -example.shape[0] :] = example
return result
def _tf_collate_batch(examples, tokenizer, pad_to_multiple_of: Optional[int] = None):
import tensorflow as tf
"""Collate `examples` into a batch, using the information in `tokenizer` for padding if necessary."""
# Tensorize if necessary.
if isinstance(examples[0], (list, tuple)):
examples = [tf.convert_to_tensor(e, dtype=tf.int64) for e in examples]
# Check if padding is necessary.
length_of_first = len(examples[0])
are_tensors_same_length = all(len(x) == length_of_first for x in examples)
if are_tensors_same_length and (pad_to_multiple_of is None or length_of_first % pad_to_multiple_of == 0):
return tf.stack(examples, axis=0)
# If yes, check if we have a `pad_token`.
if tokenizer.pad_token is None:
raise ValueError(
"You are attempting to pad samples but the tokenizer you are using"
f" ({tokenizer.__class__.__name__}) does not have a pad token."
)
# Creating the full tensor and filling it with our data.
max_length = max(len(x) for x in examples)
if pad_to_multiple_of is not None and (max_length % pad_to_multiple_of != 0):
max_length = ((max_length // pad_to_multiple_of) + 1) * pad_to_multiple_of
# result = examples[0].new_full([len(examples), max_length], tokenizer.pad_token_id)
result = []
rank = tf.rank(examples[0])
paddings = np.zeros((rank, 2), dtype=np.int32)
for example in examples:
if tokenizer.padding_side == "right":
paddings[0, 1] = max_length - len(example)
else:
paddings[0, 0] = max_length - len(example)
result.append(tf.pad(example, paddings, constant_values=tokenizer.pad_token_id))
return tf.stack(result, axis=0)
def _numpy_collate_batch(examples, tokenizer, pad_to_multiple_of: Optional[int] = None):
"""Collate `examples` into a batch, using the information in `tokenizer` for padding if necessary."""
# Tensorize if necessary.
if isinstance(examples[0], (list, tuple)):
examples = [np.array(e, dtype=np.int64) for e in examples]
# Check if padding is necessary.
length_of_first = len(examples[0])
are_tensors_same_length = all(len(x) == length_of_first for x in examples)
if are_tensors_same_length and (pad_to_multiple_of is None or length_of_first % pad_to_multiple_of == 0):
return np.stack(examples, axis=0)
# If yes, check if we have a `pad_token`.
if tokenizer.pad_token is None:
raise ValueError(
"You are attempting to pad samples but the tokenizer you are using"
f" ({tokenizer.__class__.__name__}) does not have a pad token."
)
# Creating the full tensor and filling it with our data.
max_length = max(len(x) for x in examples)
if pad_to_multiple_of is not None and (max_length % pad_to_multiple_of != 0):
max_length = ((max_length // pad_to_multiple_of) + 1) * pad_to_multiple_of
result = np.full(shape=(len(examples), max_length), fill_value=tokenizer.pad_token_id, dtype=examples[0].dtype)
for i, example in enumerate(examples):
if tokenizer.padding_side == "right":
result[i, : example.shape[0]] = example
else:
result[i, -example.shape[0] :] = example
return result
@dataclass
class DataCollatorForMultipleChoice(DataCollatorMixin):
"""
Data collator that dynamically pads a batch of nested examples for multiple choice, so that all choices
of all examples have the same length.
Args:
tokenizer ([`PreTrainedTokenizer`] or [`PreTrainedTokenizerFast`]):
The tokenizer used for encoding the data.
padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `True`):
Select a strategy to pad the returned sequences according to the model's padding side and padding index
among:
- `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single sequence
is provided).
- `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
acceptable input length for the model if that argument is not provided.
- `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different
lengths).
max_length (`int`, *optional*):
Maximum length of the returned list and optionally padding length (see above).
pad_to_multiple_of (`int`, *optional*):
Pad the sequence to a multiple of the provided value.
This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
7.5 (Volta).
return_tensors (`str`, *optional*, defaults to `"pt"`):
The type of Tensor to return. Allowable values are "np", "pt" and "tf".
"""
tokenizer: PreTrainedTokenizerBase
padding: Union[bool, str, PaddingStrategy] = True
max_length: Optional[int] = None
pad_to_multiple_of: Optional[int] = None
return_tensors: str = "pt"
def torch_call(self, examples: list[dict[str, Any]]): # Refactored implementation from the docs.
import torch
# Take labels out of the examples beforehand, because they aren't nested.
label_name = "label" if "label" in examples[0] else "labels"
labels = [example.pop(label_name) for example in examples]
batch_size = len(examples)
num_choices = len(examples[0]["input_ids"])
# Go from e.g. 2 examples of 2 choices [{input_ids: [[1], [2]]}, {input_ids: [[3], [4]]}]
# to 4 examples [{input_ids: [1]}, {input_ids: [2]}] + [{input_ids: [3]}, {input_ids: [4]}]
flat_examples = sum(
([{k: v[i] for k, v in example.items()} for i in range(num_choices)] for example in examples), start=[]
)
# Pad all choices of all examples as if you're padding any other batch of examples.
batch = self.tokenizer.pad(
flat_examples,
padding=self.padding,
max_length=self.max_length,
pad_to_multiple_of=self.pad_to_multiple_of,
return_tensors="pt",
)
# Reshape from B*C x L into B x C x L, and add the labels back in.
batch = {k: v.view(batch_size, num_choices, -1) for k, v in batch.items()}
batch["labels"] = torch.tensor(labels, dtype=torch.int64)
return batch
def tf_call(self, features): # Implementation taken from the docs.
import tensorflow as tf
label_name = "label" if "label" in features[0] else "labels"
labels = [feature.pop(label_name) for feature in features]
batch_size = len(features)
num_choices = len(features[0]["input_ids"])
flattened_features = [
[{k: v[i] for k, v in feature.items()} for i in range(num_choices)] for feature in features
]
flattened_features = sum(flattened_features, []) # Sometimes written as list(chain(*flattened_features))
batch = self.tokenizer.pad(
flattened_features,
padding=self.padding,
max_length=self.max_length,
pad_to_multiple_of=self.pad_to_multiple_of,
return_tensors="tf",
)
batch = {k: tf.reshape(v, (batch_size, num_choices, -1)) for k, v in batch.items()}
batch["labels"] = tf.convert_to_tensor(labels, dtype=tf.int64)
return batch
@dataclass
class DataCollatorForSeq2Seq:
"""
Data collator that will dynamically pad the inputs received, as well as the labels.
Args:
tokenizer ([`PreTrainedTokenizer`] or [`PreTrainedTokenizerFast`]):
The tokenizer used for encoding the data.
model ([`PreTrainedModel`], *optional*):
The model that is being trained. If set and has the *prepare_decoder_input_ids_from_labels*, use it to
prepare the *decoder_input_ids*
This is useful when using *label_smoothing* to avoid calculating loss twice.
padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `True`):
Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
among:
- `True` or `'longest'` (default): Pad to the longest sequence in the batch (or no padding if only a single
sequence is provided).
- `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
acceptable input length for the model if that argument is not provided.
- `False` or `'do_not_pad'`: No padding (i.e., can output a batch with sequences of different lengths).
max_length (`int`, *optional*):
Maximum length of the returned list and optionally padding length (see above).
pad_to_multiple_of (`int`, *optional*):
If set will pad the sequence to a multiple of the provided value.
This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
7.0 (Volta).
label_pad_token_id (`int`, *optional*, defaults to -100):
The id to use when padding the labels (-100 will be automatically ignored by PyTorch loss functions).
return_tensors (`str`, *optional*, defaults to `"pt"`):
The type of Tensor to return. Allowable values are "np", "pt" and "tf".
"""
tokenizer: PreTrainedTokenizerBase
model: Optional[Any] = None
padding: Union[bool, str, PaddingStrategy] = True
max_length: Optional[int] = None
pad_to_multiple_of: Optional[int] = None
label_pad_token_id: int = -100
return_tensors: str = "pt"
def __call__(self, features, return_tensors=None):
if return_tensors is None:
return_tensors = self.return_tensors
label_name = "label" if "label" in features[0] else "labels"
labels = [feature[label_name] for feature in features] if label_name in features[0] else None
# reconvert list[None] to None if necessary
# this might occur when we pass {..., "labels": None}
if labels is not None and all(label is None for label in labels):
labels = None
non_labels_features = [{k: v for k, v in feature.items() if k != label_name} for feature in features]
# run through tokenizer without labels to ensure no side effects
batch = pad_without_fast_tokenizer_warning(
self.tokenizer,
non_labels_features,
padding=self.padding,
max_length=self.max_length,
pad_to_multiple_of=self.pad_to_multiple_of,
return_tensors=return_tensors,
)
# we have to pad the labels manually as we cannot rely on `tokenizer.pad` and we need them to be of the same length to return tensors
no_padding = self.padding is False or self.padding == PaddingStrategy.DO_NOT_PAD
if labels is not None:
if no_padding:
if isinstance(features[0][label_name], list):
batch["labels"] = list(labels)
else:
batch["labels"] = [np.concatenate([label, []]) for label in labels]
else:
max_padding = self.padding == PaddingStrategy.MAX_LENGTH and self.max_length is not None
max_label_length = max(len(l) for l in labels) if not max_padding else self.max_length
if self.pad_to_multiple_of is not None:
max_label_length = (
(max_label_length + self.pad_to_multiple_of - 1)
// self.pad_to_multiple_of
* self.pad_to_multiple_of
)
padding_side = self.tokenizer.padding_side
if isinstance(features[0][label_name], list):
batch["labels"] = [
label + [self.label_pad_token_id] * (max_label_length - len(label))
if padding_side == "right"
else [self.label_pad_token_id] * (max_label_length - len(label)) + label
for label in labels
]
else:
batch["labels"] = [
np.concatenate(
[
label,
np.array([self.label_pad_token_id] * (max_label_length - len(label)), dtype=np.int64),
]
)
if padding_side == "right"
else np.concatenate(
[
np.array([self.label_pad_token_id] * (max_label_length - len(label)), dtype=np.int64),
label,
]
)
for label in labels
]
# reintroduce side effects via tokenizer that return respective datatypes for the `return_tensors` argument
if batch.get("labels", None) is not None:
if return_tensors == "pt":
import torch
batch["labels"] = torch.tensor(batch["labels"], dtype=torch.int64)
elif return_tensors == "tf":
import tensorflow as tf
batch["labels"] = tf.constant(batch["labels"], dtype=tf.int64)
else:
batch["labels"] = np.array(batch["labels"], dtype=np.int64)
else:
batch["labels"] = None
# prepare decoder_input_ids
if (
labels is not None
and self.model is not None
and hasattr(self.model, "prepare_decoder_input_ids_from_labels")
):
decoder_input_ids = self.model.prepare_decoder_input_ids_from_labels(labels=batch["labels"])
batch["decoder_input_ids"] = decoder_input_ids
return batch
@dataclass
class DataCollatorForLanguageModeling(DataCollatorMixin):
"""
Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
are not all of the same length.
Args:
tokenizer ([`PreTrainedTokenizer`] or [`PreTrainedTokenizerFast`]):
The tokenizer used for encoding the data.
mlm (`bool`, *optional*, defaults to `True`):
Whether or not to use masked language modeling. If set to `False`, the labels are the same as the inputs
with the padding tokens ignored (by setting them to -100). Otherwise, the labels are -100 for non-masked
tokens and the value to predict for the masked token.
mlm_probability (`float`, *optional*, defaults to 0.15):
The probability with which to (randomly) mask tokens in the input, when `mlm` is set to `True`.
mask_replace_prob (`float`, *optional*, defaults to 0.8):
The probability with which masked tokens are replaced by the tokenizer's mask token (e.g., `[MASK]`).
Defaults to 0.8, meaning 80% of the masked tokens will be replaced with `[MASK]`.
Only works when `mlm` is set to `True`.
random_replace_prob (`float`, *optional*, defaults to 0.1):
The probability with which masked tokens are replaced by random tokens from the tokenizer's vocabulary.
Defaults to 0.1, meaning 10% of the masked tokens will be replaced with random tokens. The remaining
masked tokens (1 - mask_replace_prob - random_replace_prob) are left unchanged.
Only works when `mlm` is set to `True`.
pad_to_multiple_of (`int`, *optional*):
If set, will pad the sequence to a multiple of the provided value.
return_tensors (`str`):
The type of Tensor to return. Allowable values are "np", "pt" and "tf".
seed (`int`, *optional*):
The seed to use for the random number generator for masking. If not provided, the global RNG will be used.
<Tip>
For best performance, this data collator should be used with a dataset having items that are dictionaries or
BatchEncoding, with the `"special_tokens_mask"` key, as returned by a [`PreTrainedTokenizer`] or a
[`PreTrainedTokenizerFast`] with the argument `return_special_tokens_mask=True`.
<Example Options and Expectations>
1. Default Behavior:
- `mask_replace_prob=0.8`, `random_replace_prob=0.1`.
- Expect 80% of masked tokens replaced with `[MASK]`, 10% replaced with random tokens, and 10% left unchanged.
2. All masked tokens replaced by `[MASK]`:
- `mask_replace_prob=1.0`, `random_replace_prob=0.0`.
- Expect all masked tokens to be replaced with `[MASK]`. No tokens are left unchanged or replaced with random tokens.
3. No `[MASK]` replacement, only random tokens:
- `mask_replace_prob=0.0`, `random_replace_prob=1.0`.
- Expect all masked tokens to be replaced with random tokens. No `[MASK]` replacements or unchanged tokens.
4. Balanced replacement:
- `mask_replace_prob=0.5`, `random_replace_prob=0.4`.
- Expect 50% of masked tokens replaced with `[MASK]`, 40% replaced with random tokens, and 10% left unchanged.
Note:
The sum of `mask_replace_prob` and `random_replace_prob` must not exceed 1. If their sum is less than 1, the
remaining proportion will consist of masked tokens left unchanged.
</Tip>
"""
tokenizer: PreTrainedTokenizerBase
mlm: bool = True
mlm_probability: Optional[float] = 0.15
mask_replace_prob: float = 0.8
random_replace_prob: float = 0.1
pad_to_multiple_of: Optional[int] = None
tf_experimental_compile: bool = False
return_tensors: str = "pt"
seed: Optional[int] = None
def __post_init__(self):
if self.mlm:
if self.tokenizer.mask_token is None:
raise ValueError(
"This tokenizer does not have a mask token which is necessary for masked language modeling. "
"You should pass `mlm=False` to train on causal language modeling instead."
)
if self.mlm_probability is None or self.mlm_probability < 0 or self.mlm_probability > 1:
raise ValueError("mlm_probability should be between 0 and 1.")
self.mlm_probability = float(self.mlm_probability)
if self.mask_replace_prob + self.random_replace_prob > 1:
raise ValueError("The sum of mask_replace_prob and random_replace_prob should not exceed 1")
if self.mask_replace_prob < 0 or self.mask_replace_prob > 1:
raise ValueError("mask_replace_prob should be between 0 and 1.")
if self.random_replace_prob < 0 or self.random_replace_prob > 1:
raise ValueError("random_replace_prob should be between 0 and 1.")
self.mask_replace_prob = float(self.mask_replace_prob)
self.random_replace_prob = float(self.random_replace_prob)
if self.tf_experimental_compile:
import tensorflow as tf
self.tf_mask_tokens = tf.function(self.tf_mask_tokens, jit_compile=True)
self.generator = None
def get_generator(self, seed):
if self.return_tensors == "pt":
import torch
return torch.Generator().manual_seed(seed)
elif self.return_tensors == "tf":
import tensorflow as tf
return tf.random.Generator.from_seed(seed)
else:
import numpy as np
return np.random.default_rng(seed)
def create_rng(self):
if mp.current_process().name == "MainProcess":
# If we are in the main process, we create a generator object with the seed
self.generator = self.get_generator(self.seed)
else:
# If we are in a worker process (i.e using multiprocessing), we need to set a unique seed for each
# worker's generator, generated as the main seed + the worker's ID.
# (https://pytorch.org/docs/stable/data.html#randomness-in-multi-process-data-loading)
# Only PyTorch DataLoader allows us to access the worker ID, and so we check for this.
# For other frameworks, we will throw an error.
import torch
worker_info = torch.utils.data.get_worker_info()
if worker_info is None:
error_string = (
"Worker process information is not available for seeding the generator. This may be because",
"you are using multiprocessing without using a PyTorch DataLoader. The `seed` parameter can",
"only be used when using multiprocessing with a PyTorch DataLoader. Please either use a",
"single process or use a PyTorch DataLoader with multiple workers.",
)
raise ValueError(error_string)
self.generator = self.get_generator(self.seed + worker_info.id)
@staticmethod
def tf_bernoulli(shape, probability, generator=None):
import tensorflow as tf
prob_matrix = tf.fill(shape, probability)
# if generator exists, use it to generate the random numbers
# otherwise, use the global RNG
if generator:
return tf.cast(prob_matrix - generator.uniform(shape, 0, 1) >= 0, tf.bool)
else:
return tf.cast(prob_matrix - tf.random.uniform(shape, 0, 1) >= 0, tf.bool)
def tf_mask_tokens(
self, inputs: Any, vocab_size, mask_token_id, special_tokens_mask: Optional[Any] = None
) -> tuple[Any, Any]:
"""
Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
"""
import tensorflow as tf
mask_token_id = tf.cast(mask_token_id, inputs.dtype)
input_shape = tf.shape(inputs)
# 1 for a special token, 0 for a normal token in the special tokens mask
# We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
masked_indices = self.tf_bernoulli(input_shape, self.mlm_probability, self.generator) & ~special_tokens_mask
# Replace unmasked indices with -100 in the labels since we only compute loss on masked tokens
labels = tf.where(masked_indices, inputs, -100)
# mask_replace_prob% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
indices_replaced = self.tf_bernoulli(input_shape, self.mask_replace_prob, self.generator) & masked_indices
inputs = tf.where(indices_replaced, mask_token_id, inputs)
if self.mask_replace_prob == 1 or self.random_replace_prob == 0:
return inputs, labels
remaining_prob = 1 - self.mask_replace_prob
# scaling the random_replace_prob to the remaining probability for example if
# mask_replace_prob = 0.8 and random_replace_prob = 0.1,
# then random_replace_prob_scaled = 0.1 / 0.2 = 0.5
random_replace_prob_scaled = self.random_replace_prob / remaining_prob
# random_replace_prob% of the time, we replace masked input tokens with random word
indices_random = (
self.tf_bernoulli(input_shape, random_replace_prob_scaled, self.generator)
& masked_indices
& ~indices_replaced
)
if self.generator:
random_words = self.generator.uniform(input_shape, maxval=vocab_size, dtype=inputs.dtype)
else:
random_words = tf.random.uniform(input_shape, maxval=vocab_size, dtype=inputs.dtype)
inputs = tf.where(indices_random, random_words, inputs)
# The rest of the time ((1-random_replace_prob-mask_replace_prob)% of the time) we keep the masked input tokens unchanged
return inputs, labels
def tf_call(self, examples: list[Union[list[int], Any, dict[str, Any]]]) -> dict[str, Any]:
import tensorflow as tf
if self.seed and self.generator is None:
# If we have a seed, we need to create a generator object. Subsequent calls to this function will use the same generator.
# If no seed supplied, we will use the global RNG
self.create_rng()
# Handle dict or lists with proper padding and conversion to tensor.
if isinstance(examples[0], Mapping):
batch = pad_without_fast_tokenizer_warning(
self.tokenizer, examples, return_tensors="tf", pad_to_multiple_of=self.pad_to_multiple_of
)
else:
batch = {
"input_ids": _tf_collate_batch(examples, self.tokenizer, pad_to_multiple_of=self.pad_to_multiple_of)
}
# If special token mask has been preprocessed, pop it from the dict.
special_tokens_mask = batch.pop("special_tokens_mask", None)
if self.mlm:
if special_tokens_mask is None:
special_tokens_mask = [
self.tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True)
for val in batch["input_ids"].numpy().tolist()
]
# Cannot directly create as bool
special_tokens_mask = tf.cast(tf.convert_to_tensor(special_tokens_mask, dtype=tf.int64), tf.bool)
else:
special_tokens_mask = tf.cast(special_tokens_mask, tf.bool)
batch["input_ids"], batch["labels"] = self.tf_mask_tokens(
tf.cast(batch["input_ids"], tf.int64),
special_tokens_mask=special_tokens_mask,
mask_token_id=self.tokenizer.mask_token_id,
vocab_size=len(self.tokenizer),
)
else:
labels = batch["input_ids"]
if self.tokenizer.pad_token_id is not None:
# Replace self.tokenizer.pad_token_id with -100
labels = tf.where(labels == self.tokenizer.pad_token_id, -100, labels)
else:
labels = tf.identity(labels) # Makes a copy, just in case
batch["labels"] = labels
return batch
def torch_call(self, examples: list[Union[list[int], Any, dict[str, Any]]]) -> dict[str, Any]:
# Handle dict or lists with proper padding and conversion to tensor.
if self.seed and self.generator is None:
# If we have a seed, we need to create a generator object. Subsequent calls to this function will use the same generator.
# If no seed supplied, we will use the global RNG
self.create_rng()
if isinstance(examples[0], Mapping):
batch = pad_without_fast_tokenizer_warning(
self.tokenizer, examples, return_tensors="pt", pad_to_multiple_of=self.pad_to_multiple_of
)
else:
batch = {
"input_ids": _torch_collate_batch(examples, self.tokenizer, pad_to_multiple_of=self.pad_to_multiple_of)
}
# If special token mask has been preprocessed, pop it from the dict.
special_tokens_mask = batch.pop("special_tokens_mask", None)
if self.mlm:
batch["input_ids"], batch["labels"] = self.torch_mask_tokens(
batch["input_ids"], special_tokens_mask=special_tokens_mask
)
else:
labels = batch["input_ids"].clone()
if self.tokenizer.pad_token_id is not None:
labels[labels == self.tokenizer.pad_token_id] = -100
batch["labels"] = labels
return batch
def torch_mask_tokens(self, inputs: Any, special_tokens_mask: Optional[Any] = None) -> tuple[Any, Any]:
"""
Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
"""
import torch
labels = inputs.clone()
# We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
probability_matrix = torch.full(labels.shape, self.mlm_probability)
if special_tokens_mask is None:
special_tokens_mask = [
self.tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()
]
special_tokens_mask = torch.tensor(special_tokens_mask, dtype=torch.bool)
else:
special_tokens_mask = special_tokens_mask.bool()
probability_matrix.masked_fill_(special_tokens_mask, value=0.0)
masked_indices = torch.bernoulli(probability_matrix, generator=self.generator).bool()
labels[~masked_indices] = -100 # We only compute loss on masked tokens
# mask_replace_prob% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
indices_replaced = (
torch.bernoulli(torch.full(labels.shape, self.mask_replace_prob), generator=self.generator).bool()
& masked_indices
)
inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
if self.mask_replace_prob == 1 or self.random_replace_prob == 0:
return inputs, labels
remaining_prob = 1 - self.mask_replace_prob
# scaling the random_replace_prob to the remaining probability for example if
# mask_replace_prob = 0.8 and random_replace_prob = 0.1,
# then random_replace_prob_scaled = 0.1 / 0.2 = 0.5
random_replace_prob_scaled = self.random_replace_prob / remaining_prob
# random_replace_prob% of the time, we replace masked input tokens with random word
indices_random = (
torch.bernoulli(torch.full(labels.shape, random_replace_prob_scaled), generator=self.generator).bool()
& masked_indices
& ~indices_replaced
)
random_words = torch.randint(len(self.tokenizer), labels.shape, dtype=torch.long, generator=self.generator)
inputs[indices_random] = random_words[indices_random]
# The rest of the time ((1-random_replace_prob-mask_replace_prob)% of the time) we keep the masked input tokens unchanged
return inputs, labels
def numpy_call(self, examples: list[Union[list[int], Any, dict[str, Any]]]) -> dict[str, Any]:
# Handle dict or lists with proper padding and conversion to tensor.
if self.seed and self.generator is None:
# If we have a seed, we need to create a generator object. Subsequent calls to this function will use the same generator.
# If no seed supplied, we will use the global RNG
self.create_rng()
if isinstance(examples[0], Mapping):
batch = pad_without_fast_tokenizer_warning(
self.tokenizer, examples, return_tensors="np", pad_to_multiple_of=self.pad_to_multiple_of
)
else:
batch = {
"input_ids": _numpy_collate_batch(examples, self.tokenizer, pad_to_multiple_of=self.pad_to_multiple_of)
}
# If special token mask has been preprocessed, pop it from the dict.
special_tokens_mask = batch.pop("special_tokens_mask", None)
if self.mlm:
batch["input_ids"], batch["labels"] = self.numpy_mask_tokens(
batch["input_ids"], special_tokens_mask=special_tokens_mask
)
else:
labels = np.copy(batch["input_ids"])
if self.tokenizer.pad_token_id is not None:
labels[labels == self.tokenizer.pad_token_id] = -100
batch["labels"] = labels
return batch
def numpy_mask_tokens(self, inputs: Any, special_tokens_mask: Optional[Any] = None) -> tuple[Any, Any]:
"""
Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
"""
labels = np.copy(inputs)
# We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
probability_matrix = np.full(labels.shape, self.mlm_probability)
if special_tokens_mask is None:
special_tokens_mask = [
self.tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()
]
special_tokens_mask = np.array(special_tokens_mask, dtype=bool)
else:
special_tokens_mask = special_tokens_mask.astype(bool)
probability_matrix[special_tokens_mask] = 0
# Numpy doesn't have bernoulli, so we use a binomial with 1 trial
if self.generator:
masked_indices = self.generator.binomial(1, probability_matrix, size=probability_matrix.shape).astype(bool)
else:
masked_indices = np.random.binomial(1, probability_matrix, size=probability_matrix.shape).astype(bool)
labels[~masked_indices] = -100 # We only compute loss on masked tokens
# mask_replace_prob% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
if self.generator:
indices_replaced = (
self.generator.binomial(1, self.mask_replace_prob, size=labels.shape).astype(bool) & masked_indices
)
else:
indices_replaced = (
np.random.binomial(1, self.mask_replace_prob, size=labels.shape).astype(bool) & masked_indices
)
inputs[indices_replaced] = self.tokenizer.mask_token_id
if self.mask_replace_prob == 1 or self.random_replace_prob == 0:
return inputs, labels
remaining_prob = 1 - self.mask_replace_prob
# scaling the random_replace_prob to the remaining probability for example if
# mask_replace_prob = 0.8 and random_replace_prob = 0.1,
# then random_replace_prob_scaled = 0.1 / 0.2 = 0.5
random_replace_prob_scaled = self.random_replace_prob / remaining_prob
if self.generator:
indices_random = (
self.generator.binomial(1, random_replace_prob_scaled, size=labels.shape).astype(bool)
& masked_indices
& ~indices_replaced
)
random_words = self.generator.integers(
low=0, high=len(self.tokenizer), size=np.count_nonzero(indices_random), dtype=np.int64
)
else:
indices_random = (
np.random.binomial(1, random_replace_prob_scaled, size=labels.shape).astype(bool)
& masked_indices
& ~indices_replaced
)
random_words = np.random.randint(
low=0, high=len(self.tokenizer), size=np.count_nonzero(indices_random), dtype=np.int64
)
inputs[indices_random] = random_words
# The rest of the time (10% of the time) we keep the masked input tokens unchanged
return inputs, labels
@dataclass
class DataCollatorForWholeWordMask(DataCollatorForLanguageModeling):
"""
Data collator used for language modeling that masks entire words.
- collates batches of tensors, honoring their tokenizer's pad_token
- preprocesses batches for masked language modeling
<Tip>
This collator relies on details of the implementation of subword tokenization by [`BertTokenizer`], specifically
that subword tokens are prefixed with *##*. For tokenizers that do not adhere to this scheme, this collator will
produce an output that is roughly equivalent to [`.DataCollatorForLanguageModeling`].
</Tip>"""
def torch_call(self, examples: list[Union[list[int], Any, dict[str, Any]]]) -> dict[str, Any]:
if self.seed and self.generator is None:
# If we have a seed, we need to create a generator object. Subsequent calls to this function will use the same generator.
# If no seed supplied, we will use the global RNG
self.create_rng()
if isinstance(examples[0], Mapping):
input_ids = [e["input_ids"] for e in examples]
else:
input_ids = examples
examples = [{"input_ids": e} for e in examples]
batch_input = _torch_collate_batch(input_ids, self.tokenizer, pad_to_multiple_of=self.pad_to_multiple_of)
mask_labels = []
for e in examples:
ref_tokens = []
for id in tolist(e["input_ids"]):
token = self.tokenizer._convert_id_to_token(id)
ref_tokens.append(token)
# For Chinese tokens, we need extra inf to mark sub-word, e.g [喜,欢]-> [喜，##欢]
if "chinese_ref" in e:
ref_pos = tolist(e["chinese_ref"])
len_seq = len(e["input_ids"])
for i in range(len_seq):
if i in ref_pos:
ref_tokens[i] = "##" + ref_tokens[i]
mask_labels.append(self._whole_word_mask(ref_tokens))
batch_mask = _torch_collate_batch(mask_labels, self.tokenizer, pad_to_multiple_of=self.pad_to_multiple_of)
inputs, labels = self.torch_mask_tokens(batch_input, batch_mask)
return {"input_ids": inputs, "labels": labels}
def tf_call(self, examples: list[Union[list[int], Any, dict[str, Any]]]) -> dict[str, Any]:
import tensorflow as tf
if self.seed and self.generator is None:
# If we have a seed, we need to create a generator object. Subsequent calls to this function will use the same generator.
# If no seed supplied, we will use the global RNG
self.create_rng()
if isinstance(examples[0], Mapping):
input_ids = [e["input_ids"] for e in examples]
else:
input_ids = examples
examples = [{"input_ids": e} for e in examples]
batch_input = _tf_collate_batch(input_ids, self.tokenizer, pad_to_multiple_of=self.pad_to_multiple_of)
mask_labels = []
for e in examples:
ref_tokens = []
for id in tolist(e["input_ids"]):
token = self.tokenizer._convert_id_to_token(id)
ref_tokens.append(token)
# For Chinese tokens, we need extra inf to mark sub-word, e.g [喜,欢]-> [喜，##欢]
if "chinese_ref" in e:
ref_pos = tolist(e["chinese_ref"])
len_seq = len(e["input_ids"])
for i in range(len_seq):
if i in ref_pos:
ref_tokens[i] = "##" + ref_tokens[i]
mask_labels.append(self._whole_word_mask(ref_tokens))
batch_mask = _tf_collate_batch(mask_labels, self.tokenizer, pad_to_multiple_of=self.pad_to_multiple_of)
inputs, labels = self.tf_mask_tokens(tf.cast(batch_input, tf.int64), batch_mask)
return {"input_ids": inputs, "labels": labels}
def numpy_call(self, examples: list[Union[list[int], Any, dict[str, Any]]]) -> dict[str, Any]:
if self.seed and self.generator is None:
# If we have a seed, we need to create a generator object. Subsequent calls to this function will use the same generator.
# If no seed supplied, we will use the global RNG
self.create_rng()
if isinstance(examples[0], Mapping):
input_ids = [e["input_ids"] for e in examples]
else:
input_ids = examples
examples = [{"input_ids": e} for e in examples]
batch_input = _numpy_collate_batch(input_ids, self.tokenizer, pad_to_multiple_of=self.pad_to_multiple_of)
mask_labels = []
for e in examples:
ref_tokens = []
for id in tolist(e["input_ids"]):
token = self.tokenizer._convert_id_to_token(id)
ref_tokens.append(token)
# For Chinese tokens, we need extra inf to mark sub-word, e.g [喜,欢]-> [喜，##欢]
if "chinese_ref" in e:
ref_pos = tolist(e["chinese_ref"])
len_seq = len(e["input_ids"])
for i in range(len_seq):
if i in ref_pos:
ref_tokens[i] = "##" + ref_tokens[i]
mask_labels.append(self._whole_word_mask(ref_tokens))
batch_mask = _numpy_collate_batch(mask_labels, self.tokenizer, pad_to_multiple_of=self.pad_to_multiple_of)
inputs, labels = self.numpy_mask_tokens(batch_input, batch_mask)
return {"input_ids": inputs, "labels": labels}
def _shuffle(self, cand_indexes):
# if no seed, just use random's shuffle
if self.seed is None:
random.shuffle(cand_indexes)
return cand_indexes
# if seed is provided, use the generator to shuffle
if self.return_tensors == "pt":
import torch
indices = torch.randperm(len(cand_indexes), generator=self.generator)
return [cand_indexes[i] for i in indices]
elif self.return_tensors == "tf":
import tensorflow as tf
seed = self.generator.make_seeds(2)[0]
indices = tf.random.experimental.stateless_shuffle(tf.range(len(cand_indexes)), seed=seed).numpy().tolist()
return [cand_indexes[i] for i in indices]
elif self.return_tensors == "np":
self.generator.shuffle(cand_indexes)
return cand_indexes
def _whole_word_mask(self, input_tokens: list[str], max_predictions=512):
"""
Get 0/1 labels for masked tokens with whole word mask proxy
"""
if not isinstance(self.tokenizer, (BertTokenizer, BertTokenizerFast)):
warnings.warn(
"DataCollatorForWholeWordMask is only suitable for BertTokenizer-like tokenizers. "
"Please refer to the documentation for more information."
)
cand_indexes = []
for i, token in enumerate(input_tokens):
if token == "[CLS]" or token == "[SEP]":
continue
if len(cand_indexes) >= 1 and token.startswith("##"):
cand_indexes[-1].append(i)
else:
cand_indexes.append([i])
cand_indexes = self._shuffle(cand_indexes)
num_to_predict = min(max_predictions, max(1, int(round(len(input_tokens) * self.mlm_probability))))
masked_lms = []
covered_indexes = set()
for index_set in cand_indexes:
if len(masked_lms) >= num_to_predict:
break
# If adding a whole-word mask would exceed the maximum number of
# predictions, then just skip this candidate.
if len(masked_lms) + len(index_set) > num_to_predict:
continue
for index in index_set:
covered_indexes.add(index)
masked_lms.append(index)
if len(covered_indexes) != len(masked_lms):
raise ValueError("Length of covered_indexes is not equal to length of masked_lms.")
mask_labels = [1 if i in covered_indexes else 0 for i in range(len(input_tokens))]
return mask_labels
def torch_mask_tokens(self, inputs: Any, mask_labels: Any) -> tuple[Any, Any]:
"""
Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original. Set
'mask_labels' means we use whole word mask (wwm), we directly mask idxs according to it's ref.
"""
import torch
if self.tokenizer.mask_token is None:
raise ValueError(
"This tokenizer does not have a mask token which is necessary for masked language modeling. Remove the"
" --mlm flag if you want to use this tokenizer."
)
labels = inputs.clone()
# We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa)
probability_matrix = mask_labels
special_tokens_mask = [
self.tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()
]
probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)
if self.tokenizer.pad_token is not None:
padding_mask = labels.eq(self.tokenizer.pad_token_id)
probability_matrix.masked_fill_(padding_mask, value=0.0)
masked_indices = probability_matrix.bool()
labels[~masked_indices] = -100 # We only compute loss on masked tokens
# mask_replace_prob% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
indices_replaced = (
torch.bernoulli(torch.full(labels.shape, self.mask_replace_prob), generator=self.generator).bool()
& masked_indices
)
inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
if self.mask_replace_prob == 1 or self.random_replace_prob == 0:
return inputs, labels
remaining_prob = 1 - self.mask_replace_prob
# scaling the random_replace_prob to the remaining probability for example if
# mask_replace_prob = 0.8 and random_replace_prob = 0.1,
# then random_replace_prob_scaled = 0.1 / 0.2 = 0.5
random_replace_prob_scaled = self.random_replace_prob / remaining_prob
# random_replacement_prob% of the time, we replace masked input tokens with random word
indices_random = (
torch.bernoulli(torch.full(labels.shape, random_replace_prob_scaled), generator=self.generator).bool()
& masked_indices
& ~indices_replaced
)
random_words = torch.randint(len(self.tokenizer), labels.shape, dtype=torch.long, generator=self.generator)
inputs[indices_random] = random_words[indices_random]
# The rest of the time ((1-random_replacement_prob-mask_replace_prob)% of the time) we keep the masked input tokens unchanged
return inputs, labels
def tf_mask_tokens(self, inputs: Any, mask_labels: Any) -> tuple[Any, Any]:
"""
Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original. Set
'mask_labels' means we use whole word mask (wwm), we directly mask idxs according to it's ref.
"""
import tensorflow as tf
input_shape = tf.shape(inputs)
if self.tokenizer.mask_token is None:
raise ValueError(
"This tokenizer does not have a mask token which is necessary for masked language modeling. Remove the"
" --mlm flag if you want to use this tokenizer."
)
labels = tf.identity(inputs)
# We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa)
masked_indices = tf.cast(mask_labels, tf.bool)
special_tokens_mask = [
self.tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels
]
masked_indices = masked_indices & ~tf.cast(special_tokens_mask, dtype=tf.bool)
if self.tokenizer.pad_token is not None:
padding_mask = inputs == self.tokenizer.pad_token_id
masked_indices = masked_indices & ~padding_mask
# Replace unmasked indices with -100 in the labels since we only compute loss on masked tokens
labels = tf.where(masked_indices, inputs, -100)
# mask_replace_prob% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
indices_replaced = self.tf_bernoulli(input_shape, self.mask_replace_prob, self.generator) & masked_indices
inputs = tf.where(indices_replaced, self.tokenizer.mask_token_id, inputs)
if self.mask_replace_prob == 1 or self.random_replace_prob == 0:
return inputs, labels
remaining_prob = 1 - self.mask_replace_prob
# scaling the random_replace_prob to the remaining probability for example if
# mask_replace_prob = 0.8 and random_replace_prob = 0.1,
# then random_replace_prob_scaled = 0.1 / 0.2 = 0.5
random_replace_prob_scaled = self.random_replace_prob / remaining_prob
# random_replace_prob% of the time, we replace masked input tokens with random word
indices_random = (
self.tf_bernoulli(input_shape, random_replace_prob_scaled, self.generator)
& masked_indices
& ~indices_replaced
)
if self.generator:
random_words = self.generator.uniform(input_shape, maxval=len(self.tokenizer), dtype=tf.int64)
else:
random_words = tf.random.uniform(input_shape, maxval=len(self.tokenizer), dtype=tf.int64)
inputs = tf.where(indices_random, random_words, inputs)
# The rest of the time ((1-mask_replace_prob-random_replace_prob)% of the time) we keep the masked input tokens unchanged
return inputs, labels
def numpy_mask_tokens(self, inputs: Any, mask_labels: Any) -> tuple[Any, Any]:
"""
Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original. Set
'mask_labels' means we use whole word mask (wwm), we directly mask idxs according to it's ref.
"""
if self.tokenizer.mask_token is None:
raise ValueError(
"This tokenizer does not have a mask token which is necessary for masked language modeling. Remove the"
" --mlm flag if you want to use this tokenizer."
)
labels = np.copy(inputs)
# We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa)
masked_indices = mask_labels.astype(bool)
special_tokens_mask = [
self.tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()
]
masked_indices[np.array(special_tokens_mask, dtype=bool)] = 0
if self.tokenizer.pad_token is not None:
padding_mask = labels == self.tokenizer.pad_token_id
masked_indices[padding_mask] = 0
labels[~masked_indices] = -100 # We only compute loss on masked tokens
# mask_replacement_prob% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
if self.generator:
indices_replaced = (
self.generator.binomial(1, self.mask_replace_prob, size=labels.shape).astype(bool) & masked_indices
)
else:
indices_replaced = (
np.random.binomial(1, self.mask_replace_prob, size=labels.shape).astype(bool) & masked_indices
)
inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
if self.mask_replace_prob == 1 or self.random_replace_prob == 0:
return inputs, labels
remaining_prob = 1 - self.mask_replace_prob
# scaling the random_replace_prob to the remaining probability for example if
# mask_replace_prob = 0.8 and random_replace_prob = 0.1,
# then random_replace_prob_scaled = 0.1 / 0.2 = 0.5
random_replace_prob_scaled = self.random_replace_prob / remaining_prob
if self.generator:
indices_random = (
self.generator.binomial(1, random_replace_prob_scaled, size=labels.shape).astype(bool)
& masked_indices
& ~indices_replaced
)
random_words = self.generator.integers(low=0, high=len(self.tokenizer), size=labels.shape, dtype=np.int64)
else:
indices_random = (
np.random.binomial(1, random_replace_prob_scaled, size=labels.shape).astype(bool)
& masked_indices
& ~indices_replaced
)
random_words = np.random.randint(low=0, high=len(self.tokenizer), size=labels.shape, dtype=np.int64)
inputs[indices_random] = random_words[indices_random]
# The rest of the time ((1-mask_replace_prob-random_replace_prob)% of the time) we keep the masked input tokens unchanged
return inputs, labels
def tolist(x):
if isinstance(x, list):
return x
elif hasattr(x, "numpy"): # Checks for TF tensors without needing the import
x = x.numpy()
return x.tolist()
@dataclass
class DataCollatorForSOP(DataCollatorForLanguageModeling):
"""
Data collator used for sentence order prediction task.
- collates batches of tensors, honoring their tokenizer's pad_token
- preprocesses batches for both masked language modeling and sentence order prediction
"""
def __init__(self, *args, **kwargs):
warnings.warn(
"DataCollatorForSOP is deprecated and will be removed in a future version, you can now use "
"DataCollatorForLanguageModeling instead.",
FutureWarning,
)
def __call__(self, examples: list[dict[str, Any]]) -> dict[str, Any]:
import torch
from torch.nn.utils.rnn import pad_sequence
input_ids = [example["input_ids"] for example in examples]
input_ids = _torch_collate_batch(input_ids, self.tokenizer)
input_ids, labels, attention_mask = self.mask_tokens(input_ids)
token_type_ids = [example["token_type_ids"] for example in examples]
# size of segment_ids varied because randomness, padding zero to the end as the original implementation
token_type_ids = pad_sequence(token_type_ids, batch_first=True, padding_value=self.tokenizer.pad_token_id)
sop_label_list = [example["sentence_order_label"] for example in examples]
sentence_order_label = torch.stack(sop_label_list)
return {
"input_ids": input_ids,
"labels": labels,
"attention_mask": attention_mask,
"token_type_ids": token_type_ids,
"sentence_order_label": sentence_order_label,
}
def mask_tokens(self, inputs: Any) -> tuple[Any, Any, Any]:
"""
Prepare masked tokens inputs/labels/attention_mask for masked language modeling: 80% MASK, 10% random, 10%
original. N-gram not applied yet.
"""
import torch
if self.tokenizer.mask_token is None:
raise ValueError(
"This tokenizer does not have a mask token which is necessary for masked language modeling. Remove the"
" --mlm flag if you want to use this tokenizer."
)
labels = inputs.clone()
# We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa)
probability_matrix = torch.full(labels.shape, self.mlm_probability)
special_tokens_mask = [
self.tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()
]
probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)
if self.tokenizer.pad_token is not None:
padding_mask = labels.eq(self.tokenizer.pad_token_id)
probability_matrix.masked_fill_(padding_mask, value=0.0)
masked_indices = torch.bernoulli(probability_matrix).bool()
# probability be `1` (masked), however in albert model attention mask `0` means masked, revert the value
attention_mask = (~masked_indices).float()
if self.tokenizer.pad_token is not None:
attention_padding_mask = labels.eq(self.tokenizer.pad_token_id)
attention_mask.masked_fill_(attention_padding_mask, value=1.0)
labels[~masked_indices] = -100 # We only compute loss on masked tokens, -100 is default for CE compute
# 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
indices_replaced = torch.bernoulli(torch.full(labels.shape, 0.8)).bool() & masked_indices
inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
# 10% of the time, we replace masked input tokens with random word
indices_random = torch.bernoulli(torch.full(labels.shape, 0.5)).bool() & masked_indices & ~indices_replaced
random_words = torch.randint(len(self.tokenizer), labels.shape, dtype=torch.long)
inputs[indices_random] = random_words[indices_random]
# The rest of the time (10% of the time) we keep the masked input tokens unchanged
return inputs, labels, attention_mask
@dataclass
class DataCollatorForPermutationLanguageModeling(DataCollatorMixin):
"""
Data collator used for permutation language modeling.
- collates batches of tensors, honoring their tokenizer's pad_token
- preprocesses batches for permutation language modeling with procedures specific to XLNet
"""
tokenizer: PreTrainedTokenizerBase
plm_probability: float = 1 / 6
max_span_length: int = 5 # maximum length of a span of masked tokens
return_tensors: str = "pt"
def torch_call(self, examples: list[Union[list[int], Any, dict[str, Any]]]) -> dict[str, Any]:
if isinstance(examples[0], Mapping):
examples = [e["input_ids"] for e in examples]
batch = _torch_collate_batch(examples, self.tokenizer)
inputs, perm_mask, target_mapping, labels = self.torch_mask_tokens(batch)
return {"input_ids": inputs, "perm_mask": perm_mask, "target_mapping": target_mapping, "labels": labels}
def tf_call(self, examples: list[Union[list[int], Any, dict[str, Any]]]) -> dict[str, Any]:
if isinstance(examples[0], Mapping):
examples = [e["input_ids"] for e in examples]
batch = _tf_collate_batch(examples, self.tokenizer)
inputs, perm_mask, target_mapping, labels = self.tf_mask_tokens(batch)
return {"input_ids": inputs, "perm_mask": perm_mask, "target_mapping": target_mapping, "labels": labels}
def numpy_call(self, examples: list[Union[list[int], Any, dict[str, Any]]]) -> dict[str, Any]:
if isinstance(examples[0], Mapping):
examples = [e["input_ids"] for e in examples]
batch = _numpy_collate_batch(examples, self.tokenizer)
inputs, perm_mask, target_mapping, labels = self.numpy_mask_tokens(batch)
return {"input_ids": inputs, "perm_mask": perm_mask, "target_mapping": target_mapping, "labels": labels}
def torch_mask_tokens(self, inputs: Any) -> tuple[Any, Any, Any, Any]:
"""
The masked tokens to be predicted for a particular sequence are determined by the following algorithm:
0. Start from the beginning of the sequence by setting `cur_len = 0` (number of tokens processed so far).
1. Sample a `span_length` from the interval `[1, max_span_length]` (length of span of tokens to be masked)
2. Reserve a context of length `context_length = span_length / plm_probability` to surround span to be
masked
3. Sample a starting point `start_index` from the interval `[cur_len, cur_len + context_length -
span_length]` and mask tokens `start_index:start_index + span_length`
4. Set `cur_len = cur_len + context_length`. If `cur_len < max_len` (i.e. there are tokens remaining in the
sequence to be processed), repeat from Step 1.
"""
import torch
if self.tokenizer.mask_token is None:
raise ValueError(
"This tokenizer does not have a mask token which is necessary for permutation language modeling."
" Please add a mask token if you want to use this tokenizer."
)
if inputs.size(1) % 2 != 0:
raise ValueError(
"This collator requires that sequence lengths be even to create a leakage-free perm_mask. Please see"
" relevant comments in source code for details."
)
labels = inputs.clone()
# Creating the mask and target_mapping tensors
masked_indices = torch.full(labels.shape, 0, dtype=torch.bool)
target_mapping = torch.zeros((labels.size(0), labels.size(1), labels.size(1)), dtype=torch.float32)
for i in range(labels.size(0)):
# Start from the beginning of the sequence by setting `cur_len = 0` (number of tokens processed so far).
cur_len = 0
max_len = labels.size(1)
while cur_len < max_len:
# Sample a `span_length` from the interval `[1, max_span_length]` (length of span of tokens to be masked)
span_length = torch.randint(1, self.max_span_length + 1, (1,)).item()
# Reserve a context of length `context_length = span_length / plm_probability` to surround the span to be masked
context_length = int(span_length / self.plm_probability)
# Sample a starting point `start_index` from the interval `[cur_len, cur_len + context_length - span_length]` and mask tokens `start_index:start_index + span_length`
start_index = cur_len + torch.randint(context_length - span_length + 1, (1,)).item()
masked_indices[i, start_index : start_index + span_length] = 1
# Set `cur_len = cur_len + context_length`
cur_len += context_length
# Since we're replacing non-masked tokens with -100 in the labels tensor instead of skipping them altogether,
# the i-th predict corresponds to the i-th token.
target_mapping[i] = torch.eye(labels.size(1))
special_tokens_mask = torch.tensor(
[self.tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()],
dtype=torch.bool,
)
masked_indices.masked_fill_(special_tokens_mask, value=0.0)
if self.tokenizer.pad_token is not None:
padding_mask = labels.eq(self.tokenizer.pad_token_id)
masked_indices.masked_fill_(padding_mask, value=0.0)
# Mask indicating non-functional tokens, where functional tokens are [SEP], [CLS], padding, etc.
non_func_mask = ~(padding_mask | special_tokens_mask)
inputs[masked_indices] = self.tokenizer.mask_token_id
labels[~masked_indices] = -100 # We only compute loss on masked tokens
perm_mask = torch.zeros((labels.size(0), labels.size(1), labels.size(1)), dtype=torch.float32)
for i in range(labels.size(0)):
# Generate permutation indices i.e. sample a random factorisation order for the sequence. This will
# determine which tokens a given token can attend to (encoded in `perm_mask`).
# Note: Length of token sequence being permuted has to be less than or equal to reused sequence length
# (see documentation for `mems`), otherwise information may leak through due to reuse. In this implementation,
# we assume that reused length is half of sequence length and permutation length is equal to reused length.
# This requires that the sequence length be even.
# Create a linear factorisation order
perm_index = torch.arange(labels.size(1))
# Split this into two halves, assuming that half the sequence is reused each time
perm_index = perm_index.reshape((-1, labels.size(1) // 2)).transpose(0, 1)
# Permute the two halves such that they do not cross over
perm_index = perm_index[torch.randperm(labels.size(1) // 2)]
# Flatten this out into the desired permuted factorisation order
perm_index = torch.flatten(perm_index.transpose(0, 1))
# Set the permutation indices of non-masked (non-functional) tokens to the
# smallest index (-1) so that:
# (1) They can be seen by all other positions
# (2) They cannot see masked positions, so there won't be information leak
perm_index.masked_fill_(~masked_indices[i] & non_func_mask[i], -1)
# The logic for whether the i-th token can attend on the j-th token based on the factorisation order:
# 0 (can attend): If perm_index[i] > perm_index[j] or j is neither masked nor a functional token
# 1 (cannot attend): If perm_index[i] <= perm_index[j] and j is either masked or a functional token
perm_mask[i] = (
perm_index.reshape((labels.size(1), 1)) <= perm_index.reshape((1, labels.size(1)))
) & masked_indices[i]
return inputs.long(), perm_mask, target_mapping, labels.long()
def tf_mask_tokens(self, inputs: Any) -> tuple[Any, Any, Any, Any]:
"""
The masked tokens to be predicted for a particular sequence are determined by the following algorithm:
0. Start from the beginning of the sequence by setting `cur_len = 0` (number of tokens processed so far).
1. Sample a `span_length` from the interval `[1, max_span_length]` (length of span of tokens to be masked)
2. Reserve a context of length `context_length = span_length / plm_probability` to surround span to be
masked
3. Sample a starting point `start_index` from the interval `[cur_len, cur_len + context_length -
span_length]` and mask tokens `start_index:start_index + span_length`
4. Set `cur_len = cur_len + context_length`. If `cur_len < max_len` (i.e. there are tokens remaining in the
sequence to be processed), repeat from Step 1.
"""
import tensorflow as tf
if self.tokenizer.mask_token is None:
raise ValueError(
"This tokenizer does not have a mask token which is necessary for permutation language modeling."
" Please add a mask token if you want to use this tokenizer."
)
if tf.shape(inputs)[1] % 2 != 0:
raise ValueError(
"This collator requires that sequence lengths be even to create a leakage-free perm_mask. Please see"
" relevant comments in source code for details."
)
labels = tf.identity(inputs)
# Creating the mask and target_mapping tensors
masked_indices = np.full(labels.shape.as_list(), 0, dtype=bool)
labels_shape = tf.shape(labels)
target_mapping = np.zeros((labels_shape[0], labels_shape[1], labels_shape[1]), dtype=np.float32)
for i in range(len(labels)):
# Start from the beginning of the sequence by setting `cur_len = 0` (number of tokens processed so far).
cur_len = 0
max_len = tf.shape(labels)[1]
while cur_len < max_len:
# Sample a `span_length` from the interval `[1, max_span_length]` (length of span of tokens to be masked)
span_length = randint(1, self.max_span_length + 1)
# Reserve a context of length `context_length = span_length / plm_probability` to surround the span to be masked
context_length = int(span_length / self.plm_probability)
# Sample a starting point `start_index` from the interval `[cur_len, cur_len + context_length - span_length]` and mask tokens `start_index:start_index + span_length`
start_index = cur_len + randint(0, context_length - span_length + 1)
masked_indices[i, start_index : start_index + span_length] = 1
# Set `cur_len = cur_len + context_length`
cur_len += context_length
# Since we're replacing non-masked tokens with -100 in the labels tensor instead of skipping them altogether,
# the i-th predict corresponds to the i-th token.
target_mapping[i] = np.eye(labels_shape[1])
masked_indices = tf.cast(tf.convert_to_tensor(masked_indices), dtype=tf.bool)
target_mapping = tf.convert_to_tensor(target_mapping)
special_tokens_mask = tf.convert_to_tensor(
[
self.tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True)
for val in labels.numpy().tolist()
],
)
special_tokens_mask = tf.cast(special_tokens_mask, dtype=tf.bool)
masked_indices = masked_indices & ~special_tokens_mask
if self.tokenizer.pad_token is not None:
padding_mask = labels == self.tokenizer.pad_token_id
masked_indices = masked_indices & ~padding_mask
# Mask indicating non-functional tokens, where functional tokens are [SEP], [CLS], padding, etc.
non_func_mask = ~(padding_mask | special_tokens_mask)
inputs = tf.where(masked_indices, self.tokenizer.mask_token_id, inputs)
labels = tf.where(masked_indices, labels, -100) # We only compute loss on masked tokens
perm_mask = []
for i in range(len(labels)):
# Generate permutation indices i.e. sample a random factorisation order for the sequence. This will
# determine which tokens a given token can attend to (encoded in `perm_mask`).
# Note: Length of token sequence being permuted has to be less than or equal to reused sequence length
# (see documentation for `mems`), otherwise information may leak through due to reuse. In this implementation,
# we assume that reused length is half of sequence length and permutation length is equal to reused length.
# This requires that the sequence length be even.
# Create a linear factorisation order
# tf.range is the equivalent of torch.arange
perm_index = tf.range(labels_shape[1])
# Split this into two halves, assuming that half the sequence is reused each time
perm_index = tf.transpose(tf.reshape(perm_index, (-1, labels_shape[1] // 2)))
# Permute the two halves such that they do not cross over
perm_index = tf.random.shuffle(perm_index) # Shuffles along the first dimension
# Flatten this out into the desired permuted factorisation order
perm_index = tf.reshape(tf.transpose(perm_index), (-1,))
# Set the permutation indices of non-masked (non-functional) tokens to the
# smallest index (-1) so that:
# (1) They can be seen by all other positions
# (2) They cannot see masked positions, so there won't be information leak
perm_index = tf.where(~masked_indices[i] & non_func_mask[i], -1, perm_index)
# The logic for whether the i-th token can attend on the j-th token based on the factorisation order:
# 0 (can attend): If perm_index[i] > perm_index[j] or j is neither masked nor a functional token
# 1 (cannot attend): If perm_index[i] <= perm_index[j] and j is either masked or a functional token
perm_mask.append(
(tf.reshape(perm_index, (labels_shape[1], 1)) <= tf.reshape(perm_index, (1, labels_shape[1])))
& masked_indices[i]
)
perm_mask = tf.stack(perm_mask, axis=0)
return tf.cast(inputs, tf.int64), tf.cast(perm_mask, tf.float32), target_mapping, tf.cast(labels, tf.int64)
def numpy_mask_tokens(self, inputs: Any) -> tuple[Any, Any, Any, Any]:
"""
The masked tokens to be predicted for a particular sequence are determined by the following algorithm:
0. Start from the beginning of the sequence by setting `cur_len = 0` (number of tokens processed so far).
1. Sample a `span_length` from the interval `[1, max_span_length]` (length of span of tokens to be masked)
2. Reserve a context of length `context_length = span_length / plm_probability` to surround span to be
masked
3. Sample a starting point `start_index` from the interval `[cur_len, cur_len + context_length -
span_length]` and mask tokens `start_index:start_index + span_length`
4. Set `cur_len = cur_len + context_length`. If `cur_len < max_len` (i.e. there are tokens remaining in the
sequence to be processed), repeat from Step 1.
"""
if self.tokenizer.mask_token is None:
raise ValueError(
"This tokenizer does not have a mask token which is necessary for permutation language modeling."
" Please add a mask token if you want to use this tokenizer."
)
if inputs.shape[1] % 2 != 0:
raise ValueError(
"This collator requires that sequence lengths be even to create a leakage-free perm_mask. Please see"
" relevant comments in source code for details."
)
labels = np.copy(inputs)
# Creating the mask and target_mapping tensors
masked_indices = np.full(labels.shape, 0, dtype=bool)
target_mapping = np.zeros((labels.shape[0], labels.shape[1], labels.shape[1]), dtype=np.float32)
for i in range(labels.shape[0]):
# Start from the beginning of the sequence by setting `cur_len = 0` (number of tokens processed so far).
cur_len = 0
max_len = labels.shape[1]
while cur_len < max_len:
# Sample a `span_length` from the interval `[1, max_span_length]` (length of span of tokens to be masked)
span_length = randint(1, self.max_span_length + 1)
# Reserve a context of length `context_length = span_length / plm_probability` to surround the span to be masked
context_length = int(span_length / self.plm_probability)
# Sample a starting point `start_index` from the interval `[cur_len, cur_len + context_length - span_length]` and mask tokens `start_index:start_index + span_length`
start_index = cur_len + randint(0, context_length - span_length + 1)
masked_indices[i, start_index : start_index + span_length] = 1
# Set `cur_len = cur_len + context_length`
cur_len += context_length
# Since we're replacing non-masked tokens with -100 in the labels tensor instead of skipping them altogether,
# the i-th predict corresponds to the i-th token.
target_mapping[i] = np.eye(labels.shape[1])
special_tokens_mask = np.array(
[self.tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()],
dtype=bool,
)
masked_indices[special_tokens_mask] = 0
if self.tokenizer.pad_token is not None:
padding_mask = labels == self.tokenizer.pad_token_id
masked_indices[padding_mask] = 0.0
# Mask indicating non-functional tokens, where functional tokens are [SEP], [CLS], padding, etc.
non_func_mask = ~(padding_mask | special_tokens_mask)
inputs[masked_indices] = self.tokenizer.mask_token_id
labels[~masked_indices] = -100 # We only compute loss on masked tokens
perm_mask = np.zeros((labels.shape[0], labels.shape[1], labels.shape[1]), dtype=np.float32)
for i in range(labels.shape[0]):
# Generate permutation indices i.e. sample a random factorisation order for the sequence. This will
# determine which tokens a given token can attend to (encoded in `perm_mask`).
# Note: Length of token sequence being permuted has to be less than or equal to reused sequence length
# (see documentation for `mems`), otherwise information may leak through due to reuse. In this implementation,
# we assume that reused length is half of sequence length and permutation length is equal to reused length.
# This requires that the sequence length be even.
# Create a linear factorisation order
perm_index = np.arange(labels.shape[1])
# Split this into two halves, assuming that half the sequence is reused each time
perm_index = perm_index.reshape((-1, labels.shape[1] // 2)).T
# Permute the two halves such that they do not cross over
np.random.shuffle(perm_index)
# Flatten this out into the desired permuted factorisation order
perm_index = perm_index.T.flatten()
# Set the permutation indices of non-masked (non-functional) tokens to the
# smallest index (-1) so that:
# (1) They can be seen by all other positions
# (2) They cannot see masked positions, so there won't be information leak
perm_index[~masked_indices[i] & non_func_mask[i]] = -1
# The logic for whether the i-th token can attend on the j-th token based on the factorisation order:
# 0 (can attend): If perm_index[i] > perm_index[j] or j is neither masked nor a functional token
# 1 (cannot attend): If perm_index[i] <= perm_index[j] and j is either masked or a functional token
perm_mask[i] = (
perm_index.reshape((labels.shape[1], 1)) <= perm_index.reshape((1, labels.shape[1]))
) & masked_indices[i]
return inputs.astype(np.int64), perm_mask, target_mapping, labels.astype(np.int64)
@dataclass
class DataCollatorWithFlattening(DefaultDataCollator):
"""
Data collator used for padding free approach. Does the following:
- concatenates the entire mini batch into single long sequence of shape [1, total_tokens]
- uses `separator_id` to separate sequences within the concatenated `labels`, default value is -100
- no padding will be added, returns `input_ids`, `labels` and `position_ids` by default
- optionally returns the kwargs contained in FlashAttentionKwargs
- optionally returns seq_idx indicating which sequence each token belongs to
<Tip warning={true}>
Using `DataCollatorWithFlattening` will flatten the entire mini batch into single long sequence.
Make sure your attention computation is able to handle it!
</Tip>
"""
def __init__(
self,
*args,
return_position_ids=True,
separator_id=-100,
return_flash_attn_kwargs=False,
return_seq_idx=False,
**kwargs,
):
super().__init__(*args, **kwargs)
self.return_position_ids = return_position_ids
self.separator_id = separator_id
self.return_flash_attn_kwargs = return_flash_attn_kwargs
self.return_seq_idx = return_seq_idx
self._int_64_keys = {"labels", "position_ids", "input_ids"}
self._batch_dim_keys = {"labels", "position_ids", "input_ids", "seq_idx"}
self._py_int_keys = {"max_length_q", "max_length_k"}
def __call__(self, features, return_tensors=None, separator_id=None):
if return_tensors is None:
return_tensors = self.return_tensors
if separator_id is None:
separator_id = self.separator_id
is_labels_provided = "labels" in features[0]
batch = {"input_ids": [], "labels": []}
if self.return_position_ids:
batch.update({"position_ids": []})
if self.return_seq_idx:
batch.update({"seq_idx": []})
if self.return_flash_attn_kwargs:
cu_seq_lens = [0]
max_length = 0
for seq_idx, sample in enumerate(features):
input_ids = sample["input_ids"]
batch["input_ids"] += input_ids
if is_labels_provided:
batch["labels"] += [separator_id] + sample["labels"][1:]
else:
batch["labels"] += [separator_id] + input_ids[1:]
if self.return_position_ids:
batch["position_ids"] += list(range(len(input_ids)))
if self.return_seq_idx:
batch["seq_idx"] += [seq_idx for _ in range(len(input_ids))]
if self.return_flash_attn_kwargs:
cu_seq_lens.append(cu_seq_lens[-1] + len(input_ids))
max_length = max(max_length, len(input_ids))
if self.return_flash_attn_kwargs:
batch["cu_seq_lens_q"] = batch["cu_seq_lens_k"] = cu_seq_lens
batch["max_length_q"] = batch["max_length_k"] = max_length
# FlashAttentionKwargs and seq_idx are expected to be int32s.
if return_tensors == "pt":
import torch
data_cls = torch.tensor
dtype_64 = torch.int64
dtype_32 = torch.int32
elif return_tensors == "np":
data_cls = np.array
dtype_64 = np.int64
dtype_32 = np.int32
else:
raise ValueError(f'return_tensors must be one of ("pt", "np"), {return_tensors=} not suported')
for k, v in batch.items():
if k in self._batch_dim_keys:
v = [v]
# Flash attention max_len_{q,k} are python ints
if k not in self._py_int_keys:
batch[k] = data_cls(v, dtype=dtype_64 if k in self._int_64_keys else dtype_32)
return batch
| transformers/src/transformers/data/data_collator.py/0 | {
"file_path": "transformers/src/transformers/data/data_collator.py",
"repo_id": "transformers",
"token_count": 42365
} | 420 |
# Copyright 2025 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import copy
import json
import os
from dataclasses import dataclass
from typing import Any, Union
@dataclass
class DistributedConfig:
"""
Base class for distributed configs
"""
enable_expert_parallel: bool = False
# TODO: add tp_plan, pp_plan, device_mesh etc..
@classmethod
def from_dict(cls, config_dict, **kwargs):
"""
Constructs a DistributedConfig instance from a dictionary of parameters.
Args:
config_dict (Dict[str, Any]): Dictionary containing configuration parameters.
**kwargs: Additional keyword arguments to override dictionary values.
Returns:
DistributedConfig: Instance of DistributedConfig constructed from the dictionary.
"""
config = cls(**config_dict)
to_remove = []
for key, value in kwargs.items():
if hasattr(config, key):
setattr(config, key, value)
to_remove.append(key)
for key in to_remove:
kwargs.pop(key, None)
return config
# Copied from transformers.utils.quantization_config.QuantizationConfigMixin.to_json_file
def to_json_file(self, json_file_path: Union[str, os.PathLike]):
"""
Save this instance to a JSON file.
Args:
json_file_path (`str` or `os.PathLike`):
Path to the JSON file in which this configuration instance's parameters will be saved.
use_diff (`bool`, *optional*, defaults to `True`):
If set to `True`, only the difference between the config instance and the default
`QuantizationConfig()` is serialized to JSON file.
"""
with open(json_file_path, "w", encoding="utf-8") as writer:
config_dict = self.to_dict()
json_string = json.dumps(config_dict, indent=2, sort_keys=True) + "\n"
writer.write(json_string)
def to_dict(self) -> dict[str, Any]:
"""
Serializes this instance to a Python dictionary. Returns:
`Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance.
"""
return copy.deepcopy(self.__dict__)
# Copied from transformers.utils.quantization_config.QuantizationConfigMixin.__iter__
def __iter__(self):
"""allows `dict(obj)` for situations where obj may be a dict or QuantizationConfigMixin"""
for attr, value in copy.deepcopy(self.__dict__).items():
yield attr, value
# Copied from transformers.utils.quantization_config.QuantizationConfigMixin.__repr__
def __repr__(self):
return f"{self.__class__.__name__} {self.to_json_string()}"
def to_json_string(self):
"""
Serializes this instance to a JSON formatted string.
Returns:
str: JSON formatted string representing the configuration instance.
"""
return json.dumps(self.__dict__, indent=2) + "\n"
def update(self, **kwargs):
"""
Updates attributes of this class instance with attributes from `kwargs` if they match existing attributes,
returning all the unused kwargs.
Args:
kwargs (`Dict[str, Any]`):
Dictionary of attributes to tentatively update this class.
Returns:
`Dict[str, Any]`: Dictionary containing all the key-value pairs that were not used to update the instance.
"""
to_remove = []
for key, value in kwargs.items():
if hasattr(self, key):
setattr(self, key, value)
to_remove.append(key)
# Remove all the attributes that were updated, without modifying the input dict
unused_kwargs = {key: value for key, value in kwargs.items() if key not in to_remove}
return unused_kwargs
| transformers/src/transformers/distributed/configuration_utils.py/0 | {
"file_path": "transformers/src/transformers/distributed/configuration_utils.py",
"repo_id": "transformers",
"token_count": 1710
} | 421 |
# coding=utf-8
# Copyright 2022 The HuggingFace Inc. team
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import inspect
import numpy as np
import tensorflow as tf
from ..tf_utils import stable_softmax
from ..utils import add_start_docstrings
from ..utils.logging import get_logger
logger = get_logger(__name__)
TF_LOGITS_PROCESSOR_INPUTS_DOCSTRING = r"""
Args:
input_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using [`PreTrainedTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
scores (`tf.Tensor` of shape `(batch_size, config.vocab_size)`):
Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam
search or log softmax for each vocabulary token when using beam search.
cur_len (`int`):
The current length of valid input sequence tokens. In the TF implementation, the input_ids' sequence length
is the maximum length generate can produce, and we need to know which of its tokens are valid.
kwargs (`dict[str, Any]`, *optional*):
Additional logits processor specific kwargs.
Return:
`tf.Tensor` of shape `(batch_size, config.vocab_size)`: The processed prediction scores.
"""
class TFLogitsProcessor:
"""Abstract base class for all logit processors that can be applied during generation."""
@add_start_docstrings(TF_LOGITS_PROCESSOR_INPUTS_DOCSTRING)
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
"""TF method for processing logits."""
raise NotImplementedError(
f"{self.__class__} is an abstract class. Only classes inheriting this class can be called."
)
class TFLogitsWarper:
"""Abstract base class for all logit warpers that can be applied during generation with multinomial sampling."""
@add_start_docstrings(TF_LOGITS_PROCESSOR_INPUTS_DOCSTRING)
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
"""TF method for warping logits."""
raise NotImplementedError(
f"{self.__class__} is an abstract class. Only classes inheriting this class can be called."
)
class TFLogitsProcessorList(list):
"""
This class can be used to create a list of [`TFLogitsProcessor`] to subsequently process a `scores` input tensor.
This class inherits from list and adds a specific *__call__* method to apply each [`TFLogitsProcessor`] to the
inputs.
"""
@add_start_docstrings(TF_LOGITS_PROCESSOR_INPUTS_DOCSTRING)
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int, **kwargs) -> tf.Tensor:
for processor in self:
function_args = inspect.signature(processor.__call__).parameters
if len(function_args) > 3:
if not all(arg in kwargs for arg in list(function_args.keys())[2:]):
raise ValueError(
f"Make sure that all the required parameters: {list(function_args.keys())} for "
f"{processor.__class__} are passed to the logits processor."
)
scores = processor(input_ids, scores, cur_len, **kwargs)
else:
scores = processor(input_ids, scores, cur_len)
return scores
class TFTemperatureLogitsWarper(TFLogitsWarper):
r"""
[`TFLogitsWarper`] for temperature (exponential scaling output probability distribution).
Args:
temperature (`float`):
The value used to module the logits distribution.
"""
def __init__(self, temperature: float):
if not isinstance(temperature, float) or not (temperature > 0):
raise ValueError(f"`temperature` has to be a strictly positive float, but is {temperature}")
self.temperature = temperature
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
scores = scores / self.temperature
return scores
class TFTopKLogitsWarper(TFLogitsWarper):
r"""
[`TFLogitsWarper`] that performs top-k, i.e. restricting to the k highest probability elements.
Args:
top_k (`int`):
The number of highest probability vocabulary tokens to keep for top-k-filtering.
filter_value (`float`, *optional*, defaults to -inf):
All filtered values will be set to this float value.
min_tokens_to_keep (`int`, *optional*, defaults to 1):
Minimum number of tokens that cannot be filtered.
"""
def __init__(self, top_k: int, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
if not isinstance(top_k, int) or top_k <= 0:
raise ValueError(f"`top_k` has to be a strictly positive integer, but is {top_k}")
self.top_k = max(top_k, min_tokens_to_keep)
self.filter_value = filter_value
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
top_k = min(self.top_k, scores.shape[-1]) # Safety check
# Boolean mask containing all tokens with a probability less than the last token of the top-k
indices_to_remove = scores < tf.math.top_k(scores, k=top_k)[0][..., -1:]
next_scores = tf.where(indices_to_remove, self.filter_value, scores)
return next_scores
class TFTopPLogitsWarper(TFLogitsWarper):
"""
[`TFLogitsWarper`] that performs top-p, i.e. restricting to top tokens summing to <= prob_cut_off.
Args:
top_p (`float`):
If set to < 1, only the smallest set of most probable tokens with probabilities that add up to `top_p` or
higher are kept for generation.
filter_value (`float`, *optional*, defaults to -inf):
All filtered values will be set to this float value.
min_tokens_to_keep (`int`, *optional*, defaults to 1):
Minimum number of tokens that cannot be filtered.
"""
def __init__(self, top_p: float, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
if not isinstance(top_p, float) or (top_p < 0 or top_p > 1.0):
raise ValueError(f"`top_p` has to be a float > 0 and < 1, but is {top_p}")
if not isinstance(min_tokens_to_keep, int) or (min_tokens_to_keep < 1):
raise ValueError(f"`min_tokens_to_keep` has to be a positive integer, but is {min_tokens_to_keep}")
self.top_p = top_p
self.filter_value = filter_value
self.min_tokens_to_keep = min_tokens_to_keep
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
topk_scores, topk_indices = tf.math.top_k(scores, scores.shape[-1])
mask_scores = tf.fill(scores.shape, self.filter_value)
cumulative_probs = tf.math.cumsum(stable_softmax(topk_scores, axis=-1), axis=-1)
score_mask = cumulative_probs < self.top_p
# Also include the token that is higher than top_p (the first false = shift and insert a True on the left)
score_mask = tf.concat((tf.ones([score_mask.shape[0], 1], dtype=tf.bool), score_mask[:, :-1]), axis=-1)
# Ensure min tokens to keep
score_mask = tf.concat(
(
tf.ones([score_mask.shape[0], self.min_tokens_to_keep], dtype=tf.bool),
score_mask[:, self.min_tokens_to_keep :],
),
axis=-1,
)
# Mask the values that do not fit the criteria
topk_next_scores = tf.where(score_mask, topk_scores, mask_scores)
# Undo the topk sorting: converts the 2D matrix of per-row original indices of shape (batch_size, vocab_size)
# to a 3D tensor of shape (batch_size, vocab_size, 2) containing the original score coordinate, from which we
# can scatter (i.e. `scatter_indices[row, col, :]` is a tensor containing `[row, topk_indices[row, col]]`)
scatter_rows = tf.tile(tf.expand_dims(tf.range(topk_indices.shape[0]), axis=-1), [1, topk_indices.shape[-1]])
scatter_indices = tf.stack((scatter_rows, topk_indices), axis=-1)
next_scores = tf.scatter_nd(scatter_indices, topk_next_scores, shape=topk_next_scores.shape)
return next_scores
class TFMinLengthLogitsProcessor(TFLogitsProcessor):
r"""
[`TFLogitsProcessor`] enforcing a min-length by setting EOS probability to 0.
Args:
min_length (`int`):
The minimum length below which the score of `eos_token_id` is set to `-float("Inf")`.
eos_token_id (`int`):
The id of the *end-of-sequence* token.
"""
def __init__(self, min_length: int, eos_token_id: int):
if not isinstance(min_length, int) or min_length < 0:
raise ValueError(f"`min_length` has to be a positive integer, but is {min_length}")
if not isinstance(eos_token_id, int) or eos_token_id < 0:
raise ValueError(f"`eos_token_id` has to be a positive integer, but is {eos_token_id}")
self.min_length = min_length
self.eos_token_id = eos_token_id
def _apply_eos_token_mask(self, scores: tf.Tensor) -> tf.Tensor:
eos_token_id_mask = tf.range(scores.shape[-1]) == self.eos_token_id
scores = tf.where(eos_token_id_mask, float("-inf"), scores)
return scores
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
# applies eos token masking if the first argument is true
scores = tf.cond(
tf.less(cur_len, self.min_length),
lambda: self._apply_eos_token_mask(scores),
lambda: tf.identity(scores),
)
return scores
class TFRepetitionPenaltyLogitsProcessor(TFLogitsProcessor):
r"""
[`TFLogitsProcessor`] enforcing an exponential penalty on repeated sequences.
Args:
repetition_penalty (`float`):
The parameter for repetition penalty. 1.0 means no penalty. See [this
paper](https://huggingface.co/papers/1909.05858) for more details.
"""
def __init__(self, penalty: float):
if not isinstance(penalty, float) or not (penalty > 0):
raise ValueError(f"`penalty` has to be a strictly positive float, but is {penalty}")
self.penalty = penalty
def _create_score_penalties(self, input_ids: tf.Tensor, logits: tf.Tensor) -> tf.Tensor:
# We want to populate the penalties in the positions of `input_ids`. Since XLA can't handle shapes unknown
# before runtime, `tf.unique` can't be used. Therefore, we may have redundant updates, when a given row has
# the same token multiple times.
# Gathers the penalties to apply
logit_penalties = tf.gather(logits, input_ids, axis=1, batch_dims=1)
logit_penalties = tf.where(logit_penalties > 0, 1 / self.penalty, logit_penalties)
logit_penalties = tf.where(logit_penalties < 0, self.penalty, logit_penalties)
# Scatters the penalties
token_penalties = tf.ones(logits.shape)
batch_size = input_ids.shape[0]
seq_len = tf.shape(input_ids)[1] # the sequence length has dynamic size, hence the dynamic shape
indexable_prev_input_ids = tf.concat(
(
tf.expand_dims(tf.repeat(tf.range(batch_size), seq_len), axis=-1),
tf.expand_dims(tf.reshape(input_ids, [-1]), axis=-1),
),
axis=1,
)
token_penalties = tf.tensor_scatter_nd_update(
token_penalties, indices=indexable_prev_input_ids, updates=tf.reshape(logit_penalties, [-1])
)
return token_penalties
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
score_penalties = self._create_score_penalties(input_ids[:, :cur_len], scores)
scores = tf.math.multiply(scores, score_penalties)
return scores
class TFNoBadWordsLogitsProcessor(TFLogitsProcessor):
"""
[`TFLogitsProcessor`] that enforces that specified sequences will never be sampled.
Args:
bad_words_ids (`list[list[int]]`):
List of list of token ids that are not allowed to be generated. In order to get the tokens of the words
that should not appear in the generated text, make sure to set `add_prefix_space=True` when initializing
the tokenizer, and use `tokenizer(bad_words, add_special_tokens=False).input_ids`. The `add_prefix_space`
argument is only supported for some slow tokenizers, as fast tokenizers' prefixing behaviours come from
`pre tokenizers`. Read more [here](https://huggingface.co/docs/tokenizers/api/pre-tokenizers).
eos_token_id (`int`):
The id of the *end-of-sequence* token.
"""
def __init__(self, bad_words_ids: list[list[int]], eos_token_id: int):
if not isinstance(bad_words_ids, list) or len(bad_words_ids) == 0:
raise ValueError(f"`bad_words_ids` has to be a non-empty list, but is {bad_words_ids}.")
if any(not isinstance(bad_word_ids, list) for bad_word_ids in bad_words_ids):
raise ValueError(f"`bad_words_ids` has to be a list of lists, but is {bad_words_ids}.")
if any(
any((not isinstance(token_id, (int, np.integer)) or token_id < 0) for token_id in bad_word_ids)
for bad_word_ids in bad_words_ids
):
raise ValueError(
f"Each list in `bad_words_ids` has to be a list of positive integers, but is {bad_words_ids}."
)
# stores the information about bad words in three tensors:
# 1. a rectangular tensor with the forbidden sequences (padded with `-1`), for full data comparisons
self.bad_word_seqs_ids = tf.ragged.constant(bad_words_ids).to_tensor(default_value=-1)
# 2. a tensor with the unpadded length of each forbidden sequence, for quick length comparisons
bad_word_seqs_len = [len(bad_words) for bad_words in bad_words_ids]
if any(word_len == 0 for word_len in bad_word_seqs_len):
raise ValueError(f"Banned words token sequences {bad_words_ids} cannot have an empty list")
self.bad_word_seqs_len = tf.convert_to_tensor(bad_word_seqs_len, dtype=tf.int32)
# 3. a tensor containing the last token for each sequence, for easy access to the tokens that may be banned
self.seq_forbidden_tokens = tf.convert_to_tensor([bad_words[-1] for bad_words in bad_words_ids])
def _calc_row_banned_bad_tokens(self, row_input_ids: tf.Tensor) -> tf.Tensor:
def _tokens_match(bad_word_seq_number):
def _len_one():
# If the bad sequence only has one token, always mask it
return tf.cond(
tf.math.equal(self.bad_word_seqs_len[bad_word_seq_number], 1),
lambda: tf.ones((), dtype=tf.bool),
_len_greater_than_cur_len,
)
def _len_greater_than_cur_len():
# Otherwise, if the bad sequence is longer than the current length they can't ever match
return tf.cond(
tf.math.greater(self.bad_word_seqs_len[bad_word_seq_number], tf.shape(row_input_ids)[0]),
lambda: tf.zeros((), dtype=tf.bool),
_match_found,
)
def _match_found():
# Finally, runs the actual comparison. Can only be called if the previous comparisons do not yield
# an answer (otherwise we get indexing exceptions)
compare_len = self.bad_word_seqs_len[bad_word_seq_number] - 1
return tf.cond(
tf.math.reduce_all(
tf.math.equal(
row_input_ids[-compare_len:], self.bad_word_seqs_ids[bad_word_seq_number, :compare_len]
)
),
lambda: tf.ones((), dtype=tf.bool),
lambda: tf.zeros((), dtype=tf.bool),
)
match = _len_one()
return match
# Compares the current row against all bad word sequences, obtaining a mask with the matches.
match_mask = tf.map_fn(_tokens_match, tf.range(self.bad_word_seqs_ids.shape[0]), fn_output_signature=tf.bool)
row_banned_tokens = self.seq_forbidden_tokens[match_mask]
return row_banned_tokens
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
# We want to mask some banned tokens, at a score level. Since the banned tokens depend on the previous
# `input_ids`, they may have a different length for each row, and they may even be empty for some rows.
# To remain simple and XLA-compatible, we work on a per-row fashion.
# TODO (Joao): this function might trigger XLA retracing as `cur_len` increases. Fix it if it becomes
# a frequent choke point. (make `cur_len` a tensor?)
def _get_row_updated_score(row_inputs: tuple[tf.Tensor]) -> tf.Tensor:
row_input_ids, row_score = row_inputs
banned_tokens = self._calc_row_banned_bad_tokens(row_input_ids[:cur_len])
banned_tokens_mask = tf.scatter_nd(
indices=tf.expand_dims(banned_tokens, axis=-1),
updates=tf.ones_like(banned_tokens, dtype=tf.bool),
shape=row_score.shape,
)
row_score = tf.where(banned_tokens_mask, -float("inf"), row_score)
return row_score
scores = tf.map_fn(_get_row_updated_score, (input_ids, scores), fn_output_signature=tf.float32)
return scores
class TFNoRepeatNGramLogitsProcessor(TFLogitsProcessor):
r"""
[`TFLogitsProcessor`] that enforces no repetition of n-grams. See
[Fairseq](https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345).
Args:
ngram_size (`int`):
All ngrams of size `ngram_size` can only occur once.
"""
def __init__(self, ngram_size: int):
if not isinstance(ngram_size, int) or ngram_size <= 0:
raise ValueError(f"`ngram_size` has to be a strictly positive integer, but is {ngram_size}")
self.ngram_size = ngram_size
def calc_banned_ngram_tokens(self, input_ids, num_hypos, cur_len):
# Copied from fairseq for no_repeat_ngram in beam_search
if cur_len + 1 < self.ngram_size:
# return no banned tokens if we haven't generated ngram_size tokens yet
return [[] for _ in range(num_hypos)]
generated_ngrams = [{} for _ in range(num_hypos)]
prev_input_ids = input_ids[:, :cur_len]
for idx in range(num_hypos):
gen_tokens = prev_input_ids[idx].numpy().tolist()
generated_ngram = generated_ngrams[idx]
for ngram in zip(*[gen_tokens[i:] for i in range(self.ngram_size)]):
prev_ngram_tuple = tuple(ngram[:-1])
generated_ngram[prev_ngram_tuple] = generated_ngram.get(prev_ngram_tuple, []) + [ngram[-1]]
def _get_generated_ngrams(hypo_idx):
# Before decoding the next token, prevent decoding of ngrams that have already appeared
start_idx = cur_len + 1 - self.ngram_size
ngram_idx = tuple(prev_input_ids[hypo_idx, start_idx:cur_len].numpy().tolist())
return generated_ngrams[hypo_idx].get(ngram_idx, [])
banned_tokens = [_get_generated_ngrams(hypo_idx) for hypo_idx in range(num_hypos)]
return banned_tokens
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
# TODO (joao): enable XLA on this logits processor. See discussion and attempts in
# https://github.com/huggingface/transformers/pull/16974
if not tf.executing_eagerly():
raise NotImplementedError("TFNoRepeatNGramLogitsProcessor is only implemented for eager execution.")
batch_size, vocab_size = scores.shape
banned_tokens = self.calc_banned_ngram_tokens(input_ids, batch_size, cur_len)
# create banned_tokens boolean mask
banned_tokens_indices_mask = []
for banned_tokens_slice in banned_tokens:
banned_tokens_indices_mask.append([token in banned_tokens_slice for token in range(vocab_size)])
scores = tf.where(tf.convert_to_tensor(banned_tokens_indices_mask, dtype=tf.bool), -float("inf"), scores)
return scores
class TFForcedBOSTokenLogitsProcessor(TFLogitsProcessor):
r"""
[`TFLogitsProcessor`] that enforces the specified token as the first generated token.
Args:
bos_token_id (`int`):
The id of the token to force as the first generated token.
"""
def __init__(self, bos_token_id: int):
if bos_token_id < 0:
raise ValueError(f"The forced bos token id must be a non-negative integer, got {bos_token_id}")
self.bos_token_id = bos_token_id
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
if cur_len == 1:
batch_size, num_tokens = scores.shape
# sets the score to 0 in the bos_token_id column
scores = tf.zeros((batch_size, 1))
# sets the score to -inf everywhere else
if self.bos_token_id > 0:
scores = tf.concat((tf.broadcast_to(-float("inf"), (batch_size, self.bos_token_id)), scores), axis=-1)
if self.bos_token_id < (num_tokens - 1):
scores = tf.concat(
(scores, tf.broadcast_to(-float("inf"), (batch_size, (num_tokens - 1) - self.bos_token_id))),
axis=-1,
)
return scores
class TFForcedEOSTokenLogitsProcessor(TFLogitsProcessor):
r"""
[`TFLogitsProcessor`] that enforces the specified token as the last generated token when `max_length` is reached.
Args:
max_length (`int`):
The maximum length of the sequence to be generated.
eos_token_id (`int`):
The id of the token to force as the last generated token when `max_length` is reached.
"""
def __init__(self, max_length: int, eos_token_id: int):
self.max_length = max_length
if eos_token_id < 0:
raise ValueError(f"The forced eos token id must be a non-negative integer, got {eos_token_id}")
self.eos_token_id = eos_token_id
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
if cur_len == self.max_length - 1:
batch_size, num_tokens = scores.shape
# sets the score to 0 in the eos_token_id column
scores = tf.zeros((batch_size, 1))
# sets the score to -inf everywhere else
if self.eos_token_id > 0:
scores = tf.concat((tf.broadcast_to(-float("inf"), (batch_size, self.eos_token_id)), scores), axis=-1)
if self.eos_token_id < (num_tokens - 1):
scores = tf.concat(
(scores, tf.broadcast_to(-float("inf"), (batch_size, (num_tokens - 1) - self.eos_token_id))),
axis=-1,
)
return scores
class TFSuppressTokensAtBeginLogitsProcessor(TFLogitsProcessor):
r"""
[`TFSuppressTokensAtBeginLogitsProcessor`] suppresses a list of tokens as soon as the `generate` function starts
generating using `begin_index` tokens. This should ensure that the tokens defined by `begin_suppress_tokens` at not
sampled at the beginning of the generation.
"""
def __init__(self, begin_suppress_tokens, begin_index):
self.begin_suppress_tokens = list(begin_suppress_tokens)
self.begin_index = begin_index
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
suppressed_indices = []
for token in self.begin_suppress_tokens:
if token < scores.shape[-1]: # to ensure we don't go beyond the vocab size
suppressed_indices.extend([[i, token] for i in range(scores.shape[0])])
if len(suppressed_indices) > 0:
scores = tf.cond(
tf.equal(cur_len, self.begin_index),
lambda: tf.tensor_scatter_nd_update(
scores,
indices=suppressed_indices,
updates=[-float("inf") for _ in range(scores.shape[0] * len(self.begin_suppress_tokens))],
),
lambda: scores,
)
return scores
class TFSuppressTokensLogitsProcessor(TFLogitsProcessor):
r"""This processor can be used to suppress a list of tokens. The processor will set their log probs to `-inf` so that they
are not sampled."""
def __init__(self, suppress_tokens):
self.suppress_tokens = list(suppress_tokens)
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
suppressed_indices = []
for token in self.suppress_tokens:
if token < scores.shape[-1]: # to ensure we don't go beyond the vocab size
suppressed_indices.extend([[i, token] for i in range(scores.shape[0])])
if len(suppressed_indices) > 0:
scores = tf.tensor_scatter_nd_update(
scores,
indices=[[i, token] for i in range(scores.shape[0]) for token in self.suppress_tokens],
updates=[-float("inf") for _ in range(scores.shape[0] * len(self.suppress_tokens))],
)
return scores
class TFForceTokensLogitsProcessor(TFLogitsProcessor):
r"""This processor takes a list of pairs of integers which indicates a mapping from generation indices to token
indices that will be forced before sampling. The processor will set their log probs to `0` and all other tokens to
`-inf` so that they are sampled at their corresponding index."""
def __init__(self, force_token_map: list[list[int]]):
force_token_map = dict(force_token_map)
# Converts the dictionary of format {index: token} containing the tokens to be forced to an array, where the
# index of the array corresponds to the index of the token to be forced, for XLA compatibility.
# Indexes without forced tokens will have an negative value.
force_token_array = np.ones((max(force_token_map.keys()) + 1), dtype=np.int32) * -1
for index, token in force_token_map.items():
if token is not None:
force_token_array[index] = token
self.force_token_array = tf.convert_to_tensor(force_token_array, dtype=tf.int32)
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
def _force_token(generation_idx):
batch_size = scores.shape[0]
current_token = self.force_token_array[generation_idx]
new_scores = tf.zeros_like(scores, dtype=scores.dtype) + tf.constant([scores.dtype.min])
indices = tf.stack((tf.range(batch_size), tf.tile([current_token], [batch_size])), axis=1)
updates = tf.zeros((batch_size,), dtype=scores.dtype)
new_scores = tf.tensor_scatter_nd_update(new_scores, indices, updates)
return new_scores
scores = tf.cond(
tf.greater_equal(cur_len, tf.shape(self.force_token_array)[0]),
# If the current length is geq than the length of force_token_array, the processor does nothing.
lambda: tf.identity(scores),
# Otherwise, it may force a certain token.
lambda: tf.cond(
tf.greater_equal(self.force_token_array[cur_len], 0),
# Only valid (positive) tokens are forced
lambda: _force_token(cur_len),
# Otherwise, the processor does nothing.
lambda: scores,
),
)
return scores
| transformers/src/transformers/generation/tf_logits_process.py/0 | {
"file_path": "transformers/src/transformers/generation/tf_logits_process.py",
"repo_id": "transformers",
"token_count": 12313
} | 422 |
import importlib.metadata
import inspect
import warnings
from copy import deepcopy
from inspect import signature
from packaging import version
from ..utils import (
get_available_devices,
is_accelerate_available,
is_bitsandbytes_available,
is_bitsandbytes_multi_backend_available,
is_torch_available,
logging,
)
if is_bitsandbytes_available():
import bitsandbytes as bnb
import torch
import torch.nn as nn
from ..pytorch_utils import Conv1D
if is_accelerate_available():
import accelerate
from accelerate import init_empty_weights
from accelerate.hooks import add_hook_to_module, remove_hook_from_module
from accelerate.utils import find_tied_parameters
logger = logging.get_logger(__name__)
def set_module_quantized_tensor_to_device(module, tensor_name, device, value=None, quantized_stats=None):
"""
A helper function to set a given tensor (parameter of buffer) of a module on a specific device (note that doing
`param.to(device)` creates a new tensor not linked to the parameter, which is why we need this function). The
function is adapted from `set_module_tensor_to_device` function from accelerate that is adapted to support the
class `Int8Params` from `bitsandbytes`.
Args:
module (`torch.nn.Module`):
The module in which the tensor we want to move lives.
tensor_name (`str`):
The full name of the parameter/buffer.
device (`int`, `str` or `torch.device`):
The device on which to set the tensor.
value (`torch.Tensor`, *optional*):
The value of the tensor (useful when going from the meta device to any other device).
quantized_stats (`dict[str, Any]`, *optional*):
Dict with items for either 4-bit or 8-bit serialization
"""
# Recurse if needed
if "." in tensor_name:
splits = tensor_name.split(".")
for split in splits[:-1]:
new_module = getattr(module, split)
if new_module is None:
raise ValueError(f"{module} has no attribute {split}.")
module = new_module
tensor_name = splits[-1]
if tensor_name not in module._parameters and tensor_name not in module._buffers:
raise ValueError(f"{module} does not have a parameter or a buffer named {tensor_name}.")
is_buffer = tensor_name in module._buffers
old_value = getattr(module, tensor_name)
if old_value.device == torch.device("meta") and device not in ["meta", torch.device("meta")] and value is None:
raise ValueError(f"{tensor_name} is on the meta device, we need a `value` to put in on {device}.")
prequantized_loading = quantized_stats is not None
if is_buffer or not is_bitsandbytes_available():
is_8bit = False
is_4bit = False
else:
is_4bit = hasattr(bnb.nn, "Params4bit") and isinstance(module._parameters[tensor_name], bnb.nn.Params4bit)
is_8bit = isinstance(module._parameters[tensor_name], bnb.nn.Int8Params)
if is_8bit or is_4bit:
param = module._parameters[tensor_name]
if param.device.type != "cuda":
if value is None:
new_value = old_value.to(device)
elif isinstance(value, torch.Tensor):
new_value = value.to("cpu")
else:
new_value = torch.tensor(value, device="cpu")
# Support models using `Conv1D` in place of `nn.Linear` (e.g. openai-community/gpt2) by transposing the weight matrix prior to quantization.
# Since weights are saved in the correct "orientation", we skip transposing when loading.
if issubclass(module.source_cls, Conv1D) and not prequantized_loading:
new_value = new_value.T
kwargs = old_value.__dict__
if prequantized_loading != (new_value.dtype in (torch.int8, torch.uint8)):
raise ValueError(
f"Value dtype `{new_value.dtype}` is not compatible with parameter quantization status."
)
if is_8bit:
is_8bit_serializable = version.parse(importlib.metadata.version("bitsandbytes")) > version.parse(
"0.37.2"
)
if new_value.dtype in (torch.int8, torch.uint8) and not is_8bit_serializable:
raise ValueError(
"Detected int8 weights but the version of bitsandbytes is not compatible with int8 serialization. "
"Make sure to download the latest `bitsandbytes` version. `pip install --upgrade bitsandbytes`."
)
new_value = bnb.nn.Int8Params(new_value, requires_grad=False, **kwargs).to(device)
if prequantized_loading:
setattr(new_value, "SCB", quantized_stats["SCB"].to(device))
elif is_4bit:
if prequantized_loading:
is_4bit_serializable = version.parse(importlib.metadata.version("bitsandbytes")) >= version.parse(
"0.41.3"
)
if new_value.dtype in (torch.int8, torch.uint8) and not is_4bit_serializable:
raise ValueError(
"Detected 4-bit weights but the version of bitsandbytes is not compatible with 4-bit serialization. "
"Make sure to download the latest `bitsandbytes` version. `pip install --upgrade bitsandbytes`."
)
new_value = bnb.nn.Params4bit.from_prequantized(
data=new_value,
quantized_stats=quantized_stats,
requires_grad=False,
device=device,
**kwargs,
)
else:
new_value = bnb.nn.Params4bit(new_value, requires_grad=False, **kwargs).to(device)
module._parameters[tensor_name] = new_value
else:
if value is None:
new_value = old_value.to(device)
elif isinstance(value, torch.Tensor):
new_value = value.to(device)
else:
new_value = torch.tensor(value, device=device)
if is_buffer:
module._buffers[tensor_name] = new_value
else:
new_value = nn.Parameter(new_value, requires_grad=old_value.requires_grad)
module._parameters[tensor_name] = new_value
def _replace_with_bnb_linear(
model,
modules_to_not_convert=None,
current_key_name=None,
quantization_config=None,
has_been_replaced=False,
):
"""
Private method that wraps the recursion for module replacement.
Returns the converted model and a boolean that indicates if the conversion has been successful or not.
"""
for name, module in model.named_children():
if current_key_name is None:
current_key_name = []
current_key_name.append(name)
if (isinstance(module, (nn.Linear, Conv1D))) and name not in modules_to_not_convert:
# Check if the current key is not in the `modules_to_not_convert`
current_key_name_str = ".".join(current_key_name)
if not any(
(key + "." in current_key_name_str) or (key == current_key_name_str) for key in modules_to_not_convert
):
with init_empty_weights():
if isinstance(module, Conv1D):
in_features, out_features = module.weight.shape
else:
in_features = module.in_features
out_features = module.out_features
if quantization_config.quantization_method() == "llm_int8":
model._modules[name] = bnb.nn.Linear8bitLt(
in_features,
out_features,
module.bias is not None,
has_fp16_weights=quantization_config.llm_int8_has_fp16_weight,
threshold=quantization_config.llm_int8_threshold,
)
has_been_replaced = True
else:
if (
quantization_config.llm_int8_skip_modules is not None
and name in quantization_config.llm_int8_skip_modules
):
pass
else:
extra_kwargs = (
{"quant_storage": quantization_config.bnb_4bit_quant_storage}
if "quant_storage" in list(signature(bnb.nn.Linear4bit).parameters)
else {}
)
model._modules[name] = bnb.nn.Linear4bit(
in_features,
out_features,
module.bias is not None,
quantization_config.bnb_4bit_compute_dtype,
compress_statistics=quantization_config.bnb_4bit_use_double_quant,
quant_type=quantization_config.bnb_4bit_quant_type,
**extra_kwargs,
)
has_been_replaced = True
# Store the module class in case we need to transpose the weight later
model._modules[name].source_cls = type(module)
# Force requires grad to False to avoid unexpected errors
model._modules[name].requires_grad_(False)
if len(list(module.children())) > 0:
_, has_been_replaced = _replace_with_bnb_linear(
module,
modules_to_not_convert,
current_key_name,
quantization_config,
has_been_replaced=has_been_replaced,
)
# Remove the last key for recursion
current_key_name.pop(-1)
return model, has_been_replaced
def replace_with_bnb_linear(model, modules_to_not_convert=None, current_key_name=None, quantization_config=None):
"""
A helper function to replace all `torch.nn.Linear` modules by `bnb.nn.Linear8bit` modules from the `bitsandbytes`
library. This will enable running your models using mixed int8 precision as described by the paper `LLM.int8():
8-bit Matrix Multiplication for Transformers at Scale`. Make sure `bitsandbytes` compiled with the correct CUDA
version of your hardware is installed before running this function. `pip install -i https://test.pypi.org/simple/
bitsandbytes`
The function will be run recursively and replace all `torch.nn.Linear` modules except for the `lm_head` that should
be kept as a `torch.nn.Linear` module. The replacement is done under `init_empty_weights` context manager so no
CPU/GPU memory is required to run this function. Int8 mixed-precision matrix decomposition works by separating a
matrix multiplication into two streams: (1) and systematic feature outlier stream matrix multiplied in fp16
(0.01%), (2) a regular stream of int8 matrix multiplication (99.9%). With this method, int8 inference with no
predictive degradation is possible for very large models (>=176B parameters).
Parameters:
model (`torch.nn.Module`):
Input model or `torch.nn.Module` as the function is run recursively.
modules_to_not_convert (`list[`str`]`, *optional*, defaults to `["lm_head"]`):
Names of the modules to not convert in `Linear8bitLt`. In practice we keep the `lm_head` in full precision
for numerical stability reasons.
current_key_name (`list[`str`]`, *optional*):
An array to track the current key of the recursion. This is used to check whether the current key (part of
it) is not in the list of modules to not convert (for instances modules that are offloaded to `cpu` or
`disk`).
quantization_config ('transformers.utils.quantization_config.BitsAndBytesConfig'):
To configure and manage settings related to quantization, a technique used to compress neural network models
by reducing the precision of the weights and activations, thus making models more efficient in terms of both
storage and computation.
"""
modules_to_not_convert = ["lm_head"] if modules_to_not_convert is None else modules_to_not_convert
model, has_been_replaced = _replace_with_bnb_linear(
model, modules_to_not_convert, current_key_name, quantization_config
)
if not has_been_replaced:
logger.warning(
"You are loading your model in 8bit or 4bit but no linear modules were found in your model."
" Please double check your model architecture, or submit an issue on github if you think this is"
" a bug."
)
return model
# For backward compatibility
def replace_8bit_linear(*args, **kwargs):
warnings.warn(
"`replace_8bit_linear` will be deprecated in a future version, please use `replace_with_bnb_linear` instead",
FutureWarning,
)
return replace_with_bnb_linear(*args, **kwargs)
# For backward compatibility
def set_module_8bit_tensor_to_device(*args, **kwargs):
warnings.warn(
"`set_module_8bit_tensor_to_device` will be deprecated in a future version, please use `set_module_quantized_tensor_to_device` instead",
FutureWarning,
)
return set_module_quantized_tensor_to_device(*args, **kwargs)
def get_keys_to_not_convert(model):
r"""
An utility function to get the key of the module to keep in full precision if any For example for CausalLM modules
we may want to keep the lm_head in full precision for numerical stability reasons. For other architectures, we want
to keep the tied weights of the model. The function will return a list of the keys of the modules to not convert in
int8.
Parameters:
model (`torch.nn.Module`):
Input model
"""
# Create a copy of the model and tie the weights, then
# check if it contains tied weights
tied_model = deepcopy(model) # this has 0 cost since it is done inside `init_empty_weights` context manager`
tied_model.tie_weights()
tied_params = find_tied_parameters(tied_model)
# For compatibility with Accelerate < 0.18
if isinstance(tied_params, dict):
tied_keys = sum(list(tied_params.values()), []) + list(tied_params.keys())
else:
tied_keys = sum(tied_params, [])
has_tied_params = len(tied_keys) > 0
# If there is not tied weights, we want to keep the lm_head（output_embedding) in full precision
if not has_tied_params:
output_emb = model.get_output_embeddings()
if output_emb is not None:
list_last_module = [name for name, module in model.named_modules() if id(module) == id(output_emb)]
return list_last_module
# otherwise, no tied weights, no output embedding defined, simply keep the last module in full precision
list_modules = list(model.named_parameters())
list_last_module = [list_modules[-1][0]]
# add last module together with tied weights
intersection = set(list_last_module) - set(tied_keys)
list_untouched = list(set(tied_keys)) + list(intersection)
# remove ".weight" from the keys
names_to_remove = [".weight", ".bias"]
filtered_module_names = []
for name in list_untouched:
for name_to_remove in names_to_remove:
if name_to_remove in name:
name = name.replace(name_to_remove, "")
filtered_module_names.append(name)
return filtered_module_names
# Copied from PEFT: https://github.com/huggingface/peft/blob/47b3712898539569c02ec5b3ed4a6c36811331a1/src/peft/utils/integrations.py#L41
def dequantize_bnb_weight(weight: "torch.nn.Parameter", dtype: "torch.dtype", state=None):
"""
Helper function to dequantize 4bit or 8bit bnb weights.
If the weight is not a bnb quantized weight, it will be returned as is.
"""
if not isinstance(weight, torch.nn.Parameter):
raise TypeError(f"Input weight should be of type nn.Parameter, got {type(weight)} instead")
cls_name = weight.__class__.__name__
if cls_name not in ("Params4bit", "Int8Params"):
return weight
if cls_name == "Params4bit":
output_tensor = bnb.functional.dequantize_4bit(weight.data, weight.quant_state)
logger.warning_once(
f"The model is going to be dequantized in {output_tensor.dtype} - if you want to upcast it to another dtype, make sure to pass the desired dtype when quantizing the model through `bnb_4bit_quant_type` argument of `BitsAndBytesConfig`"
)
return output_tensor.to(dtype)
if state.SCB is None:
state.SCB = weight.SCB
if hasattr(bnb.functional, "int8_vectorwise_dequant"):
# Use bitsandbytes API if available (requires v0.45.0+)
dequantized = bnb.functional.int8_vectorwise_dequant(weight.data, state.SCB)
else:
# Multiply by (scale/127) to dequantize.
dequantized = weight.data * state.SCB.view(-1, 1) * 7.874015718698502e-3
return dequantized.to(dtype)
def _create_accelerate_new_hook(old_hook):
r"""
Creates a new hook based on the old hook. Use it only if you know what you are doing !
This method is a copy of: https://github.com/huggingface/peft/blob/748f7968f3a31ec06a1c2b0328993319ad9a150a/src/peft/utils/other.py#L245
with some changes
"""
old_hook_cls = getattr(accelerate.hooks, old_hook.__class__.__name__)
old_hook_attr = old_hook.__dict__
filtered_old_hook_attr = {}
old_hook_init_signature = inspect.signature(old_hook_cls.__init__)
for k in old_hook_attr:
if k in old_hook_init_signature.parameters:
filtered_old_hook_attr[k] = old_hook_attr[k]
new_hook = old_hook_cls(**filtered_old_hook_attr)
return new_hook
def _dequantize_and_replace(
model,
dtype,
modules_to_not_convert=None,
current_key_name=None,
quantization_config=None,
has_been_replaced=False,
):
"""
Converts a quantized model into its dequantized original version. The newly converted model will have
some performance drop compared to the original model before quantization - use it only for specific usecases
such as QLoRA adapters merging.
Returns the converted model and a boolean that indicates if the conversion has been successful or not.
"""
quant_method = quantization_config.quantization_method()
target_cls = bnb.nn.Linear8bitLt if quant_method == "llm_int8" else bnb.nn.Linear4bit
for name, module in model.named_children():
if current_key_name is None:
current_key_name = []
current_key_name.append(name)
if isinstance(module, target_cls) and name not in modules_to_not_convert:
# Check if the current key is not in the `modules_to_not_convert`
current_key_name_str = ".".join(current_key_name)
if not any(
(key + "." in current_key_name_str) or (key == current_key_name_str) for key in modules_to_not_convert
):
bias = getattr(module, "bias", None)
device = module.weight.device
with init_empty_weights():
new_module = torch.nn.Linear(module.in_features, module.out_features, bias=bias is not None)
if quant_method == "llm_int8":
state = module.state
else:
state = None
new_module.weight = torch.nn.Parameter(dequantize_bnb_weight(module.weight, dtype, state))
if bias is not None:
new_module.bias = bias
# Create a new hook and attach it in case we use accelerate
if hasattr(module, "_hf_hook"):
old_hook = module._hf_hook
new_hook = _create_accelerate_new_hook(old_hook)
remove_hook_from_module(module)
add_hook_to_module(new_module, new_hook)
new_module.to(device)
model._modules[name] = new_module
has_been_replaced = True
if len(list(module.children())) > 0:
_, has_been_replaced = _dequantize_and_replace(
module,
dtype,
modules_to_not_convert,
current_key_name,
quantization_config,
has_been_replaced=has_been_replaced,
)
# Remove the last key for recursion
current_key_name.pop(-1)
return model, has_been_replaced
def dequantize_and_replace(
model,
modules_to_not_convert=None,
quantization_config=None,
):
model, has_been_replaced = _dequantize_and_replace(
model,
model.dtype,
modules_to_not_convert=modules_to_not_convert,
quantization_config=quantization_config,
)
if not has_been_replaced:
logger.warning(
"For some reason the model has not been properly dequantized. You might see unexpected behavior."
)
return model
def _validate_bnb_multi_backend_availability(raise_exception):
import bitsandbytes as bnb
bnb_supported_devices = getattr(bnb, "supported_torch_devices", set())
available_devices = set(get_available_devices())
if not available_devices.intersection(bnb_supported_devices):
if raise_exception:
err_msg = (
f"None of the available devices `available_devices = {available_devices or None}` are supported by the bitsandbytes version you have installed: `bnb_supported_devices = {bnb_supported_devices}`. "
"Please check the docs to see if the backend you intend to use is available and how to install it: https://huggingface.co/docs/bitsandbytes/main/en/installation"
)
logger.error(err_msg)
raise RuntimeError(err_msg)
logger.warning("No supported devices found for bitsandbytes multi-backend.")
return False
logger.debug("Multi-backend validation successful.")
return True
def _validate_bnb_cuda_backend_availability(raise_exception):
if not is_torch_available():
return False
import torch
if not torch.cuda.is_available():
log_msg = (
"CUDA is required but not available for bitsandbytes. Please consider installing the multi-platform enabled version of bitsandbytes, which is currently a work in progress. "
"Please check currently supported platforms and installation instructions at https://huggingface.co/docs/bitsandbytes/main/en/installation#multi-backend"
)
if raise_exception:
logger.error(log_msg)
raise RuntimeError(log_msg)
logger.warning(log_msg)
return False
logger.debug("CUDA backend validation successful.")
return True
def validate_bnb_backend_availability(raise_exception=False):
"""
Validates if the available devices are supported by bitsandbytes, optionally raising an exception if not.
"""
if not is_bitsandbytes_available():
if importlib.util.find_spec("bitsandbytes") and version.parse(
importlib.metadata.version("bitsandbytes")
) < version.parse("0.43.1"):
return _validate_bnb_cuda_backend_availability(raise_exception)
return False
if is_bitsandbytes_multi_backend_available():
return _validate_bnb_multi_backend_availability(raise_exception)
return _validate_bnb_cuda_backend_availability(raise_exception)
| transformers/src/transformers/integrations/bitsandbytes.py/0 | {
"file_path": "transformers/src/transformers/integrations/bitsandbytes.py",
"repo_id": "transformers",
"token_count": 10303
} | 423 |
# Copyright 2020 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Integrations with other Python libraries.
"""
import copy
import functools
import importlib.metadata
import importlib.util
import json
import numbers
import os
import pickle
import re
import shutil
import sys
import tempfile
from dataclasses import asdict, fields
from enum import Enum
from pathlib import Path
from typing import TYPE_CHECKING, Any, Literal, Optional, Union
import numpy as np
import packaging.version
from transformers.utils.import_utils import _is_package_available
if os.getenv("WANDB_MODE") == "offline":
print("⚙️ Running in WANDB offline mode")
from .. import PreTrainedModel, TFPreTrainedModel, TrainingArguments
from .. import __version__ as version
from ..utils import (
PushToHubMixin,
flatten_dict,
is_datasets_available,
is_pandas_available,
is_tf_available,
is_torch_available,
logging,
)
logger = logging.get_logger(__name__)
if is_torch_available():
import torch
import torch.distributed as dist
# comet_ml requires to be imported before any ML frameworks
_MIN_COMET_VERSION = "3.43.2"
try:
_comet_version = importlib.metadata.version("comet_ml")
_is_comet_installed = True
_is_comet_recent_enough = packaging.version.parse(_comet_version) >= packaging.version.parse(_MIN_COMET_VERSION)
# Check if the Comet API Key is set
import comet_ml
if comet_ml.config.get_config("comet.api_key") is not None:
_is_comet_configured = True
else:
_is_comet_configured = False
except (importlib.metadata.PackageNotFoundError, ImportError, ValueError, TypeError, AttributeError, KeyError):
_comet_version = None
_is_comet_installed = False
_is_comet_recent_enough = False
_is_comet_configured = False
_has_neptune = (
importlib.util.find_spec("neptune") is not None or importlib.util.find_spec("neptune-client") is not None
)
if TYPE_CHECKING and _has_neptune:
try:
_neptune_version = importlib.metadata.version("neptune")
logger.info(f"Neptune version {_neptune_version} available.")
except importlib.metadata.PackageNotFoundError:
try:
_neptune_version = importlib.metadata.version("neptune-client")
logger.info(f"Neptune-client version {_neptune_version} available.")
except importlib.metadata.PackageNotFoundError:
_has_neptune = False
from .. import modelcard # noqa: E402
from ..trainer_callback import ProgressCallback, TrainerCallback # noqa: E402
from ..trainer_utils import PREFIX_CHECKPOINT_DIR, BestRun, IntervalStrategy # noqa: E402
from ..training_args import ParallelMode # noqa: E402
from ..utils import ENV_VARS_TRUE_VALUES, is_torch_xla_available # noqa: E402
# Integration functions:
def is_wandb_available():
# any value of WANDB_DISABLED disables wandb
if os.getenv("WANDB_DISABLED", "").upper() in ENV_VARS_TRUE_VALUES:
logger.warning(
"Using the `WANDB_DISABLED` environment variable is deprecated and will be removed in v5. Use the "
"--report_to flag to control the integrations used for logging result (for instance --report_to none)."
)
return False
if importlib.util.find_spec("wandb") is not None:
import wandb
# wandb might still be detected by find_spec after an uninstall (leftover files or metadata), but not actually
# import correctly. To confirm it's fully installed and usable, we check for a key attribute like "run".
return hasattr(wandb, "run")
else:
return False
def is_trackio_available():
return importlib.util.find_spec("trackio") is not None
def is_clearml_available():
return importlib.util.find_spec("clearml") is not None
def is_comet_available():
if os.getenv("COMET_MODE", "").upper() == "DISABLED":
logger.warning(
"Using the `COMET_MODE=DISABLED` environment variable is deprecated and will be removed in v5. Use the "
"--report_to flag to control the integrations used for logging result (for instance --report_to none)."
)
return False
if _is_comet_installed is False:
return False
if _is_comet_recent_enough is False:
logger.warning(
"comet_ml version %s is installed, but version %s or higher is required. "
"Please update comet_ml to the latest version to enable Comet logging with pip install 'comet-ml>=%s'.",
_comet_version,
_MIN_COMET_VERSION,
_MIN_COMET_VERSION,
)
return False
if _is_comet_configured is False:
logger.warning(
"comet_ml is installed but the Comet API Key is not configured. "
"Please set the `COMET_API_KEY` environment variable to enable Comet logging. "
"Check out the documentation for other ways of configuring it: "
"https://www.comet.com/docs/v2/guides/experiment-management/configure-sdk/#set-the-api-key"
)
return False
return True
def is_tensorboard_available():
return importlib.util.find_spec("tensorboard") is not None or importlib.util.find_spec("tensorboardX") is not None
def is_optuna_available():
return importlib.util.find_spec("optuna") is not None
def is_ray_available():
return importlib.util.find_spec("ray") is not None
def is_ray_tune_available():
if not is_ray_available():
return False
return importlib.util.find_spec("ray.tune") is not None
def is_sigopt_available():
return importlib.util.find_spec("sigopt") is not None
def is_azureml_available():
if importlib.util.find_spec("azureml") is None:
return False
if importlib.util.find_spec("azureml.core") is None:
return False
return importlib.util.find_spec("azureml.core.run") is not None
def is_mlflow_available():
if os.getenv("DISABLE_MLFLOW_INTEGRATION", "FALSE").upper() == "TRUE":
return False
return importlib.util.find_spec("mlflow") is not None
def is_dagshub_available():
return None not in [importlib.util.find_spec("dagshub"), importlib.util.find_spec("mlflow")]
def is_neptune_available():
return _has_neptune
def is_codecarbon_available():
return importlib.util.find_spec("codecarbon") is not None
def is_flytekit_available():
return importlib.util.find_spec("flytekit") is not None
def is_flyte_deck_standard_available():
if not is_flytekit_available():
return False
return importlib.util.find_spec("flytekitplugins.deck") is not None
def is_dvclive_available():
return importlib.util.find_spec("dvclive") is not None
def is_swanlab_available():
return importlib.util.find_spec("swanlab") is not None
def hp_params(trial):
if is_optuna_available():
import optuna
if isinstance(trial, optuna.trial.BaseTrial):
return trial.params
if is_ray_tune_available():
if isinstance(trial, dict):
return trial
if is_sigopt_available():
if isinstance(trial, dict):
return trial
if is_wandb_available():
if isinstance(trial, dict):
return trial
raise RuntimeError(f"Unknown type for trial {trial.__class__}")
def run_hp_search_optuna(trainer, n_trials: int, direction: str, **kwargs) -> BestRun:
import optuna
from accelerate.utils.memory import release_memory
if trainer.args.process_index == 0:
def _objective(trial: optuna.Trial, checkpoint_dir=None):
checkpoint = None
if checkpoint_dir:
for subdir in os.listdir(checkpoint_dir):
if subdir.startswith(PREFIX_CHECKPOINT_DIR):
checkpoint = os.path.join(checkpoint_dir, subdir)
trainer.objective = None
if trainer.args.world_size > 1:
if trainer.args.parallel_mode != ParallelMode.DISTRIBUTED:
raise RuntimeError("only support DDP optuna HPO for ParallelMode.DISTRIBUTED currently.")
trainer.hp_space(trial)
fixed_trial = optuna.trial.FixedTrial(trial.params, trial.number)
trial_main_rank_list = [fixed_trial]
torch.distributed.broadcast_object_list(trial_main_rank_list, src=0)
trainer.train(resume_from_checkpoint=checkpoint, trial=trial)
else:
trainer.train(resume_from_checkpoint=checkpoint, trial=trial)
# If there hasn't been any evaluation during the training loop.
if getattr(trainer, "objective", None) is None:
metrics = trainer.evaluate()
trainer.objective = trainer.compute_objective(metrics)
# Free GPU memory
trainer.model_wrapped, trainer.model = release_memory(trainer.model_wrapped, trainer.model)
trainer.accelerator.clear()
return trainer.objective
timeout = kwargs.pop("timeout", None)
n_jobs = kwargs.pop("n_jobs", 1)
gc_after_trial = kwargs.pop("gc_after_trial", False)
directions = direction if isinstance(direction, list) else None
direction = None if directions is not None else direction
study = optuna.create_study(direction=direction, directions=directions, **kwargs)
study.optimize(_objective, n_trials=n_trials, timeout=timeout, n_jobs=n_jobs, gc_after_trial=gc_after_trial)
if not study._is_multi_objective():
best_trial = study.best_trial
return BestRun(str(best_trial.number), best_trial.value, best_trial.params)
else:
best_trials = study.best_trials
return [BestRun(str(best.number), best.values, best.params) for best in best_trials]
else:
for i in range(n_trials):
trainer.objective = None
trial_main_rank_list = [None]
if trainer.args.parallel_mode != ParallelMode.DISTRIBUTED:
raise RuntimeError("only support DDP optuna HPO for ParallelMode.DISTRIBUTED currently.")
torch.distributed.broadcast_object_list(trial_main_rank_list, src=0)
trainer.train(resume_from_checkpoint=None, trial=trial_main_rank_list[0])
# If there hasn't been any evaluation during the training loop.
if getattr(trainer, "objective", None) is None:
metrics = trainer.evaluate()
trainer.objective = trainer.compute_objective(metrics)
return None
def run_hp_search_ray(trainer, n_trials: int, direction: str, **kwargs) -> BestRun:
import ray
import ray.train
def _objective(trial: dict, local_trainer):
try:
from transformers.utils.notebook import NotebookProgressCallback
if local_trainer.pop_callback(NotebookProgressCallback):
local_trainer.add_callback(ProgressCallback)
except ModuleNotFoundError:
pass
local_trainer.objective = None
checkpoint = ray.train.get_checkpoint()
if checkpoint:
# Upon trial resume, the local_trainer's objective gets reset to None.
# If `local_trainer.train` is a noop (training has already reached
# the target number of epochs/steps), then this would
# trigger an unnecessary extra checkpoint at the end of training.
# -> Set the objective to a dummy value upon resume as a workaround.
local_trainer.objective = "objective"
with checkpoint.as_directory() as checkpoint_dir:
checkpoint_path = next(Path(checkpoint_dir).glob(f"{PREFIX_CHECKPOINT_DIR}*")).as_posix()
local_trainer.train(resume_from_checkpoint=checkpoint_path, trial=trial)
else:
local_trainer.train(trial=trial)
# If there hasn't been any evaluation during the training loop.
if getattr(local_trainer, "objective", None) is None:
metrics = local_trainer.evaluate()
local_trainer.objective = local_trainer.compute_objective(metrics)
metrics.update({"objective": local_trainer.objective, "done": True})
with tempfile.TemporaryDirectory() as temp_checkpoint_dir:
local_trainer._tune_save_checkpoint(checkpoint_dir=temp_checkpoint_dir)
checkpoint = ray.train.Checkpoint.from_directory(temp_checkpoint_dir)
ray.train.report(metrics, checkpoint=checkpoint)
if not trainer._memory_tracker.skip_memory_metrics:
from ..trainer_utils import TrainerMemoryTracker
logger.warning(
"Memory tracking for your Trainer is currently "
"enabled. Automatically disabling the memory tracker "
"since the memory tracker is not serializable."
)
trainer._memory_tracker = TrainerMemoryTracker(skip_memory_metrics=True)
# The model and TensorBoard writer do not pickle so we have to remove them (if they exists)
# while doing the ray hp search.
_tb_writer = trainer.pop_callback(TensorBoardCallback)
trainer.model = None
# Setup default `resources_per_trial`.
if "resources_per_trial" not in kwargs:
# Default to 1 CPU and 1 GPU (if applicable) per trial.
kwargs["resources_per_trial"] = {"cpu": 1}
if trainer.args.n_gpu > 0:
kwargs["resources_per_trial"]["gpu"] = 1
resource_msg = "1 CPU" + (" and 1 GPU" if trainer.args.n_gpu > 0 else "")
logger.info(
"No `resources_per_trial` arg was passed into "
"`hyperparameter_search`. Setting it to a default value "
f"of {resource_msg} for each trial."
)
# Make sure each trainer only uses GPUs that were allocated per trial.
gpus_per_trial = kwargs["resources_per_trial"].get("gpu", 0)
trainer.args._n_gpu = gpus_per_trial
# Setup default `progress_reporter`.
if "progress_reporter" not in kwargs:
from ray.tune import CLIReporter
kwargs["progress_reporter"] = CLIReporter(metric_columns=["objective"])
if "scheduler" in kwargs:
from ray.tune.schedulers import ASHAScheduler, HyperBandForBOHB, MedianStoppingRule, PopulationBasedTraining
# Check for `do_eval` and `eval_during_training` for schedulers that require intermediate reporting.
if isinstance(
kwargs["scheduler"], (ASHAScheduler, MedianStoppingRule, HyperBandForBOHB, PopulationBasedTraining)
) and (not trainer.args.do_eval or trainer.args.eval_strategy == IntervalStrategy.NO):
raise RuntimeError(
"You are using {cls} as a scheduler but you haven't enabled evaluation during training. "
"This means your trials will not report intermediate results to Ray Tune, and "
"can thus not be stopped early or used to exploit other trials parameters. "
"If this is what you want, do not use {cls}. If you would like to use {cls}, "
"make sure you pass `do_eval=True` and `eval_strategy='steps'` in the "
"Trainer `args`.".format(cls=type(kwargs["scheduler"]).__name__)
)
trainable = ray.tune.with_parameters(_objective, local_trainer=trainer)
@functools.wraps(trainable)
def dynamic_modules_import_trainable(*args, **kwargs):
"""
Wrapper around `tune.with_parameters` to ensure datasets_modules are loaded on each Actor.
Without this, an ImportError will be thrown. See https://github.com/huggingface/transformers/issues/11565.
Assumes that `_objective`, defined above, is a function.
"""
if is_datasets_available():
import datasets.load
dynamic_modules_path = os.path.join(datasets.load.init_dynamic_modules(), "__init__.py")
# load dynamic_modules from path
spec = importlib.util.spec_from_file_location("datasets_modules", dynamic_modules_path)
datasets_modules = importlib.util.module_from_spec(spec)
sys.modules[spec.name] = datasets_modules
spec.loader.exec_module(datasets_modules)
return trainable(*args, **kwargs)
# special attr set by tune.with_parameters
if hasattr(trainable, "__mixins__"):
dynamic_modules_import_trainable.__mixins__ = trainable.__mixins__
analysis = ray.tune.run(
dynamic_modules_import_trainable,
config=trainer.hp_space(None),
num_samples=n_trials,
**kwargs,
)
best_trial = analysis.get_best_trial(metric="objective", mode=direction[:3], scope=trainer.args.ray_scope)
best_run = BestRun(best_trial.trial_id, best_trial.last_result["objective"], best_trial.config, analysis)
if _tb_writer is not None:
trainer.add_callback(_tb_writer)
return best_run
def run_hp_search_sigopt(trainer, n_trials: int, direction: str, **kwargs) -> BestRun:
import sigopt
if trainer.args.process_index == 0:
if importlib.metadata.version("sigopt") >= "8.0.0":
sigopt.set_project("huggingface")
experiment = sigopt.create_experiment(
name="huggingface-tune",
type="offline",
parameters=trainer.hp_space(None),
metrics=[{"name": "objective", "objective": direction, "strategy": "optimize"}],
parallel_bandwidth=1,
budget=n_trials,
)
logger.info(f"created experiment: https://app.sigopt.com/experiment/{experiment.id}")
for run in experiment.loop():
with run:
trainer.objective = None
if trainer.args.world_size > 1:
if trainer.args.parallel_mode != ParallelMode.DISTRIBUTED:
raise RuntimeError("only support DDP Sigopt HPO for ParallelMode.DISTRIBUTED currently.")
trainer._hp_search_setup(run.run)
torch.distributed.broadcast_object_list(pickle.dumps(trainer.args), src=0)
trainer.train(resume_from_checkpoint=None)
else:
trainer.train(resume_from_checkpoint=None, trial=run.run)
# If there hasn't been any evaluation during the training loop.
if getattr(trainer, "objective", None) is None:
metrics = trainer.evaluate()
trainer.objective = trainer.compute_objective(metrics)
run.log_metric("objective", trainer.objective)
best = list(experiment.get_best_runs())[0]
best_run = BestRun(best.id, best.values["objective"].value, best.assignments)
else:
from sigopt import Connection
conn = Connection()
proxies = kwargs.pop("proxies", None)
if proxies is not None:
conn.set_proxies(proxies)
experiment = conn.experiments().create(
name="huggingface-tune",
parameters=trainer.hp_space(None),
metrics=[{"name": "objective", "objective": direction, "strategy": "optimize"}],
parallel_bandwidth=1,
observation_budget=n_trials,
project="huggingface",
)
logger.info(f"created experiment: https://app.sigopt.com/experiment/{experiment.id}")
while experiment.progress.observation_count < experiment.observation_budget:
suggestion = conn.experiments(experiment.id).suggestions().create()
trainer.objective = None
if trainer.args.world_size > 1:
if trainer.args.parallel_mode != ParallelMode.DISTRIBUTED:
raise RuntimeError("only support DDP Sigopt HPO for ParallelMode.DISTRIBUTED currently.")
trainer._hp_search_setup(suggestion)
torch.distributed.broadcast_object_list(pickle.dumps(trainer.args), src=0)
trainer.train(resume_from_checkpoint=None)
else:
trainer.train(resume_from_checkpoint=None, trial=suggestion)
# If there hasn't been any evaluation during the training loop.
if getattr(trainer, "objective", None) is None:
metrics = trainer.evaluate()
trainer.objective = trainer.compute_objective(metrics)
values = [{"name": "objective", "value": trainer.objective}]
obs = conn.experiments(experiment.id).observations().create(suggestion=suggestion.id, values=values)
logger.info(f"[suggestion_id, observation_id]: [{suggestion.id}, {obs.id}]")
experiment = conn.experiments(experiment.id).fetch()
best = list(conn.experiments(experiment.id).best_assignments().fetch().iterate_pages())[0]
best_run = BestRun(best.id, best.value, best.assignments)
return best_run
else:
for i in range(n_trials):
trainer.objective = None
args_main_rank = list(pickle.dumps(trainer.args))
if trainer.args.parallel_mode != ParallelMode.DISTRIBUTED:
raise RuntimeError("only support DDP Sigopt HPO for ParallelMode.DISTRIBUTED currently.")
torch.distributed.broadcast_object_list(args_main_rank, src=0)
args = pickle.loads(bytes(args_main_rank))
for key, value in asdict(args).items():
if key != "local_rank":
setattr(trainer.args, key, value)
trainer.train(resume_from_checkpoint=None)
# If there hasn't been any evaluation during the training loop.
if getattr(trainer, "objective", None) is None:
metrics = trainer.evaluate()
trainer.objective = trainer.compute_objective(metrics)
return None
def run_hp_search_wandb(trainer, n_trials: int, direction: str, **kwargs) -> BestRun:
from ..integrations import is_wandb_available
if not is_wandb_available():
raise ImportError("This function needs wandb installed: `pip install wandb`")
import wandb
# add WandbCallback if not already added in trainer callbacks
reporting_to_wandb = False
for callback in trainer.callback_handler.callbacks:
if isinstance(callback, WandbCallback):
reporting_to_wandb = True
break
if not reporting_to_wandb:
trainer.add_callback(WandbCallback())
trainer.args.report_to = ["wandb"]
best_trial = {"run_id": None, "objective": None, "hyperparameters": None}
sweep_id = kwargs.pop("sweep_id", None)
project = kwargs.pop("project", None)
name = kwargs.pop("name", None)
entity = kwargs.pop("entity", None)
metric = kwargs.pop("metric", "eval/loss")
sweep_config = trainer.hp_space(None)
sweep_config["metric"]["goal"] = direction
sweep_config["metric"]["name"] = metric
if name:
sweep_config["name"] = name
def _objective():
run = wandb.run if wandb.run else wandb.init()
trainer.state.trial_name = run.name
run.config.update({"assignments": {}, "metric": metric})
config = wandb.config
trainer.objective = None
trainer.train(resume_from_checkpoint=None, trial=vars(config)["_items"])
# If there hasn't been any evaluation during the training loop.
if getattr(trainer, "objective", None) is None:
metrics = trainer.evaluate()
trainer.objective = trainer.compute_objective(metrics)
format_metrics = rewrite_logs(metrics)
if metric not in format_metrics:
logger.warning(
f"Provided metric {metric} not found. This might result in unexpected sweeps charts. The available"
f" metrics are {format_metrics.keys()}"
)
best_score = False
if best_trial["run_id"] is not None:
if direction == "minimize":
best_score = trainer.objective < best_trial["objective"]
elif direction == "maximize":
best_score = trainer.objective > best_trial["objective"]
if best_score or best_trial["run_id"] is None:
best_trial["run_id"] = run.id
best_trial["objective"] = trainer.objective
best_trial["hyperparameters"] = dict(config)
return trainer.objective
if not sweep_id:
sweep_id = wandb.sweep(sweep_config, project=project, entity=entity)
else:
import wandb.env
if entity:
wandb.env.set_entity(entity)
wandb.env.set_project(project)
logger.info(f"wandb sweep id - {sweep_id}")
wandb.agent(sweep_id, function=_objective, count=n_trials)
return BestRun(best_trial["run_id"], best_trial["objective"], best_trial["hyperparameters"], sweep_id)
def get_available_reporting_integrations():
integrations = []
if is_azureml_available() and not is_mlflow_available():
integrations.append("azure_ml")
if is_comet_available():
integrations.append("comet_ml")
if is_dagshub_available():
integrations.append("dagshub")
if is_dvclive_available():
integrations.append("dvclive")
if is_mlflow_available():
integrations.append("mlflow")
if is_neptune_available():
integrations.append("neptune")
if is_tensorboard_available():
integrations.append("tensorboard")
if is_wandb_available():
integrations.append("wandb")
if is_codecarbon_available():
integrations.append("codecarbon")
if is_clearml_available():
integrations.append("clearml")
if is_swanlab_available():
integrations.append("swanlab")
if is_trackio_available():
integrations.append("trackio")
return integrations
def rewrite_logs(d):
new_d = {}
eval_prefix = "eval_"
eval_prefix_len = len(eval_prefix)
test_prefix = "test_"
test_prefix_len = len(test_prefix)
for k, v in d.items():
if k.startswith(eval_prefix):
new_d["eval/" + k[eval_prefix_len:]] = v
elif k.startswith(test_prefix):
new_d["test/" + k[test_prefix_len:]] = v
else:
new_d["train/" + k] = v
return new_d
class TensorBoardCallback(TrainerCallback):
"""
A [`TrainerCallback`] that sends the logs to [TensorBoard](https://www.tensorflow.org/tensorboard).
Args:
tb_writer (`SummaryWriter`, *optional*):
The writer to use. Will instantiate one if not set.
"""
def __init__(self, tb_writer=None):
has_tensorboard = is_tensorboard_available()
if not has_tensorboard:
raise RuntimeError(
"TensorBoardCallback requires tensorboard to be installed. Either update your PyTorch version or"
" install tensorboardX."
)
if has_tensorboard:
try:
from torch.utils.tensorboard import SummaryWriter # noqa: F401
self._SummaryWriter = SummaryWriter
except ImportError:
try:
from tensorboardX import SummaryWriter
self._SummaryWriter = SummaryWriter
except ImportError:
self._SummaryWriter = None
else:
self._SummaryWriter = None
self.tb_writer = tb_writer
def _init_summary_writer(self, args, log_dir=None):
log_dir = log_dir or args.logging_dir
if self._SummaryWriter is not None:
self.tb_writer = self._SummaryWriter(log_dir=log_dir)
def on_train_begin(self, args, state, control, **kwargs):
if not state.is_world_process_zero:
return
log_dir = None
if state.is_hyper_param_search:
trial_name = state.trial_name
if trial_name is not None:
log_dir = os.path.join(args.logging_dir, trial_name)
if self.tb_writer is None:
self._init_summary_writer(args, log_dir)
if self.tb_writer is not None:
self.tb_writer.add_text("args", args.to_json_string())
if "model" in kwargs:
model = kwargs["model"]
if hasattr(model, "config") and model.config is not None:
model_config_json = model.config.to_json_string()
self.tb_writer.add_text("model_config", model_config_json)
def on_log(self, args, state, control, logs=None, **kwargs):
if not state.is_world_process_zero:
return
if self.tb_writer is None:
self._init_summary_writer(args)
if self.tb_writer is not None:
logs = rewrite_logs(logs)
for k, v in logs.items():
if isinstance(v, (int, float)):
self.tb_writer.add_scalar(k, v, state.global_step)
elif isinstance(v, str):
self.tb_writer.add_text(k, v, state.global_step)
else:
logger.warning(
"Trainer is attempting to log a value of "
f'"{v}" of type {type(v)} for key "{k}" as a scalar. '
"This invocation of Tensorboard's writer.add_scalar() "
"is incorrect so we dropped this attribute."
)
self.tb_writer.flush()
def on_train_end(self, args, state, control, **kwargs):
if self.tb_writer:
self.tb_writer.close()
self.tb_writer = None
def save_model_architecture_to_file(model: Any, output_dir: str):
with open(f"{output_dir}/model_architecture.txt", "w+") as f:
if isinstance(model, PreTrainedModel):
print(model, file=f)
elif is_tf_available() and isinstance(model, TFPreTrainedModel):
def print_to_file(s):
print(s, file=f)
model.summary(print_fn=print_to_file)
elif is_torch_available() and (
isinstance(model, (torch.nn.Module, PushToHubMixin)) and hasattr(model, "base_model")
):
print(model, file=f)
class WandbLogModel(str, Enum):
"""Enum of possible log model values in W&B."""
CHECKPOINT = "checkpoint"
END = "end"
FALSE = "false"
@property
def is_enabled(self) -> bool:
"""Check if the value corresponds to a state where the `WANDB_LOG_MODEL` setting is enabled."""
return self in (WandbLogModel.CHECKPOINT, WandbLogModel.END)
@classmethod
def _missing_(cls, value: Any) -> "WandbLogModel":
if not isinstance(value, str):
raise TypeError(f"Expecting to have a string `WANDB_LOG_MODEL` setting, but got {type(value)}")
if value.upper() in ENV_VARS_TRUE_VALUES:
raise DeprecationWarning(
f"Setting `WANDB_LOG_MODEL` as {os.getenv('WANDB_LOG_MODEL')} is deprecated and will be removed in "
"version 5 of transformers. Use one of `'end'` or `'checkpoint'` instead."
)
logger.info(f"Setting `WANDB_LOG_MODEL` from {os.getenv('WANDB_LOG_MODEL')} to `end` instead")
return WandbLogModel.END
logger.warning(
f"Received unrecognized `WANDB_LOG_MODEL` setting value={value}; so disabling `WANDB_LOG_MODEL`"
)
return WandbLogModel.FALSE
class WandbCallback(TrainerCallback):
"""
A [`TrainerCallback`] that logs metrics, media, model checkpoints to [Weight and Biases](https://www.wandb.com/).
"""
def __init__(self):
has_wandb = is_wandb_available()
if not has_wandb:
# Check if wandb is actually installed but disabled via WANDB_DISABLED
if importlib.util.find_spec("wandb") is not None:
# wandb is installed but disabled
wandb_disabled = os.getenv("WANDB_DISABLED", "").upper() in ENV_VARS_TRUE_VALUES
if wandb_disabled:
raise RuntimeError(
"You specified `report_to='wandb'` but also set the `WANDB_DISABLED` environment variable.\n"
"This disables wandb logging, even though it was explicitly requested.\n\n"
"- To enable wandb logging: unset `WANDB_DISABLED`.\n"
"- To disable logging: use `report_to='none'`.\n\n"
"Note: WANDB_DISABLED is deprecated and will be removed in v5."
)
# If wandb is not installed at all, use the original error message
raise RuntimeError("WandbCallback requires wandb to be installed. Run `pip install wandb`.")
if has_wandb:
import wandb
self._wandb = wandb
self._initialized = False
self._log_model = WandbLogModel(os.getenv("WANDB_LOG_MODEL", "false"))
def setup(self, args, state, model, **kwargs):
"""
Setup the optional Weights & Biases (*wandb*) integration.
One can subclass and override this method to customize the setup if needed. Find more information
[here](https://docs.wandb.ai/guides/integrations/huggingface). You can also override the following environment
variables:
Environment:
- **WANDB_LOG_MODEL** (`str`, *optional*, defaults to `"false"`):
Whether to log model and checkpoints during training. Can be `"end"`, `"checkpoint"` or `"false"`. If set
to `"end"`, the model will be uploaded at the end of training. If set to `"checkpoint"`, the checkpoint
will be uploaded every `args.save_steps` . If set to `"false"`, the model will not be uploaded. Use along
with [`~transformers.TrainingArguments.load_best_model_at_end`] to upload best model.
<Deprecated version="5.0">
Setting `WANDB_LOG_MODEL` as `bool` will be deprecated in version 5 of 🤗 Transformers.
</Deprecated>
- **WANDB_WATCH** (`str`, *optional* defaults to `"false"`):
Can be `"gradients"`, `"all"`, `"parameters"`, or `"false"`. Set to `"all"` to log gradients and
parameters.
- **WANDB_PROJECT** (`str`, *optional*, defaults to `"huggingface"`):
Set this to a custom string to store results in a different project.
- **WANDB_DISABLED** (`bool`, *optional*, defaults to `False`):
Whether to disable wandb entirely. Set `WANDB_DISABLED=true` to disable.
"""
if self._wandb is None:
return
self._initialized = True
# prepare to handle potential configuration issues during setup
from wandb.sdk.lib.config_util import ConfigError as WandbConfigError
if state.is_world_process_zero:
logger.info(
'Automatic Weights & Biases logging enabled, to disable set os.environ["WANDB_DISABLED"] = "true"'
)
combined_dict = {**args.to_dict()}
if hasattr(model, "config") and model.config is not None:
model_config = model.config if isinstance(model.config, dict) else model.config.to_dict()
combined_dict = {**model_config, **combined_dict}
if hasattr(model, "peft_config") and model.peft_config is not None:
peft_config = model.peft_config
combined_dict = {**{"peft_config": peft_config}, **combined_dict}
trial_name = state.trial_name
init_args = {}
if trial_name is not None:
init_args["name"] = trial_name
init_args["group"] = args.run_name or args.output_dir
elif args.run_name is not None:
init_args["name"] = args.run_name
if args.run_name == args.output_dir:
self._wandb.termwarn(
"The `run_name` is currently set to the same value as `TrainingArguments.output_dir`. If this was "
"not intended, please specify a different run name by setting the `TrainingArguments.run_name` parameter.",
repeat=False,
)
if self._wandb.run is None:
self._wandb.init(
project=os.getenv("WANDB_PROJECT", "huggingface"),
**init_args,
)
# add config parameters (run may have been created manually)
self._wandb.config.update(combined_dict or {}, allow_val_change=True)
# define default x-axis (for latest wandb versions)
if getattr(self._wandb, "define_metric", None):
self._wandb.define_metric("train/global_step")
self._wandb.define_metric("*", step_metric="train/global_step", step_sync=True)
# keep track of model topology and gradients, unsupported on TPU
_watch_model = os.getenv("WANDB_WATCH", "false")
if not is_torch_xla_available() and _watch_model in ("all", "parameters", "gradients"):
self._wandb.watch(model, log=_watch_model, log_freq=max(100, state.logging_steps))
self._wandb.run._label(code="transformers_trainer")
# add number of model parameters to wandb config
try:
self._wandb.config["model/num_parameters"] = model.num_parameters()
except AttributeError:
logger.info(
"Could not log the number of model parameters in Weights & Biases due to an AttributeError."
)
except WandbConfigError:
logger.warning(
"A ConfigError was raised whilst setting the number of model parameters in Weights & Biases config."
)
# log the initial model architecture to an artifact
if self._log_model.is_enabled:
with tempfile.TemporaryDirectory() as temp_dir:
model_name = (
f"model-{self._wandb.run.id}"
if (args.run_name is None or args.run_name == args.output_dir)
else f"model-{self._wandb.run.name}"
)
model_artifact = self._wandb.Artifact(
name=model_name,
type="model",
metadata={
"model_config": model.config.to_dict() if hasattr(model, "config") else None,
"num_parameters": self._wandb.config.get("model/num_parameters"),
"initial_model": True,
},
)
# add the architecture to a separate text file
save_model_architecture_to_file(model, temp_dir)
for f in Path(temp_dir).glob("*"):
if f.is_file():
with model_artifact.new_file(f.name, mode="wb") as fa:
fa.write(f.read_bytes())
self._wandb.run.log_artifact(model_artifact, aliases=["base_model"])
badge_markdown = (
f'[<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge'
f'-28.svg" alt="Visualize in Weights & Biases" width="20'
f'0" height="32"/>]({self._wandb.run.url})'
)
modelcard.AUTOGENERATED_TRAINER_COMMENT += f"\n{badge_markdown}"
def on_train_begin(self, args, state, control, model=None, **kwargs):
if self._wandb is None:
return
hp_search = state.is_hyper_param_search
if hp_search:
self._wandb.finish()
self._initialized = False
args.run_name = None
if not self._initialized:
self.setup(args, state, model, **kwargs)
def on_train_end(self, args: TrainingArguments, state, control, model=None, processing_class=None, **kwargs):
if self._wandb is None:
return
if self._log_model.is_enabled and self._initialized and state.is_world_process_zero:
from ..trainer import Trainer
args_for_fake = copy.deepcopy(args)
args_for_fake.deepspeed = None
args_for_fake.deepspeed_plugin = None
fake_trainer = Trainer(
args=args_for_fake, model=model, processing_class=processing_class, eval_dataset=["fake"]
)
with tempfile.TemporaryDirectory() as temp_dir:
fake_trainer.save_model(temp_dir)
metadata = (
{
k: v
for k, v in dict(self._wandb.summary).items()
if isinstance(v, numbers.Number) and not k.startswith("_")
}
if not args.load_best_model_at_end
else {
f"eval/{args.metric_for_best_model}": state.best_metric,
"train/total_floss": state.total_flos,
"model/num_parameters": self._wandb.config.get("model/num_parameters"),
}
)
metadata["final_model"] = True
logger.info("Logging model artifacts. ...")
model_name = (
f"model-{self._wandb.run.id}"
if (args.run_name is None or args.run_name == args.output_dir)
else f"model-{self._wandb.run.name}"
)
# add the model architecture to a separate text file
save_model_architecture_to_file(model, temp_dir)
artifact = self._wandb.Artifact(name=model_name, type="model", metadata=metadata)
for f in Path(temp_dir).glob("*"):
if f.is_file():
with artifact.new_file(f.name, mode="wb") as fa:
fa.write(f.read_bytes())
self._wandb.run.log_artifact(artifact, aliases=["final_model"])
def on_log(self, args, state, control, model=None, logs=None, **kwargs):
single_value_scalars = [
"train_runtime",
"train_samples_per_second",
"train_steps_per_second",
"train_loss",
"total_flos",
]
if self._wandb is None:
return
if not self._initialized:
self.setup(args, state, model)
if state.is_world_process_zero:
for k, v in logs.items():
if k in single_value_scalars:
self._wandb.run.summary[k] = v
non_scalar_logs = {k: v for k, v in logs.items() if k not in single_value_scalars}
non_scalar_logs = rewrite_logs(non_scalar_logs)
self._wandb.log({**non_scalar_logs, "train/global_step": state.global_step})
def on_save(self, args, state, control, **kwargs):
if self._log_model == WandbLogModel.CHECKPOINT and self._initialized and state.is_world_process_zero:
checkpoint_metadata = {
k: v
for k, v in dict(self._wandb.summary).items()
if isinstance(v, numbers.Number) and not k.startswith("_")
}
checkpoint_metadata["model/num_parameters"] = self._wandb.config.get("model/num_parameters")
ckpt_dir = f"checkpoint-{state.global_step}"
artifact_path = os.path.join(args.output_dir, ckpt_dir)
logger.info(f"Logging checkpoint artifacts in {ckpt_dir}. ...")
checkpoint_name = (
f"model-{self._wandb.run.id}"
if (args.run_name is None or args.run_name == args.output_dir)
else f"model-{self._wandb.run.name}"
)
artifact = self._wandb.Artifact(name=checkpoint_name, type="model", metadata=checkpoint_metadata)
artifact.add_dir(artifact_path)
self._wandb.log_artifact(
artifact, aliases=[f"epoch_{round(state.epoch, 2)}", f"checkpoint_global_step_{state.global_step}"]
)
def on_predict(self, args, state, control, metrics, **kwargs):
if self._wandb is None:
return
if not self._initialized:
self.setup(args, state, **kwargs)
if state.is_world_process_zero:
metrics = rewrite_logs(metrics)
self._wandb.log(metrics)
class TrackioCallback(TrainerCallback):
"""
A [`TrainerCallback`] that logs metrics to Trackio.
It records training metrics, model (and PEFT) configuration, and GPU memory usage.
If `nvidia-ml-py` is installed, GPU power consumption is also tracked.
**Requires**:
```bash
pip install trackio
```
"""
def __init__(self):
has_trackio = is_trackio_available()
if not has_trackio:
raise RuntimeError("TrackioCallback requires trackio to be installed. Run `pip install trackio`.")
if has_trackio:
import trackio
self._trackio = trackio
self._initialized = False
def setup(self, args, state, model, **kwargs):
"""
Setup the optional Trackio integration.
To customize the setup you can also override the following environment variables:
Environment:
- **TRACKIO_PROJECT** (`str`, *optional*, defaults to `"huggingface"`):
The name of the project (can be an existing project to continue tracking or a new project to start tracking
from scratch).
- **TRACKIO_SPACE_ID** (`str`, *optional*, defaults to `None`):
If set, the project will be logged to a Hugging Face Space instead of a local directory. Should be a
complete Space name like `"username/reponame"` or `"orgname/reponame"`, or just `"reponame" in which case
the Space will be created in the currently-logged-in Hugging Face user's namespace. If the Space does not
exist, it will be created. If the Space already exists, the project will be logged to it.
"""
if state.is_world_process_zero:
combined_dict = {**args.to_dict()}
if hasattr(model, "config") and model.config is not None:
model_config = model.config if isinstance(model.config, dict) else model.config.to_dict()
combined_dict = {**model_config, **combined_dict}
if hasattr(model, "peft_config") and model.peft_config is not None:
peft_config = model.peft_config
combined_dict = {**{"peft_config": peft_config}, **combined_dict}
self._trackio.init(
project=os.getenv("TRACKIO_PROJECT", "huggingface"),
name=args.run_name,
space_id=os.getenv("TRACKIO_SPACE_ID", None),
resume="allow",
)
# Add config parameters (run may have been created manually)
self._trackio.config.update(combined_dict, allow_val_change=True)
# Add number of model parameters to trackio config
try:
self._trackio.config["model/num_parameters"] = model.num_parameters()
except AttributeError:
logger.info("Could not log the number of model parameters in Trackio due to an AttributeError.")
self._initialized = True
def on_train_begin(self, args, state, control, model=None, **kwargs):
if not self._initialized:
self.setup(args, state, model, **kwargs)
def on_train_end(self, args: TrainingArguments, state, control, model=None, processing_class=None, **kwargs):
if state.is_world_process_zero and self._initialized:
self._trackio.finish()
def on_log(self, args, state, control, model=None, logs=None, **kwargs):
single_value_scalars = [
"train_runtime",
"train_samples_per_second",
"train_steps_per_second",
"train_loss",
"total_flos",
]
if is_torch_available() and torch.cuda.is_available():
device_idx = torch.cuda.current_device()
total_memory = torch.cuda.get_device_properties(device_idx).total_memory
memory_allocated = torch.cuda.memory_allocated(device_idx)
gpu_memory_logs = {
f"gpu/{device_idx}/allocated_memory": memory_allocated / (1024**3), # GB
f"gpu/{device_idx}/memory_usage": memory_allocated / total_memory, # ratio
}
if _is_package_available("pynvml"):
power = torch.cuda.power_draw(device_idx)
gpu_memory_logs[f"gpu/{device_idx}/power"] = power / 1000 # Watts
if dist.is_available() and dist.is_initialized():
gathered_logs = [None] * dist.get_world_size()
dist.all_gather_object(gathered_logs, gpu_memory_logs)
gpu_memory_logs = {k: v for d in gathered_logs for k, v in d.items()}
else:
gpu_memory_logs = {}
if not self._initialized:
self.setup(args, state, model)
if state.is_world_process_zero:
non_scalar_logs = {k: v for k, v in logs.items() if k not in single_value_scalars}
non_scalar_logs = rewrite_logs(non_scalar_logs)
self._trackio.log({**non_scalar_logs, **gpu_memory_logs, "train/global_step": state.global_step})
def on_save(self, args, state, control, **kwargs):
return
def on_predict(self, args, state, control, metrics, **kwargs):
if self._trackio is None:
return
if not self._initialized:
self.setup(args, state, **kwargs)
if state.is_world_process_zero:
metrics = rewrite_logs(metrics)
self._trackio.log(metrics)
class CometCallback(TrainerCallback):
"""
A [`TrainerCallback`] that sends the logs to [Comet ML](https://www.comet.com/site/).
"""
def __init__(self):
if _is_comet_installed is False or _is_comet_recent_enough is False:
raise RuntimeError(
f"CometCallback requires comet-ml>={_MIN_COMET_VERSION} to be installed. Run `pip install comet-ml>={_MIN_COMET_VERSION}`."
)
self._initialized = False
self._log_assets = False
self._experiment = None
def setup(self, args, state, model):
"""
Setup the optional Comet integration.
Environment:
- **COMET_MODE** (`str`, *optional*, default to `get_or_create`):
Control whether to create and log to a new Comet experiment or append to an existing experiment.
It accepts the following values:
* `get_or_create`: Decides automatically depending if
`COMET_EXPERIMENT_KEY` is set and whether an Experiment
with that key already exists or not.
* `create`: Always create a new Comet Experiment.
* `get`: Always try to append to an Existing Comet Experiment.
Requires `COMET_EXPERIMENT_KEY` to be set.
* `ONLINE`: **deprecated**, used to create an online
Experiment. Use `COMET_START_ONLINE=1` instead.
* `OFFLINE`: **deprecated**, used to created an offline
Experiment. Use `COMET_START_ONLINE=0` instead.
* `DISABLED`: **deprecated**, used to disable Comet logging.
Use the `--report_to` flag to control the integrations used
for logging result instead.
- **COMET_PROJECT_NAME** (`str`, *optional*):
Comet project name for experiments.
- **COMET_LOG_ASSETS** (`str`, *optional*, defaults to `TRUE`):
Whether or not to log training assets (tf event logs, checkpoints, etc), to Comet. Can be `TRUE`, or
`FALSE`.
For a number of configurable items in the environment, see
[here](https://www.comet.com/docs/v2/guides/experiment-management/configure-sdk/#explore-comet-configuration-options).
"""
self._initialized = True
log_assets = os.getenv("COMET_LOG_ASSETS", "FALSE").upper()
if log_assets in {"TRUE", "1"}:
self._log_assets = True
if state.is_world_process_zero:
comet_old_mode = os.getenv("COMET_MODE")
mode = None
online = None
if comet_old_mode is not None:
comet_old_mode = comet_old_mode.lower()
if comet_old_mode == "online":
online = True
elif comet_old_mode == "offline":
online = False
elif comet_old_mode in ("get", "get_or_create", "create"):
mode = comet_old_mode
elif comet_old_mode:
logger.warning("Invalid COMET_MODE env value %r, Comet logging is disabled", comet_old_mode)
return
# For HPO, we always create a new experiment for each trial
if state.is_hyper_param_search:
if mode is not None:
logger.warning(
"Hyperparameter Search is enabled, forcing the creation of new experiments, COMET_MODE value %r is ignored",
comet_old_mode,
)
mode = "create"
import comet_ml
experiment_config = comet_ml.ExperimentConfig(name=args.run_name)
self._experiment = comet_ml.start(online=online, mode=mode, experiment_config=experiment_config)
self._experiment.__internal_api__set_model_graph__(model, framework="transformers")
params = {"args": args.to_dict()}
if hasattr(model, "config") and model.config is not None:
model_config = model.config.to_dict()
params["config"] = model_config
if hasattr(model, "peft_config") and model.peft_config is not None:
peft_config = model.peft_config
params["peft_config"] = peft_config
self._experiment.__internal_api__log_parameters__(
params, framework="transformers", source="manual", flatten_nested=True
)
if state.is_hyper_param_search:
optimization_id = getattr(state, "trial_name", None)
optimization_params = getattr(state, "trial_params", None)
self._experiment.log_optimization(optimization_id=optimization_id, parameters=optimization_params)
def on_train_begin(self, args, state, control, model=None, **kwargs):
if not self._initialized:
self.setup(args, state, model)
def on_log(self, args, state, control, model=None, logs=None, **kwargs):
if not self._initialized:
self.setup(args, state, model)
if state.is_world_process_zero:
if self._experiment is not None:
rewritten_logs = rewrite_logs(logs)
self._experiment.__internal_api__log_metrics__(
rewritten_logs, step=state.global_step, epoch=state.epoch, framework="transformers"
)
def on_train_end(self, args, state, control, **kwargs):
if self._initialized and state.is_world_process_zero:
if self._experiment is not None:
if self._log_assets is True:
logger.info("Logging checkpoints. This may take time.")
self._experiment.log_asset_folder(
args.output_dir, recursive=True, log_file_name=True, step=state.global_step
)
# We create one experiment per trial in HPO mode
if state.is_hyper_param_search:
self._experiment.clean()
self._initialized = False
def on_predict(self, args, state, control, metrics, **kwargs):
if not self._initialized:
self.setup(args, state, model=None)
if state.is_world_process_zero and self._experiment is not None:
rewritten_metrics = rewrite_logs(metrics)
self._experiment.__internal_api__log_metrics__(
rewritten_metrics, step=state.global_step, epoch=state.epoch, framework="transformers"
)
class AzureMLCallback(TrainerCallback):
"""
A [`TrainerCallback`] that sends the logs to [AzureML](https://pypi.org/project/azureml-sdk/).
"""
def __init__(self, azureml_run=None):
if not is_azureml_available():
raise RuntimeError("AzureMLCallback requires azureml to be installed. Run `pip install azureml-sdk`.")
self.azureml_run = azureml_run
def on_init_end(self, args, state, control, **kwargs):
from azureml.core.run import Run
if self.azureml_run is None and state.is_world_process_zero:
self.azureml_run = Run.get_context()
def on_log(self, args, state, control, logs=None, **kwargs):
if self.azureml_run and state.is_world_process_zero:
for k, v in logs.items():
if isinstance(v, (int, float)):
self.azureml_run.log(k, v, description=k)
class MLflowCallback(TrainerCallback):
"""
A [`TrainerCallback`] that sends the logs to [MLflow](https://www.mlflow.org/). Can be disabled by setting
environment variable `DISABLE_MLFLOW_INTEGRATION = TRUE`.
"""
def __init__(self):
if not is_mlflow_available():
raise RuntimeError("MLflowCallback requires mlflow to be installed. Run `pip install mlflow`.")
import mlflow
self._MAX_PARAM_VAL_LENGTH = mlflow.utils.validation.MAX_PARAM_VAL_LENGTH
self._MAX_PARAMS_TAGS_PER_BATCH = mlflow.utils.validation.MAX_PARAMS_TAGS_PER_BATCH
self._initialized = False
self._auto_end_run = False
self._log_artifacts = False
self._ml_flow = mlflow
def setup(self, args, state, model):
"""
Setup the optional MLflow integration.
Environment:
- **HF_MLFLOW_LOG_ARTIFACTS** (`str`, *optional*):
Whether to use MLflow `.log_artifact()` facility to log artifacts. This only makes sense if logging to a
remote server, e.g. s3 or GCS. If set to `True` or *1*, will copy each saved checkpoint on each save in
[`TrainingArguments`]'s `output_dir` to the local or remote artifact storage. Using it without a remote
storage will just copy the files to your artifact location.
- **MLFLOW_TRACKING_URI** (`str`, *optional*):
Whether to store runs at a specific path or remote server. Unset by default, which skips setting the
tracking URI entirely.
- **MLFLOW_EXPERIMENT_NAME** (`str`, *optional*, defaults to `None`):
Whether to use an MLflow experiment_name under which to launch the run. Default to `None` which will point
to the `Default` experiment in MLflow. Otherwise, it is a case sensitive name of the experiment to be
activated. If an experiment with this name does not exist, a new experiment with this name is created.
- **MLFLOW_TAGS** (`str`, *optional*):
A string dump of a dictionary of key/value pair to be added to the MLflow run as tags. Example:
`os.environ['MLFLOW_TAGS']='{"release.candidate": "RC1", "release.version": "2.2.0"}'`.
- **MLFLOW_NESTED_RUN** (`str`, *optional*):
Whether to use MLflow nested runs. If set to `True` or *1*, will create a nested run inside the current
run.
- **MLFLOW_RUN_ID** (`str`, *optional*):
Allow to reattach to an existing run which can be useful when resuming training from a checkpoint. When
`MLFLOW_RUN_ID` environment variable is set, `start_run` attempts to resume a run with the specified run ID
and other parameters are ignored.
- **MLFLOW_FLATTEN_PARAMS** (`str`, *optional*, defaults to `False`):
Whether to flatten the parameters dictionary before logging.
- **MLFLOW_MAX_LOG_PARAMS** (`int`, *optional*):
Set the maximum number of parameters to log in the run.
"""
self._log_artifacts = os.getenv("HF_MLFLOW_LOG_ARTIFACTS", "FALSE").upper() in ENV_VARS_TRUE_VALUES
self._nested_run = os.getenv("MLFLOW_NESTED_RUN", "FALSE").upper() in ENV_VARS_TRUE_VALUES
self._tracking_uri = os.getenv("MLFLOW_TRACKING_URI", None)
self._experiment_name = os.getenv("MLFLOW_EXPERIMENT_NAME", None)
self._flatten_params = os.getenv("MLFLOW_FLATTEN_PARAMS", "FALSE").upper() in ENV_VARS_TRUE_VALUES
self._run_id = os.getenv("MLFLOW_RUN_ID", None)
self._max_log_params = os.getenv("MLFLOW_MAX_LOG_PARAMS", None)
# "synchronous" flag is only available with mlflow version >= 2.8.0
# https://github.com/mlflow/mlflow/pull/9705
# https://github.com/mlflow/mlflow/releases/tag/v2.8.0
self._async_log = packaging.version.parse(self._ml_flow.__version__) >= packaging.version.parse("2.8.0")
logger.debug(
f"MLflow experiment_name={self._experiment_name}, run_name={args.run_name}, nested={self._nested_run},"
f" tracking_uri={self._tracking_uri}"
)
if state.is_world_process_zero:
if not self._ml_flow.is_tracking_uri_set():
if self._tracking_uri:
self._ml_flow.set_tracking_uri(self._tracking_uri)
logger.debug(f"MLflow tracking URI is set to {self._tracking_uri}")
else:
logger.debug(
"Environment variable `MLFLOW_TRACKING_URI` is not provided and therefore will not be"
" explicitly set."
)
else:
logger.debug(f"MLflow tracking URI is set to {self._ml_flow.get_tracking_uri()}")
if self._ml_flow.active_run() is None or self._nested_run or self._run_id:
if self._experiment_name:
# Use of set_experiment() ensure that Experiment is created if not exists
self._ml_flow.set_experiment(self._experiment_name)
self._ml_flow.start_run(run_name=args.run_name, nested=self._nested_run)
logger.debug(f"MLflow run started with run_id={self._ml_flow.active_run().info.run_id}")
self._auto_end_run = True
combined_dict = args.to_dict()
if hasattr(model, "config") and model.config is not None:
model_config = model.config.to_dict()
combined_dict = {**model_config, **combined_dict}
combined_dict = flatten_dict(combined_dict) if self._flatten_params else combined_dict
# remove params that are too long for MLflow
for name, value in list(combined_dict.items()):
# internally, all values are converted to str in MLflow
if len(str(value)) > self._MAX_PARAM_VAL_LENGTH:
logger.warning(
f'Trainer is attempting to log a value of "{value}" for key "{name}" as a parameter. MLflow\'s'
" log_param() only accepts values no longer than 250 characters so we dropped this attribute."
" You can use `MLFLOW_FLATTEN_PARAMS` environment variable to flatten the parameters and"
" avoid this message."
)
del combined_dict[name]
# MLflow cannot log more than 100 values in one go, so we have to split it
combined_dict_items = list(combined_dict.items())
if self._max_log_params and self._max_log_params.isdigit():
max_log_params = int(self._max_log_params)
if max_log_params < len(combined_dict_items):
logger.debug(
f"Reducing the number of parameters to log from {len(combined_dict_items)} to {max_log_params}."
)
combined_dict_items = combined_dict_items[:max_log_params]
for i in range(0, len(combined_dict_items), self._MAX_PARAMS_TAGS_PER_BATCH):
if self._async_log:
self._ml_flow.log_params(
dict(combined_dict_items[i : i + self._MAX_PARAMS_TAGS_PER_BATCH]), synchronous=False
)
else:
self._ml_flow.log_params(dict(combined_dict_items[i : i + self._MAX_PARAMS_TAGS_PER_BATCH]))
mlflow_tags = os.getenv("MLFLOW_TAGS", None)
if mlflow_tags:
mlflow_tags = json.loads(mlflow_tags)
self._ml_flow.set_tags(mlflow_tags)
self._initialized = True
def on_train_begin(self, args, state, control, model=None, **kwargs):
if not self._initialized:
self.setup(args, state, model)
def on_log(self, args, state, control, logs, model=None, **kwargs):
if not self._initialized:
self.setup(args, state, model)
if state.is_world_process_zero:
metrics = {}
for k, v in logs.items():
if isinstance(v, (int, float)):
metrics[k] = v
elif isinstance(v, torch.Tensor) and v.numel() == 1:
metrics[k] = v.item()
else:
logger.warning(
f'Trainer is attempting to log a value of "{v}" of type {type(v)} for key "{k}" as a metric. '
"MLflow's log_metric() only accepts float and int types so we dropped this attribute."
)
# sanitize metric names to replace unsupported characters like parentheses
sanitized_metrics = {re.sub(r"[^0-9A-Za-z_\-\.\ :/]", "_", k): v for k, v in metrics.items()}
if self._async_log:
self._ml_flow.log_metrics(metrics=sanitized_metrics, step=state.global_step, synchronous=False)
else:
self._ml_flow.log_metrics(metrics=sanitized_metrics, step=state.global_step)
def on_train_end(self, args, state, control, **kwargs):
if self._initialized and state.is_world_process_zero:
if self._auto_end_run and self._ml_flow.active_run():
self._ml_flow.end_run()
def on_save(self, args, state, control, **kwargs):
if self._initialized and state.is_world_process_zero and self._log_artifacts:
ckpt_dir = f"checkpoint-{state.global_step}"
artifact_path = os.path.join(args.output_dir, ckpt_dir)
logger.info(f"Logging checkpoint artifacts in {ckpt_dir}. This may take time.")
self._ml_flow.pyfunc.log_model(
ckpt_dir,
artifacts={"model_path": artifact_path},
python_model=self._ml_flow.pyfunc.PythonModel(),
)
def __del__(self):
# if the previous run is not terminated correctly, the fluent API will
# not let you start a new run before the previous one is killed
if (
self._auto_end_run
and callable(getattr(self._ml_flow, "active_run", None))
and self._ml_flow.active_run() is not None
):
self._ml_flow.end_run()
class DagsHubCallback(MLflowCallback):
"""
A [`TrainerCallback`] that logs to [DagsHub](https://dagshub.com/). Extends [`MLflowCallback`]
"""
def __init__(self):
super().__init__()
if not is_dagshub_available():
raise ImportError("DagsHubCallback requires dagshub to be installed. Run `pip install dagshub`.")
from dagshub.upload import Repo
self.Repo = Repo
def setup(self, *args, **kwargs):
"""
Setup the DagsHub's Logging integration.
Environment:
- **HF_DAGSHUB_LOG_ARTIFACTS** (`str`, *optional*):
Whether to save the data and model artifacts for the experiment. Default to `False`.
"""
self.log_artifacts = os.getenv("HF_DAGSHUB_LOG_ARTIFACTS", "FALSE").upper() in ENV_VARS_TRUE_VALUES
self.name = os.getenv("HF_DAGSHUB_MODEL_NAME") or "main"
self.remote = os.getenv("MLFLOW_TRACKING_URI")
self.repo = self.Repo(
owner=self.remote.split(os.sep)[-2],
name=self.remote.split(os.sep)[-1].split(".")[0],
branch=os.getenv("BRANCH") or "main",
)
self.path = Path("artifacts")
if self.remote is None:
raise RuntimeError(
"DagsHubCallback requires the `MLFLOW_TRACKING_URI` environment variable to be set. Did you run"
" `dagshub.init()`?"
)
super().setup(*args, **kwargs)
def on_train_end(self, args, state, control, **kwargs):
if self.log_artifacts:
if getattr(self, "train_dataloader", None):
torch.save(self.train_dataloader.dataset, os.path.join(args.output_dir, "dataset.pt"))
self.repo.directory(str(self.path)).add_dir(args.output_dir)
class NeptuneMissingConfiguration(Exception):
def __init__(self):
super().__init__(
"""
------ Unsupported ---- We were not able to create new runs. You provided a custom Neptune run to
`NeptuneCallback` with the `run` argument. For the integration to work fully, provide your `api_token` and
`project` by saving them as environment variables or passing them to the callback.
"""
)
class NeptuneCallback(TrainerCallback):
"""TrainerCallback that sends the logs to [Neptune](https://app.neptune.ai).
Args:
api_token (`str`, *optional*): Neptune API token obtained upon registration.
You can leave this argument out if you have saved your token to the `NEPTUNE_API_TOKEN` environment
variable (strongly recommended). See full setup instructions in the
[docs](https://docs.neptune.ai/setup/installation).
project (`str`, *optional*): Name of an existing Neptune project, in the form "workspace-name/project-name".
You can find and copy the name in Neptune from the project settings -> Properties. If None (default), the
value of the `NEPTUNE_PROJECT` environment variable is used.
name (`str`, *optional*): Custom name for the run.
base_namespace (`str`, *optional*, defaults to "finetuning"): In the Neptune run, the root namespace
that will contain all of the metadata logged by the callback.
log_parameters (`bool`, *optional*, defaults to `True`):
If True, logs all Trainer arguments and model parameters provided by the Trainer.
log_checkpoints (`str`, *optional*): If "same", uploads checkpoints whenever they are saved by the Trainer.
If "last", uploads only the most recently saved checkpoint. If "best", uploads the best checkpoint (among
the ones saved by the Trainer). If `None`, does not upload checkpoints.
run (`Run`, *optional*): Pass a Neptune run object if you want to continue logging to an existing run.
Read more about resuming runs in the [docs](https://docs.neptune.ai/logging/to_existing_object).
**neptune_run_kwargs (*optional*):
Additional keyword arguments to be passed directly to the
[`neptune.init_run()`](https://docs.neptune.ai/api/neptune#init_run) function when a new run is created.
For instructions and examples, see the [Transformers integration
guide](https://docs.neptune.ai/integrations/transformers) in the Neptune documentation.
"""
integration_version_key = "source_code/integrations/transformers"
model_parameters_key = "model_parameters"
trial_name_key = "trial"
trial_params_key = "trial_params"
trainer_parameters_key = "trainer_parameters"
flat_metrics = {"train/epoch"}
def __init__(
self,
*,
api_token: Optional[str] = None,
project: Optional[str] = None,
name: Optional[str] = None,
base_namespace: str = "finetuning",
run=None,
log_parameters: bool = True,
log_checkpoints: Optional[str] = None,
**neptune_run_kwargs,
):
if not is_neptune_available():
raise ValueError(
"NeptuneCallback requires the Neptune client library to be installed. "
"To install the library, run `pip install neptune`."
)
try:
from neptune import Run
from neptune.internal.utils import verify_type
except ImportError:
from neptune.new.internal.utils import verify_type
from neptune.new.metadata_containers.run import Run
verify_type("api_token", api_token, (str, type(None)))
verify_type("project", project, (str, type(None)))
verify_type("name", name, (str, type(None)))
verify_type("base_namespace", base_namespace, str)
verify_type("run", run, (Run, type(None)))
verify_type("log_parameters", log_parameters, bool)
verify_type("log_checkpoints", log_checkpoints, (str, type(None)))
self._base_namespace_path = base_namespace
self._log_parameters = log_parameters
self._log_checkpoints = log_checkpoints
self._initial_run: Optional[Run] = run
self._run = None
self._is_monitoring_run = False
self._run_id = None
self._force_reset_monitoring_run = False
self._init_run_kwargs = {"api_token": api_token, "project": project, "name": name, **neptune_run_kwargs}
self._volatile_checkpoints_dir = None
self._should_upload_checkpoint = self._log_checkpoints is not None
self._recent_checkpoint_path = None
if self._log_checkpoints in {"last", "best"}:
self._target_checkpoints_namespace = f"checkpoints/{self._log_checkpoints}"
self._should_clean_recently_uploaded_checkpoint = True
else:
self._target_checkpoints_namespace = "checkpoints"
self._should_clean_recently_uploaded_checkpoint = False
def _stop_run_if_exists(self):
if self._run:
self._run.stop()
del self._run
self._run = None
def _initialize_run(self, **additional_neptune_kwargs):
try:
from neptune import init_run
from neptune.exceptions import NeptuneMissingApiTokenException, NeptuneMissingProjectNameException
except ImportError:
from neptune.new import init_run
from neptune.new.exceptions import NeptuneMissingApiTokenException, NeptuneMissingProjectNameException
self._stop_run_if_exists()
try:
run_params = additional_neptune_kwargs.copy()
run_params.update(self._init_run_kwargs)
self._run = init_run(**run_params)
self._run_id = self._run["sys/id"].fetch()
except (NeptuneMissingProjectNameException, NeptuneMissingApiTokenException) as e:
raise NeptuneMissingConfiguration() from e
def _use_initial_run(self):
self._run = self._initial_run
self._is_monitoring_run = True
self._run_id = self._run["sys/id"].fetch()
self._initial_run = None
def _ensure_run_with_monitoring(self):
if self._initial_run is not None:
self._use_initial_run()
else:
if not self._force_reset_monitoring_run and self._is_monitoring_run:
return
if self._run and not self._is_monitoring_run and not self._force_reset_monitoring_run:
self._initialize_run(with_id=self._run_id)
self._is_monitoring_run = True
else:
self._initialize_run()
self._force_reset_monitoring_run = False
def _ensure_at_least_run_without_monitoring(self):
if self._initial_run is not None:
self._use_initial_run()
else:
if not self._run:
self._initialize_run(
with_id=self._run_id,
capture_stdout=False,
capture_stderr=False,
capture_hardware_metrics=False,
capture_traceback=False,
)
self._is_monitoring_run = False
@property
def run(self):
if self._run is None:
self._ensure_at_least_run_without_monitoring()
return self._run
@property
def _metadata_namespace(self):
return self.run[self._base_namespace_path]
def _log_integration_version(self):
self.run[NeptuneCallback.integration_version_key] = version
def _log_trainer_parameters(self, args):
self._metadata_namespace[NeptuneCallback.trainer_parameters_key] = args.to_sanitized_dict()
def _log_model_parameters(self, model):
from neptune.utils import stringify_unsupported
if model and hasattr(model, "config") and model.config is not None:
self._metadata_namespace[NeptuneCallback.model_parameters_key] = stringify_unsupported(
model.config.to_dict()
)
def _log_hyper_param_search_parameters(self, state):
if state and hasattr(state, "trial_name"):
self._metadata_namespace[NeptuneCallback.trial_name_key] = state.trial_name
if state and hasattr(state, "trial_params") and state.trial_params is not None:
self._metadata_namespace[NeptuneCallback.trial_params_key] = state.trial_params
def _log_model_checkpoint(self, source_directory: str, checkpoint: str):
target_path = relative_path = os.path.join(source_directory, checkpoint)
if self._volatile_checkpoints_dir is not None:
consistent_checkpoint_path = os.path.join(self._volatile_checkpoints_dir, checkpoint)
try:
# Remove leading ../ from a relative path.
cpkt_path = relative_path.replace("..", "").lstrip(os.path.sep)
copy_path = os.path.join(consistent_checkpoint_path, cpkt_path)
shutil.copytree(relative_path, copy_path)
target_path = consistent_checkpoint_path
except OSError as e:
logger.warning(
f"NeptuneCallback was unable to made a copy of checkpoint due to I/O exception: '{e}'. "
"Could fail trying to upload."
)
self._metadata_namespace[self._target_checkpoints_namespace].upload_files(target_path)
if self._should_clean_recently_uploaded_checkpoint and self._recent_checkpoint_path is not None:
self._metadata_namespace[self._target_checkpoints_namespace].delete_files(self._recent_checkpoint_path)
self._recent_checkpoint_path = relative_path
def on_init_end(self, args, state, control, **kwargs):
self._volatile_checkpoints_dir = None
if self._log_checkpoints and (args.overwrite_output_dir or args.save_total_limit is not None):
self._volatile_checkpoints_dir = tempfile.TemporaryDirectory().name
if self._log_checkpoints == "best" and not args.load_best_model_at_end:
raise ValueError("To save the best model checkpoint, the load_best_model_at_end argument must be enabled.")
def on_train_begin(self, args, state, control, model=None, **kwargs):
if not state.is_world_process_zero:
return
self._ensure_run_with_monitoring()
self._force_reset_monitoring_run = True
self._log_integration_version()
if self._log_parameters:
self._log_trainer_parameters(args)
self._log_model_parameters(model)
if state.is_hyper_param_search:
self._log_hyper_param_search_parameters(state)
def on_train_end(self, args, state, control, **kwargs):
self._stop_run_if_exists()
def __del__(self):
if self._volatile_checkpoints_dir is not None:
shutil.rmtree(self._volatile_checkpoints_dir, ignore_errors=True)
self._stop_run_if_exists()
def on_save(self, args, state, control, **kwargs):
if self._should_upload_checkpoint:
self._log_model_checkpoint(args.output_dir, f"checkpoint-{state.global_step}")
def on_evaluate(self, args, state, control, metrics=None, **kwargs):
if self._log_checkpoints == "best":
best_metric_name = args.metric_for_best_model
if not best_metric_name.startswith("eval_"):
best_metric_name = f"eval_{best_metric_name}"
metric_value = metrics.get(best_metric_name)
operator = np.greater if args.greater_is_better else np.less
self._should_upload_checkpoint = state.best_metric is None or operator(metric_value, state.best_metric)
@classmethod
def get_run(cls, trainer):
for callback in trainer.callback_handler.callbacks:
if isinstance(callback, cls):
return callback.run
raise Exception("The trainer doesn't have a NeptuneCallback configured.")
def on_log(self, args, state, control, logs: Optional[dict[str, float]] = None, **kwargs):
if not state.is_world_process_zero:
return
if logs is not None:
for name, value in rewrite_logs(logs).items():
if isinstance(value, (int, float)):
if name in NeptuneCallback.flat_metrics:
self._metadata_namespace[name] = value
else:
self._metadata_namespace[name].log(value, step=state.global_step)
class CodeCarbonCallback(TrainerCallback):
"""
A [`TrainerCallback`] that tracks the CO2 emission of training.
"""
def __init__(self):
if not is_codecarbon_available():
raise RuntimeError(
"CodeCarbonCallback requires `codecarbon` to be installed. Run `pip install codecarbon`."
)
elif torch.version.hip:
raise RuntimeError(
"CodeCarbonCallback requires `codecarbon` package, which is not compatible with AMD ROCm (https://github.com/mlco2/codecarbon/pull/490). When using the Trainer, please specify the `report_to` argument (https://huggingface.co/docs/transformers/v4.39.3/en/main_classes/trainer#transformers.TrainingArguments.report_to) to disable CodeCarbonCallback."
)
import codecarbon
self._codecarbon = codecarbon
self.tracker = None
def on_init_end(self, args, state, control, **kwargs):
if self.tracker is None and state.is_local_process_zero:
# CodeCarbon will automatically handle environment variables for configuration
self.tracker = self._codecarbon.EmissionsTracker(output_dir=args.output_dir)
def on_train_begin(self, args, state, control, model=None, **kwargs):
if self.tracker and state.is_local_process_zero:
self.tracker.start()
def on_train_end(self, args, state, control, **kwargs):
if self.tracker and state.is_local_process_zero:
self.tracker.stop()
class ClearMLCallback(TrainerCallback):
"""
A [`TrainerCallback`] that sends the logs to [ClearML](https://clear.ml/).
Environment:
- **CLEARML_PROJECT** (`str`, *optional*, defaults to `HuggingFace Transformers`):
ClearML project name.
- **CLEARML_TASK** (`str`, *optional*, defaults to `Trainer`):
ClearML task name.
- **CLEARML_LOG_MODEL** (`bool`, *optional*, defaults to `False`):
Whether to log models as artifacts during training.
"""
log_suffix = ""
_hparams_section = "Transformers"
_model_config_section = "Model Configuration"
_ignore_hparams_overrides = "_ignore_hparams_ui_overrides_"
_ignoge_model_config_overrides = "_ignore_model_config_ui_overrides_"
_model_config_description = "The configuration of model number {}."
_model_config_description_note = (
"Note that, when cloning this task and running it remotely,"
" the configuration might be applied to another model instead of this one."
" To avoid this, initialize the task externally by calling `Task.init`"
" before the `ClearMLCallback` is instantiated."
)
_train_run_counter = 0
_model_connect_counter = 0
_task_created_in_callback = False
_should_close_on_train_end = None
def __init__(self):
if is_clearml_available():
import clearml
self._clearml = clearml
else:
raise RuntimeError("ClearMLCallback requires 'clearml' to be installed. Run `pip install clearml`.")
self._initialized = False
self._clearml_task = None
self._log_model = False
self._checkpoints_saved = []
def setup(self, args, state, model, processing_class, **kwargs):
if self._clearml is None:
return
if self._initialized:
return
ClearMLCallback._train_run_counter += 1
ClearMLCallback._model_connect_counter += 1
ClearMLCallback.log_suffix = (
"" if ClearMLCallback._train_run_counter == 1 else "_" + str(ClearMLCallback._train_run_counter)
)
if state.is_world_process_zero:
logger.info("Automatic ClearML logging enabled.")
if self._clearml_task is None:
if ClearMLCallback._should_close_on_train_end is None:
if not self._clearml.Task.running_locally() or self._clearml.Task.current_task():
ClearMLCallback._should_close_on_train_end = False
else:
ClearMLCallback._should_close_on_train_end = True
# This might happen when running inside of a pipeline, where the task is already initialized
# from outside of Hugging Face
if self._clearml.Task.running_locally() and self._clearml.Task.current_task():
self._clearml_task = self._clearml.Task.current_task()
self._log_model = os.getenv(
"CLEARML_LOG_MODEL",
"FALSE" if not ClearMLCallback._task_created_in_callback else "TRUE",
).upper() in ENV_VARS_TRUE_VALUES.union({"TRUE"})
logger.info("External ClearML Task has been connected.")
else:
self._clearml_task = self._clearml.Task.init(
project_name=os.getenv("CLEARML_PROJECT", "HuggingFace Transformers"),
task_name=os.getenv("CLEARML_TASK", "Trainer"),
auto_connect_frameworks={"tensorboard": False, "pytorch": False},
output_uri=True,
)
self._log_model = os.getenv("CLEARML_LOG_MODEL", "TRUE").upper() in ENV_VARS_TRUE_VALUES.union(
{"TRUE"}
)
ClearMLCallback._task_created_in_callback = True
logger.info("ClearML Task has been initialized.")
self._initialized = True
suffixed_hparams_section = ClearMLCallback._hparams_section + ClearMLCallback.log_suffix
ignore_hparams_config_section = suffixed_hparams_section + "/" + ClearMLCallback._ignore_hparams_overrides
if self._clearml.Task.running_locally():
self._copy_training_args_as_hparams(args, suffixed_hparams_section)
self._clearml_task.set_parameter(
name=ignore_hparams_config_section,
value=True,
value_type=bool,
description=(
"If True, ignore Transformers hyperparameters overrides done in the UI/backend "
+ "when running remotely. Otherwise, the overrides will be applied when running remotely"
),
)
elif not self._clearml_task.get_parameter(ignore_hparams_config_section, default=True, cast=True):
self._clearml_task.connect(args, suffixed_hparams_section)
else:
self._copy_training_args_as_hparams(
args, ClearMLCallback._hparams_section + ClearMLCallback.log_suffix
)
if getattr(model, "config", None) is not None:
ignore_model_config_section = (
suffixed_hparams_section + "/" + ClearMLCallback._ignoge_model_config_overrides
)
configuration_object_description = ClearMLCallback._model_config_description.format(
ClearMLCallback._model_connect_counter
)
if ClearMLCallback._model_connect_counter != ClearMLCallback._train_run_counter:
configuration_object_description += " " + ClearMLCallback._model_config_description_note
if self._clearml.Task.running_locally():
self._clearml_task.set_parameter(
name=ignore_model_config_section,
value=True,
value_type=bool,
description=(
"If True, ignore Transformers model configuration overrides done in the UI/backend "
+ "when running remotely. Otherwise, the overrides will be applied when running remotely"
),
)
self._clearml_task.set_configuration_object(
name=ClearMLCallback._model_config_section + ClearMLCallback.log_suffix,
config_dict=model.config.to_dict(),
description=configuration_object_description,
)
elif not self._clearml_task.get_parameter(ignore_model_config_section, default=True, cast=True):
model.config = model.config.from_dict(
self._clearml_task.get_configuration_object_as_dict(
ClearMLCallback._model_config_section + ClearMLCallback.log_suffix
)
)
else:
self._clearml_task.set_configuration_object(
name=ClearMLCallback._model_config_section + ClearMLCallback.log_suffix,
config_dict=model.config.to_dict(),
description=configuration_object_description,
)
def on_train_begin(self, args, state, control, model=None, processing_class=None, **kwargs):
if self._clearml is None:
return
self._checkpoints_saved = []
if state.is_hyper_param_search:
self._initialized = False
if not self._initialized:
self.setup(args, state, model, processing_class, **kwargs)
def on_train_end(self, args, state, control, **kwargs):
if ClearMLCallback._should_close_on_train_end:
self._clearml_task.close()
ClearMLCallback._train_run_counter = 0
def on_log(self, args, state, control, model=None, processing_class=None, logs=None, **kwargs):
if self._clearml is None:
return
if not self._initialized:
self.setup(args, state, model, processing_class, **kwargs)
if state.is_world_process_zero:
eval_prefix = "eval_"
eval_prefix_len = len(eval_prefix)
test_prefix = "test_"
test_prefix_len = len(test_prefix)
single_value_scalars = [
"train_runtime",
"train_samples_per_second",
"train_steps_per_second",
"train_loss",
"total_flos",
"epoch",
]
for k, v in logs.items():
if isinstance(v, (int, float)):
if k in single_value_scalars:
self._clearml_task.get_logger().report_single_value(
name=k + ClearMLCallback.log_suffix, value=v
)
elif k.startswith(eval_prefix):
self._clearml_task.get_logger().report_scalar(
title="eval" + ClearMLCallback.log_suffix,
series=k[eval_prefix_len:],
value=v,
iteration=state.global_step,
)
elif k.startswith(test_prefix):
self._clearml_task.get_logger().report_scalar(
title="test" + ClearMLCallback.log_suffix,
series=k[test_prefix_len:],
value=v,
iteration=state.global_step,
)
else:
self._clearml_task.get_logger().report_scalar(
title="train" + ClearMLCallback.log_suffix,
series=k,
value=v,
iteration=state.global_step,
)
else:
logger.warning(
"Trainer is attempting to log a value of "
f'"{v}" of type {type(v)} for key "{k}" as a scalar. '
"This invocation of ClearML logger's report_scalar() "
"is incorrect so we dropped this attribute."
)
def on_save(self, args, state, control, **kwargs):
if self._log_model and self._clearml_task and state.is_world_process_zero:
ckpt_dir = f"checkpoint-{state.global_step}"
artifact_path = os.path.join(args.output_dir, ckpt_dir)
name = ckpt_dir + ClearMLCallback.log_suffix
logger.info(f"Logging checkpoint artifact `{name}`. This may take some time.")
output_model = self._clearml.OutputModel(task=self._clearml_task, name=name)
output_model.connect(task=self._clearml_task, name=name)
output_model.update_weights_package(
weights_path=artifact_path,
target_filename=ckpt_dir,
iteration=state.global_step,
auto_delete_file=False,
)
self._checkpoints_saved.append(output_model)
while args.save_total_limit and args.save_total_limit < len(self._checkpoints_saved):
try:
self._clearml.model.Model.remove(
self._checkpoints_saved[0],
delete_weights_file=True,
force=True,
raise_on_errors=True,
)
except Exception as e:
logger.warning(
f"Could not remove checkpoint `{self._checkpoints_saved[0].name}` after going over the `save_total_limit`. Error is: {e}"
)
break
self._checkpoints_saved = self._checkpoints_saved[1:]
def _copy_training_args_as_hparams(self, training_args, prefix):
as_dict = {
field.name: getattr(training_args, field.name)
for field in fields(training_args)
if field.init and not field.name.endswith("_token")
}
flat_dict = {str(k): v for k, v in self._clearml.utilities.proxy_object.flatten_dictionary(as_dict).items()}
self._clearml_task._arguments.copy_from_dict(flat_dict, prefix=prefix)
class FlyteCallback(TrainerCallback):
"""A [`TrainerCallback`] that sends the logs to [Flyte](https://flyte.org/).
NOTE: This callback only works within a Flyte task.
Args:
save_log_history (`bool`, *optional*, defaults to `True`):
When set to True, the training logs are saved as a Flyte Deck.
sync_checkpoints (`bool`, *optional*, defaults to `True`):
When set to True, checkpoints are synced with Flyte and can be used to resume training in the case of an
interruption.
Example:
```python
# Note: This example skips over some setup steps for brevity.
from flytekit import current_context, task
@task
def train_hf_transformer():
cp = current_context().checkpoint
trainer = Trainer(..., callbacks=[FlyteCallback()])
output = trainer.train(resume_from_checkpoint=cp.restore())
```
"""
def __init__(self, save_log_history: bool = True, sync_checkpoints: bool = True):
super().__init__()
if not is_flytekit_available():
raise ImportError("FlyteCallback requires flytekit to be installed. Run `pip install flytekit`.")
if not is_flyte_deck_standard_available() or not is_pandas_available():
logger.warning(
"Syncing log history requires both flytekitplugins-deck-standard and pandas to be installed. "
"Run `pip install flytekitplugins-deck-standard pandas` to enable this feature."
)
save_log_history = False
from flytekit import current_context
self.cp = current_context().checkpoint
self.save_log_history = save_log_history
self.sync_checkpoints = sync_checkpoints
def on_save(self, args, state, control, **kwargs):
if self.sync_checkpoints and state.is_world_process_zero:
ckpt_dir = f"checkpoint-{state.global_step}"
artifact_path = os.path.join(args.output_dir, ckpt_dir)
logger.info(f"Syncing checkpoint in {ckpt_dir} to Flyte. This may take time.")
self.cp.save(artifact_path)
def on_train_end(self, args, state, control, **kwargs):
if self.save_log_history:
import pandas as pd
from flytekit import Deck
from flytekitplugins.deck.renderer import TableRenderer
log_history_df = pd.DataFrame(state.log_history)
Deck("Log History", TableRenderer().to_html(log_history_df))
class DVCLiveCallback(TrainerCallback):
"""
A [`TrainerCallback`] that sends the logs to [DVCLive](https://www.dvc.org/doc/dvclive).
Use the environment variables below in `setup` to configure the integration. To customize this callback beyond
those environment variables, see [here](https://dvc.org/doc/dvclive/ml-frameworks/huggingface).
Args:
live (`dvclive.Live`, *optional*, defaults to `None`):
Optional Live instance. If None, a new instance will be created using **kwargs.
log_model (Union[Literal["all"], bool], *optional*, defaults to `None`):
Whether to use `dvclive.Live.log_artifact()` to log checkpoints created by [`Trainer`]. If set to `True`,
the final checkpoint is logged at the end of training. If set to `"all"`, the entire
[`TrainingArguments`]'s `output_dir` is logged at each checkpoint.
"""
def __init__(
self,
live: Optional[Any] = None,
log_model: Optional[Union[Literal["all"], bool]] = None,
**kwargs,
):
if not is_dvclive_available():
raise RuntimeError("DVCLiveCallback requires dvclive to be installed. Run `pip install dvclive`.")
from dvclive import Live
self._initialized = False
self.live = None
if isinstance(live, Live):
self.live = live
elif live is not None:
raise RuntimeError(f"Found class {live.__class__} for live, expected dvclive.Live")
self._log_model = log_model
if self._log_model is None:
log_model_env = os.getenv("HF_DVCLIVE_LOG_MODEL", "FALSE")
if log_model_env.upper() in ENV_VARS_TRUE_VALUES:
self._log_model = True
elif log_model_env.lower() == "all":
self._log_model = "all"
def setup(self, args, state, model):
"""
Setup the optional DVCLive integration. To customize this callback beyond the environment variables below, see
[here](https://dvc.org/doc/dvclive/ml-frameworks/huggingface).
Environment:
- **HF_DVCLIVE_LOG_MODEL** (`str`, *optional*):
Whether to use `dvclive.Live.log_artifact()` to log checkpoints created by [`Trainer`]. If set to `True` or
*1*, the final checkpoint is logged at the end of training. If set to `all`, the entire
[`TrainingArguments`]'s `output_dir` is logged at each checkpoint.
"""
from dvclive import Live
self._initialized = True
if state.is_world_process_zero:
if not self.live:
self.live = Live()
self.live.log_params(args.to_dict())
def on_train_begin(self, args, state, control, model=None, **kwargs):
if not self._initialized:
self.setup(args, state, model)
def on_log(self, args, state, control, model=None, logs=None, **kwargs):
if not self._initialized:
self.setup(args, state, model)
if state.is_world_process_zero:
from dvclive.plots import Metric
from dvclive.utils import standardize_metric_name
for key, value in logs.items():
if Metric.could_log(value):
self.live.log_metric(standardize_metric_name(key, "dvclive.huggingface"), value)
else:
logger.warning(
"Trainer is attempting to log a value of "
f'"{value}" of type {type(value)} for key "{key}" as a scalar. '
"This invocation of DVCLive's Live.log_metric() "
"is incorrect so we dropped this attribute."
)
self.live.next_step()
def on_save(self, args, state, control, **kwargs):
if self._log_model == "all" and self._initialized and state.is_world_process_zero:
self.live.log_artifact(args.output_dir)
def on_train_end(self, args, state, control, **kwargs):
if self._initialized and state.is_world_process_zero:
from transformers.trainer import Trainer
if self._log_model is True:
fake_trainer = Trainer(
args=args,
model=kwargs.get("model"),
processing_class=kwargs.get("processing_class"),
eval_dataset=["fake"],
)
name = "best" if args.load_best_model_at_end else "last"
output_dir = os.path.join(args.output_dir, name)
fake_trainer.save_model(output_dir)
self.live.log_artifact(output_dir, name=name, type="model", copy=True)
self.live.end()
class SwanLabCallback(TrainerCallback):
"""
A [`TrainerCallback`] that logs metrics, media, model checkpoints to [SwanLab](https://swanlab.cn/).
"""
def __init__(self):
if not is_swanlab_available():
raise RuntimeError("SwanLabCallback requires swanlab to be installed. Run `pip install swanlab`.")
import swanlab
self._swanlab = swanlab
self._initialized = False
self._log_model = os.getenv("SWANLAB_LOG_MODEL", None)
def setup(self, args, state, model, **kwargs):
"""
Setup the optional SwanLab (*swanlab*) integration.
One can subclass and override this method to customize the setup if needed. Find more information
[here](https://docs.swanlab.cn/guide_cloud/integration/integration-huggingface-transformers.html).
You can also override the following environment variables. Find more information about environment
variables [here](https://docs.swanlab.cn/en/api/environment-variable.html#environment-variables)
Environment:
- **SWANLAB_API_KEY** (`str`, *optional*, defaults to `None`):
Cloud API Key. During login, this environment variable is checked first. If it doesn't exist, the system
checks if the user is already logged in. If not, the login process is initiated.
- If a string is passed to the login interface, this environment variable is ignored.
- If the user is already logged in, this environment variable takes precedence over locally stored
login information.
- **SWANLAB_PROJECT** (`str`, *optional*, defaults to `None`):
Set this to a custom string to store results in a different project. If not specified, the name of the current
running directory is used.
- **SWANLAB_LOG_DIR** (`str`, *optional*, defaults to `swanlog`):
This environment variable specifies the storage path for log files when running in local mode.
By default, logs are saved in a folder named swanlog under the working directory.
- **SWANLAB_MODE** (`Literal["local", "cloud", "disabled"]`, *optional*, defaults to `cloud`):
SwanLab's parsing mode, which involves callbacks registered by the operator. Currently, there are three modes:
local, cloud, and disabled. Note: Case-sensitive. Find more information
[here](https://docs.swanlab.cn/en/api/py-init.html#swanlab-init)
- **SWANLAB_LOG_MODEL** (`str`, *optional*, defaults to `None`):
SwanLab does not currently support the save mode functionality.This feature will be available in a future
release
- **SWANLAB_WEB_HOST** (`str`, *optional*, defaults to `None`):
Web address for the SwanLab cloud environment for private version (its free)
- **SWANLAB_API_HOST** (`str`, *optional*, defaults to `None`):
API address for the SwanLab cloud environment for private version (its free)
"""
self._initialized = True
if state.is_world_process_zero:
logger.info('Automatic SwanLab logging enabled, to disable set os.environ["SWANLAB_MODE"] = "disabled"')
combined_dict = {**args.to_dict()}
if hasattr(model, "config") and model.config is not None:
model_config = model.config if isinstance(model.config, dict) else model.config.to_dict()
combined_dict = {**model_config, **combined_dict}
if hasattr(model, "peft_config") and model.peft_config is not None:
peft_config = model.peft_config
combined_dict = {**{"peft_config": peft_config}, **combined_dict}
trial_name = state.trial_name
init_args = {}
if trial_name is not None and args.run_name is not None:
init_args["experiment_name"] = f"{args.run_name}-{trial_name}"
elif args.run_name is not None:
init_args["experiment_name"] = args.run_name
elif trial_name is not None:
init_args["experiment_name"] = trial_name
init_args["project"] = os.getenv("SWANLAB_PROJECT", None)
if self._swanlab.get_run() is None:
self._swanlab.init(
**init_args,
)
# show transformers logo!
self._swanlab.config["FRAMEWORK"] = "🤗transformers"
# add config parameters (run may have been created manually)
self._swanlab.config.update(combined_dict)
# add number of model parameters to swanlab config
try:
self._swanlab.config.update({"model_num_parameters": model.num_parameters()})
# get peft model parameters
if type(model).__name__ == "PeftModel" or type(model).__name__ == "PeftMixedModel":
trainable_params, all_param = model.get_nb_trainable_parameters()
self._swanlab.config.update({"peft_model_trainable_params": trainable_params})
self._swanlab.config.update({"peft_model_all_param": all_param})
except AttributeError:
logger.info("Could not log the number of model parameters in SwanLab due to an AttributeError.")
# log the initial model architecture to an artifact
if self._log_model is not None:
logger.warning(
"SwanLab does not currently support the save mode functionality. "
"This feature will be available in a future release."
)
badge_markdown = (
f'[<img src="https://raw.githubusercontent.com/SwanHubX/assets/main/badge1.svg"'
f' alt="Visualize in SwanLab" height="28'
f'0" height="32"/>]({self._swanlab.get_run().public.cloud.exp_url})'
)
modelcard.AUTOGENERATED_TRAINER_COMMENT += f"\n{badge_markdown}"
def on_train_begin(self, args, state, control, model=None, **kwargs):
if not self._initialized:
self.setup(args, state, model, **kwargs)
def on_train_end(self, args, state, control, model=None, processing_class=None, **kwargs):
if self._log_model is not None and self._initialized and state.is_world_process_zero:
logger.warning(
"SwanLab does not currently support the save mode functionality. "
"This feature will be available in a future release."
)
def on_log(self, args, state, control, model=None, logs=None, **kwargs):
single_value_scalars = [
"train_runtime",
"train_samples_per_second",
"train_steps_per_second",
"train_loss",
"total_flos",
]
if not self._initialized:
self.setup(args, state, model)
if state.is_world_process_zero:
for k, v in logs.items():
if k in single_value_scalars:
self._swanlab.log({f"single_value/{k}": v}, step=state.global_step)
non_scalar_logs = {k: v for k, v in logs.items() if k not in single_value_scalars}
non_scalar_logs = rewrite_logs(non_scalar_logs)
self._swanlab.log({**non_scalar_logs, "train/global_step": state.global_step}, step=state.global_step)
def on_save(self, args, state, control, **kwargs):
if self._log_model is not None and self._initialized and state.is_world_process_zero:
logger.warning(
"SwanLab does not currently support the save mode functionality. "
"This feature will be available in a future release."
)
def on_predict(self, args, state, control, metrics, **kwargs):
if not self._initialized:
self.setup(args, state, **kwargs)
if state.is_world_process_zero:
metrics = rewrite_logs(metrics)
self._swanlab.log(metrics)
INTEGRATION_TO_CALLBACK = {
"azure_ml": AzureMLCallback,
"comet_ml": CometCallback,
"mlflow": MLflowCallback,
"neptune": NeptuneCallback,
"tensorboard": TensorBoardCallback,
"trackio": TrackioCallback,
"wandb": WandbCallback,
"codecarbon": CodeCarbonCallback,
"clearml": ClearMLCallback,
"dagshub": DagsHubCallback,
"flyte": FlyteCallback,
"dvclive": DVCLiveCallback,
"swanlab": SwanLabCallback,
}
def get_reporting_integration_callbacks(report_to):
if report_to is None:
return []
if isinstance(report_to, str):
if "none" == report_to:
return []
elif "all" == report_to:
report_to = get_available_reporting_integrations()
else:
report_to = [report_to]
for integration in report_to:
if integration not in INTEGRATION_TO_CALLBACK:
raise ValueError(
f"{integration} is not supported, only {', '.join(INTEGRATION_TO_CALLBACK.keys())} are supported."
)
return [INTEGRATION_TO_CALLBACK[integration] for integration in report_to]
| transformers/src/transformers/integrations/integration_utils.py/0 | {
"file_path": "transformers/src/transformers/integrations/integration_utils.py",
"repo_id": "transformers",
"token_count": 50940
} | 424 |
#include <torch/extension.h>
#include "ATen/ATen.h"
typedef at::BFloat16 bf16;
void cuda_forward(int B, int T, int C, float *w, float *u, float *k, float *v, float *y);
void cuda_forward_bf16(int B, int T, int C, float *w, bf16 *u, bf16 *k, bf16 *v, bf16 *y);
void cuda_forward_with_state(int B, int T, int C, float *w, float *u, float *k, float *v, float *y, float *s);
void cuda_forward_with_state_bf16(int B, int T, int C, float *w, bf16 *u, bf16 *k, bf16 *v, bf16 *y, float *s);
void cuda_backward(int B, int T, int C, float *w, float *u, float *k, float *v, float *y, float *gy, float *gw, float *gu, float *gk, float *gv);
void cuda_backward_bf16(int B, int T, int C, float *w, bf16 *u, bf16 *k, bf16 *v, bf16 *y, bf16 *gy, bf16 *gw, bf16 *gu, bf16 *gk, bf16 *gv);
void forward(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y) {
const int B = k.size(0);
const int T = k.size(1);
const int C = k.size(2);
cuda_forward(B, T, C, w.data_ptr<float>(), u.data_ptr<float>(), k.data_ptr<float>(), v.data_ptr<float>(), y.data_ptr<float>());
}
void forward_bf16(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y) {
const int B = k.size(0);
const int T = k.size(1);
const int C = k.size(2);
cuda_forward_bf16(B, T, C, w.data_ptr<float>(), u.data_ptr<bf16>(), k.data_ptr<bf16>(), v.data_ptr<bf16>(), y.data_ptr<bf16>());
}
void forward_with_state(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y, torch::Tensor &s) {
const int B = k.size(0);
const int T = k.size(1);
const int C = k.size(2);
cuda_forward_with_state(B, T, C, w.data_ptr<float>(), u.data_ptr<float>(), k.data_ptr<float>(), v.data_ptr<float>(), y.data_ptr<float>(), s.data_ptr<float>());
}
void forward_with_state_bf16(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y, torch::Tensor &s) {
const int B = k.size(0);
const int T = k.size(1);
const int C = k.size(2);
cuda_forward_with_state_bf16(B, T, C, w.data_ptr<float>(), u.data_ptr<bf16>(), k.data_ptr<bf16>(), v.data_ptr<bf16>(), y.data_ptr<bf16>(), s.data_ptr<float>());
}
void backward(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y, torch::Tensor &gy, torch::Tensor &gw, torch::Tensor &gu, torch::Tensor &gk, torch::Tensor &gv) {
const int B = k.size(0);
const int T = k.size(1);
const int C = k.size(2);
cuda_backward(B, T, C, w.data_ptr<float>(), u.data_ptr<float>(), k.data_ptr<float>(), v.data_ptr<float>(), y.data_ptr<float>(), gy.data_ptr<float>(), gw.data_ptr<float>(), gu.data_ptr<float>(), gk.data_ptr<float>(), gv.data_ptr<float>());
}
void backward_bf16(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y, torch::Tensor &gy, torch::Tensor &gw, torch::Tensor &gu, torch::Tensor &gk, torch::Tensor &gv) {
const int B = k.size(0);
const int T = k.size(1);
const int C = k.size(2);
cuda_backward_bf16(B, T, C, w.data_ptr<float>(), u.data_ptr<bf16>(), k.data_ptr<bf16>(), v.data_ptr<bf16>(), y.data_ptr<bf16>(),
gy.data_ptr<bf16>(), gw.data_ptr<bf16>(), gu.data_ptr<bf16>(), gk.data_ptr<bf16>(), gv.data_ptr<bf16>());
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &forward, "wkv forward");
m.def("forward_bf16", &forward_bf16, "wkv forward bf16");
m.def("forward_with_state", &forward_with_state, "wkv forward with state");
m.def("forward_with_state_bf16", &forward_with_state_bf16, "wkv forward with state bf16");
m.def("backward", &backward, "wkv backward");
m.def("backward_bf16", &backward_bf16, "wkv backward bf16");
}
TORCH_LIBRARY(wkv, m) {
m.def("forward", forward);
m.def("forward_bf16", forward_bf16);
m.def("forward_with_state", forward_with_state);
m.def("forward_with_state_bf16", forward_with_state_bf16);
m.def("backward", backward);
m.def("backward_bf16", backward_bf16);
}
| transformers/src/transformers/kernels/rwkv/wkv_op.cpp/0 | {
"file_path": "transformers/src/transformers/kernels/rwkv/wkv_op.cpp",
"repo_id": "transformers",
"token_count": 1807
} | 425 |
# coding=utf-8
# Copyright 2025 HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import itertools
from typing import Callable, Optional, Union
import torch
import torch.nn.functional as F
from .cache_utils import Cache
from .configuration_utils import PretrainedConfig
from .utils import is_torch_xpu_available, logging
from .utils.generic import GeneralInterface
from .utils.import_utils import is_torch_flex_attn_available, is_torch_greater_or_equal, is_torchdynamo_compiling
if is_torch_flex_attn_available():
from torch.nn.attention.flex_attention import _DEFAULT_SPARSE_BLOCK_SIZE as flex_default_block_size # noqa: N811
from torch.nn.attention.flex_attention import BlockMask, create_block_mask
else:
# Register a fake type to avoid crashing for annotations and `isinstance` checks
BlockMask = torch.Tensor
_is_torch_greater_or_equal_than_2_5 = is_torch_greater_or_equal("2.5", accept_dev=True)
_is_torch_greater_or_equal_than_2_6 = is_torch_greater_or_equal("2.6", accept_dev=True)
_is_torch_xpu_available = is_torch_xpu_available()
if _is_torch_greater_or_equal_than_2_6:
from torch._dynamo._trace_wrapped_higher_order_op import TransformGetItemToIndex
logger = logging.get_logger(__name__)
def and_masks(*mask_functions: list[Callable]) -> Callable:
"""Returns a mask function that is the intersection of provided mask functions"""
if not all(callable(arg) for arg in mask_functions):
raise RuntimeError(f"All inputs should be callable mask_functions: {mask_functions}")
def and_mask(batch_idx, head_idx, q_idx, kv_idx):
result = q_idx.new_ones((), dtype=torch.bool)
for mask in mask_functions:
result = result & mask(batch_idx, head_idx, q_idx, kv_idx).to(result.device)
return result
return and_mask
def or_masks(*mask_functions: list[Callable]) -> Callable:
"""Returns a mask function that is the union of provided mask functions"""
if not all(callable(arg) for arg in mask_functions):
raise RuntimeError(f"All inputs should be callable mask_functions: {mask_functions}")
def or_mask(batch_idx, head_idx, q_idx, kv_idx):
result = q_idx.new_zeros((), dtype=torch.bool)
for mask in mask_functions:
result = result | mask(batch_idx, head_idx, q_idx, kv_idx).to(result.device)
return result
return or_mask
def causal_mask_function(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
"""
This creates a basic lower-diagonal causal mask.
"""
return kv_idx <= q_idx
def sliding_window_overlay(sliding_window: int) -> Callable:
"""
This is an overlay depicting a sliding window pattern. Add it on top of a causal mask for a proper sliding
window mask.
"""
def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
return kv_idx > q_idx - sliding_window
return inner_mask
def chunked_overlay(chunk_size: int, left_padding: torch.Tensor) -> Callable:
"""
This is an overlay depicting a chunked attention pattern. Add it on top of a causal mask for a proper chunked
attention mask.
"""
def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
return (kv_idx - left_padding[batch_idx]) // chunk_size == (q_idx - left_padding[batch_idx]) // chunk_size
return inner_mask
def _legacy_chunked_overlay(chunk_size: int) -> Callable:
"""
Same as the above function, but do not correctly account for left padding tokens.
Only kept for compatibility with older torch versions (< 2.6).
"""
def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
return kv_idx // chunk_size == q_idx // chunk_size
return inner_mask
def sliding_window_causal_mask_function(sliding_window: int) -> Callable:
"""
This return the mask_function function to create a sliding window mask.
"""
return and_masks(sliding_window_overlay(sliding_window), causal_mask_function)
def chunked_causal_mask_function(chunk_size: int, left_padding: torch.Tensor) -> Callable:
"""
This return the mask_function function to create a chunked attention mask.
"""
if not _is_torch_greater_or_equal_than_2_6:
return and_masks(_legacy_chunked_overlay(chunk_size), causal_mask_function)
return and_masks(chunked_overlay(chunk_size, left_padding), causal_mask_function)
def padding_mask_function(padding_mask: torch.Tensor) -> Callable:
"""
This return the mask_function function corresponding to a 2D padding mask.
"""
def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
# Note that here the mask should ALWAYS be at least of the max `kv_index` size in the dimension 1. This is because
# we cannot pad it here in the mask_function as we don't know the final size, and we cannot try/except, as it is not
# vectorizable on accelerator devices
return padding_mask[batch_idx, kv_idx]
return inner_mask
def packed_sequence_mask_function(packed_sequence_mask: torch.Tensor) -> Callable:
"""
This return the mask_function function corresponding to a 2D packed sequence mask.
"""
def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
return packed_sequence_mask[batch_idx, q_idx] == packed_sequence_mask[batch_idx, kv_idx]
return inner_mask
def add_offsets_to_mask_function(mask_function: Callable, q_offset: int, kv_offset: int) -> Callable:
"""
This function adds the correct offsets to the `q_idx` and `kv_idx` as the torch API can only accept lengths,
not start and end indices.
"""
def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
return mask_function(batch_idx, head_idx, q_idx + q_offset, kv_idx + kv_offset)
return inner_mask
def _vmap_for_bhqkv(mask_function: Callable, bh_indices: bool = True) -> Callable:
"""
Used to vmap our mask_functions over the q_idx and kv_idx dimensions of the inputs. Optionally, vmap over
the batch and head indices as well if `bh_indices=True`.
Using vmap here allows us to keep the performance of vectorized ops, while having a single set of primitive
functions between attention interfaces (i.e. between flex and sdpa/eager, FA2 being a bit different).
Args:
mask_function (`Callable`):
The mask_function to vmap.
bh_indices (`bool`, optional):
Whether to vmap over the batch and head indices as well, or only q and kv indices.
Returns:
Callable: The vmapped function.
"""
# We vmap the function 2 times, broadcasting the [q_idx, kv_idx] dimensions
dimensions = [(None, None, None, 0), (None, None, 0, None)]
if bh_indices:
# We extend broadcasting over the [batch_idx, head_idx] dimensions
dimensions.extend([(None, 0, None, None), (0, None, None, None)])
for dims in dimensions:
mask_function = torch.vmap(mask_function, in_dims=dims, out_dims=0)
return mask_function
def prepare_padding_mask(
attention_mask: Optional[torch.Tensor], kv_length: int, kv_offset: int, _slice: bool = True
) -> Optional[torch.Tensor]:
"""
From the 2D attention mask, prepare the correct padding mask to use by potentially padding it, and slicing
according to the `kv_offset` if `_slice` is `True`.
"""
local_padding_mask = attention_mask
if attention_mask is not None:
# Pad it if necessary
if (padding_length := kv_length + kv_offset - attention_mask.shape[-1]) > 0:
local_padding_mask = torch.nn.functional.pad(attention_mask, (0, padding_length))
# For flex, we should not slice them, only use an offset
if _slice:
# Equivalent to: `local_padding_mask = attention_mask[:, kv_offset : kv_offset + kv_length]`,
# but without data-dependent slicing (i.e. torch.compile friendly)
mask_indices = torch.arange(kv_length, device=local_padding_mask.device)
mask_indices += kv_offset
local_padding_mask = local_padding_mask[:, mask_indices]
return local_padding_mask
def _ignore_causal_mask_sdpa(
padding_mask: Optional[torch.Tensor],
query_length: int,
kv_length: int,
kv_offset: int,
local_attention_size: Optional[int] = None,
) -> bool:
"""
Detects whether the causal mask can be ignored in case PyTorch's SDPA is used, rather relying on SDPA's `is_causal` argument.
In case no token is masked in the 2D `padding_mask` argument, if `query_length == 1` or
`key_value_length == query_length`, we rather rely on SDPA `is_causal` argument to use causal/non-causal masks,
allowing to dispatch to the flash attention kernel (that can otherwise not be used if a custom `attn_mask` is
passed).
"""
is_tracing = torch.jit.is_tracing() or isinstance(padding_mask, torch.fx.Proxy) or is_torchdynamo_compiling()
if padding_mask is not None and padding_mask.shape[-1] > kv_length:
mask_indices = torch.arange(kv_length, device=padding_mask.device)
mask_indices += kv_offset
padding_mask = padding_mask[:, mask_indices]
# When using `torch.export` or `torch.onnx.dynamo_export`, we must pass an example input, and `is_causal` behavior is
# hard-coded to the forward. If a user exports a model with query_length > 1, the exported model will hard-code `is_causal=True`
# which is in general wrong (see https://github.com/pytorch/pytorch/issues/108108). Thus, we only set
# `ignore_causal_mask = True` if we are not tracing
if (
not is_tracing
# only cases when lower and upper diags are the same, see https://github.com/pytorch/pytorch/issues/108108
and (query_length == 1 or (kv_length == query_length or _is_torch_xpu_available))
# in this case we need to add special patterns to the mask so cannot be skipped otherwise
and (local_attention_size is None or kv_length < local_attention_size)
# In this case, we need to add padding to the mask, so cannot be skipped otherwise
and (
padding_mask is None
or (
padding_mask.all()
if not _is_torch_xpu_available or query_length == 1
else padding_mask[:, :query_length].all()
)
)
):
return True
return False
def sdpa_mask_recent_torch(
batch_size: int,
cache_position: torch.Tensor,
kv_length: int,
kv_offset: int = 0,
mask_function: Callable = causal_mask_function,
attention_mask: Optional[torch.Tensor] = None,
local_size: Optional[int] = None,
allow_is_causal_skip: bool = True,
**kwargs,
) -> Optional[torch.Tensor]:
"""
Create a 4D boolean mask of shape `(batch_size, 1, query_length, kv_length)` where a value of True indicates that
the element should take part in the attention computation, and False that it should not.
This function can only be used with torch>=2.5, as the context manager is otherwise not available.
Args:
batch_size (`int`):
The batch size of the input sequence.
cache_position (`torch.Tensor`):
A tensor of shape (query_length,) indicating the current indices of the input sequence elements.
kv_length (`int`):
The size that the key and value states will have during the attention computation.
kv_offset (`int`, optional):
An optional offset to indicate at which first position the key and values states will refer to.
mask_function (`Callable`):
The mask factory function describing the mask pattern.
attention_mask (`torch.Tensor`, optional):
The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length)
local_size (`int`, optional):
The size of the local attention, if we do not use full attention. This is used only if `allow_is_causal_skip=True`
to try to skip mask creation if possible.
allow_is_causal_skip (`bool`, optional):
Whether to allow to return `None` for the mask under conditions where we can use the `is_causal` argument in
`torch.sdpa` instead. Default to `True`.
allow_torch_fix (`bool`, optional):
Whether to update the mask in case a query is not attending to any tokens, to solve a bug in torch's older
versions. We need an arg to skip it when using eager. By default `True`.
## Creating a simple causal mask:
To create the following causal mask:
0 ■ ⬚ ⬚ ⬚ ⬚
1 ■ ■ ⬚ ⬚ ⬚
2 ■ ■ ■ ⬚ ⬚
3 ■ ■ ■ ■ ⬚
4 ■ ■ ■ ■ ■
You can do
```python
>>> sdpa_mask(batch_size=1, cache_position=torch.arange(5), kv_length=5)
>>> tensor([[[[ True, False, False, False, False],
[ True, True, False, False, False],
[ True, True, True, False, False],
[ True, True, True, True, False],
[ True, True, True, True, True]]]])
```
## Creating a sliding window mask:
To create the following sliding window mask (`sliding_window=3`):
0 ■ ⬚ ⬚ ⬚ ⬚
1 ■ ■ ⬚ ⬚ ⬚
2 ■ ■ ■ ⬚ ⬚
3 ⬚ ■ ■ ■ ⬚
4 ⬚ ⬚ ■ ■ ■
You can do
```python
>>> sdpa_mask(batch_size=1, cache_position=torch.arange(5), kv_length=5, mask_function=sliding_window_causal_mask_function(3))
>>> tensor([[[[ True, False, False, False, False],
[ True, True, False, False, False],
[ True, True, True, False, False],
[False, True, True, True, False],
[False, False, True, True, True]]]])
```
## Creating a chunked attention mask
To create the following chunked attention mask (`chunk_size=3`):
0 ■ ⬚ ⬚ ⬚ ⬚
1 ■ ■ ⬚ ⬚ ⬚
2 ■ ■ ■ ⬚ ⬚
3 ⬚ ⬚ ⬚ ■ ⬚
4 ⬚ ⬚ ⬚ ■ ■
You can do
```python
>>> sdpa_mask(batch_size=1, cache_position=torch.arange(5), kv_length=5, mask_function=chunked_causal_mask_function(3, torch.zeros(1, dtype=int)))
>>> tensor([[[[ True, False, False, False, False],
[ True, True, False, False, False],
[ True, True, True, False, False],
[False, False, False, True, False],
[False, False, False, True, True]]]])
```
"""
q_length = cache_position.shape[0]
# Potentially pad the 2D mask, and slice it correctly
padding_mask = prepare_padding_mask(attention_mask, kv_length, kv_offset, _slice=False)
# Under specific conditions, we can avoid materializing the mask, instead relying on the `is_causal` argument
if allow_is_causal_skip and _ignore_causal_mask_sdpa(padding_mask, q_length, kv_length, kv_offset, local_size):
return None
# Similar to `kv_arange = torch.arange(start=kv_offset, end=kv_offset + kv_length, device=cache_position.device)`
# but without data-dependent slicing (i.e. torch.compile friendly)
kv_arange = torch.arange(kv_length, device=cache_position.device)
kv_arange += kv_offset
# Potentially add the padding 2D mask
if padding_mask is not None:
mask_function = and_masks(mask_function, padding_mask_function(padding_mask))
batch_arange = torch.arange(batch_size, device=cache_position.device)
head_arange = torch.arange(1, device=cache_position.device)
# This creates the 4D mask easily. Note that we need this context manager as vmap cannot handle slicing a tensor from
# scalar tensor (it internally calls `.item()` which vmap does not allow, but this context works around it
# We don't need to add an offset to the mask_function either, as we vmap directly the correct indices for k and kv indices
with TransformGetItemToIndex():
causal_mask = _vmap_for_bhqkv(mask_function)(batch_arange, head_arange, cache_position, kv_arange)
return causal_mask
def sdpa_mask_older_torch(
batch_size: int,
cache_position: torch.Tensor,
kv_length: int,
kv_offset: int = 0,
mask_function: Callable = causal_mask_function,
attention_mask: Optional[torch.Tensor] = None,
local_size: Optional[int] = None,
allow_is_causal_skip: bool = True,
allow_torch_fix: bool = True,
**kwargs,
) -> Optional[torch.Tensor]:
"""
NOTE: This function is only used when torch version is torch<2.5 - see `sdpa_mask_recent_torch` otherwise.
Create a 4D boolean mask of shape `(batch_size, 1, query_length, kv_length)` where a value of True indicates that
the element should take part in the attention computation, and False that it should not.
If `allow_torch_fix=True` (the default), rows corresponding to query tokens that do not attend
to any other tokens (due to padding) will be fully attended to instead, in order to avoid `nan` propagation (this does
not change the final result).
Args:
batch_size (`int`):
The batch size of the input sequence.
cache_position (`torch.Tensor`):
A tensor of shape (query_length,) indicating the current indices of the input sequence elements.
kv_length (`int`):
The size that the key and value states will have during the attention computation.
kv_offset (`int`, optional):
An optional offset to indicate at which first position the key and values states will refer to.
mask_function (`Callable`):
The mask factory function describing the mask pattern.
attention_mask (`torch.Tensor`, optional):
The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length)
local_size (`int`, optional):
The size of the local attention, if we do not use full attention. This is used only if `allow_is_causal_skip=True`
to try to skip mask creation if possible.
allow_is_causal_skip (`bool`, optional):
Whether to allow to return `None` for the mask under conditions where we can use the `is_causal` argument in
`torch.sdpa` instead. Default to `True`.
allow_torch_fix (`bool`, optional):
Whether to update the mask in case a query is not attending to any tokens, to solve a bug in torch's older
versions. We need an arg to skip it when using eager. By default `True`.
"""
q_length = cache_position.shape[0]
# Potentially pad the 2D mask, and slice it correctly
padding_mask = prepare_padding_mask(attention_mask, kv_length, kv_offset)
# Under specific conditions, we can avoid materializing the mask, instead relying on the `is_causal` argument
if allow_is_causal_skip and _ignore_causal_mask_sdpa(padding_mask, q_length, kv_length, kv_offset, local_size):
return None
# Similar to `kv_arange = torch.arange(start=kv_offset, end=kv_offset + kv_length, device=cache_position.device)`
# but without data-dependent slicing (i.e. torch.compile friendly)
kv_arange = torch.arange(kv_length, device=cache_position.device)
kv_arange += kv_offset
# This creates the 4D mask easily. Note that we do not include vmap over the batch_idx dimension as well,
# as vmap cannot handle slicing a tensor from scalar tensor (it internally calls `.item()` which vmap does not allow
# However, in more recent version of Pytorch, a trick was introduced to handle it - which is the reason we have
# `sdpa_mask_recent_torch`, as it allows more general `mask_function`
causal_mask = _vmap_for_bhqkv(mask_function, bh_indices=False)(None, None, cache_position, kv_arange)
causal_mask = causal_mask[None, None, :, :].expand(batch_size, -1, -1, -1)
if padding_mask is not None:
causal_mask = causal_mask * padding_mask[:, None, None, :]
# Due to a bug in versions of torch<2.5, we need to update the mask in case a query is not attending to any
# tokens (due to padding). See details in https://github.com/pytorch/pytorch/issues/110213
if not _is_torch_greater_or_equal_than_2_5 and allow_torch_fix:
causal_mask |= torch.all(~causal_mask, dim=-1, keepdim=True)
return causal_mask
# We use the version with newer torch whenever possible, as it is more general and can handle arbitrary mask functions
# (especially mask_function indexing a tensor, such as the padding mask function)
sdpa_mask = sdpa_mask_recent_torch if _is_torch_greater_or_equal_than_2_6 else sdpa_mask_older_torch
def eager_mask(
batch_size: int,
cache_position: torch.Tensor,
kv_length: int,
kv_offset: int = 0,
mask_function: Callable = causal_mask_function,
attention_mask: Optional[torch.Tensor] = None,
dtype: torch.dtype = torch.float32,
**kwargs,
) -> torch.Tensor:
"""
Create a 4D float mask of shape `(batch_size, 1, query_length, kv_length)` where a value of 0 indicates that
the element should take part in the attention computation, and -inf (minimum value for the given `dtype`) that
it should not.
Args:
batch_size (`int`):
The batch size of the input sequence.
cache_position (`torch.Tensor`):
A tensor of shape (query_length,) indicating the current indices of the input sequence elements.
kv_length (`int`):
The size that the key and value states will have during the attention computation.
kv_offset (`int`, optional):
An optional offset to indicate at which first position the key and values states will refer to.
mask_function (`Callable`):
The mask factory function describing the mask pattern.
attention_mask (`torch.Tensor`, optional):
The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length)
dtype (`torch.dtype`, optional):
The dtype to use for the mask. By default, `torch.float32`.
"""
# The masks for eager attention are simply boolean mask from sdpa, casted to 0 and -inf
_ = kwargs.pop("allow_is_causal_skip", None)
mask = sdpa_mask(
batch_size=batch_size,
cache_position=cache_position,
kv_length=kv_length,
kv_offset=kv_offset,
mask_function=mask_function,
attention_mask=attention_mask,
allow_is_causal_skip=False,
allow_torch_fix=False,
**kwargs,
)
min_dtype = torch.finfo(dtype).min
# we need 0s where the tokens should be taken into account, and -inf otherwise (mask is already of boolean type)
mask = torch.where(mask, torch.tensor(0.0, device=mask.device, dtype=dtype), min_dtype)
return mask
def flash_attention_mask(
batch_size: int,
cache_position: torch.Tensor,
kv_length: int,
kv_offset: int = 0,
mask_function: Callable = causal_mask_function,
attention_mask: Optional[torch.Tensor] = None,
**kwargs,
):
"""
Create the attention mask necessary to use FA2. Since FA2 is un-padded by definition, here we simply return
`None` if the mask is fully causal, or we return the 2D mask which will then be used to extract the seq_lens.
We just slice it in case of sliding window.
Args:
batch_size (`int`):
The batch size of the input sequence.
cache_position (`torch.Tensor`):
A tensor of shape (query_length,) indicating the current indices of the input sequence elements.
kv_length (`int`):
The size that the key and value states will have during the attention computation.
kv_offset (`int`, optional):
An optional offset to indicate at which first position the key and values states will refer to.
mask_function (`Callable`):
The mask factory function describing the mask pattern.
attention_mask (`torch.Tensor`, optional):
The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length)
"""
if attention_mask is not None:
# Here we need to slice from the right if using sliding or chunked (for full attention, this is equivalent to doing nothing)
attention_mask = attention_mask[:, -kv_length:]
# We only return an actual mask if there is at least 1 padding token, otherwise we return `None` and use `is_causal` in FA2
# (note that the attention_mask is a boolean dtype here)
if attention_mask.all():
attention_mask = None
return attention_mask
def flex_attention_mask(
batch_size: int,
cache_position: torch.Tensor,
kv_length: int,
kv_offset: int = 0,
mask_function: Callable = causal_mask_function,
attention_mask: Optional[torch.Tensor] = None,
**kwargs,
) -> BlockMask:
"""
Create a 4D block mask which is a compressed representation of the full 4D block causal mask. BlockMask is essential
for performant computation of flex attention. See: https://pytorch.org/blog/flexattention/
Args:
batch_size (`int`):
The batch size of the input sequence.
cache_position (`torch.Tensor`):
A tensor of shape (query_length,) indicating the current indices of the input sequence elements.
kv_length (`int`):
The size that the key and value states will have during the attention computation.
kv_offset (`int`, optional):
An optional offset to indicate at which first position the key and values states will refer to.
mask_function (`Callable`):
The mask factory function describing the mask pattern.
attention_mask (`torch.Tensor`, optional):
The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length)
"""
q_length, q_offset = cache_position.shape[0], cache_position[0]
# Potentially add the padding 2D mask
if attention_mask is not None:
# Older torch (2.5.x) cannot handle sequences not in multiples of 128 (default block size)
# Hence we pad to multiples of this as a minimum to ensure this
pad_len = ((attention_mask.shape[1] // flex_default_block_size) + 1) * flex_default_block_size
pad_len = pad_len - attention_mask.shape[1]
if not _is_torch_greater_or_equal_than_2_6 and pad_len > 0:
attention_mask = torch.nn.functional.pad(attention_mask, value=0, pad=(0, pad_len))
padding_mask = prepare_padding_mask(attention_mask, kv_length, kv_offset, _slice=False)
mask_function = and_masks(mask_function, padding_mask_function(padding_mask))
# Add the offsets on top (because flex interface only allows length, not start and end indices)
mask_function = add_offsets_to_mask_function(mask_function, q_offset, kv_offset)
# Finally create the block mask
block_mask = create_block_mask(
mask_mod=mask_function,
B=batch_size,
H=None,
Q_LEN=q_length,
KV_LEN=kv_length,
device=cache_position.device,
_compile=_is_torch_greater_or_equal_than_2_6,
)
return block_mask
class AttentionMaskInterface(GeneralInterface):
# Class instance object, so that a call to `register` can be reflected into all other files correctly, even if
# a new instance is created (in order to locally override a given function)
_global_mapping = {
"sdpa": sdpa_mask,
"eager": eager_mask,
"flash_attention_2": flash_attention_mask,
"flex_attention": flex_attention_mask,
}
# Global AttentionMaskInterface shared by all models which do not need to overwrite any of the existing ones
ALL_MASK_ATTENTION_FUNCTIONS: AttentionMaskInterface = AttentionMaskInterface()
def find_packed_sequence_indices(position_ids: torch.Tensor) -> torch.Tensor:
"""
Find the indices of the sequence to which each new query token in the sequence belongs when using packed
tensor format (i.e. several sequences packed in the same batch dimension).
Args:
position_ids (`torch.Tensor`)
A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences.
Returns:
A 2D tensor where each similar integer indicates that the tokens belong to the same sequence. For example, if we
pack 3 sequences of 2, 3 and 1 tokens respectively along a single batch dim, this will return [[0, 0, 1, 1, 1, 2]].
"""
# What separate different sequences is when 2 consecutive positions_ids are separated by more than 1. So
# taking the diff (by prepending the first value - 1 to keep correct indexing) and applying cumsum to the result
# gives exactly the sequence indices
# Note that we assume that a single sequence cannot span several batch dimensions, i.e. 1 single sequence
# cannot be part of the end of the first batch dim and the start of the 2nd one for example
first_dummy_value = position_ids[:, :1] - 1 # We just need the diff on this first value to be 1
position_diff = torch.diff(position_ids, prepend=first_dummy_value, dim=-1)
packed_sequence_mask = (position_diff != 1).cumsum(-1)
# Here it would be nice to return None if we did not detect packed sequence format, i.e. if `packed_sequence_mask[:, -1] == 0`
# but it causes issues with export
return packed_sequence_mask
def _preprocess_mask_arguments(
config: PretrainedConfig,
input_embeds: torch.Tensor,
attention_mask: Optional[Union[torch.Tensor, BlockMask]],
cache_position: torch.Tensor,
past_key_values: Optional[Cache],
position_ids: Optional[torch.Tensor],
layer_idx: Optional[int],
) -> tuple[bool, Optional[Union[torch.Tensor, BlockMask]], int, int]:
"""
Perform some common pre-processing of the mask arguments we get from the modeling code. Mostly determine the
key-value length and offsets, and if we should early exit or not.
Args:
config (`PretrainedConfig`):
The model config.
input_embeds (`torch.Tensor`):
The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the
batch size, query length and dtype.
attention_mask (`torch.Tensor`, optional):
The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length).
It can also be an already prepared 4D mask, in which case it is returned as-is.
cache_position (`torch.Tensor`):
A tensor of shape (query_length,) indicating the current indices of the input sequence elements.
past_key_values (`Cache`, optional):
The past key values, if we use a cache.
position_ids (`torch.Tensor`, optional)
A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences.
layer_idx (`int`, optional):
If `past_key_values` is not None, this is the layer index of the cache from which to get the key-value
length and offset. Indeed, for hybrid caches, different layers may return different lengths.
Returns:
early_exit (`bool`):
Whether we should early exit mask creation, and return the mask as-is.
attention_mask (`torch.Tensor` or `BlockMask` or `None`):
The attention mask to either return immediately, or to use in downstream mask creation.
packed_sequence_mask (`torch.Tensor`, optional):
In case we detected packed sequence format, this is a tensor where each similar integer indicates that
the tokens belong to the same sequence.
kv_length (`int`):
The size that the key and value states will have during the attention computation.
kv_offset (`int`):
An offset to indicate at which first position the key and values states will refer to.
"""
# If the mask is already 4D, simply return as-is (it was already prepared, or it is custom)
if isinstance(attention_mask, (torch.Tensor, BlockMask)) and len(attention_mask.shape) == 4:
return True, attention_mask, None, None, None
# For TGI/vLLM backends, or other custom attention without equivalent mask creation: we don't need a mask!
# Note: it's not ideal to check the `_global_mapping` attribute instead of the object itself, however otherwise
# full graph dynamo tracing (i.e. torch.export or compile with `fullgraph=True`) will fail on Python<3.11
# with `torch._dynamo.exc.Unsupported: 'inline in skipfiles:Mapping.__contains__ | __contains__, skipped
# according trace_rules.lookup SKIP_DIRS'` -- can be removed when we require Python>=3.11
if config._attn_implementation not in ALL_MASK_ATTENTION_FUNCTIONS._global_mapping:
return True, None, None, None, None
# Move the mask to correct device, and potentially switch dtype for efficiency
if attention_mask is not None and attention_mask.ndim == 2:
attention_mask = attention_mask.to(device=cache_position.device, dtype=torch.bool)
# If using a cache, it can give all information about mask sizes based on seen tokens
if past_key_values is not None:
kv_length, kv_offset = past_key_values.get_mask_sizes(cache_position, layer_idx)
# Otherwise, the sizes are simply the input sizes
else:
kv_length, kv_offset = input_embeds.shape[1], 0
# We check the position_ids for potential packed sequence format (only if the 2D attention mask is explicitly None,
# and we don't have past_key_values, i.e. generally a training setup)
packed_sequence_mask = None
if position_ids is not None and attention_mask is None and past_key_values is None:
batch_size = input_embeds.shape[0]
# The position ids are sometimes just unsqueezed, without being expanded
if batch_size != position_ids.shape[0]:
position_ids = position_ids.expand(batch_size, -1)
packed_sequence_mask = find_packed_sequence_indices(position_ids)
return False, attention_mask, packed_sequence_mask, kv_length, kv_offset
def create_causal_mask(
config: PretrainedConfig,
input_embeds: torch.Tensor,
attention_mask: Optional[torch.Tensor],
cache_position: torch.Tensor,
past_key_values: Optional[Cache],
position_ids: Optional[torch.Tensor] = None,
or_mask_function: Optional[Callable] = None,
and_mask_function: Optional[Callable] = None,
) -> Optional[Union[torch.Tensor, BlockMask]]:
"""
Create a standard causal mask based on the attention implementation used (stored in the config). If `past_key_values`
has an hybrid cache structure, this function will return the mask corresponding to one of the "full_attention" layers (to align
to what is needed in the `modeling_xxx.py` files).
Args:
config (`PretrainedConfig`):
The model config.
input_embeds (`torch.Tensor`):
The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the
batch size, query length and dtype.
attention_mask (`torch.Tensor`, optional):
The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length).
It can also be an already prepared 4D mask, in which case it is returned as-is.
cache_position (`torch.Tensor`):
A tensor of shape (query_length,) indicating the current indices of the input sequence elements.
past_key_values (`Cache`, optional):
The past key values, if we use a cache.
position_ids (`torch.Tensor`, optional)
A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences.
or_mask_function (`Callable`, optional):
An optional mask function to combine with the causal mask function (by doing the union of both). This is
useful to easily overlay another mask on top of the causal one, for example for image tokens handling.
and_mask_function (`Callable`, optional):
An optional mask function to combine with the causal mask function (by doing the intersection of both). This is
useful to easily overlay another mask on top of the causal one, for example for image tokens handling.
"""
# If we have an hybrid cache structure, here we want to create the mask for the full layers
if hasattr(past_key_values, "is_sliding") and False in past_key_values.is_sliding:
layer_idx = past_key_values.is_sliding.index(False)
else:
layer_idx = 0
early_exit, attention_mask, packed_sequence_mask, kv_length, kv_offset = _preprocess_mask_arguments(
config, input_embeds, attention_mask, cache_position, past_key_values, position_ids, layer_idx
)
if early_exit:
return attention_mask
batch_size, dtype = input_embeds.shape[0], input_embeds.dtype
mask_factory_function = causal_mask_function
mask_interface = ALL_MASK_ATTENTION_FUNCTIONS[config._attn_implementation]
# Do not allow skip if we are compiling (this is to match BC)
# TODO: cyril -> probably revisit and remove this, but a lot of tests rely on it
if _is_torch_xpu_available:
allow_is_causal_skip = True
else:
allow_is_causal_skip = not past_key_values.is_compileable if past_key_values is not None else True
# If we detected packing format
if packed_sequence_mask is not None and _is_torch_greater_or_equal_than_2_6:
mask_factory_function = and_masks(mask_factory_function, packed_sequence_mask_function(packed_sequence_mask))
allow_is_causal_skip = False
# Allow slight deviations from causal mask
if or_mask_function is not None:
if not _is_torch_greater_or_equal_than_2_6:
raise ValueError("Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6")
mask_factory_function = or_masks(mask_factory_function, or_mask_function)
allow_is_causal_skip = False
if and_mask_function is not None:
if not _is_torch_greater_or_equal_than_2_6:
raise ValueError("Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6")
mask_factory_function = and_masks(mask_factory_function, and_mask_function)
allow_is_causal_skip = False
# We now create the mask
causal_mask = mask_interface(
batch_size=batch_size,
cache_position=cache_position,
kv_length=kv_length,
kv_offset=kv_offset,
mask_function=mask_factory_function,
attention_mask=attention_mask,
allow_is_causal_skip=allow_is_causal_skip, # additional kwarg for sdpa
dtype=dtype, # Additional kwarg for eager
config=config, # Pass the config as well, in case someone wants to easily have their own mask_interface
)
return causal_mask
def create_sliding_window_causal_mask(
config: PretrainedConfig,
input_embeds: torch.Tensor,
attention_mask: Optional[torch.Tensor],
cache_position: torch.Tensor,
past_key_values: Optional[Cache],
position_ids: Optional[torch.Tensor] = None,
or_mask_function: Optional[Callable] = None,
and_mask_function: Optional[Callable] = None,
) -> Optional[Union[torch.Tensor, BlockMask]]:
"""
Create a sliding window causal mask based on the attention implementation used (stored in the config). This type
of attention pattern was mostly democratized by Mistral. If `past_key_values` has an hybrid cache structure, this
function will return the mask corresponding to one of the "sliding_attention" layers (to align to what is needed in the
`modeling_xxx.py` files).
Args:
config (`PretrainedConfig`):
The model config.
input_embeds (`torch.Tensor`):
The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the
batch size, query length and dtype.
attention_mask (`torch.Tensor`, optional):
The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length).
It can also be an already prepared 4D mask, in which case it is returned as-is.
cache_position (`torch.Tensor`):
A tensor of shape (query_length,) indicating the current indices of the input sequence elements.
past_key_values (`Cache`, optional):
The past key values, if we use a cache.
position_ids (`torch.Tensor`, optional)
A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences.
or_mask_function (`Callable`, optional):
An optional mask function to combine with the sliding causal mask function (by doing the union of both). This is
useful to easily overlay another mask on top of the sliding causal one, for example for image tokens handling.
and_mask_function (`Callable`, optional):
An optional mask function to combine with the sliding causal mask function (by doing the intersection of both). This is
useful to easily overlay another mask on top of the sliding causal one, for example for image tokens handling.
"""
# If we have an hybrid cache structure, here we want to create the mask for the sliding layers
if hasattr(past_key_values, "is_sliding") and True in past_key_values.is_sliding:
layer_idx = past_key_values.is_sliding.index(True)
else:
layer_idx = 0
early_exit, attention_mask, packed_sequence_mask, kv_length, kv_offset = _preprocess_mask_arguments(
config, input_embeds, attention_mask, cache_position, past_key_values, position_ids, layer_idx
)
if early_exit:
return attention_mask
sliding_window = getattr(config, "sliding_window", None)
if sliding_window is None:
raise ValueError("Could not find a `sliding_window` argument in the config, or it is not set")
batch_size, dtype = input_embeds.shape[0], input_embeds.dtype
mask_factory_function = sliding_window_causal_mask_function(sliding_window)
mask_interface = ALL_MASK_ATTENTION_FUNCTIONS[config._attn_implementation]
# Do not allow skip if we are compiling (this is to match BC)
# TODO: cyril -> probably revisit and remove this, but a lot of tests rely on it
allow_is_causal_skip = not past_key_values.is_compileable if past_key_values is not None else True
# If we detected packing format
if packed_sequence_mask is not None and _is_torch_greater_or_equal_than_2_6:
mask_factory_function = and_masks(mask_factory_function, packed_sequence_mask_function(packed_sequence_mask))
allow_is_causal_skip = False
# Allow slight deviations from sliding causal mask
if or_mask_function is not None:
if not _is_torch_greater_or_equal_than_2_6:
raise ValueError("Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6")
mask_factory_function = or_masks(mask_factory_function, or_mask_function)
allow_is_causal_skip = False
if and_mask_function is not None:
if not _is_torch_greater_or_equal_than_2_6:
raise ValueError("Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6")
mask_factory_function = and_masks(mask_factory_function, and_mask_function)
allow_is_causal_skip = False
# We now create the mask
causal_mask = mask_interface(
batch_size=batch_size,
cache_position=cache_position,
kv_length=kv_length,
kv_offset=kv_offset,
mask_function=mask_factory_function,
attention_mask=attention_mask,
allow_is_causal_skip=allow_is_causal_skip, # additional kwarg for sdpa
local_size=sliding_window, # Additional kwarg for sdpa
dtype=dtype, # Additional kwarg for eager
config=config, # Pass the config as well, in case someone wants to easily have their own mask_interface
)
return causal_mask
def create_chunked_causal_mask(
config: PretrainedConfig,
input_embeds: torch.Tensor,
attention_mask: Optional[torch.Tensor],
cache_position: torch.Tensor,
past_key_values: Optional[Cache],
position_ids: Optional[torch.Tensor] = None,
or_mask_function: Optional[Callable] = None,
and_mask_function: Optional[Callable] = None,
) -> Optional[Union[torch.Tensor, BlockMask]]:
"""
Create a chunked attention causal mask based on the attention implementation used (stored in the config). This type
of attention pattern was mostly democratized by Llama4. If `past_key_values` has an hybrid cache structure, this
function will return the mask corresponding to one of the "chunked_attention" layers (to align to what is needed in the
`modeling_xxx.py` files).
Args:
config (`PretrainedConfig`):
The model config.
input_embeds (`torch.Tensor`):
The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the
batch size, query length and dtype.
attention_mask (`torch.Tensor`, optional):
The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length).
It can also be an already prepared 4D mask, in which case it is returned as-is.
cache_position (`torch.Tensor`):
A tensor of shape (query_length,) indicating the current indices of the input sequence elements.
past_key_values (`Cache`, optional):
The past key values, if we use a cache.
position_ids (`torch.Tensor`, optional)
A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences.
or_mask_function (`Callable`, optional):
An optional mask function to combine with the chunked causal mask function (by doing the union of both). This is
useful to easily overlay another mask on top of the chunked causal one, for example for image tokens handling.
and_mask_function (`Callable`, optional):
An optional mask function to combine with the chunked causal mask function (by doing the intersection of both). This is
useful to easily overlay another mask on top of the chunked causal one, for example for image tokens handling.
"""
# If we have an hybrid cache structure, here we want to create the mask for the sliding layers
if hasattr(past_key_values, "is_sliding") and True in past_key_values.is_sliding:
layer_idx = past_key_values.is_sliding.index(True)
else:
layer_idx = 0
early_exit, attention_mask, packed_sequence_mask, kv_length, kv_offset = _preprocess_mask_arguments(
config, input_embeds, attention_mask, cache_position, past_key_values, position_ids, layer_idx
)
if early_exit:
return attention_mask
chunk_size = getattr(config, "attention_chunk_size", None)
if chunk_size is None:
raise ValueError("Could not find an `attention_chunk_size` argument in the config, or it is not set")
# Raise if using chunked attention on context too large with FA2
if config._attn_implementation == "flash_attention_2" and kv_length + kv_offset > chunk_size:
raise ValueError(
"Flash attention 2 cannot handle chunked attention, and the key-value length is larger than the chunk size so the "
"chunked pattern cannot be respected. You should use another `attn_implementation` when instantiating the model"
)
batch_size, dtype = input_embeds.shape[0], input_embeds.dtype
# For chunked attention and batched inputs, we need to take the number of left padding tokens into account
# to start the chunk from the actual start of the sequence for the padded sequence
if attention_mask is not None:
# Only count the left padding tokens, not all of them
left_padding_tokens = (attention_mask.cumsum(dim=-1) == torch.zeros_like(attention_mask)).sum(dim=-1)
else:
left_padding_tokens = torch.zeros(batch_size, device=cache_position.device, dtype=int)
# Raise a warning for older versions if the problematic left-padding situation arises
if (
not _is_torch_greater_or_equal_than_2_6
and kv_length + kv_offset > chunk_size
and (left_padding_tokens > 0).any()
):
logger.warning_once(
"Due to limitations of your current torch version, we cannot correctly account for the left-padding "
"when computing the chunked attention pattern. This will lead to a wrong attention mask for the padded "
"sequences. Behavior will be undefined. Please upgrade to `torch>=2.6` to solve this issue."
)
mask_factory_function = chunked_causal_mask_function(chunk_size, left_padding_tokens)
mask_interface = ALL_MASK_ATTENTION_FUNCTIONS[config._attn_implementation]
# Do not allow skip if we are compiling (this is to match BC)
# TODO: cyril -> probably revisit and remove this, but a lot of tests rely on it
allow_is_causal_skip = not past_key_values.is_compileable if past_key_values is not None else True
# If we detected packing format
if packed_sequence_mask is not None and _is_torch_greater_or_equal_than_2_6:
mask_factory_function = and_masks(mask_factory_function, packed_sequence_mask_function(packed_sequence_mask))
allow_is_causal_skip = False
# Allow slight deviations from chunked causal mask
if or_mask_function is not None:
if not _is_torch_greater_or_equal_than_2_6:
raise ValueError("Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6")
mask_factory_function = or_masks(mask_factory_function, or_mask_function)
allow_is_causal_skip = False
if and_mask_function is not None:
if not _is_torch_greater_or_equal_than_2_6:
raise ValueError("Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6")
mask_factory_function = and_masks(mask_factory_function, and_mask_function)
allow_is_causal_skip = False
# We now create the mask
causal_mask = mask_interface(
batch_size=batch_size,
cache_position=cache_position,
kv_length=kv_length,
kv_offset=kv_offset,
mask_function=mask_factory_function,
attention_mask=attention_mask,
allow_is_causal_skip=allow_is_causal_skip, # additional kwarg for sdpa
local_size=chunk_size, # Additional kwarg for sdpa
dtype=dtype, # Additional kwarg for eager
config=config, # Pass the config as well, in case someone wants to easily have their own mask_interface
)
return causal_mask
LAYER_PATTERN_TO_MASK_FUNCTION_MAPPING = {
"full_attention": create_causal_mask,
"sliding_attention": create_sliding_window_causal_mask,
"chunked_attention": create_chunked_causal_mask,
}
def create_masks_for_generate(
config: PretrainedConfig,
input_embeds: torch.Tensor,
attention_mask: Optional[torch.Tensor],
cache_position: torch.Tensor,
past_key_values: Optional[Cache],
position_ids: Optional[torch.Tensor] = None,
or_mask_function: Optional[Callable] = None,
and_mask_function: Optional[Callable] = None,
**kwargs,
):
"""
This function mimics how we create the masks in the `modeling_xxx.py` files, and is used in `generate` in order
to easily create the masks in advance, when we compile the forwards with Static caches.
Args:
config (`PretrainedConfig`):
The model config.
input_embeds (`torch.Tensor`):
The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the
batch size, query length and dtype.
attention_mask (`torch.Tensor`, optional):
The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length).
It can also be an already prepared 4D mask, in which case it is returned as-is.
cache_position (`torch.Tensor`):
A tensor of shape (query_length,) indicating the current indices of the input sequence elements.
past_key_values (`Cache`, optional):
The past key values, if we use a cache.
position_ids (`torch.Tensor`, optional)
A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences.
or_mask_function (`Callable`, optional):
An optional mask function to combine with the other mask function (by doing the union of both). This is
useful to easily overlay another mask on top of the causal one, for example for image tokens handling.
and_mask_function (`Callable`, optional):
An optional mask function to combine with the other mask function (by doing the intersection of both). This is
useful to easily overlay another mask on top of the causal one, for example for image tokens handling.
"""
# The attribute reside in the text config for composite models
effective_config = config.get_text_config()
# Prepare the mask args
mask_kwargs = {
"config": effective_config,
"input_embeds": input_embeds,
"attention_mask": attention_mask,
"cache_position": cache_position,
"past_key_values": past_key_values,
"position_ids": position_ids,
"or_mask_function": or_mask_function,
"and_mask_function": and_mask_function,
}
# If the attribute exist, we need several masks
if hasattr(effective_config, "layer_types"):
causal_masks = {}
for layer_pattern in set(effective_config.layer_types):
causal_masks[layer_pattern] = LAYER_PATTERN_TO_MASK_FUNCTION_MAPPING[layer_pattern](**mask_kwargs)
return causal_masks
# In this case, all layers are sliding
elif getattr(effective_config, "sliding_window", None) is not None:
return create_sliding_window_causal_mask(**mask_kwargs)
# In this case, all layers are chunked
elif getattr(effective_config, "attention_chunk_size", None) is not None:
return create_chunked_causal_mask(**mask_kwargs)
# All layers use standard causal attention
return create_causal_mask(**mask_kwargs)
# Below are utilities to pretty-print the different masks
# Print the matrix with words as row labels
GREEN = "\033[92m"
YELLOW = "\033[93m"
RESET = "\033[0m"
BLACK_SQUARE = "■"
WHITE_SQUARE = "⬚"
GREY_SQUARE = "∙"
LOW_TRIANGLE = "⬕"
UPPER_TRIANGLE = "⬔"
def get_style(style):
if style == "majong":
BLACK_SQUARE = "🀞" # Full block (represents "on" or active)
BLACK_SQUARE = "🀙" # Full block (represents "on" or active)
WHITE_SQUARE = "🀆" # "▒" # Light shade (represents "off" or inactive)
LOW_TRIANGLE = "🀛" # Lower left triangle (stylized indication)
UPPER_TRIANGLE = "🀛" # Upper left triangle (stylized indication)
else:
BLACK_SQUARE = "█" # Full block (represents "on" or active)
WHITE_SQUARE = "░" # "▒" # Light shade (represents "off" or inactive)
LOW_TRIANGLE = "▙" # Lower left triangle (stylized indication))
UPPER_TRIANGLE = "▜" # Upper left triangle (stylized indication)
return BLACK_SQUARE, WHITE_SQUARE, LOW_TRIANGLE, UPPER_TRIANGLE
# LOW_TRIANGLE = UPPER_TRIANGLE = "⟍" # Upper right triangle (stylized indication)
YELLOW_SQUARE = f"{YELLOW}{BLACK_SQUARE}{RESET}"
GREEN_SQUARE = f"{GREEN}{BLACK_SQUARE}{RESET}"
def tensor_to_mask_visual(original_tensor: torch.Tensor, grid_size=(20, 40), style="majong") -> str:
BLACK_SQUARE, WHITE_SQUARE, LOW_TRIANGLE, UPPER_TRIANGLE = get_style(style)
h, w = original_tensor.shape
max_h, max_w = grid_size
if not (h < max_h and w < max_w):
# Preserve aspect ratio within max grid size
aspect_ratio = 2 * w / h
if aspect_ratio > 1:
w = max_w
h = min(max_h, max(1, round(max_w / aspect_ratio)))
else:
h = max_h
w = max(1, round(max_h * aspect_ratio))
# Step 1: Rescale tensor by average pooling
tensor = original_tensor.unsqueeze(0).unsqueeze(0) # Add batch and channel dimensions
tensor = F.adaptive_avg_pool2d(tensor, output_size=(h, w))[0, 0] # Remove extra dims
else:
tensor = original_tensor
# Step 3: Build the string representation
result = []
for i in range(h):
row = ""
for j in range(w):
if tensor[i, j] == 1:
row += BLACK_SQUARE
elif tensor[i, j] == 0:
row += WHITE_SQUARE
else:
if j > 0:
if tensor[i, j - 1] == 1:
row += LOW_TRIANGLE
elif tensor[i, j - 1] == 0:
row += UPPER_TRIANGLE
else:
row += BLACK_SQUARE if tensor[i, j] == 1 else WHITE_SQUARE
else:
row += (
BLACK_SQUARE
if tensor[i, j] == 1
else (
WHITE_SQUARE
if tensor[i, j] == 0
else (UPPER_TRIANGLE if tensor[i, j + 1] == 1 else LOW_TRIANGLE)
)
)
result.append(row)
return "\n".join(result)
class AttentionMask(torch.Tensor):
def __new__(cls, data, style=None):
# Create a new instance of AttentionMask as a Tensor
cls.style = style
return torch.Tensor._make_subclass(cls, data, require_grad=False)
def __init__(self, data):
# You can initialize any additional metadata here if needed
pass
def to_string(self, grid_size=(20, 40), limit=4):
"""Returns a string representation of the block mask."""
dense_mask = self
*batch_dims, num_rows, num_cols = dense_mask.shape
total_vis = []
for idx, batch_idx in enumerate(itertools.product(*[range(i) for i in batch_dims])):
if idx == limit:
total_vis.append("...")
total_vis.append("To print out more, set AttentionMask.to_string(limit=N)")
total_vis.append("You can also index (AttentionMask[batch, head]) to choose a specific batch or head")
break
block_vis = tensor_to_mask_visual(dense_mask[batch_idx], grid_size=grid_size, style=self.style)
total_vis.append(block_vis)
total_vis.append(f"torch.Tensor(shape={tuple(self.shape)}, dtype={self.dtype})")
return "\n".join(total_vis)
def __repr__(self):
return self.to_string()
def __str__(self):
return self.to_string()
@classmethod
def from_tensor(cls, tensor: torch.Tensor, style: Optional[str] = None) -> "AttentionMask":
res = cls(tensor)
res.style = style
return res
| transformers/src/transformers/masking_utils.py/0 | {
"file_path": "transformers/src/transformers/masking_utils.py",
"repo_id": "transformers",
"token_count": 22696
} | 426 |
# Copyright 2020 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import TYPE_CHECKING
from ..utils import _LazyModule
from ..utils.import_utils import define_import_structure
if TYPE_CHECKING:
from .aimv2 import *
from .albert import *
from .align import *
from .altclip import *
from .arcee import *
from .aria import *
from .audio_spectrogram_transformer import *
from .auto import *
from .autoformer import *
from .aya_vision import *
from .bamba import *
from .bark import *
from .bart import *
from .barthez import *
from .bartpho import *
from .beit import *
from .bert import *
from .bert_generation import *
from .bert_japanese import *
from .bertweet import *
from .big_bird import *
from .bigbird_pegasus import *
from .biogpt import *
from .bit import *
from .bitnet import *
from .blenderbot import *
from .blenderbot_small import *
from .blip import *
from .blip_2 import *
from .bloom import *
from .bridgetower import *
from .bros import *
from .byt5 import *
from .camembert import *
from .canine import *
from .chameleon import *
from .chinese_clip import *
from .clap import *
from .clip import *
from .clipseg import *
from .clvp import *
from .code_llama import *
from .codegen import *
from .cohere import *
from .cohere2 import *
from .cohere2_vision import *
from .colpali import *
from .colqwen2 import *
from .conditional_detr import *
from .convbert import *
from .convnext import *
from .convnextv2 import *
from .cpm import *
from .cpmant import *
from .csm import *
from .ctrl import *
from .cvt import *
from .d_fine import *
from .dab_detr import *
from .dac import *
from .data2vec import *
from .dbrx import *
from .deberta import *
from .deberta_v2 import *
from .decision_transformer import *
from .deepseek_v2 import *
from .deepseek_v3 import *
from .deepseek_vl import *
from .deepseek_vl_hybrid import *
from .deformable_detr import *
from .deit import *
from .deprecated import *
from .depth_anything import *
from .depth_pro import *
from .detr import *
from .dia import *
from .dialogpt import *
from .diffllama import *
from .dinat import *
from .dinov2 import *
from .dinov2_with_registers import *
from .dinov3_convnext import *
from .dinov3_vit import *
from .distilbert import *
from .dit import *
from .donut import *
from .dots1 import *
from .dpr import *
from .dpt import *
from .efficientloftr import *
from .efficientnet import *
from .electra import *
from .emu3 import *
from .encodec import *
from .encoder_decoder import *
from .ernie import *
from .esm import *
from .evolla import *
from .exaone4 import *
from .falcon import *
from .falcon_h1 import *
from .falcon_mamba import *
from .fastspeech2_conformer import *
from .flaubert import *
from .flava import *
from .florence2 import *
from .fnet import *
from .focalnet import *
from .fsmt import *
from .funnel import *
from .fuyu import *
from .gemma import *
from .gemma2 import *
from .gemma3 import *
from .gemma3n import *
from .git import *
from .glm import *
from .glm4 import *
from .glpn import *
from .got_ocr2 import *
from .gpt2 import *
from .gpt_bigcode import *
from .gpt_neo import *
from .gpt_neox import *
from .gpt_neox_japanese import *
from .gpt_oss import *
from .gpt_sw3 import *
from .gptj import *
from .granite import *
from .granite_speech import *
from .granitemoe import *
from .granitemoehybrid import *
from .granitemoeshared import *
from .grounding_dino import *
from .groupvit import *
from .helium import *
from .herbert import *
from .hgnet_v2 import *
from .hiera import *
from .hubert import *
from .hunyuan_v1_dense import *
from .hunyuan_v1_moe import *
from .ibert import *
from .idefics import *
from .idefics2 import *
from .idefics3 import *
from .ijepa import *
from .imagegpt import *
from .informer import *
from .instructblip import *
from .instructblipvideo import *
from .internvl import *
from .jamba import *
from .janus import *
from .jetmoe import *
from .kosmos2 import *
from .kyutai_speech_to_text import *
from .layoutlm import *
from .layoutlmv2 import *
from .layoutlmv3 import *
from .layoutxlm import *
from .led import *
from .levit import *
from .lfm2 import *
from .lightglue import *
from .lilt import *
from .llama import *
from .llama4 import *
from .llava import *
from .llava_next import *
from .llava_next_video import *
from .llava_onevision import *
from .longformer import *
from .longt5 import *
from .luke import *
from .lxmert import *
from .m2m_100 import *
from .mamba import *
from .mamba2 import *
from .marian import *
from .markuplm import *
from .mask2former import *
from .maskformer import *
from .mbart import *
from .mbart50 import *
from .megatron_bert import *
from .megatron_gpt2 import *
from .mgp_str import *
from .mimi import *
from .minimax import *
from .mistral import *
from .mistral3 import *
from .mixtral import *
from .mlcd import *
from .mllama import *
from .mluke import *
from .mobilebert import *
from .mobilenet_v1 import *
from .mobilenet_v2 import *
from .mobilevit import *
from .mobilevitv2 import *
from .modernbert import *
from .modernbert_decoder import *
from .moonshine import *
from .moshi import *
from .mpnet import *
from .mpt import *
from .mra import *
from .mt5 import *
from .musicgen import *
from .musicgen_melody import *
from .mvp import *
from .myt5 import *
from .nemotron import *
from .nllb import *
from .nllb_moe import *
from .nougat import *
from .nystromformer import *
from .olmo import *
from .olmo2 import *
from .olmoe import *
from .omdet_turbo import *
from .oneformer import *
from .openai import *
from .opt import *
from .ovis2 import *
from .owlv2 import *
from .owlvit import *
from .paligemma import *
from .patchtsmixer import *
from .patchtst import *
from .pegasus import *
from .pegasus_x import *
from .perceiver import *
from .perception_lm import *
from .persimmon import *
from .phi import *
from .phi3 import *
from .phi4_multimodal import *
from .phimoe import *
from .phobert import *
from .pix2struct import *
from .pixtral import *
from .plbart import *
from .poolformer import *
from .pop2piano import *
from .prompt_depth_anything import *
from .prophetnet import *
from .pvt import *
from .pvt_v2 import *
from .qwen2 import *
from .qwen2_5_omni import *
from .qwen2_5_vl import *
from .qwen2_audio import *
from .qwen2_moe import *
from .qwen2_vl import *
from .qwen3 import *
from .qwen3_moe import *
from .rag import *
from .recurrent_gemma import *
from .reformer import *
from .regnet import *
from .rembert import *
from .resnet import *
from .roberta import *
from .roberta_prelayernorm import *
from .roc_bert import *
from .roformer import *
from .rt_detr import *
from .rt_detr_v2 import *
from .rwkv import *
from .sam import *
from .sam2 import *
from .sam2_video import *
from .sam_hq import *
from .seamless_m4t import *
from .seamless_m4t_v2 import *
from .seed_oss import *
from .segformer import *
from .seggpt import *
from .sew import *
from .sew_d import *
from .shieldgemma2 import *
from .siglip import *
from .siglip2 import *
from .smolvlm import *
from .speech_encoder_decoder import *
from .speech_to_text import *
from .speecht5 import *
from .splinter import *
from .squeezebert import *
from .stablelm import *
from .starcoder2 import *
from .superglue import *
from .superpoint import *
from .swiftformer import *
from .swin import *
from .swin2sr import *
from .swinv2 import *
from .switch_transformers import *
from .t5 import *
from .t5gemma import *
from .table_transformer import *
from .tapas import *
from .textnet import *
from .time_series_transformer import *
from .timesfm import *
from .timesformer import *
from .timm_backbone import *
from .timm_wrapper import *
from .trocr import *
from .tvp import *
from .udop import *
from .umt5 import *
from .unispeech import *
from .unispeech_sat import *
from .univnet import *
from .upernet import *
from .video_llava import *
from .videomae import *
from .vilt import *
from .vipllava import *
from .vision_encoder_decoder import *
from .vision_text_dual_encoder import *
from .visual_bert import *
from .vit import *
from .vit_mae import *
from .vit_msn import *
from .vitdet import *
from .vitmatte import *
from .vitpose import *
from .vitpose_backbone import *
from .vits import *
from .vivit import *
from .vjepa2 import *
from .voxtral import *
from .wav2vec2 import *
from .wav2vec2_bert import *
from .wav2vec2_conformer import *
from .wav2vec2_phoneme import *
from .wav2vec2_with_lm import *
from .wavlm import *
from .whisper import *
from .x_clip import *
from .xcodec import *
from .xglm import *
from .xlm import *
from .xlm_roberta import *
from .xlm_roberta_xl import *
from .xlnet import *
from .xlstm import *
from .xmod import *
from .yolos import *
from .yoso import *
from .zamba import *
from .zamba2 import *
from .zoedepth import *
else:
import sys
_file = globals()["__file__"]
sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)
| transformers/src/transformers/models/__init__.py/0 | {
"file_path": "transformers/src/transformers/models/__init__.py",
"repo_id": "transformers",
"token_count": 4344
} | 427 |
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Convert ALIGN checkpoints from the original repository."""
import argparse
import os
import align
import numpy as np
import requests
import tensorflow as tf
import torch
from PIL import Image
from tokenizer import Tokenizer
from transformers import (
AlignConfig,
AlignModel,
AlignProcessor,
BertConfig,
BertTokenizer,
EfficientNetConfig,
EfficientNetImageProcessor,
)
from transformers.utils import logging
logging.set_verbosity_info()
logger = logging.get_logger(__name__)
def preprocess(image):
image = tf.image.resize(image, (346, 346))
image = tf.image.crop_to_bounding_box(image, (346 - 289) // 2, (346 - 289) // 2, 289, 289)
return image
def get_align_config():
vision_config = EfficientNetConfig.from_pretrained("google/efficientnet-b7")
vision_config.image_size = 289
vision_config.hidden_dim = 640
vision_config.id2label = {"0": "LABEL_0", "1": "LABEL_1"}
vision_config.label2id = {"LABEL_0": 0, "LABEL_1": 1}
vision_config.depthwise_padding = []
text_config = BertConfig()
config = AlignConfig.from_text_vision_configs(
text_config=text_config, vision_config=vision_config, projection_dim=640
)
return config
# We will verify our results on an image of cute cats
def prepare_img():
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
im = Image.open(requests.get(url, stream=True).raw)
return im
def get_processor():
image_processor = EfficientNetImageProcessor(
do_center_crop=True,
rescale_factor=1 / 127.5,
rescale_offset=True,
do_normalize=False,
include_top=False,
resample=Image.BILINEAR,
)
tokenizer = BertTokenizer.from_pretrained("google-bert/bert-base-uncased")
tokenizer.model_max_length = 64
processor = AlignProcessor(image_processor=image_processor, tokenizer=tokenizer)
return processor
# here we list all keys to be renamed (original name on the left, our name on the right)
def rename_keys(original_param_names):
# EfficientNet image encoder
block_names = [v.split("_")[0].split("block")[1] for v in original_param_names if v.startswith("block")]
block_names = list(set(block_names))
block_names = sorted(block_names)
num_blocks = len(block_names)
block_name_mapping = {b: str(i) for b, i in zip(block_names, range(num_blocks))}
rename_keys = []
rename_keys.append(("stem_conv/kernel:0", "embeddings.convolution.weight"))
rename_keys.append(("stem_bn/gamma:0", "embeddings.batchnorm.weight"))
rename_keys.append(("stem_bn/beta:0", "embeddings.batchnorm.bias"))
rename_keys.append(("stem_bn/moving_mean:0", "embeddings.batchnorm.running_mean"))
rename_keys.append(("stem_bn/moving_variance:0", "embeddings.batchnorm.running_var"))
for b in block_names:
hf_b = block_name_mapping[b]
rename_keys.append((f"block{b}_expand_conv/kernel:0", f"encoder.blocks.{hf_b}.expansion.expand_conv.weight"))
rename_keys.append((f"block{b}_expand_bn/gamma:0", f"encoder.blocks.{hf_b}.expansion.expand_bn.weight"))
rename_keys.append((f"block{b}_expand_bn/beta:0", f"encoder.blocks.{hf_b}.expansion.expand_bn.bias"))
rename_keys.append(
(f"block{b}_expand_bn/moving_mean:0", f"encoder.blocks.{hf_b}.expansion.expand_bn.running_mean")
)
rename_keys.append(
(f"block{b}_expand_bn/moving_variance:0", f"encoder.blocks.{hf_b}.expansion.expand_bn.running_var")
)
rename_keys.append(
(f"block{b}_dwconv/depthwise_kernel:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_conv.weight")
)
rename_keys.append((f"block{b}_bn/gamma:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_norm.weight"))
rename_keys.append((f"block{b}_bn/beta:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_norm.bias"))
rename_keys.append(
(f"block{b}_bn/moving_mean:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_norm.running_mean")
)
rename_keys.append(
(f"block{b}_bn/moving_variance:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_norm.running_var")
)
rename_keys.append((f"block{b}_se_reduce/kernel:0", f"encoder.blocks.{hf_b}.squeeze_excite.reduce.weight"))
rename_keys.append((f"block{b}_se_reduce/bias:0", f"encoder.blocks.{hf_b}.squeeze_excite.reduce.bias"))
rename_keys.append((f"block{b}_se_expand/kernel:0", f"encoder.blocks.{hf_b}.squeeze_excite.expand.weight"))
rename_keys.append((f"block{b}_se_expand/bias:0", f"encoder.blocks.{hf_b}.squeeze_excite.expand.bias"))
rename_keys.append(
(f"block{b}_project_conv/kernel:0", f"encoder.blocks.{hf_b}.projection.project_conv.weight")
)
rename_keys.append((f"block{b}_project_bn/gamma:0", f"encoder.blocks.{hf_b}.projection.project_bn.weight"))
rename_keys.append((f"block{b}_project_bn/beta:0", f"encoder.blocks.{hf_b}.projection.project_bn.bias"))
rename_keys.append(
(f"block{b}_project_bn/moving_mean:0", f"encoder.blocks.{hf_b}.projection.project_bn.running_mean")
)
rename_keys.append(
(f"block{b}_project_bn/moving_variance:0", f"encoder.blocks.{hf_b}.projection.project_bn.running_var")
)
key_mapping = {}
for item in rename_keys:
if item[0] in original_param_names:
key_mapping[item[0]] = "vision_model." + item[1]
# BERT text encoder
rename_keys = []
old = "tf_bert_model/bert"
new = "text_model"
for i in range(12):
rename_keys.append(
(
f"{old}/encoder/layer_._{i}/attention/self/query/kernel:0",
f"{new}.encoder.layer.{i}.attention.self.query.weight",
)
)
rename_keys.append(
(
f"{old}/encoder/layer_._{i}/attention/self/query/bias:0",
f"{new}.encoder.layer.{i}.attention.self.query.bias",
)
)
rename_keys.append(
(
f"{old}/encoder/layer_._{i}/attention/self/key/kernel:0",
f"{new}.encoder.layer.{i}.attention.self.key.weight",
)
)
rename_keys.append(
(
f"{old}/encoder/layer_._{i}/attention/self/key/bias:0",
f"{new}.encoder.layer.{i}.attention.self.key.bias",
)
)
rename_keys.append(
(
f"{old}/encoder/layer_._{i}/attention/self/value/kernel:0",
f"{new}.encoder.layer.{i}.attention.self.value.weight",
)
)
rename_keys.append(
(
f"{old}/encoder/layer_._{i}/attention/self/value/bias:0",
f"{new}.encoder.layer.{i}.attention.self.value.bias",
)
)
rename_keys.append(
(
f"{old}/encoder/layer_._{i}/attention/output/dense/kernel:0",
f"{new}.encoder.layer.{i}.attention.output.dense.weight",
)
)
rename_keys.append(
(
f"{old}/encoder/layer_._{i}/attention/output/dense/bias:0",
f"{new}.encoder.layer.{i}.attention.output.dense.bias",
)
)
rename_keys.append(
(
f"{old}/encoder/layer_._{i}/attention/output/LayerNorm/gamma:0",
f"{new}.encoder.layer.{i}.attention.output.LayerNorm.weight",
)
)
rename_keys.append(
(
f"{old}/encoder/layer_._{i}/attention/output/LayerNorm/beta:0",
f"{new}.encoder.layer.{i}.attention.output.LayerNorm.bias",
)
)
rename_keys.append(
(
f"{old}/encoder/layer_._{i}/intermediate/dense/kernel:0",
f"{new}.encoder.layer.{i}.intermediate.dense.weight",
)
)
rename_keys.append(
(
f"{old}/encoder/layer_._{i}/intermediate/dense/bias:0",
f"{new}.encoder.layer.{i}.intermediate.dense.bias",
)
)
rename_keys.append(
(f"{old}/encoder/layer_._{i}/output/dense/kernel:0", f"{new}.encoder.layer.{i}.output.dense.weight")
)
rename_keys.append(
(f"{old}/encoder/layer_._{i}/output/dense/bias:0", f"{new}.encoder.layer.{i}.output.dense.bias")
)
rename_keys.append(
(f"{old}/encoder/layer_._{i}/output/LayerNorm/gamma:0", f"{new}.encoder.layer.{i}.output.LayerNorm.weight")
)
rename_keys.append(
(f"{old}/encoder/layer_._{i}/output/LayerNorm/beta:0", f"{new}.encoder.layer.{i}.output.LayerNorm.bias")
)
rename_keys.append((f"{old}/embeddings/word_embeddings/weight:0", f"{new}.embeddings.word_embeddings.weight"))
rename_keys.append(
(f"{old}/embeddings/position_embeddings/embeddings:0", f"{new}.embeddings.position_embeddings.weight")
)
rename_keys.append(
(f"{old}/embeddings/token_type_embeddings/embeddings:0", f"{new}.embeddings.token_type_embeddings.weight")
)
rename_keys.append((f"{old}/embeddings/LayerNorm/gamma:0", f"{new}.embeddings.LayerNorm.weight"))
rename_keys.append((f"{old}/embeddings/LayerNorm/beta:0", f"{new}.embeddings.LayerNorm.bias"))
rename_keys.append((f"{old}/pooler/dense/kernel:0", f"{new}.pooler.dense.weight"))
rename_keys.append((f"{old}/pooler/dense/bias:0", f"{new}.pooler.dense.bias"))
rename_keys.append(("dense/kernel:0", "text_projection.weight"))
rename_keys.append(("dense/bias:0", "text_projection.bias"))
rename_keys.append(("dense/bias:0", "text_projection.bias"))
rename_keys.append(("temperature:0", "temperature"))
for item in rename_keys:
if item[0] in original_param_names:
key_mapping[item[0]] = item[1]
return key_mapping
def replace_params(hf_params, tf_params, key_mapping):
list(hf_params.keys())
for key, value in tf_params.items():
if key not in key_mapping:
continue
hf_key = key_mapping[key]
if "_conv" in key and "kernel" in key:
new_hf_value = torch.from_numpy(value).permute(3, 2, 0, 1)
elif "embeddings" in key:
new_hf_value = torch.from_numpy(value)
elif "depthwise_kernel" in key:
new_hf_value = torch.from_numpy(value).permute(2, 3, 0, 1)
elif "kernel" in key:
new_hf_value = torch.from_numpy(np.transpose(value))
elif "temperature" in key:
new_hf_value = value
elif "bn/gamma" in key or "bn/beta" in key:
new_hf_value = torch.from_numpy(np.transpose(value)).squeeze()
else:
new_hf_value = torch.from_numpy(value)
# Replace HF parameters with original TF model parameters
hf_params[hf_key].copy_(new_hf_value)
@torch.no_grad()
def convert_align_checkpoint(checkpoint_path, pytorch_dump_folder_path, save_model, push_to_hub):
"""
Copy/paste/tweak model's weights to our ALIGN structure.
"""
# Load original model
seq_length = 64
tok = Tokenizer(seq_length)
original_model = align.Align("efficientnet-b7", "bert-base", 640, seq_length, tok.get_vocab_size())
original_model.compile()
original_model.load_weights(checkpoint_path)
tf_params = original_model.trainable_variables
tf_non_train_params = original_model.non_trainable_variables
tf_params = {param.name: param.numpy() for param in tf_params}
for param in tf_non_train_params:
tf_params[param.name] = param.numpy()
tf_param_names = list(tf_params.keys())
# Load HuggingFace model
config = get_align_config()
hf_model = AlignModel(config).eval()
hf_params = hf_model.state_dict()
# Create src-to-dst parameter name mapping dictionary
print("Converting parameters...")
key_mapping = rename_keys(tf_param_names)
replace_params(hf_params, tf_params, key_mapping)
# Initialize processor
processor = get_processor()
inputs = processor(
images=prepare_img(), text="A picture of a cat", padding="max_length", max_length=64, return_tensors="pt"
)
# HF model inference
hf_model.eval()
with torch.no_grad():
outputs = hf_model(**inputs)
hf_image_features = outputs.image_embeds.detach().numpy()
hf_text_features = outputs.text_embeds.detach().numpy()
# Original model inference
original_model.trainable = False
tf_image_processor = EfficientNetImageProcessor(
do_center_crop=True,
do_rescale=False,
do_normalize=False,
include_top=False,
resample=Image.BILINEAR,
)
image = tf_image_processor(images=prepare_img(), return_tensors="tf", data_format="channels_last")["pixel_values"]
text = tok(tf.constant(["A picture of a cat"]))
image_features = original_model.image_encoder(image, training=False)
text_features = original_model.text_encoder(text, training=False)
image_features = tf.nn.l2_normalize(image_features, axis=-1)
text_features = tf.nn.l2_normalize(text_features, axis=-1)
# Check whether original and HF model outputs match -> np.allclose
if not np.allclose(image_features, hf_image_features, atol=1e-3):
raise ValueError("The predicted image features are not the same.")
if not np.allclose(text_features, hf_text_features, atol=1e-3):
raise ValueError("The predicted text features are not the same.")
print("Model outputs match!")
if save_model:
# Create folder to save model
if not os.path.isdir(pytorch_dump_folder_path):
os.mkdir(pytorch_dump_folder_path)
# Save converted model and image processor
hf_model.save_pretrained(pytorch_dump_folder_path)
processor.save_pretrained(pytorch_dump_folder_path)
if push_to_hub:
# Push model and image processor to hub
print("Pushing converted ALIGN to the hub...")
processor.push_to_hub("align-base")
hf_model.push_to_hub("align-base")
if __name__ == "__main__":
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--checkpoint_path",
default="./weights/model-weights",
type=str,
help="Path to the pretrained TF ALIGN checkpoint.",
)
parser.add_argument(
"--pytorch_dump_folder_path",
default="hf_model",
type=str,
help="Path to the output PyTorch model directory.",
)
parser.add_argument("--save_model", action="store_true", help="Save model to local")
parser.add_argument("--push_to_hub", action="store_true", help="Push model and image processor to the hub")
args = parser.parse_args()
convert_align_checkpoint(args.checkpoint_path, args.pytorch_dump_folder_path, args.save_model, args.push_to_hub)
| transformers/src/transformers/models/align/convert_align_tf_to_hf.py/0 | {
"file_path": "transformers/src/transformers/models/align/convert_align_tf_to_hf.py",
"repo_id": "transformers",
"token_count": 7044
} | 428 |
# coding=utf-8
# Copyright 2024 The Rhymes-AI Teams Authors and The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from collections.abc import Iterable
from typing import Optional, Union
import numpy as np
from ...activations import ACT2FN
from ...cache_utils import Cache
from ...configuration_utils import PretrainedConfig
from ...image_processing_utils import BaseImageProcessor, BatchFeature, get_patch_output_size, select_best_resolution
from ...image_transforms import PaddingMode, convert_to_rgb, pad, resize, to_channel_dimension_format
from ...image_utils import (
ChannelDimension,
ImageInput,
PILImageResampling,
get_image_size,
infer_channel_dimension_format,
is_scaled_image,
make_flat_list_of_images,
to_numpy_array,
valid_images,
validate_preprocess_arguments,
)
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_utils import PreTrainedModel
from ...processing_utils import MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack
from ...tokenization_utils import PreTokenizedInput, TextInput
from ...utils import TensorType, TransformersKwargs, auto_docstring, can_return_tuple, logging
from ...utils.import_utils import is_torch_available
from ..auto import CONFIG_MAPPING, AutoConfig, AutoTokenizer
from ..llama.configuration_llama import LlamaConfig
from ..llama.modeling_llama import (
LlamaAttention,
LlamaDecoderLayer,
LlamaForCausalLM,
LlamaMLP,
LlamaModel,
LlamaPreTrainedModel,
LlamaRMSNorm,
)
from ..llava.modeling_llava import (
LlavaCausalLMOutputWithPast,
LlavaForConditionalGeneration,
LlavaModel,
LlavaModelOutputWithPast,
)
from ..llava_next.image_processing_llava_next import divide_to_patches
logger = logging.get_logger(__name__)
if is_torch_available():
import torch
from torch import nn
def sequential_experts_gemm(token_states, expert_weights, tokens_per_expert):
"""
Compute the matrix multiplication (GEMM) for each expert sequentially. This approach is computationally inefficient, especially when dealing with a large number of experts.
Args:
token_states (torch.Tensor): Input tensor of shape (num_tokens, in_features).
expert_weights (torch.Tensor): Weight tensor of shape (num_experts, in_features, out_features).
tokens_per_expert (torch.Tensor): Number of tokens assigned to each expert.
Returns:
torch.Tensor: Output tensor of shape (num_tokens, out_features).
"""
num_tokens = token_states.shape[0]
out_features = expert_weights.shape[-1]
output = torch.zeros(num_tokens, out_features, dtype=token_states.dtype, device=token_states.device)
cumsum_num_tokens = torch.cumsum(tokens_per_expert, dim=0)
# Insert zero at the beginning for offset index's convenience
zero_tensor = torch.zeros(1, dtype=torch.long, device=cumsum_num_tokens.device)
cumsum_num_tokens = torch.cat((zero_tensor, cumsum_num_tokens))
for expert_num in range(expert_weights.shape[0]):
start = cumsum_num_tokens[expert_num]
end = cumsum_num_tokens[expert_num + 1]
tokens = token_states[start:end]
out = torch.matmul(tokens, expert_weights[expert_num])
output[start:end] = out
return output
class AriaTextConfig(LlamaConfig):
r"""
This class handles the configuration for the text component of the Aria model.
Instantiating a configuration with the defaults will yield a similar configuration to that of the model of the Aria
[rhymes-ai/Aria](https://huggingface.co/rhymes-ai/Aria) architecture.
This class extends the LlamaConfig to include additional parameters specific to the Mixture of Experts (MoE) architecture.
Args:
vocab_size (`int`, *optional*, defaults to 32000):
Vocabulary size of the LLaMA model. Defines the number of different tokens that can be represented by the
`inputs_ids` passed when calling [`LlamaModel`]
hidden_size (`int`, *optional*, defaults to 4096):
Dimension of the hidden representations.
intermediate_size (`int`, *optional*, defaults to 4096):
The size of the MLP representations.
num_hidden_layers (`int`, *optional*, defaults to 32):
Number of hidden layers in the Transformer decoder.
num_attention_heads (`int`, *optional*, defaults to 32):
Number of attention heads for each attention layer in the Transformer decoder.
num_key_value_heads (`int`, *optional*):
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
`num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
by meanpooling all the original heads within that group. For more details, check out [this
paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to
`num_attention_heads`.
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
The non-linear activation function (function or string) in the decoder.
max_position_embeddings (`int`, *optional*, defaults to 2048):
The maximum sequence length that this model might ever be used with. Llama 1 supports up to 2048 tokens,
Llama 2 up to 4096, CodeLlama up to 16384.
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
rms_norm_eps (`float`, *optional*, defaults to 1e-06):
The epsilon used by the rms normalization layers.
use_cache (`bool`, *optional*, defaults to `True`):
Whether or not the model should return the last key/values attentions (not used by all models). Only
relevant if `config.is_decoder=True`.
pad_token_id (`int`, *optional*, defaults to 2):
Padding token id.
bos_token_id (`int`, *optional*, defaults to 1):
Beginning of stream token id.
eos_token_id (`int`, *optional*, defaults to 2):
End of stream token id.
pretraining_tp (`int`, *optional*, defaults to 1):
Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
document](https://huggingface.co/docs/transformers/main/perf_train_gpu_many#tensor-parallelism) to
understand more about it. This value is necessary to ensure exact reproducibility of the pretraining
results. Please refer to [this issue](https://github.com/pytorch/pytorch/issues/76232).
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
Whether to tie weight embeddings
rope_theta (`float`, *optional*, defaults to 10000.0):
The base period of the RoPE embeddings.
rope_scaling (`Dict`, *optional*):
Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
accordingly.
Expected contents:
`rope_type` (`str`):
The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
'llama3'], with 'default' being the original RoPE implementation.
`factor` (`float`, *optional*):
Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
most scaling types, a `factor` of x will enable the model to handle sequences of length x *
original maximum pre-trained length.
`original_max_position_embeddings` (`int`, *optional*):
Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
pretraining.
`attention_factor` (`float`, *optional*):
Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
computation. If unspecified, it defaults to value recommended by the implementation, using the
`factor` field to infer the suggested value.
`beta_fast` (`float`, *optional*):
Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
ramp function. If unspecified, it defaults to 32.
`beta_slow` (`float`, *optional*):
Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
ramp function. If unspecified, it defaults to 1.
`short_factor` (`list[float]`, *optional*):
Only used with 'longrope'. The scaling factor to be applied to short contexts (<
`original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
size divided by the number of attention heads divided by 2
`long_factor` (`list[float]`, *optional*):
Only used with 'longrope'. The scaling factor to be applied to long contexts (<
`original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
size divided by the number of attention heads divided by 2
`low_freq_factor` (`float`, *optional*):
Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
`high_freq_factor` (`float`, *optional*):
Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
attention_bias (`bool`, *optional*, defaults to `False`):
Whether to use a bias in the query, key, value and output projection layers during self-attention.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
mlp_bias (`bool`, *optional*, defaults to `False`):
Whether to use a bias in up_proj, down_proj and gate_proj layers in the MLP layers.
head_dim (`int`, *optional*):
The attention head dimension. If None, it will default to hidden_size // num_heads
moe_num_experts (`int`, *optional*, defaults to 8):
The number of experts in the MoE layer.
moe_topk (`int`, *optional*, defaults to 2):
The number of top experts to route to for each token.
moe_num_shared_experts (`int`, *optional*, defaults to 2):
The number of shared experts.
"""
model_type = "aria_text"
base_config_key = "text_config"
def __init__(
self,
intermediate_size: int = 4096,
moe_num_experts: int = 8,
moe_topk: int = 2,
moe_num_shared_experts: int = 2,
pad_token_id=2,
**super_kwargs,
):
super().__init__(pad_token_id=pad_token_id, **super_kwargs)
self.intermediate_size = intermediate_size
self.moe_num_experts = moe_num_experts
self.moe_topk = moe_topk
self.moe_num_shared_experts = moe_num_shared_experts
class AriaConfig(PretrainedConfig):
r"""
This class handles the configuration for both vision and text components of the Aria model,
as well as additional parameters for image token handling and projector mapping.
Instantiating a configuration with the defaults will yield a similar configuration to that of the model of the Aria
[rhymes-ai/Aria](https://huggingface.co/rhymes-ai/Aria) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
vision_config (`AriaVisionConfig` or `dict`, *optional*):
Configuration for the vision component.
vision_feature_layer (`int`, *optional*, defaults to -1):
The index of the layer to select the vision feature.
text_config (`AriaTextConfig` or `dict`, *optional*):
Configuration for the text component.
projector_patch_to_query_dict (`dict`, *optional*):
Mapping of patch sizes to query dimensions.
image_token_index (`int`, *optional*, defaults to 9):
Index used to represent image tokens.
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated normal initializer for initializing all weight matrices.
Attributes:
model_type (`str`):
Type of the model, set to `"aria"`.
image_token_index (`int`):
Index used to represent image tokens.
projector_patch_to_query_dict (`dict`):
Mapping of patch sizes to query dimensions.
vision_config (`AriaVisionConfig`):
Configuration for the vision component.
text_config (`AriaTextConfig`):
Configuration for the text component.
"""
model_type = "aria"
attribute_map = {
"image_token_id": "image_token_index",
}
sub_configs = {"text_config": AriaTextConfig, "vision_config": AutoConfig}
def __init__(
self,
vision_config=None,
vision_feature_layer: int = -1,
text_config: AriaTextConfig = None,
projector_patch_to_query_dict: Optional[dict] = None,
image_token_index: int = 9,
initializer_range: float = 0.02,
**kwargs,
):
self.image_token_index = image_token_index
# Convert the keys and values of projector_patch_to_query_dict to integers
# This ensures consistency even if they were provided as strings
if projector_patch_to_query_dict is None:
projector_patch_to_query_dict = {
1225: 128,
4900: 256,
}
self.projector_patch_to_query_dict = {int(k): int(v) for k, v in projector_patch_to_query_dict.items()}
self.max_value_projector_patch_to_query_dict = max(self.projector_patch_to_query_dict.values())
self.vision_feature_layer = vision_feature_layer
if isinstance(vision_config, dict):
vision_config["model_type"] = "idefics3_vision"
vision_config = CONFIG_MAPPING[vision_config["model_type"]](**vision_config)
elif vision_config is None:
vision_config = CONFIG_MAPPING["idefics3_vision"]()
self.vision_config = vision_config
self.initializer_range = initializer_range
if isinstance(text_config, dict) and "model_type" in text_config:
text_config = AriaTextConfig(**text_config)
elif text_config is None:
text_config = AriaTextConfig()
self.text_config = text_config
super().__init__(**kwargs)
class AriaTextRMSNorm(LlamaRMSNorm):
pass
class AriaProjectorMLP(nn.Module):
"""
Feed-Forward Network module for the Aria Projector.
Args:
in_features (`int`):
Input embedding dimension.
hidden_features (`int`):
Hidden dimension of the feed-forward network.
output_dim (`int`):
Output dimension.
"""
def __init__(self, in_features, hidden_features, output_dim):
super().__init__()
self.linear_in = nn.Linear(in_features, hidden_features, bias=False)
self.linear_out = nn.Linear(hidden_features, output_dim, bias=False)
self.act = ACT2FN["gelu_new"]
def forward(self, hidden_states):
hidden_states = self.act(self.linear_in(hidden_states))
hidden_states = self.linear_out(hidden_states)
return hidden_states
class AriaCrossAttention(nn.Module):
"""
Aria Cross-Attention module.
Args:
config (`AriaConfig`):
The configuration to use.
"""
def __init__(self, config: AriaConfig, dropout_rate: float = 0):
super().__init__()
hidden_size = config.vision_config.hidden_size
num_heads = config.vision_config.num_attention_heads
self.num_heads = num_heads
self.q_proj = nn.Linear(hidden_size, hidden_size, bias=False)
self.k_proj = nn.Linear(hidden_size, hidden_size, bias=False)
self.v_proj = nn.Linear(hidden_size, hidden_size, bias=False)
# Original code here: https://github.com/rhymes-ai/Aria/blob/719ff4e52b727443cba3793b0e27fe64e0244fe1/aria/model/projector.py#L48
self.multihead_attn = nn.MultiheadAttention(hidden_size, num_heads, batch_first=True)
self.linear = nn.Linear(hidden_size, hidden_size)
self.dropout = nn.Dropout(dropout_rate)
self.layer_norm = nn.LayerNorm(hidden_size)
self.layer_norm_kv = nn.LayerNorm(hidden_size)
def forward(self, key_value_states, hidden_states, attn_mask=None):
"""
Forward pass of the AriaCrossAttention module.
Args:
key_value_states (`torch.Tensor`):
Input tensor for key and value.
hidden_states (`torch.Tensor`):
Input tensor for query.
attn_mask (`torch.Tensor`, *optional*, defaults to None):
Attention mask.
Returns:
torch.Tensor:
Output tensor after cross-attention.
"""
query = self.q_proj(self.layer_norm(hidden_states))
key_value_states = self.layer_norm_kv(key_value_states)
key = self.k_proj(key_value_states)
value = self.v_proj(key_value_states)
attn_output, _ = self.multihead_attn(query, key, value, attn_mask=attn_mask)
attn_output = self.dropout(self.linear(attn_output))
return attn_output
class AriaProjector(nn.Module):
"""
Aria Projector module.
This module projects vision features into the language model's embedding space, enabling interaction between vision and language components.
Args:
config (`AriaConfig`):
Configuration object for the model.
"""
def __init__(
self,
config: AriaConfig,
):
super().__init__()
self.patch_to_query_dict = config.projector_patch_to_query_dict
self.in_features = config.vision_config.hidden_size
self.num_heads = config.vision_config.num_attention_heads
self.kv_dim = config.vision_config.hidden_size
self.hidden_features = config.text_config.hidden_size
self.output_dim = config.text_config.hidden_size
self.query = nn.Parameter(torch.zeros(config.max_value_projector_patch_to_query_dict, self.in_features))
self.cross_attn = AriaCrossAttention(config)
self.layer_norm = nn.LayerNorm(self.in_features)
self.feed_forward = AriaProjectorMLP(self.in_features, self.hidden_features, self.output_dim)
def forward(self, key_value_states: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
"""
Forward pass of the Projector module.
Args:
key_value_states (`torch.Tensor`):
Input tensor of shape (batch_size, num_patches, kv_dim).
attn_mask (`torch.Tensor`, *optional*, default is None):
Attention mask.
Returns:
`torch.Tensor`: Output tensor of shape (batch_size, query_number, output_dim).
"""
batch_size, num_patches = key_value_states.shape[0], key_value_states.shape[1]
if num_patches not in self.patch_to_query_dict:
raise KeyError(
f"Number of patches {num_patches} not found in patch_to_query_dict amongst possible values {self.patch_to_query_dict.keys()}."
)
query_num = self.patch_to_query_dict[num_patches]
queries = self.query[:query_num].unsqueeze(0).repeat(batch_size, 1, 1)
if attn_mask is not None:
attn_mask = attn_mask.repeat_interleave(self.num_heads, 0)
attn_mask = attn_mask.unsqueeze(1).expand(-1, queries.size(1), -1)
attention_out = self.cross_attn(key_value_states, queries, attn_mask=attn_mask)
out = self.feed_forward(self.layer_norm(attention_out))
return out
class AriaImageProcessor(BaseImageProcessor):
"""
A vision processor for the Aria model that handles image preprocessing.
Initialize the AriaImageProcessor.
Args:
image_mean (`list`, *optional*, defaults to [0.5, 0.5, 0.5]):
Mean values for normalization.
image_std (`list`, *optional*, defaults to [0.5, 0.5, 0.5]):
Standard deviation values for normalization.
max_image_size (`int`, *optional*, defaults to 980):
Maximum image size.
min_image_size (`int`, *optional*, defaults to 336):
Minimum image size.
split_resolutions (`list`, *optional*, defaults to a list of optimal,resolutions as tuples):
The optimal resolutions for splitting the image.
split_image (`bool`, *optional*, defaults to `False`):
Whether to split the image.
do_convert_rgb (`bool`, *optional*, defaults to `True`):
Whether to convert the image to RGB.
do_rescale (`bool`, *optional*, defaults to `True`):
Whether to rescale the image by the specified scale `rescale_factor`. Can be overridden by `do_rescale` in
the `preprocess` method.
rescale_factor (`int` or `float`, *optional*, defaults to `1/255`):
Scale factor to use if rescaling the image. Can be overridden by `rescale_factor` in the `preprocess`
method.
do_normalize (`bool`, *optional*, defaults to `True`):
Whether to normalize the image.
resample (PILImageResampling, *optional*, defaults to `BICUBIC`):
The resampling filter to use if resizing the image.
"""
model_input_names = ["pixel_values", "pixel_mask", "num_crops"]
def __init__(
self,
image_mean: Optional[list[float]] = None,
image_std: Optional[list[float]] = None,
max_image_size: int = 980,
min_image_size: int = 336,
split_resolutions: Optional[list[tuple[int, int]]] = None,
split_image: Optional[bool] = False,
do_convert_rgb: Optional[bool] = True,
do_rescale: bool = True,
rescale_factor: Union[int, float] = 1 / 255,
do_normalize: Optional[bool] = True,
resample: PILImageResampling = PILImageResampling.BICUBIC,
**kwargs,
):
super().__init__(**kwargs)
if image_mean is None:
image_mean = [0.5, 0.5, 0.5]
if image_std is None:
image_std = [0.5, 0.5, 0.5]
self.max_image_size = max_image_size
self.min_image_size = min_image_size
self.image_mean = image_mean
self.image_std = image_std
self.split_image = split_image
if split_resolutions is None:
split_resolutions = [(1, 2), (1, 3), (1, 4), (1, 5), (1, 6), (1, 7), (1, 8), (2, 4), (2, 3), (2, 2), (2, 1), (3, 1), (3, 2), (4, 1), (4, 2), (5, 1), (6, 1), (7, 1), (8, 1)] # fmt: skip
split_resolutions = [(el[0] * 490, el[1] * 490) for el in split_resolutions]
self.split_resolutions = split_resolutions
self.do_convert_rgb = do_convert_rgb
self.do_rescale = do_rescale
self.rescale_factor = rescale_factor
self.do_normalize = do_normalize
self.resample = resample
def preprocess(
self,
images: Union[ImageInput, list[ImageInput]],
image_mean: Optional[Union[float, list[float]]] = None,
image_std: Optional[Union[float, list[float]]] = None,
max_image_size: Optional[int] = None,
min_image_size: Optional[int] = None,
split_image: Optional[bool] = None,
do_convert_rgb: Optional[bool] = None,
do_rescale: Optional[bool] = None,
rescale_factor: Optional[float] = None,
do_normalize: Optional[bool] = None,
resample: PILImageResampling = None,
return_tensors: Optional[Union[str, TensorType]] = "pt",
data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
input_data_format: Optional[Union[str, ChannelDimension]] = None,
):
"""
Process a list of images.
Args:
images (ImageInput or list of ImageInput):
The input image or a list of images.
image_mean (`list`, *optional*, defaults to [0.5, 0.5, 0.5]):
Mean values for normalization.
image_std (`list`, *optional*, defaults to [0.5, 0.5, 0.5]):
Standard deviation values for normalization.
max_image_size (`int`, *optional*, defaults to `self.max_image_size` (980)):
Maximum image size.
min_image_size (`int`, *optional*, defaults to `self.min_image_size` (336)):
Minimum image size.
split_image (`bool`, *optional*, defaults to `self.split_image` (False)):
Whether to split the image.
do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb` (True)):
Whether to convert the image to RGB.
do_rescale (`bool`, *optional*, defaults to `self.do_rescale`):
Whether to rescale the image.
rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`):
Rescale factor to rescale the image by if `do_rescale` is set to `True`.
do_normalize (`bool`, *optional*, defaults to `self.do_normalize` (True)):
Whether to normalize the image.
resample (PILImageResampling, *optional*, defaults to `self.resample` (BICUBIC)):
The resampling filter to use if resizing the image.
return_tensors (`str` or `TensorType`, *optional*, defaults to "pt"):
The type of tensor to return.
data_format (`str` or `ChannelDimension`, *optional*):
The channel dimension format for the output image. Can be one of:
- `"channels_first"` or `ChannelDimension.FIRST`:
image in (num_channels, height, width) format.
- `"channels_last"` or `ChannelDimension.LAST`:
image in (height, width, num_channels) format.
If unset, will use same as the input image.
input_data_format (`str` or `ChannelDimension`, *optional*):
The channel dimension format for the input image. Can be one of:
- `"channels_first"` or `ChannelDimension.FIRST`:
image in (num_channels, height, width) format.
- `"channels_last"` or `ChannelDimension.LAST`:
image in (height, width, num_channels) format.
If unset, will use the inferred format of the input image.
Returns:
BatchFeature:
A BatchFeature object containing:
- 'pixel_values':
Tensor of processed image pixel values.
- 'pixel_mask':
Boolean pixel mask. This mask is a 2D tensor of shape (max_image_size, max_image_size) where:
- True (1) values indicate pixels that belong to the original resized image.
- False (0) values indicate pixels that are part of the padding.
The mask helps distinguish between actual image content and padded areas in subsequent processing steps.
- 'num_crops':
The maximum number of crops across all images.
"""
image_mean = image_mean if image_mean is not None else self.image_mean
image_std = image_std if image_std is not None else self.image_std
max_image_size = max_image_size if max_image_size is not None else self.max_image_size
min_image_size = min_image_size if min_image_size is not None else self.min_image_size
split_image = split_image if split_image is not None else self.split_image
do_convert_rgb = do_convert_rgb if do_convert_rgb is not None else self.do_convert_rgb
do_rescale = do_rescale if do_rescale is not None else self.do_rescale
rescale_factor = rescale_factor if rescale_factor is not None else self.rescale_factor
do_normalize = do_normalize if do_normalize is not None else self.do_normalize
resample = resample if resample is not None else self.resample
if max_image_size not in [490, 980]:
raise ValueError("max_image_size must be either 490 or 980")
images = make_flat_list_of_images(images)
if not valid_images(images):
raise ValueError(
"Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
"torch.Tensor, tf.Tensor or jax.ndarray."
)
validate_preprocess_arguments(
do_normalize=do_normalize,
image_mean=image_mean,
image_std=image_std,
resample=resample,
do_rescale=do_rescale,
rescale_factor=rescale_factor,
)
if do_convert_rgb:
images = [convert_to_rgb(image) for image in images]
# All transformations expect numpy arrays.
images = [to_numpy_array(image) for image in images]
if do_rescale and is_scaled_image(images[0]):
logger.warning_once(
"It looks like you are trying to rescale already rescaled images. If the input"
" images have pixel values between 0 and 1, set `do_rescale=False` to avoid rescaling them again."
)
if input_data_format is None:
# We assume that all images have the same channel dimension format.
input_data_format = infer_channel_dimension_format(images[0])
pixel_values = []
pixel_masks = []
num_crops = None
for image in images:
if split_image:
crop_images = self.get_image_patches(
image,
self.split_resolutions,
max_image_size,
resample,
data_format=input_data_format,
input_data_format=input_data_format,
)
else:
crop_images = [image]
if num_crops is None or len(crop_images) > num_crops:
num_crops = len(crop_images)
for crop_image in crop_images:
# At this point the scale is the rescaling factor that would bring the image to max_size in its larger dimension
h, w = get_image_size(crop_image)
scale = max_image_size / max(h, w)
if w >= h:
new_size = (max(int(h * scale), min_image_size), max_image_size) # h, w
else:
new_size = (max_image_size, max(int(w * scale), min_image_size)) # h, w
crop_image_resized = resize(
crop_image,
new_size,
resample=resample,
data_format=input_data_format,
input_data_format=input_data_format,
)
padding_bottom, padding_right = max_image_size - new_size[0], max_image_size - new_size[1]
crop_image_padded = pad(
crop_image_resized,
((0, padding_bottom), (0, padding_right)),
data_format=input_data_format,
input_data_format=input_data_format,
)
# Create a pixel mask
pixel_mask = np.zeros((max_image_size, max_image_size), dtype=bool)
pixel_mask[: new_size[0], : new_size[1]] = 1
pixel_masks.append(pixel_mask)
if do_rescale:
crop_image_padded = self.rescale(
image=crop_image_padded, scale=rescale_factor, input_data_format=input_data_format
)
if do_normalize:
crop_image_padded = self.normalize(
crop_image_padded,
self.image_mean,
self.image_std,
data_format=input_data_format,
input_data_format=input_data_format,
)
crop_image_padded = (
to_channel_dimension_format(crop_image_padded, data_format, input_data_format)
if data_format is not None
else crop_image_padded
)
pixel_values.append(crop_image_padded)
return BatchFeature(
data={
"pixel_values": np.stack(pixel_values, axis=0),
"pixel_mask": np.stack(pixel_masks, axis=0),
"num_crops": num_crops,
},
tensor_type=return_tensors,
)
def _resize_for_patching(
self, image: np.array, target_resolution: tuple, resample, input_data_format: ChannelDimension
) -> np.array:
"""
Resizes an image to a target resolution while maintaining aspect ratio.
Args:
image (np.array):
The input image.
target_resolution (tuple):
The target resolution (height, width) of the image.
resample (`PILImageResampling`):
Resampling filter to use if resizing the image.
input_data_format (`ChannelDimension` or `str`):
The channel dimension format of the input image.
Returns:
np.array: The resized and padded image.
"""
new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
# Resize the image
resized_image = resize(image, (new_height, new_width), resample=resample, input_data_format=input_data_format)
return resized_image
def _get_padding_size(self, original_resolution: tuple, target_resolution: tuple):
original_height, original_width = original_resolution
target_height, target_width = target_resolution
paste_x, r_x = divmod(target_width - original_width, 2)
paste_y, r_y = divmod(target_height - original_height, 2)
return (paste_y, paste_y + r_y), (paste_x, paste_x + r_x)
def _pad_for_patching(
self, image: np.array, target_resolution: tuple, input_data_format: ChannelDimension
) -> np.array:
"""
Pad an image to a target resolution while maintaining aspect ratio.
"""
new_resolution = get_patch_output_size(image, target_resolution, input_data_format)
padding = self._get_padding_size(new_resolution, target_resolution)
padded_image = self.pad(image, padding=padding)
return padded_image
def pad(
self,
image: np.ndarray,
padding: Union[int, tuple[int, int], Iterable[tuple[int, int]]],
mode: PaddingMode = PaddingMode.CONSTANT,
constant_values: Union[float, Iterable[float]] = 0.0,
data_format: Optional[Union[str, ChannelDimension]] = None,
input_data_format: Optional[Union[str, ChannelDimension]] = None,
) -> np.ndarray:
"""
Pads the `image` with the specified `padding` and `mode`. Padding can be in the (`height`, `width`)
dimension of in the (`num_patches`) dimension. In the second case an iterable if tuples is expected
as input.
Args:
image (`np.ndarray`):
The image to pad.
padding (`int` or `tuple[int, int]` or `Iterable[tuple[int, int]]`):
Padding to apply to the edges of the height, width axes. Can be one of three formats:
- `((before_height, after_height), (before_width, after_width))` unique pad widths for each axis.
- `((before, after),)` yields same before and after pad for height and width.
- `(pad,)` or int is a shortcut for before = after = pad width for all axes.
mode (`PaddingMode`):
The padding mode to use. Can be one of:
- `"constant"`: pads with a constant value.
- `"reflect"`: pads with the reflection of the vector mirrored on the first and last values of the
vector along each axis.
- `"replicate"`: pads with the replication of the last value on the edge of the array along each axis.
- `"symmetric"`: pads with the reflection of the vector mirrored along the edge of the array.
constant_values (`float` or `Iterable[float]`, *optional*):
The value to use for the padding if `mode` is `"constant"`.
data_format (`str` or `ChannelDimension`, *optional*):
The channel dimension format for the output image. Can be one of:
- `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
- `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
If unset, will use same as the input image.
input_data_format (`str` or `ChannelDimension`, *optional*):
The channel dimension format for the input image. Can be one of:
- `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
- `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
If unset, will use the inferred format of the input image.
Returns:
`np.ndarray`: The padded image.
"""
# call the general `pad` if padding on `height/width`, otherwise it's the `num_patched` dim
if isinstance(padding, int) or len(padding) != 4:
return pad(image, padding, mode, constant_values, data_format, input_data_format)
if input_data_format is None:
input_data_format = infer_channel_dimension_format(image)
padding_mode_mapping = {
PaddingMode.CONSTANT: "constant",
PaddingMode.REFLECT: "reflect",
PaddingMode.REPLICATE: "edge",
PaddingMode.SYMMETRIC: "symmetric",
}
image = np.pad(image, padding, mode=padding_mode_mapping[mode], constant_values=constant_values)
image = (
to_channel_dimension_format(image, data_format, input_data_format) if data_format is not None else image
)
return image
def get_image_patches(
self,
image: np.array,
grid_pinpoints: list[tuple[int, int]],
patch_size: int,
resample: PILImageResampling,
data_format: ChannelDimension,
input_data_format: ChannelDimension,
) -> list[np.array]:
"""
Process an image with variable resolutions by dividing it into patches.
Args:
image (`np.array`):
The input image to be processed.
grid_pinpoints (list[tuple[int, int]]):
A list of possible resolutions as tuples.
patch_size (`int`):
Size of the patches to divide the image into.
resample (`PILImageResampling`):
Resampling filter to use if resizing the image.
data_format (`ChannelDimension` or `str`):
The channel dimension format for the output image.
input_data_format (`ChannelDimension` or `str`):
The channel dimension format of the input image.
Returns:
`list[np.array]`: A list of NumPy arrays containing the processed image patches.
"""
if not isinstance(grid_pinpoints, list):
raise TypeError("grid_pinpoints must be a list of possible resolutions.")
possible_resolutions = grid_pinpoints
image_size = get_image_size(image, channel_dim=input_data_format)
best_resolution = select_best_resolution(image_size, possible_resolutions)
resized_image = self._resize_for_patching(
image, best_resolution, resample=resample, input_data_format=input_data_format
)
padded_image = self._pad_for_patching(resized_image, best_resolution, input_data_format=input_data_format)
patches = divide_to_patches(padded_image, patch_size=patch_size, input_data_format=input_data_format)
# make sure that all patches are in the input data format
patches = [
to_channel_dimension_format(patch, channel_dim=data_format, input_channel_dim=input_data_format)
for patch in patches
]
return patches
def get_number_of_image_patches(self, height: int, width: int, images_kwargs=None):
"""
A utility that returns number of image patches for a given image size.
Args:
height (`int`):
Height of the input image.
width (`int`):
Width of the input image.
images_kwargs (`dict`, *optional*)
Any kwargs to override defaults of the image processor.
Returns:
`int`: Number of patches per image.
"""
split_image = images_kwargs.get("split_image", self.split_image)
max_image_size = images_kwargs.get("max_image_size", self.max_image_size)
resized_height, resized_width = select_best_resolution((height, width), self.split_resolutions)
num_patches = 1 if not split_image else resized_height // max_image_size * resized_width // max_image_size
return num_patches
class AriaProcessorKwargs(ProcessingKwargs, total=False):
_defaults = {
"text_kwargs": {
"padding": False,
"return_mm_token_type_ids": False,
},
"images_kwargs": {
"max_image_size": 980,
"split_image": False,
},
"return_tensors": TensorType.PYTORCH,
}
class AriaProcessor(ProcessorMixin):
"""
AriaProcessor is a processor for the Aria model which wraps the Aria image preprocessor and the LLama slow tokenizer.
Args:
image_processor (`AriaImageProcessor`, *optional*):
The AriaImageProcessor to use for image preprocessing.
tokenizer (`PreTrainedTokenizerBase`, *optional*):
An instance of [`PreTrainedTokenizerBase`]. This should correspond with the model's text model. The tokenizer is a required input.
chat_template (`str`, *optional*):
A Jinja template which will be used to convert lists of messages in a chat into a tokenizable string.
size_conversion (`Dict`, *optional*):
A dictionary indicating size conversions for images.
"""
attributes = ["image_processor", "tokenizer"]
image_processor_class = "AriaImageProcessor"
tokenizer_class = "AutoTokenizer"
def __init__(
self,
image_processor=None,
tokenizer: Union[AutoTokenizer, str] = None,
chat_template: Optional[str] = None,
size_conversion: Optional[dict[Union[float, int], int]] = None,
):
if size_conversion is None:
size_conversion = {490: 128, 980: 256}
self.size_conversion = {int(k): v for k, v in size_conversion.items()}
self.image_token = tokenizer.image_token
self.image_token_id = tokenizer.image_token_id
if tokenizer is not None and tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.unk_token
super().__init__(image_processor, tokenizer, chat_template=chat_template)
def __call__(
self,
text: Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]],
images: Optional[ImageInput] = None,
audio=None,
videos=None,
**kwargs: Unpack[AriaProcessorKwargs],
) -> BatchFeature:
"""
Main method to prepare for the model one or several sequences(s) and image(s).
Args:
text (`TextInput`, `PreTokenizedInput`, `list[TextInput]`, `list[PreTokenizedInput]`):
The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
(pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
`is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
images (`ImageInput`):
The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
tensor. Both channels-first and channels-last formats are supported.
Returns:
[`BatchFeature`]: A [`BatchFeature`] with the following fields:
- **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
- **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
`return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
`None`).
- **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
- **pixel_mask** -- Pixel mask to be fed to a model. Returned when `images` is not `None`.
"""
output_kwargs = self._merge_kwargs(
AriaProcessorKwargs,
tokenizer_init_kwargs=self.tokenizer.init_kwargs,
**kwargs,
)
if isinstance(text, str):
text = [text]
elif not isinstance(text, list) and not isinstance(text[0], str):
raise TypeError("Invalid input text. Please provide a string, or a list of strings")
if images is not None:
image_inputs = self.image_processor(images, **output_kwargs["images_kwargs"])
# expand the image_token according to the num_crops and tokens per image
tokens_per_image = self.size_conversion[image_inputs.pixel_values.shape[2]]
prompt_strings = []
num_crops = image_inputs.pop("num_crops") * tokens_per_image
for sample in text:
sample = sample.replace(self.tokenizer.image_token, self.tokenizer.image_token * num_crops)
prompt_strings.append(sample)
else:
image_inputs = {}
prompt_strings = text
return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
return_mm_token_type_ids = output_kwargs["text_kwargs"].pop("return_mm_token_type_ids", False)
text_inputs = self.tokenizer(prompt_strings, **output_kwargs["text_kwargs"], return_tensors=None)
self._check_special_mm_tokens(prompt_strings, text_inputs, modalities=["image"])
if return_mm_token_type_ids:
array_ids = np.array(text_inputs["input_ids"])
mm_token_type_ids = np.zeros_like(text_inputs["input_ids"])
mm_token_type_ids[array_ids == self.image_token_id] = 1
text_inputs["mm_token_type_ids"] = mm_token_type_ids.tolist()
return BatchFeature(data={**text_inputs, **image_inputs}, tensor_type=return_tensors)
def _get_num_multimodal_tokens(self, image_sizes=None, **kwargs):
"""
Computes the number of placeholder tokens needed for multimodal inputs with the given sizes.
Args:
image_sizes (`list[list[int]]`, *optional*):
The input sizes formatted as (height, width) per each image.
Returns:
`MultiModalData`: A `MultiModalData` object holding number of tokens per each of the provided
input modalities, along with other useful data.
"""
vision_data = {}
if image_sizes is not None:
images_kwargs = AriaProcessorKwargs._defaults.get("images_kwargs", {})
images_kwargs.update(kwargs)
max_size = images_kwargs.get("max_image_size", None) or self.image_processor.max_image_size
num_image_patches = [
self.image_processor.get_number_of_image_patches(*image_size, images_kwargs)
for image_size in image_sizes
]
num_image_tokens = [self.size_conversion[max_size] * num_patches for num_patches in num_image_patches]
vision_data.update({"num_image_tokens": num_image_tokens, "num_image_patches": num_image_patches})
return MultiModalData(**vision_data)
@property
def model_input_names(self):
tokenizer_input_names = self.tokenizer.model_input_names
image_processor_input_names = self.image_processor.model_input_names
# Remove `num_crops`, it is popped and used only when processing. Make a copy of list when removing
# otherwise `self.image_processor.model_input_names` is also modified
image_processor_input_names = [name for name in image_processor_input_names if name != "num_crops"]
return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
class AriaSharedExpertsMLP(LlamaMLP):
"""
Shared Expert MLP for shared experts.
Unlike routed experts, shared experts process all tokens without routing.
This class reconfigures the intermediate size in comparison to the LlamaMLP.
Args:
config (`AriaTextConfig`): Configuration object for the Aria language model.
"""
def __init__(self, config: AriaTextConfig):
super().__init__(config)
self.intermediate_size = config.intermediate_size * config.moe_num_shared_experts
class AriaGroupedExpertsGemm(nn.Module):
"""
Grouped GEMM (General Matrix Multiplication) module for efficient expert computation.
This module utilizes the grouped_gemm library (https://github.com/fanshiqing/grouped_gemm)
for optimized performance. If the grouped_gemm library is not installed, it gracefully
falls back to a sequential GEMM implementation, which may be slower but ensures
functionality.
Args:
in_features (`int`):
Number of input features.
out_features (`int`):
Number of output features.
groups (`int`):
Number of expert groups.
"""
def __init__(self, in_features, out_features, groups):
super().__init__()
self.in_features = in_features
self.out_features = out_features
self.groups = groups
self.weight = nn.Parameter(torch.empty(groups, in_features, out_features))
def forward(self, input, tokens_per_expert):
"""
Perform grouped matrix multiplication.
Args:
input (`torch.Tensor`):
Input tensor of shape (num_tokens, in_features).
tokens_per_expert (`torch.Tensor`):
Number of tokens assigned to each expert.
Returns:
torch.Tensor: Output tensor of shape (num_tokens, out_features).
"""
return sequential_experts_gemm(
input,
self.weight,
tokens_per_expert.cpu(),
)
class AriaGroupedExpertsMLP(nn.Module):
"""
Grouped MLP module for Mixture of Experts.
Args:
config (`AriaTextConfig`):
Configuration object for the model.
"""
def __init__(self, config: AriaTextConfig) -> None:
super().__init__()
self.config = config
self.fc1 = AriaGroupedExpertsGemm(config.hidden_size, config.intermediate_size * 2, config.moe_num_experts)
self.fc2 = AriaGroupedExpertsGemm(config.intermediate_size, config.hidden_size, config.moe_num_experts)
def forward(self, permuted_tokens, tokens_per_expert):
"""
Forward pass of the Grouped MLP.
Args:
permuted_tokens (torch.Tensor): Permuted input tokens.
tokens_per_expert (torch.Tensor): Number of tokens assigned to each expert.
Returns:
torch.Tensor: Output tensor after passing through the MLP.
"""
fc1_output = self.fc1(permuted_tokens, tokens_per_expert)
projection, gate = torch.chunk(fc1_output, 2, dim=-1)
fc1_output = nn.functional.silu(projection) * gate
fc2_output = self.fc2(fc1_output, tokens_per_expert)
return fc2_output
# Token permutation adapted from https://github.com/NVIDIA/Megatron-LM/blob/54f1f78529cbc2b9cddad313e7f9d96ac0420a27/megatron/core/transformer/moe/token_dispatcher.py#L291-L587
class AriaTextMoELayer(nn.Module):
"""
Aria Text Mixture of Experts (MoE) Layer.
This layer applies a gating mechanism to route input tokens to different experts.
Args:
config (`AriaTextConfig`):
Configuration object for the text component of the model.
"""
def __init__(self, config: AriaTextConfig):
super().__init__()
self.router = nn.Linear(config.hidden_size, config.moe_num_experts, bias=False)
self.experts = AriaGroupedExpertsMLP(config)
self.shared_experts = AriaSharedExpertsMLP(config)
self.config = config
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
"""
Forward pass of the MoE Layer.
Args:
hidden_states (`torch.Tensor`):
Input tensor of shape (batch_size, sequence_length, hidden_size).
Returns:
torch.Tensor: Output tensor after passing through the MoE layer.
Process:
1. Route tokens to experts using the router.
2. Permute tokens based on routing decisions.
3. Process tokens through experts.
4. Unpermute and combine expert outputs.
5. Add shared expert output to the final result.
"""
original_shape = hidden_states.shape
hidden_states = hidden_states.view(-1, hidden_states.size(-1))
# Top K Routing
logits = self.router(hidden_states)
top_logits, top_indices = torch.topk(logits, k=self.config.moe_topk, dim=1)
scores = nn.functional.softmax(top_logits, dim=-1)
original_dtype = top_indices.dtype
tokens_per_expert = torch.histc(
top_indices.flatten().to(torch.float32),
bins=self.config.moe_num_experts,
min=0,
max=self.config.moe_num_experts - 1,
).to(original_dtype)
indices = top_indices
# Token permutation
flatten_indices = indices.view(-1)
sorted_indices = torch.argsort(flatten_indices)
permuted_tokens = hidden_states.index_select(0, sorted_indices // self.config.moe_topk)
# Process through experts
expert_output = self.experts(permuted_tokens, tokens_per_expert)
# Token unpermutation
unpermuted_tokens = torch.zeros(
(scores.shape[0] * self.config.moe_topk, expert_output.size(1)),
dtype=expert_output.dtype,
device=expert_output.device,
)
unpermuted_tokens.index_copy_(0, sorted_indices, expert_output)
unpermuted_tokens = unpermuted_tokens.view(-1, self.config.moe_topk, expert_output.size(1))
output = (unpermuted_tokens * scores.unsqueeze(-1)).sum(dim=1).view(original_shape)
# Add shared expert output
shared_expert_output = self.shared_experts(hidden_states.view(original_shape))
return output + shared_expert_output
class AriaTextAttention(LlamaAttention):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
pass
class AriaTextDecoderLayer(LlamaDecoderLayer):
"""
Aria Text Decoder Layer.
This class defines a single decoder layer in the language model, incorporating self-attention and Mixture of Experts (MoE) feed-forward network.
Args:
config (`AriaTextConfig`):
Configuration object for the text component of the model.
layer_idx (`int`):
Index of the layer.
"""
def __init__(self, config: AriaTextConfig, layer_idx: int):
super().__init__(config, layer_idx)
self.mlp = AriaTextMoELayer(config)
@auto_docstring
class AriaTextPreTrainedModel(PreTrainedModel):
config: AriaTextConfig
base_model_prefix = "model"
_no_split_modules = ["AriaTextDecoderLayer", "AriaGroupedExpertsGemm"]
supports_gradient_checkpointing = True
_skip_keys_device_placement = "past_key_values"
_supports_flash_attn = True
_supports_sdpa = True
_supports_attention_backend = True
_can_record_outputs = {
"hidden_states": AriaTextDecoderLayer,
"attentions": AriaTextAttention,
}
def _init_weights(self, module):
super()._init_weights(module)
if isinstance(module, AriaGroupedExpertsGemm):
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
class AriaPreTrainedModel(LlamaPreTrainedModel):
config: AriaConfig
base_model_prefix = ""
_can_compile_fullgraph = False # MoE models don't work with torch.compile (dynamic slicing)
_supports_attention_backend = True
def _init_weights(self, module):
LlamaPreTrainedModel._init_weights(self, module)
if isinstance(module, AriaProjector):
nn.init.trunc_normal_(module.query, std=self.config.initializer_range)
class AriaTextModel(LlamaModel):
def __init__(self, config: AriaTextConfig):
super().__init__(config)
self.layers = nn.ModuleList(
[AriaTextDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
)
self.gradient_checkpointing = False
self.post_init()
class AriaTextForCausalLM(AriaTextPreTrainedModel, LlamaForCausalLM):
_tied_weights_keys = ["lm_head.weight"]
def __init__(self, config: AriaTextConfig):
super().__init__(config)
self.model = AriaTextModel(config)
self.vocab_size = config.vocab_size
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
# Initialize weights and apply final processing
self.post_init()
@auto_docstring
def forward(self, **super_kwargs):
super().forward(self, **super_kwargs)
class AriaCausalLMOutputWithPast(LlavaCausalLMOutputWithPast):
pass
class AriaModelOutputWithPast(LlavaModelOutputWithPast):
pass
class AriaModel(LlavaModel):
def __init__(self, config: AriaConfig):
super().__init__(config)
self.multi_modal_projector = AriaProjector(config)
def _create_patch_attention_mask(self, pixel_mask):
if pixel_mask is None:
return None
patches_subgrid = pixel_mask.unfold(
dimension=1,
size=self.vision_tower.config.patch_size,
step=self.vision_tower.config.patch_size,
)
patches_subgrid = patches_subgrid.unfold(
dimension=2,
size=self.vision_tower.config.patch_size,
step=self.vision_tower.config.patch_size,
)
return (patches_subgrid.sum(dim=(-1, -2)) > 0).bool()
def get_image_features(
self,
pixel_values: torch.FloatTensor,
pixel_mask: Optional[torch.FloatTensor] = None,
vision_feature_layer: int = -1,
):
"""
Obtains image last hidden states from the vision tower and apply multimodal projection.
Args:
pixel_values (`torch.FloatTensor]` of shape `(batch_size, channels, height, width)`):
The tensors corresponding to the input images.
pixel_mask (`torch.FloatTensor]`, *optional*):
The tensors corresponding to the input image mask.
vision_feature_layer (`Union[int, list[int]]`, *optional*):
The index of the layer to select the vision feature. If multiple indices are provided,
the vision feature of the corresponding indices will be concatenated to form the
vision features.
Returns:
image_features (`torch.Tensor`): Image feature tensor of shape `(num_images, image_length, embed_dim)`).
"""
vision_feature_layer = (
vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
)
patch_attention_mask = self._create_patch_attention_mask(pixel_mask)
image_outputs = self.vision_tower(
pixel_values, patch_attention_mask=patch_attention_mask, output_hidden_states=True
)
image_attn_mask = None
if patch_attention_mask is not None:
flattened_mask = patch_attention_mask.flatten(1)
image_attn_mask = torch.logical_not(flattened_mask)
selected_image_feature = image_outputs.hidden_states[vision_feature_layer]
image_features = self.multi_modal_projector(selected_image_feature, attn_mask=image_attn_mask)
return image_features
def forward(
self,
input_ids: torch.LongTensor = None,
pixel_values: torch.FloatTensor = None,
pixel_mask: torch.LongTensor = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[Cache] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
use_cache: Optional[bool] = None,
cache_position: Optional[torch.LongTensor] = None,
**kwargs: Unpack[FlashAttentionKwargs],
) -> Union[tuple, AriaModelOutputWithPast]:
if inputs_embeds is None:
inputs_embeds = self.get_input_embeddings()(input_ids)
# 2. Merge text and images
if pixel_values is not None and inputs_embeds.shape[1] != 1:
image_features = self.get_image_features(
pixel_values=pixel_values,
pixel_mask=pixel_mask,
vision_feature_layer=self.config.vision_feature_layer,
)
image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
special_image_mask = self.get_placeholder_mask(
input_ids, inputs_embeds=inputs_embeds, image_features=image_features
)
inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
outputs = self.language_model(
attention_mask=attention_mask,
position_ids=position_ids,
past_key_values=past_key_values,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
cache_position=cache_position,
**kwargs,
)
return AriaModelOutputWithPast(
last_hidden_state=outputs.last_hidden_state,
past_key_values=outputs.past_key_values if use_cache else None,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
image_hidden_states=image_features if pixel_values is not None else None,
)
@auto_docstring(
custom_intro="""
Aria model for conditional generation tasks.
This model combines a vision tower, a multi-modal projector, and a language model
to perform tasks that involve both image and text inputs.
"""
)
class AriaForConditionalGeneration(LlavaForConditionalGeneration):
def get_image_features(
self,
pixel_values: torch.FloatTensor,
pixel_mask: Optional[torch.FloatTensor] = None,
vision_feature_layer: int = -1,
):
return self.model.get_image_features(
pixel_values=pixel_values,
pixel_mask=pixel_mask,
vision_feature_layer=vision_feature_layer,
)
@can_return_tuple
@auto_docstring
def forward(
self,
input_ids: torch.LongTensor = None,
pixel_values: torch.FloatTensor = None,
pixel_mask: torch.LongTensor = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[Cache] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
labels: Optional[torch.LongTensor] = None,
use_cache: Optional[bool] = None,
logits_to_keep: Union[int, torch.Tensor] = 0,
cache_position: Optional[torch.LongTensor] = None,
**kwargs: Unpack[TransformersKwargs],
) -> Union[tuple, AriaCausalLMOutputWithPast]:
r"""
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
config.vocab_size]` or `model.image_token_id` (where `model` is your instance of `AriaForConditionalGeneration`).
Tokens with indices set to `model.image_token_id` are ignored (masked), the loss is only
computed for the tokens with labels in `[0, ..., config.vocab_size]`.
Example:
```python
>>> import requests
>>> import torch
>>> from PIL import Image
>>> from io import BytesIO
>>> from transformers import AutoProcessor, AutoModel
>>> from transformers.image_utils import load_image
>>> # Note that passing the image urls (instead of the actual pil images) to the processor is also possible
>>> image1 = load_image("https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg")
>>> image2 = load_image("https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg")
>>> image3 = load_image("https://cdn.britannica.com/68/170868-050-8DDE8263/Golden-Gate-Bridge-San-Francisco.jpg")
>>> processor = AutoProcessor.from_pretrained("Rhymes-AI/Aria")
>>> model = AutoModel.from_pretrained("Rhymes-AI/Aria", dtype=torch.bfloat16, device_map="auto")
>>> # Create inputs
>>> messages = [
... {
... "role": "user",
... "content": [
... {"type": "image"},
... {"type": "text", "text": "In this image, we can see the city of New York, and more specifically the Statue of Liberty."},
... {"type": "image"},
... {"type": "text", "text": "What can we see in this image?"},
... ]
... },
... {
... "role": "user",
... "content": [
... {"type": "image"},
... {"type": "text", "text": "In which city is that bridge located?"},
... ]
... }
... ]
>>> prompts = [processor.apply_chat_template([message], add_generation_prompt=True) for message in messages]
>>> images = [[image1, image2], [image3]]
>>> inputs = processor(text=prompts, images=images, padding=True, return_tensors="pt").to(model.device)
>>> # Generate
>>> generated_ids = model.generate(**inputs, max_new_tokens=256)
>>> generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)
>>> print(generated_texts[0])
Assistant: There are buildings, trees, lights, and water visible in this image.
>>> print(generated_texts[1])
Assistant: The bridge is in San Francisco.
```"""
outputs = self.model(
input_ids=input_ids,
pixel_values=pixel_values,
pixel_mask=pixel_mask,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_values=past_key_values,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
cache_position=cache_position,
**kwargs,
)
hidden_states = outputs[0]
# Only compute necessary logits, and do not upcast them to float if we are not computing the loss
slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
logits = self.lm_head(hidden_states[:, slice_indices, :])
loss = None
if labels is not None:
loss = self.loss_function(
logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **kwargs
)
return AriaCausalLMOutputWithPast(
loss=loss,
logits=logits,
past_key_values=outputs.past_key_values,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
def prepare_inputs_for_generation(
self,
input_ids,
past_key_values=None,
inputs_embeds=None,
pixel_values=None,
pixel_mask=None,
attention_mask=None,
cache_position=None,
logits_to_keep=None,
**kwargs,
):
model_inputs = super().prepare_inputs_for_generation(
input_ids,
past_key_values=past_key_values,
inputs_embeds=inputs_embeds,
attention_mask=attention_mask,
cache_position=cache_position,
logits_to_keep=logits_to_keep,
**kwargs,
)
if cache_position[0] == 0:
# If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
# Otherwise we need pixel values to be passed to model
model_inputs["pixel_values"] = pixel_values
model_inputs["pixel_mask"] = pixel_mask
return model_inputs
__all__ = [
"AriaConfig",
"AriaTextConfig",
"AriaImageProcessor",
"AriaProcessor",
"AriaForConditionalGeneration",
"AriaPreTrainedModel",
"AriaTextPreTrainedModel",
"AriaTextModel",
"AriaModel",
"AriaTextForCausalLM",
]
| transformers/src/transformers/models/aria/modular_aria.py/0 | {
"file_path": "transformers/src/transformers/models/aria/modular_aria.py",
"repo_id": "transformers",
"token_count": 30331
} | 429 |
# coding=utf-8
# Copyright 2018 The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Auto Tokenizer class."""
import importlib
import json
import os
import warnings
from collections import OrderedDict
from typing import Any, Optional, Union
from transformers.utils.import_utils import is_mistral_common_available
from ...configuration_utils import PretrainedConfig
from ...dynamic_module_utils import get_class_from_dynamic_module, resolve_trust_remote_code
from ...modeling_gguf_pytorch_utils import load_gguf_checkpoint
from ...tokenization_utils import PreTrainedTokenizer
from ...tokenization_utils_base import TOKENIZER_CONFIG_FILE
from ...utils import (
cached_file,
extract_commit_hash,
is_g2p_en_available,
is_sentencepiece_available,
is_tokenizers_available,
logging,
)
from ..encoder_decoder import EncoderDecoderConfig
from .auto_factory import _LazyAutoMapping
from .configuration_auto import (
CONFIG_MAPPING_NAMES,
AutoConfig,
config_class_to_model_type,
model_type_to_module_name,
replace_list_option_in_docstrings,
)
if is_tokenizers_available():
from ...tokenization_utils_fast import PreTrainedTokenizerFast
else:
PreTrainedTokenizerFast = None
logger = logging.get_logger(__name__)
# Explicit rather than inferred generics to significantly improves completion suggestion performance for language servers.
TOKENIZER_MAPPING_NAMES = OrderedDict[str, tuple[Optional[str], Optional[str]]](
[
(
"aimv2",
(
"CLIPTokenizer",
"CLIPTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"albert",
(
"AlbertTokenizer" if is_sentencepiece_available() else None,
"AlbertTokenizerFast" if is_tokenizers_available() else None,
),
),
("align", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
("arcee", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("aria", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("aya_vision", (None, "CohereTokenizerFast" if is_tokenizers_available() else None)),
("bark", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
("bart", ("BartTokenizer", "BartTokenizerFast")),
(
"barthez",
(
"BarthezTokenizer" if is_sentencepiece_available() else None,
"BarthezTokenizerFast" if is_tokenizers_available() else None,
),
),
("bartpho", ("BartphoTokenizer", None)),
("bert", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
("bert-generation", ("BertGenerationTokenizer" if is_sentencepiece_available() else None, None)),
("bert-japanese", ("BertJapaneseTokenizer", None)),
("bertweet", ("BertweetTokenizer", None)),
(
"big_bird",
(
"BigBirdTokenizer" if is_sentencepiece_available() else None,
"BigBirdTokenizerFast" if is_tokenizers_available() else None,
),
),
("bigbird_pegasus", ("PegasusTokenizer", "PegasusTokenizerFast" if is_tokenizers_available() else None)),
("biogpt", ("BioGptTokenizer", None)),
("bitnet", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
("blenderbot", ("BlenderbotTokenizer", "BlenderbotTokenizerFast")),
("blenderbot-small", ("BlenderbotSmallTokenizer", None)),
("blip", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
("blip-2", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
("bloom", (None, "BloomTokenizerFast" if is_tokenizers_available() else None)),
("bridgetower", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
("bros", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
("byt5", ("ByT5Tokenizer", None)),
(
"camembert",
(
"CamembertTokenizer" if is_sentencepiece_available() else None,
"CamembertTokenizerFast" if is_tokenizers_available() else None,
),
),
("canine", ("CanineTokenizer", None)),
(
"chameleon",
(
"LlamaTokenizer" if is_sentencepiece_available() else None,
"LlamaTokenizerFast" if is_tokenizers_available() else None,
),
),
("chinese_clip", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
(
"clap",
(
"RobertaTokenizer",
"RobertaTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"clip",
(
"CLIPTokenizer",
"CLIPTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"clipseg",
(
"CLIPTokenizer",
"CLIPTokenizerFast" if is_tokenizers_available() else None,
),
),
("clvp", ("ClvpTokenizer", None)),
(
"code_llama",
(
"CodeLlamaTokenizer" if is_sentencepiece_available() else None,
"CodeLlamaTokenizerFast" if is_tokenizers_available() else None,
),
),
("codegen", ("CodeGenTokenizer", "CodeGenTokenizerFast" if is_tokenizers_available() else None)),
("cohere", (None, "CohereTokenizerFast" if is_tokenizers_available() else None)),
("cohere2", (None, "CohereTokenizerFast" if is_tokenizers_available() else None)),
("colpali", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("colqwen2", ("Qwen2Tokenizer", "Qwen2TokenizerFast" if is_tokenizers_available() else None)),
("convbert", ("ConvBertTokenizer", "ConvBertTokenizerFast" if is_tokenizers_available() else None)),
(
"cpm",
(
"CpmTokenizer" if is_sentencepiece_available() else None,
"CpmTokenizerFast" if is_tokenizers_available() else None,
),
),
("cpmant", ("CpmAntTokenizer", None)),
("ctrl", ("CTRLTokenizer", None)),
("data2vec-audio", ("Wav2Vec2CTCTokenizer", None)),
("data2vec-text", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
("dbrx", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
("deberta", ("DebertaTokenizer", "DebertaTokenizerFast" if is_tokenizers_available() else None)),
(
"deberta-v2",
(
"DebertaV2Tokenizer" if is_sentencepiece_available() else None,
"DebertaV2TokenizerFast" if is_tokenizers_available() else None,
),
),
(
"deepseek_v2",
(
"LlamaTokenizer" if is_sentencepiece_available() else None,
"LlamaTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"deepseek_v3",
(
"LlamaTokenizer" if is_sentencepiece_available() else None,
"LlamaTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"deepseek_vl",
(
"LlamaTokenizer" if is_sentencepiece_available() else None,
"LlamaTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"deepseek_vl_hybrid",
(
"LlamaTokenizer" if is_sentencepiece_available() else None,
"LlamaTokenizerFast" if is_tokenizers_available() else None,
),
),
("dia", ("DiaTokenizer", None)),
(
"diffllama",
(
"LlamaTokenizer" if is_sentencepiece_available() else None,
"LlamaTokenizerFast" if is_tokenizers_available() else None,
),
),
("distilbert", ("DistilBertTokenizer", "DistilBertTokenizerFast" if is_tokenizers_available() else None)),
(
"dpr",
(
"DPRQuestionEncoderTokenizer",
"DPRQuestionEncoderTokenizerFast" if is_tokenizers_available() else None,
),
),
("electra", ("ElectraTokenizer", "ElectraTokenizerFast" if is_tokenizers_available() else None)),
("emu3", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
("ernie", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
("ernie4_5", (None, "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("ernie4_5_moe", (None, "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("ernie_m", ("ErnieMTokenizer" if is_sentencepiece_available() else None, None)),
("esm", ("EsmTokenizer", None)),
(
"exaone4",
(
"GPT2Tokenizer" if is_tokenizers_available() else None,
"GPT2TokenizerFast" if is_tokenizers_available() else None,
),
),
("falcon", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
("falcon_mamba", (None, "GPTNeoXTokenizerFast" if is_tokenizers_available() else None)),
(
"fastspeech2_conformer",
("FastSpeech2ConformerTokenizer" if is_g2p_en_available() else None, None),
),
("flaubert", ("FlaubertTokenizer", None)),
("fnet", ("FNetTokenizer", "FNetTokenizerFast" if is_tokenizers_available() else None)),
("fsmt", ("FSMTTokenizer", None)),
("funnel", ("FunnelTokenizer", "FunnelTokenizerFast" if is_tokenizers_available() else None)),
(
"gemma",
(
"GemmaTokenizer" if is_sentencepiece_available() else None,
"GemmaTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"gemma2",
(
"GemmaTokenizer" if is_sentencepiece_available() else None,
"GemmaTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"gemma3",
(
"GemmaTokenizer" if is_sentencepiece_available() else None,
"GemmaTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"gemma3_text",
(
"GemmaTokenizer" if is_sentencepiece_available() else None,
"GemmaTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"gemma3n",
(
"GemmaTokenizer" if is_sentencepiece_available() else None,
"GemmaTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"gemma3n_text",
(
"GemmaTokenizer" if is_sentencepiece_available() else None,
"GemmaTokenizerFast" if is_tokenizers_available() else None,
),
),
("git", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
("glm", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
("glm4", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
("glm4_moe", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
("glm4v", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
("glm4v_moe", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
("gpt-sw3", ("GPTSw3Tokenizer" if is_sentencepiece_available() else None, None)),
("gpt2", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
("gpt_bigcode", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
("gpt_neo", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
("gpt_neox", (None, "GPTNeoXTokenizerFast" if is_tokenizers_available() else None)),
("gpt_neox_japanese", ("GPTNeoXJapaneseTokenizer", None)),
("gpt_oss", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
("gptj", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
("gptsan-japanese", ("GPTSanJapaneseTokenizer", None)),
("granite", ("GPT2Tokenizer", None)),
("granitemoe", ("GPT2Tokenizer", None)),
("granitemoehybrid", ("GPT2Tokenizer", None)),
("granitemoeshared", ("GPT2Tokenizer", None)),
("grounding-dino", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
("groupvit", ("CLIPTokenizer", "CLIPTokenizerFast" if is_tokenizers_available() else None)),
("helium", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
("herbert", ("HerbertTokenizer", "HerbertTokenizerFast" if is_tokenizers_available() else None)),
("hubert", ("Wav2Vec2CTCTokenizer", None)),
("ibert", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
("idefics", (None, "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("idefics2", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("idefics3", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("instructblip", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
("instructblipvideo", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
("internvl", ("Qwen2Tokenizer", "Qwen2TokenizerFast" if is_tokenizers_available() else None)),
(
"jamba",
(
"LlamaTokenizer" if is_sentencepiece_available() else None,
"LlamaTokenizerFast" if is_tokenizers_available() else None,
),
),
("janus", (None, "LlamaTokenizerFast" if is_tokenizers_available() else None)),
(
"jetmoe",
(
"LlamaTokenizer" if is_sentencepiece_available() else None,
"LlamaTokenizerFast" if is_tokenizers_available() else None,
),
),
("jukebox", ("JukeboxTokenizer", None)),
(
"kosmos-2",
(
"XLMRobertaTokenizer" if is_sentencepiece_available() else None,
"XLMRobertaTokenizerFast" if is_tokenizers_available() else None,
),
),
("kosmos-2.5", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
("layoutlm", ("LayoutLMTokenizer", "LayoutLMTokenizerFast" if is_tokenizers_available() else None)),
("layoutlmv2", ("LayoutLMv2Tokenizer", "LayoutLMv2TokenizerFast" if is_tokenizers_available() else None)),
("layoutlmv3", ("LayoutLMv3Tokenizer", "LayoutLMv3TokenizerFast" if is_tokenizers_available() else None)),
("layoutxlm", ("LayoutXLMTokenizer", "LayoutXLMTokenizerFast" if is_tokenizers_available() else None)),
("led", ("LEDTokenizer", "LEDTokenizerFast" if is_tokenizers_available() else None)),
("lilt", ("LayoutLMv3Tokenizer", "LayoutLMv3TokenizerFast" if is_tokenizers_available() else None)),
(
"llama",
(
"LlamaTokenizer" if is_sentencepiece_available() else None,
"LlamaTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"llama4",
(
"LlamaTokenizer" if is_sentencepiece_available() else None,
"LlamaTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"llama4_text",
(
"LlamaTokenizer" if is_sentencepiece_available() else None,
"LlamaTokenizerFast" if is_tokenizers_available() else None,
),
),
("llava", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("llava_next", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("llava_next_video", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("llava_onevision", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("longformer", ("LongformerTokenizer", "LongformerTokenizerFast" if is_tokenizers_available() else None)),
(
"longt5",
(
"T5Tokenizer" if is_sentencepiece_available() else None,
"T5TokenizerFast" if is_tokenizers_available() else None,
),
),
("luke", ("LukeTokenizer", None)),
("lxmert", ("LxmertTokenizer", "LxmertTokenizerFast" if is_tokenizers_available() else None)),
("m2m_100", ("M2M100Tokenizer" if is_sentencepiece_available() else None, None)),
("mamba", (None, "GPTNeoXTokenizerFast" if is_tokenizers_available() else None)),
("mamba2", (None, "GPTNeoXTokenizerFast" if is_tokenizers_available() else None)),
("marian", ("MarianTokenizer" if is_sentencepiece_available() else None, None)),
(
"mbart",
(
"MBartTokenizer" if is_sentencepiece_available() else None,
"MBartTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"mbart50",
(
"MBart50Tokenizer" if is_sentencepiece_available() else None,
"MBart50TokenizerFast" if is_tokenizers_available() else None,
),
),
("mega", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
("megatron-bert", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
(
"metaclip_2",
(
"XLMRobertaTokenizer",
"XLMRobertaTokenizerFast" if is_tokenizers_available() else None,
),
),
("mgp-str", ("MgpstrTokenizer", None)),
(
"minimax",
(
"GPT2Tokenizer" if is_sentencepiece_available() else None,
"GPT2TokenizerFast" if is_tokenizers_available() else None,
),
),
(
"mistral",
(
"MistralCommonTokenizer"
if is_mistral_common_available()
else ("LlamaTokenizer" if is_sentencepiece_available() else None),
"LlamaTokenizerFast" if is_tokenizers_available() and not is_mistral_common_available() else None,
),
),
(
"mixtral",
(
"MistralCommonTokenizer"
if is_mistral_common_available()
else ("LlamaTokenizer" if is_sentencepiece_available() else None),
"LlamaTokenizerFast" if is_tokenizers_available() and not is_mistral_common_available() else None,
),
),
("mllama", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("mluke", ("MLukeTokenizer" if is_sentencepiece_available() else None, None)),
("mm-grounding-dino", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
("mobilebert", ("MobileBertTokenizer", "MobileBertTokenizerFast" if is_tokenizers_available() else None)),
("modernbert", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
("moonshine", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
("moshi", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
("mpnet", ("MPNetTokenizer", "MPNetTokenizerFast" if is_tokenizers_available() else None)),
("mpt", (None, "GPTNeoXTokenizerFast" if is_tokenizers_available() else None)),
("mra", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
(
"mt5",
(
"MT5Tokenizer" if is_sentencepiece_available() else None,
"MT5TokenizerFast" if is_tokenizers_available() else None,
),
),
("musicgen", ("T5Tokenizer", "T5TokenizerFast" if is_tokenizers_available() else None)),
("musicgen_melody", ("T5Tokenizer", "T5TokenizerFast" if is_tokenizers_available() else None)),
("mvp", ("MvpTokenizer", "MvpTokenizerFast" if is_tokenizers_available() else None)),
("myt5", ("MyT5Tokenizer", None)),
("nemotron", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
("nezha", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
(
"nllb",
(
"NllbTokenizer" if is_sentencepiece_available() else None,
"NllbTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"nllb-moe",
(
"NllbTokenizer" if is_sentencepiece_available() else None,
"NllbTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"nystromformer",
(
"AlbertTokenizer" if is_sentencepiece_available() else None,
"AlbertTokenizerFast" if is_tokenizers_available() else None,
),
),
("olmo", (None, "GPTNeoXTokenizerFast" if is_tokenizers_available() else None)),
("olmo2", (None, "GPTNeoXTokenizerFast" if is_tokenizers_available() else None)),
("olmoe", (None, "GPTNeoXTokenizerFast" if is_tokenizers_available() else None)),
(
"omdet-turbo",
("CLIPTokenizer", "CLIPTokenizerFast" if is_tokenizers_available() else None),
),
("oneformer", ("CLIPTokenizer", "CLIPTokenizerFast" if is_tokenizers_available() else None)),
(
"openai-gpt",
("OpenAIGPTTokenizer", "OpenAIGPTTokenizerFast" if is_tokenizers_available() else None),
),
("opt", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
("owlv2", ("CLIPTokenizer", "CLIPTokenizerFast" if is_tokenizers_available() else None)),
("owlvit", ("CLIPTokenizer", "CLIPTokenizerFast" if is_tokenizers_available() else None)),
("paligemma", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
(
"pegasus",
(
"PegasusTokenizer" if is_sentencepiece_available() else None,
"PegasusTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"pegasus_x",
(
"PegasusTokenizer" if is_sentencepiece_available() else None,
"PegasusTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"perceiver",
(
"PerceiverTokenizer",
None,
),
),
(
"persimmon",
(
"LlamaTokenizer" if is_sentencepiece_available() else None,
"LlamaTokenizerFast" if is_tokenizers_available() else None,
),
),
("phi", ("CodeGenTokenizer", "CodeGenTokenizerFast" if is_tokenizers_available() else None)),
("phi3", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("phimoe", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("phobert", ("PhobertTokenizer", None)),
("pix2struct", ("T5Tokenizer", "T5TokenizerFast" if is_tokenizers_available() else None)),
(
"pixtral",
(
None,
"MistralCommonTokenizer"
if is_mistral_common_available()
else ("PreTrainedTokenizerFast" if is_tokenizers_available() else None),
),
),
("plbart", ("PLBartTokenizer" if is_sentencepiece_available() else None, None)),
("prophetnet", ("ProphetNetTokenizer", None)),
("qdqbert", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
(
"qwen2",
(
"Qwen2Tokenizer",
"Qwen2TokenizerFast" if is_tokenizers_available() else None,
),
),
("qwen2_5_omni", ("Qwen2Tokenizer", "Qwen2TokenizerFast" if is_tokenizers_available() else None)),
("qwen2_5_vl", ("Qwen2Tokenizer", "Qwen2TokenizerFast" if is_tokenizers_available() else None)),
("qwen2_audio", ("Qwen2Tokenizer", "Qwen2TokenizerFast" if is_tokenizers_available() else None)),
(
"qwen2_moe",
(
"Qwen2Tokenizer",
"Qwen2TokenizerFast" if is_tokenizers_available() else None,
),
),
("qwen2_vl", ("Qwen2Tokenizer", "Qwen2TokenizerFast" if is_tokenizers_available() else None)),
(
"qwen3",
(
"Qwen2Tokenizer",
"Qwen2TokenizerFast" if is_tokenizers_available() else None,
),
),
(
"qwen3_moe",
(
"Qwen2Tokenizer",
"Qwen2TokenizerFast" if is_tokenizers_available() else None,
),
),
("rag", ("RagTokenizer", None)),
("realm", ("RealmTokenizer", "RealmTokenizerFast" if is_tokenizers_available() else None)),
(
"recurrent_gemma",
(
"GemmaTokenizer" if is_sentencepiece_available() else None,
"GemmaTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"reformer",
(
"ReformerTokenizer" if is_sentencepiece_available() else None,
"ReformerTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"rembert",
(
"RemBertTokenizer" if is_sentencepiece_available() else None,
"RemBertTokenizerFast" if is_tokenizers_available() else None,
),
),
("retribert", ("RetriBertTokenizer", "RetriBertTokenizerFast" if is_tokenizers_available() else None)),
("roberta", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
(
"roberta-prelayernorm",
("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None),
),
("roc_bert", ("RoCBertTokenizer", None)),
("roformer", ("RoFormerTokenizer", "RoFormerTokenizerFast" if is_tokenizers_available() else None)),
("rwkv", (None, "GPTNeoXTokenizerFast" if is_tokenizers_available() else None)),
(
"seamless_m4t",
(
"SeamlessM4TTokenizer" if is_sentencepiece_available() else None,
"SeamlessM4TTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"seamless_m4t_v2",
(
"SeamlessM4TTokenizer" if is_sentencepiece_available() else None,
"SeamlessM4TTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"shieldgemma2",
(
"GemmaTokenizer" if is_sentencepiece_available() else None,
"GemmaTokenizerFast" if is_tokenizers_available() else None,
),
),
("siglip", ("SiglipTokenizer" if is_sentencepiece_available() else None, None)),
(
"siglip2",
(
"GemmaTokenizer" if is_sentencepiece_available() else None,
"GemmaTokenizerFast" if is_tokenizers_available() else None,
),
),
("smollm3", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
("speech_to_text", ("Speech2TextTokenizer" if is_sentencepiece_available() else None, None)),
("speech_to_text_2", ("Speech2Text2Tokenizer", None)),
("speecht5", ("SpeechT5Tokenizer" if is_sentencepiece_available() else None, None)),
("splinter", ("SplinterTokenizer", "SplinterTokenizerFast")),
(
"squeezebert",
("SqueezeBertTokenizer", "SqueezeBertTokenizerFast" if is_tokenizers_available() else None),
),
("stablelm", (None, "GPTNeoXTokenizerFast" if is_tokenizers_available() else None)),
("starcoder2", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
(
"switch_transformers",
(
"T5Tokenizer" if is_sentencepiece_available() else None,
"T5TokenizerFast" if is_tokenizers_available() else None,
),
),
(
"t5",
(
"T5Tokenizer" if is_sentencepiece_available() else None,
"T5TokenizerFast" if is_tokenizers_available() else None,
),
),
(
"t5gemma",
(
"GemmaTokenizer" if is_sentencepiece_available() else None,
"GemmaTokenizerFast" if is_tokenizers_available() else None,
),
),
("tapas", ("TapasTokenizer", None)),
("tapex", ("TapexTokenizer", None)),
("transfo-xl", ("TransfoXLTokenizer", None)),
("tvp", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
(
"udop",
(
"UdopTokenizer" if is_sentencepiece_available() else None,
"UdopTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"umt5",
(
"T5Tokenizer" if is_sentencepiece_available() else None,
"T5TokenizerFast" if is_tokenizers_available() else None,
),
),
("video_llava", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("vilt", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
("vipllava", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("visual_bert", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
("vits", ("VitsTokenizer", None)),
(
"voxtral",
(
"MistralCommonTokenizer"
if is_mistral_common_available()
else ("LlamaTokenizer" if is_sentencepiece_available() else None),
"LlamaTokenizerFast" if is_tokenizers_available() and not is_mistral_common_available() else None,
),
),
("wav2vec2", ("Wav2Vec2CTCTokenizer", None)),
("wav2vec2-bert", ("Wav2Vec2CTCTokenizer", None)),
("wav2vec2-conformer", ("Wav2Vec2CTCTokenizer", None)),
("wav2vec2_phoneme", ("Wav2Vec2PhonemeCTCTokenizer", None)),
("whisper", ("WhisperTokenizer", "WhisperTokenizerFast" if is_tokenizers_available() else None)),
("xclip", ("CLIPTokenizer", "CLIPTokenizerFast" if is_tokenizers_available() else None)),
(
"xglm",
(
"XGLMTokenizer" if is_sentencepiece_available() else None,
"XGLMTokenizerFast" if is_tokenizers_available() else None,
),
),
("xlm", ("XLMTokenizer", None)),
("xlm-prophetnet", ("XLMProphetNetTokenizer" if is_sentencepiece_available() else None, None)),
(
"xlm-roberta",
(
"XLMRobertaTokenizer" if is_sentencepiece_available() else None,
"XLMRobertaTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"xlm-roberta-xl",
(
"XLMRobertaTokenizer" if is_sentencepiece_available() else None,
"XLMRobertaTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"xlnet",
(
"XLNetTokenizer" if is_sentencepiece_available() else None,
"XLNetTokenizerFast" if is_tokenizers_available() else None,
),
),
("xlstm", (None, "GPTNeoXTokenizerFast" if is_tokenizers_available() else None)),
(
"xmod",
(
"XLMRobertaTokenizer" if is_sentencepiece_available() else None,
"XLMRobertaTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"yoso",
(
"AlbertTokenizer" if is_sentencepiece_available() else None,
"AlbertTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"zamba",
(
"LlamaTokenizer" if is_sentencepiece_available() else None,
"LlamaTokenizerFast" if is_tokenizers_available() else None,
),
),
(
"zamba2",
(
"LlamaTokenizer" if is_sentencepiece_available() else None,
"LlamaTokenizerFast" if is_tokenizers_available() else None,
),
),
]
)
TOKENIZER_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, TOKENIZER_MAPPING_NAMES)
CONFIG_TO_TYPE = {v: k for k, v in CONFIG_MAPPING_NAMES.items()}
def tokenizer_class_from_name(class_name: str) -> Union[type[Any], None]:
if class_name == "PreTrainedTokenizerFast":
return PreTrainedTokenizerFast
for module_name, tokenizers in TOKENIZER_MAPPING_NAMES.items():
if class_name in tokenizers:
module_name = model_type_to_module_name(module_name)
if module_name in ["mistral", "mixtral"] and class_name == "MistralCommonTokenizer":
module = importlib.import_module(".tokenization_mistral_common", "transformers")
else:
module = importlib.import_module(f".{module_name}", "transformers.models")
try:
return getattr(module, class_name)
except AttributeError:
continue
for tokenizers in TOKENIZER_MAPPING._extra_content.values():
for tokenizer in tokenizers:
if getattr(tokenizer, "__name__", None) == class_name:
return tokenizer
# We did not fine the class, but maybe it's because a dep is missing. In that case, the class will be in the main
# init and we return the proper dummy to get an appropriate error message.
main_module = importlib.import_module("transformers")
if hasattr(main_module, class_name):
return getattr(main_module, class_name)
return None
def get_tokenizer_config(
pretrained_model_name_or_path: Union[str, os.PathLike[str]],
cache_dir: Optional[Union[str, os.PathLike[str]]] = None,
force_download: bool = False,
resume_download: Optional[bool] = None,
proxies: Optional[dict[str, str]] = None,
token: Optional[Union[bool, str]] = None,
revision: Optional[str] = None,
local_files_only: bool = False,
subfolder: str = "",
**kwargs,
) -> dict[str, Any]:
"""
Loads the tokenizer configuration from a pretrained model tokenizer configuration.
Args:
pretrained_model_name_or_path (`str` or `os.PathLike`):
This can be either:
- a string, the *model id* of a pretrained model configuration hosted inside a model repo on
huggingface.co.
- a path to a *directory* containing a configuration file saved using the
[`~PreTrainedTokenizer.save_pretrained`] method, e.g., `./my_model_directory/`.
cache_dir (`str` or `os.PathLike`, *optional*):
Path to a directory in which a downloaded pretrained model configuration should be cached if the standard
cache should not be used.
force_download (`bool`, *optional*, defaults to `False`):
Whether or not to force to (re-)download the configuration files and override the cached versions if they
exist.
resume_download:
Deprecated and ignored. All downloads are now resumed by default when possible.
Will be removed in v5 of Transformers.
proxies (`dict[str, str]`, *optional*):
A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
token (`str` or *bool*, *optional*):
The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
when running `hf auth login` (stored in `~/.huggingface`).
revision (`str`, *optional*, defaults to `"main"`):
The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
identifier allowed by git.
local_files_only (`bool`, *optional*, defaults to `False`):
If `True`, will only try to load the tokenizer configuration from local files.
subfolder (`str`, *optional*, defaults to `""`):
In case the tokenizer config is located inside a subfolder of the model repo on huggingface.co, you can
specify the folder name here.
<Tip>
Passing `token=True` is required when you want to use a private model.
</Tip>
Returns:
`dict`: The configuration of the tokenizer.
Examples:
```python
# Download configuration from huggingface.co and cache.
tokenizer_config = get_tokenizer_config("google-bert/bert-base-uncased")
# This model does not have a tokenizer config so the result will be an empty dict.
tokenizer_config = get_tokenizer_config("FacebookAI/xlm-roberta-base")
# Save a pretrained tokenizer locally and you can reload its config
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-cased")
tokenizer.save_pretrained("tokenizer-test")
tokenizer_config = get_tokenizer_config("tokenizer-test")
```"""
use_auth_token = kwargs.pop("use_auth_token", None)
if use_auth_token is not None:
warnings.warn(
"The `use_auth_token` argument is deprecated and will be removed in v5 of Transformers. Please use `token` instead.",
FutureWarning,
)
if token is not None:
raise ValueError("`token` and `use_auth_token` are both specified. Please set only the argument `token`.")
token = use_auth_token
commit_hash = kwargs.get("_commit_hash")
resolved_config_file = cached_file(
pretrained_model_name_or_path,
TOKENIZER_CONFIG_FILE,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
token=token,
revision=revision,
local_files_only=local_files_only,
subfolder=subfolder,
_raise_exceptions_for_gated_repo=False,
_raise_exceptions_for_missing_entries=False,
_raise_exceptions_for_connection_errors=False,
_commit_hash=commit_hash,
)
if resolved_config_file is None:
logger.info("Could not locate the tokenizer configuration file, will try to use the model config instead.")
return {}
commit_hash = extract_commit_hash(resolved_config_file, commit_hash)
with open(resolved_config_file, encoding="utf-8") as reader:
result = json.load(reader)
result["_commit_hash"] = commit_hash
return result
class AutoTokenizer:
r"""
This is a generic tokenizer class that will be instantiated as one of the tokenizer classes of the library when
created with the [`AutoTokenizer.from_pretrained`] class method.
This class cannot be instantiated directly using `__init__()` (throws an error).
"""
def __init__(self):
raise OSError(
"AutoTokenizer is designed to be instantiated "
"using the `AutoTokenizer.from_pretrained(pretrained_model_name_or_path)` method."
)
@classmethod
@replace_list_option_in_docstrings(TOKENIZER_MAPPING_NAMES)
def from_pretrained(cls, pretrained_model_name_or_path, *inputs, **kwargs):
r"""
Instantiate one of the tokenizer classes of the library from a pretrained model vocabulary.
The tokenizer class to instantiate is selected based on the `model_type` property of the config object (either
passed as an argument or loaded from `pretrained_model_name_or_path` if possible), or when it's missing, by
falling back to using pattern matching on `pretrained_model_name_or_path`:
List options
Params:
pretrained_model_name_or_path (`str` or `os.PathLike`):
Can be either:
- A string, the *model id* of a predefined tokenizer hosted inside a model repo on huggingface.co.
- A path to a *directory* containing vocabulary files required by the tokenizer, for instance saved
using the [`~PreTrainedTokenizer.save_pretrained`] method, e.g., `./my_model_directory/`.
- A path or url to a single saved vocabulary file if and only if the tokenizer only requires a
single vocabulary file (like Bert or XLNet), e.g.: `./my_model_directory/vocab.txt`. (Not
applicable to all derived classes)
inputs (additional positional arguments, *optional*):
Will be passed along to the Tokenizer `__init__()` method.
config ([`PretrainedConfig`], *optional*)
The configuration object used to determine the tokenizer class to instantiate.
cache_dir (`str` or `os.PathLike`, *optional*):
Path to a directory in which a downloaded pretrained model configuration should be cached if the
standard cache should not be used.
force_download (`bool`, *optional*, defaults to `False`):
Whether or not to force the (re-)download the model weights and configuration files and override the
cached versions if they exist.
resume_download:
Deprecated and ignored. All downloads are now resumed by default when possible.
Will be removed in v5 of Transformers.
proxies (`dict[str, str]`, *optional*):
A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
revision (`str`, *optional*, defaults to `"main"`):
The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
identifier allowed by git.
subfolder (`str`, *optional*):
In case the relevant files are located inside a subfolder of the model repo on huggingface.co (e.g. for
facebook/rag-token-base), specify it here.
use_fast (`bool`, *optional*, defaults to `True`):
Use a [fast Rust-based tokenizer](https://huggingface.co/docs/tokenizers/index) if it is supported for
a given model. If a fast tokenizer is not available for a given model, a normal Python-based tokenizer
is returned instead.
tokenizer_type (`str`, *optional*):
Tokenizer type to be loaded.
trust_remote_code (`bool`, *optional*, defaults to `False`):
Whether or not to allow for custom models defined on the Hub in their own modeling files. This option
should only be set to `True` for repositories you trust and in which you have read the code, as it will
execute code present on the Hub on your local machine.
kwargs (additional keyword arguments, *optional*):
Will be passed to the Tokenizer `__init__()` method. Can be used to set special tokens like
`bos_token`, `eos_token`, `unk_token`, `sep_token`, `pad_token`, `cls_token`, `mask_token`,
`additional_special_tokens`. See parameters in the `__init__()` for more details.
Examples:
```python
>>> from transformers import AutoTokenizer
>>> # Download vocabulary from huggingface.co and cache.
>>> tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-uncased")
>>> # Download vocabulary from huggingface.co (user-uploaded) and cache.
>>> tokenizer = AutoTokenizer.from_pretrained("dbmdz/bert-base-german-cased")
>>> # If vocabulary files are in a directory (e.g. tokenizer was saved using *save_pretrained('./test/saved_model/')*)
>>> # tokenizer = AutoTokenizer.from_pretrained("./test/bert_saved_model/")
>>> # Download vocabulary from huggingface.co and define model-specific arguments
>>> tokenizer = AutoTokenizer.from_pretrained("FacebookAI/roberta-base", add_prefix_space=True)
```"""
use_auth_token = kwargs.pop("use_auth_token", None)
if use_auth_token is not None:
warnings.warn(
"The `use_auth_token` argument is deprecated and will be removed in v5 of Transformers. Please use `token` instead.",
FutureWarning,
)
if kwargs.get("token") is not None:
raise ValueError(
"`token` and `use_auth_token` are both specified. Please set only the argument `token`."
)
kwargs["token"] = use_auth_token
config = kwargs.pop("config", None)
kwargs["_from_auto"] = True
use_fast = kwargs.pop("use_fast", True)
tokenizer_type = kwargs.pop("tokenizer_type", None)
trust_remote_code = kwargs.pop("trust_remote_code", None)
gguf_file = kwargs.get("gguf_file")
# First, let's see whether the tokenizer_type is passed so that we can leverage it
if tokenizer_type is not None:
tokenizer_class = None
tokenizer_class_tuple = TOKENIZER_MAPPING_NAMES.get(tokenizer_type, None)
if tokenizer_class_tuple is None:
raise ValueError(
f"Passed `tokenizer_type` {tokenizer_type} does not exist. `tokenizer_type` should be one of "
f"{', '.join(c for c in TOKENIZER_MAPPING_NAMES)}."
)
tokenizer_class_name, tokenizer_fast_class_name = tokenizer_class_tuple
if use_fast:
if tokenizer_fast_class_name is not None:
tokenizer_class = tokenizer_class_from_name(tokenizer_fast_class_name)
else:
logger.warning(
"`use_fast` is set to `True` but the tokenizer class does not have a fast version. "
" Falling back to the slow version."
)
if tokenizer_class is None:
tokenizer_class = tokenizer_class_from_name(tokenizer_class_name)
if tokenizer_class is None:
raise ValueError(f"Tokenizer class {tokenizer_class_name} is not currently imported.")
return tokenizer_class.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
# Next, let's try to use the tokenizer_config file to get the tokenizer class.
tokenizer_config = get_tokenizer_config(pretrained_model_name_or_path, **kwargs)
if "_commit_hash" in tokenizer_config:
kwargs["_commit_hash"] = tokenizer_config["_commit_hash"]
config_tokenizer_class = tokenizer_config.get("tokenizer_class")
tokenizer_auto_map = None
if "auto_map" in tokenizer_config:
if isinstance(tokenizer_config["auto_map"], (tuple, list)):
# Legacy format for dynamic tokenizers
tokenizer_auto_map = tokenizer_config["auto_map"]
else:
tokenizer_auto_map = tokenizer_config["auto_map"].get("AutoTokenizer", None)
# If that did not work, let's try to use the config.
if config_tokenizer_class is None:
if not isinstance(config, PretrainedConfig):
if gguf_file:
gguf_path = cached_file(pretrained_model_name_or_path, gguf_file, **kwargs)
config_dict = load_gguf_checkpoint(gguf_path, return_tensors=False)["config"]
config = AutoConfig.for_model(**config_dict)
else:
config = AutoConfig.from_pretrained(
pretrained_model_name_or_path, trust_remote_code=trust_remote_code, **kwargs
)
config_tokenizer_class = config.tokenizer_class
if hasattr(config, "auto_map") and "AutoTokenizer" in config.auto_map:
tokenizer_auto_map = config.auto_map["AutoTokenizer"]
has_remote_code = tokenizer_auto_map is not None
has_local_code = type(config) in TOKENIZER_MAPPING or (
config_tokenizer_class is not None
and (
tokenizer_class_from_name(config_tokenizer_class) is not None
or tokenizer_class_from_name(config_tokenizer_class + "Fast") is not None
)
)
if has_remote_code:
if use_fast and tokenizer_auto_map[1] is not None:
class_ref = tokenizer_auto_map[1]
else:
class_ref = tokenizer_auto_map[0]
if "--" in class_ref:
upstream_repo = class_ref.split("--")[0]
else:
upstream_repo = None
trust_remote_code = resolve_trust_remote_code(
trust_remote_code, pretrained_model_name_or_path, has_local_code, has_remote_code, upstream_repo
)
if has_remote_code and trust_remote_code:
tokenizer_class = get_class_from_dynamic_module(class_ref, pretrained_model_name_or_path, **kwargs)
_ = kwargs.pop("code_revision", None)
tokenizer_class.register_for_auto_class()
return tokenizer_class.from_pretrained(
pretrained_model_name_or_path, *inputs, trust_remote_code=trust_remote_code, **kwargs
)
elif config_tokenizer_class is not None:
tokenizer_class = None
if use_fast and not config_tokenizer_class.endswith("Fast"):
tokenizer_class_candidate = f"{config_tokenizer_class}Fast"
tokenizer_class = tokenizer_class_from_name(tokenizer_class_candidate)
if tokenizer_class is None:
tokenizer_class_candidate = config_tokenizer_class
tokenizer_class = tokenizer_class_from_name(tokenizer_class_candidate)
if tokenizer_class is None:
raise ValueError(
f"Tokenizer class {tokenizer_class_candidate} does not exist or is not currently imported."
)
return tokenizer_class.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
# Otherwise we have to be creative.
# if model is an encoder decoder, the encoder tokenizer class is used by default
if isinstance(config, EncoderDecoderConfig):
if type(config.decoder) is not type(config.encoder): # noqa: E721
logger.warning(
f"The encoder model config class: {config.encoder.__class__} is different from the decoder model "
f"config class: {config.decoder.__class__}. It is not recommended to use the "
"`AutoTokenizer.from_pretrained()` method in this case. Please use the encoder and decoder "
"specific tokenizer classes."
)
config = config.encoder
model_type = config_class_to_model_type(type(config).__name__)
if model_type is not None:
tokenizer_class_py, tokenizer_class_fast = TOKENIZER_MAPPING[type(config)]
if tokenizer_class_fast and (use_fast or tokenizer_class_py is None):
return tokenizer_class_fast.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
else:
if tokenizer_class_py is not None:
return tokenizer_class_py.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
else:
raise ValueError(
"This tokenizer cannot be instantiated. Please make sure you have `sentencepiece` installed "
"in order to use this tokenizer."
)
raise ValueError(
f"Unrecognized configuration class {config.__class__} to build an AutoTokenizer.\n"
f"Model type should be one of {', '.join(c.__name__ for c in TOKENIZER_MAPPING)}."
)
@staticmethod
def register(config_class, slow_tokenizer_class=None, fast_tokenizer_class=None, exist_ok=False):
"""
Register a new tokenizer in this mapping.
Args:
config_class ([`PretrainedConfig`]):
The configuration corresponding to the model to register.
slow_tokenizer_class ([`PretrainedTokenizer`], *optional*):
The slow tokenizer to register.
fast_tokenizer_class ([`PretrainedTokenizerFast`], *optional*):
The fast tokenizer to register.
"""
if slow_tokenizer_class is None and fast_tokenizer_class is None:
raise ValueError("You need to pass either a `slow_tokenizer_class` or a `fast_tokenizer_class")
if slow_tokenizer_class is not None and issubclass(slow_tokenizer_class, PreTrainedTokenizerFast):
raise ValueError("You passed a fast tokenizer in the `slow_tokenizer_class`.")
if fast_tokenizer_class is not None and issubclass(fast_tokenizer_class, PreTrainedTokenizer):
raise ValueError("You passed a slow tokenizer in the `fast_tokenizer_class`.")
if (
slow_tokenizer_class is not None
and fast_tokenizer_class is not None
and issubclass(fast_tokenizer_class, PreTrainedTokenizerFast)
and fast_tokenizer_class.slow_tokenizer_class != slow_tokenizer_class
):
raise ValueError(
"The fast tokenizer class you are passing has a `slow_tokenizer_class` attribute that is not "
"consistent with the slow tokenizer class you passed (fast tokenizer has "
f"{fast_tokenizer_class.slow_tokenizer_class} and you passed {slow_tokenizer_class}. Fix one of those "
"so they match!"
)
# Avoid resetting a set slow/fast tokenizer if we are passing just the other ones.
if config_class in TOKENIZER_MAPPING._extra_content:
existing_slow, existing_fast = TOKENIZER_MAPPING[config_class]
if slow_tokenizer_class is None:
slow_tokenizer_class = existing_slow
if fast_tokenizer_class is None:
fast_tokenizer_class = existing_fast
TOKENIZER_MAPPING.register(config_class, (slow_tokenizer_class, fast_tokenizer_class), exist_ok=exist_ok)
__all__ = ["TOKENIZER_MAPPING", "AutoTokenizer"]
| transformers/src/transformers/models/auto/tokenization_auto.py/0 | {
"file_path": "transformers/src/transformers/models/auto/tokenization_auto.py",
"repo_id": "transformers",
"token_count": 25878
} | 430 |
# coding=utf-8
# Copyright 2023 The Suno AI Authors and The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""BARK model configuration"""
from typing import Optional
from ...configuration_utils import PretrainedConfig
from ...utils import add_start_docstrings, logging
from ..auto import CONFIG_MAPPING, AutoConfig
logger = logging.get_logger(__name__)
BARK_SUBMODELCONFIG_START_DOCSTRING = """
This is the configuration class to store the configuration of a [`{model}`]. It is used to instantiate the model
according to the specified arguments, defining the model architecture. Instantiating a configuration with the
defaults will yield a similar configuration to that of the Bark [suno/bark](https://huggingface.co/suno/bark)
architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
block_size (`int`, *optional*, defaults to 1024):
The maximum sequence length that this model might ever be used with. Typically set this to something large
just in case (e.g., 512 or 1024 or 2048).
input_vocab_size (`int`, *optional*, defaults to 10_048):
Vocabulary size of a Bark sub-model. Defines the number of different tokens that can be represented by the
`inputs_ids` passed when calling [`{model}`]. Defaults to 10_048 but should be carefully thought with
regards to the chosen sub-model.
output_vocab_size (`int`, *optional*, defaults to 10_048):
Output vocabulary size of a Bark sub-model. Defines the number of different tokens that can be represented
by the: `output_ids` when passing forward a [`{model}`]. Defaults to 10_048 but should be carefully thought
with regards to the chosen sub-model.
num_layers (`int`, *optional*, defaults to 12):
Number of hidden layers in the given sub-model.
num_heads (`int`, *optional*, defaults to 12):
Number of attention heads for each attention layer in the Transformer architecture.
hidden_size (`int`, *optional*, defaults to 768):
Dimensionality of the "intermediate" (often named feed-forward) layer in the architecture.
dropout (`float`, *optional*, defaults to 0.0):
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
bias (`bool`, *optional*, defaults to `True`):
Whether or not to use bias in the linear layers and layer norm layers.
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
use_cache (`bool`, *optional*, defaults to `True`):
Whether or not the model should return the last key/values attentions (not used by all models).
"""
class BarkSubModelConfig(PretrainedConfig):
keys_to_ignore_at_inference = ["past_key_values"]
attribute_map = {
"num_attention_heads": "num_heads",
"num_hidden_layers": "num_layers",
"vocab_size": "input_vocab_size",
"window_size": "block_size",
}
def __init__(
self,
block_size=1024,
input_vocab_size=10_048,
output_vocab_size=10_048,
num_layers=12,
num_heads=12,
hidden_size=768,
dropout=0.0,
bias=True, # True: bias in Linears and LayerNorms, like GPT-2. False: a bit better and faster
initializer_range=0.02,
use_cache=True,
**kwargs,
):
self.block_size = block_size
self.input_vocab_size = input_vocab_size
self.output_vocab_size = output_vocab_size
self.num_layers = num_layers
self.num_heads = num_heads
self.hidden_size = hidden_size
self.dropout = dropout
self.bias = bias
self.use_cache = use_cache
self.initializer_range = initializer_range
super().__init__(**kwargs)
@add_start_docstrings(
BARK_SUBMODELCONFIG_START_DOCSTRING.format(config="BarkSemanticConfig", model="BarkSemanticModel"),
"""
Example:
```python
>>> from transformers import BarkSemanticConfig, BarkSemanticModel
>>> # Initializing a Bark sub-module style configuration
>>> configuration = BarkSemanticConfig()
>>> # Initializing a model (with random weights) from the suno/bark style configuration
>>> model = BarkSemanticModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```""",
)
class BarkSemanticConfig(BarkSubModelConfig):
model_type = "semantic"
base_config_key = "semantic_config"
@add_start_docstrings(
BARK_SUBMODELCONFIG_START_DOCSTRING.format(config="BarkCoarseConfig", model="BarkCoarseModel"),
"""
Example:
```python
>>> from transformers import BarkCoarseConfig, BarkCoarseModel
>>> # Initializing a Bark sub-module style configuration
>>> configuration = BarkCoarseConfig()
>>> # Initializing a model (with random weights) from the suno/bark style configuration
>>> model = BarkCoarseModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```""",
)
class BarkCoarseConfig(BarkSubModelConfig):
model_type = "coarse_acoustics"
base_config_key = "coarse_acoustics_config"
@add_start_docstrings(
BARK_SUBMODELCONFIG_START_DOCSTRING.format(config="BarkFineConfig", model="BarkFineModel"),
"""
n_codes_total (`int`, *optional*, defaults to 8):
The total number of audio codebooks predicted. Used in the fine acoustics sub-model.
n_codes_given (`int`, *optional*, defaults to 1):
The number of audio codebooks predicted in the coarse acoustics sub-model. Used in the acoustics
sub-models.
Example:
```python
>>> from transformers import BarkFineConfig, BarkFineModel
>>> # Initializing a Bark sub-module style configuration
>>> configuration = BarkFineConfig()
>>> # Initializing a model (with random weights) from the suno/bark style configuration
>>> model = BarkFineModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```""",
)
class BarkFineConfig(BarkSubModelConfig):
model_type = "fine_acoustics"
base_config_key = "fine_acoustics_config"
def __init__(self, tie_word_embeddings=True, n_codes_total=8, n_codes_given=1, **kwargs):
self.n_codes_total = n_codes_total
self.n_codes_given = n_codes_given
super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
class BarkConfig(PretrainedConfig):
"""
This is the configuration class to store the configuration of a [`BarkModel`]. It is used to instantiate a Bark
model according to the specified sub-models configurations, defining the model architecture.
Instantiating a configuration with the defaults will yield a similar configuration to that of the Bark
[suno/bark](https://huggingface.co/suno/bark) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
semantic_config ([`BarkSemanticConfig`], *optional*):
Configuration of the underlying semantic sub-model.
coarse_acoustics_config ([`BarkCoarseConfig`], *optional*):
Configuration of the underlying coarse acoustics sub-model.
fine_acoustics_config ([`BarkFineConfig`], *optional*):
Configuration of the underlying fine acoustics sub-model.
codec_config ([`AutoConfig`], *optional*):
Configuration of the underlying codec sub-model.
Example:
```python
>>> from transformers import (
... BarkSemanticConfig,
... BarkCoarseConfig,
... BarkFineConfig,
... BarkModel,
... BarkConfig,
... AutoConfig,
... )
>>> # Initializing Bark sub-modules configurations.
>>> semantic_config = BarkSemanticConfig()
>>> coarse_acoustics_config = BarkCoarseConfig()
>>> fine_acoustics_config = BarkFineConfig()
>>> codec_config = AutoConfig.from_pretrained("facebook/encodec_24khz")
>>> # Initializing a Bark module style configuration
>>> configuration = BarkConfig.from_sub_model_configs(
... semantic_config, coarse_acoustics_config, fine_acoustics_config, codec_config
... )
>>> # Initializing a model (with random weights)
>>> model = BarkModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```
"""
model_type = "bark"
sub_configs = {
"semantic_config": BarkSemanticConfig,
"coarse_acoustics_config": BarkCoarseConfig,
"fine_acoustics_config": BarkFineConfig,
"codec_config": AutoConfig,
}
def __init__(
self,
semantic_config: Optional[dict] = None,
coarse_acoustics_config: Optional[dict] = None,
fine_acoustics_config: Optional[dict] = None,
codec_config: Optional[dict] = None,
initializer_range=0.02,
**kwargs,
):
if semantic_config is None:
semantic_config = {}
logger.info("semantic_config is None. initializing the semantic model with default values.")
if coarse_acoustics_config is None:
coarse_acoustics_config = {}
logger.info("coarse_acoustics_config is None. initializing the coarse model with default values.")
if fine_acoustics_config is None:
fine_acoustics_config = {}
logger.info("fine_acoustics_config is None. initializing the fine model with default values.")
if codec_config is None:
codec_config = {}
logger.info("codec_config is None. initializing the codec model with default values.")
self.semantic_config = BarkSemanticConfig(**semantic_config)
self.coarse_acoustics_config = BarkCoarseConfig(**coarse_acoustics_config)
self.fine_acoustics_config = BarkFineConfig(**fine_acoustics_config)
codec_model_type = codec_config.get("model_type", "encodec")
self.codec_config = CONFIG_MAPPING[codec_model_type](**codec_config)
self.initializer_range = initializer_range
super().__init__(**kwargs)
@classmethod
def from_sub_model_configs(
cls,
semantic_config: BarkSemanticConfig,
coarse_acoustics_config: BarkCoarseConfig,
fine_acoustics_config: BarkFineConfig,
codec_config: PretrainedConfig,
**kwargs,
):
r"""
Instantiate a [`BarkConfig`] (or a derived class) from bark sub-models configuration.
Returns:
[`BarkConfig`]: An instance of a configuration object
"""
return cls(
semantic_config=semantic_config.to_dict(),
coarse_acoustics_config=coarse_acoustics_config.to_dict(),
fine_acoustics_config=fine_acoustics_config.to_dict(),
codec_config=codec_config.to_dict(),
**kwargs,
)
__all__ = ["BarkCoarseConfig", "BarkConfig", "BarkFineConfig", "BarkSemanticConfig"]
| transformers/src/transformers/models/bark/configuration_bark.py/0 | {
"file_path": "transformers/src/transformers/models/bark/configuration_bark.py",
"repo_id": "transformers",
"token_count": 4329
} | 431 |
# coding=utf-8
# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""BiT model configuration"""
from ...configuration_utils import PretrainedConfig
from ...utils import logging
from ...utils.backbone_utils import BackboneConfigMixin, get_aligned_output_features_output_indices
logger = logging.get_logger(__name__)
class BitConfig(BackboneConfigMixin, PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`BitModel`]. It is used to instantiate an BiT
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
defaults will yield a similar configuration to that of the BiT
[google/bit-50](https://huggingface.co/google/bit-50) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
num_channels (`int`, *optional*, defaults to 3):
The number of input channels.
embedding_size (`int`, *optional*, defaults to 64):
Dimensionality (hidden size) for the embedding layer.
hidden_sizes (`list[int]`, *optional*, defaults to `[256, 512, 1024, 2048]`):
Dimensionality (hidden size) at each stage.
depths (`list[int]`, *optional*, defaults to `[3, 4, 6, 3]`):
Depth (number of layers) for each stage.
layer_type (`str`, *optional*, defaults to `"preactivation"`):
The layer to use, it can be either `"preactivation"` or `"bottleneck"`.
hidden_act (`str`, *optional*, defaults to `"relu"`):
The non-linear activation function in each block. If string, `"gelu"`, `"relu"`, `"selu"` and `"gelu_new"`
are supported.
global_padding (`str`, *optional*):
Padding strategy to use for the convolutional layers. Can be either `"valid"`, `"same"`, or `None`.
num_groups (`int`, *optional*, defaults to 32):
Number of groups used for the `BitGroupNormActivation` layers.
drop_path_rate (`float`, *optional*, defaults to 0.0):
The drop path rate for the stochastic depth.
embedding_dynamic_padding (`bool`, *optional*, defaults to `False`):
Whether or not to make use of dynamic padding for the embedding layer.
output_stride (`int`, *optional*, defaults to 32):
The output stride of the model.
width_factor (`int`, *optional*, defaults to 1):
The width factor for the model.
out_features (`list[str]`, *optional*):
If used as backbone, list of features to output. Can be any of `"stem"`, `"stage1"`, `"stage2"`, etc.
(depending on how many stages the model has). If unset and `out_indices` is set, will default to the
corresponding stages. If unset and `out_indices` is unset, will default to the last stage. Must be in the
same order as defined in the `stage_names` attribute.
out_indices (`list[int]`, *optional*):
If used as backbone, list of indices of features to output. Can be any of 0, 1, 2, etc. (depending on how
many stages the model has). If unset and `out_features` is set, will default to the corresponding stages.
If unset and `out_features` is unset, will default to the last stage. Must be in the
same order as defined in the `stage_names` attribute.
Example:
```python
>>> from transformers import BitConfig, BitModel
>>> # Initializing a BiT bit-50 style configuration
>>> configuration = BitConfig()
>>> # Initializing a model (with random weights) from the bit-50 style configuration
>>> model = BitModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```
"""
model_type = "bit"
layer_types = ["preactivation", "bottleneck"]
supported_padding = ["SAME", "VALID"]
def __init__(
self,
num_channels=3,
embedding_size=64,
hidden_sizes=[256, 512, 1024, 2048],
depths=[3, 4, 6, 3],
layer_type="preactivation",
hidden_act="relu",
global_padding=None,
num_groups=32,
drop_path_rate=0.0,
embedding_dynamic_padding=False,
output_stride=32,
width_factor=1,
out_features=None,
out_indices=None,
**kwargs,
):
super().__init__(**kwargs)
if layer_type not in self.layer_types:
raise ValueError(f"layer_type={layer_type} is not one of {','.join(self.layer_types)}")
if global_padding is not None:
if global_padding.upper() in self.supported_padding:
global_padding = global_padding.upper()
else:
raise ValueError(f"Padding strategy {global_padding} not supported")
self.num_channels = num_channels
self.embedding_size = embedding_size
self.hidden_sizes = hidden_sizes
self.depths = depths
self.layer_type = layer_type
self.hidden_act = hidden_act
self.global_padding = global_padding
self.num_groups = num_groups
self.drop_path_rate = drop_path_rate
self.embedding_dynamic_padding = embedding_dynamic_padding
self.output_stride = output_stride
self.width_factor = width_factor
self.stage_names = ["stem"] + [f"stage{idx}" for idx in range(1, len(depths) + 1)]
self._out_features, self._out_indices = get_aligned_output_features_output_indices(
out_features=out_features, out_indices=out_indices, stage_names=self.stage_names
)
__all__ = ["BitConfig"]
| transformers/src/transformers/models/bit/configuration_bit.py/0 | {
"file_path": "transformers/src/transformers/models/bit/configuration_bit.py",
"repo_id": "transformers",
"token_count": 2362
} | 432 |
# coding=utf-8
# Copyright 2023 The Salesforce Team Authors and The HuggingFace Team. All rights reserved.
#
# Licensed under the BSD-3-clause license (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://opensource.org/licenses/BSD-3-Clause
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import annotations
import math
import tensorflow as tf
from ...modeling_tf_outputs import (
TFBaseModelOutputWithPastAndCrossAttentions,
TFBaseModelOutputWithPoolingAndCrossAttentions,
TFCausalLMOutputWithCrossAttentions,
)
from ...modeling_tf_utils import (
TFModelInputType,
TFPreTrainedModel,
get_initializer,
get_tf_activation,
keras,
keras_serializable,
shape_list,
unpack_inputs,
)
from ...tf_utils import check_embeddings_within_bounds, invert_attention_mask, stable_softmax
from ...utils import add_start_docstrings_to_model_forward, logging
from .configuration_blip import BlipTextConfig
logger = logging.get_logger(__name__)
BLIP_TEXT_INPUTS_DOCSTRING = r"""
Args:
input_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
it.
Indices can be obtained using [`AutoProcessor`]. See [`BlipProcessor.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
position_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
config.max_position_embeddings - 1]`.
[What are position IDs?](../glossary#position-ids)
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
# Adapted from https://github.com/salesforce/BLIP/blob/main/models/med.py#L52
class TFBlipTextEmbeddings(keras.layers.Layer):
"""Construct the embeddings from word and position embeddings."""
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.word_embeddings = keras.layers.Embedding(
config.vocab_size,
config.hidden_size,
embeddings_initializer=get_initializer(config.initializer_range),
name="word_embeddings",
)
self.position_embeddings = keras.layers.Embedding(
config.max_position_embeddings,
config.hidden_size,
embeddings_initializer=get_initializer(config.initializer_range),
name="position_embeddings",
)
# self.LayerNorm is not snake-cased to stick with PyTorch model variable name and be able to load
# any TensorFlow checkpoint file
self.LayerNorm = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.dropout = keras.layers.Dropout(config.hidden_dropout_prob, name="dropout")
self.position_ids = tf.expand_dims(tf.range(config.max_position_embeddings), 0)
self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
self.config = config
def call(self, input_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0, training=None):
if input_ids is not None:
input_shape = tf.shape(input_ids)
else:
input_shape = tf.shape(inputs_embeds)[:-1]
seq_length = input_shape[1]
if position_ids is None:
position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
if inputs_embeds is None:
check_embeddings_within_bounds(input_ids, self.config.vocab_size)
inputs_embeds = self.word_embeddings(input_ids)
embeddings = inputs_embeds
if self.position_embedding_type == "absolute":
position_embeddings = self.position_embeddings(position_ids)
embeddings += position_embeddings
embeddings = self.LayerNorm(embeddings)
embeddings = self.dropout(embeddings, training=training)
return embeddings
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "word_embeddings", None) is not None:
with tf.name_scope(self.word_embeddings.name):
self.word_embeddings.build(None)
if getattr(self, "position_embeddings", None) is not None:
with tf.name_scope(self.position_embeddings.name):
self.position_embeddings.build(None)
if getattr(self, "LayerNorm", None) is not None:
with tf.name_scope(self.LayerNorm.name):
self.LayerNorm.build([None, None, self.config.hidden_size])
if getattr(self, "dropout", None) is not None:
with tf.name_scope(self.dropout.name):
self.dropout.build(None)
# Adapted from https://github.com/salesforce/BLIP/blob/main/models/med.py#L97
class TFBlipTextSelfAttention(keras.layers.Layer):
def __init__(self, config, is_cross_attention, **kwargs):
super().__init__(**kwargs)
self.config = config
if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
raise ValueError(
"The hidden size (%d) is not a multiple of the number of attention heads (%d)"
% (config.hidden_size, config.num_attention_heads)
)
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
self.all_head_size = self.num_attention_heads * self.attention_head_size
self.query = keras.layers.Dense(
self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="query"
)
self.key = keras.layers.Dense(
self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="key"
)
self.value = keras.layers.Dense(
self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="value"
)
self.dropout = keras.layers.Dropout(config.attention_probs_dropout_prob)
self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
self.max_position_embeddings = config.max_position_embeddings
self.distance_embedding = keras.layers.Embedding(
2 * config.max_position_embeddings - 1, self.attention_head_size
)
self.is_cross_attention = is_cross_attention
def transpose_for_scores(self, x):
new_x_shape = tf.concat(
[tf.shape(x)[:-1], tf.constant([self.num_attention_heads, self.attention_head_size], dtype=tf.int32)],
axis=0,
)
x = tf.reshape(x, new_x_shape)
return tf.transpose(x, perm=(0, 2, 1, 3))
def call(
self,
hidden_states,
attention_mask=None,
head_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
past_key_value=None,
output_attentions=False,
training=None,
):
mixed_query_layer = self.query(hidden_states)
# If this is instantiated as a cross-attention module, the keys
# and values come from an encoder; the attention mask needs to be
# such that the encoder's padding tokens are not attended to.
is_cross_attention = encoder_hidden_states is not None
if is_cross_attention:
key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
attention_mask = encoder_attention_mask
elif past_key_value is not None:
key_layer = self.transpose_for_scores(self.key(hidden_states))
value_layer = self.transpose_for_scores(self.value(hidden_states))
key_layer = tf.concat([past_key_value[0], key_layer], axis=2)
value_layer = tf.concat([past_key_value[1], value_layer], axis=2)
else:
key_layer = self.transpose_for_scores(self.key(hidden_states))
value_layer = self.transpose_for_scores(self.value(hidden_states))
query_layer = self.transpose_for_scores(mixed_query_layer)
past_key_value = (key_layer, value_layer)
# Take the dot product between "query" and "key" to get the raw attention scores.
attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
seq_length = shape_list(hidden_states)[1]
position_ids_l = tf.expand_dims(tf.range(seq_length, dtype=tf.int64, device=hidden_states.device), 1)
position_ids_r = tf.expand_dims(tf.range(seq_length, dtype=tf.int64, device=hidden_states.device), 0)
distance = position_ids_l - position_ids_r
positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
positional_embedding = tf.cast(positional_embedding, query_layer.dtype) # fp16 compatibility
if self.position_embedding_type == "relative_key":
relative_position_scores = tf.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
attention_scores = attention_scores + relative_position_scores
elif self.position_embedding_type == "relative_key_query":
relative_position_scores_query = tf.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
relative_position_scores_key = tf.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
attention_scores = attention_scores / math.sqrt(self.attention_head_size)
if attention_mask is not None:
# Apply the attention mask is (precomputed for all layers in BlipTextModel forward() function)
attention_scores = attention_scores + tf.cast(attention_mask, attention_scores.dtype)
# Normalize the attention scores to probabilities.
attention_probs = stable_softmax(attention_scores, axis=-1)
# This is actually dropping out entire tokens to attend to, which might
# seem a bit unusual, but is taken from the original Transformer paper.
attention_probs_dropped = self.dropout(attention_probs, training=training)
# Mask heads if we want to
if head_mask is not None:
attention_probs_dropped = attention_probs_dropped * head_mask
context_layer = attention_probs_dropped @ value_layer
context_layer = tf.transpose(context_layer, perm=(0, 2, 1, 3))
new_context_layer_shape = shape_list(context_layer)[:-2] + [self.all_head_size]
context_layer = tf.reshape(context_layer, new_context_layer_shape)
outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
outputs = outputs + (past_key_value,)
return outputs
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "query", None) is not None:
with tf.name_scope(self.query.name):
self.query.build([None, None, self.config.hidden_size])
if self.is_cross_attention:
if getattr(self, "key", None) is not None:
with tf.name_scope(self.key.name):
self.key.build([None, None, self.config.encoder_hidden_size])
if getattr(self, "value", None) is not None:
with tf.name_scope(self.value.name):
self.value.build([None, None, self.config.encoder_hidden_size])
else:
if getattr(self, "key", None) is not None:
with tf.name_scope(self.key.name):
self.key.build([None, None, self.config.hidden_size])
if getattr(self, "value", None) is not None:
with tf.name_scope(self.value.name):
self.value.build([None, None, self.config.hidden_size])
class TFBlipTextSelfOutput(keras.layers.Layer):
def __init__(self, config: BlipTextConfig, **kwargs):
super().__init__(**kwargs)
self.dense = keras.layers.Dense(
units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
self.LayerNorm = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.dropout = keras.layers.Dropout(rate=config.hidden_dropout_prob)
self.config = config
def call(self, hidden_states: tf.Tensor, input_tensor: tf.Tensor, training: bool | None = None) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.dropout(inputs=hidden_states, training=training)
hidden_states = self.LayerNorm(inputs=hidden_states + input_tensor)
return hidden_states
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "dense", None) is not None:
with tf.name_scope(self.dense.name):
self.dense.build([None, None, self.config.hidden_size])
if getattr(self, "LayerNorm", None) is not None:
with tf.name_scope(self.LayerNorm.name):
self.LayerNorm.build([None, None, self.config.hidden_size])
# Adapted from https://github.com/salesforce/BLIP/blob/main/models/med.py#242
class TFBlipTextAttention(keras.layers.Layer):
def __init__(self, config, is_cross_attention=False, **kwargs):
super().__init__(**kwargs)
self.self = TFBlipTextSelfAttention(config, is_cross_attention, name="self")
# "output" is a protected attribute on TF models
self.self_output = TFBlipTextSelfOutput(config, name="output")
def call(
self,
hidden_states: tf.Tensor,
attention_mask: tf.Tensor | None = None,
head_mask: tf.Tensor | None = None,
encoder_hidden_states: tf.Tensor | None = None,
encoder_attention_mask: tf.Tensor | None = None,
past_key_value: tuple[tuple[tf.Tensor]] | None = None,
output_attentions: bool | None = False,
training: bool | None = None,
):
self_outputs = self.self(
hidden_states,
attention_mask,
head_mask,
encoder_hidden_states,
encoder_attention_mask,
past_key_value,
output_attentions,
training=training,
)
attention_output = self.self_output(self_outputs[0], hidden_states, training=training)
outputs = (attention_output,) + self_outputs[1:] # add attentions if we output them
return outputs
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "self", None) is not None:
with tf.name_scope(self.self.name):
self.self.build(None)
if getattr(self, "self_output", None) is not None:
with tf.name_scope(self.self_output.name):
self.self_output.build(None)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertIntermediate with Bert->BlipText
class TFBlipTextIntermediate(keras.layers.Layer):
def __init__(self, config: BlipTextConfig, **kwargs):
super().__init__(**kwargs)
self.dense = keras.layers.Dense(
units=config.intermediate_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
if isinstance(config.hidden_act, str):
self.intermediate_act_fn = get_tf_activation(config.hidden_act)
else:
self.intermediate_act_fn = config.hidden_act
self.config = config
def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.intermediate_act_fn(hidden_states)
return hidden_states
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "dense", None) is not None:
with tf.name_scope(self.dense.name):
self.dense.build([None, None, self.config.hidden_size])
class TFBlipTextOutput(keras.layers.Layer):
def __init__(self, config: BlipTextConfig, **kwargs):
super().__init__(**kwargs)
self.dense = keras.layers.Dense(
units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
self.LayerNorm = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.dropout = keras.layers.Dropout(rate=config.hidden_dropout_prob)
self.config = config
def call(self, hidden_states: tf.Tensor, input_tensor: tf.Tensor, training: bool = False) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.dropout(inputs=hidden_states, training=training)
hidden_states = self.LayerNorm(inputs=hidden_states + input_tensor)
return hidden_states
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "dense", None) is not None:
with tf.name_scope(self.dense.name):
self.dense.build([None, None, self.config.intermediate_size])
if getattr(self, "LayerNorm", None) is not None:
with tf.name_scope(self.LayerNorm.name):
self.LayerNorm.build([None, None, self.config.hidden_size])
class TFBlipTextLayer(keras.layers.Layer):
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.config = config
self.attention = TFBlipTextAttention(config, name="attention")
if self.config.is_decoder:
self.crossattention = TFBlipTextAttention(
config, is_cross_attention=self.config.is_decoder, name="crossattention"
)
self.intermediate = TFBlipTextIntermediate(config, name="intermediate")
self.self_output = TFBlipTextOutput(config, name="output")
def call(
self,
hidden_states,
attention_mask=None,
head_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
past_key_value=None,
output_attentions=False,
training=None,
):
# decoder uni-directional self-attention cached key/values tuple is at positions 1,2
self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
self_attention_outputs = self.attention(
hidden_states,
attention_mask,
head_mask,
output_attentions=output_attentions,
past_key_value=self_attn_past_key_value,
training=training,
)
attention_output = self_attention_outputs[0]
outputs = self_attention_outputs[1:-1]
present_key_value = self_attention_outputs[-1]
if encoder_hidden_states is not None:
cross_attention_outputs = self.crossattention(
attention_output,
attention_mask,
head_mask,
encoder_hidden_states,
encoder_attention_mask,
output_attentions=output_attentions,
training=training,
)
attention_output = cross_attention_outputs[0]
outputs = outputs + cross_attention_outputs[1:-1] # add cross attentions if we output attention weights
intermediate_output = self.intermediate(attention_output)
layer_output = self.self_output(intermediate_output, attention_output, training=training)
outputs = (layer_output,) + outputs
outputs = outputs + (present_key_value,)
return outputs
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "attention", None) is not None:
with tf.name_scope(self.attention.name):
self.attention.build(None)
if getattr(self, "intermediate", None) is not None:
with tf.name_scope(self.intermediate.name):
self.intermediate.build(None)
if getattr(self, "self_output", None) is not None:
with tf.name_scope(self.self_output.name):
self.self_output.build(None)
if getattr(self, "crossattention", None) is not None:
with tf.name_scope(self.crossattention.name):
self.crossattention.build(None)
# Adapted from https://github.com/salesforce/BLIP/blob/main/models/med.py#L386
@keras_serializable
class TFBlipTextEncoder(keras.layers.Layer):
config_class = BlipTextConfig
def __init__(self, config, name=None, **kwargs):
super().__init__(name=name, **kwargs)
self.config = config
self.layer = [TFBlipTextLayer(config, name=f"layer_._{i}") for i in range(config.num_hidden_layers)]
@unpack_inputs
def call(
self,
hidden_states,
attention_mask=None,
head_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
past_key_values=None,
use_cache=None,
output_attentions=False,
output_hidden_states=False,
return_dict=True,
training=None,
):
all_hidden_states = () if output_hidden_states else None
all_self_attentions = () if output_attentions else None
all_cross_attentions = () if output_attentions and self.config.is_decoder else None
next_decoder_cache = () if use_cache else None
for i in range(self.config.num_hidden_layers):
layer_module = self.layer[i]
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
layer_head_mask = head_mask[i] if head_mask is not None else None
past_key_value = past_key_values[i] if past_key_values is not None else None
layer_outputs = layer_module(
hidden_states,
attention_mask,
layer_head_mask,
encoder_hidden_states,
encoder_attention_mask,
past_key_value,
output_attentions,
training=training,
)
hidden_states = layer_outputs[0]
if use_cache:
next_decoder_cache += (layer_outputs[-1],)
if output_attentions:
all_self_attentions = all_self_attentions + (layer_outputs[1],)
all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if not return_dict:
return tuple(
v
for v in [
hidden_states,
next_decoder_cache,
all_hidden_states,
all_self_attentions,
all_cross_attentions,
]
if v is not None
)
return TFBaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=hidden_states,
past_key_values=next_decoder_cache,
hidden_states=all_hidden_states,
attentions=all_self_attentions,
cross_attentions=all_cross_attentions,
)
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "layer", None) is not None:
for layer in self.layer:
with tf.name_scope(layer.name):
layer.build(None)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertPooler with Bert->BlipText
class TFBlipTextPooler(keras.layers.Layer):
def __init__(self, config: BlipTextConfig, **kwargs):
super().__init__(**kwargs)
self.dense = keras.layers.Dense(
units=config.hidden_size,
kernel_initializer=get_initializer(config.initializer_range),
activation="tanh",
name="dense",
)
self.config = config
def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
# We "pool" the model by simply taking the hidden state corresponding
# to the first token.
first_token_tensor = hidden_states[:, 0]
pooled_output = self.dense(inputs=first_token_tensor)
return pooled_output
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "dense", None) is not None:
with tf.name_scope(self.dense.name):
self.dense.build([None, None, self.config.hidden_size])
# Copied from transformers.models.bert.modeling_tf_bert.TFBertPredictionHeadTransform with Bert->BlipText
class TFBlipTextPredictionHeadTransform(keras.layers.Layer):
def __init__(self, config: BlipTextConfig, **kwargs):
super().__init__(**kwargs)
self.dense = keras.layers.Dense(
units=config.hidden_size,
kernel_initializer=get_initializer(config.initializer_range),
name="dense",
)
if isinstance(config.hidden_act, str):
self.transform_act_fn = get_tf_activation(config.hidden_act)
else:
self.transform_act_fn = config.hidden_act
self.LayerNorm = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.config = config
def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.transform_act_fn(hidden_states)
hidden_states = self.LayerNorm(inputs=hidden_states)
return hidden_states
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "dense", None) is not None:
with tf.name_scope(self.dense.name):
self.dense.build([None, None, self.config.hidden_size])
if getattr(self, "LayerNorm", None) is not None:
with tf.name_scope(self.LayerNorm.name):
self.LayerNorm.build([None, None, self.config.hidden_size])
class TFBlipTextLMPredictionHead(keras.layers.Layer):
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.transform = TFBlipTextPredictionHeadTransform(config, name="transform")
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
self.decoder = keras.layers.Dense(
config.vocab_size,
kernel_initializer=get_initializer(config.initializer_range),
name="decoder",
use_bias=False,
)
self.config = config
def build(self, input_shape=None):
self.bias = self.add_weight(name="bias", shape=(self.config.vocab_size,), initializer="zeros", trainable=True)
if self.built:
return
self.built = True
if getattr(self, "transform", None) is not None:
with tf.name_scope(self.transform.name):
self.transform.build(None)
if getattr(self, "decoder", None) is not None:
with tf.name_scope(self.decoder.name):
self.decoder.build([None, None, self.config.hidden_size])
def call(self, hidden_states):
hidden_states = self.transform(hidden_states)
hidden_states = self.decoder(hidden_states) + self.bias
return hidden_states
class TFBlipTextOnlyMLMHead(keras.layers.Layer):
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.predictions = TFBlipTextLMPredictionHead(config, name="predictions")
def call(self, sequence_output: tf.Tensor) -> tf.Tensor:
prediction_scores = self.predictions(sequence_output)
return prediction_scores
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "predictions", None) is not None:
with tf.name_scope(self.predictions.name):
self.predictions.build(None)
# Adapted from https://github.com/salesforce/BLIP/blob/main/models/med.py#L548
class TFBlipTextPreTrainedModel(TFPreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = BlipTextConfig
base_model_prefix = "bert"
_keys_to_ignore_on_load_missing = [r"position_ids"]
# Adapted from https://github.com/salesforce/BLIP/blob/3a29b7410476bf5f2ba0955827390eb6ea1f4f9d/models/med.py#L571
class TFBlipTextModel(TFBlipTextPreTrainedModel):
"""
The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
cross-attention is added between the self-attention layers, following the architecture described in [Attention is
all you need](https://huggingface.co/papers/1706.03762) by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,
Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin. argument and `is_decoder` set to `True`; an
`encoder_hidden_states` is then expected as an input to the forward pass.
"""
def __init__(self, config, add_pooling_layer=True, name=None, **kwargs):
super().__init__(config, name=name, **kwargs)
self.config = config
self.embeddings = TFBlipTextEmbeddings(config, name="embeddings")
self.encoder = TFBlipTextEncoder(config, name="encoder")
self.pooler = TFBlipTextPooler(config, name="pooler") if add_pooling_layer else None
def get_input_embeddings(self):
return self.embeddings.word_embeddings
def set_input_embeddings(self, value):
self.embeddings.word_embeddings = value
@tf.function
def get_extended_attention_mask(
self, attention_mask: tf.Tensor, input_shape: tuple[int], is_decoder: bool
) -> tf.Tensor:
"""
Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
Arguments:
attention_mask (`tf.Tensor`):
Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
input_shape (`tuple[int]`):
The shape of the input to the model.
is_decoder (`bool`):
Whether the model is used as a decoder.
Returns:
`tf.Tensor` The extended attention mask, with the same dtype as `attention_mask.dtype`.
"""
# We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
# ourselves in which case we just need to make it broadcastable to all heads.
if not isinstance(attention_mask, tf.Tensor):
attention_mask = tf.convert_to_tensor(attention_mask) # Catches NumPy inputs that haven't been cast yet
if attention_mask.shape.rank == 3:
extended_attention_mask = attention_mask[:, None, :, :]
elif attention_mask.shape.rank == 2:
# Provided a padding mask of dimensions [batch_size, seq_length]
# - if the model is a decoder, apply a causal mask in addition to the padding mask
# - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
if is_decoder:
batch_size, seq_length = input_shape
seq_ids = tf.range(seq_length, dtype=attention_mask.dtype)
causal_mask = tf.broadcast_to(seq_ids, (batch_size, seq_length, seq_length)) <= seq_ids[None, :, None]
# in case past_key_values are used we need to add a prefix ones mask to the causal mask
if shape_list(causal_mask)[1] < shape_list(attention_mask)[1]:
prefix_seq_len = tf.shape(attention_mask)[1] - tf.shape(causal_mask)[1]
causal_mask = tf.concat(
[
tf.ones((batch_size, seq_length, prefix_seq_len), dtype=causal_mask.dtype),
causal_mask,
],
axis=-1,
)
extended_attention_mask = (
tf.cast(causal_mask[:, None, :, :], attention_mask.dtype) * attention_mask[:, None, None, :]
)
else:
extended_attention_mask = attention_mask[:, None, None, :]
else:
raise ValueError(
f"Wrong shape for input_ids (shape {input_shape}) or attention_mask (shape {attention_mask.shape})"
)
# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
# masked positions, this operation will create a tensor which is 0.0 for
# positions we want to attend and -10000.0 for masked positions.
# Since we are adding it to the raw scores before the softmax, this is
# effectively the same as removing these entirely.
extended_attention_mask = tf.cast(extended_attention_mask, self.dtype) # fp16 compatibility
extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
return extended_attention_mask
@add_start_docstrings_to_model_forward(BLIP_TEXT_INPUTS_DOCSTRING)
@unpack_inputs
def call(
self,
input_ids: TFModelInputType | None = None,
attention_mask: tf.Tensor | None = None,
position_ids: tf.Tensor | None = None,
head_mask: tf.Tensor | None = None,
inputs_embeds: tf.Tensor | None = None,
encoder_embeds: tf.Tensor | None = None,
encoder_hidden_states: tf.Tensor | None = None,
encoder_attention_mask: tf.Tensor | None = None,
past_key_values: tuple[tuple[tf.Tensor]] | None = None,
use_cache: bool | None = None,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
return_dict: bool | None = None,
is_decoder: bool = False,
training: bool = False,
) -> tuple[tf.Tensor] | TFBaseModelOutputWithPoolingAndCrossAttentions:
r"""
encoder_hidden_states (`tf.Tensor`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
the model is configured as a decoder.
encoder_attention_mask (`tf.Tensor`, *optional*):
Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
past_key_values (`tuple(tuple(tf.Tensor))`, *optional*):
Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
`decoder_input_ids` of shape `(batch_size, sequence_length)`.
use_cache (`bool`, *optional*):
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
`past_key_values`).
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if is_decoder:
use_cache = use_cache if use_cache is not None else self.config.use_cache
else:
use_cache = False
if input_ids is not None and inputs_embeds is not None:
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
elif input_ids is not None:
input_shape = shape_list(input_ids)
batch_size, seq_length = input_shape
elif inputs_embeds is not None:
input_shape = shape_list(inputs_embeds)[:-1]
batch_size, seq_length = input_shape
elif encoder_embeds is not None:
input_shape = shape_list(encoder_embeds)[:-1]
batch_size, seq_length = input_shape
else:
raise ValueError("You have to specify either input_ids or inputs_embeds or encoder_embeds")
# past_key_values_length
past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
if attention_mask is None:
attention_mask = tf.ones((batch_size, seq_length + past_key_values_length))
# We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
# ourselves in which case we just need to make it broadcastable to all heads.
extended_attention_mask: tf.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, is_decoder)
# If a 2D or 3D attention mask is provided for the cross-attention
# we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
if encoder_hidden_states is not None:
if isinstance(encoder_hidden_states, list):
encoder_batch_size, encoder_sequence_length, _ = shape_list(encoder_hidden_states[0])
else:
encoder_batch_size, encoder_sequence_length, _ = shape_list(encoder_hidden_states)
encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
if isinstance(encoder_attention_mask, list):
encoder_extended_attention_mask = [invert_attention_mask(mask) for mask in encoder_attention_mask]
elif encoder_attention_mask is None:
encoder_attention_mask = tf.ones(encoder_hidden_shape)
encoder_extended_attention_mask = invert_attention_mask(encoder_attention_mask)
else:
encoder_extended_attention_mask = invert_attention_mask(encoder_attention_mask)
else:
encoder_extended_attention_mask = None
# Prepare head mask if needed
# 1.0 in head_mask indicate we keep the head
# attention_probs has shape bsz x n_heads x N x N
# input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
# and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
if encoder_embeds is None:
embedding_output = self.embeddings(
input_ids=input_ids,
position_ids=position_ids,
inputs_embeds=inputs_embeds,
past_key_values_length=past_key_values_length,
)
else:
embedding_output = encoder_embeds
encoder_outputs = self.encoder(
embedding_output,
attention_mask=extended_attention_mask,
head_mask=head_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_extended_attention_mask,
past_key_values=past_key_values,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
training=training,
)
sequence_output = encoder_outputs[0]
pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
if not return_dict:
return (sequence_output, pooled_output) + encoder_outputs[1:]
return TFBaseModelOutputWithPoolingAndCrossAttentions(
last_hidden_state=sequence_output,
pooler_output=pooled_output,
past_key_values=encoder_outputs.past_key_values,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
cross_attentions=encoder_outputs.cross_attentions,
)
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "embeddings", None) is not None:
with tf.name_scope(self.embeddings.name):
self.embeddings.build(None)
if getattr(self, "encoder", None) is not None:
with tf.name_scope(self.encoder.name):
self.encoder.build(None)
if getattr(self, "pooler", None) is not None:
with tf.name_scope(self.pooler.name):
self.pooler.build(None)
# Adapted from https://github.com/salesforce/BLIP/blob/main/models/med.py#L811
class TFBlipTextLMHeadModel(TFBlipTextPreTrainedModel):
_keys_to_ignore_on_load_unexpected = [r"pooler"]
_keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
def __init__(self, config, **kwargs):
super().__init__(config, **kwargs)
self.bert = TFBlipTextModel(config, add_pooling_layer=False, name="bert")
self.cls = TFBlipTextOnlyMLMHead(config, name="cls")
self.label_smoothing = config.label_smoothing
def get_output_embeddings(self):
return self.cls.predictions.decoder
def set_output_embeddings(self, new_embeddings):
self.cls.predictions.decoder = new_embeddings
@add_start_docstrings_to_model_forward(BLIP_TEXT_INPUTS_DOCSTRING)
@unpack_inputs
def call(
self,
input_ids=None,
attention_mask=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
labels=None,
past_key_values=None,
use_cache=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
return_logits=False,
is_decoder=True,
training=None,
):
r"""
encoder_hidden_states (`tf.Tensor`, *optional*): Sequence of
hidden-states at the output of the last layer of the encoder. Used in the cross-attention if the model is
configured as a decoder.
encoder_attention_mask (`tf.Tensor`, *optional*):
Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
labels (`tf.Tensor`, *optional*):
Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
`[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are
ignored (masked), the loss is only computed for the tokens with labels n `[0, ..., config.vocab_size]`
past_key_values (`tuple(tuple(tf.Tensor))`, *optional*):
Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
`decoder_input_ids` of shape `(batch_size, sequence_length)`.
use_cache (`bool`, *optional*):
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
`past_key_values`).
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if labels is not None:
use_cache = False
outputs = self.bert(
input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
is_decoder=is_decoder,
training=training,
)
sequence_output = outputs[0]
prediction_scores = self.cls(sequence_output)
if return_logits:
return prediction_scores[:, :-1, :]
lm_loss = None
if labels is not None:
# we are doing next-token prediction; shift prediction scores and input ids by one
shifted_prediction_scores = prediction_scores[:, :-1, :]
shifted_prediction_scores = tf.reshape(shifted_prediction_scores, (-1, self.config.vocab_size))
labels = labels[:, 1:]
labels = tf.reshape(labels, (-1,))
# Keras won't give us label smoothing for sparse CE, so we de-sparsify things here
# Use relu to clamp masked labels at 0 to avoid NaN (we will be zeroing those out later anyway)
one_hot_labels = tf.one_hot(tf.nn.relu(labels), depth=self.config.vocab_size, dtype=tf.float32)
loss_fct = keras.losses.CategoricalCrossentropy(
from_logits=True, label_smoothing=self.label_smoothing, reduction="none"
)
masked_positions = tf.cast(tf.not_equal(labels, -100), dtype=tf.float32)
lm_loss = loss_fct(one_hot_labels, shifted_prediction_scores)
lm_loss *= masked_positions
lm_loss = tf.reduce_sum(lm_loss, axis=0) / tf.math.count_nonzero(masked_positions, dtype=tf.float32)
if not return_dict:
output = (prediction_scores,) + outputs[2:]
return ((lm_loss,) + output) if lm_loss is not None else output
return TFCausalLMOutputWithCrossAttentions(
loss=lm_loss,
logits=prediction_scores,
past_key_values=outputs.past_key_values,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
cross_attentions=outputs.cross_attentions,
)
def prepare_inputs_for_generation(self, input_ids, past_key_values=None, attention_mask=None, **model_kwargs):
input_shape = input_ids.shape
# if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
if attention_mask is None:
attention_mask = input_ids.new_ones(input_shape)
# cut decoder_input_ids if past_key_values is used
if past_key_values is not None:
input_ids = input_ids[:, -1:]
return {
"input_ids": input_ids,
"attention_mask": attention_mask,
"past_key_values": past_key_values,
"encoder_hidden_states": model_kwargs.get("encoder_hidden_states"),
"encoder_attention_mask": model_kwargs.get("encoder_attention_mask"),
"is_decoder": True,
}
def _reorder_cache(self, past_key_values, beam_idx):
reordered_past = ()
for layer_past in past_key_values:
reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
return reordered_past
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "bert", None) is not None:
with tf.name_scope(self.bert.name):
self.bert.build(None)
if getattr(self, "cls", None) is not None:
with tf.name_scope(self.cls.name):
self.cls.build(None)
__all__ = ["TFBlipTextLMHeadModel", "TFBlipTextModel", "TFBlipTextPreTrainedModel"]
| transformers/src/transformers/models/blip/modeling_tf_blip_text.py/0 | {
"file_path": "transformers/src/transformers/models/blip/modeling_tf_blip_text.py",
"repo_id": "transformers",
"token_count": 21887
} | 433 |
# coding=utf-8
# Copyright 2025 The Intel Labs Team Authors, The Microsoft Research Team Authors and HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Fast Image processor class for BridgeTower."""
from collections.abc import Iterable
from typing import Optional, Union
from ...image_processing_utils_fast import (
BaseImageProcessorFast,
BatchFeature,
DefaultFastImageProcessorKwargs,
ImageInput,
SizeDict,
TensorType,
Unpack,
get_max_height_width,
group_images_by_shape,
reorder_images,
)
from ...image_utils import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD, PILImageResampling
from ...utils import auto_docstring, is_torch_available, is_torchvision_available, is_torchvision_v2_available
if is_torch_available():
import torch
if is_torchvision_available():
if is_torchvision_v2_available():
from torchvision.transforms.v2 import functional as F
else:
from torchvision.transforms import functional as F
def make_pixel_mask(
image: "torch.Tensor",
output_size: tuple[int, int],
) -> "torch.Tensor":
"""
Make a pixel mask for the image, where 1 indicates a valid pixel and 0 indicates padding.
Args:
image (`np.ndarray`):
Image to make the pixel mask for.
output_size (`tuple[int, int]`):
Output size of the mask.
"""
input_height, input_width = image.shape[-2:]
batch_size = image.size(0)
mask = torch.zeros((batch_size, *output_size), dtype=torch.long)
mask[:input_height, :input_width] = 1
return mask
def get_resize_output_image_size(
input_image: "torch.Tensor",
shorter: int = 800,
longer: int = 1333,
size_divisor: int = 32,
) -> tuple[int, int]:
input_height, input_width = input_image.shape[-2:]
min_size, max_size = shorter, longer
scale = min_size / min(input_height, input_width)
if input_height < input_width:
new_height = min_size
new_width = scale * input_width
else:
new_height = scale * input_height
new_width = min_size
if max(new_height, new_width) > max_size:
scale = max_size / max(new_height, new_width)
new_height = scale * new_height
new_width = scale * new_width
new_height, new_width = int(new_height + 0.5), int(new_width + 0.5)
new_height = new_height // size_divisor * size_divisor
new_width = new_width // size_divisor * size_divisor
return new_height, new_width
class BridgeTowerFastImageProcessorKwargs(DefaultFastImageProcessorKwargs):
"""
Args:
size_divisor (`int`, *optional*, defaults to 32):
The size by which to make sure both the height and width can be divided. Only has an effect if `do_resize`
is set to `True`. Can be overridden by the `size_divisor` parameter in the `preprocess` method.
do_pad (`bool`, *optional*, defaults to `True`):
Whether to pad the image to the `(max_height, max_width)` of the images in the batch. Can be overridden by
the `do_pad` parameter in the `preprocess` method.
"""
size_divisor: Optional[int]
do_pad: Optional[bool]
@auto_docstring
class BridgeTowerImageProcessorFast(BaseImageProcessorFast):
resample = PILImageResampling.BICUBIC
image_mean = OPENAI_CLIP_MEAN
image_std = OPENAI_CLIP_STD
size = {"shortest_edge": 288}
default_to_square = False
crop_size = {"shortest_edge": 288}
do_resize = True
do_center_crop = True
do_rescale = True
do_normalize = True
do_pad = True
size_divisor = 32
valid_kwargs = BridgeTowerFastImageProcessorKwargs
model_input_names = ["pixel_values", "pixel_mask"]
def __init__(self, **kwargs: Unpack[BridgeTowerFastImageProcessorKwargs]):
super().__init__(**kwargs)
@auto_docstring
def preprocess(self, images: ImageInput, **kwargs: Unpack[BridgeTowerFastImageProcessorKwargs]) -> BatchFeature:
return super().preprocess(images, **kwargs)
def resize(
self,
image: "torch.Tensor",
size: SizeDict,
size_divisor: int = 32,
interpolation: "F.InterpolationMode" = None,
antialias: bool = True,
**kwargs,
) -> "torch.Tensor":
"""
Resize an image.
Resizes the shorter side of the image to `size["shortest_edge"]` while preserving the aspect ratio. If the
longer side is larger than the max size `(int(`size["shortest_edge"]` * 1333 / 800))`, the longer side is then
resized to the max size while preserving the aspect ratio.
Args:
image (`torch.Tensor`):
Image to resize.
size (`SizeDict`):
Dictionary in the format `{"height": int, "width": int}` specifying the size of the output image.
size_divisor (`int`, *optional*, defaults to 32):
The image is resized to a size that is a multiple of this value.
resample (`InterpolationMode`, *optional*, defaults to `InterpolationMode.BILINEAR`):
`InterpolationMode` filter to use when resizing the image e.g. `InterpolationMode.BICUBIC`.
Returns:
`torch.Tensor`: The resized image.
"""
interpolation = interpolation if interpolation is not None else F.InterpolationMode.BILINEAR
if not size.shortest_edge:
raise ValueError(f"The `size` dictionary must contain the key `shortest_edge`. Got {size.keys()}")
shorter = size.shortest_edge
longer = int(1333 / 800 * shorter)
output_height, output_width = get_resize_output_image_size(
image,
shorter=shorter,
longer=longer,
size_divisor=size_divisor,
)
return super().resize(
image=image,
size=SizeDict(height=output_height, width=output_width),
interpolation=interpolation,
antialias=antialias,
)
def center_crop(
self,
image: "torch.Tensor",
size: dict[str, int],
**kwargs,
) -> "torch.Tensor":
"""
Center crop an image to `(size["height"], size["width"])`. If the input size is smaller than `crop_size` along
any edge, the image is padded with 0's and then center cropped.
Args:
image (`torch.Tensor`):
Image to center crop.
size (`dict[str, int]`):
Size of the output image in the form `{"height": h, "width": w}`.
"""
output_size = size.shortest_edge
return F.center_crop(
image,
output_size=(output_size, output_size),
**kwargs,
)
def _pad_image(
self,
image: "torch.Tensor",
output_size: tuple[int, int],
constant_values: Union[float, Iterable[float]] = 0,
) -> "torch.Tensor":
"""
Pad an image with zeros to the given size.
"""
input_height, input_width = image.shape[-2:]
output_height, output_width = output_size
pad_bottom = output_height - input_height
pad_right = output_width - input_width
padding = (0, 0, pad_right, pad_bottom)
padded_image = F.pad(
image,
padding,
fill=constant_values,
)
return padded_image
def pad(
self,
images: list["torch.Tensor"],
constant_values: Union[float, Iterable[float]] = 0,
return_pixel_mask: bool = True,
disable_grouping: Optional[bool] = False,
) -> tuple:
"""
Pads a batch of images to the bottom and right of the image with zeros to the size of largest height and width
in the batch and optionally returns their corresponding pixel mask.
Args:
image (`torch.Tensor`):
Image to pad.
constant_values (`float` or `Iterable[float]`, *optional*):
The value to use for the padding if `mode` is `"constant"`.
return_pixel_mask (`bool`, *optional*, defaults to `True`):
Whether to return a pixel mask.
disable_grouping (`bool`, *optional*, defaults to `False`):
Whether to disable grouping of images by size.
return_tensors (`str` or `TensorType`, *optional*):
The type of tensors to return. Can be one of:
- Unset: Return a list of `np.ndarray`.
- `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
- `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
- `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
- `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
"""
pad_size = get_max_height_width(images)
grouped_images, grouped_images_index = group_images_by_shape(images, disable_grouping=disable_grouping)
processed_images_grouped = {}
processed_masks_grouped = {}
for shape, stacked_images in grouped_images.items():
stacked_images = self._pad_image(
stacked_images,
pad_size,
constant_values=constant_values,
)
processed_images_grouped[shape] = stacked_images
if return_pixel_mask:
stacked_masks = make_pixel_mask(image=stacked_images, output_size=pad_size)
processed_masks_grouped[shape] = stacked_masks
processed_images = reorder_images(processed_images_grouped, grouped_images_index)
processed_masks = None
if return_pixel_mask:
processed_masks = reorder_images(processed_masks_grouped, grouped_images_index)
return processed_images, processed_masks
def _preprocess(
self,
images: list["torch.Tensor"],
do_resize: bool,
size: SizeDict,
size_divisor: Optional[int],
interpolation: Optional["F.InterpolationMode"],
do_pad: bool,
do_center_crop: bool,
crop_size: SizeDict,
do_rescale: bool,
rescale_factor: float,
do_normalize: bool,
image_mean: Optional[Union[float, list[float]]],
image_std: Optional[Union[float, list[float]]],
disable_grouping: Optional[bool],
return_tensors: Optional[Union[str, TensorType]],
**kwargs,
) -> BatchFeature:
# Group images by size for batched resizing
grouped_images, grouped_images_index = group_images_by_shape(images, disable_grouping=disable_grouping)
resized_images_grouped = {}
for shape, stacked_images in grouped_images.items():
if do_resize:
stacked_images = self.resize(
image=stacked_images, size=size, size_divisor=size_divisor, interpolation=interpolation
)
resized_images_grouped[shape] = stacked_images
resized_images = reorder_images(resized_images_grouped, grouped_images_index)
# Group images by size for further processing
# Needed in case do_resize is False, or resize returns images with different sizes
grouped_images, grouped_images_index = group_images_by_shape(resized_images, disable_grouping=disable_grouping)
processed_images_grouped = {}
for shape, stacked_images in grouped_images.items():
if do_center_crop:
stacked_images = self.center_crop(stacked_images, crop_size)
# Fused rescale and normalize
stacked_images = self.rescale_and_normalize(
stacked_images, do_rescale, rescale_factor, do_normalize, image_mean, image_std
)
processed_images_grouped[shape] = stacked_images
processed_images = reorder_images(processed_images_grouped, grouped_images_index)
data = {}
if do_pad:
processed_images, processed_masks = self.pad(
processed_images, return_pixel_mask=True, disable_grouping=disable_grouping
)
processed_masks = torch.stack(processed_masks, dim=0) if return_tensors else processed_masks
data["pixel_mask"] = processed_masks
processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
data["pixel_values"] = processed_images
return BatchFeature(data=data, tensor_type=return_tensors)
def to_dict(self):
encoder_dict = super().to_dict()
encoder_dict.pop("_valid_processor_keys", None)
encoder_dict.pop("crop_size", None)
return encoder_dict
__all__ = ["BridgeTowerImageProcessorFast"]
| transformers/src/transformers/models/bridgetower/image_processing_bridgetower_fast.py/0 | {
"file_path": "transformers/src/transformers/models/bridgetower/image_processing_bridgetower_fast.py",
"repo_id": "transformers",
"token_count": 5634
} | 434 |
# coding=utf-8
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch CLVP model."""
import copy
import math
from dataclasses import dataclass
from typing import Callable, Optional, Union
import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import CrossEntropyLoss
from ...activations import ACT2FN, get_activation
from ...cache_utils import Cache, DynamicCache
from ...generation import GenerationConfig, GenerationMixin
from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
from ...modeling_outputs import (
BaseModelOutput,
BaseModelOutputWithPastAndCrossAttentions,
BaseModelOutputWithPooling,
CausalLMOutputWithCrossAttentions,
)
from ...modeling_utils import PreTrainedModel
from ...pytorch_utils import Conv1D, isin_mps_friendly
from ...utils import (
ModelOutput,
auto_docstring,
logging,
)
from ...utils.deprecation import deprecate_kwarg
from .configuration_clvp import (
ClvpConfig,
ClvpDecoderConfig,
ClvpEncoderConfig,
)
logger = logging.get_logger(__name__)
# Copied from transformers.models.clip.modeling_clip.contrastive_loss
def contrastive_loss(logits: torch.Tensor) -> torch.Tensor:
return nn.functional.cross_entropy(logits, torch.arange(len(logits), device=logits.device))
# Copied from transformers.models.clip.modeling_clip.clip_loss with clip->clvp, image_loss->speech_loss
def clvp_loss(similarity: torch.Tensor) -> torch.Tensor:
caption_loss = contrastive_loss(similarity)
speech_loss = contrastive_loss(similarity.t())
return (caption_loss + speech_loss) / 2.0
# Copied from transformers.models.llama.modeling_llama.rotate_half
def rotate_half(x):
"""Rotates half the hidden dims of the input."""
x1 = x[..., : x.shape[-1] // 2]
x2 = x[..., x.shape[-1] // 2 :]
return torch.cat((-x2, x1), dim=-1)
def apply_rotary_pos_emb(q, k, v, cos, sin, position_ids, unsqueeze_dim=1):
"""Applies Rotary Position Embedding to the query and key tensors.
Args:
q (`torch.Tensor`): The query tensor.
k (`torch.Tensor`): The key tensor.
cos (`torch.Tensor`): The cosine part of the rotary embedding.
sin (`torch.Tensor`): The sine part of the rotary embedding.
position_ids (`torch.Tensor`):
The position indices of the tokens corresponding to the query and key tensors. For example, this can be
used to pass offsetted position ids when working with a KV-cache.
unsqueeze_dim (`int`, *optional*, defaults to 1):
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
Returns:
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
"""
cos = cos[position_ids].unsqueeze(unsqueeze_dim)
sin = sin[position_ids].unsqueeze(unsqueeze_dim)
q_embed = (q * cos) + (rotate_half(q) * sin)
k_embed = (k * cos) + (rotate_half(k) * sin)
v_embed = (v * cos) + (rotate_half(v) * sin)
return q_embed, k_embed, v_embed
def _pad_extra_bos_eos_tokens(
input_ids,
attention_mask=None,
pad_token_id=0,
bos_token_id=255,
eos_token_id=0,
add_bos_token=True,
add_eos_token=True,
):
"""
This method adds extra bos and eos tokens to input_ids and accordingly modifies the attention_mask which is used in
`ClvpConditioningEncoder` and the generation loop of the `ClvpModelForConditionalGeneration`.
"""
# add the bos token at the beginning
if add_bos_token:
input_ids = torch.nn.functional.pad(input_ids, (1, 0), value=bos_token_id)
attention_mask = (
torch.nn.functional.pad(attention_mask, (1, 0), value=1) if attention_mask is not None else attention_mask
)
modified_input_ids = input_ids
if add_eos_token:
modified_input_ids = torch.zeros(
(input_ids.shape[0], input_ids.shape[1] + 1), dtype=input_ids.dtype, device=input_ids.device
)
for i, each_input_id in enumerate(input_ids):
# locate where the valid tokens end and then add the eos token
if isin_mps_friendly(each_input_id, pad_token_id).sum():
pos = torch.where(each_input_id == pad_token_id)[0].min()
modified_input_ids[i] = torch.concatenate(
[each_input_id[:pos], torch.tensor([eos_token_id], device=input_ids.device), each_input_id[pos:]]
)
else:
# if there are no pad tokens present, then add eos to the end
modified_input_ids[i] = torch.nn.functional.pad(each_input_id, (0, 1), value=eos_token_id)
attention_mask = (
torch.nn.functional.pad(attention_mask, (1, 0), value=1) if attention_mask is not None else attention_mask
)
return modified_input_ids, attention_mask
@dataclass
@auto_docstring(
custom_intro="""
Base class for CLVP encoder's outputs that contains a pooling of the last hidden states as well as a projection
output (a linear layer on top of the pooled output).
"""
)
class ClvpEncoderOutput(ModelOutput):
r"""
embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)`, *optional*, returned when model is initialized with `with_projection=True`):
The embeddings obtained by applying the projection layer to the pooler_output.
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
The hidden state of the last layer of the model.
pooler_output (`torch.FloatTensor` of shape `(batch_size, hidden_size)`):
Pooled output of the `last_hidden_state`.
"""
embeds: Optional[torch.FloatTensor] = None
last_hidden_state: Optional[torch.FloatTensor] = None
pooler_output: Optional[torch.FloatTensor] = None
hidden_states: Optional[tuple[torch.FloatTensor]] = None
attentions: Optional[tuple[torch.FloatTensor]] = None
@dataclass
@auto_docstring
class ClvpOutput(ModelOutput):
r"""
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
Contrastive loss for speech-text similarity.
speech_ids (`torch.LongTensor`, *optional*):
speech_ids (or speech candidates) generated by the `ClvpForCausalLM` model.
logits_per_speech (`torch.FloatTensor` of shape `(speech_batch_size, text_batch_size)`):
The scaled dot product scores between `speech_embeds` and `text_embeds`. This represents the speech-text
similarity scores.
logits_per_text (`torch.FloatTensor` of shape `(text_batch_size, speech_batch_size)`):
The scaled dot product scores between `text_embeds` and `speech_embeds`. This represents the text-speech
similarity scores.
text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
The text embeddings obtained by applying the projection layer to the pooled output of the text encoder
model.
speech_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
The speech embeddings obtained by applying the projection layer to the pooled output of the speech encoder
model.
text_model_output (`BaseModelOutputWithPooling`):
The pooled output of the `last_hidden_state` of the text encoder Model.
speech_model_output (`BaseModelOutputWithPooling`):
The pooled output of the `last_hidden_state` of the speech encoder Model.
decoder_hidden_states (`torch.FloatTensor`, *optional*):
The hidden states of the decoder model.
text_encoder_hidden_states (`torch.FloatTensor`, *optional*):
The hidden states of the text encoder model.
speech_encoder_hidden_states (`torch.FloatTensor`, *optional*):
The hidden states of the speech encoder model.
"""
loss: Optional[torch.FloatTensor] = None
speech_ids: Optional[torch.LongTensor] = None
logits_per_speech: Optional[torch.FloatTensor] = None
logits_per_text: Optional[torch.FloatTensor] = None
text_embeds: Optional[torch.FloatTensor] = None
speech_embeds: Optional[torch.FloatTensor] = None
text_model_output: BaseModelOutputWithPooling = None
speech_model_output: BaseModelOutputWithPooling = None
decoder_hidden_states: Optional[torch.FloatTensor] = None
text_encoder_hidden_states: Optional[torch.FloatTensor] = None
speech_encoder_hidden_states: Optional[torch.FloatTensor] = None
# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Clvp
class ClvpRMSNorm(nn.Module):
def __init__(self, hidden_size, eps=1e-6):
"""
ClvpRMSNorm is equivalent to T5LayerNorm
"""
super().__init__()
self.weight = nn.Parameter(torch.ones(hidden_size))
self.variance_epsilon = eps
def forward(self, hidden_states):
input_dtype = hidden_states.dtype
hidden_states = hidden_states.to(torch.float32)
variance = hidden_states.pow(2).mean(-1, keepdim=True)
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
return self.weight * hidden_states.to(input_dtype)
def extra_repr(self):
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
class ClvpRotaryPositionalEmbedding(nn.Module):
"""
Rotary Position Embedding Class for CLVP. It was proposed in the paper 'ROFORMER: ENHANCED TRANSFORMER WITH ROTARY
POSITION EMBEDDING', Please see https://huggingface.co/papers/2104.09864v1.pdf .
"""
def __init__(self, config):
super().__init__()
dim = max(config.projection_dim // (config.num_attention_heads * 2), 32)
inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, dtype=torch.int64).float() / dim))
self.register_buffer("inv_freq", inv_freq)
self.cached_sequence_length = None
self.cached_rotary_positional_embedding = None
def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
sequence_length = hidden_states.shape[1]
if sequence_length == self.cached_sequence_length and self.cached_rotary_positional_embedding is not None:
return self.cached_rotary_positional_embedding
self.cached_sequence_length = sequence_length
time_stamps = torch.arange(sequence_length, device=hidden_states.device).type_as(self.inv_freq)
freqs = torch.einsum("i,j->ij", time_stamps, self.inv_freq)
embeddings = torch.cat((freqs, freqs), dim=-1)
self.cached_rotary_positional_embedding = embeddings.unsqueeze(0)
return self.cached_rotary_positional_embedding
class ClvpSelfAttention(nn.Module):
"""
Multi-headed attention to combine Absolute and Rotary Positional Embeddings into a single Attention module.
"""
def __init__(self, config, layer_idx=None):
super().__init__()
self.config = config
self.embed_dim = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = self.embed_dim // self.num_heads
if self.head_dim * self.num_heads != self.embed_dim:
raise ValueError(
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
f" {self.num_heads})."
)
self.scale = self.head_dim**-0.5
self.dropout = config.attention_dropout
self.layer_idx = layer_idx
if hasattr(config, "max_position_embeddings"):
max_positions = config.max_position_embeddings
bias = torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool))
bias = bias.view(1, 1, max_positions, max_positions)
self.register_buffer("bias", bias, persistent=False)
self.k_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.use_attention_bias)
self.v_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.use_attention_bias)
self.q_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.use_attention_bias)
self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
@deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
def forward(
self,
hidden_states: torch.FloatTensor,
rotary_pos_emb: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[Cache] = None,
use_cache: Optional[bool] = False,
head_mask: Optional[torch.FloatTensor] = None,
output_attentions: Optional[bool] = False,
cache_position: Optional[torch.Tensor] = None,
) -> tuple[torch.FloatTensor, Optional[torch.FloatTensor], Optional[tuple[torch.FloatTensor]]]:
# Raise error when position_ids is None but rotary_pos_emb is provided, because we need that when applying
# rotary_pos_emb to query and key states.
if rotary_pos_emb is not None and position_ids is None:
raise ValueError("`position_ids` must be provided when `rotary_pos_emb` is not None.")
bsz, _, embed_dim = hidden_states.size()
# get query proj
query_states = self._shape(self.q_proj(hidden_states), -1, bsz) * self.scale
key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
if past_key_values is not None:
key_states, value_states = past_key_values.update(
key_states, value_states, self.layer_idx, {"cache_position": cache_position}
)
if rotary_pos_emb is not None:
rotary_emb_dim = rotary_pos_emb.shape[-1]
# Partial rotary embedding
query_rot, query_pass = (
query_states[..., :rotary_emb_dim],
query_states[..., rotary_emb_dim:],
)
key_rot, key_pass = (
key_states[..., :rotary_emb_dim],
key_states[..., rotary_emb_dim:],
)
value_rot, value_pass = (
value_states[..., :rotary_emb_dim],
value_states[..., rotary_emb_dim:],
)
cos, sin = rotary_pos_emb.cos().squeeze(0), rotary_pos_emb.sin().squeeze(0)
query_rot, key_rot, value_rot = apply_rotary_pos_emb(query_rot, key_rot, value_rot, cos, sin, position_ids)
# [batch_size, num_heads, seq_length, head_dim]
query_states = torch.cat((query_rot, query_pass), dim=-1)
key_states = torch.cat((key_rot, key_pass), dim=-1)
value_states = torch.cat((value_rot, value_pass), dim=-1)
tgt_len = query_states.shape[2]
src_len = key_states.shape[2]
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3))
if attention_mask is not None:
if attention_mask.size() != (bsz, 1, tgt_len, src_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
)
attn_weights = attn_weights + attention_mask
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
# Mask heads if we want to
if head_mask is not None:
attn_weights = attn_weights * head_mask
attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
attn_output = torch.matmul(attn_probs, value_states)
if attn_output.size() != (bsz, self.num_heads, tgt_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
f" {attn_output.size()}"
)
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
attn_output = self.out_proj(attn_output)
return attn_output, attn_weights
class ClvpGatedLinearUnit(nn.Module):
"""
`ClvpGatedLinearUnit` uses the second half of the `hidden_states` to act as a gate for the first half of the
`hidden_states` which controls the flow of data from the first of the tensor.
"""
def __init__(self, config):
super().__init__()
self.activation_fn = ACT2FN[config.hidden_act]
self.proj = nn.Linear(config.hidden_size, config.intermediate_size * 2)
def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
hidden_states, gate = self.proj(hidden_states).chunk(2, dim=-1)
return hidden_states * self.activation_fn(gate)
class ClvpEncoderMLP(nn.Module):
"""
This MLP is used in CLVP speech or text encoder models.
"""
def __init__(self, config):
super().__init__()
self.config = config
self.fc1 = ClvpGatedLinearUnit(config)
self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
self.dropout_layer = nn.Dropout(config.dropout)
def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
hidden_states = self.fc1(hidden_states)
hidden_states = self.dropout_layer(hidden_states)
hidden_states = self.fc2(hidden_states)
return hidden_states
class ClvpEncoderLayer(nn.Module):
def __init__(self, config: ClvpConfig):
super().__init__()
self.config = config
self.embed_dim = config.hidden_size
self.self_attn = ClvpSelfAttention(config)
self.mlp = ClvpEncoderMLP(config)
self.input_rmsnorm = ClvpRMSNorm(self.embed_dim, eps=config.layer_norm_eps)
self.post_attention_rmsnorm = ClvpRMSNorm(self.embed_dim, eps=config.layer_norm_eps)
def forward(
self,
hidden_states: torch.FloatTensor,
rotary_pos_emb: torch.FloatTensor,
attention_mask: torch.LongTensor,
position_ids: torch.LongTensor,
output_attentions: Optional[bool] = False,
) -> tuple[torch.FloatTensor]:
"""
Args:
hidden_states (`torch.FloatTensor` of shape `(batch, seq_len, embed_dim)`):
input to the layer.
rotary_pos_emb (`torch.FloatTensor`):
rotary position embeddings generated by `ClvpRotaryPositionalEmbedding` module.
attention_mask (`torch.FloatTensor` of shape `(batch, 1, tgt_len, src_len)`):
attention mask where padding elements are indicated by very large negative values.
position_ids (`torch.LongTensor`):
Denotes position ids of the input tokens.
output_attentions (`bool`, *optional*, defaults to `False`):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
"""
residual = hidden_states
hidden_states = self.input_rmsnorm(hidden_states)
hidden_states, attn_weights = self.self_attn(
hidden_states=hidden_states,
rotary_pos_emb=rotary_pos_emb,
attention_mask=attention_mask,
position_ids=position_ids,
output_attentions=output_attentions,
)
hidden_states = residual + hidden_states
residual = hidden_states
hidden_states = self.post_attention_rmsnorm(hidden_states)
hidden_states = self.mlp(hidden_states)
hidden_states = residual + hidden_states
return hidden_states, attn_weights
# Copied from transformers.models.xlm.modeling_xlm.XLMSequenceSummary with XLM->Clvp
class ClvpSequenceSummary(nn.Module):
r"""
Compute a single vector summary of a sequence hidden states.
Args:
config ([`ClvpConfig`]):
The config used by the model. Relevant arguments in the config class of the model are (refer to the actual
config class of your model for the default values it uses):
- **summary_type** (`str`) -- The method to use to make this summary. Accepted values are:
- `"last"` -- Take the last token hidden state (like XLNet)
- `"first"` -- Take the first token hidden state (like Bert)
- `"mean"` -- Take the mean of all tokens hidden states
- `"cls_index"` -- Supply a Tensor of classification token position (GPT/GPT-2)
- `"attn"` -- Not implemented now, use multi-head attention
- **summary_use_proj** (`bool`) -- Add a projection after the vector extraction.
- **summary_proj_to_labels** (`bool`) -- If `True`, the projection outputs to `config.num_labels` classes
(otherwise to `config.hidden_size`).
- **summary_activation** (`Optional[str]`) -- Set to `"tanh"` to add a tanh activation to the output,
another string or `None` will add no activation.
- **summary_first_dropout** (`float`) -- Optional dropout probability before the projection and activation.
- **summary_last_dropout** (`float`)-- Optional dropout probability after the projection and activation.
"""
def __init__(self, config: ClvpConfig):
super().__init__()
self.summary_type = getattr(config, "summary_type", "last")
if self.summary_type == "attn":
# We should use a standard multi-head attention module with absolute positional embedding for that.
# Cf. https://github.com/zihangdai/xlnet/blob/master/modeling.py#L253-L276
# We can probably just use the multi-head attention module of PyTorch >=1.1.0
raise NotImplementedError
self.summary = nn.Identity()
if hasattr(config, "summary_use_proj") and config.summary_use_proj:
if hasattr(config, "summary_proj_to_labels") and config.summary_proj_to_labels and config.num_labels > 0:
num_classes = config.num_labels
else:
num_classes = config.hidden_size
self.summary = nn.Linear(config.hidden_size, num_classes)
activation_string = getattr(config, "summary_activation", None)
self.activation: Callable = get_activation(activation_string) if activation_string else nn.Identity()
self.first_dropout = nn.Identity()
if hasattr(config, "summary_first_dropout") and config.summary_first_dropout > 0:
self.first_dropout = nn.Dropout(config.summary_first_dropout)
self.last_dropout = nn.Identity()
if hasattr(config, "summary_last_dropout") and config.summary_last_dropout > 0:
self.last_dropout = nn.Dropout(config.summary_last_dropout)
def forward(
self, hidden_states: torch.FloatTensor, cls_index: Optional[torch.LongTensor] = None
) -> torch.FloatTensor:
"""
Compute a single vector summary of a sequence hidden states.
Args:
hidden_states (`torch.FloatTensor` of shape `[batch_size, seq_len, hidden_size]`):
The hidden states of the last layer.
cls_index (`torch.LongTensor` of shape `[batch_size]` or `[batch_size, ...]` where ... are optional leading dimensions of `hidden_states`, *optional*):
Used if `summary_type == "cls_index"` and takes the last token of the sequence as classification token.
Returns:
`torch.FloatTensor`: The summary of the sequence hidden states.
"""
if self.summary_type == "last":
output = hidden_states[:, -1]
elif self.summary_type == "first":
output = hidden_states[:, 0]
elif self.summary_type == "mean":
output = hidden_states.mean(dim=1)
elif self.summary_type == "cls_index":
if cls_index is None:
cls_index = torch.full_like(
hidden_states[..., :1, :],
hidden_states.shape[-2] - 1,
dtype=torch.long,
)
else:
cls_index = cls_index.unsqueeze(-1).unsqueeze(-1)
cls_index = cls_index.expand((-1,) * (cls_index.dim() - 1) + (hidden_states.size(-1),))
# shape of cls_index: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states
output = hidden_states.gather(-2, cls_index).squeeze(-2) # shape (bsz, XX, hidden_size)
elif self.summary_type == "attn":
raise NotImplementedError
output = self.first_dropout(output)
output = self.summary(output)
output = self.activation(output)
output = self.last_dropout(output)
return output
# Copied from transformers.models.gpt2.modeling_gpt2.GPT2MLP with GPT2->ClvpDecoderMLP
class ClvpDecoderMLP(nn.Module):
def __init__(self, intermediate_size, config):
super().__init__()
embed_dim = config.hidden_size
self.c_fc = Conv1D(intermediate_size, embed_dim)
self.c_proj = Conv1D(embed_dim, intermediate_size)
self.act = ACT2FN[config.activation_function]
self.dropout = nn.Dropout(config.resid_pdrop)
def forward(self, hidden_states: Optional[tuple[torch.FloatTensor]]) -> torch.FloatTensor:
hidden_states = self.c_fc(hidden_states)
hidden_states = self.act(hidden_states)
hidden_states = self.c_proj(hidden_states)
hidden_states = self.dropout(hidden_states)
return hidden_states
class ClvpDecoderLayer(nn.Module):
def __init__(self, config, layer_idx=None):
super().__init__()
hidden_size = config.hidden_size
inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
self.input_layernorm = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
self.attn = ClvpSelfAttention(config, layer_idx=layer_idx)
self.post_attention_layernorm = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
self.mlp = ClvpDecoderMLP(inner_dim, config)
@deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
def forward(
self,
hidden_states: Optional[tuple[torch.FloatTensor]],
past_key_values: Optional[Cache] = None,
attention_mask: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
head_mask: Optional[torch.FloatTensor] = None,
use_cache: Optional[bool] = False,
output_attentions: Optional[bool] = False,
cache_position: Optional[torch.Tensor] = None,
) -> Union[tuple[torch.Tensor], Optional[tuple[torch.Tensor, tuple[torch.FloatTensor, ...]]]]:
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
attn_outputs = self.attn(
hidden_states,
past_key_values=past_key_values,
attention_mask=attention_mask,
position_ids=position_ids,
head_mask=head_mask,
use_cache=use_cache,
output_attentions=output_attentions,
cache_position=cache_position,
)
attn_output = attn_outputs[0]
# residual connection
hidden_states = attn_output + residual
residual = hidden_states
hidden_states = self.post_attention_layernorm(hidden_states)
feed_forward_hidden_states = self.mlp(hidden_states)
# residual connection
hidden_states = residual + feed_forward_hidden_states
return (hidden_states,) + attn_outputs[1:]
class ClvpConditioningEncoder(nn.Module):
"""
This class processes the log-mel spectrograms(extracted by the Feature Extractor) and text tokens(produced by the
tokenizer) as inputs for the decoder model.
First each log-mel spectrogram is processed into a single vector which captures valuable characteristics from each
of them, then the text tokens are converted into token embeddings and position embeddings are added afterwards.
Both of these vectors are concatenated and then passed to the decoder model.
The text tokens helps to incorporate the "text information" and the log-mel spectrogram is used to specify the
"voice characteristics" into the generated mel tokens.
"""
def __init__(self, config: ClvpConfig):
super().__init__()
self.text_config = config.text_config
self.decoder_config = config.decoder_config
self.text_token_embedding = nn.Embedding(self.text_config.vocab_size, self.decoder_config.hidden_size)
self.text_position_embedding = nn.Embedding(
self.decoder_config.max_text_tokens, self.decoder_config.hidden_size
)
self.mel_conv = nn.Conv1d(self.decoder_config.feature_size, self.decoder_config.hidden_size, kernel_size=1)
# define group norms to be used before each attention layer
num_groups = self.compute_groupnorm_groups(self.decoder_config.hidden_size)
self.group_norms = nn.ModuleList(
[
nn.GroupNorm(num_groups, self.decoder_config.hidden_size, eps=1e-5, affine=True)
for _ in range(self.decoder_config.num_mel_attn_blocks)
]
)
# define the attention layers
self.mel_attn_blocks = nn.ModuleList(
[ClvpSelfAttention(self.decoder_config) for _ in range(self.decoder_config.num_mel_attn_blocks)]
)
self.gradient_checkpointing = False
def compute_groupnorm_groups(self, channels: int, groups: int = 32):
"""
Calculates the value of `num_groups` for nn.GroupNorm. This logic is taken from the official tortoise
repository. link :
https://github.com/neonbjb/tortoise-tts/blob/4003544b6ff4b68c09856e04d3eff9da26d023c2/tortoise/models/arch_util.py#L26
"""
if channels <= 16:
groups = 8
elif channels <= 64:
groups = 16
while channels % groups != 0:
groups = int(groups / 2)
if groups <= 2:
raise ValueError(
f"Number of groups for the GroupNorm must be greater than 2, but it is {groups}."
f"Please consider using a different `hidden_size`"
)
return groups
def forward(
self,
input_features: torch.FloatTensor,
input_ids: Optional[torch.LongTensor] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.LongTensor] = None,
):
# process text
if input_ids is not None and inputs_embeds is not None:
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
elif input_ids is not None:
batch_size, seq_length = input_ids.size()
elif inputs_embeds is not None:
batch_size, seq_length = inputs_embeds.size()[:-1]
else:
raise ValueError("You have to specify either input_ids or inputs_embeds")
# construct attention mask if not given
if attention_mask is None:
attention_mask = torch.ones([batch_size, seq_length], dtype=torch.long, device=input_ids.device)
# We add bos and eos input_ids in the modeling file instead of the tokenizer file to keep the logic simple
# This logic is specific to ClvpConditioningEncoder and not used by other modules.
input_ids, attention_mask = _pad_extra_bos_eos_tokens(
input_ids,
attention_mask,
bos_token_id=self.text_config.bos_token_id,
eos_token_id=self.text_config.eos_token_id,
)
inputs_embeds = self.text_token_embedding(input_ids)
position_ids = attention_mask.cumsum(-1) - 1
position_embeds = self.text_position_embedding(position_ids)
text_embeds = inputs_embeds + position_embeds
if self.gradient_checkpointing and self.training:
# process each log-mel spectrogram into a single vector
mel_spec = torch.utils.checkpoint.checkpoint(self.mel_conv, input_features)
for i, mel_attn_block in enumerate(self.mel_attn_blocks):
residual_mel_spec = mel_spec.transpose(1, 2)
mel_spec = torch.utils.checkpoint.checkpoint(self.group_norms[i], mel_spec).transpose(1, 2)
mel_spec = torch.utils.checkpoint.checkpoint(mel_attn_block, mel_spec)[0] + residual_mel_spec
mel_spec = mel_spec.transpose(1, 2)
else:
# process each log-mel spectrogram into a single vector
mel_spec = self.mel_conv(input_features)
for i, mel_attn_block in enumerate(self.mel_attn_blocks):
residual_mel_spec = mel_spec.transpose(1, 2)
mel_spec = self.group_norms[i](mel_spec).transpose(1, 2)
mel_spec = mel_attn_block(mel_spec)[0] + residual_mel_spec
mel_spec = mel_spec.transpose(1, 2)
mel_spec = mel_spec[:, :, 0]
mel_spec = mel_spec.unsqueeze(1)
# repeat if there is either (1 text vs N audios) or (N texts vs 1 audio)
if text_embeds.shape[0] == 1 and mel_spec.shape[0] != 1:
text_embeds = text_embeds.repeat(mel_spec.shape[0], 1, 1)
elif text_embeds.shape[0] != 1 and mel_spec.shape[0] == 1:
mel_spec = mel_spec.repeat(text_embeds.shape[0], 1, 1)
# If there is N texts and M audios we will raise error since the number of text and audio must be same.
elif text_embeds.shape[0] != mel_spec.shape[0]:
raise ValueError(
f"The number of texts and number of audios must be same. "
f"Found {text_embeds.shape[0]} texts vs {mel_spec.shape[0]} audios"
)
return torch.concat([mel_spec, text_embeds], dim=1)
@auto_docstring
class ClvpPreTrainedModel(PreTrainedModel):
config: ClvpConfig
base_model_prefix = "clvp"
supports_gradient_checkpointing = True
_skip_keys_device_placement = "past_key_values"
def _init_weights(self, module: nn.Module):
"""Initialize the weights"""
factor = self.config.initializer_factor
if isinstance(module, nn.Embedding):
module.weight.data.normal_(mean=0.0, std=factor * 0.02)
elif isinstance(module, (nn.Linear, Conv1D, nn.Conv1d)):
module.weight.data.normal_(mean=0.0, std=factor * 0.02)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, ClvpRMSNorm):
module.weight.data.fill_(1.0)
elif isinstance(module, ClvpEncoderMLP):
in_proj_std = (module.config.hidden_size**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor
fc_std = (2 * module.config.hidden_size) ** -0.5 * factor
nn.init.normal_(module.fc1.proj.weight if getattr(module.fc1, "proj") else module.fc1.weight, std=fc_std)
nn.init.normal_(module.fc2.weight, std=in_proj_std)
elif isinstance(module, ClvpEncoder):
config = self.config.get_text_config()
factor = config.initializer_factor
module.projection.weight.data.normal_(mean=0.0, std=factor * (config.hidden_size**-0.5))
elif isinstance(module, ClvpConditioningEncoder):
module.mel_conv.weight.data.normal_(mean=0.0, std=factor)
module.mel_conv.bias.data.zero_()
elif isinstance(module, ClvpForCausalLM):
for name, p in module.named_parameters():
if name == "c_proj.weight":
p.data.normal_(
mean=0.0, std=(self.config.initializer_range / math.sqrt(2 * self.config.num_hidden_layers))
)
elif isinstance(module, ClvpModelForConditionalGeneration):
module.logit_scale.data.fill_(self.config.logit_scale_init_value)
if isinstance(module, (nn.LayerNorm, nn.GroupNorm)):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
class ClvpEncoder(ClvpPreTrainedModel):
"""
Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
[`ClvpEncoderLayer`].
Args:
config: ClvpConfig
"""
def __init__(self, config: ClvpConfig):
super().__init__(config)
self.config = config
self.token_embedding = nn.Embedding(config.vocab_size, config.hidden_size)
self.rotary_pos_emb = ClvpRotaryPositionalEmbedding(config) if config.use_rotary_embedding else None
self.layers = nn.ModuleList([ClvpEncoderLayer(config) for _ in range(config.num_hidden_layers)])
self.sequence_summary = ClvpSequenceSummary(config)
self.final_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
self.projection = nn.Linear(config.hidden_size, config.projection_dim, bias=False)
self.gradient_checkpointing = False
self.post_init()
def get_input_embeddings(self):
return self.token_embedding
def set_input_embeddings(self, value):
self.token_embedding = value
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
inputs_embeds: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[tuple, BaseModelOutput]:
r"""
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
input embeddings for the model. This bypasses the model's internal embedding lookup matrix.
attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
position_ids (`torch.LongTensor`, *optional*):
Denotes the position ids of `input_ids`.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
for more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if input_ids is not None and inputs_embeds is not None:
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
elif input_ids is not None:
self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
input_shape = input_ids.size()
input_ids = input_ids.view(-1, input_shape[-1])
inputs_embeds = self.token_embedding(input_ids)
elif inputs_embeds is not None:
input_shape = inputs_embeds.size()[:-1]
else:
raise ValueError("You have to specify either input_ids or inputs_embeds")
# expand attention_mask and create position_ids if needed
if attention_mask is not None:
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
attention_mask = _prepare_4d_attention_mask(attention_mask, inputs_embeds.dtype)
if position_ids is None:
device = input_ids.device if input_ids is not None else inputs_embeds.device
position_ids = torch.arange(input_shape[1], dtype=torch.long, device=device)
position_ids = position_ids.unsqueeze(0)
encoder_states = () if output_hidden_states else None
all_attentions = () if output_attentions else None
rotary_pos_emb = self.rotary_pos_emb(inputs_embeds) if self.rotary_pos_emb is not None else None
hidden_states = inputs_embeds
for idx, encoder_layer in enumerate(self.layers):
if output_hidden_states:
encoder_states = encoder_states + (hidden_states,)
if self.gradient_checkpointing and self.training:
layer_outputs = torch.utils.checkpoint.checkpoint(
encoder_layer.__call__,
hidden_states,
rotary_pos_emb,
attention_mask,
position_ids,
)
else:
layer_outputs = encoder_layer(
hidden_states,
rotary_pos_emb,
attention_mask,
position_ids,
output_attentions=output_attentions,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_attentions = all_attentions + (layer_outputs[1],)
if output_hidden_states:
encoder_states = encoder_states + (hidden_states,)
last_hidden_state = hidden_states
last_hidden_state = self.final_layer_norm(last_hidden_state)
# take the mean over axis 1 and get pooled output
pooled_output = self.sequence_summary(last_hidden_state)
# apply the projection layer
embeds = self.projection(pooled_output)
if not return_dict:
return tuple(
v for v in [embeds, last_hidden_state, pooled_output, encoder_states, all_attentions] if v is not None
)
return ClvpEncoderOutput(
embeds=embeds,
last_hidden_state=last_hidden_state,
pooler_output=pooled_output,
hidden_states=encoder_states,
attentions=all_attentions,
)
class ClvpDecoder(ClvpPreTrainedModel):
"""
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`ClvpDecoderLayer`]
"""
def __init__(self, config):
super().__init__(config)
self.config = config
self.input_embeds_layer = nn.Embedding(self.config.vocab_size, self.config.hidden_size)
self.position_embeds_layer = nn.Embedding(self.config.max_position_embeddings, self.config.hidden_size)
self.drop = nn.Dropout(self.config.embd_pdrop)
self.layers = nn.ModuleList(
[ClvpDecoderLayer(self.config, layer_idx=i) for i in range(self.config.num_hidden_layers)]
)
self.layer_norm = nn.LayerNorm(self.config.hidden_size, eps=self.config.layer_norm_epsilon)
self.gradient_checkpointing = False
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self):
return self.input_embeds_layer
def set_input_embeddings(self, new_embeddings):
self.input_embeds_layer = new_embeddings
def _prune_heads(self, heads_to_prune):
"""
Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
"""
for layer, heads in heads_to_prune.items():
self.layers[layer].attn.prune_heads(heads)
@auto_docstring
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
token_type_ids: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
head_mask: Optional[torch.FloatTensor] = None,
past_key_values: Optional[tuple[tuple[torch.Tensor]]] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
cache_position: Optional[torch.Tensor] = None,
) -> Union[tuple, BaseModelOutputWithPastAndCrossAttentions]:
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
use_cache = use_cache if use_cache is not None else self.config.use_cache
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if input_ids is not None and inputs_embeds is not None:
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
elif input_ids is not None:
self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
input_shape = input_ids.size()
input_ids = input_ids.view(-1, input_shape[-1])
input_ids.shape[0]
elif inputs_embeds is not None:
input_shape = inputs_embeds.size()[:-1]
inputs_embeds.shape[0]
else:
raise ValueError("You have to specify either input_ids or inputs_embeds")
device = input_ids.device if input_ids is not None else inputs_embeds.device
if token_type_ids is not None:
token_type_ids = token_type_ids.view(-1, input_shape[-1])
if self.gradient_checkpointing and self.training:
if use_cache:
logger.warning_once(
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
)
use_cache = False
if use_cache and past_key_values is None:
past_key_values = DynamicCache()
if use_cache and isinstance(past_key_values, tuple):
logger.warning_once(
"Passing a tuple of `past_key_values` is deprecated and will be removed in Transformers v4.58.0. "
"You should pass an instance of `DynamicCache` instead, e.g. "
"`past_key_values=DynamicCache.from_legacy_cache(past_key_values)`."
)
past_key_values = DynamicCache.from_legacy_cache(past_key_values)
past_key_values_length = past_key_values.get_seq_length() if past_key_values is not None else 0
if position_ids is None:
position_ids = torch.arange(
past_key_values_length, input_shape[-1] + past_key_values_length, dtype=torch.long, device=device
)
position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
if inputs_embeds is None:
inputs_embeds = self.input_embeds_layer(input_ids)
position_embeds = self.position_embeds_layer(position_ids)
inputs_embeds = inputs_embeds + position_embeds
attention_mask = _prepare_4d_causal_attention_mask(
attention_mask, input_shape, inputs_embeds, past_key_values_length
)
# Prepare head mask if needed
# 1.0 in head_mask indicate we keep the head
# attention_probs has shape bsz x num_attention_heads x N x N
# head_mask has shape num_hidden_layers x batch x num_attention_heads x N x N
head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
hidden_states = inputs_embeds
if token_type_ids is not None:
token_type_embeds = self.input_embeds_layer(token_type_ids)
hidden_states = hidden_states + token_type_embeds
hidden_states = self.drop(hidden_states)
output_shape = (-1,) + input_shape[1:] + (hidden_states.size(-1),)
all_self_attentions = () if output_attentions else None
all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
all_hidden_states = () if output_hidden_states else None
for i, block in enumerate(self.layers):
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if self.gradient_checkpointing and self.training:
outputs = torch.utils.checkpoint.checkpoint(
block.__call__,
hidden_states,
None,
attention_mask,
position_ids,
head_mask[i],
cache_position,
)
else:
outputs = block(
hidden_states,
past_key_values=past_key_values,
attention_mask=attention_mask,
position_ids=position_ids,
head_mask=head_mask[i],
use_cache=use_cache,
output_attentions=output_attentions,
cache_position=cache_position,
)
hidden_states = outputs[0]
if output_attentions:
all_self_attentions = all_self_attentions + (outputs[1],)
if self.config.add_cross_attention:
all_cross_attentions = all_cross_attentions + (outputs[2],)
hidden_states = self.layer_norm(hidden_states)
hidden_states = hidden_states.view(output_shape)
# Add last hidden state
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if not return_dict:
return tuple(
v
for v in [hidden_states, past_key_values, all_hidden_states, all_self_attentions, all_cross_attentions]
if v is not None
)
return BaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=hidden_states,
past_key_values=past_key_values,
hidden_states=all_hidden_states,
attentions=all_self_attentions,
cross_attentions=all_cross_attentions,
)
@auto_docstring
class ClvpModel(ClvpPreTrainedModel):
def __init__(self, config: ClvpDecoderConfig):
super().__init__(config)
self.config = config
self.decoder = ClvpDecoder(self.config)
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self):
return self.decoder.input_embeds_layer
def set_input_embeddings(self, value):
self.decoder.input_embeds_layer = value
def get_decoder(self):
return self.decoder
@auto_docstring
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
token_type_ids: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
head_mask: Optional[torch.FloatTensor] = None,
past_key_values: Optional[tuple[tuple[torch.Tensor]]] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
cache_position: Optional[torch.Tensor] = None,
) -> Union[tuple, BaseModelOutputWithPastAndCrossAttentions]:
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
use_cache = use_cache if use_cache is not None else self.config.use_cache
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# decoder outputs consists of (dec_features, past_key_values, dec_hidden, dec_attn)
decoder_outputs = self.decoder(
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
past_key_values=past_key_values,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
cache_position=cache_position,
)
if not return_dict:
return decoder_outputs
return BaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=decoder_outputs.last_hidden_state,
past_key_values=decoder_outputs.past_key_values,
hidden_states=decoder_outputs.hidden_states,
attentions=decoder_outputs.attentions,
cross_attentions=decoder_outputs.cross_attentions,
)
@auto_docstring(
custom_intro="""
The CLVP decoder model with a language modelling head on top.
"""
)
class ClvpForCausalLM(ClvpPreTrainedModel, GenerationMixin):
def __init__(self, config):
super().__init__(config)
self.config = config
self.model = ClvpModel(self.config)
self.final_norm = nn.LayerNorm(self.config.hidden_size)
self.lm_head = nn.Linear(self.config.hidden_size, self.config.vocab_size, bias=True)
# Initialize weights and apply final processing
self.post_init()
def get_output_embeddings(self):
# NOTE: get_output_embeddings() must return None to prevent accidental weight tying.
# See e.g. https://github.com/huggingface/transformers/pull/39339#discussion_r2219126400
return None
def get_input_embeddings(self):
return self.model.decoder.input_embeds_layer
def set_input_embeddings(self, new_embeddings):
self.model.decoder.input_embeds_layer = new_embeddings
def _prepare_model_inputs(
self,
inputs: Optional[torch.Tensor] = None,
bos_token_id: Optional[int] = None,
model_kwargs: Optional[dict[str, torch.Tensor]] = None,
) -> tuple[torch.Tensor, Optional[str], dict[str, torch.Tensor]]:
"""
This function extracts the model-specific `inputs` for generation.
"""
input_name = self.main_input_name
model_kwargs = {k: v for k, v in model_kwargs.items() if v is not None}
inputs_kwarg = model_kwargs.pop(input_name, None)
if inputs_kwarg is not None and inputs is not None:
raise ValueError(
f"`inputs`: {inputs}` were passed alongside {input_name} which is not allowed."
f"Make sure to either pass {inputs} or {input_name}=..."
)
elif inputs_kwarg is not None:
inputs = inputs_kwarg
if input_name == "input_ids" and "inputs_embeds" in model_kwargs:
model_kwargs["input_ids"] = self._maybe_initialize_input_ids_for_generation(
inputs, bos_token_id, model_kwargs=model_kwargs
)
inputs, input_name = model_kwargs["inputs_embeds"], "inputs_embeds"
# Check if conditioning_embeds are provided or not, if yes then concatenate the bos_token_id at the end of the conditioning_embeds.
# Then we must subtract the positional_ids because during the forward pass it will be added anyways, so we must cancel them out here.
conditioning_embeds = model_kwargs.get("conditioning_embeds")
if conditioning_embeds is not None:
mel_start_token_embedding = self.model.decoder.input_embeds_layer(
torch.full(
(conditioning_embeds.shape[0], 1),
fill_value=self.config.bos_token_id,
device=conditioning_embeds.device,
)
)
mel_start_token_embedding += self.model.decoder.position_embeds_layer(
torch.full((conditioning_embeds.shape[0], 1), fill_value=0, device=conditioning_embeds.device)
)
conditioning_embeds = torch.concat([conditioning_embeds, mel_start_token_embedding], dim=1)
# subtract the positional_ids here
if hasattr(model_kwargs, "attention_mask"):
position_ids = model_kwargs["attention_mask"].long().cumsum(-1) - 1
else:
position_ids = torch.arange(
0, conditioning_embeds.shape[1], dtype=torch.long, device=conditioning_embeds.device
)
position_ids = position_ids.unsqueeze(0).repeat(conditioning_embeds.shape[0], 1)
model_kwargs["inputs_embeds"] = conditioning_embeds - self.model.decoder.position_embeds_layer(
position_ids
)
model_kwargs["input_ids"] = (
torch.ones((model_kwargs["inputs_embeds"].shape[0], 1), dtype=torch.long, device=self.device)
* self.config.bos_token_id
)
return model_kwargs["inputs_embeds"], "inputs_embeds", model_kwargs
inputs = self._maybe_initialize_input_ids_for_generation(inputs, bos_token_id, model_kwargs)
return inputs, input_name, model_kwargs
def prepare_inputs_for_generation(
self,
input_ids,
past_key_values=None,
inputs_embeds=None,
conditioning_embeds=None,
cache_position=None,
**kwargs,
):
# Overwritten: has `conditioning_embeds`-related logic
input_ids_length = input_ids.shape[-1]
model_inputs = super().prepare_inputs_for_generation(
input_ids,
past_key_values=past_key_values,
inputs_embeds=inputs_embeds,
cache_position=cache_position,
**kwargs,
)
if conditioning_embeds is not None and cache_position[0] != 0:
model_inputs["position_ids"] = torch.tensor([input_ids_length], dtype=torch.long, device=input_ids.device)
return model_inputs
@auto_docstring
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[tuple[tuple[torch.Tensor]]] = None,
attention_mask: Optional[torch.FloatTensor] = None,
token_type_ids: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
head_mask: Optional[torch.FloatTensor] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
labels: Optional[torch.LongTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
cache_position: Optional[torch.Tensor] = None,
) -> Union[tuple, CausalLMOutputWithCrossAttentions]:
r"""
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
`labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
use_cache = use_cache if use_cache is not None else self.config.use_cache
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.model(
input_ids=input_ids,
past_key_values=past_key_values,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
cache_position=cache_position,
)
hidden_states = outputs[0]
lm_logits = self.final_norm(hidden_states)
lm_logits = self.lm_head(lm_logits)
loss = None
if labels is not None:
labels = labels.to(lm_logits.device)
# Shift so that tokens < n predict n
shift_logits = lm_logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
# Flatten the tokens
loss_fct = CrossEntropyLoss()
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
if not return_dict:
output = (lm_logits,) + outputs[1:]
return ((loss,) + output) if loss is not None else output
return CausalLMOutputWithCrossAttentions(
loss=loss,
logits=lm_logits,
past_key_values=outputs.past_key_values,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
cross_attentions=outputs.cross_attentions,
)
@auto_docstring(
custom_intro="""
The composite CLVP model with a text encoder, speech encoder and speech decoder model.
"""
)
class ClvpModelForConditionalGeneration(ClvpPreTrainedModel, GenerationMixin):
config: ClvpConfig
def __init__(self, config: ClvpConfig):
super().__init__(config)
if not isinstance(config.text_config, ClvpEncoderConfig):
raise TypeError(
"config.text_config is expected to be of type `ClvpEncoderConfig` but is of type"
f" {type(config.text_config)}."
)
if not isinstance(config.speech_config, ClvpEncoderConfig):
raise TypeError(
"config.speech_config is expected to be of type `ClvpEncoderConfig` but is of type"
f" {type(config.speech_config)}."
)
if not isinstance(config.decoder_config, ClvpDecoderConfig):
raise TypeError(
"config.decoder_config is expected to be of type `ClvpDecoderConfig` but is of type"
f" {type(config.decoder_config)}."
)
self.conditioning_encoder = ClvpConditioningEncoder(config)
self.speech_decoder_model = ClvpForCausalLM(config.decoder_config)
self.text_encoder_model = ClvpEncoder(config.text_config)
self.speech_encoder_model = ClvpEncoder(config.speech_config)
self.logit_scale = nn.Parameter(torch.tensor(self.config.logit_scale_init_value))
# Initialize weights and apply final processing
self.post_init()
# taken from the original repo,
# link : https://github.com/neonbjb/tortoise-tts/blob/4003544b6ff4b68c09856e04d3eff9da26d023c2/tortoise/api.py#L117
def fix_speech_decoder_output(self, speech_ids: torch.LongTensor) -> torch.LongTensor:
"""
This method modifies the output of the decoder model, such as replacing the `eos_token_id` and changing the
last few tokens of each sequence.
Args:
speech_ids (`torch.LongTensor`):
This refers to the output of the decoder model.
"""
decoder_fixing_codes = self.config.decoder_config.decoder_fixing_codes
speech_ids = speech_ids[:, 1:]
stop_token_indices = torch.where(speech_ids == self.speech_decoder_model.config.eos_token_id, 1, 0)
speech_ids = torch.masked_fill(speech_ids, mask=stop_token_indices.bool(), value=decoder_fixing_codes[0])
for i, each_seq_stop_token_index in enumerate(stop_token_indices):
# This means that no stop tokens were found so the sentence was still being generated, in that case we don't need
# to apply any padding so just skip to the next sequence of tokens.
if each_seq_stop_token_index.sum() == 0:
continue
stm = each_seq_stop_token_index.argmax()
speech_ids[i, stm:] = decoder_fixing_codes[0]
if stm - 3 < speech_ids.shape[1]:
speech_ids[i, -3:] = torch.tensor(
[decoder_fixing_codes[1:]], device=speech_ids.device, dtype=torch.long
)
return speech_ids
def get_text_features(
self,
input_ids: Optional[torch.LongTensor] = None,
text_encoder_inputs_embeds: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.LongTensor] = None,
) -> torch.FloatTensor:
r"""
This method can be used to extract text_embeds from a text. The text embeddings obtained by applying the
projection layer to the pooled output of the CLVP text encoder model.
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
provide it.
[What are input IDs?](../glossary#input-ids)
text_encoder_inputs_embeds (`torch.FloatTensor`, *optional*):
inputs_embeds for the text encoder model passed in place of `input_ids`.
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
Returns:
`torch.FloatTensor` of shape `(batch_size, output_dim)`:
The text embeddings obtained by applying the projection layer to the pooled output of the CLVP Text
Model.
Examples:
```python
>>> from transformers import ClvpProcessor, ClvpModelForConditionalGeneration
>>> # Define the Text
>>> text = "This is an example text."
>>> # Define processor and model
>>> processor = ClvpProcessor.from_pretrained("susnato/clvp_dev")
>>> model = ClvpModelForConditionalGeneration.from_pretrained("susnato/clvp_dev")
>>> # Generate processor output and text embeds
>>> processor_output = processor(text=text, return_tensors="pt")
>>> text_embeds = model.get_text_features(input_ids=processor_output["input_ids"])
```
"""
outputs = self.text_encoder_model(
input_ids=input_ids,
inputs_embeds=text_encoder_inputs_embeds,
attention_mask=attention_mask,
)
return outputs[0]
def get_speech_features(
self,
speech_ids: Optional[torch.LongTensor] = None,
input_ids: Optional[torch.LongTensor] = None,
input_features: Optional[torch.FloatTensor] = None,
conditioning_encoder_inputs_embeds: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.Tensor] = None,
generation_config: Optional[GenerationConfig] = None,
**kwargs,
) -> torch.FloatTensor:
r"""
This method can be used to extract speech_embeds. The speech embeddings are obtained by applying the speech
model on speech_ids. If speech_ids is not present but both input_ids and input_features are given then the
decoder model will be used to first generate the speech_ids and then applying the speech model.
Args:
speech_ids (`torch.LongTensor` of shape `(batch_size, num_speech_ids)`, *optional*):
Speech Tokens. Padding will be ignored by default should you provide it. If speech_ids are provided
then input_ids and input_features will be automatically ignored.
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Input text Tokens. Processed from the [`ClvpTokenizer`]. If speech_ids is not provided, then input_ids
and input_features will be used.
conditioning_encoder_inputs_embeds (`torch.FloatTensor`, *optional*):
inputs_embeds for `ClvpConditioningEncoder`. Can be used in place of `input_ids`.
attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding speech token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
generation_config (`GenerationConfig`, *optional*):
generation config to control the generation of speech_ids if they are not provided.
Returns:
`torch.FloatTensor` of shape `(batch_size, output_dim)`:
The speech embeddings obtained by applying the projection layer to the pooled output of the CLVP Speech
Model.
Examples:
```python
>>> import datasets
>>> from transformers import ClvpProcessor, ClvpModelForConditionalGeneration
>>> # Define the Text and Load the Audio (We are taking an audio example from HuggingFace Hub using `datasets` library)
>>> text = "This is an example text."
>>> ds = datasets.load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
>>> ds = ds.cast_column("audio", datasets.Audio(sampling_rate=22050))
>>> audio = ds.sort("id")["audio"][0]
>>> audio_sample, sr = audio["array"], audio["sampling_rate"]
>>> # Define processor and model
>>> processor = ClvpProcessor.from_pretrained("susnato/clvp_dev")
>>> model = ClvpModelForConditionalGeneration.from_pretrained("susnato/clvp_dev")
>>> # Generate processor output and model output
>>> processor_output = processor(raw_speech=audio_sample, sampling_rate=sr, text=text, return_tensors="pt")
>>> speech_embeds = model.get_speech_features(
... input_ids=processor_output["input_ids"], input_features=processor_output["input_features"]
... )
```
"""
if speech_ids is None:
if (input_ids is None and conditioning_encoder_inputs_embeds is None) or input_features is None:
raise ValueError(
"Either speech_ids or input_ids/conditioning_encoder_inputs_embeds and input_features must be provided."
)
if generation_config is None:
generation_config = self.generation_config
generation_config.update(**kwargs)
conditioning_embeds = self.conditioning_encoder(
input_features=input_features,
input_ids=input_ids,
inputs_embeds=conditioning_encoder_inputs_embeds,
attention_mask=attention_mask,
)
speech_ids = self.speech_decoder_model.generate(
conditioning_embeds=conditioning_embeds,
generation_config=generation_config,
)
speech_ids = self.fix_speech_decoder_output(speech_ids[0])
outputs = self.speech_encoder_model(
input_ids=speech_ids,
attention_mask=attention_mask,
)
return outputs[0]
@auto_docstring
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
input_features: Optional[torch.FloatTensor] = None,
conditioning_encoder_inputs_embeds: Optional[torch.FloatTensor] = None,
text_encoder_inputs_embeds: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.LongTensor] = None,
return_loss: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
output_attentions: Optional[bool] = False,
return_dict: Optional[bool] = None,
cache_position: Optional[torch.Tensor] = None,
) -> Union[tuple, ClvpOutput]:
r"""
conditioning_encoder_inputs_embeds (`torch.FloatTensor`, *optional*):
inputs_embeds for `ClvpConditioningEncoder`. Can be used in place of `input_ids`.
text_encoder_inputs_embeds (`torch.FloatTensor`, *optional*):
inputs_embeds for the text encoder model passed in place of `input_ids`.
return_loss (`bool`, *optional*):
Whether or not to return the contrastive loss.
Examples:
```python
>>> import datasets
>>> from transformers import ClvpProcessor, ClvpModelForConditionalGeneration
>>> # Define the Text and Load the Audio (We are taking an audio example from HuggingFace Hub using `datasets` library)
>>> text = "This is an example text."
>>> ds = datasets.load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
>>> ds = ds.cast_column("audio", datasets.Audio(sampling_rate=22050))
>>> audio = ds.sort("id")["audio"][0]
>>> audio_sample, sr = audio["array"], audio["sampling_rate"]
>>> # Define processor and model
>>> processor = ClvpProcessor.from_pretrained("susnato/clvp_dev")
>>> model = ClvpModelForConditionalGeneration.from_pretrained("susnato/clvp_dev")
>>> # processor outputs and model outputs
>>> processor_output = processor(raw_speech=audio_sample, sampling_rate=sr, text=text, return_tensors="pt")
>>> outputs = model(
... input_ids=processor_output["input_ids"],
... input_features=processor_output["input_features"],
... return_dict=True,
... )
```
"""
# Use CLVP model's config for some fields (if specified) instead of those of speech & text components.
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
conditioning_embeds = self.conditioning_encoder(
input_features=input_features,
input_ids=input_ids,
inputs_embeds=conditioning_encoder_inputs_embeds,
attention_mask=attention_mask,
)
decoder_outputs = self.speech_decoder_model(
inputs_embeds=conditioning_embeds,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
cache_position=cache_position,
)
speech_ids = decoder_outputs[0]
# since we will get the embeds of shape `(batch_size, seq_len, embedding_dim)` during the forward pass
# we must convert it to tokens, to make it compaitable with speech_transformer
if speech_ids.ndim == 3:
speech_ids = speech_ids.argmax(2)
speech_ids = self.fix_speech_decoder_output(speech_ids)
speech_outputs = self.speech_encoder_model(
input_ids=speech_ids,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
text_outputs = self.text_encoder_model(
input_ids=input_ids,
inputs_embeds=text_encoder_inputs_embeds,
attention_mask=attention_mask,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
speech_embeds = speech_outputs[0]
text_embeds = text_outputs[0]
# normalized features
speech_embeds = speech_embeds / speech_embeds.norm(p=2, dim=-1, keepdim=True)
text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True)
# cosine similarity as logits
logit_scale = self.logit_scale.exp()
logits_per_text = torch.matmul(text_embeds, speech_embeds.t()) * logit_scale
logits_per_speech = logits_per_text.t()
loss = None
if return_loss:
loss = clvp_loss(logits_per_text)
if not return_dict:
output = (
logits_per_speech,
logits_per_text,
text_embeds,
speech_embeds,
text_outputs[2],
speech_outputs[2],
)
if output_hidden_states:
output += (
decoder_outputs[-1],
text_outputs[-1],
speech_outputs[-1],
)
return ((loss,) + output) if loss is not None else output
return ClvpOutput(
loss=loss,
logits_per_speech=logits_per_speech,
logits_per_text=logits_per_text,
text_embeds=text_embeds,
speech_embeds=speech_embeds,
text_model_output=text_outputs[2],
speech_model_output=speech_outputs[2],
decoder_hidden_states=decoder_outputs.hidden_states,
text_encoder_hidden_states=text_outputs.hidden_states,
speech_encoder_hidden_states=speech_outputs.hidden_states,
)
@torch.no_grad()
def generate(
self,
input_ids: Optional[torch.LongTensor] = None,
input_features: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.LongTensor] = None,
generation_config: Optional[GenerationConfig] = None,
pad_to_max_mel_tokens: Optional[int] = None,
output_hidden_states: Optional[bool] = None,
**kwargs,
):
"""
Generate method for `ClvpModelForConditionalGeneration`, this method calls the `generate` method of
`ClvpForCausalLM` and then uses those generated `speech_ids` to process `text_embeds` and `speech_embeds` using
`ClvpEncoder`.
Args:
input_ids (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
Input text Tokens. Processed from the [`ClvpTokenizer`].
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding text token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
generation_config (`~generation.GenerationConfig`, *optional*):
The generation configuration to be used as base parametrization for the generation call. `**kwargs`
passed to generate matching the attributes of `generation_config` will override them. If
`generation_config` is not provided, the default will be used, which had the following loading
priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model
configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s
default values, whose documentation should be checked to parameterize generation.
pad_to_max_mel_tokens (`int`, *optional*):
Pads generated speech_ids to the specified value. This is to implement the same logic from the official
repo, link: https://github.com/neonbjb/tortoise-tts/blob/80f89987a5abda5e2b082618cd74f9c7411141dc/tortoise/api.py#L430
and to make sure the logits are same.
This does not affect generation quality so please don't consider using it since it is less efficient.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of decoder model, text encoder and speech encoder models.
Returns:
`ClvpOutput` or tuple: A `ClvpOutput` (if `return_dict_in_generate=True` or when
`config.return_dict_in_generate=True`) or a tuple.
"""
# If the input sequences are larger than (self.config.decoder_config.max_text_tokens - 3) then raise error,
# because we need to add 3 tokens ( 1 bos tokens and 2 eos tokens) to the input_ids in ClvpConditioningEncoder to
# properly sample
sequence_length = input_ids.shape[-1]
if sequence_length > (self.config.decoder_config.max_text_tokens - 3):
raise ValueError(
f"Maximum sequence length reached! Found input_ids of length {sequence_length}."
f"Please make sure that the maximum length of input_ids is {self.config.decoder_config.max_text_tokens - 3}"
)
if generation_config is None:
generation_config = self.generation_config
generation_config = copy.deepcopy(generation_config)
model_kwargs = generation_config.update(**kwargs) # All unused kwargs must be model kwargs
generation_config.validate()
self._validate_model_kwargs(model_kwargs.copy())
# pad input_ids as specified in the original repo
# link: https://github.com/neonbjb/tortoise-tts/blob/80f89987a5abda5e2b082618cd74f9c7411141dc/tortoise/api.py#L380
input_ids, attention_mask = _pad_extra_bos_eos_tokens(
input_ids,
attention_mask,
add_bos_token=False,
bos_token_id=self.config.text_config.bos_token_id,
eos_token_id=self.config.text_config.eos_token_id,
)
conditioning_embeds = self.conditioning_encoder(
input_features=input_features,
input_ids=input_ids,
attention_mask=attention_mask,
)
decoder_outputs = self.speech_decoder_model.generate(
conditioning_embeds=conditioning_embeds,
generation_config=generation_config,
output_hidden_states=output_hidden_states,
return_dict=generation_config.return_dict_in_generate,
)
if isinstance(decoder_outputs, ModelOutput):
speech_ids = decoder_outputs.sequences
# pad to pad_to_max_mel_tokens if given, to replicate the original repo logic
# link: https://github.com/neonbjb/tortoise-tts/blob/80f89987a5abda5e2b082618cd74f9c7411141dc/tortoise/api.py#L430
if pad_to_max_mel_tokens is not None:
padding_needed = pad_to_max_mel_tokens - speech_ids.shape[-1]
speech_ids = torch.nn.functional.pad(
speech_ids, (0, padding_needed), value=self.generation_config.eos_token_id
)
speech_ids = self.fix_speech_decoder_output(speech_ids)
speech_outputs = self.speech_encoder_model(
input_ids=speech_ids,
output_hidden_states=output_hidden_states,
return_dict=generation_config.return_dict_in_generate,
)
text_outputs = self.text_encoder_model(
input_ids=input_ids,
attention_mask=attention_mask,
output_hidden_states=output_hidden_states,
return_dict=generation_config.return_dict_in_generate,
)
speech_embeds = speech_outputs[0]
text_embeds = text_outputs[0]
# normalized features
speech_embeds = speech_embeds / speech_embeds.norm(p=2, dim=-1, keepdim=True)
text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True)
# cosine similarity as logits
logit_scale = self.logit_scale.exp()
logits_per_text = torch.matmul(text_embeds, speech_embeds.t()) * logit_scale
logits_per_speech = logits_per_text.t()
if not generation_config.return_dict_in_generate:
output = (
speech_ids,
logits_per_speech,
logits_per_text,
text_embeds,
speech_embeds,
text_outputs[2],
speech_outputs[2],
)
if output_hidden_states:
output += (
decoder_outputs[-1],
text_outputs[-1],
speech_outputs[-1],
)
return output
return ClvpOutput(
speech_ids=speech_ids,
logits_per_speech=logits_per_speech,
logits_per_text=logits_per_text,
text_embeds=text_embeds,
speech_embeds=speech_embeds,
text_model_output=text_outputs[2],
speech_model_output=speech_outputs[2],
decoder_hidden_states=decoder_outputs.hidden_states,
text_encoder_hidden_states=text_outputs.hidden_states,
speech_encoder_hidden_states=speech_outputs.hidden_states,
)
__all__ = [
"ClvpModelForConditionalGeneration",
"ClvpForCausalLM",
"ClvpModel",
"ClvpPreTrainedModel",
"ClvpEncoder",
"ClvpDecoder",
]
| transformers/src/transformers/models/clvp/modeling_clvp.py/0 | {
"file_path": "transformers/src/transformers/models/clvp/modeling_clvp.py",
"repo_id": "transformers",
"token_count": 37891
} | 435 |
# coding=utf-8
# Copyright 2024 Cohere team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# This file is based on the tokenization_llama_fast.py file in transformers
import pickle
from typing import Literal, Union
from tokenizers import processors
from ...tokenization_utils_base import BatchEncoding
from ...tokenization_utils_fast import PreTrainedTokenizerFast
from ...utils import logging
logger = logging.get_logger(__name__)
VOCAB_FILES_NAMES = {"tokenizer_file": "tokenizer.json"}
PRETRAINED_VOCAB_FILES_MAP = {
"tokenizer_file": {
"Cohere/Command-nightly": "https://huggingface.co/Cohere/Command-nightly/blob/main/tokenizer.json",
},
}
# fmt: off
DEFAULT_SYSTEM_PROMPT = "You are Command-R, a brilliant, sophisticated, AI-assistant trained to assist human users by providing thorough responses. You are trained by Cohere."
DEFAULT_RAG_PREAMBLE = """## Task and Context
You help people answer their questions and other requests interactively. You will be asked a very wide array of requests on all kinds of topics. You will be equipped with a wide range of search engines or similar tools to help you, which you use to research your answer. You should focus on serving the user's needs as best you can, which will be wide-ranging.
## Style Guide
Unless the user asks for a different style of answer, you should answer in full sentences, using proper grammar and spelling."""
# fmt: on
class CohereTokenizerFast(PreTrainedTokenizerFast):
"""
Construct a Cohere tokenizer. Based on byte-level Byte-Pair-Encoding.
This uses notably ByteFallback and NFC normalization.
```python
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("CohereForAI/c4ai-command-r-v01")
>>> tokenizer.encode("Hello this is a test")
[5, 28339, 2075, 1801, 1671, 3282]
```
If you want to change the `bos_token` or the `eos_token`, make sure to specify them when initializing the model, or
call `tokenizer.update_post_processor()` to make sure that the post-processing is correctly done (otherwise the
values of the first token and final token of an encoded sequence will not be correct). For more details, checkout
[post-processors] (https://huggingface.co/docs/tokenizers/api/post-processors) documentation.
You can get around that behavior by passing `add_prefix_space=True` when instantiating this tokenizer, but since
the model was not pretrained this way, it might yield a decrease in performance.
<Tip>
When used with `is_split_into_words=True`, this tokenizer needs to be instantiated with `add_prefix_space=True`.
</Tip>
This tokenizer inherits from [`PreTrainedTokenizerFast`] which contains most of the main methods. Users should
refer to this superclass for more information regarding those methods.
Args:
vocab_file (`str`, *optional*):
Path to the vocabulary file.
merges_file (`str`, *optional*):
Path to the merges file.
tokenizer_file (`str`, *optional*):
[tokenizers](https://github.com/huggingface/tokenizers) file (generally has a .json extension) that
contains everything needed to load the tokenizer.
clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):
Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
extra spaces.
unk_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<UNK>"`):
The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
token instead.
bos_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<BOS_TOKEN>"`):
The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
eos_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<|END_OF_TURN_TOKEN|>"`):
The end of sequence token.
add_bos_token (`bool`, *optional*, defaults to `True`):
Whether or not to add an `bos_token` at the start of sequences.
add_eos_token (`bool`, *optional*, defaults to `False`):
Whether or not to add an `eos_token` at the end of sequences.
use_default_system_prompt (`bool`, *optional*, defaults to `False`):
Whether or not the default system prompt for Cohere tokenizer should be used.
add_prefix_space (`bool`, *optional*, defaults to `False`):
Whether or not the tokenizer should automatically add a prefix space
"""
vocab_files_names = VOCAB_FILES_NAMES
pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
padding_side = "left"
model_input_names = ["input_ids", "attention_mask"]
slow_tokenizer_class = None
# No `max_model_input_sizes`
def __init__(
self,
vocab_file=None,
merges_file=None,
tokenizer_file=None,
clean_up_tokenization_spaces=False,
unk_token="<UNK>",
bos_token="<BOS_TOKEN>",
eos_token="<|END_OF_TURN_TOKEN|>",
add_bos_token=True,
add_eos_token=False,
use_default_system_prompt=False,
add_prefix_space=False,
**kwargs,
):
super().__init__(
vocab_file=vocab_file,
merges_file=merges_file,
tokenizer_file=tokenizer_file,
clean_up_tokenization_spaces=clean_up_tokenization_spaces,
unk_token=unk_token,
bos_token=bos_token,
eos_token=eos_token,
add_bos_token=add_bos_token,
add_eos_token=add_eos_token,
use_default_system_prompt=use_default_system_prompt,
add_prefix_space=add_prefix_space,
**kwargs,
)
self._add_bos_token = add_bos_token
self._add_eos_token = add_eos_token
self.update_post_processor()
self.use_default_system_prompt = use_default_system_prompt
self.vocab_file = vocab_file
self.grounded_generation_template = kwargs.pop("grounded_generation_template", None)
self.tool_use_template = kwargs.pop("tool_use_template", None)
# TODO @ArthurZucker this can only work one way for now, to update later-on. Tests should also properly
# check this as they were green before.
pre_tok_state = pickle.dumps(self.backend_tokenizer.pre_tokenizer)
decoder_state = pickle.dumps(self.backend_tokenizer.decoder)
if add_prefix_space:
pre_tok_state = pre_tok_state.replace(b'"add_prefix_space":false', b'"add_prefix_space": true')
decoder_state = decoder_state.replace(b'"add_prefix_space":false', b'"add_prefix_space": true')
self.backend_tokenizer.pre_tokenizer = pickle.loads(pre_tok_state)
self.backend_tokenizer.decoder = pickle.loads(decoder_state)
self.add_prefix_space = add_prefix_space
def _batch_encode_plus(self, *args, **kwargs) -> BatchEncoding:
is_split_into_words = kwargs.get("is_split_into_words", False)
if not (self.add_prefix_space or not is_split_into_words):
raise Exception(
f"You need to instantiate {self.__class__.__name__} with add_prefix_space=True to use it with"
" pretokenized inputs."
)
return super()._batch_encode_plus(*args, **kwargs)
def _encode_plus(self, *args, **kwargs) -> BatchEncoding:
is_split_into_words = kwargs.get("is_split_into_words", False)
if not (self.add_prefix_space or not is_split_into_words):
raise Exception(
f"You need to instantiate {self.__class__.__name__} with add_prefix_space=True to use it with"
" pretokenized inputs."
)
return super()._encode_plus(*args, **kwargs)
def update_post_processor(self):
"""
Updates the underlying post processor with the current `bos_token` and `eos_token`.
"""
bos = self.bos_token
bos_token_id = self.bos_token_id
if bos is None and self.add_bos_token:
raise ValueError("add_bos_token = True but bos_token = None")
eos = self.eos_token
eos_token_id = self.eos_token_id
if eos is None and self.add_eos_token:
raise ValueError("add_eos_token = True but eos_token = None")
single = f"{(bos + ':0 ') if self.add_bos_token else ''}$A:0{(' ' + eos + ':0') if self.add_eos_token else ''}"
pair = f"{single}{(' ' + bos + ':1') if self.add_bos_token else ''} $B:1{(' ' + eos + ':1') if self.add_eos_token else ''}"
special_tokens = []
if self.add_bos_token:
special_tokens.append((bos, bos_token_id))
if self.add_eos_token:
special_tokens.append((eos, eos_token_id))
self._tokenizer.post_processor = processors.TemplateProcessing(
single=single, pair=pair, special_tokens=special_tokens
)
@property
def add_eos_token(self):
return self._add_eos_token
@property
def add_bos_token(self):
return self._add_bos_token
@add_eos_token.setter
def add_eos_token(self, value):
self._add_eos_token = value
self.update_post_processor()
@add_bos_token.setter
def add_bos_token(self, value):
self._add_bos_token = value
self.update_post_processor()
def apply_tool_use_template(
self,
conversation: Union[list[dict[str, str]]],
tools: list[dict],
**kwargs,
) -> Union[str, list[int]]:
"""Create a Command-R tool-use prompt.
Once rendered, the prompt instructs the model to generate a list of actions to perform on a set of user supplied tools
to help carry out the user's requests.
Conceptually, this works in the same way as `apply_chat_format`, but takes an additional `tools` parameter.
Converts a chat in the form of a list of dictionaries with `"role"` and `"content"` keys and a list of available
tools for the model to use into a prompt string, or a list of token ids.
This method will use the tokenizer's `default_tool_use_template` template specified at the class level.
You can override the default template using the `tool_use_template` kwarg but the quality of your results may decrease.
Args:
conversation (Union[list[dict[str, str]]]): A list of dicts
with "role" and "content" keys, representing the chat history so far.
tools (list[Dict]): a list of tools to render into the prompt for the model to choose from.
See an example at the bottom of the docstring.
The format should be:
* name (str): The name of the tool to be called. Valid names contain only the characters a-z,
A-Z, 0-9, _ and must not begin with a digit.
* description (str): The description of what the tool does, the model uses the description to
choose when and how to call the function.
* parameter_definitions (list[Dict]): The input parameters of the tool. Accepts a dictionary
where the key is the name of the parameter and the value is the parameter spec.
Valid parameter names contain only the characters a-z, A-Z, 0-9, _ and must not begin with a digit.
Parameter specs are as follows:
* description (str): The description of the parameter.
* type (str): the type of the parameter - most effective for python builtin data types, such as 'str', 'bool'
* required: boolean: Denotes whether the parameter is always present (required) or not. Defaults to not required.
add_generation_prompt (bool, *optional*): Whether to end the prompt with the token(s) that indicate
the start of an assistant message. This is useful when you want to generate a response from the model.
Note that this argument will be passed to the chat template, and so it must be supported in the
template for this argument to have any effect.
tokenize (`bool`, defaults to `True`):
Whether to tokenize the output. If `False`, the output will be a string.
padding (`bool`, defaults to `False`):
Whether to pad sequences to the maximum length. Has no effect if tokenize is `False`.
truncation (`bool`, defaults to `False`):
Whether to truncate sequences at the maximum length. Has no effect if tokenize is `False`.
max_length (`int`, *optional*):
Maximum length (in tokens) to use for padding or truncation. Has no effect if tokenize is `False`. If
not specified, the tokenizer's `max_length` attribute will be used as a default.
return_tensors (`str` or [`~utils.TensorType`], *optional*):
If set, will return tensors of a particular framework. Has no effect if tokenize is `False`. Acceptable
values are:
- `'tf'`: Return TensorFlow `tf.Tensor` objects.
- `'pt'`: Return PyTorch `torch.Tensor` objects.
- `'np'`: Return NumPy `np.ndarray` objects.
- `'jax'`: Return JAX `jnp.ndarray` objects.
return_dict (`bool`, *optional*, defaults to `False`):
Whether to return a dictionary with named outputs. Has no effect if tokenize is `False`.
**tokenizer_kwargs: Additional kwargs to pass to the tokenizer.
Returns:
`str`: A rendered prompt string.
or if tokenize=True:
`list[int]`: A list of token ids representing the tokenized chat so far, including control tokens. This
output is ready to pass to the model, either directly or via methods like `generate()`.
Examples:
```python
>> tokenizer = CohereTokenizerFast.from_pretrained("CohereForAI/c4ai-command-r-v01")
>> tools = [
{
"name": "internet_search",
"description": "Returns a list of relevant document snippets for a textual query retrieved from the internet",
"parameter_definitions": {
"query": {
"description": "Query to search the internet with",
"type": "str",
"required": True
}
}
},
{
"name': "directly_answer",
"description": "Calls a standard (un-augmented) AI chatbot to generate a response given the conversation history",
"parameter_definitions": {}
}
]
>> conversation = [
{"role": "user", "content": "Whats the biggest penguin in the world?"}
]
>> # render the prompt, ready for user to inspect, or for input into the model:
>> prompt = tokenizer.apply_tool_use_template(conversation, tools=tools, tokenize=False, add_generation_prompt=True)
>> print(prompt)
<BOS_TOKEN><|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|># Safety Preamble
The instructions in this section override those in the task description and style guide sections. Don't answer questions that are harmful or immoral.
# System Preamble
## Basic Rules
You are a powerful conversational AI trained by Cohere to help people. You are augmented by a number of tools, and your job is to use and consume the output of these tools to best help the user. You will see a conversation history between yourself and a user, ending with an utterance from the user. You will then see a specific instruction instructing you what kind of response to generate. When you answer the user's requests, you cite your sources in your answers, according to those instructions.
# User Preamble
## Task and Context
You help people answer their questions and other requests interactively. You will be asked a very wide array of requests on all kinds of topics. You will be equipped with a wide range of search engines or similar tools to help you, which you use to research your answer. You should focus on serving the user's needs as best you can, which will be wide-ranging.
## Style Guide
Unless the user asks for a different style of answer, you should answer in full sentences, using proper grammar and spelling.
## Available Tools
Here is a list of tools that you have available to you:
\\`\\`\\`python
def internet_search(query: str) -> list[Dict]:
\"\"\"Returns a list of relevant document snippets for a textual query retrieved from the internet
Args:
query (str): Query to search the internet with
\"\"\"
pass
\\`\\`\\`
\\`\\`\\`python
def directly_answer() -> list[Dict]:
\"\"\"Calls a standard (un-augmented) AI chatbot to generate a response given the conversation history
\"\"\"
pass
\\`\\`\\`<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Whats the biggest penguin in the world?<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>Write 'Action:' followed by a json-formatted list of actions that you want to perform in order to produce a good response to the user's last input. You can use any of the supplied tools any number of times, but you should aim to execute the minimum number of necessary actions for the input. You should use the `directly-answer` tool if calling the other tools is unnecessary. The list of actions you want to call should be formatted as a list of json objects, for example:
\\`\\`\\`json
[
{
"tool_name": title of the tool in the specification,
"parameters": a dict of parameters to input into the tool as they are defined in the specs, or {} if it takes no parameters
}
]\\`\\`\\`<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>
```
>> inputs = tokenizer.encode(prompt, add_special_tokens=False, return_tensors='pt')
>> outputs = model.generate(inputs, max_new_tokens=128)
>> print(tokenizer.decode(outputs[0]))
Action: ```json
[
{
"tool_name": "internet_search",
"parameters": {
"query": "biggest penguin in the world"
}
}
]
```
"""
return self.apply_chat_template(
conversation,
chat_template="tool_use",
tools=tools,
**kwargs,
)
def apply_grounded_generation_template(
self,
conversation: Union[list[dict[str, str]]],
documents: list[dict],
citation_mode: Literal["fast", "accurate"] = "accurate",
**kwargs,
) -> Union[str, list[int]]:
"""Create a Command-R grounded generation (aka RAG) prompt.
Once rendered, the prompt instructs the model to generate a response with citations in, based on supplied documents.
Conceptually, this works in the same way as `apply_chat_format`, but takes additional `documents`
and parameter `citation_mode` parameters.
Converts a list of dictionaries with `"role"` and `"content"` keys and a list of
documents for the model to ground its response on into a prompt string, or a list of token ids.
This method will use the tokenizer's `grounded_generation_template` template specified at the class level.
You can override the default template using the `grounded_generation_template` kwarg but the quality of your results may decrease.
Args:
conversation (Union[list[dict[str, str]]]): A list of dicts
with "role" and "content" keys, representing the chat history so far.
documents (list[dict[str, str]): A list of dicts, representing documents or tool outputs to ground your
generation on. A document is a semistructured dict, with a string to string mapping. Common fields are
`url`, `title`, `snippet` etc but should be descriptive of the key. They will get rendered into the prompt.
citation_mode: either "accurate" (prompt the model to generate an answer first, then rewrite it with citation
spans in) or "fast", where the prompt instructs the model to generate an answer with citations in directly.
The former has higher quality citations, the latter requires fewer tokens to be generated.
add_generation_prompt (bool, *optional*): Whether to end the prompt with the token(s) that indicate
the start of an assistant message. This is useful when you want to generate a response from the model.
Note that this argument will be passed to the chat template, and so it must be supported in the
template for this argument to have any effect.
tokenize (`bool`, defaults to `True`):
Whether to tokenize the output. If `False`, the output will be a string.
padding (`bool`, defaults to `False`):
Whether to pad sequences to the maximum length. Has no effect if tokenize is `False`.
truncation (`bool`, defaults to `False`):
Whether to truncate sequences at the maximum length. Has no effect if tokenize is `False`.
max_length (`int`, *optional*):
Maximum length (in tokens) to use for padding or truncation. Has no effect if tokenize is `False`. If
not specified, the tokenizer's `max_length` attribute will be used as a default.
return_tensors (`str` or [`~utils.TensorType`], *optional*):
If set, will return tensors of a particular framework. Has no effect if tokenize is `False`. Acceptable
values are:
- `'tf'`: Return TensorFlow `tf.Tensor` objects.
- `'pt'`: Return PyTorch `torch.Tensor` objects.
- `'np'`: Return NumPy `np.ndarray` objects.
- `'jax'`: Return JAX `jnp.ndarray` objects.
return_dict (`bool`, *optional*, defaults to `False`):
Whether to return a dictionary with named outputs. Has no effect if tokenize is `False`.
**tokenizer_kwargs: Additional kwargs to pass to the tokenizer.
Returns:
`str`: A rendered prompt string.
or if tokenize=True:
`list[int]`: A list of token ids representing the tokenized chat so far, including control tokens. This
output is ready to pass to the model, either directly or via methods like `generate()`.
Examples:
```python
>> tokenizer = CohereTokenizerFast.from_pretrained('CohereForAI/c4ai-command-r-v01')
>> # define documents:
>> documents = [
{ "title": "Tall penguins", "text": "Emperor penguins are the tallest." },
{ "title": "Penguin habitats", "text": "Emperor penguins only live in Antarctica."}
]
>> # define a conversation:
>> conversation = [
{"role": "user", "content": "Whats the biggest penguin in the world?"}
]
>> # render the prompt, ready for user to inspect, or for input into the model:
>> grounded_generation_prompt = tokenizer.apply_grounded_generation_template(conversation, documents=documents, tokenize=False, add_generation_prompt=True)
>> print(grounded_generation_prompt)
<BOS_TOKEN><|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|># Safety Preamble
The instructions in this section override those in the task description and style guide sections. Don't answer questions that are harmful or immoral.
## Basic Rules
You are a powerful conversational AI trained by Cohere to help people. You are augmented by a number of tools, and your job is to use and consume the output of these tools to best help the user. You will see a conversation history between yourself and a user, ending with an utterance from the user. You will then see a specific instruction instructing you what kind of response to generate. When you answer the user's requests, you cite your sources in your answers, according to those instructions.
# User Preamble
## Task and Context
You help people answer their questions and other requests interactively. You will be asked a very wide array of requests on all kinds of topics. You will be equipped with a wide range of search engines or similar tools to help you, which you use to research your answer. You should focus on serving the user's needs as best you can, which will be wide-ranging.
## Style Guide
Unless the user asks for a different style of answer, you should answer in full sentences, using proper grammar and spelling.<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Whats the biggest penguin in the world?<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|><results>
Document: 0
title: Tall penguins
text: Emperor penguins are the tallest.
Document: 1
title: Penguin habitats
text: Emperor penguins only live in Antarctica.
</results><|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>Carefully perform the following instructions, in order, starting each with a new line.
Firstly, Decide which of the retrieved documents are relevant to the user's last input by writing 'Relevant Documents:' followed by comma-separated list of document numbers. If none are relevant, you should instead write 'None'.
Secondly, Decide which of the retrieved documents contain facts that should be cited in a good answer to the user's last input by writing 'Cited Documents:' followed a comma-separated list of document numbers. If you dont want to cite any of them, you should instead write 'None'.
Thirdly, Write 'Answer:' followed by a response to the user's last input in high quality natural english. Use the retrieved documents to help you. Do not insert any citations or grounding markup.
Finally, Write 'Grounded answer:' followed by a response to the user's last input in high quality natural english. Use the symbols <co: doc> and </co: doc> to indicate when a fact comes from a document in the search result, e.g <co: 0>my fact</co: 0> for a fact from document 0.<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>'''
```
>> inputs = tokenizer.encode(prompt, add_special_tokens=False, return_tensors='pt')
>> outputs = model.generate(inputs, max_new_tokens=128)
>> print(tokenizer.decode(outputs[0]))
Relevant Documents: 0,1
Cited Documents: 0,1
Answer: The Emperor Penguin is the tallest or biggest penguin in the world. It is a bird that lives only in Antarctica and grows to a height of around 122 centimetres.
Grounded answer: The <co: 0>Emperor Penguin</co: 0> is the <co: 0>tallest</co: 0> or biggest penguin in the world. It is a bird that <co: 1>lives only in Antarctica</co: 1> and <co: 0>grows to a height of around 122 centimetres.</co: 0>
"""
return self.apply_chat_template(
conversation,
chat_template="rag",
documents=documents,
citation_mode=citation_mode,
**kwargs,
)
# TODO ArthurZ let's rely on the template processor instead, refactor all fast tokenizers
def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
bos_token_id = [self.bos_token_id] if self.add_bos_token else []
eos_token_id = [self.eos_token_id] if self.add_eos_token else []
output = bos_token_id + token_ids_0 + eos_token_id
if token_ids_1 is not None:
output = output + bos_token_id + token_ids_1 + eos_token_id
return output
__all__ = ["CohereTokenizerFast"]
| transformers/src/transformers/models/cohere/tokenization_cohere_fast.py/0 | {
"file_path": "transformers/src/transformers/models/cohere/tokenization_cohere_fast.py",
"repo_id": "transformers",
"token_count": 10999
} | 436 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.