text
stringlengths 7
1.24M
| id
stringlengths 14
166
| metadata
dict | __index_level_0__
int64 0
519
|
|---|---|---|---|
/// Text Generation Inference Webserver
pub mod config;
pub mod infer;
pub mod server;
pub mod validation;
#[cfg(feature = "kserve")]
mod kserve;
pub mod logging;
pub mod usage_stats;
use serde::{Deserialize, Serialize};
use tracing::warn;
use utoipa::ToSchema;
use validation::Validation;
#[derive(PartialEq)]
pub enum Attention {
Paged,
FlashDecoding,
FlashInfer,
}
impl Attention {
pub fn block_size(&self) -> u32 {
match self {
Attention::FlashDecoding => 256,
Attention::FlashInfer => 1,
Attention::Paged => 16,
}
}
}
#[derive(Debug)]
pub struct ParseError;
impl std::fmt::Display for ParseError {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
write!(f, "Cannot parse attention value")
}
}
impl std::error::Error for ParseError {}
impl std::str::FromStr for Attention {
type Err = ParseError;
fn from_str(s: &str) -> Result<Self, Self::Err> {
match s {
"paged" => Ok(Attention::Paged),
"flashdecoding" => Ok(Attention::FlashDecoding),
"flashinfer" => Ok(Attention::FlashInfer),
_ => Err(ParseError),
}
}
}
#[derive(Clone, Deserialize, ToSchema)]
pub(crate) struct VertexInstance {
#[schema(example = "What is Deep Learning?")]
pub inputs: String,
#[schema(nullable = true, default = "null", example = "null")]
pub parameters: Option<GenerateParameters>,
}
#[derive(Deserialize, ToSchema)]
pub(crate) struct VertexRequest {
#[serde(rename = "instances")]
pub instances: Vec<VertexInstance>,
}
#[derive(Clone, Deserialize, ToSchema, Serialize)]
pub(crate) struct VertexResponse {
pub predictions: Vec<String>,
}
/// Hub type
#[derive(Clone, Debug, Deserialize)]
pub struct HubModelInfo {
#[serde(rename(deserialize = "id"))]
pub model_id: String,
pub sha: Option<String>,
pub pipeline_tag: Option<String>,
}
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
pub struct ChatTemplate {
name: String,
template: String,
}
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
#[serde(untagged)]
pub enum ChatTemplateVersions {
Single(String),
Multiple(Vec<ChatTemplate>),
}
use std::path::Path;
#[derive(Debug, Clone, Serialize, Deserialize, Default)]
pub struct HubTokenizerConfig {
pub chat_template: Option<ChatTemplateVersions>,
pub completion_template: Option<String>,
pub bos_token: Option<TokenizerConfigToken>,
pub eos_token: Option<TokenizerConfigToken>,
pub tokenizer_class: Option<String>,
pub add_bos_token: Option<bool>,
pub add_eos_token: Option<bool>,
}
impl HubTokenizerConfig {
pub fn from_file<P: AsRef<Path>>(filename: P) -> Option<Self> {
std::fs::read_to_string(filename)
.ok()
.and_then(|content| serde_json::from_str(&content).ok())
}
}
#[derive(Debug, Clone, Deserialize, Serialize, PartialEq)]
#[serde(untagged)]
pub enum TokenizerConfigToken {
String(String),
Object { content: String },
}
impl TokenizerConfigToken {
pub fn as_str(&self) -> &str {
match self {
TokenizerConfigToken::String(s) => s,
TokenizerConfigToken::Object { content } => content,
}
}
}
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(tag = "processor_class")]
pub enum HubPreprocessorConfig {
Idefics2Processor(Idefics2Preprocessor),
}
impl HubPreprocessorConfig {
pub fn from_file<P: AsRef<std::path::Path>>(filename: P) -> Option<Self> {
let content = std::fs::read_to_string(filename).ok()?;
serde_json::from_str(&content).ok()
}
}
#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Idefics2Preprocessor {
#[serde(default)]
do_image_splitting: bool,
}
#[derive(Debug, Clone, Deserialize, Default)]
pub struct HubProcessorConfig {
pub chat_template: Option<ChatTemplateVersions>,
pub image_seq_len: usize,
pub processor_class: Option<String>,
}
impl HubProcessorConfig {
pub fn from_file<P: AsRef<Path>>(filename: P) -> Option<Self> {
std::fs::read_to_string(filename)
.ok()
.and_then(|content| serde_json::from_str(&content).ok())
}
}
#[derive(Clone, Debug, Deserialize, ToSchema, Serialize)]
#[serde(tag = "type", content = "value")]
pub(crate) enum GrammarType {
/// A string that represents a [JSON Schema](https://json-schema.org/).
///
/// JSON Schema is a declarative language that allows to annotate JSON documents
/// with types and descriptions.
#[serde(rename = "json")]
#[serde(alias = "json_object")]
#[schema(example = json ! ({"properties": {"location":{"type": "string"}}}))]
Json(serde_json::Value),
#[serde(rename = "regex")]
Regex(String),
}
#[derive(Clone, Debug, Serialize, ToSchema)]
pub struct Info {
/// Model info
#[schema(example = "bigscience/blomm-560m")]
pub model_id: String,
#[schema(nullable = true, example = "e985a63cdc139290c5f700ff1929f0b5942cced2")]
pub model_sha: Option<String>,
// #[schema(example = "torch.float16")]
// pub model_dtype: String,
// #[schema(example = "cuda")]
// pub model_device_type: String,
#[schema(nullable = true, example = "text-generation")]
pub model_pipeline_tag: Option<String>,
/// Router Parameters
#[schema(example = "128")]
pub max_concurrent_requests: usize,
#[schema(example = "2")]
pub max_best_of: usize,
#[schema(example = "4")]
pub max_stop_sequences: usize,
#[schema(example = "1024")]
pub max_input_tokens: usize,
#[schema(example = "2048")]
pub max_total_tokens: usize,
#[schema(example = "2")]
pub validation_workers: usize,
#[schema(example = "32")]
pub max_client_batch_size: usize,
/// Router Info
#[schema(example = "text-generation-router")]
pub router: &'static str,
#[schema(example = "0.5.0")]
pub version: &'static str,
#[schema(nullable = true, example = "null")]
pub sha: Option<&'static str>,
#[schema(nullable = true, example = "null")]
pub docker_label: Option<&'static str>,
}
#[derive(Clone, Debug, Deserialize, ToSchema, Default)]
pub(crate) struct GenerateParameters {
/// Generate best_of sequences and return the one if the highest token logprobs.
#[serde(default)]
#[schema(exclusive_minimum = 0, nullable = true, default = "null", example = 1)]
pub best_of: Option<usize>,
/// The value used to module the logits distribution.
#[serde(default)]
#[schema(
exclusive_minimum = 0.0,
nullable = true,
default = "null",
example = 0.5
)]
pub temperature: Option<f32>,
/// The parameter for repetition penalty. 1.0 means no penalty.
/// See [this paper](https://arxiv.org/pdf/1909.05858.pdf) for more details.
#[serde(default)]
#[schema(
exclusive_minimum = 0.0,
nullable = true,
default = "null",
example = 1.03
)]
pub repetition_penalty: Option<f32>,
/// The parameter for frequency penalty. 1.0 means no penalty
/// Penalize new tokens based on their existing frequency in the text so far,
/// decreasing the model's likelihood to repeat the same line verbatim.
#[serde(default)]
#[schema(
exclusive_minimum = -2.0,
nullable = true,
default = "null",
example = 0.1
)]
pub frequency_penalty: Option<f32>,
/// The number of highest probability vocabulary tokens to keep for top-k-filtering.
#[serde(default)]
#[schema(exclusive_minimum = 0, nullable = true, default = "null", example = 10)]
pub top_k: Option<i32>,
/// Top-p value for nucleus sampling.
#[serde(default)]
#[schema(
exclusive_minimum = 0.0,
maximum = 1.0,
nullable = true,
default = "null",
example = 0.95
)]
pub top_p: Option<f32>,
/// Typical Decoding mass
/// See [Typical Decoding for Natural Language Generation](https://arxiv.org/abs/2202.00666) for more information.
#[serde(default)]
#[schema(
exclusive_minimum = 0.0,
maximum = 1.0,
nullable = true,
default = "null",
example = 0.95
)]
pub typical_p: Option<f32>,
/// Activate logits sampling.
#[serde(default)]
#[schema(default = "false", example = true)]
pub do_sample: bool,
/// Maximum number of tokens to generate.
#[serde(default = "default_max_new_tokens")]
#[schema(nullable = true, default = "100", example = "20")]
pub max_new_tokens: Option<u32>,
/// Whether to prepend the prompt to the generated text
#[serde(default)]
#[schema(nullable = true, default = "null", example = false)]
pub return_full_text: Option<bool>,
/// Stop generating tokens if a member of `stop` is generated.
#[serde(default)]
#[schema(inline, max_items = 4, example = json ! (["photographer"]))]
pub stop: Vec<String>,
/// Truncate inputs tokens to the given size.
#[serde(default)]
#[schema(nullable = true, default = "null", example = "null")]
pub truncate: Option<usize>,
/// Watermarking with [A Watermark for Large Language Models](https://arxiv.org/abs/2301.10226).
#[serde(default)]
#[schema(default = "false", example = true)]
pub watermark: bool,
/// Whether to return generation details.
#[serde(default)]
#[schema(default = "true")]
pub details: bool,
/// Whether to return decoder input token logprobs and ids.
#[serde(default)]
#[schema(default = "false")]
pub decoder_input_details: bool,
/// Random sampling seed.
#[serde(default)]
#[schema(
exclusive_minimum = 0,
nullable = true,
default = "null",
example = "null"
)]
pub seed: Option<u64>,
/// The number of highest probability vocabulary tokens to keep for top-n-filtering.
#[serde(default)]
#[schema(exclusive_minimum = 0, nullable = true, default = "null", example = 5)]
pub top_n_tokens: Option<u32>,
/// Grammar constraints for the generation.
#[serde(default)]
#[schema(nullable = true, default = "null", example = "null")]
pub grammar: Option<GrammarType>,
/// Lora adapter id
#[serde(default)]
#[schema(nullable = true, default = "null", example = "null")]
pub adapter_id: Option<String>,
}
fn default_max_new_tokens() -> Option<u32> {
Some(100)
}
fn default_parameters() -> GenerateParameters {
GenerateParameters {
best_of: None,
temperature: None,
repetition_penalty: None,
frequency_penalty: None,
top_k: None,
top_p: None,
typical_p: None,
do_sample: true,
max_new_tokens: default_max_new_tokens(),
return_full_text: None,
stop: Vec::new(),
truncate: None,
watermark: false,
details: false,
decoder_input_details: false,
seed: None,
top_n_tokens: None,
grammar: None,
adapter_id: None,
}
}
#[derive(Clone, Deserialize, Serialize, ToSchema, Debug)]
#[serde(try_from = "PromptDeserializer")]
pub struct Prompt(pub Vec<String>);
#[derive(Deserialize)]
#[serde(untagged)]
enum PromptDeserializer {
Single(String),
Multiple(Vec<String>),
}
impl TryFrom<PromptDeserializer> for Prompt {
type Error = String;
fn try_from(value: PromptDeserializer) -> Result<Self, Self::Error> {
match value {
PromptDeserializer::Single(s) => Ok(Prompt(vec![s])),
PromptDeserializer::Multiple(v) => {
if v.is_empty() {
Err(
"Empty array detected. Do not use an empty array for the prompt."
.to_string(),
)
} else {
Ok(Prompt(v))
}
}
}
}
}
#[derive(Clone, Deserialize, Serialize, ToSchema, Debug)]
pub struct CompletionRequest {
/// UNUSED
#[schema(example = "mistralai/Mistral-7B-Instruct-v0.2")]
/// ID of the model to use. See the model endpoint compatibility table for details on which models work with the Chat API.
pub model: Option<String>,
/// The prompt to generate completions for.
#[schema(example = "What is Deep Learning?")]
pub prompt: Prompt,
/// The maximum number of tokens that can be generated in the chat completion.
#[serde(default)]
#[schema(default = "32")]
pub max_tokens: Option<u32>,
/// What sampling temperature to use, between 0 and 2. Higher values like 0.8 will make the output more random, while
/// lower values like 0.2 will make it more focused and deterministic. We generally recommend altering this or `top_p` but not both.
#[serde(default)]
#[schema(nullable = true, example = 1.0)]
pub temperature: Option<f32>,
/// An alternative to sampling with temperature, called nucleus sampling, where the model considers the results of the
/// tokens with top_p probability mass. So 0.1 means only the tokens comprising the top 10% probability mass are considered.
#[serde(default)]
#[schema(nullable = true, example = 0.95)]
pub top_p: Option<f32>,
#[serde(default = "bool::default")]
pub stream: bool,
#[schema(nullable = true, example = 42)]
pub seed: Option<u64>,
/// The text to append to the prompt. This is useful for completing sentences or generating a paragraph of text.
/// please see the completion_template field in the model's tokenizer_config.json file for completion template.
#[serde(default)]
pub suffix: Option<String>,
#[serde(default)]
pub repetition_penalty: Option<f32>,
/// Number between -2.0 and 2.0. Positive values penalize new tokens based on their existing frequency in the text so far,
/// decreasing the model's likelihood to repeat the same line verbatim.
#[serde(default)]
#[schema(example = "1.0")]
pub frequency_penalty: Option<f32>,
/// Up to 4 sequences where the API will stop generating further tokens.
#[serde(default)]
#[schema(nullable = true, example = "null")]
pub stop: Option<Vec<String>>,
}
#[derive(Clone, Serialize, ToSchema)]
#[serde(tag = "object")]
enum Completion {
#[serde(rename = "text_completion")]
Chunk(Chunk),
#[serde(rename = "text_completion")]
Final(CompletionFinal),
}
#[derive(Clone, Deserialize, Serialize, ToSchema, Default)]
pub(crate) struct CompletionFinal {
pub id: String,
#[schema(example = "1706270835")]
pub created: u64,
#[schema(example = "mistralai/Mistral-7B-Instruct-v0.2")]
pub model: String,
pub system_fingerprint: String,
pub choices: Vec<CompletionComplete>,
pub usage: Usage,
}
#[derive(Clone, Deserialize, Serialize, ToSchema)]
pub(crate) struct CompletionComplete {
pub index: u32,
pub text: String,
pub logprobs: Option<Vec<f32>>,
pub finish_reason: String,
}
#[derive(Clone, Deserialize, Serialize, ToSchema)]
pub(crate) struct Chunk {
pub id: String,
pub created: u64,
pub choices: Vec<CompletionComplete>,
pub model: String,
pub system_fingerprint: String,
}
#[derive(Clone, Deserialize, Serialize, ToSchema)]
pub(crate) struct ChatCompletion {
pub id: String,
#[schema(example = "1706270835")]
pub created: u64,
#[schema(example = "mistralai/Mistral-7B-Instruct-v0.2")]
pub model: String,
pub system_fingerprint: String,
pub choices: Vec<ChatCompletionComplete>,
pub usage: Usage,
}
#[derive(Clone, Deserialize, Serialize, ToSchema)]
pub(crate) struct ChatCompletionComplete {
pub index: u32,
pub message: OutputMessage,
pub logprobs: Option<ChatCompletionLogprobs>,
pub finish_reason: String,
}
#[derive(Clone, Deserialize, Serialize, ToSchema)]
pub(crate) struct ChatCompletionLogprobs {
content: Vec<ChatCompletionLogprob>,
}
impl From<(Token, Vec<Token>)> for ChatCompletionLogprobs {
fn from(value: (Token, Vec<Token>)) -> Self {
let (token, top_tokens) = value;
Self {
content: vec![ChatCompletionLogprob {
token: token.text,
logprob: token.logprob,
top_logprobs: top_tokens
.into_iter()
.map(|t| ChatCompletionTopLogprob {
token: t.text,
logprob: t.logprob,
})
.collect(),
}],
}
}
}
impl From<(Vec<Token>, Vec<Vec<Token>>)> for ChatCompletionLogprobs {
fn from(value: (Vec<Token>, Vec<Vec<Token>>)) -> Self {
let (tokens, top_tokens) = value;
// Create an iterator that produces None for top_tokens once it's exhausted
let top_tokens_iter = top_tokens
.into_iter()
.map(Some)
.chain(std::iter::repeat(None));
let content = tokens
.into_iter()
.zip(top_tokens_iter)
.map(|(t, top_t_option)| ChatCompletionLogprob {
token: t.text,
logprob: t.logprob,
top_logprobs: match top_t_option {
Some(top_t) => top_t
.into_iter()
.map(|t| ChatCompletionTopLogprob {
token: t.text,
logprob: t.logprob,
})
.collect(),
None => vec![], // Handle the case where there are no top tokens
},
})
.collect();
Self { content }
}
}
#[derive(Clone, Deserialize, Serialize, ToSchema)]
pub(crate) struct ChatCompletionLogprob {
token: String,
logprob: f32,
top_logprobs: Vec<ChatCompletionTopLogprob>,
}
#[derive(Clone, Deserialize, Serialize, ToSchema)]
pub(crate) struct ChatCompletionTopLogprob {
token: String,
logprob: f32,
}
#[derive(Clone, Deserialize, Serialize, ToSchema, Default)]
pub(crate) struct Usage {
pub prompt_tokens: u32,
pub completion_tokens: u32,
pub total_tokens: u32,
}
#[derive(Clone, Serialize, ToSchema)]
#[serde(tag = "object")]
enum CompletionType {
#[serde(rename = "chat.completion.chunk")]
ChatCompletionChunk(ChatCompletionChunk),
#[serde(rename = "chat.completion")]
ChatCompletion(ChatCompletion),
}
impl ChatCompletion {
pub(crate) fn new(
model: String,
system_fingerprint: String,
output: Option<String>,
created: u64,
details: Details,
return_logprobs: bool,
tool_calls: Option<Vec<ToolCall>>,
) -> Self {
let message = match (output, tool_calls) {
(Some(content), None) => OutputMessage::ChatMessage(TextMessage {
role: "assistant".into(),
content,
}),
(None, Some(tool_calls)) => OutputMessage::ToolCall(ToolCallMessage {
role: "assistant".to_string(),
tool_calls,
}),
(Some(output), Some(_)) => {
warn!("Received both chat and tool call");
OutputMessage::ChatMessage(TextMessage {
role: "assistant".into(),
content: output,
})
}
(None, None) => {
warn!("Didn't receive an answer");
OutputMessage::ChatMessage(TextMessage {
role: "assistant".into(),
content: "".to_string(),
})
}
};
Self {
id: String::new(),
created,
model,
system_fingerprint,
choices: vec![ChatCompletionComplete {
index: 0,
message,
logprobs: return_logprobs
.then(|| ChatCompletionLogprobs::from((details.tokens, details.top_tokens))),
finish_reason: details.finish_reason.format(true),
}],
usage: Usage {
prompt_tokens: details.prefill.len() as u32,
completion_tokens: details.generated_tokens,
total_tokens: details.prefill.len() as u32 + details.generated_tokens,
},
}
}
}
#[derive(Clone, Serialize, ToSchema)]
pub(crate) struct ChatCompletionChunk {
pub id: String,
#[schema(example = "1706270978")]
pub created: u64,
#[schema(example = "mistralai/Mistral-7B-Instruct-v0.2")]
pub model: String,
pub system_fingerprint: String,
pub choices: Vec<ChatCompletionChoice>,
}
#[derive(Clone, Serialize, ToSchema)]
pub(crate) struct ChatCompletionChoice {
pub index: u32,
pub delta: ChatCompletionDelta,
pub logprobs: Option<ChatCompletionLogprobs>,
pub finish_reason: Option<String>,
}
#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
pub struct ToolCallDelta {
#[schema(example = "assistant")]
role: String,
tool_calls: DeltaToolCall,
}
#[derive(Clone, Debug, Serialize, ToSchema)]
#[serde(untagged)]
enum ChatCompletionDelta {
Chat(TextMessage),
Tool(ToolCallDelta),
}
#[derive(Clone, Deserialize, Serialize, ToSchema, Debug, PartialEq)]
pub(crate) struct DeltaToolCall {
pub index: u32,
pub id: String,
pub r#type: String,
pub function: Function,
}
#[derive(Clone, Deserialize, Serialize, ToSchema, Debug, PartialEq)]
pub(crate) struct Function {
pub name: Option<String>,
pub arguments: String,
}
#[allow(clippy::too_many_arguments)]
impl ChatCompletionChunk {
pub(crate) fn new(
model: String,
system_fingerprint: String,
delta: Option<String>,
tool_calls: Option<Vec<String>>,
created: u64,
logprobs: Option<ChatCompletionLogprobs>,
finish_reason: Option<String>,
) -> Self {
let delta = match (delta, tool_calls) {
(Some(delta), _) => ChatCompletionDelta::Chat(TextMessage {
role: "assistant".to_string(),
content: delta,
}),
(None, Some(tool_calls)) => ChatCompletionDelta::Tool(ToolCallDelta {
role: "assistant".to_string(),
tool_calls: DeltaToolCall {
index: 0,
id: String::new(),
r#type: "function".to_string(),
function: Function {
name: None,
arguments: tool_calls[0].to_string(),
},
},
}),
(None, None) => ChatCompletionDelta::Chat(TextMessage {
role: "assistant".to_string(),
content: "".to_string(),
}),
};
Self {
id: String::new(),
created,
model,
system_fingerprint,
choices: vec![ChatCompletionChoice {
index: 0,
delta,
logprobs,
finish_reason,
}],
}
}
}
#[derive(Clone, Deserialize, ToSchema, Serialize)]
pub(crate) struct ChatRequest {
#[schema(example = "mistralai/Mistral-7B-Instruct-v0.2")]
/// [UNUSED] ID of the model to use. See the model endpoint compatibility table for details on which models work with the Chat API.
pub model: Option<String>,
/// A list of messages comprising the conversation so far.
#[schema(example = "[{\"role\": \"user\", \"content\": \"What is Deep Learning?\"}]")]
pub messages: Vec<Message>,
/// Number between -2.0 and 2.0. Positive values penalize new tokens based on their existing frequency in the text so far,
/// decreasing the model's likelihood to repeat the same line verbatim.
#[serde(default)]
#[schema(example = "1.0")]
pub frequency_penalty: Option<f32>,
/// UNUSED
/// Modify the likelihood of specified tokens appearing in the completion. Accepts a JSON object that maps tokens
/// (specified by their token ID in the tokenizer) to an associated bias value from -100 to 100. Mathematically,
/// the bias is added to the logits generated by the model prior to sampling. The exact effect will vary per model,
/// but values between -1 and 1 should decrease or increase likelihood of selection; values like -100 or 100 should
/// result in a ban or exclusive selection of the relevant token.
#[serde(default)]
pub logit_bias: Option<Vec<f32>>,
/// Whether to return log probabilities of the output tokens or not. If true, returns the log probabilities of each
/// output token returned in the content of message.
#[serde(default)]
#[schema(example = "false")]
pub logprobs: Option<bool>,
/// An integer between 0 and 5 specifying the number of most likely tokens to return at each token position, each with
/// an associated log probability. logprobs must be set to true if this parameter is used.
#[serde(default)]
#[schema(example = "5")]
pub top_logprobs: Option<u32>,
/// The maximum number of tokens that can be generated in the chat completion.
#[serde(default)]
#[schema(example = "32")]
pub max_tokens: Option<u32>,
/// UNUSED
/// How many chat completion choices to generate for each input message. Note that you will be charged based on the
/// number of generated tokens across all of the choices. Keep n as 1 to minimize costs.
#[serde(default)]
#[schema(nullable = true, example = "2")]
pub n: Option<u32>,
/// Number between -2.0 and 2.0. Positive values penalize new tokens based on whether they appear in the text so far,
/// increasing the model's likelihood to talk about new topics
#[serde(default)]
#[schema(nullable = true, example = 0.1)]
pub presence_penalty: Option<f32>,
/// Up to 4 sequences where the API will stop generating further tokens.
#[serde(default)]
#[schema(nullable = true, example = "null")]
pub stop: Option<Vec<String>>,
#[serde(default = "bool::default")]
pub stream: bool,
#[schema(nullable = true, example = 42)]
pub seed: Option<u64>,
/// What sampling temperature to use, between 0 and 2. Higher values like 0.8 will make the output more random, while
/// lower values like 0.2 will make it more focused and deterministic.
///
/// We generally recommend altering this or `top_p` but not both.
#[serde(default)]
#[schema(nullable = true, example = 1.0)]
pub temperature: Option<f32>,
/// An alternative to sampling with temperature, called nucleus sampling, where the model considers the results of the
/// tokens with top_p probability mass. So 0.1 means only the tokens comprising the top 10% probability mass are considered.
#[serde(default)]
#[schema(nullable = true, example = 0.95)]
pub top_p: Option<f32>,
/// A list of tools the model may call. Currently, only functions are supported as a tool. Use this to provide a list of
/// functions the model may generate JSON inputs for.
#[serde(default)]
#[schema(nullable = true, example = "null")]
pub tools: Option<Vec<Tool>>,
/// A prompt to be appended before the tools
#[serde(default)]
#[schema(
nullable = true,
example = "Given the functions available, please respond with a JSON for a function call with its proper arguments that best answers the given prompt. Respond in the format {name: function name, parameters: dictionary of argument name and its value}.Do not use variables."
)]
pub tool_prompt: Option<String>,
/// A specific tool to use. If not provided, the model will default to use any of the tools provided in the tools parameter.
#[serde(default)]
#[schema(nullable = true, example = "null")]
pub tool_choice: ToolChoice,
/// Response format constraints for the generation.
///
/// NOTE: A request can use `response_format` OR `tools` but not both.
#[serde(default)]
#[schema(nullable = true, default = "null", example = "null")]
pub response_format: Option<GrammarType>,
/// A guideline to be used in the chat_template
#[serde(default)]
#[schema(nullable = true, default = "null", example = "null")]
pub guideline: Option<String>,
}
pub fn default_tool_prompt() -> String {
"\nGiven the functions available, please respond with a JSON for a function call with its proper arguments that best answers the given prompt. Respond in the format {name: function name, parameters: dictionary of argument name and its value}.Do not use variables.\n".to_string()
}
#[derive(Clone, Debug, Deserialize, PartialEq, Serialize, ToSchema)]
#[serde(untagged)]
pub enum ToolType {
OneOf,
FunctionName(String),
Function { function: FunctionName },
NoTool,
}
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, ToSchema)]
pub struct FunctionName {
pub name: String,
}
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, Default, ToSchema)]
#[serde(from = "ToolTypeDeserializer")]
pub struct ToolChoice(pub Option<ToolType>);
#[derive(Deserialize)]
#[serde(untagged)]
enum ToolTypeDeserializer {
String(String),
ToolType(ToolType),
}
impl From<ToolTypeDeserializer> for ToolChoice {
fn from(value: ToolTypeDeserializer) -> Self {
match value {
ToolTypeDeserializer::String(s) => match s.as_str() {
"none" => ToolChoice(Some(ToolType::NoTool)),
"auto" => ToolChoice(Some(ToolType::OneOf)),
_ => ToolChoice(Some(ToolType::FunctionName(s))),
},
ToolTypeDeserializer::ToolType(tool_type) => ToolChoice(Some(tool_type)),
}
}
}
#[derive(Debug, Deserialize, Serialize, ToSchema, PartialEq)]
pub struct JsonSchemaTool {
#[serde(flatten)]
functions_map: FunctionsMap,
properties: Properties,
}
#[derive(Debug, Serialize, Deserialize, PartialEq)]
struct FunctionsMap {
#[serde(rename = "$functions")]
functions: std::collections::HashMap<String, serde_json::Value>,
}
#[derive(Debug, Serialize, Deserialize, PartialEq)]
struct FunctionRef {
#[serde(rename = "$ref")]
ref_path: String,
}
#[derive(Debug, Serialize, Deserialize, PartialEq)]
struct Properties {
#[serde(serialize_with = "serialize_function")]
function: Vec<FunctionRef>,
}
fn serialize_function<S>(functions: &Vec<FunctionRef>, serializer: S) -> Result<S::Ok, S::Error>
where
S: serde::Serializer,
{
use serde::ser::SerializeStruct;
let mut state = serializer.serialize_struct("Function", 1)?;
state.serialize_field("anyOf", functions)?;
state.end()
}
#[derive(Clone, Debug, Deserialize, Serialize, ToSchema, Default, PartialEq)]
pub(crate) struct FunctionDefinition {
#[serde(default)]
pub description: Option<String>,
pub name: String,
#[serde(alias = "parameters")]
pub arguments: serde_json::Value,
}
#[derive(Clone, Debug, Deserialize, Serialize, ToSchema)]
pub(crate) struct Tool {
// The type of the tool. Currently, only 'function' is supported.
#[schema(example = "function")]
pub r#type: String,
// Grab the tool as generic JSON for debugging purposes.
pub function: FunctionDefinition,
}
#[derive(Clone, Serialize, Deserialize, Default)]
pub(crate) struct ChatTemplateInputs<'a> {
messages: Vec<TextMessage>,
bos_token: Option<&'a str>,
eos_token: Option<&'a str>,
add_generation_prompt: bool,
tools: Option<Vec<Tool>>,
guideline: Option<&'a str>,
}
#[derive(Clone, Deserialize, Serialize, ToSchema, Default, Debug, PartialEq)]
pub(crate) struct ToolCall {
pub id: String,
pub r#type: String,
pub function: FunctionDefinition,
}
#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
pub struct Url {
url: String,
}
#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
#[serde(tag = "type")]
#[serde(rename_all = "snake_case")]
pub enum MessageChunk {
Text { text: String },
ImageUrl { image_url: Url },
}
#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
pub struct Message {
#[schema(example = "user")]
role: String,
#[schema(example = "My name is David and I")]
pub content: MessageContent,
#[serde(default, skip_serializing_if = "Option::is_none")]
#[schema(example = "\"David\"")]
name: Option<String>,
}
#[derive(Clone, Deserialize, Serialize, ToSchema, Debug, PartialEq)]
#[serde(untagged)]
pub enum MessageContent {
SingleText(String),
MultipleChunks(Vec<MessageChunk>),
}
// Pushing a chunk to a single text message will convert it to a multiple chunks message
impl MessageContent {
pub fn push(&mut self, chunk: MessageChunk) {
match self {
MessageContent::SingleText(text) => {
*self = MessageContent::MultipleChunks(vec![
MessageChunk::Text { text: text.clone() },
chunk,
]);
}
MessageContent::MultipleChunks(chunks) => {
chunks.push(chunk);
}
}
}
}
#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
pub struct TextMessage {
#[schema(example = "user")]
pub role: String,
#[schema(example = "My name is David and I")]
pub content: String,
}
impl From<Message> for TextMessage {
fn from(value: Message) -> Self {
TextMessage {
role: value.role,
content: match value.content {
MessageContent::SingleText(text) => text,
MessageContent::MultipleChunks(chunks) => chunks
.into_iter()
.map(|chunk| match chunk {
MessageChunk::Text { text } => text,
MessageChunk::ImageUrl { image_url } => format!("", image_url.url),
})
.collect::<Vec<_>>()
.join(""),
},
}
}
}
#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
pub struct ToolCallMessage {
#[schema(example = "assistant")]
role: String,
tool_calls: Vec<ToolCall>,
}
#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
#[serde(untagged)]
pub(crate) enum OutputMessage {
ChatMessage(TextMessage),
ToolCall(ToolCallMessage),
}
#[derive(Clone, Debug, Deserialize, ToSchema)]
pub(crate) struct GenerateRequest {
#[schema(example = "My name is Olivier and I")]
pub inputs: String,
#[serde(default = "default_parameters")]
pub parameters: GenerateParameters,
/// This is used internally because some requests
/// already contain the templated input therefore
/// we shouldn't add the special tokens.
#[serde(default = "default_true", skip)]
pub add_special_tokens: bool,
}
fn default_true() -> bool {
true
}
#[derive(Clone, Debug, Deserialize, ToSchema)]
pub(crate) struct CompatGenerateRequest {
#[schema(example = "My name is Olivier and I")]
pub inputs: String,
#[serde(default = "default_parameters")]
pub parameters: GenerateParameters,
#[serde(default)]
#[schema(default = "false")]
pub stream: bool,
}
impl From<CompatGenerateRequest> for GenerateRequest {
fn from(req: CompatGenerateRequest) -> Self {
Self {
inputs: req.inputs,
add_special_tokens: true,
parameters: req.parameters,
}
}
}
#[derive(Debug, Serialize, ToSchema)]
pub struct PrefillToken {
#[schema(example = 0)]
pub id: u32,
#[schema(example = "test")]
pub text: String,
#[schema(nullable = true, example = - 0.34)]
pub logprob: f32,
}
#[derive(Debug, Serialize, ToSchema, Clone)]
pub struct Token {
#[schema(example = 0)]
pub id: u32,
#[schema(example = "test")]
pub text: String,
#[schema(nullable = true, example = - 0.34)]
pub logprob: f32,
#[schema(example = "false")]
pub special: bool,
}
#[derive(Debug, Serialize, ToSchema)]
pub struct SimpleToken {
#[schema(example = 0)]
id: u32,
#[schema(example = "test")]
text: String,
#[schema(example = 0)]
start: usize,
#[schema(example = 2)]
stop: usize,
}
#[derive(Debug, Serialize, ToSchema)]
#[serde(rename_all(serialize = "snake_case"))]
#[schema(example = "Length")]
pub enum FinishReason {
#[schema(rename = "length")]
Length,
#[serde(rename = "eos_token")]
#[schema(rename = "eos_token")]
EndOfSequenceToken,
#[schema(rename = "stop_sequence")]
StopSequence,
}
impl std::fmt::Display for FinishReason {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
match self {
FinishReason::Length => write!(f, "length"),
FinishReason::EndOfSequenceToken => write!(f, "eos_token"),
FinishReason::StopSequence => write!(f, "stop_sequence"),
}
}
}
impl FinishReason {
pub fn format(&self, use_stop: bool) -> String {
match self {
FinishReason::EndOfSequenceToken if use_stop => "stop".to_string(),
_ => self.to_string(),
}
}
}
#[derive(Serialize, ToSchema)]
pub(crate) struct BestOfSequence {
#[schema(example = "test")]
pub generated_text: String,
#[schema(example = "length")]
pub finish_reason: FinishReason,
#[schema(example = 1)]
pub generated_tokens: u32,
#[schema(nullable = true, example = 42)]
pub seed: Option<u64>,
pub prefill: Vec<PrefillToken>,
pub tokens: Vec<Token>,
#[serde(skip_serializing_if = "Vec::is_empty")]
pub top_tokens: Vec<Vec<Token>>,
}
#[derive(Serialize, ToSchema)]
pub(crate) struct Details {
#[schema(example = "length")]
pub finish_reason: FinishReason,
#[schema(example = 1)]
pub generated_tokens: u32,
#[schema(nullable = true, example = 42)]
pub seed: Option<u64>,
pub prefill: Vec<PrefillToken>,
pub tokens: Vec<Token>,
#[serde(skip_serializing_if = "Option::is_none")]
pub best_of_sequences: Option<Vec<BestOfSequence>>,
#[serde(skip_serializing_if = "Vec::is_empty")]
pub top_tokens: Vec<Vec<Token>>,
}
#[derive(Serialize, ToSchema)]
pub(crate) struct GenerateResponse {
#[schema(example = "test")]
pub generated_text: String,
#[serde(skip_serializing_if = "Option::is_none")]
pub details: Option<Details>,
}
#[derive(Serialize, ToSchema)]
pub(crate) struct ChatTokenizeResponse {
pub(crate) tokenize_response: TokenizeResponse,
pub(crate) templated_text: String,
}
#[derive(Serialize, ToSchema)]
#[serde(transparent)]
pub(crate) struct TokenizeResponse(Vec<SimpleToken>);
#[derive(Serialize, ToSchema)]
pub(crate) struct StreamDetails {
#[schema(example = "length")]
pub finish_reason: FinishReason,
#[schema(example = 1)]
pub generated_tokens: u32,
#[schema(nullable = true, example = 42)]
pub seed: Option<u64>,
#[schema(example = 1)]
pub input_length: u32,
}
#[derive(Serialize, ToSchema)]
pub(crate) struct StreamResponse {
pub index: u32,
pub token: Token,
#[serde(skip_serializing_if = "Vec::is_empty")]
pub top_tokens: Vec<Token>,
#[schema(nullable = true, default = "null", example = "test")]
pub generated_text: Option<String>,
#[schema(nullable = true, default = "null")]
pub details: Option<StreamDetails>,
}
#[derive(Serialize, ToSchema)]
pub(crate) struct ErrorResponse {
pub error: String,
pub error_type: String,
}
#[derive(Serialize, Deserialize, ToSchema)]
pub(crate) struct ModelInfo {
#[schema(example = "gpt2")]
pub id: String,
#[schema(example = "model")]
pub object: String,
#[schema(example = 1686935002)]
pub created: u64,
#[schema(example = "openai")]
pub owned_by: String,
}
#[derive(Serialize, Deserialize, ToSchema)]
pub(crate) struct ModelsInfo {
#[schema(example = "list")]
pub object: String,
pub data: Vec<ModelInfo>,
}
impl Default for ModelsInfo {
fn default() -> Self {
ModelsInfo {
object: "list".to_string(),
data: Vec::new(),
}
}
}
#[cfg(test)]
mod tests {
use super::*;
use serde_json::json;
use tokenizers::Tokenizer;
pub(crate) async fn get_tokenizer() -> Tokenizer {
let api = hf_hub::api::sync::Api::new().unwrap();
let repo = api.model("gpt2".to_string());
let filename = repo.get("tokenizer.json").unwrap();
Tokenizer::from_file(filename).unwrap()
}
#[test]
fn test_hub_nested_tokens_tokenizer_config() {
// this is a subset of the tokenizer.json file
// in this case we expect the tokens to be encoded as simple strings
let json_content = r#"{
"chat_template": "test",
"bos_token": "<｜begin▁of▁sentence｜>",
"eos_token": "<｜end▁of▁sentence｜>"
}"#;
let config: HubTokenizerConfig = serde_json::from_str(json_content).unwrap();
// check that we successfully parsed the tokens
assert_eq!(
config.chat_template,
Some(ChatTemplateVersions::Single("test".to_string()))
);
assert_eq!(
config.bos_token,
Some(TokenizerConfigToken::String(
"<｜begin▁of▁sentence｜>".to_string()
))
);
assert_eq!(
config.eos_token,
Some(TokenizerConfigToken::String(
"<｜end▁of▁sentence｜>".to_string()
))
);
// in this case we expect the tokens to be encoded as structured tokens
// we want the content of the structured token
let json_content = r#"{
"chat_template": "test",
"bos_token": {
"__type": "AddedToken",
"content": "<｜begin▁of▁sentence｜>",
"lstrip": false,
"normalized": true,
"rstrip": false,
"single_word": false
},
"eos_token": {
"__type": "AddedToken",
"content": "<｜end▁of▁sentence｜>",
"lstrip": false,
"normalized": true,
"rstrip": false,
"single_word": false
}
}"#;
let config: HubTokenizerConfig = serde_json::from_str(json_content).unwrap();
// check that we successfully parsed the tokens
assert_eq!(
config.chat_template,
Some(ChatTemplateVersions::Single("test".to_string()))
);
assert_eq!(
config.bos_token,
Some(TokenizerConfigToken::Object {
content: "<｜begin▁of▁sentence｜>".to_string()
})
);
assert_eq!(
config.eos_token,
Some(TokenizerConfigToken::Object {
content: "<｜end▁of▁sentence｜>".to_string()
})
);
}
#[test]
fn test_chat_simple_string() {
let json = json!({
"model": "",
"messages": [{
"role": "user",
"content": "What is Deep Learning?"
}]
});
let request: ChatRequest = serde_json::from_str(json.to_string().as_str()).unwrap();
assert_eq!(
request.messages[0],
Message {
role: "user".to_string(),
content: MessageContent::SingleText("What is Deep Learning?".to_string()),
name: None
}
);
}
#[test]
fn test_message_content_append() {
let mut content = MessageContent::SingleText("Initial text".to_string());
let chunk = MessageChunk::Text {
text: "Additional text".to_string(),
};
content.push(chunk);
match content {
MessageContent::MultipleChunks(chunks) => {
assert_eq!(chunks.len(), 2);
assert_eq!(
chunks[0],
MessageChunk::Text {
text: "Initial text".to_string()
}
);
assert_eq!(
chunks[1],
MessageChunk::Text {
text: "Additional text".to_string()
}
);
}
_ => panic!("Expected MultipleChunks, but got a different variant"),
}
}
#[test]
fn test_chat_request() {
let json = json!({
"model": "",
"messages": [{
"role": "user",
"content": [
{"type": "text", "text": "Whats in this image?"},
{"type": "image_url", "image_url": {"url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png"}},
]
}]
});
let request: ChatRequest = serde_json::from_str(json.to_string().as_str()).unwrap();
assert_eq!(
request.messages[0],
Message{
role: "user".to_string(),
content: MessageContent::MultipleChunks(vec![
MessageChunk::Text { text: "Whats in this image?".to_string() },
MessageChunk::ImageUrl { image_url: Url { url: "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png".to_string() }},
]),
name: None
}
);
}
#[test]
fn text_message_convert() {
let message = Message{
role: "user".to_string(),
content: MessageContent::MultipleChunks(vec![
MessageChunk::Text { text: "Whats in this image?".to_string() },
MessageChunk::ImageUrl { image_url: Url { url: "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png".to_string() } }
]),
name: None
};
let textmsg: TextMessage = message.into();
assert_eq!(textmsg.content, "Whats in this image?");
}
#[test]
fn openai_output() {
let message = OutputMessage::ChatMessage(TextMessage {
role: "assistant".to_string(),
content: "This is the answer".to_string(),
});
let serialized = serde_json::to_string(&message).unwrap();
assert_eq!(
serialized,
r#"{"role":"assistant","content":"This is the answer"}"#
);
let message = OutputMessage::ToolCall(ToolCallMessage {
role: "assistant".to_string(),
tool_calls: vec![ToolCall {
id: "0".to_string(),
r#type: "function".to_string(),
function: FunctionDefinition {
description: None,
name: "myfn".to_string(),
arguments: json!({
"format": "csv"
}),
},
}],
});
let serialized = serde_json::to_string(&message).unwrap();
assert_eq!(
serialized,
r#"{"role":"assistant","tool_calls":[{"id":"0","type":"function","function":{"description":null,"name":"myfn","arguments":{"format":"csv"}}}]}"#
);
}
}
|
text-generation-inference/router/src/lib.rs/0
|
{
"file_path": "text-generation-inference/router/src/lib.rs",
"repo_id": "text-generation-inference",
"token_count": 20587
}
| 237
|
install-flashinfer:
pip install flashinfer==0.1.5 -i https://flashinfer.ai/whl/cu124/torch2.4
|
text-generation-inference/server/Makefile-flashinfer/0
|
{
"file_path": "text-generation-inference/server/Makefile-flashinfer",
"repo_id": "text-generation-inference",
"token_count": 42
}
| 238
|
// Adapted from turboderp exllama: https://github.com/turboderp/exllama
#ifndef _q4_matrix_cuh
#define _q4_matrix_cuh
#include <cuda_runtime.h>
#include <cuda_fp16.h>
#include <cstdint>
class Q4Matrix
{
public:
int device;
int height;
int width;
int groups;
int groupsize;
uint32_t* cuda_qweight = NULL;
uint32_t* cuda_qzeros = NULL;
half* cuda_scales = NULL;
uint32_t* cuda_x_map = NULL;
Q4Matrix
(
const int _height,
const int _width,
const int _groups,
uint32_t* _qweight,
uint32_t* _qzeros,
half* _scales,
uint32_t* _g_idx,
const int _device
);
~Q4Matrix();
void reconstruct(half* out);
private:
void make_sequential(const uint32_t* cpu_g_idx);
};
void g_q4_keep_matrix(Q4Matrix* m);
void g_q4_free_matrices();
#endif
|
text-generation-inference/server/exllama_kernels/exllama_kernels/cuda_func/q4_matrix.cuh/0
|
{
"file_path": "text-generation-inference/server/exllama_kernels/exllama_kernels/cuda_func/q4_matrix.cuh",
"repo_id": "text-generation-inference",
"token_count": 420
}
| 239
|
#ifndef _q_matrix_cuh
#define _q_matrix_cuh
#include <cuda_runtime.h>
#include <cuda_fp16.h>
#include <cstdint>
#include <cstdio>
#define MAX_SUPERGROUPS 16
class QMatrix
{
public:
int device;
bool is_gptq;
int height;
int width;
int groups;
int gptq_groupsize;
int rows_8;
int rows_6;
int rows_5;
int rows_4;
int rows_3;
int rows_2;
uint32_t* cuda_q_weight = NULL;
uint16_t* cuda_q_perm = NULL;
uint16_t* cuda_q_invperm = NULL;
uint32_t* cuda_q_scale = NULL;
half* cuda_q_scale_max = NULL;
uint16_t* cuda_q_groups = NULL;
uint16_t* cuda_q_group_map = NULL;
uint32_t* cuda_gptq_qzeros = NULL;
half* cuda_gptq_scales = NULL;
half* temp_dq;
bool failed;
QMatrix
(
const int _device,
const int _height,
const int _width,
const int _groups,
uint32_t* _q_weight,
uint16_t* _q_perm,
uint16_t* _q_invperm,
uint32_t* _q_scale,
half* _q_scale_max,
uint16_t* _q_groups,
uint16_t* _q_group_map,
uint32_t* _gptq_qzeros,
half* _gptq_scales,
uint32_t* _gptq_g_idx,
half* _temp_dq
);
~QMatrix();
void reconstruct(half* out);
bool make_sequential(const uint32_t* cpu_g_idx);
private:
};
#endif
|
text-generation-inference/server/exllamav2_kernels/exllamav2_kernels/cuda/q_matrix.cuh/0
|
{
"file_path": "text-generation-inference/server/exllamav2_kernels/exllamav2_kernels/cuda/q_matrix.cuh",
"repo_id": "text-generation-inference",
"token_count": 702
}
| 240
|
import pytest
import os
from text_generation_server.pb import generate_pb2
os.environ["USE_PREFIX_CACHING"] = "1"
os.environ["ATTENTION"] = "flashinfer"
@pytest.fixture
def default_pb_parameters():
return generate_pb2.NextTokenChooserParameters(
temperature=1.0,
repetition_penalty=1.0,
top_k=0,
top_p=1.0,
typical_p=1.0,
do_sample=False,
)
@pytest.fixture
def default_pb_stop_parameters():
return generate_pb2.StoppingCriteriaParameters(stop_sequences=[], max_new_tokens=10)
|
text-generation-inference/server/tests/conftest.py/0
|
{
"file_path": "text-generation-inference/server/tests/conftest.py",
"repo_id": "text-generation-inference",
"token_count": 237
}
| 241
|
# Origin: https://github.com/predibase/lorax
# Path: lorax/server/lorax_server/adapters/lora.py
# License: Apache License Version 2.0, January 2004
from collections import defaultdict
from dataclasses import dataclass
from typing import Dict, List, Optional, Set, Tuple, Type, Union
import torch
from peft import LoraConfig as _LoraConfig
from torch.distributed import ProcessGroup
from text_generation_server.adapters.config import AdapterConfig, ModuleMap
from text_generation_server.adapters.weights import (
AdapterBatchMetadata,
AdapterWeights,
BatchAdapterWeights,
)
from text_generation_server.utils.sgmv import (
BGMV_MAX_RANK,
MAX_RANK_CUSTOM,
get_tmp_tensors,
orient_for_rank,
pad_rank,
use_cutlass_shrink,
)
def get_start_stop_idxs_for_rank(offset, size, rank, world_size):
block_size = size // world_size
start = offset + rank * block_size
stop = offset + (rank + 1) * block_size
return start, stop
def shard_on_dim(
t: torch.Tensor, dim: int, process_group: torch.distributed.ProcessGroup
):
world_size = process_group.size()
rank = process_group.rank()
size = t.shape[dim]
start, stop = get_start_stop_idxs_for_rank(0, size, rank, world_size)
if dim == 0:
tensor = t[start:stop]
elif dim == 1:
tensor = t[:, start:stop]
else:
raise NotImplementedError("Let's make that generic when needed")
return tensor
def shard_lora_weights(
weights_a: List[torch.Tensor],
weights_b: List[torch.Tensor],
split_dim: int,
process_group: ProcessGroup,
) -> Tuple[List[torch.Tensor], List[torch.Tensor]]:
# [hidden_size, r]
weights_a = [
shard_on_dim(w, dim=split_dim, process_group=process_group) for w in weights_a
]
# [r, hidden_size]
weights_b = [shard_on_dim(w, dim=1, process_group=process_group) for w in weights_b]
return weights_a, weights_b
@dataclass
class LoraConfig(AdapterConfig):
r: int
target_modules: Optional[Union[List[str], str]]
fan_in_fan_out: bool
lora_alpha: int
use_rslora: bool
def map_weights_for_model(
self,
adapter_weights: Dict[int, AdapterWeights],
weight_names: Tuple[str],
) -> Tuple[ModuleMap, Set[str]]:
adapter_weight_names = set()
module_map = {}
for weight_name in weight_names:
lora_a_name = f"base_model.model.{weight_name}.lora_A.weight"
lora_b_name = f"base_model.model.{weight_name}.lora_B.weight"
if lora_a_name not in adapter_weights or lora_b_name not in adapter_weights:
continue
module_map[weight_name] = {
"lora_A": (adapter_weights[lora_a_name], lora_a_name),
"lora_B": (adapter_weights[lora_b_name], lora_b_name),
}
adapter_weight_names.add(lora_a_name)
adapter_weight_names.add(lora_b_name)
return module_map, adapter_weight_names
@classmethod
def load(cls, adapter_id: str, api_token: str) -> "LoraConfig":
hf_config = _LoraConfig.from_pretrained(adapter_id, token=api_token)
return cls(
base_model_name_or_path=hf_config.base_model_name_or_path,
r=hf_config.r,
target_modules=hf_config.target_modules,
fan_in_fan_out=hf_config.fan_in_fan_out,
lora_alpha=hf_config.lora_alpha,
use_rslora=(
hf_config.use_rslora if hasattr(hf_config, "use_rslora") else False
),
)
class LoraWeights(AdapterWeights):
"""LoRA weights for a single adapter merged across all layers."""
def __init__(
self,
weights_a: List[torch.Tensor],
weights_b: List[torch.Tensor],
adapter_config: LoraConfig,
):
self.lora_a_r = weights_a[0].size(1) if len(weights_a) > 0 else 1
self.lora_b_r = weights_b[0].size(0) if len(weights_a) > 0 else 1
self._use_cutlass_shrink = use_cutlass_shrink(self.lora_a_r)
self._is_transposed = False
# [num_layers, hidden_size, r]
weights_a = [orient_for_rank(w, w.size(1)).contiguous() for w in weights_a]
self._weights_a = torch.stack(weights_a)
# [num_layers, r, hidden_size]
self._weights_b = torch.stack(weights_b)
self.adapter_config = adapter_config
@property
def weights_a(self) -> torch.Tensor:
if self._is_transposed:
self._transpose_weights()
return self._weights_a
@property
def weights_b(self) -> torch.Tensor:
if self._is_transposed:
self._transpose_weights()
return self._weights_b
@property
def weights_a_t(self) -> torch.Tensor:
if not self._is_transposed:
self._transpose_weights()
return self._weights_a
@property
def weights_b_t(self) -> torch.Tensor:
if not self._is_transposed:
self._transpose_weights()
return self._weights_b
def _transpose_weights(self):
if self._use_cutlass_shrink:
# If we're not using the cutlass shrink, then both SGMV and BGMV use the same orientation
self._weights_a = self._weights_a.transpose(1, 2).contiguous()
self._weights_b = self._weights_b.transpose(1, 2).contiguous()
self._is_transposed = not self._is_transposed
@classmethod
def get_batch_types(cls) -> List[Type[BatchAdapterWeights]]:
return [BatchLoraWeights]
# prepare pre-loaded lora weights for use in the model.
#
# this method processes and organizes lora weights for a specific layer type across all layers:
# - uses `config` (LoraConfig) to apply lora-specific settings like scaling factor.
# - retrieves weights from `module_map` based on the `layer_type`.
# - processes `nlayers` number of layers.
# - converts weights to the specified `dtype`.
# - shards weights across `world_size` number of processes using the `process_group`.
# - maps weights to specific layers using `target_to_layer`.
# - tracks `unused_weight_names` to identify any unused weights.
#
# the method handles weight transposition, scaling, and padding to ensure compatibility
# with SGMV or BGMV operations.
@classmethod
def prepare_weights(
cls,
config: LoraConfig,
module_map: Dict[str, Dict],
layer_type: str,
unused_weight_names: Set[str],
nlayers: int,
dtype: torch.dtype,
world_size: int,
process_group: ProcessGroup,
target_to_layer: Dict[str, Tuple[str, torch.Tensor]],
) -> Optional[AdapterWeights]:
lora_a_list = [None] * nlayers
lora_b_list = [None] * nlayers
for layer_id in range(nlayers):
key = (layer_id, layer_type)
weight_name, layer = target_to_layer[key]
base_weight = layer.base_layer.linear.weight
base_device = base_weight.device
if weight_name not in module_map:
# There is no LoRA weight for this layer type in the adapter
return None
lora_a, lora_a_name = module_map[weight_name]["lora_A"]
lora_a = lora_a.to(base_device, dtype)
lora_b, lora_b_name = module_map[weight_name]["lora_B"]
lora_b = lora_b.to(base_device, dtype)
scale = get_scaling_factor(
config.lora_alpha,
config.r,
uses_rslora=config.use_rslora,
)
unused_weight_names.discard(lora_a_name)
unused_weight_names.discard(lora_b_name)
# Merge scaling factor into lora_b due to associativity of matrix multiplication:
# (A * B) * C = A * (B * C)
lora_a_list[layer_id] = lora_a.transpose(0, 1)
lora_b_list[layer_id] = lora_b.transpose(0, 1) * scale
# pad lora ranks to be compatible with sgmv
lora_a_list = [pad_rank(w, dim=1, world_size=world_size) for w in lora_a_list]
lora_b_list = [pad_rank(w, dim=0, world_size=world_size) for w in lora_b_list]
if lora_a_list:
# update rank if it was padded
padded_rank = lora_a_list[0].size(1)
config.r = padded_rank
return LoraWeights(
*shard_lora_weights(
weights_a=lora_a_list,
weights_b=lora_b_list,
split_dim=0 if layer_type in {"o_proj", "down_proj", "lm_head"} else 1,
process_group=process_group,
),
config,
)
@dataclass
class RankSegments:
rank: int
lora_a_ptr: torch.Tensor
lora_b_ptr: torch.Tensor
# prefill (sgmv)
tmp_shrink: torch.Tensor
tmp_expand: torch.Tensor
segment_starts: torch.Tensor
segment_ends: torch.Tensor
# decode (bgmv)
indices: torch.Tensor
@dataclass
class BatchLoraWeights(BatchAdapterWeights):
lora_a: Dict[int, torch.Tensor]
lora_b: Dict[int, torch.Tensor]
adapter_index_configs: Dict[int, LoraConfig]
rank_data: Dict[int, RankSegments]
use_sgmv: bool
def has_adapter(self, adapter_index: int) -> bool:
return adapter_index in self.adapter_index_configs
def can_vectorize(self, pg: ProcessGroup) -> bool:
return all(
rank_data.rank // pg.size() <= MAX_RANK_CUSTOM
for rank_data in self.rank_data.values()
)
@classmethod
def load(
self,
adapter_weights: Dict[int, AdapterWeights],
meta: AdapterBatchMetadata,
prefill: bool,
prefill_head_indices: Optional[torch.Tensor],
) -> Optional["BatchLoraWeights"]:
adapter_weights = {k: _convert_lora(v) for k, v in adapter_weights.items()}
adapter_weights = {
k: v for k, v in adapter_weights.items() if isinstance(v, LoraWeights)
}
if not adapter_weights:
return None
first_weights = next(iter(adapter_weights.values()))
device = first_weights.weights_a.device
segment_indices = meta.segment_indices
lora_a = {
idx: adapter_weights[idx].weights_a
for idx in segment_indices
if idx in adapter_weights
}
lora_b = {
idx: adapter_weights[idx].weights_b
for idx in segment_indices
if idx in adapter_weights
}
max_rank = max(
(
adapter_weights[idx].lora_a_r
for idx in segment_indices
if idx in adapter_weights
),
default=0,
)
if prefill or max_rank > BGMV_MAX_RANK:
use_sgmv = True
lora_a_ptr = torch.tensor(
[
(
adapter_weights[idx].weights_a.data_ptr()
if idx in adapter_weights
else 0
)
for idx in segment_indices
],
dtype=torch.int64,
device=device,
)
lora_b_ptr = torch.tensor(
[
(
adapter_weights[idx].weights_b.data_ptr()
if idx in adapter_weights
else 0
)
for idx in segment_indices
],
dtype=torch.int64,
device=device,
)
else:
use_sgmv = False
lora_a_ptr = torch.tensor(
[
(
adapter_weights[idx].weights_a_t.data_ptr()
if idx in adapter_weights
else 0
)
for idx in segment_indices
],
dtype=torch.int64,
device=device,
)
lora_b_ptr = torch.tensor(
[
(
adapter_weights[idx].weights_b_t.data_ptr()
if idx in adapter_weights
else 0
)
for idx in segment_indices
],
dtype=torch.int64,
device=device,
)
adapter_index_configs = {
idx: adapter_weights[idx].adapter_config
for idx in segment_indices
if idx in adapter_weights
}
adapter_to_segment = {v: k for k, v in enumerate(segment_indices)}
rank_indices = defaultdict(list)
for segment_idx, adapter_idx in enumerate(segment_indices):
if adapter_idx not in adapter_weights:
continue
rank_indices[adapter_weights[adapter_idx].lora_a_r].append(segment_idx)
if prefill_head_indices is not None:
j, prefill_head_segment_starts, prefill_head_segment_ends = 1, [0], [0]
for head_index in prefill_head_indices:
# j cannot go out of bounds as that would mean there are tokens without corresponding adapters
if head_index < meta.adapter_segments[j]:
prefill_head_segment_ends[-1] += 1
else:
prefill_head_segment_starts.append(prefill_head_segment_ends[-1])
prefill_head_segment_ends.append(prefill_head_segment_ends[-1] + 1)
j += 1
rank_data = {}
for rank, indices in rank_indices.items():
tmp_shrink = None
tmp_expand = None
segment_starts = None
segment_ends = None
batch_indices = None
if use_sgmv:
lora_a_ptr_indices = lora_a_ptr[indices]
tmp_shrink, tmp_expand = get_tmp_tensors(
lora_a_ptr_indices.size(0), rank, device
)
segment_starts = meta.adapter_segments[indices]
segment_ends = meta.adapter_segments[[i + 1 for i in indices]]
if prefill_head_indices is not None:
for i, segment_index in enumerate(indices):
segment_starts[i] = prefill_head_segment_starts[segment_index]
segment_ends[i] = prefill_head_segment_ends[segment_index]
else:
rank_indices = set(indices)
batch_indices = [
adapter_to_segment[idx] for idx in meta.adapter_indices.tolist()
]
batch_indices = [
idx if idx in rank_indices else -1 for idx in batch_indices
]
batch_indices = torch.tensor(
batch_indices, dtype=torch.int64, device=device
)
rank_data[rank] = RankSegments(
rank=rank,
tmp_shrink=tmp_shrink,
tmp_expand=tmp_expand,
lora_a_ptr=lora_a_ptr[indices],
lora_b_ptr=lora_b_ptr[indices],
segment_starts=segment_starts,
segment_ends=segment_ends,
indices=batch_indices,
)
return BatchLoraWeights(
lora_a=lora_a,
lora_b=lora_b,
adapter_index_configs=adapter_index_configs,
rank_data=rank_data,
use_sgmv=use_sgmv,
)
def get_scaling_factor(
lora_alpha: int,
r: int,
uses_rslora: bool = False,
) -> float:
"""Computes the scaling factor for the lora weights."""
if uses_rslora:
return lora_alpha / (r**0.5)
return lora_alpha / r
def _convert_lora(v: AdapterWeights) -> AdapterWeights:
if hasattr(v, "lora_weights"):
return v.lora_weights
return v
|
text-generation-inference/server/text_generation_server/adapters/lora.py/0
|
{
"file_path": "text-generation-inference/server/text_generation_server/adapters/lora.py",
"repo_id": "text-generation-inference",
"token_count": 8028
}
| 242
|
from accelerate import init_empty_weights
import torch
@classmethod
def load_conv2d(cls, prefix, weights, in_channels, out_channels, kernel_size, stride):
weight = weights.get_tensor(f"{prefix}.weight")
bias = weights.get_tensor(f"{prefix}.bias")
with init_empty_weights():
conv2d = cls(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
)
conv2d.weight = torch.nn.Parameter(weight)
conv2d.bias = torch.nn.Parameter(bias)
return conv2d
@classmethod
def load_conv2d_no_bias(
cls, prefix, weights, in_channels, out_channels, kernel_size, stride
):
weight = weights.get_tensor(f"{prefix}.weight")
with init_empty_weights():
conv2d = cls(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
)
conv2d.weight = torch.nn.Parameter(weight)
conv2d.bias = None
return conv2d
torch.nn.Conv2d.load = load_conv2d
torch.nn.Conv2d.load_no_bias = load_conv2d_no_bias
|
text-generation-inference/server/text_generation_server/layers/conv.py/0
|
{
"file_path": "text-generation-inference/server/text_generation_server/layers/conv.py",
"repo_id": "text-generation-inference",
"token_count": 518
}
| 243
|
from dataclasses import dataclass
from typing import List, Optional, Union
import numpy
import torch
import torch.nn as nn
from loguru import logger
from text_generation_server.layers.marlin.util import (
_check_marlin_kernels,
marlin_zero_points,
permute_scales,
unpack_cols,
)
from text_generation_server.utils.import_utils import SYSTEM
from text_generation_server.utils.log import log_once
from text_generation_server.utils.weights import Weight, Weights, WeightsLoader
try:
import marlin_kernels
except ImportError:
marlin_kernels = None
try:
major, _minor = torch.cuda.get_device_capability()
has_sm_8_0 = major >= 8
except Exception:
has_sm_8_0 = False
GPTQ_MARLIN_BITS = [4, 8]
GPTQ_MARLIN_GROUP_SIZES = [-1, 32, 64, 128]
MARLIN_TILE_SIZE = 16
def can_use_gptq_marlin(
*, bits: int, groupsize: int, quant_method: str, quantize: str, sym: bool
) -> bool:
return (
SYSTEM == "cuda"
and marlin_kernels is not None
and has_sm_8_0
and quantize in {"awq", "gptq"}
and quant_method in {"awq", "gptq"}
and bits in GPTQ_MARLIN_BITS
and groupsize in GPTQ_MARLIN_GROUP_SIZES
# We only suppord asymmetric quantization for AWQ.
and (sym or quant_method == "awq")
)
class GPTQMarlinWeightsLoader(WeightsLoader):
"""
Loader for using GPTQ- and AWQ-quantized weights with Marlin kernels.
"""
def __init__(
self,
*,
bits: int,
desc_act: bool,
groupsize: int,
quant_method: str,
quantize: str,
sym: bool,
):
self.bits = bits
self.desc_act = desc_act
self.groupsize = groupsize
self.quant_method = quant_method
self.quantize = quantize
self.sym = sym
def get_weights(self, weights: Weights, prefix: str):
log_once(logger.info, "Using GPTQ-Marlin kernels")
try:
qweight = weights.get_tensor(f"{prefix}.qweight")
except RuntimeError:
raise RuntimeError(
f"Cannot load `{self.quantize}` weight for GPTQ -> Marlin repacking, make sure the model is already quantized"
)
if not self.sym:
qzeros = weights.get_tensor(f"{prefix}.qzeros")
else:
qzeros = None
if self.quant_method == "awq":
g_idx = None
else:
g_idx = weights.get_tensor(f"{prefix}.g_idx")
scales = weights.get_tensor(f"{prefix}.scales")
return repack_gptq_for_marlin(
qweight=qweight,
scales=scales,
qzeros=qzeros,
g_idx=g_idx,
bits=self.bits,
desc_act=self.desc_act,
groupsize=self.groupsize,
quant_method=self.quant_method,
sym=self.sym,
sharded_infeatures=False,
)
def get_weights_col_packed(
self,
weights: Weights,
prefix: str,
block_sizes: Union[int, List[int]],
):
try:
qweight = weights.get_packed_sharded(
f"{prefix}.qweight", dim=1, block_sizes=block_sizes
)
except RuntimeError:
raise RuntimeError(
f"Cannot load `{self.quantize}` weight, make sure the model is already quantized."
)
scales = weights.get_packed_sharded(
f"{prefix}.scales", dim=1, block_sizes=block_sizes
)
scales = scales.to(dtype=weights.dtype)
if not self.sym:
qzeros = weights.get_packed_sharded(
f"{prefix}.qzeros", dim=1, block_sizes=block_sizes
)
else:
qzeros = None
if self.quant_method == "awq":
g_idx = None
else:
g_idx = weights.get_tensor(f"{prefix}.g_idx")
return repack_gptq_for_marlin(
qweight=qweight,
scales=scales,
qzeros=qzeros,
g_idx=g_idx,
bits=self.bits,
desc_act=self.desc_act,
groupsize=self.groupsize,
quant_method=self.quant_method,
sym=self.sym,
sharded_infeatures=False,
)
def get_multi_weights_col(self, weights: Weights, prefixes: List[str], dim: int):
try:
qweight = torch.cat(
[weights.get_sharded(f"{p}.qweight", dim=1) for p in prefixes], dim=1
)
except RuntimeError:
raise RuntimeError(
f"Cannot load `{self.quantize}` weight, make sure the model is already quantized"
)
scales = torch.cat(
[weights.get_sharded(f"{p}.scales", dim=1) for p in prefixes], dim=1
)
if not self.sym:
qzeros = torch.cat(
[weights.get_sharded(f"{p}.qzeros", dim=1) for p in prefixes], dim=1
)
else:
qzeros = None
if self.quant_method == "awq":
g_idx = None
else:
w = [weights.get_tensor(f"{p}.g_idx") for p in prefixes]
for w2 in w[1:]:
torch.testing.assert_close(w2, w[0])
g_idx = w[0]
return repack_gptq_for_marlin(
qweight=qweight,
scales=scales,
qzeros=qzeros,
g_idx=g_idx,
bits=self.bits,
desc_act=self.desc_act,
groupsize=self.groupsize,
quant_method=self.quant_method,
sym=self.sym,
sharded_infeatures=False,
)
def get_weights_row(self, weights: Weights, prefix: str):
log_once(logger.info, "Using GPTQ-Marlin kernels")
try:
qweight = weights.get_sharded(f"{prefix}.qweight", dim=0)
except RuntimeError:
raise RuntimeError(
f"Cannot load `{self.quantize}` weight for GPTQ -> Marlin repacking, make sure the model is already quantized"
)
if not self.sym:
if self.desc_act or self.groupsize == -1:
qzeros = weights.get_tensor(f"{prefix}.qzeros")
else:
qzeros = weights.get_sharded(f"{prefix}.qzeros", dim=0)
else:
qzeros = None
if self.quant_method == "awq":
g_idx = None
else:
g_idx = weights.get_sharded(f"{prefix}.g_idx", dim=0)
if self.desc_act or self.groupsize == -1:
scales = weights.get_tensor(f"{prefix}.scales")
else:
scales = weights.get_sharded(f"{prefix}.scales", dim=0)
sharded_in_features = weights.process_group.size() > 1
return repack_gptq_for_marlin(
qweight=qweight,
scales=scales,
qzeros=qzeros,
g_idx=g_idx,
bits=self.bits,
desc_act=self.desc_act,
groupsize=self.groupsize,
quant_method=self.quant_method,
sym=self.sym,
sharded_infeatures=sharded_in_features,
)
def _get_gptq_params(self, weights: Weights):
if weights._has_tensor("gptq_bits") and weights._has_tensor("gptq_groupsize"):
self.bits = weights.get_tensor("gptq_bits").item()
self.groupsize = weights.get_tensor("gptq_groupsize").item()
self.desc_act = False
# `server quantize` used asymmetric quantization unconditionally
# before the `gptq_sym` setting tensor was added.
self.sym = (
weights.get_tensor("gptq_sym").item()
if weights._has_tensor("gptq_sym")
else False
)
self.quant_method = "gptq"
@dataclass
class GPTQMarlinWeight(Weight):
"""
Repacked GPTQ Marlin weights.
"""
qweight: torch.Tensor
qzeros: torch.Tensor
scales: torch.Tensor
g_idx: torch.Tensor
perm: torch.Tensor
bits: int
is_full_k: bool
def __post_init__(self):
assert self.qweight.dtype == torch.int32
assert self.scales.dtype == torch.float16
assert self.g_idx.dtype == torch.int32
assert self.perm.dtype == torch.int32
def get_linear(self, bias: torch.Tensor):
return GPTQMarlinLinear(
weight=self,
bias=bias,
)
def repack_gptq_for_marlin(
*,
qweight: torch.Tensor,
qzeros: Optional[torch.Tensor],
scales: torch.Tensor,
g_idx: Optional[torch.Tensor],
bits: int,
desc_act: bool,
groupsize: int,
quant_method: str,
sym: bool,
sharded_infeatures: bool,
) -> GPTQMarlinWeight:
"""Convert GPTQ weights to a layout that's compatible with GPTQ-Marlin kernels."""
_check_marlin_kernels()
assert marlin_kernels is not None
if bits not in GPTQ_MARLIN_BITS:
supported_bits = ", ".join(str(b) for b in GPTQ_MARLIN_BITS)
raise RuntimeError(
f"Repacking {bits}-bit GPTQ weights as Marlin is not supported, must be one of: {supported_bits}"
)
if groupsize not in GPTQ_MARLIN_GROUP_SIZES:
supported_sizes = ", ".join(str(b) for b in GPTQ_MARLIN_GROUP_SIZES)
raise RuntimeError(
f"Repacking GPTQ weights with group size {groupsize} as Marlin is not supported, must be one of: {supported_sizes}"
)
if not (sym or quant_method == "awq"):
raise RuntimeError(
"Repacking GPTQ weights with asymmetric quantization as Marlin is not supported."
)
log_once(logger.info, f"Converting {quant_method} model to Marlin packing format.")
weights_per_int = 32 // bits
in_features = qweight.shape[0]
out_features = qweight.shape[1]
# AWQ uses column packing, GPTQ uses row packing
if quant_method == "awq":
out_features *= weights_per_int
else:
in_features *= weights_per_int
if in_features % groupsize != 0:
raise ValueError(
f"Number of input features ({in_features}) not divisible by group size ({groupsize})"
)
if g_idx is not None and desc_act and groupsize != -1:
perm = torch.argsort(g_idx).to(torch.int)
g_idx = g_idx[perm]
else:
perm = torch.empty(0, dtype=torch.int, device=qweight.device)
g_idx = torch.empty(0, dtype=torch.int, device=qweight.device)
if quant_method == "awq":
repacked = marlin_kernels.awq_marlin_repack(
qweight, in_features, out_features, bits
)
if qzeros is not None:
qzeros = awq_to_marlin_zero_points(
qzeros,
in_features // groupsize,
out_features,
bits,
)
else:
repacked = marlin_kernels.gptq_marlin_repack(
qweight, perm, in_features, out_features, bits
)
if qzeros is None:
qzeros = torch.empty(0, dtype=torch.int, device=qweight.device)
scales = permute_scales(scales)
is_full_k = not (desc_act and sharded_infeatures)
return GPTQMarlinWeight(
qweight=repacked,
qzeros=qzeros,
scales=scales,
g_idx=g_idx,
perm=perm,
bits=bits,
is_full_k=is_full_k,
)
class GPTQMarlinLinear(nn.Module):
"""
Linear layer for GPTQ weights that were converted for the GPTQ-Marlin
kernels.
"""
def __init__(
self,
*,
weight: GPTQMarlinWeight,
bias: Optional[torch.Tensor],
):
super().__init__()
_check_marlin_kernels()
assert marlin_kernels is not None
in_features = weight.qweight.shape[0] * MARLIN_TILE_SIZE
out_features = weight.scales.shape[1]
_check_valid_shape(in_features=in_features, out_features=out_features)
self.bits = weight.bits
self.is_full_k = weight.is_full_k
self.qweight = weight.qweight
self.qzeros = weight.qzeros
self.scales = weight.scales
self.g_idx = weight.g_idx
self.perm = weight.perm
if bias is not None:
self.bias = bias
else:
self.bias = None
self.workspace = torch.zeros(
out_features // 64 * 16, dtype=torch.int, device=weight.qweight.device
)
def forward(self, A: torch.Tensor) -> torch.Tensor:
assert marlin_kernels is not None
A_flat = A.view(-1, A.shape[-1])
C = marlin_kernels.gptq_marlin_gemm(
A_flat,
self.qweight,
self.scales,
self.qzeros,
self.g_idx,
self.perm,
self.workspace,
self.bits,
A_flat.shape[0],
self.scales.shape[1],
A_flat.shape[1],
self.is_full_k,
self.qzeros.numel() > 0,
True,
)
C = C.reshape(A.shape[:-1] + (self.scales.shape[1],))
if self.bias is not None:
C += self.bias
return C
def awq_to_marlin_zero_points(
q_zp_packed: torch.Tensor, size_k: int, size_n: int, num_bits: int
) -> torch.Tensor:
# AWQ zero-points are quantized and packed on the column dim.
# In addition, the values are permuted based on dequantizer.
# Here we undo both of these, and then apply marlin permutation
# and pack it back.
q_zp = unpack_cols(q_zp_packed, num_bits, size_k, size_n)
# Undo interleaving (use argsort(..) to get inverse perm)
if num_bits == 4:
undo_interleave = numpy.argsort(numpy.array([0, 2, 4, 6, 1, 3, 5, 7]))
elif num_bits == 8:
undo_interleave = numpy.argsort(numpy.array([0, 2, 1, 3]))
else:
raise Exception("num_bits must be 4 or 8, got {}".format(num_bits))
q_zp = q_zp.reshape((-1, len(undo_interleave)))[:, undo_interleave].ravel()
q_zp = q_zp.reshape((-1, size_n)).contiguous()
marlin_zp = marlin_zero_points(q_zp, size_k, size_n, num_bits)
return marlin_zp
def _check_valid_shape(in_features: int, out_features: int):
if (in_features % 128 != 0 or out_features % 64 != 0) and (
in_features % 64 != 0 or out_features % 128 != 0
):
raise ValueError(
f"The GPTQ Marlin kernel does not have a valid thread configuration for weight matrix with shape ({out_features}, {in_features})."
" The shape elements must be divisible by (128, 64) or (64, 128)."
)
|
text-generation-inference/server/text_generation_server/layers/marlin/gptq.py/0
|
{
"file_path": "text-generation-inference/server/text_generation_server/layers/marlin/gptq.py",
"repo_id": "text-generation-inference",
"token_count": 7135
}
| 244
|
# coding=utf-8
# Copyright 2023, 2024 DeepSeek-AI 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 typing import Any, Dict, List, Optional, Tuple
import torch
import torch.distributed
from text_generation_server.layers import (
FastLinear,
SpeculativeHead,
TensorParallelColumnLinear,
TensorParallelEmbedding,
TensorParallelRowLinear,
get_linear,
)
from text_generation_server.layers.attention import (
attention,
paged_attention,
reshape_and_cache,
Seqlen,
)
from text_generation_server.layers.layernorm import FastRMSNorm
from text_generation_server.layers.rotary import PositionRotaryEmbedding, get_mscale
from text_generation_server.utils.import_utils import SYSTEM
from text_generation_server.utils.weights import Weights
from torch import nn
from transformers.activations import ACT2FN
from transformers.configuration_utils import PretrainedConfig
if SYSTEM == "rocm":
try:
from vllm import _custom_C
except Exception as e:
raise ImportError(f"Could not load `vllm._custom_C`. Full error: {e}")
class DeepseekV2Config(PretrainedConfig):
def __init__(
self,
vocab_size=102400,
hidden_size=4096,
intermediate_size=11008,
moe_intermediate_size=1407,
num_hidden_layers=30,
num_attention_heads=32,
num_key_value_heads=32,
n_shared_experts=2,
n_routed_experts=160,
ep_size=1,
routed_scaling_factor=1.0,
kv_lora_rank=512,
q_lora_rank=1536,
qk_rope_head_dim=64,
v_head_dim=128,
qk_nope_head_dim=128,
topk_method="gready",
n_group=8,
topk_group=3,
num_experts_per_tok=6,
moe_layer_freq=1,
first_k_dense_replace=0,
norm_topk_prob=False,
scoring_func="softmax",
aux_loss_alpha=0.001,
seq_aux=True,
hidden_act="silu",
max_position_embeddings=2048,
initializer_range=0.02,
rms_norm_eps=1e-6,
use_cache=True,
pad_token_id=None,
bos_token_id=100000,
eos_token_id=100001,
pretraining_tp=1,
tie_word_embeddings=False,
rope_theta=10000.0,
rope_scaling=None,
attention_bias=False,
attention_dropout=0.0,
**kwargs,
):
self.vocab_size = vocab_size
self.max_position_embeddings = max_position_embeddings
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.moe_intermediate_size = moe_intermediate_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.n_shared_experts = n_shared_experts
self.n_routed_experts = n_routed_experts
self.ep_size = ep_size
self.routed_scaling_factor = routed_scaling_factor
self.kv_lora_rank = kv_lora_rank
self.q_lora_rank = q_lora_rank
self.qk_rope_head_dim = qk_rope_head_dim
self.v_head_dim = v_head_dim
self.qk_nope_head_dim = qk_nope_head_dim
self.topk_method = topk_method
self.n_group = n_group
self.topk_group = topk_group
self.num_experts_per_tok = num_experts_per_tok
self.moe_layer_freq = moe_layer_freq
self.first_k_dense_replace = first_k_dense_replace
self.norm_topk_prob = norm_topk_prob
self.scoring_func = scoring_func
self.aux_loss_alpha = aux_loss_alpha
self.seq_aux = seq_aux
# for backward compatibility
if num_key_value_heads is None:
num_key_value_heads = num_attention_heads
self.num_key_value_heads = num_key_value_heads
self.hidden_act = hidden_act
self.initializer_range = initializer_range
self.rms_norm_eps = rms_norm_eps
self.pretraining_tp = pretraining_tp
self.use_cache = use_cache
self.rope_theta = rope_theta
self.rope_scaling = rope_scaling
self.attention_bias = attention_bias
self.attention_dropout = attention_dropout
tie_word_embeddings = kwargs.pop("tie_word_embeddings", False)
if tie_word_embeddings:
raise ValueError(
"tie_word_embeddings is not supported for Deepseek V2 models."
)
if ep_size != 1:
raise ValueError(
f"Currently only ep_size == 1 is supported for Deepseek V2 models, was {ep_size}"
)
super().__init__(
pad_token_id=pad_token_id,
bos_token_id=bos_token_id,
eos_token_id=eos_token_id,
tie_word_embeddings=tie_word_embeddings,
**kwargs,
)
def _load_experts(config, prefix: str, mat: str, weights: Weights):
if config.quantize is not None:
raise NotImplementedError(
"Deepseek V2 does not support weight quantization yet."
)
assert mat in ["gate_proj", "up_proj", "down_proj"]
world_size = weights.process_group.size()
rank = weights.process_group.rank()
assert (
config.moe_intermediate_size % world_size == 0
), f"The chosen size {config.moe_intermediate_size} is not compatible with sharding on {world_size} shards"
block_size = config.moe_intermediate_size // world_size
start = rank * block_size
stop = (rank + 1) * block_size
tensor = torch.empty(
(config.n_routed_experts * block_size, config.hidden_size),
dtype=weights.dtype,
device=weights.device,
)
for i in range(config.n_routed_experts):
slice_ = weights._get_slice(f"{prefix}.{i}.{mat}.weight")
if mat == "down_proj":
expert_slice = slice_[:, start:stop].t().contiguous()
else:
expert_slice = slice_[start:stop]
tensor[i * block_size : (i + 1) * block_size] = expert_slice.to(
dtype=weights.dtype
).to(device=weights.device)
return tensor
class DeepseekV2Attention(torch.nn.Module):
def __init__(
self,
prefix: str,
config,
weights: Weights,
):
super().__init__()
self.num_heads = config.num_attention_heads
self.hidden_size = config.hidden_size
self.kv_lora_rank = config.kv_lora_rank
self.q_lora_rank = config.q_lora_rank
self.qk_nope_head_dim = config.qk_nope_head_dim
self.qk_rope_head_dim = config.qk_rope_head_dim
self.head_size = config.qk_nope_head_dim + config.qk_rope_head_dim
self.value_head_size = config.v_head_dim
self.head_pad_size = max(self.head_size, self.value_head_size)
self.rotary_emb = PositionRotaryEmbedding.static(
config=config,
dim=self.qk_rope_head_dim,
base=config.rope_theta,
device=weights.device,
)
mscale = get_mscale(
self.rotary_emb.scaling_factor, self.rotary_emb.mscale_all_dim
)
self.softmax_scale = self.head_size**-0.5 * mscale * mscale
if self.num_heads % weights.process_group.size() != 0:
raise ValueError(
f"`num_heads` must be divisible by `num_shards` (got `num_heads`: {self.num_heads} "
f"and `num_shards`: {weights.process_group.size()}"
)
self.num_heads = self.num_heads // weights.process_group.size()
self.num_key_value_heads = (
config.num_key_value_heads // weights.process_group.size()
)
if self.q_lora_rank is None:
self.q_proj = TensorParallelColumnLinear.load(
config,
prefix=f"{prefix}.q_proj",
weights=weights,
bias=config.attention_bias,
)
else:
self.q_a_proj = get_linear(
weight=weights.get_weights(f"{prefix}.q_a_proj"),
bias=(
weights.get_tensor(f"{prefix}.q_a_proj.bias")
if config.attention_bias
else None
),
)
self.q_a_layernorm = FastRMSNorm.load(
prefix=f"{prefix}.q_a_layernorm",
weights=weights,
eps=config.rms_norm_eps,
)
self.q_b_proj = TensorParallelColumnLinear.load(
config,
prefix=f"{prefix}.q_b_proj",
weights=weights,
bias=config.attention_bias,
)
self.kv_a_proj_with_mqa = get_linear(
weight=weights.get_weights(f"{prefix}.kv_a_proj_with_mqa"),
bias=(
weights.get_tensor(f"{prefix}.kv_a_proj_with_mqa.bias")
if config.attention_bias
else None
),
)
self.kv_a_layernorm = FastRMSNorm.load(
prefix=f"{prefix}.kv_a_layernorm", weights=weights, eps=config.rms_norm_eps
)
self.kv_b_proj = TensorParallelColumnLinear.load(
config,
prefix=f"{prefix}.kv_b_proj",
weights=weights,
bias=config.attention_bias,
)
self.o_proj = TensorParallelRowLinear.load(
config,
prefix=f"{prefix}.o_proj",
weights=weights,
bias=False,
)
self.num_groups = self.num_heads // self.num_key_value_heads
self.kv_head_mapping = torch.arange(
0, self.num_key_value_heads, dtype=torch.int32, device=weights.device
).repeat_interleave(self.num_groups)
def forward(
self,
hidden_states: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
cu_seqlen_prefill: torch.Tensor,
kv_cache: Tuple[torch.Tensor, torch.Tensor],
block_tables: torch.Tensor,
slots: torch.Tensor,
seqlen: Seqlen,
max_s: int,
):
if self.q_lora_rank is None:
query = self.q_proj(hidden_states)
else:
query = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states))[0])
query = query.view(-1, self.num_heads, self.head_size)
_, query_pe = torch.split(
query, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1
)
compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
compressed_kv, key_pe = torch.split(
compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
)
key_pe = key_pe.view(-1, 1, self.qk_rope_head_dim)
kv = self.kv_b_proj(self.kv_a_layernorm(compressed_kv.contiguous())[0]).view(
-1, self.num_key_value_heads, self.qk_nope_head_dim + self.value_head_size
)
key_nope, value = torch.split(
kv, [self.qk_nope_head_dim, self.value_head_size], dim=-1
)
batch_size, heads, head_dim = query_pe.shape
query_pe = (
query_pe.view(batch_size, heads, head_dim // 2, 2)
.transpose(2, 3)
.reshape(batch_size, heads, head_dim)
)
batch_size, heads, head_dim = key_pe.shape
key_pe = (
key_pe.view(batch_size, heads, head_dim // 2, 2)
.transpose(2, 3)
.reshape(batch_size, heads, head_dim)
)
self.rotary_emb(query_pe, key_pe, cos, sin)
query[..., self.qk_nope_head_dim :] = query_pe
key = torch.empty_like(query)
key[..., : self.qk_nope_head_dim] = key_nope
key[..., self.qk_nope_head_dim :] = key_pe
# We need to pad the heads because Flash Attention does not support
# qk and v with different head sizes.
query = torch.nn.functional.pad(
query, (0, self.head_pad_size - self.head_size), value=0
)
key = torch.nn.functional.pad(
key, (0, self.head_pad_size - self.head_size), value=0
)
value = torch.nn.functional.pad(
value, (0, self.head_pad_size - self.value_head_size), value=0
)
reshape_and_cache(key, value, kv_cache[0], kv_cache[1], slots)
# Prefill
if cu_seqlen_prefill is not None:
# flash attention
attn_output = attention(
query,
kv_cache[0],
kv_cache[1],
seqlen,
block_tables,
self.softmax_scale,
)
# Decode
else:
attn_output = paged_attention(
query,
kv_cache[0],
kv_cache[1],
self.kv_head_mapping,
self.softmax_scale,
block_tables,
seqlen,
max_s,
)
# Remove padding.
attn_output = attn_output[..., : self.value_head_size]
return self.o_proj(
attn_output.reshape(-1, self.num_heads * self.value_head_size)
)
class DeepseekV2MLP(nn.Module):
def __init__(self, prefix: str, config, weights, intermediate_size: int):
super().__init__()
self.hidden_act = config.hidden_act
if self.hidden_act != "silu":
# Bail out because MoE only supports silu.
raise NotImplementedError(
"Currently only `silu` is supported as an activation for Deepseek V2."
)
self.act = ACT2FN[self.hidden_act]
self.gate_up_proj = TensorParallelColumnLinear.load_multi(
config,
prefixes=[f"{prefix}.gate_proj", f"{prefix}.up_proj"],
weights=weights,
dim=0,
bias=False,
)
self.down_proj = TensorParallelRowLinear.load(
config,
prefix=f"{prefix}.down_proj",
weights=weights,
bias=False,
)
self.intermediate_size = intermediate_size // weights.process_group.size()
# TODO: This is a hotfix to be removed & properly refactored.
self.quantize = config.quantize
def forward(self, hidden_states: torch.Tensor, reduce: bool = True):
if (
SYSTEM == "rocm"
and self.hidden_act == "silu"
and hidden_states.shape[0] == 1
and not self.quantize
):
out = torch.empty(
hidden_states.shape[0],
self.intermediate_size,
dtype=hidden_states.dtype,
device="cuda",
)
_custom_C.LLMM_Silu(self.gate_up_proj.linear.weight, hidden_states, out, 8)
return self.down_proj(out, reduce=reduce)
else:
gate_up_states = self.gate_up_proj(hidden_states)
gate_up_states = gate_up_states.view(-1, 2, self.intermediate_size)
return self.down_proj(
self.act(gate_up_states[:, 0]) * gate_up_states[:, 1], reduce=reduce
)
class BlockSparseMoE(nn.Module):
def __init__(self, prefix, config: DeepseekV2Config, weights):
super().__init__()
self.hidden_dim = config.hidden_size
self.moe_intermediate_size = (
config.moe_intermediate_size // weights.process_group.size()
)
self.n_routed_experts = config.n_routed_experts
self.n_expert_group = config.n_group
self.topk_group = config.topk_group
self.top_k = config.num_experts_per_tok
self.norm_topk_prob = config.norm_topk_prob
self.routed_scaling_factor = config.routed_scaling_factor
gate_proj = _load_experts(
config, f"{prefix}.experts", "gate_proj", weights
).view(self.n_routed_experts, self.moe_intermediate_size, self.hidden_dim)
up_proj = _load_experts(config, f"{prefix}.experts", "up_proj", weights).view(
self.n_routed_experts, self.moe_intermediate_size, self.hidden_dim
)
self.gate_up_proj = torch.cat([gate_proj, up_proj], dim=1)
self.down_proj = (
_load_experts(config, f"{prefix}.experts", "down_proj", weights)
.view(self.n_routed_experts, self.moe_intermediate_size, self.hidden_dim)
.transpose(1, 2)
.contiguous()
)
# Gating
self.gate = FastLinear.load(config, f"{prefix}.gate", weights, bias=False)
if config.n_shared_experts is not None:
self.shared_experts = DeepseekV2MLP(
prefix=f"{prefix}.shared_experts",
config=config,
weights=weights,
intermediate_size=config.moe_intermediate_size
* config.n_shared_experts,
)
else:
self.shared_experts = None
self.process_group = weights.process_group
def forward(self, x: torch.Tensor) -> torch.Tensor:
if self.shared_experts is not None:
shared_output = self.shared_experts(x, reduce=False)
else:
shared_output = None
router_logits = self.gate(x)
topk_weights, topk_ids = grouped_topk(
x,
router_logits,
self.top_k,
renormalize=self.norm_topk_prob,
num_expert_group=self.n_expert_group,
topk_group=self.topk_group,
)
out = (
fused_experts(
x,
self.gate_up_proj,
self.down_proj,
topk_weights,
topk_ids,
inplace=True,
)
* self.routed_scaling_factor
)
if shared_output is not None:
out = out + shared_output
# Reduce sum
if self.process_group.size() > 1:
torch.distributed.all_reduce(out, group=self.process_group)
return out.view(*x.shape)
class DenseMoE(nn.Module):
def __init__(self, prefix: str, config: DeepseekV2Config, weights: Weights):
super().__init__()
self.hidden_dim = config.hidden_size
self.moe_intermediate_size = config.moe_intermediate_size
self.n_routed_experts = config.n_routed_experts
self.n_expert_group = config.n_group
self.topk_group = config.topk_group
self.top_k = config.num_experts_per_tok
self.norm_topk_prob = config.norm_topk_prob
self.routed_scaling_factor = config.routed_scaling_factor
# Gating
#
# Seems like no one quantizes the gate.
self.gate = FastLinear.load(config, f"{prefix}.gate", weights, bias=False)
self.experts = [
DeepseekV2MLP(
f"{prefix}.experts.{i}", config, weights, self.moe_intermediate_size
)
for i in range(self.n_routed_experts)
]
if config.n_shared_experts is not None:
self.shared_experts = DeepseekV2MLP(
prefix=f"{prefix}.shared_experts",
config=config,
weights=weights,
intermediate_size=config.moe_intermediate_size
* config.n_shared_experts,
)
else:
self.shared_experts = None
self.process_group = weights.process_group
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
x: (sequence_length, model_dim)
gate_logits: (sequence_length, n_experts)
"""
# optional reshape
input_shape = x.shape
x = x.view(-1, input_shape[-1])
if self.shared_experts is not None:
shared_output = self.shared_experts(x, reduce=False)
else:
shared_output = None
# gate_logits: (sequence_length, n_experts)
router_logits = self.gate(x)
topk_weights, topk_ids = grouped_topk(
x,
router_logits,
self.top_k,
renormalize=self.norm_topk_prob,
num_expert_group=self.n_expert_group,
topk_group=self.topk_group,
)
out = self.moe_infer_gpu(x, topk_ids, topk_weights) * self.routed_scaling_factor
if shared_output is not None:
out = out + shared_output
# Reduce sum
if self.process_group.size() > 1:
torch.distributed.all_reduce(out, group=self.process_group)
return out
def moe_infer_gpu(
self, x: torch.Tensor, topk_ids: torch.Tensor, topk_weight: torch.Tensor
):
weights = torch.zeros(
topk_ids.shape[0], len(self.experts), dtype=x.dtype, device=x.device
)
weights.scatter_(1, topk_ids, topk_weight)
out = x.new_zeros(x.shape[0], self.hidden_dim)
for i, expert in enumerate(self.experts):
# Add expert output to out with masking
out += expert(x, reduce=False) * weights[:, i].view(-1, 1)
return out
class DeepseekV2Layer(nn.Module):
def __init__(self, prefix, layer_id, config, weights):
super().__init__()
prefix = f"{prefix}.layers.{layer_id}"
self.self_attn = DeepseekV2Attention(
prefix=f"{prefix}.self_attn",
config=config,
weights=weights,
)
if (
config.n_routed_experts is not None
and layer_id >= config.first_k_dense_replace
and layer_id % config.moe_layer_freq == 0
):
moe_cls = BlockSparseMoE if config.quantize is None else DenseMoE
self.mlp = moe_cls(f"{prefix}.mlp", config, weights)
else:
self.mlp = DeepseekV2MLP(
prefix=f"{prefix}.mlp",
config=config,
weights=weights,
intermediate_size=config.intermediate_size,
)
self.input_layernorm = FastRMSNorm.load(
prefix=f"{prefix}.input_layernorm", weights=weights, eps=config.rms_norm_eps
)
self.post_attention_layernorm = FastRMSNorm.load(
prefix=f"{prefix}.post_attention_layernorm",
weights=weights,
eps=config.rms_norm_eps,
)
def forward(
self,
hidden_states: torch.Tensor,
residual: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
cu_seqlen_prefill: torch.Tensor,
kv_cache,
block_tables: torch.Tensor,
slots: torch.Tensor,
seqlen: Seqlen,
max_s: int,
):
normed_hidden_states, residual = self.input_layernorm(hidden_states, residual)
# Self Attention
attn_output = self.self_attn(
normed_hidden_states,
cos,
sin,
cu_seqlen_prefill,
kv_cache,
block_tables,
slots,
seqlen,
max_s,
)
# faster post attention rms norm
normed_attn_res_output, residual = self.post_attention_layernorm(
attn_output, residual
)
output = self.mlp(normed_attn_res_output)
return output, residual
class DeepseekV2Model(torch.nn.Module):
def __init__(self, prefix: str, config, weights: Weights):
super().__init__()
self.embed_tokens = TensorParallelEmbedding(
prefix=f"{prefix}.embed_tokens", weights=weights
)
self.layers = nn.ModuleList(
[
DeepseekV2Layer(
prefix,
layer_id,
config,
weights,
)
for layer_id in range(config.num_hidden_layers)
]
)
self.norm = FastRMSNorm.load(
prefix=f"{prefix}.norm", weights=weights, eps=config.rms_norm_eps
)
self.head_size = self.layers[0].self_attn.head_size
self.num_heads = self.layers[0].self_attn.num_heads
self.num_key_value_heads = self.layers[0].self_attn.num_key_value_heads
def forward(
self,
input_ids: torch.Tensor,
position_ids: torch.Tensor,
cu_seqlen_prefill: Optional[torch.Tensor],
kv_cache: List[Tuple[torch.Tensor, torch.Tensor]],
block_tables: torch.Tensor,
slots: torch.Tensor,
seqlen: Seqlen,
max_s: int,
) -> torch.Tensor:
hidden_states = self.embed_tokens(input_ids)
# Get rotary cos and sin for this forward
# Avoid to index in each layer
cos, sin = self.layers[0].self_attn.rotary_emb.get_cos_sin(
position_ids, max_s, hidden_states.dtype
)
residual = None
for i, layer in enumerate(self.layers):
hidden_states, residual = layer(
hidden_states,
residual,
cos,
sin,
cu_seqlen_prefill,
kv_cache[i],
block_tables,
slots,
seqlen,
max_s,
)
hidden_states, _ = self.norm(hidden_states, residual)
return hidden_states
class FlashDeepseekV2ForCausalLM(torch.nn.Module):
def __init__(self, prefix: str, config, weights: Weights):
super().__init__()
self.model = DeepseekV2Model(
"model" if not prefix else f"{prefix}.model", config, weights
)
self.lm_head = SpeculativeHead.load(
config,
prefix="lm_head" if not prefix else f"{prefix}.lm_head",
weights=weights,
)
def forward(
self,
input_ids: torch.Tensor,
position_ids: torch.Tensor,
cu_seqlen_prefill: Optional[torch.Tensor],
kv_cache: List[Tuple[torch.Tensor, torch.Tensor]],
block_tables: torch.Tensor,
slots: torch.Tensor,
seqlen: Seqlen,
max_s: int,
prefill_cache_indices: Optional[torch.Tensor],
lm_head_indices: Optional[torch.Tensor] = None,
adapter_data: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
hidden_states = self.model(
input_ids,
position_ids,
cu_seqlen_prefill,
kv_cache,
block_tables,
slots,
seqlen,
max_s,
)
if lm_head_indices is not None:
hidden_states = hidden_states[lm_head_indices]
logits, speculative_logits = self.lm_head(hidden_states)
return logits, speculative_logits
# Functions below are from vLLM:
#
# https://github.com/vllm-project/vllm/blob/f7160d946a0a07703e72d81ba9ecf3913f192605/vllm/model_executor/layers/fused_moe/fused_moe.py#L397
#
# Remove after we have synced our version with upstream.
def grouped_topk(
hidden_states: torch.Tensor,
gating_output: torch.Tensor,
topk: int,
renormalize: bool,
num_expert_group: int = 0,
topk_group: int = 0,
) -> Tuple[torch.Tensor, torch.Tensor]:
scores = torch.softmax(gating_output, dim=-1)
num_token = scores.shape[0]
group_scores = (
scores.view(num_token, num_expert_group, -1).max(dim=-1).values
) # [n, n_group]
group_idx = torch.topk(group_scores, k=topk_group, dim=-1, sorted=False)[
1
] # [n, top_k_group]
group_mask = torch.zeros_like(group_scores) # [n, n_group]
group_mask.scatter_(1, group_idx, 1) # [n, n_group]
score_mask = (
group_mask.unsqueeze(-1)
.expand(num_token, num_expert_group, scores.shape[-1] // num_expert_group)
.reshape(num_token, -1)
) # [n, e]
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
topk_weights, topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)
if renormalize:
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
return topk_weights, topk_ids
def get_default_config(
M: int,
E: int,
N: int,
K: int,
topk: int,
dtype: Optional[str],
) -> Dict[str, int]:
config = {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
}
if M <= E:
config = {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
}
return config
def fused_experts(
hidden_states: torch.Tensor,
w1: torch.Tensor,
w2: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
inplace: bool = False,
override_config: Optional[Dict[str, Any]] = None,
use_fp8: bool = False,
w1_scale: Optional[torch.Tensor] = None,
w2_scale: Optional[torch.Tensor] = None,
a1_scale: Optional[torch.Tensor] = None,
a2_scale: Optional[torch.Tensor] = None,
):
# Check constraints.
assert hidden_states.shape[1] == w1.shape[2], "Hidden size mismatch"
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
assert w2.is_contiguous(), "Expert weights2 must be contiguous"
assert hidden_states.dtype in [torch.float32, torch.float16, torch.bfloat16]
import triton.language as tl
from vllm import _custom_ops as ops
from vllm.model_executor.layers.fused_moe.fused_moe import (
get_moe_configs,
invoke_fused_moe_kernel,
moe_align_block_size,
)
M, _ = hidden_states.shape
E, N, _ = w1.shape
if override_config:
config = override_config
else:
# First try to load optimal config from the file
configs = get_moe_configs(E, w2.shape[2], "float8" if use_fp8 else None)
if configs:
# If an optimal configuration map has been found, look up the
# optimal config
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
else:
# Else use the default config
config = get_default_config(
M, E, N, w1.shape[2], topk_ids.shape[1], "float8" if use_fp8 else None
)
intermediate_cache1 = torch.empty(
(M, topk_ids.shape[1], N),
device=hidden_states.device,
dtype=hidden_states.dtype,
)
intermediate_cache2 = torch.empty(
(M * topk_ids.shape[1], N // 2),
device=hidden_states.device,
dtype=hidden_states.dtype,
)
intermediate_cache3 = torch.empty(
(M, topk_ids.shape[1], w2.shape[1]),
device=hidden_states.device,
dtype=hidden_states.dtype,
)
sorted_token_ids, expert_ids, num_tokens_post_padded = moe_align_block_size(
topk_ids, config["BLOCK_SIZE_M"], E
)
compute_type = tl.bfloat16 if hidden_states.dtype == torch.bfloat16 else tl.float16
invoke_fused_moe_kernel(
hidden_states,
w1,
intermediate_cache1,
a1_scale,
w1_scale,
topk_weights,
topk_ids,
sorted_token_ids,
expert_ids,
num_tokens_post_padded,
False,
topk_ids.shape[1],
config,
compute_type=compute_type,
use_fp8=use_fp8,
)
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
invoke_fused_moe_kernel(
intermediate_cache2,
w2,
intermediate_cache3,
a2_scale,
w2_scale,
topk_weights,
topk_ids,
sorted_token_ids,
expert_ids,
num_tokens_post_padded,
True,
1,
config,
compute_type=compute_type,
use_fp8=use_fp8,
)
if inplace:
return torch.sum(
intermediate_cache3.view(*intermediate_cache3.shape),
dim=1,
out=hidden_states,
)
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
text-generation-inference/server/text_generation_server/models/custom_modeling/flash_deepseek_v2_modeling.py/0
|
{
"file_path": "text-generation-inference/server/text_generation_server/models/custom_modeling/flash_deepseek_v2_modeling.py",
"repo_id": "text-generation-inference",
"token_count": 16319
}
| 245
|
import re
import torch
import torch.distributed
from transformers import (
PreTrainedTokenizerBase,
)
from text_generation_server.models.causal_lm import CausalLMBatch
from text_generation_server.pb import generate_pb2
from text_generation_server.utils import (
NextTokenChooser,
StoppingCriteria,
)
from text_generation_server.utils.chunks import concat_text_chunks
# CREDIT: Papers with code => https://github.com/paperswithcode/galai/blob/main/galai/utils.py
# we split individual characters inside special tokens like [START_DNA]
CUSTOM_SEQ_RE = re.compile(r"(\[START_(DNA|SMILES|I_SMILES|AMINO)])(.*?)(\[END_\2])")
# token added to implement a custom sequence tokenization. This token is added at
# corpus cleaning step and removed in pretokenization. The digits are added to increase the chance
# that they do not occur in the corpus. The digits are escaped so that the token does not appear
# literally in the source code in case we ever include it in the training data.
SPLIT_MARKER = f"SPL{1}T-TH{1}S-Pl3A5E"
def _insert_split_marker(m: re.Match):
"""
Applies split marker based on a regex match of special tokens such as
[START_DNA].
Parameters
----------
n : str
Input text to split
Returns
----------
str - the text with the split token added
"""
start_token, _, sequence, end_token = m.groups()
sequence = re.sub(r"(.)", rf"{SPLIT_MARKER}\1", sequence, flags=re.DOTALL)
return f"{start_token}{sequence}{SPLIT_MARKER}{end_token}"
def escape_custom_split_sequence(text):
"""
Applies custom splitting to the text for GALILEO's tokenization
Parameters
----------
text : str
Input text to split
Returns
----------
str - the text with the split token added
"""
return CUSTOM_SEQ_RE.sub(_insert_split_marker, text)
# END CREDIT
class GalacticaCausalLMBatch(CausalLMBatch):
@classmethod
def from_pb(
cls,
pb: generate_pb2.Batch,
tokenizer: PreTrainedTokenizerBase,
dtype: torch.dtype,
device: torch.device,
) -> "GalacticaCausalLMBatch":
inputs = []
next_token_choosers = []
stopping_criterias = []
prefix_offsets = []
top_n_tokens = []
read_offsets = []
requests_idx_mapping = {}
# Parse batch
max_truncation = 0
padding_right_offset = 0
max_decode_tokens = 0
for i, r in enumerate(pb.requests):
requests_idx_mapping[r.id] = i
# Add escape_custom_split_sequence to the CausalLMBatch logic
inputs.append(
escape_custom_split_sequence(concat_text_chunks(r.input_chunks.chunks))
)
next_token_choosers.append(
NextTokenChooser.from_pb(r.parameters, device, tokenizer)
)
stopping_criteria = StoppingCriteria.from_pb(
r.stopping_parameters, tokenizer
)
stopping_criterias.append(stopping_criteria)
top_n_tokens.append(r.top_n_tokens)
max_truncation = max(max_truncation, r.truncate)
max_decode_tokens += stopping_criteria.max_new_tokens
padding_right_offset = max(
padding_right_offset, stopping_criteria.max_new_tokens
)
tokenized_inputs = tokenizer(
inputs,
return_tensors="pt",
padding=True,
return_token_type_ids=False,
truncation=True,
max_length=max_truncation,
).to(device)
for _ in pb.requests:
input_len = tokenized_inputs["input_ids"].shape[1]
prefix_offsets.append(0)
read_offsets.append(input_len)
input_lengths = tokenized_inputs["attention_mask"].sum(1)
max_input_length = input_lengths.max()
input_ids = tokenized_inputs["input_ids"]
# Allocate maximum attention_mask
attention_mask = input_ids.new_zeros(
(pb.size, max_input_length + padding_right_offset)
)
# Copy tokenizer attention_mask into fully allocated attention_mask
attention_mask[:, :max_input_length] = tokenized_inputs["attention_mask"]
position_ids = tokenized_inputs["attention_mask"].long().cumsum(-1) - 1
position_ids.masked_fill_(tokenized_inputs["attention_mask"] == 0, 1)
all_input_ids = tokenized_inputs["input_ids"].T.split(1, dim=1)
top_n_tokens_tensor = torch.tensor(
top_n_tokens, device=device, dtype=torch.int64
)
max_tokens = len(inputs) * max_input_length + max_decode_tokens
return cls(
batch_id=pb.id,
requests=pb.requests,
requests_idx_mapping=requests_idx_mapping,
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_values=None,
all_input_ids=list(all_input_ids),
input_lengths=input_lengths.tolist(),
prefix_offsets=prefix_offsets,
read_offsets=read_offsets,
next_token_choosers=next_token_choosers,
stopping_criterias=stopping_criterias,
top_n_tokens=top_n_tokens,
top_n_tokens_tensor=top_n_tokens_tensor,
max_input_length=max_input_length.item(),
padding_right_offset=padding_right_offset,
max_tokens=max_tokens,
)
|
text-generation-inference/server/text_generation_server/models/galactica.py/0
|
{
"file_path": "text-generation-inference/server/text_generation_server/models/galactica.py",
"repo_id": "text-generation-inference",
"token_count": 2499
}
| 246
|
import datetime
import torch
import os
from loguru import logger
from pathlib import Path
from safetensors.torch import save_file, load_file, _find_shared_tensors, _is_complete
from typing import List, Dict
from collections import defaultdict
def _remove_duplicate_names(
state_dict: Dict[str, torch.Tensor],
*,
preferred_names: List[str] = None,
discard_names: List[str] = None,
) -> Dict[str, List[str]]:
if preferred_names is None:
preferred_names = []
preferred_names = set(preferred_names)
if discard_names is None:
discard_names = []
discard_names = set(discard_names)
shareds = _find_shared_tensors(state_dict)
to_remove = defaultdict(list)
for shared in shareds:
complete_names = set(
[name for name in shared if _is_complete(state_dict[name])]
)
if not complete_names:
if len(shared) == 1:
# Force contiguous
name = list(shared)[0]
state_dict[name] = state_dict[name].clone()
complete_names = {name}
else:
raise RuntimeError(
f"Error while trying to find names to remove to save state dict, but found no suitable name to keep for saving amongst: {shared}. None is covering the entire storage.Refusing to save/load the model since you could be storing much more memory than needed. Please refer to https://huggingface.co/docs/safetensors/torch_shared_tensors for more information. Or open an issue."
)
keep_name = sorted(list(complete_names))[0]
# Mecanism to preferentially select keys to keep
# coming from the on-disk file to allow
# loading models saved with a different choice
# of keep_name
preferred = complete_names.difference(discard_names)
if preferred:
keep_name = sorted(list(preferred))[0]
if preferred_names:
preferred = preferred_names.intersection(complete_names)
if preferred:
keep_name = sorted(list(preferred))[0]
for name in sorted(shared):
if name != keep_name:
to_remove[keep_name].append(name)
return to_remove
def convert_file(pt_file: Path, sf_file: Path, discard_names: List[str]):
"""
Convert a pytorch file to a safetensors file
This will remove duplicate tensors from the file.
Unfortunately, this might not respect *transformers* convention.
Forcing us to check for potentially different keys during load when looking
for specific tensors (making tensor sharing explicit).
"""
loaded = torch.load(pt_file, map_location="cpu", weights_only=True)
if "state_dict" in loaded:
loaded = loaded["state_dict"]
to_removes = _remove_duplicate_names(loaded, discard_names=discard_names)
metadata = {"format": "pt"}
for kept_name, to_remove_group in to_removes.items():
for to_remove in to_remove_group:
if to_remove not in metadata:
metadata[to_remove] = kept_name
del loaded[to_remove]
# Force tensors to be contiguous
loaded = {k: v.contiguous() for k, v in loaded.items()}
dirname = os.path.dirname(sf_file)
os.makedirs(dirname, exist_ok=True)
save_file(loaded, sf_file, metadata=metadata)
reloaded = load_file(sf_file)
for k in loaded:
pt_tensor = loaded[k]
sf_tensor = reloaded[k]
if not torch.equal(pt_tensor, sf_tensor):
raise RuntimeError(f"The output tensors do not match for key {k}")
def convert_files(pt_files: List[Path], sf_files: List[Path], discard_names: List[str]):
assert len(pt_files) == len(sf_files)
N = len(pt_files)
# We do this instead of using tqdm because we want to parse the logs with the launcher
for i, (pt_file, sf_file) in enumerate(zip(pt_files, sf_files)):
# Skip blacklisted files
if (
"arguments" in pt_file.name
or "args" in pt_file.name
or "training" in pt_file.name
):
continue
start = datetime.datetime.now()
convert_file(pt_file, sf_file, discard_names)
elapsed = datetime.datetime.now() - start
logger.info(f"Convert: [{i + 1}/{N}] -- Took: {elapsed}")
|
text-generation-inference/server/text_generation_server/utils/convert.py/0
|
{
"file_path": "text-generation-inference/server/text_generation_server/utils/convert.py",
"repo_id": "text-generation-inference",
"token_count": 1775
}
| 247
|
#!/bin/bash
ldconfig 2>/dev/null || echo 'unable to refresh ld cache, not a big deal in most cases'
text-generation-launcher $@
|
text-generation-inference/tgi-entrypoint.sh/0
|
{
"file_path": "text-generation-inference/tgi-entrypoint.sh",
"repo_id": "text-generation-inference",
"token_count": 45
}
| 248
|
# This CITATION.cff file was generated with cffinit.
# Visit https://bit.ly/cffinit to generate yours today!
cff-version: 1.2.0
title: HuggingFace's Tokenizers
message: >-
Fast State-of-the-Art Tokenizers optimized for Research
and Production.
type: software
authors:
- given-names: Anthony
family-names: Moi
email: m.anthony.moi@gmail.com
affiliation: HuggingFace
- given-names: Nicolas
family-names: Patry
affiliation: HuggingFace
repository-code: 'https://github.com/huggingface/tokenizers'
url: 'https://github.com/huggingface/tokenizers'
repository: 'https://huggingface.co'
abstract: >-
Fast State-of-the-Art Tokenizers optimized for Research
and Production.
keywords:
- Rust
- Tokenizer
- NLP
license: Apache-2.0
commit: 37372b6
version: 0.13.4
date-released: '2023-04-05'
|
tokenizers/CITATION.cff/0
|
{
"file_path": "tokenizers/CITATION.cff",
"repo_id": "tokenizers",
"token_count": 293
}
| 249
|
<p align="center">
<br>
<img src="https://huggingface.co/landing/assets/tokenizers/tokenizers-logo.png" width="600"/>
<br>
<p>
<p align="center">
<a href="https://badge.fury.io/js/tokenizers">
<img alt="Build" src="https://badge.fury.io/js/tokenizers.svg">
</a>
<a href="https://github.com/huggingface/tokenizers/blob/master/LICENSE">
<img alt="GitHub" src="https://img.shields.io/github/license/huggingface/tokenizers.svg?color=blue">
</a>
</p>
<br>
NodeJS implementation of today's most used tokenizers, with a focus on performance and
versatility. Bindings over the [Rust](https://github.com/huggingface/tokenizers/tree/master/tokenizers) implementation.
If you are interested in the High-level design, you can go check it there.
## Main features
- Train new vocabularies and tokenize using 4 pre-made tokenizers (Bert WordPiece and the 3
most common BPE versions).
- Extremely fast (both training and tokenization), thanks to the Rust implementation. Takes
less than 20 seconds to tokenize a GB of text on a server's CPU.
- Easy to use, but also extremely versatile.
- Designed for research and production.
- Normalization comes with alignments tracking. It's always possible to get the part of the
original sentence that corresponds to a given token.
- Does all the pre-processing: Truncate, Pad, add the special tokens your model needs.
## Installation
```bash
npm install tokenizers@latest
```
## Basic example
```ts
import { Tokenizer } from "tokenizers";
const tokenizer = await Tokenizer.fromFile("tokenizer.json");
const wpEncoded = await tokenizer.encode("Who is John?");
console.log(wpEncoded.getLength());
console.log(wpEncoded.getTokens());
console.log(wpEncoded.getIds());
console.log(wpEncoded.getAttentionMask());
console.log(wpEncoded.getOffsets());
console.log(wpEncoded.getOverflowing());
console.log(wpEncoded.getSpecialTokensMask());
console.log(wpEncoded.getTypeIds());
console.log(wpEncoded.getWordIds());
```
## License
[Apache License 2.0](../../LICENSE)
|
tokenizers/bindings/node/README.md/0
|
{
"file_path": "tokenizers/bindings/node/README.md",
"repo_id": "tokenizers",
"token_count": 651
}
| 250
|
/* eslint-disable @typescript-eslint/no-explicit-any */
/* eslint-disable @typescript-eslint/no-empty-function */
import { TruncationStrategy, BPE, Encoding, AddedToken, Tokenizer } from '../../'
// jest.mock('../../bindings/tokenizer');
// jest.mock('../../bindings/models', () => ({
// __esModule: true,
// Model: jest.fn()
// }));
// Or:
// jest.mock('../../bindings/models', () => {
// return require('../../bindings/__mocks__/models');
// });
// const TokenizerMock = mocked(Tokenizer);
describe('AddedToken', () => {
it('instantiates with only content', () => {
const addToken = new AddedToken('test', false)
expect(addToken.constructor.name).toEqual('AddedToken')
})
it('instantiates with empty options', () => {
const addToken = new AddedToken('test', false, {})
expect(addToken.constructor.name).toEqual('AddedToken')
})
it('instantiates with options', () => {
const addToken = new AddedToken('test', false, {
leftStrip: true,
rightStrip: true,
singleWord: true,
})
expect(addToken.constructor.name).toEqual('AddedToken')
})
describe('getContent', () => {
it('returns the string content of AddedToken', () => {
const addedToken = new AddedToken('test', false)
expect(addedToken.getContent()).toEqual('test')
})
})
})
describe('Tokenizer', () => {
it('has expected methods', () => {
const model = BPE.empty()
const tokenizer = new Tokenizer(model)
expect(typeof Tokenizer.fromFile).toBe('function')
expect(typeof Tokenizer.fromString).toBe('function')
// expect(typeof Tokenizer.fromPretrained).toBe('function')
expect(typeof tokenizer.addSpecialTokens).toBe('function')
expect(typeof tokenizer.addTokens).toBe('function')
expect(typeof tokenizer.decode).toBe('function')
expect(typeof tokenizer.decodeBatch).toBe('function')
expect(typeof tokenizer.disablePadding).toBe('function')
expect(typeof tokenizer.disableTruncation).toBe('function')
expect(typeof tokenizer.encode).toBe('function')
expect(typeof tokenizer.encodeBatch).toBe('function')
expect(typeof tokenizer.getDecoder).toBe('function')
expect(typeof tokenizer.getNormalizer).toBe('function')
expect(typeof tokenizer.getPostProcessor).toBe('function')
expect(typeof tokenizer.getPreTokenizer).toBe('function')
expect(typeof tokenizer.getVocab).toBe('function')
expect(typeof tokenizer.getVocabSize).toBe('function')
expect(typeof tokenizer.idToToken).toBe('function')
expect(typeof tokenizer.runningTasks).toBe('function')
expect(typeof tokenizer.save).toBe('function')
expect(typeof tokenizer.setDecoder).toBe('function')
expect(typeof tokenizer.setModel).toBe('function')
expect(typeof tokenizer.setNormalizer).toBe('function')
expect(typeof tokenizer.setPadding).toBe('function')
expect(typeof tokenizer.setPostProcessor).toBe('function')
expect(typeof tokenizer.setPreTokenizer).toBe('function')
expect(typeof tokenizer.setTruncation).toBe('function')
expect(typeof tokenizer.tokenToId).toBe('function')
expect(typeof tokenizer.toString).toBe('function')
expect(typeof tokenizer.train).toBe('function')
})
// it('can be instantiated from the hub', async () => {
// let tokenizer: Tokenizer
// let output: Encoding
// tokenizer = Tokenizer.fromPretrained('bert-base-cased')
// output = await tokenizer.encode('Hey there dear friend!', null, { addSpecialTokens: false })
// expect(output.getTokens()).toEqual(['Hey', 'there', 'dear', 'friend', '!'])
// tokenizer = Tokenizer.fromPretrained('anthony/tokenizers-test')
// output = await tokenizer.encode('Hey there dear friend!', null, { addSpecialTokens: false })
// expect(output.getTokens()).toEqual(['hey', 'there', 'dear', 'friend', '!'])
// tokenizer = Tokenizer.fromPretrained('anthony/tokenizers-test', {
// revision: 'gpt-2',
// })
// output = await tokenizer.encode('Hey there dear friend!', null, { addSpecialTokens: false })
// expect(output.getTokens()).toEqual(['Hey', 'Ġthere', 'Ġdear', 'Ġfriend', '!'])
// }, 10000)
describe('addTokens', () => {
it('accepts a list of string as new tokens when initial model is empty', () => {
const model = BPE.empty()
const tokenizer = new Tokenizer(model)
const nbAdd = tokenizer.addTokens(['my', 'name', 'is', 'john', 'pair'])
expect(nbAdd).toBe(5)
})
it('accepts a list of AddedToken as new tokens when initial model is empty', () => {
const model = BPE.empty()
const tokenizer = new Tokenizer(model)
const addedToken = new AddedToken('test', false)
const nbAdd = tokenizer.addAddedTokens([addedToken])
expect(nbAdd).toBe(1)
})
})
describe('encode', () => {
let tokenizer: Tokenizer
beforeEach(() => {
// Clear all instances and calls to constructor and all methods:
// TokenizerMock.mockClear();
const model = BPE.empty()
tokenizer = new Tokenizer(model)
tokenizer.addTokens(['my', 'name', 'is', 'john', 'pair'])
})
it('accepts a pair of strings as parameters', async () => {
const encoding = await tokenizer.encode('my name is john', 'pair')
expect(encoding).toBeDefined()
})
it('accepts a string with a null pair', async () => {
const encoding = await tokenizer.encode('my name is john', null)
expect(encoding).toBeDefined()
})
// TODO
// it("throws if we try to encode a pre-tokenized string without isPretokenized=true", async () => {
// await expect((encode as any)(["my", "name", "is", "john"], null)).rejects.toThrow(
// "encode with isPreTokenized=false expect string"
// );
// });
// it("accepts a pre-tokenized string as parameter", async () => {
// const encoding = await tokenizer.encode(["my", "name", "is", "john"], undefined, {
// isPretokenized: true,
// });
// expect(encoding).toBeDefined();
// });
// it("throws if we try to encodeBatch pre-tokenized strings without isPretokenized=true", async () => {
// await expect((encodeBatch as any)([["my", "name", "is", "john"]])).rejects.toThrow(
// "encodeBatch with isPretokenized=false expects input to be `EncodeInput[]` " +
// "with `EncodeInput = string | [string, string]`"
// );
// });
// it("accepts a pre-tokenized input in encodeBatch", async () => {
// const encoding = await tokenizer.encodeBatch([["my", "name", "is", "john"]], {
// isPretokenized: true,
// });
// expect(encoding).toBeDefined();
// });
it('Encodes correctly if called with only one argument', async () => {
const encoded = await tokenizer.encode('my name is john')
expect(encoded.getIds()).toEqual([0, 1, 2, 3])
})
it('returns an Encoding', async () => {
const encoding = await tokenizer.encode('my name is john', 'pair')
expect(encoding.getAttentionMask()).toEqual([1, 1, 1, 1, 1])
const ids = encoding.getIds()
expect(Array.isArray(ids)).toBe(true)
expect(ids).toHaveLength(5)
for (const id of ids) {
expect(typeof id).toBe('number')
}
expect(encoding.getOffsets()).toEqual([
[0, 2],
[3, 7],
[8, 10],
[11, 15],
[0, 4],
])
expect(encoding.getOverflowing()).toEqual([])
expect(encoding.getSpecialTokensMask()).toEqual([0, 0, 0, 0, 0])
expect(encoding.getTokens()).toEqual(['my', 'name', 'is', 'john', 'pair'])
expect(encoding.getTypeIds()).toEqual([0, 0, 0, 0, 1])
})
describe('when truncation is enabled', () => {
it('truncates with default if no truncation options provided', async () => {
tokenizer.setTruncation(2)
const singleEncoding = await tokenizer.encode('my name is john', null)
expect(singleEncoding.getTokens()).toEqual(['my', 'name'])
const pairEncoding = await tokenizer.encode('my name is john', 'pair')
expect(pairEncoding.getTokens()).toEqual(['my', 'pair'])
})
it('throws an error with strategy `only_second` and no pair is encoded', async () => {
tokenizer.setTruncation(2, { strategy: TruncationStrategy.OnlySecond })
await expect(tokenizer.encode('my name is john', null)).rejects.toThrow(
'Truncation error: Second sequence not provided',
)
})
})
describe('when padding is enabled', () => {
it('does not pad anything with default options', async () => {
tokenizer.setPadding()
const singleEncoding = await tokenizer.encode('my name', null)
expect(singleEncoding.getTokens()).toEqual(['my', 'name'])
const pairEncoding = await tokenizer.encode('my name', 'pair')
expect(pairEncoding.getTokens()).toEqual(['my', 'name', 'pair'])
})
it('pads to the right by default', async () => {
tokenizer.setPadding({ maxLength: 5 })
const singleEncoding = await tokenizer.encode('my name', null)
expect(singleEncoding.getTokens()).toEqual(['my', 'name', '[PAD]', '[PAD]', '[PAD]'])
const pairEncoding = await tokenizer.encode('my name', 'pair')
expect(pairEncoding.getTokens()).toEqual(['my', 'name', 'pair', '[PAD]', '[PAD]'])
})
it('pads to multiple of the given value', async () => {
tokenizer.setPadding({ padToMultipleOf: 8 })
const singleEncoding = await tokenizer.encode('my name', null)
expect(singleEncoding.getTokens()).toHaveLength(8)
const pairEncoding = await tokenizer.encode('my name', 'pair')
expect(pairEncoding.getTokens()).toHaveLength(8)
})
})
})
describe('decode', () => {
let tokenizer: Tokenizer
beforeEach(() => {
const model = BPE.empty()
tokenizer = new Tokenizer(model)
tokenizer.addTokens(['my', 'name', 'is', 'john', 'pair'])
})
it('has its callback called with the decoded string', async () => {
const decode = tokenizer.decode.bind(tokenizer)
expect(await decode([0, 1, 2, 3], true)).toEqual('my name is john')
})
})
describe('decodeBatch', () => {
let tokenizer: Tokenizer
beforeEach(() => {
const model = BPE.empty()
tokenizer = new Tokenizer(model)
tokenizer.addTokens(['my', 'name', 'is', 'john', 'pair'])
})
it('has its callback called with the decoded string', async () => {
const decodeBatch = tokenizer.decodeBatch.bind(tokenizer)
expect(await decodeBatch([[0, 1, 2, 3], [4]], true)).toEqual(['my name is john', 'pair'])
})
})
describe('getVocab', () => {
it('accepts `undefined` as parameter', () => {
const model = BPE.empty()
const tokenizer = new Tokenizer(model)
expect(tokenizer.getVocab(undefined)).toBeDefined()
})
it('returns the vocabulary', () => {
const model = BPE.empty()
const tokenizer = new Tokenizer(model)
tokenizer.addTokens(['my', 'name', 'is', 'john'])
expect(tokenizer.getVocab(true)).toEqual({
my: 0,
name: 1,
is: 2,
john: 3,
})
})
})
describe('getVocabSize', () => {
it('accepts `undefined` as parameter', () => {
const model = BPE.empty()
const tokenizer = new Tokenizer(model)
expect(tokenizer.getVocabSize(undefined)).toBeDefined()
})
})
describe('setTruncation', () => {
it('returns the full truncation configuration', () => {
const model = BPE.empty()
const tokenizer = new Tokenizer(model)
tokenizer.setTruncation(2)
// TODO Return type is weird
// const expectedConfig: TruncationOptions = {
// maxLength: 2,
// strategy: TruncationStrategy.LongestFirst,
// stride: 0,
// direction: TruncationDirection.Right,
// };
// expect(truncation).toEqual(expectedConfig);
})
})
describe('setPadding', () => {
it('returns the full padding params', () => {
const model = BPE.empty()
const tokenizer = new Tokenizer(model)
tokenizer.setPadding()
// TODO Return type is weird
// const expectedConfig: PaddingOptions = {
// direction: PaddingDirection.Right,
// padId: 0,
// padToken: "[PAD]",
// padTypeId: 0,
// };
// expect(padding).toEqual(expectedConfig);
})
})
describe('postProcess', () => {
let tokenizer: Tokenizer
let firstEncoding: Encoding
let secondEncoding: Encoding
beforeAll(() => {
const model = BPE.empty()
tokenizer = new Tokenizer(model)
tokenizer.addTokens(['my', 'name', 'is', 'john', 'pair'])
})
beforeEach(async () => {
firstEncoding = await tokenizer.encode('my name is john', null)
secondEncoding = await tokenizer.encode('pair', null)
tokenizer.setTruncation(2)
tokenizer.setPadding({ maxLength: 5 })
})
it('returns correctly with a single Encoding param', () => {
const encoding = tokenizer.postProcess(firstEncoding)
expect(encoding.getTokens()).toEqual(['my', 'name', '[PAD]', '[PAD]', '[PAD]'])
})
it('returns correctly with `undefined` as second and third parameters', () => {
const encoding = tokenizer.postProcess(firstEncoding, undefined, undefined)
expect(encoding.getTokens()).toEqual(['my', 'name', '[PAD]', '[PAD]', '[PAD]'])
})
it('returns correctly with 2 encodings', () => {
const encoding = tokenizer.postProcess(firstEncoding, secondEncoding)
expect(encoding.getTokens()).toEqual(['my', 'pair', '[PAD]', '[PAD]', '[PAD]'])
})
})
})
|
tokenizers/bindings/node/lib/bindings/tokenizer.test.ts/0
|
{
"file_path": "tokenizers/bindings/node/lib/bindings/tokenizer.test.ts",
"repo_id": "tokenizers",
"token_count": 5268
}
| 251
|
# `tokenizers-linux-arm64-musl`
This is the **aarch64-unknown-linux-musl** binary for `tokenizers`
|
tokenizers/bindings/node/npm/linux-arm64-musl/README.md/0
|
{
"file_path": "tokenizers/bindings/node/npm/linux-arm64-musl/README.md",
"repo_id": "tokenizers",
"token_count": 37
}
| 252
|
use crate::tokenizer::PaddingOptions;
use napi::bindgen_prelude::*;
use napi_derive::napi;
use tokenizers::utils::truncation::TruncationDirection;
use tokenizers::Encoding;
#[napi(js_name = "Encoding")]
#[derive(Clone, Default)]
pub struct JsEncoding {
pub(crate) encoding: Option<Encoding>,
}
impl From<Encoding> for JsEncoding {
fn from(value: Encoding) -> Self {
Self {
encoding: Some(value),
}
}
}
impl TryFrom<JsEncoding> for Encoding {
type Error = Error;
fn try_from(value: JsEncoding) -> Result<Self> {
value
.encoding
.ok_or(Error::from_reason("Uninitialized encoding".to_string()))
}
}
#[napi(string_enum, js_name = "TruncationDirection")]
pub enum JsTruncationDirection {
Left,
Right,
}
impl From<JsTruncationDirection> for TruncationDirection {
fn from(value: JsTruncationDirection) -> Self {
match value {
JsTruncationDirection::Left => TruncationDirection::Left,
JsTruncationDirection::Right => TruncationDirection::Right,
}
}
}
impl TryFrom<String> for JsTruncationDirection {
type Error = Error;
fn try_from(value: String) -> Result<JsTruncationDirection> {
match value.as_str() {
"left" => Ok(JsTruncationDirection::Left),
"right" => Ok(JsTruncationDirection::Right),
s => Err(Error::from_reason(format!(
"{s:?} is not a valid direction"
))),
}
}
}
#[napi(string_enum, js_name = "TruncationStrategy")]
pub enum JsTruncationStrategy {
LongestFirst,
OnlyFirst,
OnlySecond,
}
impl From<JsTruncationStrategy> for tokenizers::TruncationStrategy {
fn from(value: JsTruncationStrategy) -> Self {
match value {
JsTruncationStrategy::LongestFirst => tokenizers::TruncationStrategy::LongestFirst,
JsTruncationStrategy::OnlyFirst => tokenizers::TruncationStrategy::OnlyFirst,
JsTruncationStrategy::OnlySecond => tokenizers::TruncationStrategy::OnlySecond,
}
}
}
#[napi]
impl JsEncoding {
#[napi(constructor)]
pub fn new() -> Self {
Self { encoding: None }
}
#[napi]
pub fn get_length(&self) -> u32 {
self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.get_ids()
.len() as u32
}
#[napi]
pub fn get_n_sequences(&self) -> u32 {
self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.n_sequences() as u32
}
#[napi]
pub fn get_ids(&self) -> Vec<u32> {
self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.get_ids()
.to_vec()
}
#[napi]
pub fn get_type_ids(&self) -> Vec<u32> {
self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.get_type_ids()
.to_vec()
}
#[napi]
pub fn get_attention_mask(&self) -> Vec<u32> {
self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.get_attention_mask()
.to_vec()
}
#[napi]
pub fn get_special_tokens_mask(&self) -> Vec<u32> {
self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.get_special_tokens_mask()
.to_vec()
}
#[napi]
pub fn get_tokens(&self) -> Vec<String> {
self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.get_tokens()
.to_vec()
}
#[napi]
pub fn get_offsets(&self) -> Vec<Vec<u32>> {
self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.get_offsets()
.iter()
.map(|(a, b)| vec![*a as u32, *b as u32])
.collect()
}
#[napi]
pub fn get_word_ids(&self) -> Vec<Option<u32>> {
self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.get_word_ids()
.to_vec()
}
#[napi]
pub fn char_to_token(&self, pos: u32, seq_id: Option<u32>) -> Option<u32> {
let seq_id = seq_id.unwrap_or(0);
self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.char_to_token(pos as usize, seq_id as usize)
.map(|i| i as u32)
}
#[napi]
pub fn char_to_word(&self, pos: u32, seq_id: Option<u32>) -> Option<u32> {
let seq_id = seq_id.unwrap_or(0);
self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.char_to_word(pos as usize, seq_id as usize)
}
#[napi]
pub fn pad(&mut self, length: u32, options: Option<PaddingOptions>) -> Result<()> {
let params: tokenizers::PaddingParams = options.unwrap_or_default().try_into()?;
self.encoding.as_mut().expect("Uninitialized Encoding").pad(
length as usize,
params.pad_id,
params.pad_type_id,
¶ms.pad_token,
params.direction,
);
Ok(())
}
#[napi]
pub fn truncate(
&mut self,
length: u32,
stride: Option<u32>,
direction: Option<Either<String, JsTruncationDirection>>,
) -> Result<()> {
let stride = stride.unwrap_or_default();
let direction = match direction {
None => TruncationDirection::Left,
Some(Either::A(s)) => match s.as_str() {
"left" => TruncationDirection::Left,
"right" => TruncationDirection::Right,
d => {
return Err(Error::from_reason(format!(
"{d} is not a valid truncation direction"
)));
}
},
Some(Either::B(t)) => t.into(),
};
self
.encoding
.as_mut()
.expect("Uninitialized Encoding")
.truncate(length as usize, stride as usize, direction);
Ok(())
}
#[napi(ts_return_type = "[number, number] | null | undefined")]
pub fn word_to_tokens(&self, env: Env, word: u32, seq_id: Option<u32>) -> Result<Option<Array>> {
let seq_id = seq_id.unwrap_or(0);
if let Some((a, b)) = self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.word_to_tokens(word, seq_id as usize)
{
let mut arr = env.create_array(2)?;
arr.set(0, env.create_uint32(a as u32)?)?;
arr.set(1, env.create_uint32(b as u32)?)?;
Ok(Some(arr))
} else {
Ok(None)
}
}
#[napi(ts_return_type = "[number, number] | null | undefined")]
pub fn word_to_chars(&self, env: Env, word: u32, seq_id: Option<u32>) -> Result<Option<Array>> {
let seq_id = seq_id.unwrap_or(0);
if let Some((a, b)) = self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.word_to_chars(word, seq_id as usize)
{
let mut arr = env.create_array(2)?;
arr.set(0, env.create_uint32(a as u32)?)?;
arr.set(1, env.create_uint32(b as u32)?)?;
Ok(Some(arr))
} else {
Ok(None)
}
}
#[napi(ts_return_type = "[number, [number, number]] | null | undefined")]
pub fn token_to_chars(&self, env: Env, token: u32) -> Result<Option<Array>> {
if let Some((_, (start, stop))) = self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.token_to_chars(token as usize)
{
let mut offsets = env.create_array(2)?;
offsets.set(0, env.create_uint32(start as u32)?)?;
offsets.set(1, env.create_uint32(stop as u32)?)?;
Ok(Some(offsets))
} else {
Ok(None)
}
}
#[napi]
pub fn token_to_word(&self, token: u32) -> Result<Option<u32>> {
if let Some((_, index)) = self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.token_to_word(token as usize)
{
Ok(Some(index))
} else {
Ok(None)
}
}
#[napi]
pub fn get_overflowing(&self) -> Vec<JsEncoding> {
self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.get_overflowing()
.clone()
.into_iter()
.map(|enc| JsEncoding {
encoding: Some(enc),
})
.collect()
}
#[napi]
pub fn get_sequence_ids(&self) -> Vec<Option<u32>> {
self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.get_sequence_ids()
.into_iter()
.map(|s| s.map(|id| id as u32))
.collect()
}
#[napi]
pub fn token_to_sequence(&self, token: u32) -> Option<u32> {
self
.encoding
.as_ref()
.expect("Uninitialized Encoding")
.token_to_sequence(token as usize)
.map(|s| s as u32)
}
}
|
tokenizers/bindings/node/src/encoding.rs/0
|
{
"file_path": "tokenizers/bindings/node/src/encoding.rs",
"repo_id": "tokenizers",
"token_count": 3778
}
| 253
|
from .. import decoders
Decoder = decoders.Decoder
ByteLevel = decoders.ByteLevel
Replace = decoders.Replace
WordPiece = decoders.WordPiece
ByteFallback = decoders.ByteFallback
Fuse = decoders.Fuse
Strip = decoders.Strip
Metaspace = decoders.Metaspace
BPEDecoder = decoders.BPEDecoder
CTC = decoders.CTC
Sequence = decoders.Sequence
|
tokenizers/bindings/python/py_src/tokenizers/decoders/__init__.py/0
|
{
"file_path": "tokenizers/bindings/python/py_src/tokenizers/decoders/__init__.py",
"repo_id": "tokenizers",
"token_count": 128
}
| 254
|
# Generated content DO NOT EDIT
class PostProcessor:
"""
Base class for all post-processors
This class is not supposed to be instantiated directly. Instead, any implementation of
a PostProcessor will return an instance of this class when instantiated.
"""
def num_special_tokens_to_add(self, is_pair):
"""
Return the number of special tokens that would be added for single/pair sentences.
Args:
is_pair (:obj:`bool`):
Whether the input would be a pair of sequences
Returns:
:obj:`int`: The number of tokens to add
"""
pass
def process(self, encoding, pair=None, add_special_tokens=True):
"""
Post-process the given encodings, generating the final one
Args:
encoding (:class:`~tokenizers.Encoding`):
The encoding for the first sequence
pair (:class:`~tokenizers.Encoding`, `optional`):
The encoding for the pair sequence
add_special_tokens (:obj:`bool`):
Whether to add the special tokens
Return:
:class:`~tokenizers.Encoding`: The final encoding
"""
pass
class BertProcessing(PostProcessor):
"""
This post-processor takes care of adding the special tokens needed by
a Bert model:
- a SEP token
- a CLS token
Args:
sep (:obj:`Tuple[str, int]`):
A tuple with the string representation of the SEP token, and its id
cls (:obj:`Tuple[str, int]`):
A tuple with the string representation of the CLS token, and its id
"""
def __init__(self, sep, cls):
pass
def num_special_tokens_to_add(self, is_pair):
"""
Return the number of special tokens that would be added for single/pair sentences.
Args:
is_pair (:obj:`bool`):
Whether the input would be a pair of sequences
Returns:
:obj:`int`: The number of tokens to add
"""
pass
def process(self, encoding, pair=None, add_special_tokens=True):
"""
Post-process the given encodings, generating the final one
Args:
encoding (:class:`~tokenizers.Encoding`):
The encoding for the first sequence
pair (:class:`~tokenizers.Encoding`, `optional`):
The encoding for the pair sequence
add_special_tokens (:obj:`bool`):
Whether to add the special tokens
Return:
:class:`~tokenizers.Encoding`: The final encoding
"""
pass
class ByteLevel(PostProcessor):
"""
This post-processor takes care of trimming the offsets.
By default, the ByteLevel BPE might include whitespaces in the produced tokens. If you don't
want the offsets to include these whitespaces, then this PostProcessor must be used.
Args:
trim_offsets (:obj:`bool`):
Whether to trim the whitespaces from the produced offsets.
"""
def __init__(self, trim_offsets=True):
pass
def num_special_tokens_to_add(self, is_pair):
"""
Return the number of special tokens that would be added for single/pair sentences.
Args:
is_pair (:obj:`bool`):
Whether the input would be a pair of sequences
Returns:
:obj:`int`: The number of tokens to add
"""
pass
def process(self, encoding, pair=None, add_special_tokens=True):
"""
Post-process the given encodings, generating the final one
Args:
encoding (:class:`~tokenizers.Encoding`):
The encoding for the first sequence
pair (:class:`~tokenizers.Encoding`, `optional`):
The encoding for the pair sequence
add_special_tokens (:obj:`bool`):
Whether to add the special tokens
Return:
:class:`~tokenizers.Encoding`: The final encoding
"""
pass
class RobertaProcessing(PostProcessor):
"""
This post-processor takes care of adding the special tokens needed by
a Roberta model:
- a SEP token
- a CLS token
It also takes care of trimming the offsets.
By default, the ByteLevel BPE might include whitespaces in the produced tokens. If you don't
want the offsets to include these whitespaces, then this PostProcessor should be initialized
with :obj:`trim_offsets=True`
Args:
sep (:obj:`Tuple[str, int]`):
A tuple with the string representation of the SEP token, and its id
cls (:obj:`Tuple[str, int]`):
A tuple with the string representation of the CLS token, and its id
trim_offsets (:obj:`bool`, `optional`, defaults to :obj:`True`):
Whether to trim the whitespaces from the produced offsets.
add_prefix_space (:obj:`bool`, `optional`, defaults to :obj:`True`):
Whether the add_prefix_space option was enabled during pre-tokenization. This
is relevant because it defines the way the offsets are trimmed out.
"""
def __init__(self, sep, cls, trim_offsets=True, add_prefix_space=True):
pass
def num_special_tokens_to_add(self, is_pair):
"""
Return the number of special tokens that would be added for single/pair sentences.
Args:
is_pair (:obj:`bool`):
Whether the input would be a pair of sequences
Returns:
:obj:`int`: The number of tokens to add
"""
pass
def process(self, encoding, pair=None, add_special_tokens=True):
"""
Post-process the given encodings, generating the final one
Args:
encoding (:class:`~tokenizers.Encoding`):
The encoding for the first sequence
pair (:class:`~tokenizers.Encoding`, `optional`):
The encoding for the pair sequence
add_special_tokens (:obj:`bool`):
Whether to add the special tokens
Return:
:class:`~tokenizers.Encoding`: The final encoding
"""
pass
class Sequence(PostProcessor):
"""
Sequence Processor
Args:
processors (:obj:`List[PostProcessor]`)
The processors that need to be chained
"""
def __init__(self, processors):
pass
def num_special_tokens_to_add(self, is_pair):
"""
Return the number of special tokens that would be added for single/pair sentences.
Args:
is_pair (:obj:`bool`):
Whether the input would be a pair of sequences
Returns:
:obj:`int`: The number of tokens to add
"""
pass
def process(self, encoding, pair=None, add_special_tokens=True):
"""
Post-process the given encodings, generating the final one
Args:
encoding (:class:`~tokenizers.Encoding`):
The encoding for the first sequence
pair (:class:`~tokenizers.Encoding`, `optional`):
The encoding for the pair sequence
add_special_tokens (:obj:`bool`):
Whether to add the special tokens
Return:
:class:`~tokenizers.Encoding`: The final encoding
"""
pass
class TemplateProcessing(PostProcessor):
"""
Provides a way to specify templates in order to add the special tokens to each
input sequence as relevant.
Let's take :obj:`BERT` tokenizer as an example. It uses two special tokens, used to
delimitate each sequence. :obj:`[CLS]` is always used at the beginning of the first
sequence, and :obj:`[SEP]` is added at the end of both the first, and the pair
sequences. The final result looks like this:
- Single sequence: :obj:`[CLS] Hello there [SEP]`
- Pair sequences: :obj:`[CLS] My name is Anthony [SEP] What is my name? [SEP]`
With the type ids as following::
[CLS] ... [SEP] ... [SEP]
0 0 0 1 1
You can achieve such behavior using a TemplateProcessing::
TemplateProcessing(
single="[CLS] $0 [SEP]",
pair="[CLS] $A [SEP] $B:1 [SEP]:1",
special_tokens=[("[CLS]", 1), ("[SEP]", 0)],
)
In this example, each input sequence is identified using a ``$`` construct. This identifier
lets us specify each input sequence, and the type_id to use. When nothing is specified,
it uses the default values. Here are the different ways to specify it:
- Specifying the sequence, with default ``type_id == 0``: ``$A`` or ``$B``
- Specifying the `type_id` with default ``sequence == A``: ``$0``, ``$1``, ``$2``, ...
- Specifying both: ``$A:0``, ``$B:1``, ...
The same construct is used for special tokens: ``<identifier>(:<type_id>)?``.
**Warning**: You must ensure that you are giving the correct tokens/ids as these
will be added to the Encoding without any further check. If the given ids correspond
to something totally different in a `Tokenizer` using this `PostProcessor`, it
might lead to unexpected results.
Args:
single (:obj:`Template`):
The template used for single sequences
pair (:obj:`Template`):
The template used when both sequences are specified
special_tokens (:obj:`Tokens`):
The list of special tokens used in each sequences
Types:
Template (:obj:`str` or :obj:`List`):
- If a :obj:`str` is provided, the whitespace is used as delimiter between tokens
- If a :obj:`List[str]` is provided, a list of tokens
Tokens (:obj:`List[Union[Tuple[int, str], Tuple[str, int], dict]]`):
- A :obj:`Tuple` with both a token and its associated ID, in any order
- A :obj:`dict` with the following keys:
- "id": :obj:`str` => The special token id, as specified in the Template
- "ids": :obj:`List[int]` => The associated IDs
- "tokens": :obj:`List[str]` => The associated tokens
The given dict expects the provided :obj:`ids` and :obj:`tokens` lists to have
the same length.
"""
def __init__(self, single, pair, special_tokens):
pass
def num_special_tokens_to_add(self, is_pair):
"""
Return the number of special tokens that would be added for single/pair sentences.
Args:
is_pair (:obj:`bool`):
Whether the input would be a pair of sequences
Returns:
:obj:`int`: The number of tokens to add
"""
pass
def process(self, encoding, pair=None, add_special_tokens=True):
"""
Post-process the given encodings, generating the final one
Args:
encoding (:class:`~tokenizers.Encoding`):
The encoding for the first sequence
pair (:class:`~tokenizers.Encoding`, `optional`):
The encoding for the pair sequence
add_special_tokens (:obj:`bool`):
Whether to add the special tokens
Return:
:class:`~tokenizers.Encoding`: The final encoding
"""
pass
|
tokenizers/bindings/python/py_src/tokenizers/processors/__init__.pyi/0
|
{
"file_path": "tokenizers/bindings/python/py_src/tokenizers/processors/__init__.pyi",
"repo_id": "tokenizers",
"token_count": 4779
}
| 255
|
use std::collections::HashMap;
use std::path::{Path, PathBuf};
use std::sync::{Arc, RwLock};
use crate::token::PyToken;
use crate::trainers::PyTrainer;
use pyo3::exceptions;
use pyo3::prelude::*;
use pyo3::types::*;
use serde::{Deserialize, Serialize};
use tk::models::bpe::{BpeBuilder, Merges, Vocab, BPE};
use tk::models::unigram::Unigram;
use tk::models::wordlevel::WordLevel;
use tk::models::wordpiece::{WordPiece, WordPieceBuilder};
use tk::models::ModelWrapper;
use tk::{Model, Token};
use tokenizers as tk;
use super::error::{deprecation_warning, ToPyResult};
/// Base class for all models
///
/// The model represents the actual tokenization algorithm. This is the part that
/// will contain and manage the learned vocabulary.
///
/// This class cannot be constructed directly. Please use one of the concrete models.
#[pyclass(module = "tokenizers.models", name = "Model", subclass)]
#[derive(Clone, Serialize, Deserialize)]
#[serde(transparent)]
pub struct PyModel {
pub model: Arc<RwLock<ModelWrapper>>,
}
impl PyModel {
pub(crate) fn get_as_subtype(&self, py: Python<'_>) -> PyResult<PyObject> {
let base = self.clone();
Ok(match *self.model.as_ref().read().unwrap() {
ModelWrapper::BPE(_) => Py::new(py, (PyBPE {}, base))?.into_py(py),
ModelWrapper::WordPiece(_) => Py::new(py, (PyWordPiece {}, base))?.into_py(py),
ModelWrapper::WordLevel(_) => Py::new(py, (PyWordLevel {}, base))?.into_py(py),
ModelWrapper::Unigram(_) => Py::new(py, (PyUnigram {}, base))?.into_py(py),
})
}
}
impl Model for PyModel {
type Trainer = PyTrainer;
fn tokenize(&self, tokens: &str) -> tk::Result<Vec<Token>> {
self.model.read().unwrap().tokenize(tokens)
}
fn token_to_id(&self, token: &str) -> Option<u32> {
self.model.read().unwrap().token_to_id(token)
}
fn id_to_token(&self, id: u32) -> Option<String> {
self.model.read().unwrap().id_to_token(id)
}
fn get_vocab(&self) -> HashMap<String, u32> {
self.model.read().unwrap().get_vocab()
}
fn get_vocab_size(&self) -> usize {
self.model.read().unwrap().get_vocab_size()
}
fn save(&self, folder: &Path, name: Option<&str>) -> tk::Result<Vec<PathBuf>> {
self.model.read().unwrap().save(folder, name)
}
fn get_trainer(&self) -> Self::Trainer {
self.model.read().unwrap().get_trainer().into()
}
}
impl<I> From<I> for PyModel
where
I: Into<ModelWrapper>,
{
fn from(model: I) -> Self {
Self {
model: Arc::new(RwLock::new(model.into())),
}
}
}
#[pymethods]
impl PyModel {
#[new]
#[pyo3(text_signature = None)]
fn __new__() -> Self {
// Instantiate a default empty model. This doesn't really make sense, but we need
// to be able to instantiate an empty model for pickle capabilities.
PyModel {
model: Arc::new(RwLock::new(BPE::default().into())),
}
}
fn __getstate__(&self, py: Python) -> PyResult<PyObject> {
let data = serde_json::to_string(&self.model).map_err(|e| {
exceptions::PyException::new_err(format!(
"Error while attempting to pickle Model: {}",
e
))
})?;
Ok(PyBytes::new_bound(py, data.as_bytes()).to_object(py))
}
fn __setstate__(&mut self, py: Python, state: PyObject) -> PyResult<()> {
match state.extract::<&PyBytes>(py) {
Ok(s) => {
self.model = serde_json::from_slice(s.as_bytes()).map_err(|e| {
exceptions::PyException::new_err(format!(
"Error while attempting to unpickle Model: {}",
e
))
})?;
Ok(())
}
Err(e) => Err(e),
}
}
/// Tokenize a sequence
///
/// Args:
/// sequence (:obj:`str`):
/// A sequence to tokenize
///
/// Returns:
/// A :obj:`List` of :class:`~tokenizers.Token`: The generated tokens
#[pyo3(text_signature = "(self, sequence)")]
fn tokenize(&self, sequence: &str) -> PyResult<Vec<PyToken>> {
Ok(ToPyResult(self.model.read().unwrap().tokenize(sequence))
.into_py()?
.into_iter()
.map(|t| t.into())
.collect())
}
/// Get the ID associated to a token
///
/// Args:
/// token (:obj:`str`):
/// A token to convert to an ID
///
/// Returns:
/// :obj:`int`: The ID associated to the token
#[pyo3(text_signature = "(self, tokens)")]
fn token_to_id(&self, token: &str) -> Option<u32> {
self.model.read().unwrap().token_to_id(token)
}
/// Get the token associated to an ID
///
/// Args:
/// id (:obj:`int`):
/// An ID to convert to a token
///
/// Returns:
/// :obj:`str`: The token associated to the ID
#[pyo3(text_signature = "(self, id)")]
fn id_to_token(&self, id: u32) -> Option<String> {
self.model.read().unwrap().id_to_token(id)
}
/// Save the current model
///
/// Save the current model in the given folder, using the given prefix for the various
/// files that will get created.
/// Any file with the same name that already exists in this folder will be overwritten.
///
/// Args:
/// folder (:obj:`str`):
/// The path to the target folder in which to save the various files
///
/// prefix (:obj:`str`, `optional`):
/// An optional prefix, used to prefix each file name
///
/// Returns:
/// :obj:`List[str]`: The list of saved files
#[pyo3(text_signature = "(self, folder, prefix)")]
fn save<'a>(
&self,
py: Python<'_>,
folder: &str,
mut prefix: Option<&'a str>,
name: Option<&'a str>,
) -> PyResult<Vec<String>> {
if name.is_some() {
deprecation_warning(
py,
"0.10.0",
"Parameter `name` of Model.save has been renamed `prefix`",
)?;
if prefix.is_none() {
prefix = name;
}
}
let saved: PyResult<Vec<_>> =
ToPyResult(self.model.read().unwrap().save(Path::new(folder), prefix)).into();
Ok(saved?
.into_iter()
.map(|path| path.to_string_lossy().into_owned())
.collect())
}
/// Get the associated :class:`~tokenizers.trainers.Trainer`
///
/// Retrieve the :class:`~tokenizers.trainers.Trainer` associated to this
/// :class:`~tokenizers.models.Model`.
///
/// Returns:
/// :class:`~tokenizers.trainers.Trainer`: The Trainer used to train this model
#[pyo3(text_signature = "(self)")]
fn get_trainer(&self, py: Python<'_>) -> PyResult<PyObject> {
PyTrainer::from(self.model.read().unwrap().get_trainer()).get_as_subtype(py)
}
fn __repr__(&self) -> PyResult<String> {
crate::utils::serde_pyo3::repr(self)
.map_err(|e| exceptions::PyException::new_err(e.to_string()))
}
fn __str__(&self) -> PyResult<String> {
crate::utils::serde_pyo3::to_string(self)
.map_err(|e| exceptions::PyException::new_err(e.to_string()))
}
}
/// An implementation of the BPE (Byte-Pair Encoding) algorithm
///
/// Args:
/// vocab (:obj:`Dict[str, int]`, `optional`):
/// A dictionary of string keys and their ids :obj:`{"am": 0,...}`
///
/// merges (:obj:`List[Tuple[str, str]]`, `optional`):
/// A list of pairs of tokens (:obj:`Tuple[str, str]`) :obj:`[("a", "b"),...]`
///
/// cache_capacity (:obj:`int`, `optional`):
/// The number of words that the BPE cache can contain. The cache allows
/// to speed-up the process by keeping the result of the merge operations
/// for a number of words.
///
/// dropout (:obj:`float`, `optional`):
/// A float between 0 and 1 that represents the BPE dropout to use.
///
/// unk_token (:obj:`str`, `optional`):
/// The unknown token to be used by the model.
///
/// continuing_subword_prefix (:obj:`str`, `optional`):
/// The prefix to attach to subword units that don't represent a beginning of word.
///
/// end_of_word_suffix (:obj:`str`, `optional`):
/// The suffix to attach to subword units that represent an end of word.
///
/// fuse_unk (:obj:`bool`, `optional`):
/// Whether to fuse any subsequent unknown tokens into a single one
///
/// byte_fallback (:obj:`bool`, `optional`):
/// Whether to use spm byte-fallback trick (defaults to False)
///
/// ignore_merges (:obj:`bool`, `optional`):
/// Whether or not to match tokens with the vocab before using merges.
#[pyclass(extends=PyModel, module = "tokenizers.models", name = "BPE")]
pub struct PyBPE {}
impl PyBPE {
fn with_builder(
mut builder: BpeBuilder,
kwargs: Option<&Bound<'_, PyDict>>,
) -> PyResult<(Self, PyModel)> {
if let Some(kwargs) = kwargs {
for (key, value) in kwargs {
let key: &str = key.extract()?;
match key {
"cache_capacity" => builder = builder.cache_capacity(value.extract()?),
"dropout" => {
if let Some(dropout) = value.extract()? {
builder = builder.dropout(dropout);
}
}
"unk_token" => {
if let Some(unk) = value.extract()? {
builder = builder.unk_token(unk);
}
}
"continuing_subword_prefix" => {
builder = builder.continuing_subword_prefix(value.extract()?)
}
"end_of_word_suffix" => builder = builder.end_of_word_suffix(value.extract()?),
"fuse_unk" => builder = builder.fuse_unk(value.extract()?),
"byte_fallback" => builder = builder.byte_fallback(value.extract()?),
"ignore_merges" => builder = builder.ignore_merges(value.extract()?),
_ => println!("Ignored unknown kwarg option {}", key),
};
}
}
match builder.build() {
Err(e) => Err(exceptions::PyException::new_err(format!(
"Error while initializing BPE: {}",
e
))),
Ok(bpe) => Ok((PyBPE {}, bpe.into())),
}
}
}
macro_rules! getter {
($self: ident, $variant: ident, $($name: tt)+) => {{
let super_ = $self.as_ref();
let model = super_.model.read().unwrap();
if let ModelWrapper::$variant(ref mo) = *model {
mo.$($name)+
} else {
unreachable!()
}
}};
}
macro_rules! setter {
($self: ident, $variant: ident, $name: ident, $value: expr) => {{
let super_ = $self.as_ref();
let mut model = super_.model.write().unwrap();
if let ModelWrapper::$variant(ref mut mo) = *model {
mo.$name = $value;
}
}};
}
#[derive(FromPyObject)]
enum PyVocab {
Vocab(Vocab),
Filename(String),
}
#[derive(FromPyObject)]
enum PyMerges {
Merges(Merges),
Filename(String),
}
#[pymethods]
impl PyBPE {
#[getter]
fn get_dropout(self_: PyRef<Self>) -> Option<f32> {
getter!(self_, BPE, dropout)
}
#[setter]
fn set_dropout(self_: PyRef<Self>, dropout: Option<f32>) {
setter!(self_, BPE, dropout, dropout);
}
#[getter]
fn get_unk_token(self_: PyRef<Self>) -> Option<String> {
getter!(self_, BPE, unk_token.clone())
}
#[setter]
fn set_unk_token(self_: PyRef<Self>, unk_token: Option<String>) {
setter!(self_, BPE, unk_token, unk_token);
}
#[getter]
fn get_continuing_subword_prefix(self_: PyRef<Self>) -> Option<String> {
getter!(self_, BPE, continuing_subword_prefix.clone())
}
#[setter]
fn set_continuing_subword_prefix(
self_: PyRef<Self>,
continuing_subword_prefix: Option<String>,
) {
setter!(
self_,
BPE,
continuing_subword_prefix,
continuing_subword_prefix
);
}
#[getter]
fn get_end_of_word_suffix(self_: PyRef<Self>) -> Option<String> {
getter!(self_, BPE, end_of_word_suffix.clone())
}
#[setter]
fn set_end_of_word_suffix(self_: PyRef<Self>, end_of_word_suffix: Option<String>) {
setter!(self_, BPE, end_of_word_suffix, end_of_word_suffix);
}
#[getter]
fn get_fuse_unk(self_: PyRef<Self>) -> bool {
getter!(self_, BPE, fuse_unk)
}
#[setter]
fn set_fuse_unk(self_: PyRef<Self>, fuse_unk: bool) {
setter!(self_, BPE, fuse_unk, fuse_unk);
}
#[getter]
fn get_byte_fallback(self_: PyRef<Self>) -> bool {
getter!(self_, BPE, byte_fallback)
}
#[setter]
fn set_byte_fallback(self_: PyRef<Self>, byte_fallback: bool) {
setter!(self_, BPE, byte_fallback, byte_fallback);
}
#[getter]
fn get_ignore_merges(self_: PyRef<Self>) -> bool {
getter!(self_, BPE, ignore_merges)
}
#[setter]
fn set_ignore_merges(self_: PyRef<Self>, ignore_merges: bool) {
setter!(self_, BPE, ignore_merges, ignore_merges);
}
#[new]
#[pyo3(
signature = (vocab=None, merges=None, **kwargs),
text_signature = "(self, vocab=None, merges=None, cache_capacity=None, dropout=None, unk_token=None, continuing_subword_prefix=None, end_of_word_suffix=None, fuse_unk=None, byte_fallback=False, ignore_merges=False)")]
fn new(
py: Python<'_>,
vocab: Option<PyVocab>,
merges: Option<PyMerges>,
kwargs: Option<&Bound<'_, PyDict>>,
) -> PyResult<(Self, PyModel)> {
if (vocab.is_some() && merges.is_none()) || (vocab.is_none() && merges.is_some()) {
return Err(exceptions::PyValueError::new_err(
"`vocab` and `merges` must be both specified",
));
}
let mut builder = BPE::builder();
if let (Some(vocab), Some(merges)) = (vocab, merges) {
match (vocab, merges) {
(PyVocab::Vocab(vocab), PyMerges::Merges(merges)) => {
builder = builder.vocab_and_merges(vocab, merges);
}
(PyVocab::Filename(vocab_filename), PyMerges::Filename(merges_filename)) => {
deprecation_warning(
py,
"0.9.0",
"BPE.__init__ will not create from files anymore, try `BPE.from_file` instead",
)?;
builder =
builder.files(vocab_filename.to_string(), merges_filename.to_string());
}
_ => {
return Err(exceptions::PyValueError::new_err(
"`vocab` and `merges` must be both be from memory or both filenames",
));
}
}
}
PyBPE::with_builder(builder, kwargs)
}
/// Read a :obj:`vocab.json` and a :obj:`merges.txt` files
///
/// This method provides a way to read and parse the content of these files,
/// returning the relevant data structures. If you want to instantiate some BPE models
/// from memory, this method gives you the expected input from the standard files.
///
/// Args:
/// vocab (:obj:`str`):
/// The path to a :obj:`vocab.json` file
///
/// merges (:obj:`str`):
/// The path to a :obj:`merges.txt` file
///
/// Returns:
/// A :obj:`Tuple` with the vocab and the merges:
/// The vocabulary and merges loaded into memory
#[staticmethod]
#[pyo3(text_signature = "(self, vocab, merges)")]
fn read_file(vocab: &str, merges: &str) -> PyResult<(Vocab, Merges)> {
BPE::read_file(vocab, merges).map_err(|e| {
exceptions::PyException::new_err(format!(
"Error while reading vocab & merges files: {}",
e
))
})
}
/// Instantiate a BPE model from the given files.
///
/// This method is roughly equivalent to doing::
///
/// vocab, merges = BPE.read_file(vocab_filename, merges_filename)
/// bpe = BPE(vocab, merges)
///
/// If you don't need to keep the :obj:`vocab, merges` values lying around,
/// this method is more optimized than manually calling
/// :meth:`~tokenizers.models.BPE.read_file` to initialize a :class:`~tokenizers.models.BPE`
///
/// Args:
/// vocab (:obj:`str`):
/// The path to a :obj:`vocab.json` file
///
/// merges (:obj:`str`):
/// The path to a :obj:`merges.txt` file
///
/// Returns:
/// :class:`~tokenizers.models.BPE`: An instance of BPE loaded from these files
#[classmethod]
#[pyo3(signature = (vocab, merges, **kwargs))]
#[pyo3(text_signature = "(cls, vocab, merge, **kwargs)")]
fn from_file(
_cls: &Bound<'_, PyType>,
py: Python,
vocab: &str,
merges: &str,
kwargs: Option<&Bound<'_, PyDict>>,
) -> PyResult<Py<Self>> {
let (vocab, merges) = BPE::read_file(vocab, merges).map_err(|e| {
exceptions::PyException::new_err(format!("Error while reading BPE files: {}", e))
})?;
Py::new(
py,
PyBPE::new(
py,
Some(PyVocab::Vocab(vocab)),
Some(PyMerges::Merges(merges)),
kwargs,
)?,
)
}
}
/// An implementation of the WordPiece algorithm
///
/// Args:
/// vocab (:obj:`Dict[str, int]`, `optional`):
/// A dictionary of string keys and their ids :obj:`{"am": 0,...}`
///
/// unk_token (:obj:`str`, `optional`):
/// The unknown token to be used by the model.
///
/// max_input_chars_per_word (:obj:`int`, `optional`):
/// The maximum number of characters to authorize in a single word.
#[pyclass(extends=PyModel, module = "tokenizers.models", name = "WordPiece")]
pub struct PyWordPiece {}
impl PyWordPiece {
fn with_builder(
mut builder: WordPieceBuilder,
kwargs: Option<&Bound<'_, PyDict>>,
) -> PyResult<(Self, PyModel)> {
if let Some(kwargs) = kwargs {
for (key, val) in kwargs {
let key: &str = key.extract()?;
match key {
"unk_token" => {
builder = builder.unk_token(val.extract()?);
}
"max_input_chars_per_word" => {
builder = builder.max_input_chars_per_word(val.extract()?);
}
"continuing_subword_prefix" => {
builder = builder.continuing_subword_prefix(val.extract()?);
}
_ => println!("Ignored unknown kwargs option {}", key),
}
}
}
match builder.build() {
Err(e) => Err(exceptions::PyException::new_err(format!(
"Error while initializing WordPiece: {}",
e
))),
Ok(wordpiece) => Ok((PyWordPiece {}, wordpiece.into())),
}
}
}
#[pymethods]
impl PyWordPiece {
#[getter]
fn get_unk_token(self_: PyRef<Self>) -> String {
getter!(self_, WordPiece, unk_token.clone())
}
#[setter]
fn set_unk_token(self_: PyRef<Self>, unk_token: String) {
setter!(self_, WordPiece, unk_token, unk_token);
}
#[getter]
fn get_continuing_subword_prefix(self_: PyRef<Self>) -> String {
getter!(self_, WordPiece, continuing_subword_prefix.clone())
}
#[setter]
fn set_continuing_subword_prefix(self_: PyRef<Self>, continuing_subword_prefix: String) {
setter!(
self_,
WordPiece,
continuing_subword_prefix,
continuing_subword_prefix
);
}
#[getter]
fn get_max_input_chars_per_word(self_: PyRef<Self>) -> usize {
getter!(self_, WordPiece, max_input_chars_per_word)
}
#[setter]
fn set_max_input_chars_per_word(self_: PyRef<Self>, max: usize) {
setter!(self_, WordPiece, max_input_chars_per_word, max);
}
#[new]
#[pyo3(signature = (vocab=None, **kwargs), text_signature = "(self, vocab, unk_token, max_input_chars_per_word)")]
fn new(
py: Python<'_>,
vocab: Option<PyVocab>,
kwargs: Option<&Bound<'_, PyDict>>,
) -> PyResult<(Self, PyModel)> {
let mut builder = WordPiece::builder();
if let Some(vocab) = vocab {
match vocab {
PyVocab::Vocab(vocab) => {
builder = builder.vocab(vocab);
}
PyVocab::Filename(vocab_filename) => {
deprecation_warning(
py,
"0.9.0",
"WordPiece.__init__ will not create from files anymore, try `WordPiece.from_file` instead",
)?;
builder = builder.files(vocab_filename.to_string());
}
}
}
PyWordPiece::with_builder(builder, kwargs)
}
/// Read a :obj:`vocab.txt` file
///
/// This method provides a way to read and parse the content of a standard `vocab.txt`
/// file as used by the WordPiece Model, returning the relevant data structures. If you
/// want to instantiate some WordPiece models from memory, this method gives you the
/// expected input from the standard files.
///
/// Args:
/// vocab (:obj:`str`):
/// The path to a :obj:`vocab.txt` file
///
/// Returns:
/// :obj:`Dict[str, int]`: The vocabulary as a :obj:`dict`
#[staticmethod]
#[pyo3(text_signature = "(vocab)")]
fn read_file(vocab: &str) -> PyResult<Vocab> {
WordPiece::read_file(vocab).map_err(|e| {
exceptions::PyException::new_err(format!("Error while reading WordPiece file: {}", e))
})
}
/// Instantiate a WordPiece model from the given file
///
/// This method is roughly equivalent to doing::
///
/// vocab = WordPiece.read_file(vocab_filename)
/// wordpiece = WordPiece(vocab)
///
/// If you don't need to keep the :obj:`vocab` values lying around, this method is
/// more optimized than manually calling :meth:`~tokenizers.models.WordPiece.read_file` to
/// initialize a :class:`~tokenizers.models.WordPiece`
///
/// Args:
/// vocab (:obj:`str`):
/// The path to a :obj:`vocab.txt` file
///
/// Returns:
/// :class:`~tokenizers.models.WordPiece`: An instance of WordPiece loaded from file
#[classmethod]
#[pyo3(signature = (vocab, **kwargs))]
#[pyo3(text_signature = "(vocab, **kwargs)")]
fn from_file(
_cls: &Bound<'_, PyType>,
py: Python,
vocab: &str,
kwargs: Option<&Bound<'_, PyDict>>,
) -> PyResult<Py<Self>> {
let vocab = WordPiece::read_file(vocab).map_err(|e| {
exceptions::PyException::new_err(format!("Error while reading WordPiece file: {}", e))
})?;
Py::new(
py,
PyWordPiece::new(py, Some(PyVocab::Vocab(vocab)), kwargs)?,
)
}
}
/// An implementation of the WordLevel algorithm
///
/// Most simple tokenizer model based on mapping tokens to their corresponding id.
///
/// Args:
/// vocab (:obj:`str`, `optional`):
/// A dictionary of string keys and their ids :obj:`{"am": 0,...}`
///
/// unk_token (:obj:`str`, `optional`):
/// The unknown token to be used by the model.
#[pyclass(extends=PyModel, module = "tokenizers.models", name = "WordLevel")]
pub struct PyWordLevel {}
#[pymethods]
impl PyWordLevel {
#[getter]
fn get_unk_token(self_: PyRef<Self>) -> String {
getter!(self_, WordLevel, unk_token.clone())
}
#[setter]
fn set_unk_token(self_: PyRef<Self>, unk_token: String) {
setter!(self_, WordLevel, unk_token, unk_token);
}
#[new]
#[pyo3(signature = (vocab=None, unk_token = None), text_signature = "(self, vocab, unk_token)")]
fn new(
py: Python<'_>,
vocab: Option<PyVocab>,
unk_token: Option<String>,
) -> PyResult<(Self, PyModel)> {
let mut builder = WordLevel::builder();
if let Some(vocab) = vocab {
match vocab {
PyVocab::Vocab(vocab) => {
builder = builder.vocab(vocab);
}
PyVocab::Filename(vocab_filename) => {
deprecation_warning(
py,
"0.9.0",
"WordLevel.__init__ will not create from files anymore, \
try `WordLevel.from_file` instead",
)?;
builder = builder.files(vocab_filename.to_string());
}
};
}
if let Some(unk_token) = unk_token {
builder = builder.unk_token(unk_token);
}
Ok((
PyWordLevel {},
builder
.build()
.map_err(|e| exceptions::PyException::new_err(e.to_string()))?
.into(),
))
}
/// Read a :obj:`vocab.json`
///
/// This method provides a way to read and parse the content of a vocabulary file,
/// returning the relevant data structures. If you want to instantiate some WordLevel models
/// from memory, this method gives you the expected input from the standard files.
///
/// Args:
/// vocab (:obj:`str`):
/// The path to a :obj:`vocab.json` file
///
/// Returns:
/// :obj:`Dict[str, int]`: The vocabulary as a :obj:`dict`
#[staticmethod]
#[pyo3(text_signature = "(vocab)")]
fn read_file(vocab: &str) -> PyResult<Vocab> {
WordLevel::read_file(vocab).map_err(|e| {
exceptions::PyException::new_err(format!("Error while reading WordLevel file: {}", e))
})
}
/// Instantiate a WordLevel model from the given file
///
/// This method is roughly equivalent to doing::
///
/// vocab = WordLevel.read_file(vocab_filename)
/// wordlevel = WordLevel(vocab)
///
/// If you don't need to keep the :obj:`vocab` values lying around, this method is
/// more optimized than manually calling :meth:`~tokenizers.models.WordLevel.read_file` to
/// initialize a :class:`~tokenizers.models.WordLevel`
///
/// Args:
/// vocab (:obj:`str`):
/// The path to a :obj:`vocab.json` file
///
/// Returns:
/// :class:`~tokenizers.models.WordLevel`: An instance of WordLevel loaded from file
#[classmethod]
#[pyo3(signature = (vocab, unk_token = None))]
#[pyo3(text_signature = "(vocab, unk_token)")]
fn from_file(
_cls: &Bound<'_, PyType>,
py: Python,
vocab: &str,
unk_token: Option<String>,
) -> PyResult<Py<Self>> {
let vocab = WordLevel::read_file(vocab).map_err(|e| {
exceptions::PyException::new_err(format!("Error while reading WordLevel file: {}", e))
})?;
Py::new(
py,
PyWordLevel::new(py, Some(PyVocab::Vocab(vocab)), unk_token)?,
)
}
}
/// An implementation of the Unigram algorithm
///
/// Args:
/// vocab (:obj:`List[Tuple[str, float]]`, `optional`, `optional`):
/// A list of vocabulary items and their relative score [("am", -0.2442),...]
#[pyclass(extends=PyModel, module = "tokenizers.models", name = "Unigram")]
pub struct PyUnigram {}
#[pymethods]
impl PyUnigram {
#[new]
#[pyo3(text_signature = "(self, vocab, unk_id, byte_fallback)")]
fn new(
vocab: Option<Vec<(String, f64)>>,
unk_id: Option<usize>,
byte_fallback: Option<bool>,
) -> PyResult<(Self, PyModel)> {
match (vocab, unk_id, byte_fallback) {
(Some(vocab), unk_id, byte_fallback) => {
let model =
Unigram::from(vocab, unk_id, byte_fallback.unwrap_or(false)).map_err(|e| {
exceptions::PyException::new_err(format!(
"Error while loading Unigram: {}",
e
))
})?;
Ok((PyUnigram {}, model.into()))
}
(None, None, _) => Ok((PyUnigram {}, Unigram::default().into())),
_ => Err(exceptions::PyValueError::new_err(
"`vocab` and `unk_id` must be both specified",
)),
}
}
}
/// Models Module
#[pymodule]
pub fn models(m: &Bound<'_, PyModule>) -> PyResult<()> {
m.add_class::<PyModel>()?;
m.add_class::<PyBPE>()?;
m.add_class::<PyWordPiece>()?;
m.add_class::<PyWordLevel>()?;
m.add_class::<PyUnigram>()?;
Ok(())
}
#[cfg(test)]
mod test {
use crate::models::PyModel;
use pyo3::prelude::*;
use tk::models::bpe::BPE;
use tk::models::ModelWrapper;
#[test]
fn get_subtype() {
Python::with_gil(|py| {
let py_model = PyModel::from(BPE::default());
let py_bpe = py_model.get_as_subtype(py).unwrap();
assert_eq!("BPE", py_bpe.bind(py).get_type().qualname().unwrap());
})
}
#[test]
fn serialize() {
let rs_bpe = BPE::default();
let rs_bpe_ser = serde_json::to_string(&rs_bpe).unwrap();
let rs_wrapper: ModelWrapper = rs_bpe.into();
let rs_wrapper_ser = serde_json::to_string(&rs_wrapper).unwrap();
let py_model = PyModel::from(rs_wrapper);
let py_ser = serde_json::to_string(&py_model).unwrap();
assert_eq!(py_ser, rs_bpe_ser);
assert_eq!(py_ser, rs_wrapper_ser);
let py_model: PyModel = serde_json::from_str(&rs_bpe_ser).unwrap();
match *py_model.model.as_ref().read().unwrap() {
ModelWrapper::BPE(_) => (),
_ => panic!("Expected Bert postprocessor."),
};
let py_model: PyModel = serde_json::from_str(&rs_wrapper_ser).unwrap();
match *py_model.model.as_ref().read().unwrap() {
ModelWrapper::BPE(_) => (),
_ => panic!("Expected Bert postprocessor."),
};
}
}
|
tokenizers/bindings/python/src/models.rs/0
|
{
"file_path": "tokenizers/bindings/python/src/models.rs",
"repo_id": "tokenizers",
"token_count": 14893
}
| 256
|
from tokenizers import ByteLevelBPETokenizer
from ..utils import data_dir, multiprocessing_with_parallelism, roberta_files
class TestByteLevelBPE:
def test_basic_encode(self, roberta_files):
tokenizer = ByteLevelBPETokenizer.from_file(roberta_files["vocab"], roberta_files["merges"])
output = tokenizer.encode("The quick brown fox jumps over the lazy dog")
assert output.ids == [133, 2119, 6219, 23602, 13855, 81, 5, 22414, 2335]
assert output.tokens == [
"The",
"Ġquick",
"Ġbrown",
"Ġfox",
"Ġjumps",
"Ġover",
"Ġthe",
"Ġlazy",
"Ġdog",
]
assert output.offsets == [
(0, 3),
(3, 9),
(9, 15),
(15, 19),
(19, 25),
(25, 30),
(30, 34),
(34, 39),
(39, 43),
]
def test_add_prefix_space(self, roberta_files):
tokenizer = ByteLevelBPETokenizer.from_file(
roberta_files["vocab"], roberta_files["merges"], add_prefix_space=True
)
output = tokenizer.encode("The quick brown fox jumps over the lazy dog")
assert output.ids == [20, 2119, 6219, 23602, 13855, 81, 5, 22414, 2335]
assert output.tokens == [
"ĠThe",
"Ġquick",
"Ġbrown",
"Ġfox",
"Ġjumps",
"Ġover",
"Ġthe",
"Ġlazy",
"Ġdog",
]
assert output.offsets == [
(0, 3),
(3, 9),
(9, 15),
(15, 19),
(19, 25),
(25, 30),
(30, 34),
(34, 39),
(39, 43),
]
def test_lowerspace(self, roberta_files):
tokenizer = ByteLevelBPETokenizer.from_file(
roberta_files["vocab"],
roberta_files["merges"],
add_prefix_space=True,
lowercase=True,
)
output = tokenizer.encode("The Quick Brown Fox Jumps Over The Lazy Dog")
assert output.ids == [5, 2119, 6219, 23602, 13855, 81, 5, 22414, 2335]
assert output.tokens == [
"Ġthe",
"Ġquick",
"Ġbrown",
"Ġfox",
"Ġjumps",
"Ġover",
"Ġthe",
"Ġlazy",
"Ġdog",
]
def test_multiprocessing_with_parallelism(self, roberta_files):
tokenizer = ByteLevelBPETokenizer.from_file(roberta_files["vocab"], roberta_files["merges"])
multiprocessing_with_parallelism(tokenizer, False)
multiprocessing_with_parallelism(tokenizer, True)
def test_train_from_iterator(self):
text = ["A first sentence", "Another sentence", "And a last one"]
tokenizer = ByteLevelBPETokenizer()
tokenizer.train_from_iterator(text, show_progress=False)
output = tokenizer.encode("A sentence")
assert output.tokens == ["A", "Ġsentence"]
|
tokenizers/bindings/python/tests/implementations/test_byte_level_bpe.py/0
|
{
"file_path": "tokenizers/bindings/python/tests/implementations/test_byte_level_bpe.py",
"repo_id": "tokenizers",
"token_count": 1653
}
| 257
|
# Pre-tokenizers
<tokenizerslangcontent>
<python>
## BertPreTokenizer
[[autodoc]] tokenizers.pre_tokenizers.BertPreTokenizer
## ByteLevel
[[autodoc]] tokenizers.pre_tokenizers.ByteLevel
## CharDelimiterSplit
[[autodoc]] tokenizers.pre_tokenizers.CharDelimiterSplit
## Digits
[[autodoc]] tokenizers.pre_tokenizers.Digits
## Metaspace
[[autodoc]] tokenizers.pre_tokenizers.Metaspace
## PreTokenizer
[[autodoc]] tokenizers.pre_tokenizers.PreTokenizer
## Punctuation
[[autodoc]] tokenizers.pre_tokenizers.Punctuation
## Sequence
[[autodoc]] tokenizers.pre_tokenizers.Sequence
## Split
[[autodoc]] tokenizers.pre_tokenizers.Split
## UnicodeScripts
[[autodoc]] tokenizers.pre_tokenizers.UnicodeScripts
## Whitespace
[[autodoc]] tokenizers.pre_tokenizers.Whitespace
## WhitespaceSplit
[[autodoc]] tokenizers.pre_tokenizers.WhitespaceSplit
</python>
<rust>
The Rust API Reference is available directly on the [Docs.rs](https://docs.rs/tokenizers/latest/tokenizers/) website.
</rust>
<node>
The node API has not been documented yet.
</node>
</tokenizerslangcontent>
|
tokenizers/docs/source-doc-builder/api/pre-tokenizers.mdx/0
|
{
"file_path": "tokenizers/docs/source-doc-builder/api/pre-tokenizers.mdx",
"repo_id": "tokenizers",
"token_count": 371
}
| 258
|
The tokenization pipeline
====================================================================================================
When calling :entity:`Tokenizer.encode` or :entity:`Tokenizer.encode_batch`, the input text(s) go
through the following pipeline:
- :ref:`normalization`
- :ref:`pre-tokenization`
- :ref:`model`
- :ref:`post-processing`
We'll see in details what happens during each of those steps in detail, as well as when you want to
:ref:`decode <decoding>` some token ids, and how the 🤗 Tokenizers library allows you to customize
each of those steps to your needs. If you're already familiar with those steps and want to learn by
seeing some code, jump to :ref:`our BERT from scratch example <example>`.
For the examples that require a :entity:`Tokenizer`, we will use the tokenizer we trained
in the :doc:`quicktour`, which you can load with:
.. only:: python
.. literalinclude:: ../../bindings/python/tests/documentation/test_pipeline.py
:language: python
:start-after: START reload_tokenizer
:end-before: END reload_tokenizer
:dedent: 8
.. only:: rust
.. literalinclude:: ../../tokenizers/tests/documentation.rs
:language: rust
:start-after: START pipeline_reload_tokenizer
:end-before: END pipeline_reload_tokenizer
:dedent: 4
.. only:: node
.. literalinclude:: ../../bindings/node/examples/documentation/pipeline.test.ts
:language: javascript
:start-after: START reload_tokenizer
:end-before: END reload_tokenizer
:dedent: 8
.. _normalization:
Normalization
----------------------------------------------------------------------------------------------------
Normalization is, in a nutshell, a set of operations you apply to a raw string to make it less
random or "cleaner". Common operations include stripping whitespace, removing accented characters
or lowercasing all text. If you're familiar with `Unicode normalization
<https://unicode.org/reports/tr15>`__, it is also a very common normalization operation applied
in most tokenizers.
Each normalization operation is represented in the 🤗 Tokenizers library by a
:entity:`Normalizer`, and you can combine several of those by using a
:entity:`normalizers.Sequence`. Here is a normalizer applying NFD Unicode normalization
and removing accents as an example:
.. only:: python
.. literalinclude:: ../../bindings/python/tests/documentation/test_pipeline.py
:language: python
:start-after: START setup_normalizer
:end-before: END setup_normalizer
:dedent: 8
.. only:: rust
.. literalinclude:: ../../tokenizers/tests/documentation.rs
:language: rust
:start-after: START pipeline_setup_normalizer
:end-before: END pipeline_setup_normalizer
:dedent: 4
.. only:: node
.. literalinclude:: ../../bindings/node/examples/documentation/pipeline.test.ts
:language: javascript
:start-after: START setup_normalizer
:end-before: END setup_normalizer
:dedent: 8
You can manually test that normalizer by applying it to any string:
.. only:: python
.. literalinclude:: ../../bindings/python/tests/documentation/test_pipeline.py
:language: python
:start-after: START test_normalizer
:end-before: END test_normalizer
:dedent: 8
.. only:: rust
.. literalinclude:: ../../tokenizers/tests/documentation.rs
:language: rust
:start-after: START pipeline_test_normalizer
:end-before: END pipeline_test_normalizer
:dedent: 4
.. only:: node
.. literalinclude:: ../../bindings/node/examples/documentation/pipeline.test.ts
:language: javascript
:start-after: START test_normalizer
:end-before: END test_normalizer
:dedent: 8
When building a :entity:`Tokenizer`, you can customize its normalizer by just changing
the corresponding attribute:
.. only:: python
.. literalinclude:: ../../bindings/python/tests/documentation/test_pipeline.py
:language: python
:start-after: START replace_normalizer
:end-before: END replace_normalizer
:dedent: 8
.. only:: rust
.. literalinclude:: ../../tokenizers/tests/documentation.rs
:language: rust
:start-after: START pipeline_replace_normalizer
:end-before: END pipeline_replace_normalizer
:dedent: 4
.. only:: node
.. literalinclude:: ../../bindings/node/examples/documentation/pipeline.test.ts
:language: javascript
:start-after: START replace_normalizer
:end-before: END replace_normalizer
:dedent: 8
Of course, if you change the way a tokenizer applies normalization, you should probably retrain it
from scratch afterward.
.. _pre-tokenization:
Pre-Tokenization
----------------------------------------------------------------------------------------------------
Pre-tokenization is the act of splitting a text into smaller objects that give an upper bound to
what your tokens will be at the end of training. A good way to think of this is that the
pre-tokenizer will split your text into "words" and then, your final tokens will be parts of those
words.
An easy way to pre-tokenize inputs is to split on spaces and punctuations, which is done by the
:entity:`pre_tokenizers.Whitespace` pre-tokenizer:
.. only:: python
.. literalinclude:: ../../bindings/python/tests/documentation/test_pipeline.py
:language: python
:start-after: START setup_pre_tokenizer
:end-before: END setup_pre_tokenizer
:dedent: 8
.. only:: rust
.. literalinclude:: ../../tokenizers/tests/documentation.rs
:language: rust
:start-after: START pipeline_setup_pre_tokenizer
:end-before: END pipeline_setup_pre_tokenizer
:dedent: 4
.. only:: node
.. literalinclude:: ../../bindings/node/examples/documentation/pipeline.test.ts
:language: javascript
:start-after: START setup_pre_tokenizer
:end-before: END setup_pre_tokenizer
:dedent: 8
The output is a list of tuples, with each tuple containing one word and its span in the original
sentence (which is used to determine the final :obj:`offsets` of our :entity:`Encoding`).
Note that splitting on punctuation will split contractions like :obj:`"I'm"` in this example.
You can combine together any :entity:`PreTokenizer` together. For
instance, here is a pre-tokenizer that will split on space, punctuation and digits, separating
numbers in their individual digits:
.. only:: python
.. literalinclude:: ../../bindings/python/tests/documentation/test_pipeline.py
:language: python
:start-after: START combine_pre_tokenizer
:end-before: END combine_pre_tokenizer
:dedent: 8
.. only:: rust
.. literalinclude:: ../../tokenizers/tests/documentation.rs
:language: rust
:start-after: START pipeline_combine_pre_tokenizer
:end-before: END pipeline_combine_pre_tokenizer
:dedent: 4
.. only:: node
.. literalinclude:: ../../bindings/node/examples/documentation/pipeline.test.ts
:language: javascript
:start-after: START combine_pre_tokenizer
:end-before: END combine_pre_tokenizer
:dedent: 8
As we saw in the :doc:`quicktour`, you can customize the pre-tokenizer of a
:entity:`Tokenizer` by just changing the corresponding attribute:
.. only:: python
.. literalinclude:: ../../bindings/python/tests/documentation/test_pipeline.py
:language: python
:start-after: START replace_pre_tokenizer
:end-before: END replace_pre_tokenizer
:dedent: 8
.. only:: rust
.. literalinclude:: ../../tokenizers/tests/documentation.rs
:language: rust
:start-after: START pipeline_replace_pre_tokenizer
:end-before: END pipeline_replace_pre_tokenizer
:dedent: 4
.. only:: node
.. literalinclude:: ../../bindings/node/examples/documentation/pipeline.test.ts
:language: javascript
:start-after: START replace_pre_tokenizer
:end-before: END replace_pre_tokenizer
:dedent: 8
Of course, if you change the way the pre-tokenizer, you should probably retrain your tokenizer from
scratch afterward.
.. _model:
The Model
----------------------------------------------------------------------------------------------------
Once the input texts are normalized and pre-tokenized, the :entity:`Tokenizer` applies the model on
the pre-tokens. This is the part of the pipeline that needs training on your corpus (or that has
been trained if you are using a pretrained tokenizer).
The role of the model is to split your "words" into tokens, using the rules it has learned. It's
also responsible for mapping those tokens to their corresponding IDs in the vocabulary of the model.
This model is passed along when intializing the :entity:`Tokenizer` so you already know
how to customize this part. Currently, the 🤗 Tokenizers library supports:
- :entity:`models.BPE`
- :entity:`models.Unigram`
- :entity:`models.WordLevel`
- :entity:`models.WordPiece`
For more details about each model and its behavior, you can check `here <components#models>`__
.. _post-processing:
Post-Processing
----------------------------------------------------------------------------------------------------
Post-processing is the last step of the tokenization pipeline, to perform any additional
transformation to the :entity:`Encoding` before it's returned, like adding potential
special tokens.
As we saw in the quick tour, we can customize the post processor of a :entity:`Tokenizer`
by setting the corresponding attribute. For instance, here is how we can post-process to make the
inputs suitable for the BERT model:
.. only:: python
.. literalinclude:: ../../bindings/python/tests/documentation/test_pipeline.py
:language: python
:start-after: START setup_processor
:end-before: END setup_processor
:dedent: 8
.. only:: rust
.. literalinclude:: ../../tokenizers/tests/documentation.rs
:language: rust
:start-after: START pipeline_setup_processor
:end-before: END pipeline_setup_processor
:dedent: 4
.. only:: node
.. literalinclude:: ../../bindings/node/examples/documentation/pipeline.test.ts
:language: javascript
:start-after: START setup_processor
:end-before: END setup_processor
:dedent: 8
Note that contrarily to the pre-tokenizer or the normalizer, you don't need to retrain a tokenizer
after changing its post-processor.
.. _example:
All together: a BERT tokenizer from scratch
----------------------------------------------------------------------------------------------------
Let's put all those pieces together to build a BERT tokenizer. First, BERT relies on WordPiece, so
we instantiate a new :entity:`Tokenizer` with this model:
.. only:: python
.. literalinclude:: ../../bindings/python/tests/documentation/test_pipeline.py
:language: python
:start-after: START bert_setup_tokenizer
:end-before: END bert_setup_tokenizer
:dedent: 8
.. only:: rust
.. literalinclude:: ../../tokenizers/tests/documentation.rs
:language: rust
:start-after: START bert_setup_tokenizer
:end-before: END bert_setup_tokenizer
:dedent: 4
.. only:: node
.. literalinclude:: ../../bindings/node/examples/documentation/pipeline.test.ts
:language: javascript
:start-after: START bert_setup_tokenizer
:end-before: END bert_setup_tokenizer
:dedent: 8
Then we know that BERT preprocesses texts by removing accents and lowercasing. We also use a unicode
normalizer:
.. only:: python
.. literalinclude:: ../../bindings/python/tests/documentation/test_pipeline.py
:language: python
:start-after: START bert_setup_normalizer
:end-before: END bert_setup_normalizer
:dedent: 8
.. only:: rust
.. literalinclude:: ../../tokenizers/tests/documentation.rs
:language: rust
:start-after: START bert_setup_normalizer
:end-before: END bert_setup_normalizer
:dedent: 4
.. only:: node
.. literalinclude:: ../../bindings/node/examples/documentation/pipeline.test.ts
:language: javascript
:start-after: START bert_setup_normalizer
:end-before: END bert_setup_normalizer
:dedent: 8
The pre-tokenizer is just splitting on whitespace and punctuation:
.. only:: python
.. literalinclude:: ../../bindings/python/tests/documentation/test_pipeline.py
:language: python
:start-after: START bert_setup_pre_tokenizer
:end-before: END bert_setup_pre_tokenizer
:dedent: 8
.. only:: rust
.. literalinclude:: ../../tokenizers/tests/documentation.rs
:language: rust
:start-after: START bert_setup_pre_tokenizer
:end-before: END bert_setup_pre_tokenizer
:dedent: 4
.. only:: node
.. literalinclude:: ../../bindings/node/examples/documentation/pipeline.test.ts
:language: javascript
:start-after: START bert_setup_pre_tokenizer
:end-before: END bert_setup_pre_tokenizer
:dedent: 8
And the post-processing uses the template we saw in the previous section:
.. only:: python
.. literalinclude:: ../../bindings/python/tests/documentation/test_pipeline.py
:language: python
:start-after: START bert_setup_processor
:end-before: END bert_setup_processor
:dedent: 8
.. only:: rust
.. literalinclude:: ../../tokenizers/tests/documentation.rs
:language: rust
:start-after: START bert_setup_processor
:end-before: END bert_setup_processor
:dedent: 4
.. only:: node
.. literalinclude:: ../../bindings/node/examples/documentation/pipeline.test.ts
:language: javascript
:start-after: START bert_setup_processor
:end-before: END bert_setup_processor
:dedent: 8
We can use this tokenizer and train on it on wikitext like in the :doc:`quicktour`:
.. only:: python
.. literalinclude:: ../../bindings/python/tests/documentation/test_pipeline.py
:language: python
:start-after: START bert_train_tokenizer
:end-before: END bert_train_tokenizer
:dedent: 8
.. only:: rust
.. literalinclude:: ../../tokenizers/tests/documentation.rs
:language: rust
:start-after: START bert_train_tokenizer
:end-before: END bert_train_tokenizer
:dedent: 4
.. only:: node
.. literalinclude:: ../../bindings/node/examples/documentation/pipeline.test.ts
:language: javascript
:start-after: START bert_train_tokenizer
:end-before: END bert_train_tokenizer
:dedent: 8
.. _decoding:
Decoding
----------------------------------------------------------------------------------------------------
.. entities:: python
bert_tokenizer
:obj:`bert_tokenizer`
.. entities:: rust
bert_tokenizer
:obj:`bert_tokenizer`
.. entities:: node
bert_tokenizer
:obj:`bertTokenizer`
On top of encoding the input texts, a :entity:`Tokenizer` also has an API for decoding,
that is converting IDs generated by your model back to a text. This is done by the methods
:entity:`Tokenizer.decode` (for one predicted text) and :entity:`Tokenizer.decode_batch` (for a
batch of predictions).
The `decoder` will first convert the IDs back to tokens (using the tokenizer's vocabulary) and
remove all special tokens, then join those tokens with spaces:
.. only:: python
.. literalinclude:: ../../bindings/python/tests/documentation/test_pipeline.py
:language: python
:start-after: START test_decoding
:end-before: END test_decoding
:dedent: 8
.. only:: rust
.. literalinclude:: ../../tokenizers/tests/documentation.rs
:language: rust
:start-after: START pipeline_test_decoding
:end-before: END pipeline_test_decoding
:dedent: 4
.. only:: node
.. literalinclude:: ../../bindings/node/examples/documentation/pipeline.test.ts
:language: javascript
:start-after: START test_decoding
:end-before: END test_decoding
:dedent: 8
If you used a model that added special characters to represent subtokens of a given "word" (like
the :obj:`"##"` in WordPiece) you will need to customize the `decoder` to treat them properly. If we
take our previous :entity:`bert_tokenizer` for instance the default decoding will give:
.. only:: python
.. literalinclude:: ../../bindings/python/tests/documentation/test_pipeline.py
:language: python
:start-after: START bert_test_decoding
:end-before: END bert_test_decoding
:dedent: 8
.. only:: rust
.. literalinclude:: ../../tokenizers/tests/documentation.rs
:language: rust
:start-after: START bert_test_decoding
:end-before: END bert_test_decoding
:dedent: 4
.. only:: node
.. literalinclude:: ../../bindings/node/examples/documentation/pipeline.test.ts
:language: javascript
:start-after: START bert_test_decoding
:end-before: END bert_test_decoding
:dedent: 8
But by changing it to a proper decoder, we get:
.. only:: python
.. literalinclude:: ../../bindings/python/tests/documentation/test_pipeline.py
:language: python
:start-after: START bert_proper_decoding
:end-before: END bert_proper_decoding
:dedent: 8
.. only:: rust
.. literalinclude:: ../../tokenizers/tests/documentation.rs
:language: rust
:start-after: START bert_proper_decoding
:end-before: END bert_proper_decoding
:dedent: 4
.. only:: node
.. literalinclude:: ../../bindings/node/examples/documentation/pipeline.test.ts
:language: javascript
:start-after: START bert_proper_decoding
:end-before: END bert_proper_decoding
:dedent: 8
|
tokenizers/docs/source/pipeline.rst/0
|
{
"file_path": "tokenizers/docs/source/pipeline.rst",
"repo_id": "tokenizers",
"token_count": 6323
}
| 259
|
use tokenizers::models::wordpiece::WordPiece;
use tokenizers::{AddedToken, Tokenizer};
fn main() {
let start = std::time::Instant::now();
let mut tokenizer = Tokenizer::new(WordPiece::default());
// Mix special and not special
// You can make sure ids are in order, and special status is correct.
let tokens: Vec<_> = (0..120_000)
.map(|i| AddedToken::from(format!("[SPECIAL_{}]", i), i % 2 == 0))
.collect();
tokenizer.add_tokens(&tokens);
tokenizer.save("_tok.json", true).unwrap();
println!("Save took {:?}", start.elapsed());
let start = std::time::Instant::now();
let _tok = Tokenizer::from_file("_tok.json").unwrap();
println!("Took {:?}", start.elapsed());
std::fs::remove_file("_tok.json").unwrap();
}
|
tokenizers/tokenizers/examples/serialization.rs/0
|
{
"file_path": "tokenizers/tokenizers/examples/serialization.rs",
"repo_id": "tokenizers",
"token_count": 300
}
| 260
|
#![allow(clippy::map_entry)]
use super::{Pair, WithFirstLastIterator, Word, BPE};
use crate::parallelism::*;
use crate::tokenizer::{AddedToken, Result, Trainer};
use crate::utils::progress::{ProgressBar, ProgressStyle};
use serde::{Deserialize, Serialize};
use std::cmp::Ordering;
use std::collections::{BinaryHeap, HashMap, HashSet};
#[derive(Debug, Eq)]
struct Merge {
pair: Pair,
count: u64,
pos: HashSet<usize>,
}
impl PartialEq for Merge {
fn eq(&self, other: &Self) -> bool {
self.count == other.count && self.pair == other.pair
}
}
impl PartialOrd for Merge {
fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
Some(self.cmp(other))
}
}
impl Ord for Merge {
fn cmp(&self, other: &Self) -> Ordering {
if self.count != other.count {
self.count.cmp(&other.count)
} else {
// Here we want ascending order
other.pair.cmp(&self.pair)
}
}
}
struct Config {
min_frequency: u64,
vocab_size: usize,
show_progress: bool,
special_tokens: Vec<AddedToken>,
limit_alphabet: Option<usize>,
initial_alphabet: HashSet<char>,
continuing_subword_prefix: Option<String>,
end_of_word_suffix: Option<String>,
max_token_length: Option<usize>,
}
/// A `BpeTrainerBuilder` can be used to create a `BpeTrainer` with a custom
/// configuration.
pub struct BpeTrainerBuilder {
config: Config,
}
impl Default for BpeTrainerBuilder {
fn default() -> Self {
Self {
config: Config {
min_frequency: 0,
vocab_size: 30000,
show_progress: true,
special_tokens: vec![],
limit_alphabet: None,
initial_alphabet: HashSet::new(),
continuing_subword_prefix: None,
end_of_word_suffix: None,
max_token_length: None,
},
}
}
}
impl BpeTrainerBuilder {
/// Constructs a new `BpeTrainerBuilder`
pub fn new() -> Self {
Self::default()
}
/// Set the expected minimum frequency
#[must_use]
pub fn min_frequency(mut self, frequency: u64) -> Self {
self.config.min_frequency = frequency;
self
}
/// Set the vocabulary size
#[must_use]
pub fn vocab_size(mut self, size: usize) -> Self {
self.config.vocab_size = size;
self
}
/// Set whether to show progress
#[must_use]
pub fn show_progress(mut self, show: bool) -> Self {
self.config.show_progress = show;
self
}
/// Set the special tokens
#[must_use]
pub fn special_tokens(mut self, tokens: Vec<AddedToken>) -> Self {
self.config.special_tokens = tokens;
self
}
/// Set whether to limit the alphabet
#[must_use]
pub fn limit_alphabet(mut self, limit: usize) -> Self {
self.config.limit_alphabet = Some(limit);
self
}
/// Set the initial alphabet
#[must_use]
pub fn initial_alphabet(mut self, alphabet: HashSet<char>) -> Self {
self.config.initial_alphabet = alphabet;
self
}
/// Set the continuing_subword_prefix
#[must_use]
pub fn continuing_subword_prefix(mut self, prefix: String) -> Self {
self.config.continuing_subword_prefix = Some(prefix);
self
}
/// Set the end_of_word_suffix
#[must_use]
pub fn end_of_word_suffix(mut self, suffix: String) -> Self {
self.config.end_of_word_suffix = Some(suffix);
self
}
/// Set max_token_length
#[must_use]
pub fn max_token_length(mut self, max_token_length: Option<usize>) -> Self {
self.config.max_token_length = max_token_length;
self
}
/// Constructs the final BpeTrainer
pub fn build(self) -> BpeTrainer {
BpeTrainer {
min_frequency: self.config.min_frequency,
vocab_size: self.config.vocab_size,
show_progress: self.config.show_progress,
special_tokens: self.config.special_tokens,
limit_alphabet: self.config.limit_alphabet,
initial_alphabet: self.config.initial_alphabet,
continuing_subword_prefix: self.config.continuing_subword_prefix,
end_of_word_suffix: self.config.end_of_word_suffix,
max_token_length: self.config.max_token_length,
words: HashMap::new(),
}
}
}
/// In charge of training a `BPE` model
///
/// # Examples
///
/// ```
/// use tokenizers::tokenizer::Trainer;
/// use tokenizers::models::bpe::{BPE, BpeTrainer};
///
/// let sequences = vec![ "Hello", "World" ];
///
/// let mut trainer = BpeTrainer::default();
/// trainer.feed(sequences.iter(), |s| Ok(vec![s.to_owned()]));
///
/// let mut model = BPE::default();
/// let special_tokens = trainer.train(&mut model).unwrap();
/// ```
#[non_exhaustive]
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, Eq)]
pub struct BpeTrainer {
/// The minimum frequency a pair must have to produce a merge operation
pub min_frequency: u64,
/// The target vocabulary size
pub vocab_size: usize,
/// Whether to show progress while training
pub show_progress: bool,
/// A list of special tokens that the model should know of
pub special_tokens: Vec<AddedToken>,
/// Whether to limit the number of initial tokens that can be kept before computing merges
pub limit_alphabet: Option<usize>,
/// The initial alphabet we want absolutely to include. This allows to cover
/// some characters that are not necessarily in the training set
pub initial_alphabet: HashSet<char>,
/// An optional prefix to use on any subword that exist only behind another one
pub continuing_subword_prefix: Option<String>,
/// An optional suffix to caracterize and end-of-word subword
pub end_of_word_suffix: Option<String>,
/// An optional parameter to limit the max length of any single token
pub max_token_length: Option<usize>,
words: HashMap<String, u64>,
}
impl Default for BpeTrainer {
fn default() -> Self {
Self::builder().build()
}
}
impl BpeTrainer {
pub fn new(min_frequency: u64, vocab_size: usize) -> Self {
Self {
min_frequency,
vocab_size,
..Default::default()
}
}
pub fn builder() -> BpeTrainerBuilder {
BpeTrainerBuilder::new()
}
/// Setup a progress bar if asked to show progress
fn setup_progress(&self) -> Option<ProgressBar> {
if self.show_progress {
let p = ProgressBar::new(0);
p.set_style(
ProgressStyle::default_bar()
.template("[{elapsed_precise}] {msg:<30!} {wide_bar} {pos:<9!}/{len:>9!}")
.expect("Invalid progress template"),
);
Some(p)
} else {
None
}
}
/// Set the progress bar in the finish state
fn finalize_progress(&self, p: &Option<ProgressBar>, final_len: usize) {
if let Some(p) = p {
p.set_length(final_len as u64);
p.finish();
println!();
}
}
/// Update the progress bar with the new provided length and message
fn update_progress(&self, p: &Option<ProgressBar>, len: usize, message: &'static str) {
if let Some(p) = p {
p.set_message(message);
p.set_length(len as u64);
p.reset();
}
}
/// Add the provided special tokens to the initial vocabulary
fn add_special_tokens(&self, w2id: &mut HashMap<String, u32>, id2w: &mut Vec<String>) {
for token in &self.special_tokens {
if !w2id.contains_key(&token.content) {
id2w.push(token.content.to_owned());
w2id.insert(token.content.to_owned(), (id2w.len() - 1) as u32);
}
}
}
/// Compute the initial alphabet and limit it if relevant
fn compute_alphabet(
&self,
wc: &HashMap<String, u64>,
w2id: &mut HashMap<String, u32>,
id2w: &mut Vec<String>,
) {
// Compute the alphabet from seen words
let mut alphabet: HashMap<char, usize> = HashMap::new();
for (word, count) in wc {
for c in word.chars() {
alphabet
.entry(c)
.and_modify(|cnt| *cnt += *count as usize)
.or_insert(*count as usize);
}
}
// Also include anything from the provided initial alphabet
for c in &self.initial_alphabet {
alphabet
.entry(*c)
.and_modify(|cnt| *cnt = usize::MAX)
.or_insert(usize::MAX);
}
let mut kept = alphabet.iter().collect::<Vec<_>>();
// Compute the number of chars to remove from the alphabet
// If `limit_alphabet < initial_alphabet.len()`, some of these initial characters
// will be removed
let to_remove = self
.limit_alphabet
.map(|limit| {
if alphabet.len() > limit {
alphabet.len() - limit
} else {
0
}
})
.unwrap_or(0);
// Remove the unwanted chars
if to_remove > 0 {
kept.sort_unstable_by_key(|k| *k.1);
kept.drain(..to_remove);
}
// Keep the initial alphabet (sorted for determinism)
kept.sort_unstable_by_key(|k| (*k.0) as u32);
kept.into_iter().for_each(|(c, _)| {
let s = c.to_string();
if !w2id.contains_key(&s) {
id2w.push(s.clone());
w2id.insert(s, (id2w.len() - 1) as u32);
}
});
}
/// Tokenize words and add subwords to the vocabulary when relevant
fn tokenize_words(
&self,
wc: &HashMap<String, u64>,
w2id: &mut HashMap<String, u32>,
id2w: &mut Vec<String>,
p: &Option<ProgressBar>,
) -> (Vec<Word>, Vec<u64>) {
let mut words: Vec<Word> = Vec::with_capacity(wc.len());
let mut counts: Vec<u64> = Vec::with_capacity(wc.len());
for (word, count) in wc {
let mut current_word = Word::new();
counts.push(*count);
for (is_first, is_last, c) in word.chars().with_first_and_last() {
let mut s = c.to_string();
if w2id.contains_key(&s) {
// Found the initial char in the authorized alphabet
// Add the `continuing_subword_prefix` if relevant
if !is_first {
if let Some(prefix) = &self.continuing_subword_prefix {
s = format!("{}{}", prefix, s);
}
}
// Add the `end_of_word_suffix` if relevant
if is_last {
if let Some(suffix) = &self.end_of_word_suffix {
s = format!("{}{}", s, suffix);
}
}
// Insert the new formed string if necessary
if !w2id.contains_key(&s) {
id2w.push(s.clone());
w2id.insert(s.clone(), (id2w.len() - 1) as u32);
}
current_word.add(w2id[&s], 1); // We do not care about the len here
}
}
words.push(current_word);
if let Some(p) = p {
p.inc(1);
}
}
(words, counts)
}
fn count_pairs(
&self,
words: &[Word],
counts: &[u64],
p: &Option<ProgressBar>,
) -> (HashMap<Pair, i32>, HashMap<Pair, HashSet<usize>>) {
words
.maybe_par_iter()
.enumerate()
.map(|(i, word)| {
let mut pair_counts = HashMap::new();
let mut where_to_update: HashMap<Pair, HashSet<usize>> = HashMap::new();
for window in word.get_chars().windows(2) {
let cur_pair: Pair = (window[0], window[1]);
// Initialize pair_counts and where_to_update for this pair if we just saw it
if !pair_counts.contains_key(&cur_pair) {
pair_counts.insert(cur_pair, 0);
}
// Then update counts
let count = counts[i];
where_to_update
.entry(cur_pair)
.and_modify(|h| {
h.insert(i);
})
.or_insert_with(|| {
let mut h = HashSet::new();
h.insert(i);
h
});
*pair_counts.get_mut(&cur_pair).unwrap() += count as i32;
}
if let Some(p) = &p {
p.inc(1);
}
(pair_counts, where_to_update)
})
.reduce(
|| (HashMap::new(), HashMap::new()),
|(mut pair_counts, mut where_to_update), (pc, wtu)| {
for (k, v) in pc {
pair_counts.entry(k).and_modify(|c| *c += v).or_insert(v);
}
for (k, v) in wtu {
where_to_update
.entry(k)
.and_modify(|set| *set = set.union(&v).copied().collect())
.or_insert(v);
}
(pair_counts, where_to_update)
},
)
}
pub fn do_train(
&self,
word_counts: &HashMap<String, u64>,
model: &mut BPE,
) -> Result<Vec<AddedToken>> {
let mut word_to_id: HashMap<String, u32> = HashMap::with_capacity(self.vocab_size);
let mut id_to_word: Vec<String> = Vec::with_capacity(self.vocab_size);
let max_token_length: usize = self.max_token_length.unwrap_or(usize::MAX);
let progress = self.setup_progress();
//
// 1. Add all special tokens to the vocabulary
//
self.add_special_tokens(&mut word_to_id, &mut id_to_word);
//
// 2. Compute the initial alphabet
//
self.compute_alphabet(word_counts, &mut word_to_id, &mut id_to_word);
//
// 3. Tokenize words
//
self.update_progress(&progress, word_counts.len(), "Tokenize words");
let (words, counts) =
self.tokenize_words(word_counts, &mut word_to_id, &mut id_to_word, &progress);
self.finalize_progress(&progress, words.len());
//
// 4. Count pairs in words
//
self.update_progress(&progress, words.len(), "Count pairs");
let (mut pair_counts, mut where_to_update) = self.count_pairs(&words, &counts, &progress);
// Insert them in the queue
let mut queue = BinaryHeap::with_capacity(pair_counts.len());
where_to_update.drain().for_each(|(pair, pos)| {
let count = pair_counts[&pair];
if count > 0 {
queue.push(Merge {
pair,
count: count as u64,
pos,
});
}
});
self.finalize_progress(&progress, words.len());
//
// 5. Do merges
//
self.update_progress(&progress, self.vocab_size, "Compute merges");
let mut merges: Vec<(Pair, u32)> = vec![];
loop {
// Stop as soon as we have a big enough vocabulary
if word_to_id.len() >= self.vocab_size {
break;
}
if queue.is_empty() {
break;
}
let mut top = queue.pop().unwrap();
if top.count != pair_counts[&top.pair] as u64 {
top.count = pair_counts[&top.pair] as u64;
queue.push(top);
continue;
}
if top.count < 1 || self.min_frequency > top.count {
break;
}
let part_a = &id_to_word[top.pair.0 as usize];
let mut part_b = id_to_word[top.pair.1 as usize].to_owned();
// Build new token
if let Some(prefix) = &self.continuing_subword_prefix {
if part_b.starts_with(prefix) {
let prefix_byte_len = prefix.chars().map(|c| c.len_utf8()).sum();
part_b = part_b[prefix_byte_len..].to_string();
}
}
let new_token = format!("{}{}", part_a, part_b);
// implement sentencepiece-like merge.
// if this code were to be merged, integrate a way in the python bindings to communicate this variable
// default should be 0/None to maintain previous behavior. 16 is the spm default.
// Insert new token if it does not already exist
let new_token_id = word_to_id
.get(&new_token)
.copied()
.unwrap_or(id_to_word.len() as u32);
if !word_to_id.contains_key(&new_token) {
id_to_word.push(new_token.clone());
word_to_id.insert(new_token.clone(), new_token_id);
}
merges.push((top.pair, new_token_id));
// Merge the new pair in every words
let changes = top
.pos
.maybe_par_iter()
.flat_map(|&i| {
let word = &words[i] as *const _ as *mut Word;
// We can merge each of these words in parallel here because each position
// can be there only once (HashSet). So this is safe.
unsafe {
// let word: &mut Word = &mut (*word);
(*word)
.merge(top.pair.0, top.pair.1, new_token_id, max_token_length)
.into_iter()
.map(|c| (c, i))
.collect::<Vec<_>>()
}
})
.collect::<Vec<_>>();
// Introduce new formed pairs
for ((pair, change), iw) in changes {
let count = change * counts[iw] as i32;
pair_counts
.entry(pair)
.and_modify(|c| *c += count)
.or_insert(count);
if change > 0 {
where_to_update
.entry(pair)
.and_modify(|h| {
h.insert(iw);
})
.or_insert_with(|| {
let mut h = HashSet::new();
h.insert(iw);
h
});
}
}
where_to_update.drain().for_each(|(pair, pos)| {
let count = pair_counts[&pair];
if count > 0 {
queue.push(Merge {
pair,
count: count as u64,
pos,
});
}
});
if let Some(p) = &progress {
p.inc(1);
}
}
self.finalize_progress(&progress, merges.len());
// Transfer new vocab & options to model
model.vocab = word_to_id;
model.vocab_r = model
.vocab
.iter()
.map(|(key, val)| (*val, key.to_owned()))
.collect();
model.merges = merges
.into_iter()
.enumerate()
.map(|(i, (pair, new_token_id))| (pair, (i as u32, new_token_id)))
.collect();
if let Some(prefix) = &self.continuing_subword_prefix {
model.continuing_subword_prefix = Some(prefix.to_owned());
} else {
model.continuing_subword_prefix = None;
}
if let Some(suffix) = &self.end_of_word_suffix {
model.end_of_word_suffix = Some(suffix.to_owned());
} else {
model.end_of_word_suffix = None;
}
Ok(self.special_tokens.clone())
}
}
impl Trainer for BpeTrainer {
type Model = BPE;
/// Train a BPE model
fn train(&self, model: &mut BPE) -> Result<Vec<AddedToken>> {
self.do_train(&self.words, model)
}
/// Whether we should show progress
fn should_show_progress(&self) -> bool {
self.show_progress
}
fn feed<I, S, F>(&mut self, iterator: I, process: F) -> Result<()>
where
I: Iterator<Item = S> + Send,
S: AsRef<str> + Send,
F: Fn(&str) -> Result<Vec<String>> + Sync,
{
let words: Result<HashMap<String, u64>> = iterator
.maybe_par_bridge()
.map(|sequence| {
let words = process(sequence.as_ref())?;
let mut map = HashMap::new();
for word in words {
map.entry(word).and_modify(|c| *c += 1).or_insert(1);
}
Ok(map)
})
.reduce(
|| Ok(HashMap::new()),
|acc, ws| {
let mut acc = acc?;
for (k, v) in ws? {
acc.entry(k).and_modify(|c| *c += v).or_insert(v);
}
Ok(acc)
},
);
self.words = words?;
Ok(())
}
}
#[cfg(test)]
mod tests {
use super::{BpeTrainer, Pair, BPE};
use std::collections::HashMap;
#[test]
fn test_train() {
let word_counts: HashMap<String, u64> = [
("roses".into(), 1),
("are".into(), 2),
("red".into(), 1),
("voilets".into(), 1),
("blue".into(), 1),
("BERT".into(), 1),
("is".into(), 2),
("big".into(), 1),
("and".into(), 1),
("so".into(), 1),
("GPT-2".into(), 1),
]
.iter()
.cloned()
.collect();
let trainer = BpeTrainer::builder()
.show_progress(false)
.min_frequency(2)
.build();
let mut model = BPE::default();
trainer.do_train(&word_counts, &mut model).unwrap();
// Vocab should contain all of the characters from the `word_counts` mapping
// as well as three merges: 're', 'are', and 'is'.
let expected_vocab: HashMap<String, u32> = [
("-".into(), 0),
("2".into(), 1),
("B".into(), 2),
("E".into(), 3),
("G".into(), 4),
("P".into(), 5),
("R".into(), 6),
("T".into(), 7),
("a".into(), 8),
("b".into(), 9),
("d".into(), 10),
("e".into(), 11),
("g".into(), 12),
("i".into(), 13),
("l".into(), 14),
("n".into(), 15),
("o".into(), 16),
("r".into(), 17),
("s".into(), 18),
("t".into(), 19),
("u".into(), 20),
("v".into(), 21),
("re".into(), 22),
("are".into(), 23),
("is".into(), 24),
]
.iter()
.cloned()
.collect();
assert_eq!(model.vocab, expected_vocab);
// The keys in `merges` are pairs of symbols, the values are tuples of (rank, id),
// where 'rank' determines the order in which this merge will be applied during
// tokenization, and 'id' is the vocab id of the symbol resulting from merging
// the pair of symbols in the corresponding key.
let expected_merges: HashMap<Pair, (u32, u32)> = [
((17, 11), (0, 22)), // 'r' + 'e' -> 're'
((8, 22), (1, 23)), // 'a' + 're' -> 'are'
((13, 18), (2, 24)), // 'i' + 's' -> 'is'
]
.iter()
.cloned()
.collect();
assert_eq!(model.merges, expected_merges);
}
#[test]
fn bpe_test_max_token_length_16() {
/* bpe_test_max_token_length series of tests test the max_token_length flag of bpetrainer
// this is the more robust version that only tests max length of learned tokens
// (pre) tokenizer settings or vocab can be easily modified when necessary
*/
let max_token_length = 16;
let long_word_counts: HashMap<String, u64> = [
("singlelongtokenwithoutcasechange", 2),
("singleLongTokenWithCamelCaseChange", 2),
("Longsingletokenwithpunctu@t!onwithin", 2),
("Anotherlongsingletokenwithnumberw1th1n", 2),
("짧은한글문자열짧은한", 2), // korean 10 char
("긴한글문자열긴한글문자열긴한글문", 2), // korean 16 char
("短字符串短字符串短字", 2), //simplified chinese 10 char
("长字符串长字符串长字符串长字符串", 2), // simp. chinese 16 char
("短い文字列短い文字列", 2), // japanese 10 char
("長い文字列長い文字列長い文字列長", 2), // japanese 16 char
("so", 2),
("GPT-2", 2),
]
.iter()
.map(|(key, value)| (key.to_string(), *value))
.collect();
let trainer = BpeTrainer::builder()
.max_token_length(Some(max_token_length))
.show_progress(false)
.min_frequency(0)
.build();
let mut model = BPE::default();
trainer.do_train(&long_word_counts, &mut model).unwrap();
let vocab = model.get_vocab();
for token in vocab.keys() {
assert!(
token.chars().count() <= max_token_length,
"token too long : {} , chars().count() = {}",
token,
token.chars().count()
)
}
}
#[test]
fn bpe_test_max_token_length_direct_assert() {
/* more direct version of bpe_test_max_token_length test
// directly compares tokens with known expected values.
// maybe unstable depending on specific settings or changes.
*/
let long_word_counts: HashMap<String, u64> = [
("sin", 2),
("Sin", 2),
("Lon", 2),
("Ano", 2),
("짧은한", 2),
("긴한글", 2),
("短字符", 2),
("长字符", 2),
("短い文", 2),
("長い文", 2),
("so", 2),
("GP", 2),
]
.iter()
.map(|(key, value)| (key.to_string(), *value))
.collect();
let trainer = BpeTrainer::builder()
.max_token_length(Some(2))
.show_progress(false)
.min_frequency(0)
.build();
let mut model = BPE::default();
trainer.do_train(&long_word_counts, &mut model).unwrap();
let trained_vocab: HashMap<String, u32> = model.get_vocab();
let expected_vocab: HashMap<String, u32> = [
("短", 12),
("n", 6),
("i", 5),
("s", 8),
("字符", 23),
("長", 14),
("긴", 17),
("い文", 22),
("L", 2),
("in", 21),
("o", 7),
("은한", 29),
("S", 4),
("P", 3),
("so", 27),
("符", 13),
("文", 11),
("字", 10),
("짧", 19),
("GP", 25),
("글", 16),
("G", 1),
("An", 24),
("长", 15),
("A", 0),
("Lo", 26),
("긴한", 28),
("い", 9),
("한", 20),
("은", 18),
]
.iter()
.cloned()
.map(|(k, v)| (k.to_string(), v))
.collect();
assert_eq!(trained_vocab, expected_vocab)
}
}
|
tokenizers/tokenizers/src/models/bpe/trainer.rs/0
|
{
"file_path": "tokenizers/tokenizers/src/models/bpe/trainer.rs",
"repo_id": "tokenizers",
"token_count": 15113
}
| 261
|
use crate::processors::byte_level::bytes_char;
use crate::tokenizer::{NormalizedString, Normalizer, Result};
use crate::utils::macro_rules_attribute;
use std::collections::{HashMap, HashSet};
#[derive(Clone, Debug)]
#[macro_rules_attribute(impl_serde_type!)]
pub struct ByteLevel;
lazy_static! {
static ref BYTES_CHAR: HashMap<u8, char> = bytes_char();
static ref CHAR_BYTES: HashMap<char, u8> =
bytes_char().into_iter().map(|(c, b)| (b, c)).collect();
}
impl Default for ByteLevel {
fn default() -> Self {
Self::new()
}
}
impl ByteLevel {
pub fn new() -> Self {
Self {}
}
pub fn alphabet() -> HashSet<char> {
BYTES_CHAR.values().copied().collect()
}
}
impl Normalizer for ByteLevel {
/// Strip the normalized string inplace
fn normalize(&self, normalized: &mut NormalizedString) -> Result<()> {
if !normalized.is_empty() {
let s = normalized.get();
let mut transformations: Vec<(char, isize)> = Vec::with_capacity(s.len());
let mut i = 0;
for cur_char in s.chars() {
let size = cur_char.len_utf8();
let bytes = s[i..i + size].as_bytes();
i += size;
transformations.extend(
bytes
.iter()
.enumerate()
.map(|(i, b)| (BYTES_CHAR[b], isize::from(i > 0))),
);
}
normalized.transform(transformations, 0);
}
Ok(())
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_byte_level_normalize() {
let original = "Hello 我今天能为你做什么";
let normalized = "HelloĠæĪijä»Ĭ天èĥ½ä¸ºä½łåģļä»Ģä¹Ī";
assert_ne!(original, normalized);
let mut n = NormalizedString::from(original);
let byte_level = ByteLevel::new();
byte_level.normalize(&mut n).unwrap();
assert_eq!(&n.get(), &normalized);
assert_eq!(
n,
NormalizedString::new(
original.to_string(),
normalized.to_string(),
vec![
(0, 1),
(1, 2),
(2, 3),
(3, 4),
(4, 5),
(5, 6),
(5, 6),
(6, 9),
(6, 9),
(6, 9),
(6, 9),
(6, 9),
(6, 9),
(9, 12),
(9, 12),
(9, 12),
(9, 12),
(9, 12),
(9, 12),
(12, 15),
(12, 15),
(12, 15),
(12, 15),
(12, 15),
(12, 15),
(15, 18),
(15, 18),
(15, 18),
(15, 18),
(15, 18),
(15, 18),
(18, 21),
(18, 21),
(18, 21),
(18, 21),
(18, 21),
(18, 21),
(21, 24),
(21, 24),
(21, 24),
(21, 24),
(21, 24),
(21, 24),
(24, 27),
(24, 27),
(24, 27),
(24, 27),
(24, 27),
(24, 27),
(27, 30),
(27, 30),
(27, 30),
(27, 30),
(27, 30),
(27, 30),
(30, 33),
(30, 33),
(30, 33),
(30, 33),
(30, 33),
(30, 33)
],
0
)
);
assert_eq!(
n.alignments_original(),
vec![
(0, 1),
(1, 2),
(2, 3),
(3, 4),
(4, 5),
(5, 7),
(7, 13),
(7, 13),
(7, 13),
(13, 19),
(13, 19),
(13, 19),
(19, 25),
(19, 25),
(19, 25),
(25, 31),
(25, 31),
(25, 31),
(31, 37),
(31, 37),
(31, 37),
(37, 43),
(37, 43),
(37, 43),
(43, 49),
(43, 49),
(43, 49),
(49, 55),
(49, 55),
(49, 55),
(55, 61),
(55, 61),
(55, 61)
]
);
}
}
|
tokenizers/tokenizers/src/normalizers/byte_level.rs/0
|
{
"file_path": "tokenizers/tokenizers/src/normalizers/byte_level.rs",
"repo_id": "tokenizers",
"token_count": 3445
}
| 262
|
use crate::utils::SysRegex;
use serde::{Deserialize, Deserializer, Serialize};
use crate::tokenizer::{
pattern::Invert, PreTokenizedString, PreTokenizer, Result, SplitDelimiterBehavior,
};
/// Represents the different patterns that `Split` can use
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, Eq)]
pub enum SplitPattern {
String(String),
Regex(String),
}
impl From<String> for SplitPattern {
fn from(v: String) -> Self {
Self::String(v)
}
}
impl From<&str> for SplitPattern {
fn from(v: &str) -> Self {
Self::String(v.to_owned())
}
}
#[derive(Debug, Serialize)]
#[serde(tag = "type")]
pub struct Split {
pattern: SplitPattern,
#[serde(skip)]
regex: SysRegex,
behavior: SplitDelimiterBehavior,
invert: bool,
}
impl<'de> Deserialize<'de> for Split {
fn deserialize<D>(deserializer: D) -> std::result::Result<Self, D::Error>
where
D: Deserializer<'de>,
{
#[derive(Deserialize)]
enum Type {
Split,
}
#[derive(Deserialize)]
pub struct SplitHelper {
#[serde(rename = "type")]
_type: Type,
pattern: SplitPattern,
behavior: SplitDelimiterBehavior,
invert: bool,
}
let helper = SplitHelper::deserialize(deserializer)?;
Self::new(helper.pattern, helper.behavior, helper.invert).map_err(serde::de::Error::custom)
}
}
impl Clone for Split {
fn clone(&self) -> Self {
Self::new(self.pattern.clone(), self.behavior, self.invert).unwrap()
}
}
impl PartialEq for Split {
fn eq(&self, other: &Self) -> bool {
self.pattern == other.pattern
&& self.behavior == other.behavior
&& self.invert == other.invert
}
}
impl Split {
pub fn new<I: Into<SplitPattern>>(
pattern: I,
behavior: SplitDelimiterBehavior,
invert: bool,
) -> Result<Self> {
let pattern: SplitPattern = pattern.into();
let regex = match &pattern {
SplitPattern::String(s) => SysRegex::new(®ex::escape(s))?,
SplitPattern::Regex(r) => SysRegex::new(r)?,
};
Ok(Self {
pattern,
regex,
behavior,
invert,
})
}
}
impl PreTokenizer for Split {
fn pre_tokenize(&self, pretokenized: &mut PreTokenizedString) -> Result<()> {
if self.invert {
pretokenized.split(|_, normalized| normalized.split(Invert(&self.regex), self.behavior))
} else {
pretokenized.split(|_, normalized| normalized.split(&self.regex, self.behavior))
}
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::{OffsetReferential, OffsetType, PreTokenizer};
use SplitDelimiterBehavior::*;
#[test]
fn basic() {
let tests = vec![
(
Removed,
"How are you doing?",
vec![
("How", (0, 3)),
("are", (4, 7)),
("you", (8, 11)),
("doing", (12, 17)),
("?", (17, 18)),
],
),
(
Isolated,
"How are you doing?",
vec![
("How", (0, 3)),
(" ", (3, 4)),
("are", (4, 7)),
(" ", (7, 8)),
("you", (8, 11)),
(" ", (11, 12)),
("doing", (12, 17)),
("?", (17, 18)),
],
),
(
MergedWithPrevious,
"How are you doing?",
vec![
("How ", (0, 4)),
("are ", (4, 8)),
("you ", (8, 12)),
("doing", (12, 17)),
("?", (17, 18)),
],
),
(
MergedWithNext,
"How are you doing?",
vec![
("How", (0, 3)),
(" are", (3, 7)),
(" you", (7, 11)),
(" doing", (11, 17)),
("?", (17, 18)),
],
),
(
Contiguous,
"How are you doing?",
vec![
("How", (0, 3)),
(" ", (3, 4)),
("are", (4, 7)),
(" ", (7, 8)),
("you", (8, 11)),
(" ", (11, 12)),
("doing?", (12, 18)),
],
),
];
// use whitespace regex
let regex = SplitPattern::Regex(r"\w+|[^\w\s]+".into());
for (behavior, s, res) in tests {
let mut pretokenized = PreTokenizedString::from(s);
let pretok = Split::new(regex.clone(), behavior, true).unwrap();
pretok.pre_tokenize(&mut pretokenized).unwrap();
assert_eq!(
pretokenized
.get_splits(OffsetReferential::Original, OffsetType::Byte)
.into_iter()
.map(|(s, o, _)| (s, o))
.collect::<Vec<_>>(),
res
);
}
}
#[test]
fn regex_string() {
let mut pretok_str_for_regex = PreTokenizedString::from("Hey, man!");
let mut pretok_str_for_string = pretok_str_for_regex.clone();
// pre-tokenizer splits on " " - one from Regex, one from string
let pretokenizer_regex = Split::new(
SplitPattern::Regex(r"\s+".into()),
SplitDelimiterBehavior::Removed,
false,
)
.unwrap();
let pretokenizer_string = Split::new(" ", SplitDelimiterBehavior::Removed, false).unwrap();
pretokenizer_regex
.pre_tokenize(&mut pretok_str_for_regex)
.unwrap();
pretokenizer_string
.pre_tokenize(&mut pretok_str_for_string)
.unwrap();
assert_eq!(pretok_str_for_regex, pretok_str_for_string);
}
#[test]
fn invert() {
let mut pretok_str = PreTokenizedString::from("Hello Hello Hello");
let mut pretok_str_for_invert = pretok_str.clone();
// one pre-tokenizer splits on " " - one splits inverted on "Hello"
let pretokenizer = Split::new(" ", SplitDelimiterBehavior::Removed, false).unwrap();
let pretokenizer_invert =
Split::new("Hello", SplitDelimiterBehavior::Removed, true).unwrap();
pretokenizer.pre_tokenize(&mut pretok_str).unwrap();
pretokenizer_invert
.pre_tokenize(&mut pretok_str_for_invert)
.unwrap();
assert_eq!(pretok_str, pretok_str_for_invert);
}
#[test]
fn serialization() {
use SplitDelimiterBehavior::*;
let split = Split::new("Hello", Removed, true).unwrap();
let split_s =
r#"{"type":"Split","pattern":{"String":"Hello"},"behavior":"Removed","invert":true}"#;
assert_eq!(serde_json::to_string(&split).unwrap(), split_s);
assert_eq!(serde_json::from_str::<Split>(split_s).unwrap(), split);
let split = Split::new(SplitPattern::Regex(r"\s+".into()), Isolated, false).unwrap();
let split_s =
r#"{"type":"Split","pattern":{"Regex":"\\s+"},"behavior":"Isolated","invert":false}"#;
assert_eq!(serde_json::to_string(&split).unwrap(), split_s);
assert_eq!(serde_json::from_str::<Split>(split_s).unwrap(), split);
}
}
|
tokenizers/tokenizers/src/pre_tokenizers/split.rs/0
|
{
"file_path": "tokenizers/tokenizers/src/pre_tokenizers/split.rs",
"repo_id": "tokenizers",
"token_count": 4038
}
| 263
|
use std::marker::PhantomData;
use serde::{
self,
de::{Error, MapAccess, Visitor},
ser::SerializeStruct,
Deserialize, Deserializer, Serialize, Serializer,
};
use super::{added_vocabulary::AddedTokenWithId, TokenizerImpl};
use crate::{Decoder, Model, Normalizer, PostProcessor, PreTokenizer, TokenizerBuilder};
static SERIALIZATION_VERSION: &str = "1.0";
impl<M, N, PT, PP, D> Serialize for TokenizerImpl<M, N, PT, PP, D>
where
M: Serialize,
N: Serialize,
PT: Serialize,
PP: Serialize,
D: Serialize,
{
fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
where
S: Serializer,
{
let mut tokenizer = serializer.serialize_struct("Tokenizer", 9)?;
// Start by adding the current version
tokenizer.serialize_field("version", SERIALIZATION_VERSION)?;
// Params
tokenizer.serialize_field("truncation", &self.truncation)?;
tokenizer.serialize_field("padding", &self.padding)?;
// Added tokens
tokenizer.serialize_field("added_tokens", &self.added_vocabulary)?;
// Then add our parts
tokenizer.serialize_field("normalizer", &self.normalizer)?;
tokenizer.serialize_field("pre_tokenizer", &self.pre_tokenizer)?;
tokenizer.serialize_field("post_processor", &self.post_processor)?;
tokenizer.serialize_field("decoder", &self.decoder)?;
tokenizer.serialize_field("model", &self.model)?;
tokenizer.end()
}
}
impl<'de, M, N, PT, PP, D> Deserialize<'de> for TokenizerImpl<M, N, PT, PP, D>
where
M: Deserialize<'de> + Model,
N: Deserialize<'de> + Normalizer,
PT: Deserialize<'de> + PreTokenizer,
PP: Deserialize<'de> + PostProcessor,
D: Deserialize<'de> + Decoder,
{
fn deserialize<De>(deserializer: De) -> Result<Self, De::Error>
where
De: Deserializer<'de>,
{
deserializer.deserialize_struct(
"Tokenizer",
&[
"version",
"truncation",
"padding",
"added_tokens",
"normalizer",
"pre_tokenizer",
"post_processor",
"decoder",
"model",
],
TokenizerVisitor(
PhantomData,
PhantomData,
PhantomData,
PhantomData,
PhantomData,
),
)
}
}
struct TokenizerVisitor<M, N, PT, PP, D>(
PhantomData<M>,
PhantomData<N>,
PhantomData<PT>,
PhantomData<PP>,
PhantomData<D>,
);
impl<'de, M, N, PT, PP, D> Visitor<'de> for TokenizerVisitor<M, N, PT, PP, D>
where
M: Deserialize<'de> + Model,
N: Deserialize<'de> + Normalizer,
PT: Deserialize<'de> + PreTokenizer,
PP: Deserialize<'de> + PostProcessor,
D: Deserialize<'de> + Decoder,
{
type Value = TokenizerImpl<M, N, PT, PP, D>;
fn expecting(&self, fmt: &mut std::fmt::Formatter) -> std::fmt::Result {
write!(fmt, "struct Tokenizer")
}
fn visit_map<V>(self, mut map: V) -> Result<Self::Value, V::Error>
where
V: MapAccess<'de>,
{
let mut builder = TokenizerBuilder::new();
let mut tokens: Vec<AddedTokenWithId> = vec![];
while let Some(key) = map.next_key::<String>()? {
match key.as_ref() {
"version" => {
let v: String = map.next_value()?;
if &v != "1.0" {
return Err(Error::custom(format!("Unknown tokenizer version '{}'", v)));
}
}
"truncation" => {
builder = builder.with_truncation(map.next_value()?);
}
"padding" => {
builder = builder.with_padding(map.next_value()?);
}
"added_tokens" => {
tokens = map.next_value()?;
}
"normalizer" => {
builder = builder.with_normalizer(map.next_value()?);
}
"pre_tokenizer" => {
builder = builder.with_pre_tokenizer(map.next_value()?);
}
"model" => {
builder = builder.with_model(map.next_value()?);
}
"decoder" => {
builder = builder.with_decoder(map.next_value()?);
}
"post_processor" => {
builder = builder.with_post_processor(map.next_value()?);
}
_ => {}
};
}
let mut tokenizer = builder
.build()
.map_err(|e| V::Error::custom(e.to_string()))?;
// We take care of deserializing the added_tokens (instead of `AddedVocabulary` directly
// because it let us check that associated IDs are still good, and warn the user otherwise
for token in &tokens {
// Warn the user if the id is different than expected
let received_id = tokenizer.token_to_id(&token.token.content);
if let Some(rid) = received_id {
if rid != token.id {
warn!(
"Warning: Token '{}' was expected to have ID '{}' but was given ID '{}'",
token.token.content,
token.id,
rid.to_string()
);
}
}
}
let added_tokens: Vec<_> = tokens.into_iter().map(|token| token.token).collect();
tokenizer.add_tokens(&added_tokens[..]);
Ok(tokenizer)
}
}
#[cfg(test)]
mod tests {
use crate::tokenizer::Tokenizer;
use std::str::FromStr;
#[test]
fn test_deserialization_serialization_invariant() {
let tok_json = r#"{
"version": "1.0",
"truncation": null,
"padding": null,
"added_tokens": [
{
"id": 0,
"content": "[SPECIAL_0]",
"single_word": false,
"lstrip": false,
"rstrip": false,
"normalized": false,
"special": true
},
{
"id": 1,
"content": "[SPECIAL_1]",
"single_word": false,
"lstrip": false,
"rstrip": false,
"normalized": true,
"special": false
},
{
"id": 2,
"content": "[SPECIAL_2]",
"single_word": false,
"lstrip": false,
"rstrip": false,
"normalized": false,
"special": true
}
],
"normalizer": null,
"pre_tokenizer": null,
"post_processor": null,
"decoder": null,
"model": {
"type": "WordPiece",
"unk_token": "[UNK]",
"continuing_subword_prefix": "",
"max_input_chars_per_word": 100,
"vocab": {}
}
}"#;
let tokenizer = Tokenizer::from_str(tok_json).unwrap();
let tok_str = serde_json::to_string_pretty(&tokenizer).unwrap();
// It should be exactly the same as above
assert_eq!(tok_str, tok_json);
}
#[cfg(feature = "http")]
#[test]
fn test_from_pretrained() {
tracing_subscriber::fmt()
.with_max_level(tracing::Level::DEBUG)
.with_target(false)
.init();
let _ = Tokenizer::from_pretrained("Qwen/Qwen2-7B-Instruct", None);
warn!("This should be the first warning");
}
}
|
tokenizers/tokenizers/src/tokenizer/serialization.rs/0
|
{
"file_path": "tokenizers/tokenizers/src/tokenizer/serialization.rs",
"repo_id": "tokenizers",
"token_count": 3739
}
| 264
|
mod common;
use common::*;
use tokenizers::decoders::byte_level::ByteLevel;
use tokenizers::decoders::DecoderWrapper;
use tokenizers::models::bpe::BPE;
use tokenizers::models::wordlevel::WordLevel;
use tokenizers::models::wordpiece::WordPiece;
use tokenizers::models::ModelWrapper;
use tokenizers::normalizers::bert::BertNormalizer;
use tokenizers::normalizers::unicode::{NFC, NFKC};
use tokenizers::normalizers::NormalizerWrapper;
use tokenizers::pre_tokenizers::bert::BertPreTokenizer;
use tokenizers::pre_tokenizers::delimiter::CharDelimiterSplit;
use tokenizers::pre_tokenizers::split::{Split, SplitPattern};
use tokenizers::pre_tokenizers::whitespace::Whitespace;
use tokenizers::pre_tokenizers::PreTokenizerWrapper;
use tokenizers::processors::bert::BertProcessing;
use tokenizers::processors::PostProcessorWrapper;
use tokenizers::{SplitDelimiterBehavior, Tokenizer, TokenizerImpl};
#[test]
fn bpe_serde() {
let bpe = get_byte_level_bpe();
let ser = serde_json::to_string(&bpe).unwrap();
let de = serde_json::from_str(&ser).unwrap();
assert_eq!(bpe, de);
}
#[test]
fn wordpiece_serde() {
let wordpiece = get_bert_wordpiece();
let ser = serde_json::to_string(&wordpiece).unwrap();
let de = serde_json::from_str(&ser).unwrap();
assert_eq!(wordpiece, de);
}
#[test]
fn wordlevel_serde() {
let wordlevel = WordLevel::from_file("data/gpt2-vocab.json", "<unk>".into()).unwrap();
let ser = serde_json::to_string(&wordlevel).unwrap();
let de = serde_json::from_str(&ser).unwrap();
assert_eq!(wordlevel, de);
}
#[test]
fn normalizers() {
// Test unit struct
let nfc = NFC;
let nfc_ser = serde_json::to_string(&nfc).unwrap();
assert_eq!(nfc_ser, r#"{"type":"NFC"}"#);
// empty struct can deserialize from self
serde_json::from_str::<NFC>(&nfc_ser).unwrap();
let err: Result<NFKC, _> = serde_json::from_str(&nfc_ser);
assert!(err.is_err(), "NFKC shouldn't be deserializable from NFC");
// wrapper can can deserialize from inner
let nfc_wrapped: NormalizerWrapper = serde_json::from_str(&nfc_ser).unwrap();
match &nfc_wrapped {
NormalizerWrapper::NFC(_) => (),
_ => panic!("NFC wrapped with incorrect variant"),
}
let ser_wrapped = serde_json::to_string(&nfc_wrapped).unwrap();
assert_eq!(ser_wrapped, nfc_ser);
// Test non-empty roundtrip
let bert = BertNormalizer::default();
let bert_ser = serde_json::to_string(&bert).unwrap();
assert_eq!(
bert_ser,
r#"{"type":"BertNormalizer","clean_text":true,"handle_chinese_chars":true,"strip_accents":null,"lowercase":true}"#
);
// make sure we can deserialize to self
serde_json::from_str::<BertNormalizer>(&bert_ser).unwrap();
// wrapper can deserialize from inner serialization
let bert_wrapped: NormalizerWrapper = serde_json::from_str(&bert_ser).unwrap();
match &bert_wrapped {
NormalizerWrapper::BertNormalizer(_) => (),
_ => panic!("BertNormalizer wrapped with incorrect variant"),
}
// wrapped serializes same way as inner
let ser_wrapped = serde_json::to_string(&bert_wrapped).unwrap();
assert_eq!(ser_wrapped, bert_ser);
}
#[test]
fn processors() {
let bert = BertProcessing::new(("SEP".into(), 0), ("CLS".into(), 0));
let bert_ser = serde_json::to_string(&bert).unwrap();
assert_eq!(
bert_ser,
r#"{"type":"BertProcessing","sep":["SEP",0],"cls":["CLS",0]}"#
);
serde_json::from_str::<BertProcessing>(&bert_ser).unwrap();
let bert_wrapped: PostProcessorWrapper = serde_json::from_str(&bert_ser).unwrap();
match &bert_wrapped {
PostProcessorWrapper::Bert(_) => (),
_ => panic!("Bert wrapped with incorrect variant"),
}
let ser_wrapped = serde_json::to_string(&bert_wrapped).unwrap();
assert_eq!(ser_wrapped, bert_ser);
}
#[test]
fn pretoks() {
// Test unit struct
let bert = BertPreTokenizer;
let bert_ser = serde_json::to_string(&bert).unwrap();
assert_eq!(bert_ser, r#"{"type":"BertPreTokenizer"}"#);
// empty struct can deserialize from self
serde_json::from_str::<BertPreTokenizer>(&bert_ser).unwrap();
let err: Result<Whitespace, _> = serde_json::from_str(&bert_ser);
assert!(
err.is_err(),
"Whitespace shouldn't be deserializable from BertPreTokenizer"
);
// wrapper can can deserialize from inner
let bert_wrapped: PreTokenizerWrapper = serde_json::from_str(&bert_ser).unwrap();
match &bert_wrapped {
PreTokenizerWrapper::BertPreTokenizer(_) => (),
_ => panic!("Bert wrapped with incorrect variant"),
}
let ser_wrapped = serde_json::to_string(&bert_wrapped).unwrap();
assert_eq!(ser_wrapped, bert_ser);
// Test non-empty roundtrip
let ch = CharDelimiterSplit::new(' ');
let ch_ser = serde_json::to_string(&ch).unwrap();
assert_eq!(ch_ser, r#"{"type":"CharDelimiterSplit","delimiter":" "}"#);
// make sure we can deserialize to self
serde_json::from_str::<CharDelimiterSplit>(&ch_ser).unwrap();
// wrapper can deserialize from inner serialization
let ch_wrapped: PreTokenizerWrapper = serde_json::from_str(&ch_ser).unwrap();
match &ch_wrapped {
PreTokenizerWrapper::Delimiter(_) => (),
_ => panic!("CharDelimiterSplit wrapped with incorrect variant"),
}
// wrapped serializes same way as inner
let ser_wrapped = serde_json::to_string(&ch_wrapped).unwrap();
assert_eq!(ser_wrapped, ch_ser);
let wsp = Whitespace {};
let wsp_ser = serde_json::to_string(&wsp).unwrap();
assert_eq!(wsp_ser, r#"{"type":"Whitespace"}"#);
serde_json::from_str::<Whitespace>(&wsp_ser).unwrap();
let err: Result<BertPreTokenizer, _> = serde_json::from_str(&wsp_ser);
assert!(
err.is_err(),
"BertPreTokenizer shouldn't be deserializable from Whitespace"
);
let pattern: SplitPattern = "[SEP]".into();
let pretok = Split::new(pattern, SplitDelimiterBehavior::Isolated, false).unwrap();
let pretok_str = serde_json::to_string(&pretok).unwrap();
assert_eq!(
pretok_str,
r#"{"type":"Split","pattern":{"String":"[SEP]"},"behavior":"Isolated","invert":false}"#
);
assert_eq!(serde_json::from_str::<Split>(&pretok_str).unwrap(), pretok);
let pattern = SplitPattern::Regex("[SEP]".to_string());
let pretok = Split::new(pattern, SplitDelimiterBehavior::Isolated, false).unwrap();
let pretok_str = serde_json::to_string(&pretok).unwrap();
assert_eq!(
pretok_str,
r#"{"type":"Split","pattern":{"Regex":"[SEP]"},"behavior":"Isolated","invert":false}"#
);
assert_eq!(serde_json::from_str::<Split>(&pretok_str).unwrap(), pretok);
}
#[test]
fn decoders() {
let byte_level = ByteLevel::default();
let byte_level_ser = serde_json::to_string(&byte_level).unwrap();
assert_eq!(
byte_level_ser,
r#"{"type":"ByteLevel","add_prefix_space":true,"trim_offsets":true,"use_regex":true}"#
);
serde_json::from_str::<ByteLevel>(&byte_level_ser).unwrap();
let byte_level_wrapper: DecoderWrapper = serde_json::from_str(&byte_level_ser).unwrap();
match &byte_level_wrapper {
DecoderWrapper::ByteLevel(_) => (),
_ => panic!("ByteLevel wrapped with incorrect variant"),
}
let ser_wrapped = serde_json::to_string(&byte_level_wrapper).unwrap();
assert_eq!(ser_wrapped, byte_level_ser);
}
#[test]
fn models() {
let bpe = BPE::default();
let bpe_ser = serde_json::to_string(&bpe).unwrap();
serde_json::from_str::<BPE>(&bpe_ser).unwrap();
let bpe_wrapper: ModelWrapper = serde_json::from_str(&bpe_ser).unwrap();
match &bpe_wrapper {
ModelWrapper::BPE(_) => (),
_ => panic!("BPE wrapped with incorrect variant"),
}
let ser_wrapped = serde_json::to_string(&bpe_wrapper).unwrap();
assert_eq!(ser_wrapped, bpe_ser);
}
#[test]
fn tokenizer() {
let wordpiece = WordPiece::default();
let mut tokenizer = Tokenizer::new(wordpiece);
tokenizer.with_normalizer(Some(NFC));
let ser = serde_json::to_string(&tokenizer).unwrap();
let _: Tokenizer = serde_json::from_str(&ser).unwrap();
let unwrapped_nfc_tok: TokenizerImpl<
WordPiece,
NFC,
PreTokenizerWrapper,
PostProcessorWrapper,
DecoderWrapper,
> = serde_json::from_str(&ser).unwrap();
assert_eq!(serde_json::to_string(&unwrapped_nfc_tok).unwrap(), ser);
let err: Result<
TokenizerImpl<WordPiece, NFKC, PreTokenizerWrapper, PostProcessorWrapper, DecoderWrapper>,
_,
> = serde_json::from_str(&ser);
assert!(err.is_err(), "NFKC shouldn't be deserializable from NFC");
let de: TokenizerImpl<
WordPiece,
NormalizerWrapper,
PreTokenizerWrapper,
PostProcessorWrapper,
DecoderWrapper,
> = serde_json::from_str(&ser).unwrap();
assert_eq!(serde_json::to_string(&de).unwrap(), ser);
}
#[test]
fn bpe_with_dropout_serde() {
let mut bpe = BPE::default();
bpe.dropout = Some(0.1);
let ser = serde_json::to_string(&bpe).unwrap();
let de = serde_json::from_str(&ser).unwrap();
assert_eq!(bpe, de);
// set dropout to 0.0 (which is analogous to None) and reserialize
bpe.dropout = Some(0.0);
let ser = serde_json::to_string(&bpe).unwrap();
let de = serde_json::from_str(&ser).unwrap();
assert_eq!(bpe, de);
}
#[test]
fn test_deserialize_long_file() {
let _tokenizer = Tokenizer::from_file("data/albert-base-v1-tokenizer.json").unwrap();
}
|
tokenizers/tokenizers/tests/serialization.rs/0
|
{
"file_path": "tokenizers/tokenizers/tests/serialization.rs",
"repo_id": "tokenizers",
"token_count": 3890
}
| 265
|
# 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.
# tests directory-specific settings - this file is run automatically
# by pytest before any tests are run
import doctest
import sys
import warnings
from os.path import abspath, dirname, join
import _pytest
import pytest
from transformers.testing_utils import HfDoctestModule, HfDocTestParser
NOT_DEVICE_TESTS = {
"test_tokenization",
"test_processor",
"test_processing",
"test_beam_constraints",
"test_configuration_utils",
"test_data_collator",
"test_trainer_callback",
"test_trainer_utils",
"test_feature_extraction",
"test_image_processing",
"test_image_processor",
"test_image_transforms",
"test_optimization",
"test_retrieval",
"test_config",
"test_from_pretrained_no_checkpoint",
"test_keep_in_fp32_modules",
"test_gradient_checkpointing_backward_compatibility",
"test_gradient_checkpointing_enable_disable",
"test_save_load_fast_init_from_base",
"test_fast_init_context_manager",
"test_fast_init_tied_embeddings",
"test_save_load_fast_init_to_base",
"test_torch_save_load",
"test_initialization",
"test_forward_signature",
"test_model_get_set_embeddings",
"test_model_main_input_name",
"test_correct_missing_keys",
"test_tie_model_weights",
"test_can_use_safetensors",
"test_load_save_without_tied_weights",
"test_tied_weights_keys",
"test_model_weights_reload_no_missing_tied_weights",
"test_pt_tf_model_equivalence",
"test_mismatched_shapes_have_properly_initialized_weights",
"test_matched_shapes_have_loaded_weights_when_some_mismatched_shapes_exist",
"test_model_is_small",
"test_tf_from_pt_safetensors",
"test_flax_from_pt_safetensors",
"ModelTest::test_pipeline_", # None of the pipeline tests from PipelineTesterMixin (of which XxxModelTest inherits from) are running on device
"ModelTester::test_pipeline_",
"/repo_utils/",
"/utils/",
"/agents/",
}
# allow having multiple repository checkouts and not needing to remember to rerun
# `pip install -e '.[dev]'` when switching between checkouts and running tests.
git_repo_path = abspath(join(dirname(__file__), "src"))
sys.path.insert(1, git_repo_path)
# silence FutureWarning warnings in tests since often we can't act on them until
# they become normal warnings - i.e. the tests still need to test the current functionality
warnings.simplefilter(action="ignore", category=FutureWarning)
def pytest_configure(config):
config.addinivalue_line(
"markers", "is_pt_tf_cross_test: mark test to run only when PT and TF interactions are tested"
)
config.addinivalue_line(
"markers", "is_pt_flax_cross_test: mark test to run only when PT and FLAX interactions are tested"
)
config.addinivalue_line("markers", "is_pipeline_test: mark test to run only when pipelines are tested")
config.addinivalue_line("markers", "is_staging_test: mark test to run only in the staging environment")
config.addinivalue_line("markers", "accelerate_tests: mark test that require accelerate")
config.addinivalue_line("markers", "agent_tests: mark the agent tests that are run on their specific schedule")
config.addinivalue_line("markers", "not_device_test: mark the tests always running on cpu")
def pytest_collection_modifyitems(items):
for item in items:
if any(test_name in item.nodeid for test_name in NOT_DEVICE_TESTS):
item.add_marker(pytest.mark.not_device_test)
def pytest_addoption(parser):
from transformers.testing_utils import pytest_addoption_shared
pytest_addoption_shared(parser)
def pytest_terminal_summary(terminalreporter):
from transformers.testing_utils import pytest_terminal_summary_main
make_reports = terminalreporter.config.getoption("--make-reports")
if make_reports:
pytest_terminal_summary_main(terminalreporter, id=make_reports)
def pytest_sessionfinish(session, exitstatus):
# If no tests are collected, pytest exists with code 5, which makes the CI fail.
if exitstatus == 5:
session.exitstatus = 0
# Doctest custom flag to ignore output.
IGNORE_RESULT = doctest.register_optionflag("IGNORE_RESULT")
OutputChecker = doctest.OutputChecker
class CustomOutputChecker(OutputChecker):
def check_output(self, want, got, optionflags):
if IGNORE_RESULT & optionflags:
return True
return OutputChecker.check_output(self, want, got, optionflags)
doctest.OutputChecker = CustomOutputChecker
_pytest.doctest.DoctestModule = HfDoctestModule
doctest.DocTestParser = HfDocTestParser
|
transformers/conftest.py/0
|
{
"file_path": "transformers/conftest.py",
"repo_id": "transformers",
"token_count": 1806
}
| 266
|
FROM nvidia/cuda:10.2-cudnn7-devel-ubuntu18.04
LABEL maintainer="Hugging Face"
LABEL repository="transformers"
RUN apt update && \
apt install -y bash \
build-essential \
git \
curl \
ca-certificates \
python3 \
python3-pip && \
rm -rf /var/lib/apt/lists
RUN python3 -m pip install --no-cache-dir --upgrade pip && \
python3 -m pip install --no-cache-dir \
jupyter \
tensorflow \
torch
RUN git clone https://github.com/NVIDIA/apex
RUN cd apex && \
python3 setup.py install && \
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
WORKDIR /workspace
COPY . transformers/
RUN cd transformers/ && \
python3 -m pip install --no-cache-dir .
CMD ["/bin/bash"]
|
transformers/docker/transformers-gpu/Dockerfile/0
|
{
"file_path": "transformers/docker/transformers-gpu/Dockerfile",
"repo_id": "transformers",
"token_count": 397
}
| 267
|
<!--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.
-->
# Trainieren mit einem Skript
Neben den 🤗 Transformers [notebooks](./notebooks) gibt es auch Beispielskripte, die zeigen, wie man ein Modell für eine Aufgabe mit [PyTorch](https://github.com/huggingface/transformers/tree/main/examples/pytorch), [TensorFlow](https://github.com/huggingface/transformers/tree/main/examples/tensorflow) oder [JAX/Flax](https://github.com/huggingface/transformers/tree/main/examples/flax) trainiert.
Sie werden auch Skripte finden, die wir in unseren [Forschungsprojekten](https://github.com/huggingface/transformers/tree/main/examples/research_projects) und [Legacy-Beispielen](https://github.com/huggingface/transformers/tree/main/examples/legacy) verwendet haben und die größtenteils von der Community stammen. Diese Skripte werden nicht aktiv gepflegt und erfordern eine bestimmte Version von 🤗 Transformers, die höchstwahrscheinlich nicht mit der neuesten Version der Bibliothek kompatibel ist.
Es wird nicht erwartet, dass die Beispielskripte bei jedem Problem sofort funktionieren. Möglicherweise müssen Sie das Skript an das Problem anpassen, das Sie zu lösen versuchen. Um Ihnen dabei zu helfen, legen die meisten Skripte vollständig offen, wie die Daten vorverarbeitet werden, so dass Sie sie nach Bedarf für Ihren Anwendungsfall bearbeiten können.
Für jede Funktion, die Sie in einem Beispielskript implementieren möchten, diskutieren Sie bitte im [Forum](https://discuss.huggingface.co/) oder in einem [issue](https://github.com/huggingface/transformers/issues), bevor Sie einen Pull Request einreichen. Wir freuen uns zwar über Fehlerkorrekturen, aber es ist unwahrscheinlich, dass wir einen Pull Request zusammenführen, der mehr Funktionalität auf Kosten der Lesbarkeit hinzufügt.
Diese Anleitung zeigt Ihnen, wie Sie ein Beispiel für ein Trainingsskript zur Zusammenfassung in [PyTorch](https://github.com/huggingface/transformers/tree/main/examples/pytorch/summarization) und [TensorFlow](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/summarization) ausführen können. Sofern nicht anders angegeben, sollten alle Beispiele mit beiden Frameworks funktionieren.
## Einrichtung
Um die neueste Version der Beispielskripte erfolgreich auszuführen, **müssen Sie 🤗 Transformers aus dem Quellcode** in einer neuen virtuellen Umgebung installieren:
```bash
git clone https://github.com/huggingface/transformers
cd transformers
pip install .
```
Für ältere Versionen der Beispielskripte klicken Sie auf die Umschalttaste unten:
<details>
<summary>Beispiele für ältere Versionen von 🤗 Transformers</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>
Dann stellen Sie Ihren aktuellen Klon von 🤗 Transformers auf eine bestimmte Version um, z.B. v3.5.1:
```bash
git checkout tags/v3.5.1
```
Nachdem Sie die richtige Bibliotheksversion eingerichtet haben, navigieren Sie zu dem Beispielordner Ihrer Wahl und installieren die beispielspezifischen Anforderungen:
```bash
pip install -r requirements.txt
```
## Ein Skript ausführen
<frameworkcontent>
<pt>
Das Beispielskript lädt einen Datensatz aus der 🤗 [Datasets](https://huggingface.co/docs/datasets/) Bibliothek herunter und verarbeitet ihn vor. Dann nimmt das Skript eine Feinabstimmung eines Datensatzes mit dem [Trainer](https://huggingface.co/docs/transformers/main_classes/trainer) auf einer Architektur vor, die eine Zusammenfassung unterstützt. Das folgende Beispiel zeigt, wie die Feinabstimmung von [T5-small](https://huggingface.co/google-t5/t5-small) auf dem Datensatz [CNN/DailyMail](https://huggingface.co/datasets/cnn_dailymail) durchgeführt wird. Das T5-Modell benötigt aufgrund der Art und Weise, wie es trainiert wurde, ein zusätzliches Argument `source_prefix`. Mit dieser Eingabeaufforderung weiß T5, dass es sich um eine Zusammenfassungsaufgabe handelt.
```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>
Das Beispielskript lädt einen Datensatz aus der 🤗 [Datasets](https://huggingface.co/docs/datasets/) Bibliothek herunter und verarbeitet ihn vor. Anschließend nimmt das Skript die Feinabstimmung eines Datensatzes mit Keras auf einer Architektur vor, die die Zusammenfassung unterstützt. Das folgende Beispiel zeigt, wie die Feinabstimmung von [T5-small](https://huggingface.co/google-t5/t5-small) auf dem [CNN/DailyMail](https://huggingface.co/datasets/cnn_dailymail) Datensatz durchgeführt wird. Das T5-Modell benötigt aufgrund der Art und Weise, wie es trainiert wurde, ein zusätzliches Argument `source_prefix`. Mit dieser Eingabeaufforderung weiß T5, dass es sich um eine Zusammenfassungsaufgabe handelt.
```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>
## Verteiltes Training und gemischte Präzision
Der [Trainer](https://huggingface.co/docs/transformers/main_classes/trainer) unterstützt verteiltes Training und gemischte Präzision, d.h. Sie können ihn auch in einem Skript verwenden. So aktivieren Sie diese beiden Funktionen:
- Fügen Sie das Argument `fp16` hinzu, um gemischte Genauigkeit zu aktivieren.
- Legen Sie die Anzahl der zu verwendenden GPUs mit dem Argument `nproc_per_node` fest.
```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-Skripte verwenden eine [`MirroredStrategy`](https://www.tensorflow.org/guide/distributed_training#mirroredstrategy) für verteiltes Training, und Sie müssen dem Trainingsskript keine zusätzlichen Argumente hinzufügen. Das TensorFlow-Skript verwendet standardmäßig mehrere GPUs, wenn diese verfügbar sind.
## Ein Skript auf einer TPU ausführen
<frameworkcontent>
<pt>
Tensor Processing Units (TPUs) sind speziell für die Beschleunigung der Leistung konzipiert. PyTorch unterstützt TPUs mit dem [XLA](https://www.tensorflow.org/xla) Deep Learning Compiler (siehe [hier](https://github.com/pytorch/xla/blob/master/README.md) für weitere Details). Um eine TPU zu verwenden, starten Sie das Skript `xla_spawn.py` und verwenden das Argument `num_cores`, um die Anzahl der TPU-Kerne festzulegen, die Sie verwenden möchten.
```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) sind speziell für die Beschleunigung der Leistung konzipiert. TensorFlow Skripte verwenden eine [`TPUStrategy`](https://www.tensorflow.org/guide/distributed_training#tpustrategy) für das Training auf TPUs. Um eine TPU zu verwenden, übergeben Sie den Namen der TPU-Ressource an das Argument `tpu`.
```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>
## Führen Sie ein Skript mit 🤗 Accelerate aus.
🤗 [Accelerate](https://huggingface.co/docs/accelerate) ist eine reine PyTorch-Bibliothek, die eine einheitliche Methode für das Training eines Modells auf verschiedenen Arten von Setups (nur CPU, mehrere GPUs, TPUs) bietet und dabei die vollständige Transparenz der PyTorch-Trainingsschleife beibehält. Stellen Sie sicher, dass Sie 🤗 Accelerate installiert haben, wenn Sie es nicht bereits haben:
> Hinweis: Da Accelerate schnell weiterentwickelt wird, muss die Git-Version von Accelerate installiert sein, um die Skripte auszuführen.
```bash
pip install git+https://github.com/huggingface/accelerate
```
Anstelle des Skripts `run_summarization.py` müssen Sie das Skript `run_summarization_no_trainer.py` verwenden. Die von Accelerate unterstützten Skripte haben eine Datei `task_no_trainer.py` im Ordner. Beginnen Sie mit dem folgenden Befehl, um eine Konfigurationsdatei zu erstellen und zu speichern:
```bash
accelerate config
```
Testen Sie Ihre Einrichtung, um sicherzustellen, dass sie korrekt konfiguriert ist:
```bash
accelerate test
```
Jetzt sind Sie bereit, das Training zu starten:
```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
```
## Verwenden Sie einen benutzerdefinierten Datensatz
Das Verdichtungsskript unterstützt benutzerdefinierte Datensätze, solange es sich um eine CSV- oder JSON-Line-Datei handelt. Wenn Sie Ihren eigenen Datensatz verwenden, müssen Sie mehrere zusätzliche Argumente angeben:
- `train_file` und `validation_file` geben den Pfad zu Ihren Trainings- und Validierungsdateien an.
- `text_column` ist der Eingabetext, der zusammengefasst werden soll.
- Summary_column" ist der auszugebende Zieltext.
Ein Zusammenfassungsskript, das einen benutzerdefinierten Datensatz verwendet, würde wie folgt aussehen:
```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
```
## Testen Sie ein Skript
Es ist oft eine gute Idee, Ihr Skript an einer kleineren Anzahl von Beispielen für Datensätze auszuführen, um sicherzustellen, dass alles wie erwartet funktioniert, bevor Sie sich auf einen ganzen Datensatz festlegen, dessen Fertigstellung Stunden dauern kann. Verwenden Sie die folgenden Argumente, um den Datensatz auf eine maximale Anzahl von Stichproben zu beschränken:
- `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
```
Nicht alle Beispielskripte unterstützen das Argument `max_predict_samples`. Wenn Sie sich nicht sicher sind, ob Ihr Skript dieses Argument unterstützt, fügen Sie das Argument `-h` hinzu, um dies zu überprüfen:
```bash
examples/pytorch/summarization/run_summarization.py -h
```
## Training vom Kontrollpunkt fortsetzen
Eine weitere hilfreiche Option, die Sie aktivieren können, ist die Wiederaufnahme des Trainings von einem früheren Kontrollpunkt aus. Auf diese Weise können Sie im Falle einer Unterbrechung Ihres Trainings dort weitermachen, wo Sie aufgehört haben, ohne von vorne beginnen zu müssen. Es gibt zwei Methoden, um das Training von einem Kontrollpunkt aus wieder aufzunehmen.
Die erste Methode verwendet das Argument `output_dir previous_output_dir`, um das Training ab dem letzten in `output_dir` gespeicherten Kontrollpunkt wieder aufzunehmen. In diesem Fall sollten Sie `overwrite_output_dir` entfernen:
```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
```
Die zweite Methode verwendet das Argument `Resume_from_checkpoint path_to_specific_checkpoint`, um das Training ab einem bestimmten Checkpoint-Ordner wieder aufzunehmen.
```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
```
## Teilen Sie Ihr Modell
Alle Skripte können Ihr endgültiges Modell in den [Model Hub](https://huggingface.co/models) hochladen. Stellen Sie sicher, dass Sie bei Hugging Face angemeldet sind, bevor Sie beginnen:
```bash
huggingface-cli login
```
Dann fügen Sie dem Skript das Argument `push_to_hub` hinzu. Mit diesem Argument wird ein Repository mit Ihrem Hugging Face-Benutzernamen und dem in `output_dir` angegebenen Ordnernamen erstellt.
Wenn Sie Ihrem Repository einen bestimmten Namen geben möchten, fügen Sie ihn mit dem Argument `push_to_hub_model_id` hinzu. Das Repository wird automatisch unter Ihrem Namensraum aufgeführt.
Das folgende Beispiel zeigt, wie Sie ein Modell mit einem bestimmten Repository-Namen hochladen können:
```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/de/run_scripts.md/0
|
{
"file_path": "transformers/docs/source/de/run_scripts.md",
"repo_id": "transformers",
"token_count": 7515
}
| 268
|
<!--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.
-->
# Chat Templates
## Introduction
An increasingly common use case for LLMs is **chat**. In a chat context, rather than continuing a single string
of text (as is the case with a standard language model), the model instead continues a conversation that consists
of one or more **messages**, each of which includes a **role**, like "user" or "assistant", as well as message text.
Much like tokenization, different models expect very different input formats for chat. This is the reason we added
**chat templates** as a feature. Chat templates are part of the tokenizer. They specify how to convert conversations,
represented as lists of messages, into a single tokenizable string in the format that the model expects.
Let's make this concrete with a quick example using the `BlenderBot` model. BlenderBot has an extremely simple default
template, which mostly just adds whitespace between rounds of dialogue:
```python
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
>>> chat = [
... {"role": "user", "content": "Hello, how are you?"},
... {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
... {"role": "user", "content": "I'd like to show off how chat templating works!"},
... ]
>>> tokenizer.apply_chat_template(chat, tokenize=False)
" Hello, how are you? I'm doing great. How can I help you today? I'd like to show off how chat templating works!</s>"
```
Notice how the entire chat is condensed into a single string. If we use `tokenize=True`, which is the default setting,
that string will also be tokenized for us. To see a more complex template in action, though, let's use the
`mistralai/Mistral-7B-Instruct-v0.1` model.
```python
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.1")
>>> chat = [
... {"role": "user", "content": "Hello, how are you?"},
... {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
... {"role": "user", "content": "I'd like to show off how chat templating works!"},
... ]
>>> tokenizer.apply_chat_template(chat, tokenize=False)
"<s>[INST] Hello, how are you? [/INST]I'm doing great. How can I help you today?</s> [INST] I'd like to show off how chat templating works! [/INST]"
```
Note that this time, the tokenizer has added the control tokens [INST] and [/INST] to indicate the start and end of
user messages (but not assistant messages!). Mistral-instruct was trained with these tokens, but BlenderBot was not.
## How do I use chat templates?
As you can see in the example above, chat templates are easy to use. Simply build a list of messages, with `role`
and `content` keys, and then pass it to the [`~PreTrainedTokenizer.apply_chat_template`] method. Once you do that,
you'll get output that's ready to go! When using chat templates as input for model generation, it's also a good idea
to use `add_generation_prompt=True` to add a [generation prompt](#what-are-generation-prompts).
Here's an example of preparing input for `model.generate()`, using the `Zephyr` assistant model:
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
checkpoint = "HuggingFaceH4/zephyr-7b-beta"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint) # You may want to use bfloat16 and/or move to GPU here
messages = [
{
"role": "system",
"content": "You are a friendly chatbot who always responds in the style of a pirate",
},
{"role": "user", "content": "How many helicopters can a human eat in one sitting?"},
]
tokenized_chat = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt")
print(tokenizer.decode(tokenized_chat[0]))
```
This will yield a string in the input format that Zephyr expects.
```text
<|system|>
You are a friendly chatbot who always responds in the style of a pirate</s>
<|user|>
How many helicopters can a human eat in one sitting?</s>
<|assistant|>
```
Now that our input is formatted correctly for Zephyr, we can use the model to generate a response to the user's question:
```python
outputs = model.generate(tokenized_chat, max_new_tokens=128)
print(tokenizer.decode(outputs[0]))
```
This will yield:
```text
<|system|>
You are a friendly chatbot who always responds in the style of a pirate</s>
<|user|>
How many helicopters can a human eat in one sitting?</s>
<|assistant|>
Matey, I'm afraid I must inform ye that humans cannot eat helicopters. Helicopters are not food, they are flying machines. Food is meant to be eaten, like a hearty plate o' grog, a savory bowl o' stew, or a delicious loaf o' bread. But helicopters, they be for transportin' and movin' around, not for eatin'. So, I'd say none, me hearties. None at all.
```
Arr, 'twas easy after all!
## Is there an automated pipeline for chat?
Yes, there is! Our text generation pipelines support chat inputs, which makes it easy to use chat models. In the past,
we used to use a dedicated "ConversationalPipeline" class, but this has now been deprecated and its functionality
has been merged into the [`TextGenerationPipeline`]. Let's try the `Zephyr` example again, but this time using
a pipeline:
```python
from transformers import pipeline
pipe = pipeline("text-generation", "HuggingFaceH4/zephyr-7b-beta")
messages = [
{
"role": "system",
"content": "You are a friendly chatbot who always responds in the style of a pirate",
},
{"role": "user", "content": "How many helicopters can a human eat in one sitting?"},
]
print(pipe(messages, max_new_tokens=128)[0]['generated_text'][-1]) # Print the assistant's response
```
```text
{'role': 'assistant', 'content': "Matey, I'm afraid I must inform ye that humans cannot eat helicopters. Helicopters are not food, they are flying machines. Food is meant to be eaten, like a hearty plate o' grog, a savory bowl o' stew, or a delicious loaf o' bread. But helicopters, they be for transportin' and movin' around, not for eatin'. So, I'd say none, me hearties. None at all."}
```
The pipeline will take care of all the details of tokenization and calling `apply_chat_template` for you -
once the model has a chat template, all you need to do is initialize the pipeline and pass it the list of messages!
## What are "generation prompts"?
You may have noticed that the `apply_chat_template` method has an `add_generation_prompt` argument. This argument tells
the template to add tokens that indicate the start of a bot response. For example, consider the following chat:
```python
messages = [
{"role": "user", "content": "Hi there!"},
{"role": "assistant", "content": "Nice to meet you!"},
{"role": "user", "content": "Can I ask a question?"}
]
```
Here's what this will look like without a generation prompt, using the ChatML template we saw in the Zephyr example:
```python
tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)
"""<|im_start|>user
Hi there!<|im_end|>
<|im_start|>assistant
Nice to meet you!<|im_end|>
<|im_start|>user
Can I ask a question?<|im_end|>
"""
```
And here's what it looks like **with** a generation prompt:
```python
tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
"""<|im_start|>user
Hi there!<|im_end|>
<|im_start|>assistant
Nice to meet you!<|im_end|>
<|im_start|>user
Can I ask a question?<|im_end|>
<|im_start|>assistant
"""
```
Note that this time, we've added the tokens that indicate the start of a bot response. This ensures that when the model
generates text it will write a bot response instead of doing something unexpected, like continuing the user's
message. Remember, chat models are still just language models - they're trained to continue text, and chat is just a
special kind of text to them! You need to guide them with appropriate control tokens, so they know what they're
supposed to be doing.
Not all models require generation prompts. Some models, like BlenderBot and LLaMA, don't have any
special tokens before bot responses. In these cases, the `add_generation_prompt` argument will have no effect. The exact
effect that `add_generation_prompt` has will depend on the template being used.
## Can I use chat templates in training?
Yes! This is a good way to ensure that the chat template matches the tokens the model sees during training.
We recommend that you apply the chat template as a preprocessing step for your dataset. After this, you
can simply continue like any other language model training task. When training, you should usually set
`add_generation_prompt=False`, because the added tokens to prompt an assistant response will not be helpful during
training. Let's see an example:
```python
from transformers import AutoTokenizer
from datasets import Dataset
tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-beta")
chat1 = [
{"role": "user", "content": "Which is bigger, the moon or the sun?"},
{"role": "assistant", "content": "The sun."}
]
chat2 = [
{"role": "user", "content": "Which is bigger, a virus or a bacterium?"},
{"role": "assistant", "content": "A bacterium."}
]
dataset = Dataset.from_dict({"chat": [chat1, chat2]})
dataset = dataset.map(lambda x: {"formatted_chat": tokenizer.apply_chat_template(x["chat"], tokenize=False, add_generation_prompt=False)})
print(dataset['formatted_chat'][0])
```
And we get:
```text
<|user|>
Which is bigger, the moon or the sun?</s>
<|assistant|>
The sun.</s>
```
From here, just continue training like you would with a standard language modelling task, using the `formatted_chat` column.
<Tip>
By default, some tokenizers add special tokens like `<bos>` and `<eos>` to text they tokenize. Chat templates should
already include all the special tokens they need, and so additional special tokens will often be incorrect or
duplicated, which will hurt model performance.
Therefore, if you format text with `apply_chat_template(tokenize=False)`, you should set the argument
`add_special_tokens=False` when you tokenize that text later. If you use `apply_chat_template(tokenize=True)`, you don't need to worry about this!
</Tip>
## Advanced: Extra inputs to chat templates
The only argument that `apply_chat_template` requires is `messages`. However, you can pass any keyword
argument to `apply_chat_template` and it will be accessible inside the template. This gives you a lot of freedom to use
chat templates for many things. There are no restrictions on the names or the format of these arguments - you can pass
strings, lists, dicts or whatever else you want.
That said, there are some common use-cases for these extra arguments,
such as passing tools for function calling, or documents for retrieval-augmented generation. In these common cases,
we have some opinionated recommendations about what the names and formats of these arguments should be, which are
described in the sections below. We encourage model authors to make their chat templates compatible with this format,
to make it easy to transfer tool-calling code between models.
## Advanced: Tool use / function calling
"Tool use" LLMs can choose to call functions as external tools before generating an answer. When passing tools
to a tool-use model, you can simply pass a list of functions to the `tools` argument:
```python
import datetime
def current_time():
"""Get the current local time as a string."""
return str(datetime.now())
def multiply(a: float, b: float):
"""
A function that multiplies two numbers
Args:
a: The first number to multiply
b: The second number to multiply
"""
return a * b
tools = [current_time, multiply]
model_input = tokenizer.apply_chat_template(
messages,
tools=tools
)
```
In order for this to work correctly, you should write your functions in the format above, so that they can be parsed
correctly as tools. Specifically, you should follow these rules:
- The function should have a descriptive name
- Every argument must have a type hint
- The function must have a docstring in the standard Google style (in other words, an initial function description
followed by an `Args:` block that describes the arguments, unless the function does not have any arguments.
- Do not include types in the `Args:` block. In other words, write `a: The first number to multiply`, not
`a (int): The first number to multiply`. Type hints should go in the function header instead.
- The function can have a return type and a `Returns:` block in the docstring. However, these are optional
because most tool-use models ignore them.
### Passing tool results to the model
The sample code above is enough to list the available tools for your model, but what happens if it wants to actually use
one? If that happens, you should:
1. Parse the model's output to get the tool name(s) and arguments.
2. Add the model's tool call(s) to the conversation.
3. Call the corresponding function(s) with those arguments.
4. Add the result(s) to the conversation
### A complete tool use example
Let's walk through a tool use example, step by step. For this example, we will use an 8B `Hermes-2-Pro` model,
as it is one of the highest-performing tool-use models in its size category at the time of writing. If you have the
memory, you can consider using a larger model instead like [Command-R](https://huggingface.co/CohereForAI/c4ai-command-r-v01)
or [Mixtral-8x22B](https://huggingface.co/mistralai/Mixtral-8x22B-Instruct-v0.1), both of which also support tool use
and offer even stronger performance.
First, let's load our model and tokenizer:
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
checkpoint = "NousResearch/Hermes-2-Pro-Llama-3-8B"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint, torch_dtype=torch.bfloat16, device_map="auto")
```
Next, let's define a list of tools:
```python
def get_current_temperature(location: str, unit: str) -> float:
"""
Get the current temperature at a location.
Args:
location: The location to get the temperature for, in the format "City, Country"
unit: The unit to return the temperature in. (choices: ["celsius", "fahrenheit"])
Returns:
The current temperature at the specified location in the specified units, as a float.
"""
return 22. # A real function should probably actually get the temperature!
def get_current_wind_speed(location: str) -> float:
"""
Get the current wind speed in km/h at a given location.
Args:
location: The location to get the temperature for, in the format "City, Country"
Returns:
The current wind speed at the given location in km/h, as a float.
"""
return 6. # A real function should probably actually get the wind speed!
tools = [get_current_temperature, get_current_wind_speed]
```
Now, let's set up a conversation for our bot:
```python
messages = [
{"role": "system", "content": "You are a bot that responds to weather queries. You should reply with the unit used in the queried location."},
{"role": "user", "content": "Hey, what's the temperature in Paris right now?"}
]
```
Now, let's apply the chat template and generate a response:
```python
inputs = tokenizer.apply_chat_template(messages, tools=tools, add_generation_prompt=True, return_dict=True, return_tensors="pt")
inputs = {k: v.to(model.device) for k, v in inputs.items()}
out = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(out[0][len(inputs["input_ids"][0]):]))
```
And we get:
```text
<tool_call>
{"arguments": {"location": "Paris, France", "unit": "celsius"}, "name": "get_current_temperature"}
</tool_call><|im_end|>
```
The model has called the function with valid arguments, in the format requested by the function docstring. It has
inferred that we're most likely referring to the Paris in France, and it remembered that, as the home of SI units,
the temperature in France should certainly be displayed in Celsius.
<Tip>
The output format above is specific to the `Hermes-2-Pro` model we're using in this example. Other models may emit different
tool call formats, and you may need to do some manual parsing at this step. For example, `Llama-3.1` models will emit
slightly different JSON, with `parameters` instead of `arguments`. Regardless of the format the model outputs, you
should add the tool call to the conversation in the format below, with `tool_calls`, `function` and `arguments` keys.
</Tip>
Next, let's append the model's tool call to the conversation.
```python
tool_call = {"name": "get_current_temperature", "arguments": {"location": "Paris, France", "unit": "celsius"}}
messages.append({"role": "assistant", "tool_calls": [{"type": "function", "function": tool_call}]})
```
Now that we've added the tool call to the conversation, we can call the function and append the result to the
conversation. Since we're just using a dummy function for this example that always returns 22.0, we can just append
that result directly.
```python
messages.append({"role": "tool", "name": "get_current_temperature", "content": "22.0"})
```
<Tip>
Some model architectures, notably Mistral/Mixtral, also require a `tool_call_id` here, which should be
9 randomly-generated alphanumeric characters, and assigned to the `id` key of the tool call
dictionary. The same key should also be assigned to the `tool_call_id` key of the tool response dictionary below, so
that tool calls can be matched to tool responses. So, for Mistral/Mixtral models, the code above would be:
```python
tool_call_id = "9Ae3bDc2F" # Random ID, 9 alphanumeric characters
tool_call = {"name": "get_current_temperature", "arguments": {"location": "Paris, France", "unit": "celsius"}}
messages.append({"role": "assistant", "tool_calls": [{"type": "function", "id": tool_call_id, "function": tool_call}]})
```
and
```python
messages.append({"role": "tool", "tool_call_id": tool_call_id, "name": "get_current_temperature", "content": "22.0"})
```
</Tip>
Finally, let's let the assistant read the function outputs and continue chatting with the user:
```python
inputs = tokenizer.apply_chat_template(messages, tools=tools, add_generation_prompt=True, return_dict=True, return_tensors="pt")
inputs = {k: v.to(model.device) for k, v in inputs.items()}
out = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(out[0][len(inputs["input_ids"][0]):]))
```
And we get:
```text
The current temperature in Paris, France is 22.0 ° Celsius.<|im_end|>
```
Although this was a simple demo with dummy tools and a single call, the same technique works with
multiple real tools and longer conversations. This can be a powerful way to extend the capabilities of conversational
agents with real-time information, computational tools like calculators, or access to large databases.
### Understanding tool schemas
Each function you pass to the `tools` argument of `apply_chat_template` is converted into a
[JSON schema](https://json-schema.org/learn/getting-started-step-by-step). These schemas
are then passed to the model chat template. In other words, tool-use models do not see your functions directly, and they
never see the actual code inside them. What they care about is the function **definitions** and the **arguments** they
need to pass to them - they care about what the tools do and how to use them, not how they work! It is up to you
to read their outputs, detect if they have requested to use a tool, pass their arguments to the tool function, and
return the response in the chat.
Generating JSON schemas to pass to the template should be automatic and invisible as long as your functions
follow the specification above, but if you encounter problems, or you simply want more control over the conversion,
you can handle the conversion manually. Here is an example of a manual schema conversion.
```python
from transformers.utils import get_json_schema
def multiply(a: float, b: float):
"""
A function that multiplies two numbers
Args:
a: The first number to multiply
b: The second number to multiply
"""
return a * b
schema = get_json_schema(multiply)
print(schema)
```
This will yield:
```json
{
"type": "function",
"function": {
"name": "multiply",
"description": "A function that multiplies two numbers",
"parameters": {
"type": "object",
"properties": {
"a": {
"type": "number",
"description": "The first number to multiply"
},
"b": {
"type": "number",
"description": "The second number to multiply"
}
},
"required": ["a", "b"]
}
}
}
```
If you wish, you can edit these schemas, or even write them from scratch yourself without using `get_json_schema` at
all. JSON schemas can be passed directly to the `tools` argument of
`apply_chat_template` - this gives you a lot of power to define precise schemas for more complex functions. Be careful,
though - the more complex your schemas, the more likely the model is to get confused when dealing with them! We
recommend simple function signatures where possible, keeping arguments (and especially complex, nested arguments)
to a minimum.
Here is an example of defining schemas by hand, and passing them directly to `apply_chat_template`:
```python
# A simple function that takes no arguments
current_time = {
"type": "function",
"function": {
"name": "current_time",
"description": "Get the current local time as a string.",
"parameters": {
'type': 'object',
'properties': {}
}
}
}
# A more complete function that takes two numerical arguments
multiply = {
'type': 'function',
'function': {
'name': 'multiply',
'description': 'A function that multiplies two numbers',
'parameters': {
'type': 'object',
'properties': {
'a': {
'type': 'number',
'description': 'The first number to multiply'
},
'b': {
'type': 'number', 'description': 'The second number to multiply'
}
},
'required': ['a', 'b']
}
}
}
model_input = tokenizer.apply_chat_template(
messages,
tools = [current_time, multiply]
)
```
## Advanced: Retrieval-augmented generation
"Retrieval-augmented generation" or "RAG" LLMs can search a corpus of documents for information before responding
to a query. This allows models to vastly expand their knowledge base beyond their limited context size. Our
recommendation for RAG models is that their template
should accept a `documents` argument. This should be a list of documents, where each "document"
is a single dict with `title` and `contents` keys, both of which are strings. Because this format is much simpler
than the JSON schemas used for tools, no helper functions are necessary.
Here's an example of a RAG template in action:
```python
document1 = {
"title": "The Moon: Our Age-Old Foe",
"contents": "Man has always dreamed of destroying the moon. In this essay, I shall..."
}
document2 = {
"title": "The Sun: Our Age-Old Friend",
"contents": "Although often underappreciated, the sun provides several notable benefits..."
}
model_input = tokenizer.apply_chat_template(
messages,
documents=[document1, document2]
)
```
## Advanced: How do chat templates work?
The chat template for a model is stored on the `tokenizer.chat_template` attribute. If no chat template is set, the
default template for that model class is used instead. Let's take a look at the template for `BlenderBot`:
```python
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
>>> tokenizer.chat_template
"{% for message in messages %}{% if message['role'] == 'user' %}{{ ' ' }}{% endif %}{{ message['content'] }}{% if not loop.last %}{{ ' ' }}{% endif %}{% endfor %}{{ eos_token }}"
```
That's kind of intimidating. Let's clean it up a little to make it more readable. In the process, though, we also make
sure that the newlines and indentation we add don't end up being included in the template output - see the tip on
[trimming whitespace](#trimming-whitespace) below!
```
{%- for message in messages %}
{%- if message['role'] == 'user' %}
{{- ' ' }}
{%- endif %}
{{- message['content'] }}
{%- if not loop.last %}
{{- ' ' }}
{%- endif %}
{%- endfor %}
{{- eos_token }}
```
If you've never seen one of these before, this is a [Jinja template](https://jinja.palletsprojects.com/en/3.1.x/templates/).
Jinja is a templating language that allows you to write simple code that generates text. In many ways, the code and
syntax resembles Python. In pure Python, this template would look something like this:
```python
for idx, message in enumerate(messages):
if message['role'] == 'user':
print(' ')
print(message['content'])
if not idx == len(messages) - 1: # Check for the last message in the conversation
print(' ')
print(eos_token)
```
Effectively, the template does three things:
1. For each message, if the message is a user message, add a blank space before it, otherwise print nothing.
2. Add the message content
3. If the message is not the last message, add two spaces after it. After the final message, print the EOS token.
This is a pretty simple template - it doesn't add any control tokens, and it doesn't support "system" messages, which
are a common way to give the model directives about how it should behave in the subsequent conversation.
But Jinja gives you a lot of flexibility to do those things! Let's see a Jinja template that can format inputs
similarly to the way LLaMA formats them (note that the real LLaMA template includes handling for default system
messages and slightly different system message handling in general - don't use this one in your actual code!)
```
{%- for message in messages %}
{%- if message['role'] == 'user' %}
{{- bos_token + '[INST] ' + message['content'] + ' [/INST]' }}
{%- elif message['role'] == 'system' %}
{{- '<<SYS>>\\n' + message['content'] + '\\n<</SYS>>\\n\\n' }}
{%- elif message['role'] == 'assistant' %}
{{- ' ' + message['content'] + ' ' + eos_token }}
{%- endif %}
{%- endfor %}
```
Hopefully if you stare at this for a little bit you can see what this template is doing - it adds specific tokens based
on the "role" of each message, which represents who sent it. User, assistant and system messages are clearly
distinguishable to the model because of the tokens they're wrapped in.
## Advanced: Adding and editing chat templates
### How do I create a chat template?
Simple, just write a jinja template and set `tokenizer.chat_template`. You may find it easier to start with an
existing template from another model and simply edit it for your needs! For example, we could take the LLaMA template
above and add "[ASST]" and "[/ASST]" to assistant messages:
```
{%- for message in messages %}
{%- if message['role'] == 'user' %}
{{- bos_token + '[INST] ' + message['content'].strip() + ' [/INST]' }}
{%- elif message['role'] == 'system' %}
{{- '<<SYS>>\\n' + message['content'].strip() + '\\n<</SYS>>\\n\\n' }}
{%- elif message['role'] == 'assistant' %}
{{- '[ASST] ' + message['content'] + ' [/ASST]' + eos_token }}
{%- endif %}
{%- endfor %}
```
Now, simply set the `tokenizer.chat_template` attribute. Next time you use [`~PreTrainedTokenizer.apply_chat_template`], it will
use your new template! This attribute will be saved in the `tokenizer_config.json` file, so you can use
[`~utils.PushToHubMixin.push_to_hub`] to upload your new template to the Hub and make sure everyone's using the right
template for your model!
```python
template = tokenizer.chat_template
template = template.replace("SYS", "SYSTEM") # Change the system token
tokenizer.chat_template = template # Set the new template
tokenizer.push_to_hub("model_name") # Upload your new template to the Hub!
```
The method [`~PreTrainedTokenizer.apply_chat_template`] which uses your chat template is called by the [`TextGenerationPipeline`] class, so
once you set the correct chat template, your model will automatically become compatible with [`TextGenerationPipeline`].
<Tip>
If you're fine-tuning a model for chat, in addition to setting a chat template, you should probably add any new chat
control tokens as special tokens in the tokenizer. Special tokens are never split,
ensuring that your control tokens are always handled as single tokens rather than being tokenized in pieces. You
should also set the tokenizer's `eos_token` attribute to the token that marks the end of assistant generations in your
template. This will ensure that text generation tools can correctly figure out when to stop generating text.
</Tip>
### Why do some models have multiple templates?
Some models use different templates for different use cases. For example, they might use one template for normal chat
and another for tool-use, or retrieval-augmented generation. In these cases, `tokenizer.chat_template` is a dictionary.
This can cause some confusion, and where possible, we recommend using a single template for all use-cases. You can use
Jinja statements like `if tools is defined` and `{% macro %}` definitions to easily wrap multiple code paths in a
single template.
When a tokenizer has multiple templates, `tokenizer.chat_template` will be a `dict`, where each key is the name
of a template. The `apply_chat_template` method has special handling for certain template names: Specifically, it will
look for a template named `default` in most cases, and will raise an error if it can't find one. However, if a template
named `tool_use` exists when the user has passed a `tools` argument, it will use that instead. To access templates
with other names, pass the name of the template you want to the `chat_template` argument of
`apply_chat_template()`.
We find that this can be a bit confusing for users, though - so if you're writing a template yourself, we recommend
trying to put it all in a single template where possible!
### What template should I use?
When setting the template for a model that's already been trained for chat, you should ensure that the template
exactly matches the message formatting that the model saw during training, or else you will probably experience
performance degradation. This is true even if you're training the model further - you will probably get the best
performance if you keep the chat tokens constant. This is very analogous to tokenization - you generally get the
best performance for inference or fine-tuning when you precisely match the tokenization used during training.
If you're training a model from scratch, or fine-tuning a base language model for chat, on the other hand,
you have a lot of freedom to choose an appropriate template! LLMs are smart enough to learn to handle lots of different
input formats. One popular choice is the `ChatML` format, and this is a good, flexible choice for many use-cases.
It looks like this:
```
{%- for message in messages %}
{{- '<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n' }}
{%- endfor %}
```
If you like this one, here it is in one-liner form, ready to copy into your code. The one-liner also includes
handy support for [generation prompts](#what-are-generation-prompts), but note that it doesn't add BOS or EOS tokens!
If your model expects those, they won't be added automatically by `apply_chat_template` - in other words, the
text will be tokenized with `add_special_tokens=False`. This is to avoid potential conflicts between the template and
the `add_special_tokens` logic. If your model expects special tokens, make sure to add them to the template!
```python
tokenizer.chat_template = "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}"
```
This template wraps each message in `<|im_start|>` and `<|im_end|>` tokens, and simply writes the role as a string, which
allows for flexibility in the roles you train with. The output looks like this:
```text
<|im_start|>system
You are a helpful chatbot that will do its best not to say anything so stupid that people tweet about it.<|im_end|>
<|im_start|>user
How are you?<|im_end|>
<|im_start|>assistant
I'm doing great!<|im_end|>
```
The "user", "system" and "assistant" roles are the standard for chat, and we recommend using them when it makes sense,
particularly if you want your model to operate well with [`TextGenerationPipeline`]. However, you are not limited
to these roles - templating is extremely flexible, and any string can be a role.
### I want to add some chat templates! How should I get started?
If you have any chat models, you should set their `tokenizer.chat_template` attribute and test it using
[`~PreTrainedTokenizer.apply_chat_template`], then push the updated tokenizer to the Hub. This applies even if you're
not the model owner - if you're using a model with an empty chat template, or one that's still using the default class
template, please open a [pull request](https://huggingface.co/docs/hub/repositories-pull-requests-discussions) to the model repository so that this attribute can be set properly!
Once the attribute is set, that's it, you're done! `tokenizer.apply_chat_template` will now work correctly for that
model, which means it is also automatically supported in places like `TextGenerationPipeline`!
By ensuring that models have this attribute, we can make sure that the whole community gets to use the full power of
open-source models. Formatting mismatches have been haunting the field and silently harming performance for too long -
it's time to put an end to them!
## Advanced: Template writing tips
<Tip>
The easiest way to get started with writing Jinja templates is to take a look at some existing ones. You can use
`print(tokenizer.chat_template)` for any chat model to see what template it's using. In general, models that support tool use have
much more complex templates than other models - so when you're just getting started, they're probably a bad example
to learn from! You can also take a look at the
[Jinja documentation](https://jinja.palletsprojects.com/en/3.1.x/templates/#synopsis) for details
of general Jinja formatting and syntax.
</Tip>
Jinja templates in `transformers` are identical to Jinja templates elsewhere. The main thing to know is that
the conversation history will be accessible inside your template as a variable called `messages`.
You will be able to access `messages` in your template just like you can in Python, which means you can loop over
it with `{% for message in messages %}` or access individual messages with `{{ messages[0] }}`, for example.
You can also use the following tips to write clean, efficient Jinja templates:
### Trimming whitespace
By default, Jinja will print any whitespace that comes before or after a block. This can be a problem for chat
templates, which generally want to be very precise with whitespace! To avoid this, we strongly recommend writing
your templates like this:
```
{%- for message in messages %}
{{- message['role'] + message['content'] }}
{%- endfor %}
```
rather than like this:
```
{% for message in messages %}
{{ message['role'] + message['content'] }}
{% endfor %}
```
Adding `-` will strip any whitespace that comes before the block. The second example looks innocent, but the newline
and indentation may end up being included in the output, which is probably not what you want!
### Special variables
Inside your template, you will have access several special variables. The most important of these is `messages`,
which contains the chat history as a list of message dicts. However, there are several others. Not every
variable will be used in every template. The most common other variables are:
- `tools` contains a list of tools in JSON schema format. Will be `None` or undefined if no tools are passed.
- `documents` contains a list of documents in the format `{"title": "Title", "contents": "Contents"}`, used for retrieval-augmented generation. Will be `None` or undefined if no documents are passed.
- `add_generation_prompt` is a bool that is `True` if the user has requested a generation prompt, and `False` otherwise. If this is set, your template should add the header for an assistant message to the end of the conversation. If your model doesn't have a specific header for assistant messages, you can ignore this flag.
- **Special tokens** like `bos_token` and `eos_token`. These are extracted from `tokenizer.special_tokens_map`. The exact tokens available inside each template will differ depending on the parent tokenizer.
<Tip>
You can actually pass any `kwarg` to `apply_chat_template`, and it will be accessible inside the template as a variable. In general,
we recommend trying to stick to the core variables above, as it will make your model harder to use if users have
to write custom code to pass model-specific `kwargs`. However, we're aware that this field moves quickly, so if you
have a new use-case that doesn't fit in the core API, feel free to use a new `kwarg` for it! If a new `kwarg`
becomes common we may promote it into the core API and create a standard, documented format for it.
</Tip>
### Callable functions
There is also a short list of callable functions available to you inside your templates. These are:
- `raise_exception(msg)`: Raises a `TemplateException`. This is useful for debugging, and for telling users when they're
doing something that your template doesn't support.
- `strftime_now(format_str)`: Equivalent to `datetime.now().strftime(format_str)` in Python. This is used for getting
the current date/time in a specific format, which is sometimes included in system messages.
### Compatibility with non-Python Jinja
There are multiple implementations of Jinja in various languages. They generally have the same syntax,
but a key difference is that when you're writing a template in Python you can use Python methods, such as
`.lower()` on strings or `.items()` on dicts. This will break if someone tries to use your template on a non-Python
implementation of Jinja. Non-Python implementations are particularly common in deployment environments, where JS
and Rust are very popular.
Don't panic, though! There are a few easy changes you can make to your templates to ensure they're compatible across
all implementations of Jinja:
- Replace Python methods with Jinja filters. These usually have the same name, for example `string.lower()` becomes
`string|lower`, and `dict.items()` becomes `dict|items`. One notable change is that `string.strip()` becomes `string|trim`.
See the [list of built-in filters](https://jinja.palletsprojects.com/en/3.1.x/templates/#builtin-filters)
in the Jinja documentation for more.
- Replace `True`, `False` and `None`, which are Python-specific, with `true`, `false` and `none`.
- Directly rendering a dict or list may give different results in other implementations (for example, string entries
might change from single-quoted to double-quoted). Adding the `tojson` filter can help to ensure consistency here.
### Writing and debugging larger templates
When this feature was introduced, most templates were quite small, the Jinja equivalent of a "one-liner" script.
However, with new models and features like tool-use and RAG, some templates can be 100 lines long or more. When
writing templates like these, it's a good idea to write them in a separate file, using a text editor. You can easily
extract a chat template to a file:
```python
open("template.jinja", "w").write(tokenizer.chat_template)
```
Or load the edited template back into the tokenizer:
```python
tokenizer.chat_template = open("template.jinja").read()
```
As an added bonus, when you write a long, multi-line template in a separate file, line numbers in that file will
exactly correspond to line numbers in template parsing or execution errors. This will make it much easier to
identify the source of issues.
|
transformers/docs/source/en/chat_templating.md/0
|
{
"file_path": "transformers/docs/source/en/chat_templating.md",
"repo_id": "transformers",
"token_count": 11885
}
| 269
|
<!--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.
-->
# Utilities for `FeatureExtractors`
This page lists all the utility functions that can be used by the audio [`FeatureExtractor`] in order to compute special features from a raw audio using common algorithms such as *Short Time Fourier Transform* or *log mel spectrogram*.
Most of those are only useful if you are studying the code of the audio processors in the library.
## Audio Transformations
[[autodoc]] audio_utils.hertz_to_mel
[[autodoc]] audio_utils.mel_to_hertz
[[autodoc]] audio_utils.mel_filter_bank
[[autodoc]] audio_utils.optimal_fft_length
[[autodoc]] audio_utils.window_function
[[autodoc]] audio_utils.spectrogram
[[autodoc]] audio_utils.power_to_db
[[autodoc]] audio_utils.amplitude_to_db
|
transformers/docs/source/en/internal/audio_utils.md/0
|
{
"file_path": "transformers/docs/source/en/internal/audio_utils.md",
"repo_id": "transformers",
"token_count": 405
}
| 270
|
<!--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.
-->
# Trainer
The [`Trainer`] class provides an API for feature-complete training in PyTorch, and it supports distributed training on multiple GPUs/TPUs, mixed precision for [NVIDIA GPUs](https://nvidia.github.io/apex/), [AMD GPUs](https://rocm.docs.amd.com/en/latest/rocm.html), and [`torch.amp`](https://pytorch.org/docs/stable/amp.html) for PyTorch. [`Trainer`] goes hand-in-hand with the [`TrainingArguments`] class, which offers a wide range of options to customize how a model is trained. Together, these two classes provide a complete training API.
[`Seq2SeqTrainer`] and [`Seq2SeqTrainingArguments`] inherit from the [`Trainer`] and [`TrainingArgument`] classes and they're adapted for training models for sequence-to-sequence tasks such as summarization or translation.
<Tip warning={true}>
The [`Trainer`] class is optimized for 🤗 Transformers models and can have surprising behaviors
when used with other models. When using it with your own model, make sure:
- your model always return tuples or subclasses of [`~utils.ModelOutput`]
- your model can compute the loss if a `labels` argument is provided and that loss is returned as the first
element of the tuple (if your model returns tuples)
- your model can accept multiple label arguments (use `label_names` in [`TrainingArguments`] to indicate their name to the [`Trainer`]) but none of them should be named `"label"`
</Tip>
## Trainer[[api-reference]]
[[autodoc]] Trainer
- all
## Seq2SeqTrainer
[[autodoc]] Seq2SeqTrainer
- evaluate
- predict
## TrainingArguments
[[autodoc]] TrainingArguments
- all
## Seq2SeqTrainingArguments
[[autodoc]] Seq2SeqTrainingArguments
- all
|
transformers/docs/source/en/main_classes/trainer.md/0
|
{
"file_path": "transformers/docs/source/en/main_classes/trainer.md",
"repo_id": "transformers",
"token_count": 689
}
| 271
|
<!--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.
-->
# BigBird
## Overview
The BigBird model was proposed in [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by
Zaheer, Manzil and Guruganesh, Guru and Dubey, Kumar Avinava and Ainslie, Joshua and Alberti, Chris and Ontanon,
Santiago and Pham, Philip and Ravula, Anirudh and Wang, Qifan and Yang, Li and others. BigBird, is a sparse-attention
based transformer which extends Transformer based models, such as BERT to much longer sequences. In addition to sparse
attention, BigBird also applies global attention as well as random attention to the input sequence. Theoretically, it
has been shown that applying sparse, global, and random attention approximates full attention, while being
computationally much more efficient for longer sequences. As a consequence of the capability to handle longer context,
BigBird has shown improved performance on various long document NLP tasks, such as question answering and
summarization, compared to BERT or RoBERTa.
The abstract from the paper is the following:
*Transformers-based models, such as BERT, have been one of the most successful deep learning models for NLP.
Unfortunately, one of their core limitations is the quadratic dependency (mainly in terms of memory) on the sequence
length due to their full attention mechanism. To remedy this, we propose, BigBird, a sparse attention mechanism that
reduces this quadratic dependency to linear. We show that BigBird is a universal approximator of sequence functions and
is Turing complete, thereby preserving these properties of the quadratic, full attention model. Along the way, our
theoretical analysis reveals some of the benefits of having O(1) global tokens (such as CLS), that attend to the entire
sequence as part of the sparse attention mechanism. The proposed sparse attention can handle sequences of length up to
8x of what was previously possible using similar hardware. As a consequence of the capability to handle longer context,
BigBird drastically improves performance on various NLP tasks such as question answering and summarization. We also
propose novel applications to genomics data.*
This model was contributed by [vasudevgupta](https://huggingface.co/vasudevgupta). The original code can be found
[here](https://github.com/google-research/bigbird).
## Usage tips
- For an in-detail explanation on how BigBird's attention works, see [this blog post](https://huggingface.co/blog/big-bird).
- BigBird comes with 2 implementations: **original_full** & **block_sparse**. For the sequence length < 1024, using
**original_full** is advised as there is no benefit in using **block_sparse** attention.
- The code currently uses window size of 3 blocks and 2 global blocks.
- Sequence length must be divisible by block size.
- Current implementation supports only **ITC**.
- Current implementation doesn't support **num_random_blocks = 0**
- BigBird is a model with absolute position embeddings so it's usually advised to pad the inputs on the right rather than
the left.
## Resources
- [Text classification task guide](../tasks/sequence_classification)
- [Token classification task guide](../tasks/token_classification)
- [Question answering task guide](../tasks/question_answering)
- [Causal language modeling task guide](../tasks/language_modeling)
- [Masked language modeling task guide](../tasks/masked_language_modeling)
- [Multiple choice task guide](../tasks/multiple_choice)
## BigBirdConfig
[[autodoc]] BigBirdConfig
## BigBirdTokenizer
[[autodoc]] BigBirdTokenizer
- build_inputs_with_special_tokens
- get_special_tokens_mask
- create_token_type_ids_from_sequences
- save_vocabulary
## BigBirdTokenizerFast
[[autodoc]] BigBirdTokenizerFast
## BigBird specific outputs
[[autodoc]] models.big_bird.modeling_big_bird.BigBirdForPreTrainingOutput
<frameworkcontent>
<pt>
## BigBirdModel
[[autodoc]] BigBirdModel
- forward
## BigBirdForPreTraining
[[autodoc]] BigBirdForPreTraining
- forward
## BigBirdForCausalLM
[[autodoc]] BigBirdForCausalLM
- forward
## BigBirdForMaskedLM
[[autodoc]] BigBirdForMaskedLM
- forward
## BigBirdForSequenceClassification
[[autodoc]] BigBirdForSequenceClassification
- forward
## BigBirdForMultipleChoice
[[autodoc]] BigBirdForMultipleChoice
- forward
## BigBirdForTokenClassification
[[autodoc]] BigBirdForTokenClassification
- forward
## BigBirdForQuestionAnswering
[[autodoc]] BigBirdForQuestionAnswering
- forward
</pt>
<jax>
## FlaxBigBirdModel
[[autodoc]] FlaxBigBirdModel
- __call__
## FlaxBigBirdForPreTraining
[[autodoc]] FlaxBigBirdForPreTraining
- __call__
## FlaxBigBirdForCausalLM
[[autodoc]] FlaxBigBirdForCausalLM
- __call__
## FlaxBigBirdForMaskedLM
[[autodoc]] FlaxBigBirdForMaskedLM
- __call__
## FlaxBigBirdForSequenceClassification
[[autodoc]] FlaxBigBirdForSequenceClassification
- __call__
## FlaxBigBirdForMultipleChoice
[[autodoc]] FlaxBigBirdForMultipleChoice
- __call__
## FlaxBigBirdForTokenClassification
[[autodoc]] FlaxBigBirdForTokenClassification
- __call__
## FlaxBigBirdForQuestionAnswering
[[autodoc]] FlaxBigBirdForQuestionAnswering
- __call__
</jax>
</frameworkcontent>
|
transformers/docs/source/en/model_doc/big_bird.md/0
|
{
"file_path": "transformers/docs/source/en/model_doc/big_bird.md",
"repo_id": "transformers",
"token_count": 1682
}
| 272
|
<!--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.
-->
# DistilBERT
<div class="flex flex-wrap space-x-1">
<a href="https://huggingface.co/models?filter=distilbert">
<img alt="Models" src="https://img.shields.io/badge/All_model_pages-distilbert-blueviolet">
</a>
<a href="https://huggingface.co/spaces/docs-demos/distilbert-base-uncased">
<img alt="Spaces" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue">
</a>
<a href="https://huggingface.co/papers/1910.01108">
<img alt="Paper page" src="https://img.shields.io/badge/Paper%20page-1910.01108-green">
</a>
</div>
## Overview
The DistilBERT model was proposed in the blog post [Smaller, faster, cheaper, lighter: Introducing DistilBERT, a
distilled version of BERT](https://medium.com/huggingface/distilbert-8cf3380435b5), and the paper [DistilBERT, a
distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108). DistilBERT is a
small, fast, cheap and light Transformer model trained by distilling BERT base. It has 40% less parameters than
*google-bert/bert-base-uncased*, runs 60% faster while preserving over 95% of BERT's performances as measured on the GLUE language
understanding benchmark.
The abstract from the paper is the following:
*As Transfer Learning from large-scale pre-trained models becomes more prevalent in Natural Language Processing (NLP),
operating these large models in on-the-edge and/or under constrained computational training or inference budgets
remains challenging. In this work, we propose a method to pre-train a smaller general-purpose language representation
model, called DistilBERT, which can then be fine-tuned with good performances on a wide range of tasks like its larger
counterparts. While most prior work investigated the use of distillation for building task-specific models, we leverage
knowledge distillation during the pretraining phase and show that it is possible to reduce the size of a BERT model by
40%, while retaining 97% of its language understanding capabilities and being 60% faster. To leverage the inductive
biases learned by larger models during pretraining, we introduce a triple loss combining language modeling,
distillation and cosine-distance losses. Our smaller, faster and lighter model is cheaper to pre-train and we
demonstrate its capabilities for on-device computations in a proof-of-concept experiment and a comparative on-device
study.*
This model was contributed by [victorsanh](https://huggingface.co/victorsanh). This model jax version was
contributed by [kamalkraj](https://huggingface.co/kamalkraj). The original code can be found [here](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation).
## Usage tips
- DistilBERT doesn't have `token_type_ids`, you don't need to indicate which token belongs to which segment. Just
separate your segments with the separation token `tokenizer.sep_token` (or `[SEP]`).
- DistilBERT doesn't have options to select the input positions (`position_ids` input). This could be added if
necessary though, just let us know if you need this option.
- Same as BERT but smaller. Trained by distillation of the pretrained BERT model, meaning it’s been trained to predict the same probabilities as the larger model. The actual objective is a combination of:
* finding the same probabilities as the teacher model
* predicting the masked tokens correctly (but no next-sentence objective)
* a cosine similarity between the hidden states of the student and the teacher model
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with DistilBERT. 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.
<PipelineTag pipeline="text-classification"/>
- A blog post on [Getting Started with Sentiment Analysis using Python](https://huggingface.co/blog/sentiment-analysis-python) with DistilBERT.
- A blog post on how to [train DistilBERT with Blurr for sequence classification](https://huggingface.co/blog/fastai).
- A blog post on how to use [Ray to tune DistilBERT hyperparameters](https://huggingface.co/blog/ray-tune).
- A blog post on how to [train DistilBERT with Hugging Face and Amazon SageMaker](https://huggingface.co/blog/the-partnership-amazon-sagemaker-and-hugging-face).
- A notebook on how to [finetune DistilBERT for multi-label classification](https://colab.research.google.com/github/DhavalTaunk08/Transformers_scripts/blob/master/Transformers_multilabel_distilbert.ipynb). 🌎
- A notebook on how to [finetune DistilBERT for multiclass classification with PyTorch](https://colab.research.google.com/github/abhimishra91/transformers-tutorials/blob/master/transformers_multiclass_classification.ipynb). 🌎
- A notebook on how to [finetune DistilBERT for text classification in TensorFlow](https://colab.research.google.com/github/peterbayerle/huggingface_notebook/blob/main/distilbert_tf.ipynb). 🌎
- [`DistilBertForSequenceClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/text-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/text_classification.ipynb).
- [`TFDistilBertForSequenceClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/text-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/text_classification-tf.ipynb).
- [`FlaxDistilBertForSequenceClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/flax/text-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/text_classification_flax.ipynb).
- [Text classification task guide](../tasks/sequence_classification)
<PipelineTag pipeline="token-classification"/>
- [`DistilBertForTokenClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/token-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/token_classification.ipynb).
- [`TFDistilBertForTokenClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/token-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/token_classification-tf.ipynb).
- [`FlaxDistilBertForTokenClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/flax/token-classification).
- [Token classification](https://huggingface.co/course/chapter7/2?fw=pt) chapter of the 🤗 Hugging Face Course.
- [Token classification task guide](../tasks/token_classification)
<PipelineTag pipeline="fill-mask"/>
- [`DistilBertForMaskedLM`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/language-modeling#robertabertdistilbert-and-masked-language-modeling) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling.ipynb).
- [`TFDistilBertForMaskedLM`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/language-modeling#run_mlmpy) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling-tf.ipynb).
- [`FlaxDistilBertForMaskedLM`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/flax/language-modeling#masked-language-modeling) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/masked_language_modeling_flax.ipynb).
- [Masked language modeling](https://huggingface.co/course/chapter7/3?fw=pt) chapter of the 🤗 Hugging Face Course.
- [Masked language modeling task guide](../tasks/masked_language_modeling)
<PipelineTag pipeline="question-answering"/>
- [`DistilBertForQuestionAnswering`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/question-answering) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/question_answering.ipynb).
- [`TFDistilBertForQuestionAnswering`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/question-answering) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/question_answering-tf.ipynb).
- [`FlaxDistilBertForQuestionAnswering`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/flax/question-answering).
- [Question answering](https://huggingface.co/course/chapter7/7?fw=pt) chapter of the 🤗 Hugging Face Course.
- [Question answering task guide](../tasks/question_answering)
**Multiple choice**
- [`DistilBertForMultipleChoice`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/multiple-choice) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/multiple_choice.ipynb).
- [`TFDistilBertForMultipleChoice`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/multiple-choice) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/multiple_choice-tf.ipynb).
- [Multiple choice task guide](../tasks/multiple_choice)
⚗️ Optimization
- A blog post on how to [quantize DistilBERT with 🤗 Optimum and Intel](https://huggingface.co/blog/intel).
- A blog post on how [Optimizing Transformers for GPUs with 🤗 Optimum](https://www.philschmid.de/optimizing-transformers-with-optimum-gpu).
- A blog post on [Optimizing Transformers with Hugging Face Optimum](https://www.philschmid.de/optimizing-transformers-with-optimum).
⚡️ Inference
- A blog post on how to [Accelerate BERT inference with Hugging Face Transformers and AWS Inferentia](https://huggingface.co/blog/bert-inferentia-sagemaker) with DistilBERT.
- A blog post on [Serverless Inference with Hugging Face's Transformers, DistilBERT and Amazon SageMaker](https://www.philschmid.de/sagemaker-serverless-huggingface-distilbert).
🚀 Deploy
- A blog post on how to [deploy DistilBERT on Google Cloud](https://huggingface.co/blog/how-to-deploy-a-pipeline-to-google-clouds).
- A blog post on how to [deploy DistilBERT with Amazon SageMaker](https://huggingface.co/blog/deploy-hugging-face-models-easily-with-amazon-sagemaker).
- A blog post on how to [Deploy BERT with Hugging Face Transformers, Amazon SageMaker and Terraform module](https://www.philschmid.de/terraform-huggingface-amazon-sagemaker).
## Combining DistilBERT and Flash Attention 2
First, make sure to install the latest version of Flash Attention 2 to include the sliding window attention feature.
```bash
pip install -U flash-attn --no-build-isolation
```
Make also sure that you have a hardware that is compatible with Flash-Attention 2. Read more about it in the official documentation of flash-attn repository. Make also sure to load your model in half-precision (e.g. `torch.float16`)
To load and run a model using Flash Attention 2, refer to the snippet below:
```python
>>> import torch
>>> from transformers import AutoTokenizer, AutoModel
>>> device = "cuda" # the device to load the model onto
>>> tokenizer = AutoTokenizer.from_pretrained('distilbert/distilbert-base-uncased')
>>> model = AutoModel.from_pretrained("distilbert/distilbert-base-uncased", torch_dtype=torch.float16, attn_implementation="flash_attention_2")
>>> text = "Replace me by any text you'd like."
>>> encoded_input = tokenizer(text, return_tensors='pt').to(device)
>>> model.to(device)
>>> output = model(**encoded_input)
```
## DistilBertConfig
[[autodoc]] DistilBertConfig
## DistilBertTokenizer
[[autodoc]] DistilBertTokenizer
## DistilBertTokenizerFast
[[autodoc]] DistilBertTokenizerFast
<frameworkcontent>
<pt>
## DistilBertModel
[[autodoc]] DistilBertModel
- forward
## DistilBertForMaskedLM
[[autodoc]] DistilBertForMaskedLM
- forward
## DistilBertForSequenceClassification
[[autodoc]] DistilBertForSequenceClassification
- forward
## DistilBertForMultipleChoice
[[autodoc]] DistilBertForMultipleChoice
- forward
## DistilBertForTokenClassification
[[autodoc]] DistilBertForTokenClassification
- forward
## DistilBertForQuestionAnswering
[[autodoc]] DistilBertForQuestionAnswering
- forward
</pt>
<tf>
## TFDistilBertModel
[[autodoc]] TFDistilBertModel
- call
## TFDistilBertForMaskedLM
[[autodoc]] TFDistilBertForMaskedLM
- call
## TFDistilBertForSequenceClassification
[[autodoc]] TFDistilBertForSequenceClassification
- call
## TFDistilBertForMultipleChoice
[[autodoc]] TFDistilBertForMultipleChoice
- call
## TFDistilBertForTokenClassification
[[autodoc]] TFDistilBertForTokenClassification
- call
## TFDistilBertForQuestionAnswering
[[autodoc]] TFDistilBertForQuestionAnswering
- call
</tf>
<jax>
## FlaxDistilBertModel
[[autodoc]] FlaxDistilBertModel
- __call__
## FlaxDistilBertForMaskedLM
[[autodoc]] FlaxDistilBertForMaskedLM
- __call__
## FlaxDistilBertForSequenceClassification
[[autodoc]] FlaxDistilBertForSequenceClassification
- __call__
## FlaxDistilBertForMultipleChoice
[[autodoc]] FlaxDistilBertForMultipleChoice
- __call__
## FlaxDistilBertForTokenClassification
[[autodoc]] FlaxDistilBertForTokenClassification
- __call__
## FlaxDistilBertForQuestionAnswering
[[autodoc]] FlaxDistilBertForQuestionAnswering
- __call__
</jax>
</frameworkcontent>
|
transformers/docs/source/en/model_doc/distilbert.md/0
|
{
"file_path": "transformers/docs/source/en/model_doc/distilbert.md",
"repo_id": "transformers",
"token_count": 4631
}
| 273
|
<!--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.
-->
# FLAN-T5
## Overview
FLAN-T5 was released in the paper [Scaling Instruction-Finetuned Language Models](https://arxiv.org/pdf/2210.11416.pdf) - 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": 781
}
| 274
|
<!--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.
-->
# GPT Neo
## Overview
The GPTNeo model was released in the [EleutherAI/gpt-neo](https://github.com/EleutherAI/gpt-neo) repository by Sid
Black, Stella Biderman, Leo Gao, Phil Wang and Connor Leahy. It is a GPT2 like causal language model trained on the
[Pile](https://pile.eleuther.ai/) dataset.
The architecture is similar to GPT2 except that GPT Neo uses local attention in every other layer with a window size of
256 tokens.
This model was contributed by [valhalla](https://huggingface.co/valhalla).
## Usage example
The `generate()` method can be used to generate text using GPT Neo model.
```python
>>> from transformers import GPTNeoForCausalLM, GPT2Tokenizer
>>> model = GPTNeoForCausalLM.from_pretrained("EleutherAI/gpt-neo-1.3B")
>>> tokenizer = GPT2Tokenizer.from_pretrained("EleutherAI/gpt-neo-1.3B")
>>> prompt = (
... "In a shocking finding, scientists discovered a herd of unicorns living in a remote, "
... "previously unexplored valley, in the Andes Mountains. Even more surprising to the "
... "researchers was the fact that the unicorns spoke perfect English."
... )
>>> input_ids = tokenizer(prompt, return_tensors="pt").input_ids
>>> gen_tokens = model.generate(
... input_ids,
... do_sample=True,
... temperature=0.9,
... max_length=100,
... )
>>> gen_text = tokenizer.batch_decode(gen_tokens)[0]
```
## Combining GPT-Neo and Flash Attention 2
First, make sure to install the latest version of Flash Attention 2 to include the sliding window attention feature, and make sure your hardware is compatible with Flash-Attention 2. More details are available [here](https://huggingface.co/docs/transformers/perf_infer_gpu_one#flashattention-2) concerning the installation.
Make sure as well to load your model in half-precision (e.g. `torch.float16`).
To load and run a model using Flash Attention 2, refer to the snippet below:
```python
>>> import torch
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> device = "cuda" # the device to load the model onto
>>> model = AutoModelForCausalLM.from_pretrained("EleutherAI/gpt-neo-2.7B", torch_dtype=torch.float16, attn_implementation="flash_attention_2")
>>> tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neo-2.7B")
>>> prompt = "def hello_world():"
>>> model_inputs = tokenizer([prompt], return_tensors="pt").to(device)
>>> model.to(device)
>>> generated_ids = model.generate(**model_inputs, max_new_tokens=100, do_sample=True)
>>> tokenizer.batch_decode(generated_ids)[0]
"def hello_world():\n >>> run_script("hello.py")\n >>> exit(0)\n<|endoftext|>"
```
### Expected speedups
Below is an expected speedup diagram that compares pure inference time between the native implementation in transformers using `EleutherAI/gpt-neo-2.7B` checkpoint and the Flash Attention 2 version of the model.
Note that for GPT-Neo it is not possible to train / run on very long context as the max [position embeddings](https://huggingface.co/EleutherAI/gpt-neo-2.7B/blob/main/config.json#L58 ) is limited to 2048 - but this is applicable to all gpt-neo models and not specific to FA-2
<div style="text-align: center">
<img src="https://user-images.githubusercontent.com/49240599/272241893-b1c66b75-3a48-4265-bc47-688448568b3d.png">
</div>
## Resources
- [Text classification task guide](../tasks/sequence_classification)
- [Causal language modeling task guide](../tasks/language_modeling)
## GPTNeoConfig
[[autodoc]] GPTNeoConfig
<frameworkcontent>
<pt>
## GPTNeoModel
[[autodoc]] GPTNeoModel
- forward
## GPTNeoForCausalLM
[[autodoc]] GPTNeoForCausalLM
- forward
## GPTNeoForQuestionAnswering
[[autodoc]] GPTNeoForQuestionAnswering
- forward
## GPTNeoForSequenceClassification
[[autodoc]] GPTNeoForSequenceClassification
- forward
## GPTNeoForTokenClassification
[[autodoc]] GPTNeoForTokenClassification
- forward
</pt>
<jax>
## FlaxGPTNeoModel
[[autodoc]] FlaxGPTNeoModel
- __call__
## FlaxGPTNeoForCausalLM
[[autodoc]] FlaxGPTNeoForCausalLM
- __call__
</jax>
</frameworkcontent>
|
transformers/docs/source/en/model_doc/gpt_neo.md/0
|
{
"file_path": "transformers/docs/source/en/model_doc/gpt_neo.md",
"repo_id": "transformers",
"token_count": 1582
}
| 275
|
<!--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.
-->
# Llama3
```py3
import transformers
import torch
model_id = "meta-llama/Meta-Llama-3-8B"
pipeline = transformers.pipeline("text-generation", model=model_id, model_kwargs={"torch_dtype": torch.bfloat16}, device_map="auto")
pipeline("Hey how are you doing today?")
```
## Overview
The Llama3 model was proposed in [Introducing Meta Llama 3: The most capable openly available LLM to date](https://ai.meta.com/blog/meta-llama-3/) by the meta AI team.
The abstract from the blogpost is the following:
*Today, we’re excited to share the first two models of the next generation of Llama, Meta Llama 3, available for broad use. This release features pretrained and instruction-fine-tuned language models with 8B and 70B parameters that can support a broad range of use cases. This next generation of Llama demonstrates state-of-the-art performance on a wide range of industry benchmarks and offers new capabilities, including improved reasoning. We believe these are the best open source models of their class, period. In support of our longstanding open approach, we’re putting Llama 3 in the hands of the community. We want to kickstart the next wave of innovation in AI across the stack—from applications to developer tools to evals to inference optimizations and more. We can’t wait to see what you build and look forward to your feedback.*
Checkout all Llama3 model checkpoints [here](https://huggingface.co/models?search=llama3).
The original code of the authors can be found [here](https://github.com/meta-llama/llama3).
## Usage tips
<Tip warning={true}>
The `Llama3` models were trained using `bfloat16`, but the original inference uses `float16`. The checkpoints uploaded on the Hub use `torch_dtype = 'float16'`, which will be
used by the `AutoModel` API to cast the checkpoints from `torch.float32` to `torch.float16`.
The `dtype` of the online weights is mostly irrelevant unless you are using `torch_dtype="auto"` when initializing a model using `model = AutoModelForCausalLM.from_pretrained("path", torch_dtype = "auto")`. The reason is that the model will first be downloaded ( using the `dtype` of the checkpoints online), then it will be casted to the default `dtype` of `torch` (becomes `torch.float32`), and finally, if there is a `torch_dtype` provided in the config, it will be used.
Training the model in `float16` is not recommended and is known to produce `nan`; as such, the model should be trained in `bfloat16`.
</Tip>
Tips:
- Weights for the Llama3 models can be obtained by filling out [this form](https://ai.meta.com/resources/models-and-libraries/llama-downloads/)
- The architecture is exactly the same as Llama2.
- The tokenizer is a BPE model based on [tiktoken](https://github.com/openai/tiktoken) (vs the one based on sentencepiece implementation for Llama2). The main difference that it ignores BPE merge rules when an input token is part of the vocab. This means that if no merge exist to produce `"hugging"`, instead of having the smallest units, like `["hug","ging"] form 2 tokens, if `"hugging"` is part of the vocab, it will be automatically returned as a token.
- The original model uses `pad_id = -1` which means that there is no padding token. We can't have the same logic, make sure to add a padding token using `tokenizer.add_special_tokens({"pad_token":"<pad>"})` and resize the token embedding accordingly. You should also set the `model.config.pad_token_id`. The `embed_tokens` layer of the model is initialized with `self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.config.padding_idx)`, which makes sure that encoding the padding token will output zeros, so passing it when initializing is recommended.
- The original checkpoint can be converted using the [conversion script](https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/convert_llama_weights_to_hf.py). The script can be called with the following (example) command:
```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 --llama_version 3
```
- After conversion, the model and tokenizer can be loaded via:
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("/output/path")
model = AutoModelForCausalLM.from_pretrained("/output/path")
```
Note that executing the script requires enough CPU RAM to host the whole model in float16 precision (even if the biggest versions
come in several checkpoints they each contain a part of each weight of the model, so we need to load them all in RAM). For the 75B model, it's thus 145GB of RAM needed.
- When using Flash Attention 2 via `attn_implementation="flash_attention_2"`, don't pass `torch_dtype` to the `from_pretrained` class method and use Automatic Mixed-Precision training. When using `Trainer`, it is simply specifying either `fp16` or `bf16` to `True`. Otherwise, make sure you are using `torch.autocast`. This is required because the Flash Attention only support `fp16` and `bf16` data type.
## Resources
A ton of cool resources are already available on the documentation page of [Llama2](./llama2), inviting contributors to add new resources curated for Llama3 here! 🤗
|
transformers/docs/source/en/model_doc/llama3.md/0
|
{
"file_path": "transformers/docs/source/en/model_doc/llama3.md",
"repo_id": "transformers",
"token_count": 1664
}
| 276
|
<!--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.
-->
# MPNet
## Overview
The MPNet model was proposed in [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu.
MPNet adopts a novel pre-training method, named masked and permuted language modeling, to inherit the advantages of
masked language modeling and permuted language modeling for natural language understanding.
The abstract from the paper is the following:
*BERT adopts masked language modeling (MLM) for pre-training and is one of the most successful pre-training models.
Since BERT neglects dependency among predicted tokens, XLNet introduces permuted language modeling (PLM) for
pre-training to address this problem. However, XLNet does not leverage the full position information of a sentence and
thus suffers from position discrepancy between pre-training and fine-tuning. In this paper, we propose MPNet, a novel
pre-training method that inherits the advantages of BERT and XLNet and avoids their limitations. MPNet leverages the
dependency among predicted tokens through permuted language modeling (vs. MLM in BERT), and takes auxiliary position
information as input to make the model see a full sentence and thus reducing the position discrepancy (vs. PLM in
XLNet). We pre-train MPNet on a large-scale dataset (over 160GB text corpora) and fine-tune on a variety of
down-streaming tasks (GLUE, SQuAD, etc). Experimental results show that MPNet outperforms MLM and PLM by a large
margin, and achieves better results on these tasks compared with previous state-of-the-art pre-trained methods (e.g.,
BERT, XLNet, RoBERTa) under the same model setting.*
The original code can be found [here](https://github.com/microsoft/MPNet).
## Usage tips
MPNet doesn't have `token_type_ids`, you don't need to indicate which token belongs to which segment. Just
separate your segments with the separation token `tokenizer.sep_token` (or `[sep]`).
## Resources
- [Text classification task guide](../tasks/sequence_classification)
- [Token classification task guide](../tasks/token_classification)
- [Question answering task guide](../tasks/question_answering)
- [Masked language modeling task guide](../tasks/masked_language_modeling)
- [Multiple choice task guide](../tasks/multiple_choice)
## MPNetConfig
[[autodoc]] MPNetConfig
## MPNetTokenizer
[[autodoc]] MPNetTokenizer
- build_inputs_with_special_tokens
- get_special_tokens_mask
- create_token_type_ids_from_sequences
- save_vocabulary
## MPNetTokenizerFast
[[autodoc]] MPNetTokenizerFast
<frameworkcontent>
<pt>
## MPNetModel
[[autodoc]] MPNetModel
- forward
## MPNetForMaskedLM
[[autodoc]] MPNetForMaskedLM
- forward
## MPNetForSequenceClassification
[[autodoc]] MPNetForSequenceClassification
- forward
## MPNetForMultipleChoice
[[autodoc]] MPNetForMultipleChoice
- forward
## MPNetForTokenClassification
[[autodoc]] MPNetForTokenClassification
- forward
## MPNetForQuestionAnswering
[[autodoc]] MPNetForQuestionAnswering
- forward
</pt>
<tf>
## TFMPNetModel
[[autodoc]] TFMPNetModel
- call
## TFMPNetForMaskedLM
[[autodoc]] TFMPNetForMaskedLM
- call
## TFMPNetForSequenceClassification
[[autodoc]] TFMPNetForSequenceClassification
- call
## TFMPNetForMultipleChoice
[[autodoc]] TFMPNetForMultipleChoice
- call
## TFMPNetForTokenClassification
[[autodoc]] TFMPNetForTokenClassification
- call
## TFMPNetForQuestionAnswering
[[autodoc]] TFMPNetForQuestionAnswering
- call
</tf>
</frameworkcontent>
|
transformers/docs/source/en/model_doc/mpnet.md/0
|
{
"file_path": "transformers/docs/source/en/model_doc/mpnet.md",
"repo_id": "transformers",
"token_count": 1255
}
| 277
|
<!--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.
-->
# Open-Llama
<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: v4.31.0.
You can do so by running the following command: `pip install -U transformers==4.31.0`.
</Tip>
<Tip warning={true}>
This model differs from the [OpenLLaMA models](https://huggingface.co/models?search=openllama) on the Hugging Face Hub, which primarily use the [LLaMA](llama) architecture.
</Tip>
## Overview
The Open-Llama model was proposed in the open source Open-Llama project by community developer s-JoL.
The model is mainly based on LLaMA with some modifications, incorporating memory-efficient attention from Xformers, stable embedding from Bloom, and shared input-output embedding from PaLM.
And the model is pre-trained on both Chinese and English, which gives it better performance on Chinese language tasks.
This model was contributed by [s-JoL](https://huggingface.co/s-JoL).
The original code was released on GitHub by [s-JoL](https://github.com/s-JoL), but is now removed.
## OpenLlamaConfig
[[autodoc]] OpenLlamaConfig
## OpenLlamaModel
[[autodoc]] OpenLlamaModel
- forward
## OpenLlamaForCausalLM
[[autodoc]] OpenLlamaForCausalLM
- forward
## OpenLlamaForSequenceClassification
[[autodoc]] OpenLlamaForSequenceClassification
- forward
|
transformers/docs/source/en/model_doc/open-llama.md/0
|
{
"file_path": "transformers/docs/source/en/model_doc/open-llama.md",
"repo_id": "transformers",
"token_count": 620
}
| 278
|
<!--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.
-->
# PLBart
## Overview
The PLBART model was proposed in [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) by Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang.
This is a BART-like model which can be used to perform code-summarization, code-generation, and code-translation tasks. The pre-trained model `plbart-base` has been trained using multilingual denoising task
on Java, Python and English.
According to the abstract
*Code summarization and generation empower conversion between programming language (PL) and natural language (NL),
while code translation avails the migration of legacy code from one PL to another. This paper introduces PLBART,
a sequence-to-sequence model capable of performing a broad spectrum of program and language understanding and generation tasks.
PLBART is pre-trained on an extensive collection of Java and Python functions and associated NL text via denoising autoencoding.
Experiments on code summarization in the English language, code generation, and code translation in seven programming languages
show that PLBART outperforms or rivals state-of-the-art models. Moreover, experiments on discriminative tasks, e.g., program
repair, clone detection, and vulnerable code detection, demonstrate PLBART's effectiveness in program understanding.
Furthermore, analysis reveals that PLBART learns program syntax, style (e.g., identifier naming convention), logical flow
(e.g., if block inside an else block is equivalent to else if block) that are crucial to program semantics and thus excels
even with limited annotations.*
This model was contributed by [gchhablani](https://huggingface.co/gchhablani). The Authors' code can be found [here](https://github.com/wasiahmad/PLBART).
## Usage examples
PLBart is a multilingual encoder-decoder (sequence-to-sequence) model primarily intended for code-to-text, text-to-code, code-to-code tasks. As the
model is multilingual it expects the sequences in a different format. A special language id token is added in both the
source and target text. The source text format is `X [eos, src_lang_code]` where `X` is the source text. The
target text format is `[tgt_lang_code] X [eos]`. `bos` is never used.
However, for fine-tuning, in some cases no language token is provided in cases where a single language is used. Please refer to [the paper](https://arxiv.org/abs/2103.06333) to learn more about this.
In cases where the language code is needed, the regular [`~PLBartTokenizer.__call__`] will encode source text format
when you pass texts as the first argument or with the keyword argument `text`, and will encode target text format if
it's passed with the `text_target` keyword argument.
### Supervised training
```python
>>> from transformers import PLBartForConditionalGeneration, PLBartTokenizer
>>> tokenizer = PLBartTokenizer.from_pretrained("uclanlp/plbart-base", src_lang="en_XX", tgt_lang="python")
>>> example_python_phrase = "def maximum(a,b,c):NEW_LINE_INDENTreturn max([a,b,c])"
>>> expected_translation_english = "Returns the maximum value of a b c."
>>> inputs = tokenizer(example_python_phrase, text_target=expected_translation_english, return_tensors="pt")
>>> model(**inputs)
```
### Generation
While generating the target text set the `decoder_start_token_id` to the target language id. The following
example shows how to translate Python to English using the `uclanlp/plbart-python-en_XX` model.
```python
>>> from transformers import PLBartForConditionalGeneration, PLBartTokenizer
>>> tokenizer = PLBartTokenizer.from_pretrained("uclanlp/plbart-python-en_XX", src_lang="python", tgt_lang="en_XX")
>>> example_python_phrase = "def maximum(a,b,c):NEW_LINE_INDENTreturn max([a,b,c])"
>>> inputs = tokenizer(example_python_phrase, return_tensors="pt")
>>> model = PLBartForConditionalGeneration.from_pretrained("uclanlp/plbart-python-en_XX")
>>> translated_tokens = model.generate(**inputs, decoder_start_token_id=tokenizer.lang_code_to_id["en_XX"])
>>> tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]
"Returns the maximum value of a b c."
```
## Resources
- [Text classification task guide](../tasks/sequence_classification)
- [Causal language modeling task guide](../tasks/language_modeling)
- [Translation task guide](../tasks/translation)
- [Summarization task guide](../tasks/summarization)
## PLBartConfig
[[autodoc]] PLBartConfig
## PLBartTokenizer
[[autodoc]] PLBartTokenizer
- build_inputs_with_special_tokens
## PLBartModel
[[autodoc]] PLBartModel
- forward
## PLBartForConditionalGeneration
[[autodoc]] PLBartForConditionalGeneration
- forward
## PLBartForSequenceClassification
[[autodoc]] PLBartForSequenceClassification
- forward
## PLBartForCausalLM
[[autodoc]] PLBartForCausalLM
- forward
|
transformers/docs/source/en/model_doc/plbart.md/0
|
{
"file_path": "transformers/docs/source/en/model_doc/plbart.md",
"repo_id": "transformers",
"token_count": 1586
}
| 279
|
<!--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.
-->
# RemBERT
## Overview
The RemBERT model was proposed in [Rethinking Embedding Coupling in Pre-trained Language Models](https://arxiv.org/abs/2010.12821) by Hyung Won Chung, Thibault Févry, Henry Tsai, Melvin Johnson, Sebastian Ruder.
The abstract from the paper is the following:
*We re-evaluate the standard practice of sharing weights between input and output embeddings in state-of-the-art
pre-trained language models. We show that decoupled embeddings provide increased modeling flexibility, allowing us to
significantly improve the efficiency of parameter allocation in the input embedding of multilingual models. By
reallocating the input embedding parameters in the Transformer layers, we achieve dramatically better performance on
standard natural language understanding tasks with the same number of parameters during fine-tuning. We also show that
allocating additional capacity to the output embedding provides benefits to the model that persist through the
fine-tuning stage even though the output embedding is discarded after pre-training. Our analysis shows that larger
output embeddings prevent the model's last layers from overspecializing to the pre-training task and encourage
Transformer representations to be more general and more transferable to other tasks and languages. Harnessing these
findings, we are able to train models that achieve strong performance on the XTREME benchmark without increasing the
number of parameters at the fine-tuning stage.*
## Usage tips
For fine-tuning, RemBERT can be thought of as a bigger version of mBERT with an ALBERT-like factorization of the
embedding layer. The embeddings are not tied in pre-training, in contrast with BERT, which enables smaller input
embeddings (preserved during fine-tuning) and bigger output embeddings (discarded at fine-tuning). The tokenizer is
also similar to the Albert one rather than the BERT one.
## Resources
- [Text classification task guide](../tasks/sequence_classification)
- [Token classification task guide](../tasks/token_classification)
- [Question answering task guide](../tasks/question_answering)
- [Causal language modeling task guide](../tasks/language_modeling)
- [Masked language modeling task guide](../tasks/masked_language_modeling)
- [Multiple choice task guide](../tasks/multiple_choice)
## RemBertConfig
[[autodoc]] RemBertConfig
## RemBertTokenizer
[[autodoc]] RemBertTokenizer
- build_inputs_with_special_tokens
- get_special_tokens_mask
- create_token_type_ids_from_sequences
- save_vocabulary
## RemBertTokenizerFast
[[autodoc]] RemBertTokenizerFast
- build_inputs_with_special_tokens
- get_special_tokens_mask
- create_token_type_ids_from_sequences
- save_vocabulary
<frameworkcontent>
<pt>
## RemBertModel
[[autodoc]] RemBertModel
- forward
## RemBertForCausalLM
[[autodoc]] RemBertForCausalLM
- forward
## RemBertForMaskedLM
[[autodoc]] RemBertForMaskedLM
- forward
## RemBertForSequenceClassification
[[autodoc]] RemBertForSequenceClassification
- forward
## RemBertForMultipleChoice
[[autodoc]] RemBertForMultipleChoice
- forward
## RemBertForTokenClassification
[[autodoc]] RemBertForTokenClassification
- forward
## RemBertForQuestionAnswering
[[autodoc]] RemBertForQuestionAnswering
- forward
</pt>
<tf>
## TFRemBertModel
[[autodoc]] TFRemBertModel
- call
## TFRemBertForMaskedLM
[[autodoc]] TFRemBertForMaskedLM
- call
## TFRemBertForCausalLM
[[autodoc]] TFRemBertForCausalLM
- call
## TFRemBertForSequenceClassification
[[autodoc]] TFRemBertForSequenceClassification
- call
## TFRemBertForMultipleChoice
[[autodoc]] TFRemBertForMultipleChoice
- call
## TFRemBertForTokenClassification
[[autodoc]] TFRemBertForTokenClassification
- call
## TFRemBertForQuestionAnswering
[[autodoc]] TFRemBertForQuestionAnswering
- call
</tf>
</frameworkcontent>
|
transformers/docs/source/en/model_doc/rembert.md/0
|
{
"file_path": "transformers/docs/source/en/model_doc/rembert.md",
"repo_id": "transformers",
"token_count": 1363
}
| 280
|
<!--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.
-->
# SigLIP
## Overview
The SigLIP model was proposed in [Sigmoid Loss for Language Image Pre-Training](https://arxiv.org/abs/2303.15343) by Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, Lucas Beyer. SigLIP proposes to replace the loss function used in [CLIP](clip) by a simple pairwise sigmoid loss. This results in better performance in terms of zero-shot classification accuracy on ImageNet.
The abstract from the paper is the following:
*We propose a simple pairwise Sigmoid loss for Language-Image Pre-training (SigLIP). Unlike standard contrastive learning with softmax normalization, the sigmoid loss operates solely on image-text pairs and does not require a global view of the pairwise similarities for normalization. The sigmoid loss simultaneously allows further scaling up the batch size, while also performing better at smaller batch sizes. Combined with Locked-image Tuning, with only four TPUv4 chips, we train a SigLiT model that achieves 84.5% ImageNet zero-shot accuracy in two days. The disentanglement of the batch size from the loss further allows us to study the impact of examples vs pairs and negative to positive ratio. Finally, we push the batch size to the extreme, up to one million, and find that the benefits of growing batch size quickly diminish, with a more reasonable batch size of 32k being sufficient.*
## Usage tips
- Usage of SigLIP is similar to [CLIP](clip). The main difference is the training loss, which does not require a global view of all the pairwise similarities of images and texts within a batch. One needs to apply the sigmoid activation function to the logits, rather than the softmax.
- Training is supported but does not use `torch.distributed` utilities which may limit the scalability of batch size. However, DDP and FDSP works on single-node multi-gpu setup.
- When using the standalone [`SiglipTokenizer`] or [`SiglipProcessor`], make sure to pass `padding="max_length"` as that's how the model was trained.
- To get the same results as the pipeline, a prompt template of "This is a photo of {label}." should be used.
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/siglip_table.jpeg"
alt="drawing" width="600"/>
<small> SigLIP evaluation results compared to CLIP. Taken from the <a href="https://arxiv.org/abs/2303.15343">original paper</a>.</small>
This model was contributed by [nielsr](https://huggingface.co/nielsr).
The original code can be found [here](https://github.com/google-research/big_vision/tree/main).
## Usage example
There are 2 main ways to use SigLIP: either using the pipeline API, which abstracts away all the complexity for you, or by using the `SiglipModel` class yourself.
### Pipeline API
The pipeline allows to use the model in a few lines of code:
```python
>>> from transformers import pipeline
>>> from PIL import Image
>>> import requests
>>> # load pipe
>>> image_classifier = pipeline(task="zero-shot-image-classification", model="google/siglip-base-patch16-224")
>>> # load image
>>> url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> # inference
>>> candidate_labels = ["2 cats", "a plane", "a remote"]
>>> outputs = image_classifier(image, candidate_labels=candidate_labels)
>>> outputs = [{"score": round(output["score"], 4), "label": output["label"] } for output in outputs]
>>> print(outputs)
[{'score': 0.1979, 'label': '2 cats'}, {'score': 0.0, 'label': 'a remote'}, {'score': 0.0, 'label': 'a plane'}]
```
### Using the model yourself
If you want to do the pre- and postprocessing yourself, here's how to do that:
```python
>>> from PIL import Image
>>> import requests
>>> from transformers import AutoProcessor, AutoModel
>>> import torch
>>> model = AutoModel.from_pretrained("google/siglip-base-patch16-224")
>>> processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> candidate_labels = ["2 cats", "2 dogs"]
# follows the pipeline prompt template to get same results
>>> candidate_labels = [f'This is a photo of {label}.' for label in candidate_labels]
>>> # important: we pass `padding=max_length` since the model was trained with this
>>> inputs = processor(text=texts, images=image, padding="max_length", return_tensors="pt")
>>> with torch.no_grad():
... outputs = model(**inputs)
>>> logits_per_image = outputs.logits_per_image
>>> probs = torch.sigmoid(logits_per_image) # these are the probabilities
>>> print(f"{probs[0][0]:.1%} that image 0 is '{texts[0]}'")
31.9% that image 0 is 'a photo of 2 cats'
```
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with SigLIP.
- [Zero-shot image classification task guide](../tasks/zero_shot_image_classification_md)
- Demo notebooks for SigLIP can be found [here](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/SigLIP). 🌎
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.
## Combining SigLIP and Flash Attention 2
First, make sure to install the latest version of Flash Attention 2.
```bash
pip install -U flash-attn --no-build-isolation
```
Make also sure that you have a hardware that is compatible with Flash-Attention 2. Read more about it in the official documentation of flash-attn repository. Make also sure to load your model in half-precision (e.g. `torch.float16``)
To load and run a model using Flash Attention 2, refer to the snippet below:
```python
>>> import torch
>>> import requests
>>> from PIL import Image
>>> from transformers import SiglipProcessor, SiglipModel
>>> device = "cuda" # the device to load the model onto
>>> model = SiglipModel.from_pretrained(
... "google/siglip-so400m-patch14-384",
... attn_implementation="flash_attention_2",
... torch_dtype=torch.float16,
... device_map=device,
... )
>>> processor = SiglipProcessor.from_pretrained("google/siglip-so400m-patch14-384")
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> candidate_labels = ["2 cats", "2 dogs"]
# follows the pipeline prompt template to get same results
>>> candidate_labels = [f'This is a photo of {label}.' for label in candidate_labels]
# important: we pass `padding=max_length` since the model was trained with this
>>> inputs = processor(text=candidate_labels, images=image, padding="max_length", return_tensors="pt")
>>> inputs.to(device)
>>> with torch.no_grad():
... with torch.autocast(device):
... outputs = model(**inputs)
>>> logits_per_image = outputs.logits_per_image
>>> probs = torch.sigmoid(logits_per_image) # these are the probabilities
>>> print(f"{probs[0][0]:.1%} that image 0 is '{candidate_labels[0]}'")
51.3% that image 0 is 'This is a photo of 2 cats.'
```
## Using Scaled Dot Product Attention (SDPA)
PyTorch includes a native scaled dot-product attention (SDPA) operator as part of `torch.nn.functional`. This function
encompasses several implementations that can be applied depending on the inputs and the hardware in use. See the
[official documentation](https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html)
or the [GPU Inference](https://huggingface.co/docs/transformers/main/en/perf_infer_gpu_one#pytorch-scaled-dot-product-attention)
page for more information.
You may set `attn_implementation="sdpa"` in `from_pretrained()` to explicitly request SDPA to be used. Make sure you have `torch>=2.1.1`.
```python
>>> from transformers import SiglipModel
>>> model = SiglipModel.from_pretrained(
... "google/siglip-so400m-patch14-384",
... attn_implementation="sdpa",
... torch_dtype=torch.float16,
... device_map=device,
... )
```
For the best speedups, we recommend loading the model in half-precision (e.g. `torch.float16` or `torch.bfloat16`).
## Expected speedups
Below is an expected speedup diagram that compares inference time between the native implementation in transformers using `google/siglip-so400m-patch14-384` checkpoint in `float16` precision and the Flash Attention 2 / SDPA version of the model using different batch sizes.
<div style="text-align: center">
<img src="https://i.imgur.com/cWm4rsn.png">
</div>
## SiglipConfig
[[autodoc]] SiglipConfig
- from_text_vision_configs
## SiglipTextConfig
[[autodoc]] SiglipTextConfig
## SiglipVisionConfig
[[autodoc]] SiglipVisionConfig
## SiglipTokenizer
[[autodoc]] SiglipTokenizer
- build_inputs_with_special_tokens
- get_special_tokens_mask
- create_token_type_ids_from_sequences
- save_vocabulary
## SiglipImageProcessor
[[autodoc]] SiglipImageProcessor
- preprocess
## SiglipProcessor
[[autodoc]] SiglipProcessor
## SiglipModel
[[autodoc]] SiglipModel
- forward
- get_text_features
- get_image_features
## SiglipTextModel
[[autodoc]] SiglipTextModel
- forward
## SiglipVisionModel
[[autodoc]] SiglipVisionModel
- forward
## SiglipForImageClassification
[[autodoc]] SiglipForImageClassification
- forward
|
transformers/docs/source/en/model_doc/siglip.md/0
|
{
"file_path": "transformers/docs/source/en/model_doc/siglip.md",
"repo_id": "transformers",
"token_count": 3066
}
| 281
|
<!--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.
-->
# T5v1.1
## Overview
T5v1.1 was released in the [google-research/text-to-text-transfer-transformer](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511)
repository by Colin Raffel et al. It's an improved version of the original T5 model.
This model was contributed by [patrickvonplaten](https://huggingface.co/patrickvonplaten). The original code can be
found [here](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511).
## Usage tips
One can directly plug in the weights of T5v1.1 into a T5 model, like so:
```python
>>> from transformers import T5ForConditionalGeneration
>>> model = T5ForConditionalGeneration.from_pretrained("google/t5-v1_1-base")
```
T5 Version 1.1 includes the following improvements compared to the original T5 model:
- GEGLU activation in the feed-forward hidden layer, rather than ReLU. See [this paper](https://arxiv.org/abs/2002.05202).
- Dropout was turned off in pre-training (quality win). Dropout should be re-enabled during fine-tuning.
- Pre-trained on C4 only without mixing in the downstream tasks.
- No parameter sharing between the embedding and classifier layer.
- "xl" and "xxl" replace "3B" and "11B". The model shapes are a bit different - larger `d_model` and smaller
`num_heads` and `d_ff`.
Note: T5 Version 1.1 was only pre-trained on [C4](https://huggingface.co/datasets/c4) excluding any supervised
training. Therefore, this model has to be fine-tuned before it is usable on a downstream task, unlike the original T5
model. Since t5v1.1 was pre-trained unsupervisedly, there's no real advantage to using a task prefix during single-task
fine-tuning. If you are doing multi-task fine-tuning, you should use a prefix.
Google has released the following variants:
- [google/t5-v1_1-small](https://huggingface.co/google/t5-v1_1-small)
- [google/t5-v1_1-base](https://huggingface.co/google/t5-v1_1-base)
- [google/t5-v1_1-large](https://huggingface.co/google/t5-v1_1-large)
- [google/t5-v1_1-xl](https://huggingface.co/google/t5-v1_1-xl)
- [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl).
<Tip>
Refer to [T5's documentation page](t5) for all API reference, tips, code examples and notebooks.
</Tip>
|
transformers/docs/source/en/model_doc/t5v1.1.md/0
|
{
"file_path": "transformers/docs/source/en/model_doc/t5v1.1.md",
"repo_id": "transformers",
"token_count": 967
}
| 282
|
<!--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.
-->
# XLS-R
## Overview
The XLS-R model was proposed in [XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale](https://arxiv.org/abs/2111.09296) by Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman
Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, Michael Auli.
The abstract from the paper is the following:
*This paper presents XLS-R, a large-scale model for cross-lingual speech representation learning based on wav2vec 2.0.
We train models with up to 2B parameters on nearly half a million hours of publicly available speech audio in 128
languages, an order of magnitude more public data than the largest known prior work. Our evaluation covers a wide range
of tasks, domains, data regimes and languages, both high and low-resource. On the CoVoST-2 speech translation
benchmark, we improve the previous state of the art by an average of 7.4 BLEU over 21 translation directions into
English. For speech recognition, XLS-R improves over the best known prior work on BABEL, MLS, CommonVoice as well as
VoxPopuli, lowering error rates by 14-34% relative on average. XLS-R also sets a new state of the art on VoxLingua107
language identification. Moreover, we show that with sufficient model size, cross-lingual pretraining can outperform
English-only pretraining when translating English speech into other languages, a setting which favors monolingual
pretraining. We hope XLS-R can help to improve speech processing tasks for many more languages of the world.*
Relevant checkpoints can be found under https://huggingface.co/models?other=xls_r.
The original code can be found [here](https://github.com/pytorch/fairseq/tree/master/fairseq/models/wav2vec).
## Usage tips
- XLS-R is a speech model that accepts a float array corresponding to the raw waveform of the speech signal.
- XLS-R model was trained using connectionist temporal classification (CTC) so the model output has to be decoded using
[`Wav2Vec2CTCTokenizer`].
<Tip>
XLS-R's architecture is based on the Wav2Vec2 model, refer to [Wav2Vec2's documentation page](wav2vec2) for API reference.
</Tip>
|
transformers/docs/source/en/model_doc/xls_r.md/0
|
{
"file_path": "transformers/docs/source/en/model_doc/xls_r.md",
"repo_id": "transformers",
"token_count": 782
}
| 283
|
<!--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()
This guide aims to provide a benchmark on the inference speed-ups introduced with [`torch.compile()`](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) for [computer vision models in 🤗 Transformers](https://huggingface.co/models?pipeline_tag=image-classification&library=transformers&sort=trending).
## Benefits of torch.compile
Depending on the model and the GPU, `torch.compile()` yields up to 30% speed-up during inference. To use `torch.compile()`, simply install any version of `torch` above 2.0.
Compiling a model takes time, so it's useful if you are compiling the model only once instead of every time you infer.
To compile any computer vision model of your choice, call `torch.compile()` on the model as shown below:
```diff
from transformers import AutoModelForImageClassification
model = AutoModelForImageClassification.from_pretrained(MODEL_ID).to("cuda")
+ model = torch.compile(model)
```
`compile()` comes with multiple modes for compiling, which essentially differ in compilation time and inference overhead. `max-autotune` takes longer than `reduce-overhead` but results in faster inference. Default mode is fastest for compilation but is not as efficient compared to `reduce-overhead` for inference time. In this guide, we used the default mode. You can learn more about it [here](https://pytorch.org/get-started/pytorch-2.0/#user-experience).
We benchmarked `torch.compile` with different computer vision models, tasks, types of hardware, and batch sizes on `torch` version 2.0.1.
## Benchmarking code
Below you can find the benchmarking code for each task. We warm up the GPU before inference and take the mean time of 300 inferences, using the same image each time.
### Image Classification with ViT
```python
import torch
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").to("cuda")
model = torch.compile(model)
processed_input = processor(image, return_tensors='pt').to(device="cuda")
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").to("cuda")
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("cuda")
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").to("cuda")
model = torch.compile(model)
seg_inputs = processor(images=image, return_tensors="pt").to("cuda")
with torch.no_grad():
_ = model(**seg_inputs)
```
Below you can find the list of the models we benchmarked.
**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/microsoft/resnet-50)
**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)
Below you can find visualization of inference durations with and without `torch.compile()` and percentage improvements for each model in different hardware and batch sizes.
<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>
Below you can find inference durations in milliseconds for each model with and without `compile()`. Note that OwlViT results in OOM in larger batch sizes.
### 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
We also benchmarked on PyTorch nightly (2.1.0dev, find the wheel [here](https://download.pytorch.org/whl/nightly/cu118)) and observed improvement in latency both for uncompiled and compiled models.
### 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
We benchmarked `reduce-overhead` compilation mode for A100 and T4 in Nightly.
### 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/en/perf_torch_compile.md/0
|
{
"file_path": "transformers/docs/source/en/perf_torch_compile.md",
"repo_id": "transformers",
"token_count": 5865
}
| 284
|
<!--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
[bitsandbytes](https://github.com/TimDettmers/bitsandbytes) is the easiest option for quantizing a model to 8 and 4-bit. 8-bit quantization multiplies outliers in fp16 with non-outliers in int8, converts the non-outlier values back to fp16, and then adds them together to return the weights in fp16. This reduces the degradative effect outlier values have on a model's performance. 4-bit quantization compresses a model even further, and it is commonly used with [QLoRA](https://hf.co/papers/2305.14314) to finetune quantized LLMs.
To use bitsandbytes, make sure you have the following libraries installed:
<hfoptions id="bnb">
<hfoption id="8-bit">
```bash
pip install transformers accelerate bitsandbytes>0.37.0
```
</hfoption>
<hfoption id="4-bit">
```bash
pip install bitsandbytes>=0.39.0
pip install --upgrade accelerate transformers
```
</hfoption>
</hfoptions>
Now you can quantize a model by passing a `BitsAndBytesConfig` to [`~PreTrainedModel.from_pretrained`] method. This works for any model in any modality, as long as it supports loading with Accelerate and contains `torch.nn.Linear` layers.
<hfoptions id="bnb">
<hfoption id="8-bit">
Quantizing a model in 8-bit halves the memory-usage, and for large models, set `device_map="auto"` to efficiently use the GPUs available:
```py
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(load_in_8bit=True)
model_8bit = AutoModelForCausalLM.from_pretrained(
"bigscience/bloom-1b7",
quantization_config=quantization_config
)
```
By default, all the other modules such as `torch.nn.LayerNorm` are converted to `torch.float16`. You can change the data type of these modules with the `torch_dtype` parameter if you want:
```py
import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(load_in_8bit=True)
model_8bit = AutoModelForCausalLM.from_pretrained(
"facebook/opt-350m",
quantization_config=quantization_config,
torch_dtype=torch.float32
)
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 the [`~PreTrainedModel.push_to_hub`] method. 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",
quantization_config=quantization_config
)
tokenizer = AutoTokenizer.from_pretrained("bigscience/bloom-560m")
model.push_to_hub("bloom-560m-8bit")
```
</hfoption>
<hfoption id="4-bit">
Quantizing a model in 4-bit reduces your memory-usage by 4x, and for large models, set `device_map="auto"` to efficiently use the GPUs available:
```py
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(load_in_4bit=True)
model_4bit = AutoModelForCausalLM.from_pretrained(
"bigscience/bloom-1b7",
quantization_config=quantization_config
)
```
By default, all the other modules such as `torch.nn.LayerNorm` are converted to `torch.float16`. You can change the data type of these modules with the `torch_dtype` parameter if you want:
```py
import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(load_in_4bit=True)
model_4bit = AutoModelForCausalLM.from_pretrained(
"facebook/opt-350m",
quantization_config=quantization_config,
torch_dtype=torch.float32
)
model_4bit.model.decoder.layers[-1].final_layer_norm.weight.dtype
```
If you have `bitsandbytes>=0.41.3`, you can serialize 4-bit models and push them on Hugging Face Hub. Simply call `model.push_to_hub()` after loading it in 4-bit precision. You can also save the serialized 4-bit models locally with `model.save_pretrained()` command.
</hfoption>
</hfoptions>
<Tip warning={true}>
Training with 8-bit and 4-bit weights are only supported for training *extra* parameters.
</Tip>
You can check your memory footprint with the `get_memory_footprint` method:
```py
print(model.get_memory_footprint())
```
Quantized models can be loaded from the [`~PreTrainedModel.from_pretrained`] method without needing to specify the `load_in_8bit` or `load_in_4bit` parameters:
```py
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("{your_username}/bloom-560m-8bit", device_map="auto")
```
## 8-bit (LLM.int8() algorithm)
<Tip>
Learn more about the details of 8-bit quantization in this [blog post](https://huggingface.co/blog/hf-bitsandbytes-integration)!
</Tip>
This section explores some of the specific features of 8-bit models, such as offloading, outlier thresholds, skipping module conversion, and finetuning.
### Offloading
8-bit models can offload weights between the CPU and GPU to support fitting very large models into memory. The weights dispatched to the CPU are actually stored in **float32**, and aren't converted to 8-bit. For example, to enable offloading for the [bigscience/bloom-1b7](https://huggingface.co/bigscience/bloom-1b7) model, start by creating a [`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 you'll dispatch 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",
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, we recommend experimenting with the `llm_int8_threshold` parameter in [`BitsAndBytesConfig`]:
```py
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
model_id = "bigscience/bloom-1b7"
quantization_config = BitsAndBytesConfig(
llm_int8_threshold=10,
)
model_8bit = AutoModelForCausalLM.from_pretrained(
model_id,
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 which 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,
device_map="auto",
quantization_config=quantization_config,
)
```
### Finetuning
With the [PEFT](https://github.com/huggingface/peft) library, you can finetune 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'll automatically load your model on a GPU. However, you can still customize the device map with the `device_map` parameter if you want to (`device_map="auto"` should only be used for inference).
## 4-bit (QLoRA algorithm)
<Tip>
Try 4-bit quantization in this [notebook](https://colab.research.google.com/drive/1ge2F1QSK8Q7h0hn3YKuBCOAS0bK8E0wf) and learn more about it's details in this [blog post](https://huggingface.co/blog/4bit-transformers-bitsandbytes).
</Tip>
This section explores some of the specific features of 4-bit models, such as changing the compute data type, using the Normal Float 4 (NF4) data type, and using nested quantization.
### Compute data type
To speedup computation, you can change the data type from float32 (the default value) to bf16 using the `bnb_4bit_compute_dtype` parameter in [`BitsAndBytesConfig`]:
```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. This can be configured with the `bnb_4bit_quant_type` parameter in the [`BitsAndBytesConfig`]:
```py
from transformers import BitsAndBytesConfig
nf4_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
)
model_nf4 = AutoModelForCausalLM.from_pretrained(model_id, 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 `torch_dtype` values.
### Nested quantization
Nested quantization is a technique that can save additional memory at no additional performance cost. This feature performs a second quantization of the already quantized weights to save an addition 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 enabling 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", quantization_config=double_quant_config)
```
## Dequantizing `bitsandbytes` models
Once quantized, you can dequantize the model to the original precision but this might result in a small quality loss of the model. Make sure you have enough GPU RAM to fit the dequantized model.
```python
from transformers import AutoModelForCausalLM, BitsAndBytesConfig, AutoTokenizer
model_id = "facebook/opt-125m"
model = AutoModelForCausalLM.from_pretrained(model_id, BitsAndBytesConfig(load_in_4bit=True))
tokenizer = AutoTokenizer.from_pretrained(model_id)
model.dequantize()
text = tokenizer("Hello my name is", return_tensors="pt").to(0)
out = model.generate(**text)
print(tokenizer.decode(out[0]))
```
|
transformers/docs/source/en/quantization/bitsandbytes.md/0
|
{
"file_path": "transformers/docs/source/en/quantization/bitsandbytes.md",
"repo_id": "transformers",
"token_count": 3848
}
| 285
|
<!--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 classification
[[open-in-colab]]
<Youtube id="KWwzcmG98Ds"/>
Audio classification - just like with text - assigns a class label output from the input data. The only difference is instead of text inputs, you have raw audio waveforms. Some practical applications of audio classification include identifying speaker intent, language classification, and even animal species by their sounds.
This guide will show you how to:
1. Finetune [Wav2Vec2](https://huggingface.co/facebook/wav2vec2-base) on the [MInDS-14](https://huggingface.co/datasets/PolyAI/minds14) dataset to classify speaker intent.
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/audio-classification)
</Tip>
Before you begin, make sure you have all the necessary libraries installed:
```bash
pip install transformers datasets evaluate
```
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 MInDS-14 dataset
Start by loading the MInDS-14 dataset from the 🤗 Datasets library:
```py
>>> from datasets import load_dataset, Audio
>>> minds = load_dataset("PolyAI/minds14", name="en-US", split="train")
```
Split the dataset's `train` split into a smaller train and test set with the [`~datasets.Dataset.train_test_split`] method. This'll give you a chance to experiment and make sure everything works before spending more time on the full dataset.
```py
>>> minds = minds.train_test_split(test_size=0.2)
```
Then take a look at the dataset:
```py
>>> minds
DatasetDict({
train: Dataset({
features: ['path', 'audio', 'transcription', 'english_transcription', 'intent_class', 'lang_id'],
num_rows: 450
})
test: Dataset({
features: ['path', 'audio', 'transcription', 'english_transcription', 'intent_class', 'lang_id'],
num_rows: 113
})
})
```
While the dataset contains a lot of useful information, like `lang_id` and `english_transcription`, you'll focus on the `audio` and `intent_class` in this guide. Remove the other columns with the [`~datasets.Dataset.remove_columns`] method:
```py
>>> minds = minds.remove_columns(["path", "transcription", "english_transcription", "lang_id"])
```
Take a look at an example now:
```py
>>> minds["train"][0]
{'audio': {'array': array([ 0. , 0. , 0. , ..., -0.00048828,
-0.00024414, -0.00024414], dtype=float32),
'path': '/root/.cache/huggingface/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~APP_ERROR/602b9a5fbb1e6d0fbce91f52.wav',
'sampling_rate': 8000},
'intent_class': 2}
```
There are two fields:
- `audio`: a 1-dimensional `array` of the speech signal that must be called to load and resample the audio file.
- `intent_class`: represents the class id of the speaker's intent.
To make it easier for the model to get the label name from the label id, create a dictionary that maps the label name to an integer and vice versa:
```py
>>> labels = minds["train"].features["intent_class"].names
>>> label2id, id2label = dict(), dict()
>>> for i, label in enumerate(labels):
... label2id[label] = str(i)
... id2label[str(i)] = label
```
Now you can convert the label id to a label name:
```py
>>> id2label[str(2)]
'app_error'
```
## Preprocess
The next step is to load a Wav2Vec2 feature extractor to process the audio signal:
```py
>>> from transformers import AutoFeatureExtractor
>>> feature_extractor = AutoFeatureExtractor.from_pretrained("facebook/wav2vec2-base")
```
The MInDS-14 dataset has a sampling rate of 8000khz (you can find this information in it's [dataset card](https://huggingface.co/datasets/PolyAI/minds14)), which means you'll need to resample the dataset to 16000kHz to use the pretrained Wav2Vec2 model:
```py
>>> minds = minds.cast_column("audio", Audio(sampling_rate=16_000))
>>> minds["train"][0]
{'audio': {'array': array([ 2.2098757e-05, 4.6582241e-05, -2.2803260e-05, ...,
-2.8419291e-04, -2.3305941e-04, -1.1425107e-04], dtype=float32),
'path': '/root/.cache/huggingface/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~APP_ERROR/602b9a5fbb1e6d0fbce91f52.wav',
'sampling_rate': 16000},
'intent_class': 2}
```
Now create a preprocessing function that:
1. Calls the `audio` column to load, and if necessary, resample the audio file.
2. Checks if the sampling rate of the audio file matches the sampling rate of the audio data a model was pretrained with. You can find this information in the Wav2Vec2 [model card](https://huggingface.co/facebook/wav2vec2-base).
3. Set a maximum input length to batch longer inputs without truncating them.
```py
>>> def preprocess_function(examples):
... audio_arrays = [x["array"] for x in examples["audio"]]
... inputs = feature_extractor(
... audio_arrays, sampling_rate=feature_extractor.sampling_rate, max_length=16000, truncation=True
... )
... return inputs
```
To apply the preprocessing function over the entire dataset, use 🤗 Datasets [`~datasets.Dataset.map`] function. You can speed up `map` by setting `batched=True` to process multiple elements of the dataset at once. Remove the columns you don't need, and rename `intent_class` to `label` because that's the name the model expects:
```py
>>> encoded_minds = minds.map(preprocess_function, remove_columns="audio", batched=True)
>>> encoded_minds = encoded_minds.rename_column("intent_class", "label")
```
## 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 [accuracy](https://huggingface.co/spaces/evaluate-metric/accuracy) 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
>>> accuracy = evaluate.load("accuracy")
```
Then create a function that passes your predictions and labels to [`~evaluate.EvaluationModule.compute`] to calculate the accuracy:
```py
>>> import numpy as np
>>> def compute_metrics(eval_pred):
... predictions = np.argmax(eval_pred.predictions, axis=1)
... return accuracy.compute(predictions=predictions, references=eval_pred.label_ids)
```
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 Wav2Vec2 with [`AutoModelForAudioClassification`] along with the number of expected labels, and the label mappings:
```py
>>> from transformers import AutoModelForAudioClassification, TrainingArguments, Trainer
>>> num_labels = len(id2label)
>>> model = AutoModelForAudioClassification.from_pretrained(
... "facebook/wav2vec2-base", num_labels=num_labels, label2id=label2id, id2label=id2label
... )
```
At this point, only three steps remain:
1. Define your training hyperparameters in [`TrainingArguments`]. 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 accuracy and save the training checkpoint.
2. Pass the training arguments to [`Trainer`] along with the model, dataset, tokenizer, data collator, and `compute_metrics` function.
3. Call [`~Trainer.train`] to finetune your model.
```py
>>> training_args = TrainingArguments(
... output_dir="my_awesome_mind_model",
... eval_strategy="epoch",
... save_strategy="epoch",
... learning_rate=3e-5,
... per_device_train_batch_size=32,
... gradient_accumulation_steps=4,
... per_device_eval_batch_size=32,
... num_train_epochs=10,
... warmup_ratio=0.1,
... logging_steps=10,
... load_best_model_at_end=True,
... metric_for_best_model="accuracy",
... push_to_hub=True,
... )
>>> trainer = Trainer(
... model=model,
... args=training_args,
... train_dataset=encoded_minds["train"],
... eval_dataset=encoded_minds["test"],
... tokenizer=feature_extractor,
... 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>
</frameworkcontent>
<Tip>
For a more in-depth example of how to finetune a model for audio classification, take a look at the corresponding [PyTorch notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/audio_classification.ipynb).
</Tip>
## Inference
Great, now that you've finetuned a model, you can use it for inference!
Load an audio file you'd like to run inference on. Remember to resample the sampling rate of the audio file to match the sampling rate of the model if you need to!
```py
>>> from datasets import load_dataset, Audio
>>> dataset = load_dataset("PolyAI/minds14", name="en-US", split="train")
>>> dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))
>>> sampling_rate = dataset.features["audio"].sampling_rate
>>> audio_file = dataset[0]["audio"]["path"]
```
The simplest way to try out your finetuned model for inference is to use it in a [`pipeline`]. Instantiate a `pipeline` for audio classification with your model, and pass your audio file to it:
```py
>>> from transformers import pipeline
>>> classifier = pipeline("audio-classification", model="stevhliu/my_awesome_minds_model")
>>> classifier(audio_file)
[
{'score': 0.09766869246959686, 'label': 'cash_deposit'},
{'score': 0.07998877018690109, 'label': 'app_error'},
{'score': 0.0781070664525032, 'label': 'joint_account'},
{'score': 0.07667109370231628, 'label': 'pay_bill'},
{'score': 0.0755252093076706, 'label': 'balance'}
]
```
You can also manually replicate the results of the `pipeline` if you'd like:
<frameworkcontent>
<pt>
Load a feature extractor to preprocess the audio file and return the `input` as PyTorch tensors:
```py
>>> from transformers import AutoFeatureExtractor
>>> feature_extractor = AutoFeatureExtractor.from_pretrained("stevhliu/my_awesome_minds_model")
>>> inputs = feature_extractor(dataset[0]["audio"]["array"], sampling_rate=sampling_rate, return_tensors="pt")
```
Pass your inputs to the model and return the logits:
```py
>>> from transformers import AutoModelForAudioClassification
>>> model = AutoModelForAudioClassification.from_pretrained("stevhliu/my_awesome_minds_model")
>>> with torch.no_grad():
... logits = model(**inputs).logits
```
Get the class with the highest probability, and use the model's `id2label` mapping to convert it to a label:
```py
>>> import torch
>>> predicted_class_ids = torch.argmax(logits).item()
>>> predicted_label = model.config.id2label[predicted_class_ids]
>>> predicted_label
'cash_deposit'
```
</pt>
</frameworkcontent>
|
transformers/docs/source/en/tasks/audio_classification.md/0
|
{
"file_path": "transformers/docs/source/en/tasks/audio_classification.md",
"repo_id": "transformers",
"token_count": 4050
}
| 286
|
<!--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.
-->
# Question answering
[[open-in-colab]]
<Youtube id="ajPx5LwJD-I"/>
Question answering tasks return an answer given a question. If you've ever asked a virtual assistant like Alexa, Siri or Google what the weather is, then you've used a question answering model before. There are two common types of question answering tasks:
- Extractive: extract the answer from the given context.
- Abstractive: generate an answer from the context that correctly answers the question.
This guide will show you how to:
1. Finetune [DistilBERT](https://huggingface.co/distilbert/distilbert-base-uncased) on the [SQuAD](https://huggingface.co/datasets/squad) dataset for extractive question answering.
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/question-answering)
</Tip>
Before you begin, make sure you have all the necessary libraries installed:
```bash
pip install transformers datasets evaluate
```
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 SQuAD dataset
Start by loading a smaller subset of the SQuAD dataset from the 🤗 Datasets library. This'll give you a chance to experiment and make sure everything works before spending more time training on the full dataset.
```py
>>> from datasets import load_dataset
>>> squad = load_dataset("squad", split="train[:5000]")
```
Split the dataset's `train` split into a train and test set with the [`~datasets.Dataset.train_test_split`] method:
```py
>>> squad = squad.train_test_split(test_size=0.2)
```
Then take a look at an example:
```py
>>> squad["train"][0]
{'answers': {'answer_start': [515], 'text': ['Saint Bernadette Soubirous']},
'context': 'Architecturally, the school has a Catholic character. Atop the Main Building\'s gold dome is a golden statue of the Virgin Mary. Immediately in front of the Main Building and facing it, is a copper statue of Christ with arms upraised with the legend "Venite Ad Me Omnes". Next to the Main Building is the Basilica of the Sacred Heart. Immediately behind the basilica is the Grotto, a Marian place of prayer and reflection. It is a replica of the grotto at Lourdes, France where the Virgin Mary reputedly appeared to Saint Bernadette Soubirous in 1858. At the end of the main drive (and in a direct line that connects through 3 statues and the Gold Dome), is a simple, modern stone statue of Mary.',
'id': '5733be284776f41900661182',
'question': 'To whom did the Virgin Mary allegedly appear in 1858 in Lourdes France?',
'title': 'University_of_Notre_Dame'
}
```
There are several important fields here:
- `answers`: the starting location of the answer token and the answer text.
- `context`: background information from which the model needs to extract the answer.
- `question`: the question a model should answer.
## Preprocess
<Youtube id="qgaM0weJHpA"/>
The next step is to load a DistilBERT tokenizer to process the `question` and `context` fields:
```py
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("distilbert/distilbert-base-uncased")
```
There are a few preprocessing steps particular to question answering tasks you should be aware of:
1. Some examples in a dataset may have a very long `context` that exceeds the maximum input length of the model. To deal with longer sequences, truncate only the `context` by setting `truncation="only_second"`.
2. Next, map the start and end positions of the answer to the original `context` by setting
`return_offset_mapping=True`.
3. With the mapping in hand, now you can find the start and end tokens of the answer. Use the [`~tokenizers.Encoding.sequence_ids`] method to
find which part of the offset corresponds to the `question` and which corresponds to the `context`.
Here is how you can create a function to truncate and map the start and end tokens of the `answer` to the `context`:
```py
>>> def preprocess_function(examples):
... questions = [q.strip() for q in examples["question"]]
... inputs = tokenizer(
... questions,
... examples["context"],
... max_length=384,
... truncation="only_second",
... return_offsets_mapping=True,
... padding="max_length",
... )
... offset_mapping = inputs.pop("offset_mapping")
... answers = examples["answers"]
... start_positions = []
... end_positions = []
... for i, offset in enumerate(offset_mapping):
... answer = answers[i]
... start_char = answer["answer_start"][0]
... end_char = answer["answer_start"][0] + len(answer["text"][0])
... sequence_ids = inputs.sequence_ids(i)
... # Find the start and end of the context
... idx = 0
... while sequence_ids[idx] != 1:
... idx += 1
... context_start = idx
... while sequence_ids[idx] == 1:
... idx += 1
... context_end = idx - 1
... # If the answer is not fully inside the context, label it (0, 0)
... if offset[context_start][0] > end_char or offset[context_end][1] < start_char:
... start_positions.append(0)
... end_positions.append(0)
... else:
... # Otherwise it's the start and end token positions
... idx = context_start
... while idx <= context_end and offset[idx][0] <= start_char:
... idx += 1
... start_positions.append(idx - 1)
... idx = context_end
... while idx >= context_start and offset[idx][1] >= end_char:
... idx -= 1
... end_positions.append(idx + 1)
... inputs["start_positions"] = start_positions
... inputs["end_positions"] = end_positions
... return inputs
```
To apply the preprocessing function over the entire dataset, use 🤗 Datasets [`~datasets.Dataset.map`] function. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once. Remove any columns you don't need:
```py
>>> tokenized_squad = squad.map(preprocess_function, batched=True, remove_columns=squad["train"].column_names)
```
Now create a batch of examples using [`DefaultDataCollator`]. Unlike other data collators in 🤗 Transformers, the [`DefaultDataCollator`] does not apply any additional preprocessing such as padding.
<frameworkcontent>
<pt>
```py
>>> from transformers import DefaultDataCollator
>>> data_collator = DefaultDataCollator()
```
</pt>
<tf>
```py
>>> from transformers import DefaultDataCollator
>>> data_collator = DefaultDataCollator(return_tensors="tf")
```
</tf>
</frameworkcontent>
## 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 DistilBERT with [`AutoModelForQuestionAnswering`]:
```py
>>> from transformers import AutoModelForQuestionAnswering, TrainingArguments, Trainer
>>> model = AutoModelForQuestionAnswering.from_pretrained("distilbert/distilbert-base-uncased")
```
At this point, only three steps remain:
1. Define your training hyperparameters in [`TrainingArguments`]. 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).
2. Pass the training arguments to [`Trainer`] along with the model, dataset, tokenizer, and data collator.
3. Call [`~Trainer.train`] to finetune your model.
```py
>>> training_args = TrainingArguments(
... output_dir="my_awesome_qa_model",
... eval_strategy="epoch",
... learning_rate=2e-5,
... per_device_train_batch_size=16,
... per_device_eval_batch_size=16,
... num_train_epochs=3,
... weight_decay=0.01,
... push_to_hub=True,
... )
>>> trainer = Trainer(
... model=model,
... args=training_args,
... train_dataset=tokenized_squad["train"],
... eval_dataset=tokenized_squad["test"],
... tokenizer=tokenizer,
... data_collator=data_collator,
... )
>>> 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 create_optimizer
>>> batch_size = 16
>>> num_epochs = 2
>>> total_train_steps = (len(tokenized_squad["train"]) // batch_size) * num_epochs
>>> optimizer, schedule = create_optimizer(
... init_lr=2e-5,
... num_warmup_steps=0,
... num_train_steps=total_train_steps,
... )
```
Then you can load DistilBERT with [`TFAutoModelForQuestionAnswering`]:
```py
>>> from transformers import TFAutoModelForQuestionAnswering
>>> model = TFAutoModelForQuestionAnswering.from_pretrained("distilbert/distilbert-base-uncased")
```
Convert your datasets to the `tf.data.Dataset` format with [`~transformers.TFPreTrainedModel.prepare_tf_dataset`]:
```py
>>> tf_train_set = model.prepare_tf_dataset(
... tokenized_squad["train"],
... shuffle=True,
... batch_size=16,
... collate_fn=data_collator,
... )
>>> tf_validation_set = model.prepare_tf_dataset(
... tokenized_squad["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):
```py
>>> import tensorflow as tf
>>> model.compile(optimizer=optimizer)
```
The last thing to setup before you start training is to provide a way to push your model to the Hub. This can be done by specifying where to push your model and tokenizer in the [`~transformers.PushToHubCallback`]:
```py
>>> from transformers.keras_callbacks import PushToHubCallback
>>> callback = PushToHubCallback(
... output_dir="my_awesome_qa_model",
... tokenizer=tokenizer,
... )
```
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 callback to finetune the model:
```py
>>> model.fit(x=tf_train_set, validation_data=tf_validation_set, epochs=3, callbacks=[callback])
```
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 question answering, take a look at the corresponding
[PyTorch notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/question_answering.ipynb)
or [TensorFlow notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/question_answering-tf.ipynb).
</Tip>
## Evaluate
Evaluation for question answering requires a significant amount of postprocessing. To avoid taking up too much of your time, this guide skips the evaluation step. The [`Trainer`] still calculates the evaluation loss during training so you're not completely in the dark about your model's performance.
If have more time and you're interested in how to evaluate your model for question answering, take a look at the [Question answering](https://huggingface.co/course/chapter7/7?fw=pt#post-processing) chapter from the 🤗 Hugging Face Course!
## Inference
Great, now that you've finetuned a model, you can use it for inference!
Come up with a question and some context you'd like the model to predict:
```py
>>> question = "How many programming languages does BLOOM support?"
>>> context = "BLOOM has 176 billion parameters and can generate text in 46 languages natural languages and 13 programming languages."
```
The simplest way to try out your finetuned model for inference is to use it in a [`pipeline`]. Instantiate a `pipeline` for question answering with your model, and pass your text to it:
```py
>>> from transformers import pipeline
>>> question_answerer = pipeline("question-answering", model="my_awesome_qa_model")
>>> question_answerer(question=question, context=context)
{'score': 0.2058267742395401,
'start': 10,
'end': 95,
'answer': '176 billion parameters and can generate text in 46 languages natural languages and 13'}
```
You can also manually replicate the results of the `pipeline` if you'd like:
<frameworkcontent>
<pt>
Tokenize the text and return PyTorch tensors:
```py
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("my_awesome_qa_model")
>>> inputs = tokenizer(question, context, return_tensors="pt")
```
Pass your inputs to the model and return the `logits`:
```py
>>> import torch
>>> from transformers import AutoModelForQuestionAnswering
>>> model = AutoModelForQuestionAnswering.from_pretrained("my_awesome_qa_model")
>>> with torch.no_grad():
... outputs = model(**inputs)
```
Get the highest probability from the model output for the start and end positions:
```py
>>> answer_start_index = outputs.start_logits.argmax()
>>> answer_end_index = outputs.end_logits.argmax()
```
Decode the predicted tokens to get the answer:
```py
>>> predict_answer_tokens = inputs.input_ids[0, answer_start_index : answer_end_index + 1]
>>> tokenizer.decode(predict_answer_tokens)
'176 billion parameters and can generate text in 46 languages natural languages and 13'
```
</pt>
<tf>
Tokenize the text and return TensorFlow tensors:
```py
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("my_awesome_qa_model")
>>> inputs = tokenizer(question, text, return_tensors="tf")
```
Pass your inputs to the model and return the `logits`:
```py
>>> from transformers import TFAutoModelForQuestionAnswering
>>> model = TFAutoModelForQuestionAnswering.from_pretrained("my_awesome_qa_model")
>>> outputs = model(**inputs)
```
Get the highest probability from the model output for the start and end positions:
```py
>>> answer_start_index = int(tf.math.argmax(outputs.start_logits, axis=-1)[0])
>>> answer_end_index = int(tf.math.argmax(outputs.end_logits, axis=-1)[0])
```
Decode the predicted tokens to get the answer:
```py
>>> predict_answer_tokens = inputs.input_ids[0, answer_start_index : answer_end_index + 1]
>>> tokenizer.decode(predict_answer_tokens)
'176 billion parameters and can generate text in 46 languages natural languages and 13'
```
</tf>
</frameworkcontent>
|
transformers/docs/source/en/tasks/question_answering.md/0
|
{
"file_path": "transformers/docs/source/en/tasks/question_answering.md",
"repo_id": "transformers",
"token_count": 5079
}
| 287
|
<!--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.
-->
# Summary of the tokenizers
[[open-in-colab]]
On this page, we will have a closer look at tokenization.
<Youtube id="VFp38yj8h3A"/>
As we saw in [the preprocessing tutorial](preprocessing), tokenizing a text is splitting it into words or
subwords, which then are converted to ids through a look-up table. Converting words or subwords to ids is
straightforward, so in this summary, we will focus on splitting a text into words or subwords (i.e. tokenizing a text).
More specifically, we will look at the three main types of tokenizers used in 🤗 Transformers: [Byte-Pair Encoding
(BPE)](#byte-pair-encoding), [WordPiece](#wordpiece), and [SentencePiece](#sentencepiece), and show examples
of which tokenizer type is used by which model.
Note that on each model page, you can look at the documentation of the associated tokenizer to know which tokenizer
type was used by the pretrained model. For instance, if we look at [`BertTokenizer`], we can see
that the model uses [WordPiece](#wordpiece).
## Introduction
Splitting a text into smaller chunks is a task that is harder than it looks, and there are multiple ways of doing so.
For instance, let's look at the sentence `"Don't you love 🤗 Transformers? We sure do."`
<Youtube id="nhJxYji1aho"/>
A simple way of tokenizing this text is to split it by spaces, which would give:
```
["Don't", "you", "love", "🤗", "Transformers?", "We", "sure", "do."]
```
This is a sensible first step, but if we look at the tokens `"Transformers?"` and `"do."`, we notice that the
punctuation is attached to the words `"Transformer"` and `"do"`, which is suboptimal. We should take the
punctuation into account so that a model does not have to learn a different representation of a word and every possible
punctuation symbol that could follow it, which would explode the number of representations the model has to learn.
Taking punctuation into account, tokenizing our exemplary text would give:
```
["Don", "'", "t", "you", "love", "🤗", "Transformers", "?", "We", "sure", "do", "."]
```
Better. However, it is disadvantageous, how the tokenization dealt with the word `"Don't"`. `"Don't"` stands for
`"do not"`, so it would be better tokenized as `["Do", "n't"]`. This is where things start getting complicated, and
part of the reason each model has its own tokenizer type. Depending on the rules we apply for tokenizing a text, a
different tokenized output is generated for the same text. A pretrained model only performs properly if you feed it an
input that was tokenized with the same rules that were used to tokenize its training data.
[spaCy](https://spacy.io/) and [Moses](http://www.statmt.org/moses/?n=Development.GetStarted) are two popular
rule-based tokenizers. Applying them on our example, *spaCy* and *Moses* would output something like:
```
["Do", "n't", "you", "love", "🤗", "Transformers", "?", "We", "sure", "do", "."]
```
As can be seen space and punctuation tokenization, as well as rule-based tokenization, is used here. Space and
punctuation tokenization and rule-based tokenization are both examples of word tokenization, which is loosely defined
as splitting sentences into words. While it's the most intuitive way to split texts into smaller chunks, this
tokenization method can lead to problems for massive text corpora. In this case, space and punctuation tokenization
usually generates a very big vocabulary (the set of all unique words and tokens used). *E.g.*, [Transformer XL](model_doc/transfo-xl) uses space and punctuation tokenization, resulting in a vocabulary size of 267,735!
Such a big vocabulary size forces the model to have an enormous embedding matrix as the input and output layer, which
causes both an increased memory and time complexity. In general, transformers models rarely have a vocabulary size
greater than 50,000, especially if they are pretrained only on a single language.
So if simple space and punctuation tokenization is unsatisfactory, why not simply tokenize on characters?
<Youtube id="ssLq_EK2jLE"/>
While character tokenization is very simple and would greatly reduce memory and time complexity it makes it much harder
for the model to learn meaningful input representations. *E.g.* learning a meaningful context-independent
representation for the letter `"t"` is much harder than learning a context-independent representation for the word
`"today"`. Therefore, character tokenization is often accompanied by a loss of performance. So to get the best of
both worlds, transformers models use a hybrid between word-level and character-level tokenization called **subword**
tokenization.
## Subword tokenization
<Youtube id="zHvTiHr506c"/>
Subword tokenization algorithms rely on the principle that frequently used words should not be split into smaller
subwords, but rare words should be decomposed into meaningful subwords. For instance `"annoyingly"` might be
considered a rare word and could be decomposed into `"annoying"` and `"ly"`. Both `"annoying"` and `"ly"` as
stand-alone subwords would appear more frequently while at the same time the meaning of `"annoyingly"` is kept by the
composite meaning of `"annoying"` and `"ly"`. This is especially useful in agglutinative languages such as Turkish,
where you can form (almost) arbitrarily long complex words by stringing together subwords.
Subword tokenization allows the model to have a reasonable vocabulary size while being able to learn meaningful
context-independent representations. In addition, subword tokenization enables the model to process words it has never
seen before, by decomposing them into known subwords. For instance, the [`~transformers.BertTokenizer`] tokenizes
`"I have a new GPU!"` as follows:
```py
>>> from transformers import BertTokenizer
>>> tokenizer = BertTokenizer.from_pretrained("google-bert/bert-base-uncased")
>>> tokenizer.tokenize("I have a new GPU!")
["i", "have", "a", "new", "gp", "##u", "!"]
```
Because we are considering the uncased model, the sentence was lowercased first. We can see that the words `["i", "have", "a", "new"]` are present in the tokenizer's vocabulary, but the word `"gpu"` is not. Consequently, the
tokenizer splits `"gpu"` into known subwords: `["gp" and "##u"]`. `"##"` means that the rest of the token should
be attached to the previous one, without space (for decoding or reversal of the tokenization).
As another example, [`~transformers.XLNetTokenizer`] tokenizes our previously exemplary text as follows:
```py
>>> from transformers import XLNetTokenizer
>>> tokenizer = XLNetTokenizer.from_pretrained("xlnet/xlnet-base-cased")
>>> tokenizer.tokenize("Don't you love 🤗 Transformers? We sure do.")
["▁Don", "'", "t", "▁you", "▁love", "▁", "🤗", "▁", "Transform", "ers", "?", "▁We", "▁sure", "▁do", "."]
```
We'll get back to the meaning of those `"▁"` when we look at [SentencePiece](#sentencepiece). As one can see,
the rare word `"Transformers"` has been split into the more frequent subwords `"Transform"` and `"ers"`.
Let's now look at how the different subword tokenization algorithms work. Note that all of those tokenization
algorithms rely on some form of training which is usually done on the corpus the corresponding model will be trained
on.
<a id='byte-pair-encoding'></a>
### Byte-Pair Encoding (BPE)
Byte-Pair Encoding (BPE) was introduced in [Neural Machine Translation of Rare Words with Subword Units (Sennrich et
al., 2015)](https://arxiv.org/abs/1508.07909). BPE relies on a pre-tokenizer that splits the training data into
words. Pretokenization can be as simple as space tokenization, e.g. [GPT-2](model_doc/gpt2), [RoBERTa](model_doc/roberta). More advanced pre-tokenization include rule-based tokenization, e.g. [XLM](model_doc/xlm),
[FlauBERT](model_doc/flaubert) which uses Moses for most languages, or [GPT](model_doc/openai-gpt) which uses
spaCy and ftfy, to count the frequency of each word in the training corpus.
After pre-tokenization, a set of unique words has been created and the frequency with which each word occurred in the
training data has been determined. Next, BPE creates a base vocabulary consisting of all symbols that occur in the set
of unique words and learns merge rules to form a new symbol from two symbols of the base vocabulary. It does so until
the vocabulary has attained the desired vocabulary size. Note that the desired vocabulary size is a hyperparameter to
define before training the tokenizer.
As an example, let's assume that after pre-tokenization, the following set of words including their frequency has been
determined:
```
("hug", 10), ("pug", 5), ("pun", 12), ("bun", 4), ("hugs", 5)
```
Consequently, the base vocabulary is `["b", "g", "h", "n", "p", "s", "u"]`. Splitting all words into symbols of the
base vocabulary, we obtain:
```
("h" "u" "g", 10), ("p" "u" "g", 5), ("p" "u" "n", 12), ("b" "u" "n", 4), ("h" "u" "g" "s", 5)
```
BPE then counts the frequency of each possible symbol pair and picks the symbol pair that occurs most frequently. In
the example above `"h"` followed by `"u"` is present _10 + 5 = 15_ times (10 times in the 10 occurrences of
`"hug"`, 5 times in the 5 occurrences of `"hugs"`). However, the most frequent symbol pair is `"u"` followed by
`"g"`, occurring _10 + 5 + 5 = 20_ times in total. Thus, the first merge rule the tokenizer learns is to group all
`"u"` symbols followed by a `"g"` symbol together. Next, `"ug"` is added to the vocabulary. The set of words then
becomes
```
("h" "ug", 10), ("p" "ug", 5), ("p" "u" "n", 12), ("b" "u" "n", 4), ("h" "ug" "s", 5)
```
BPE then identifies the next most common symbol pair. It's `"u"` followed by `"n"`, which occurs 16 times. `"u"`,
`"n"` is merged to `"un"` and added to the vocabulary. The next most frequent symbol pair is `"h"` followed by
`"ug"`, occurring 15 times. Again the pair is merged and `"hug"` can be added to the vocabulary.
At this stage, the vocabulary is `["b", "g", "h", "n", "p", "s", "u", "ug", "un", "hug"]` and our set of unique words
is represented as
```
("hug", 10), ("p" "ug", 5), ("p" "un", 12), ("b" "un", 4), ("hug" "s", 5)
```
Assuming, that the Byte-Pair Encoding training would stop at this point, the learned merge rules would then be applied
to new words (as long as those new words do not include symbols that were not in the base vocabulary). For instance,
the word `"bug"` would be tokenized to `["b", "ug"]` but `"mug"` would be tokenized as `["<unk>", "ug"]` since
the symbol `"m"` is not in the base vocabulary. In general, single letters such as `"m"` are not replaced by the
`"<unk>"` symbol because the training data usually includes at least one occurrence of each letter, but it is likely
to happen for very special characters like emojis.
As mentioned earlier, the vocabulary size, *i.e.* the base vocabulary size + the number of merges, is a hyperparameter
to choose. For instance [GPT](model_doc/openai-gpt) has a vocabulary size of 40,478 since they have 478 base characters
and chose to stop training after 40,000 merges.
#### Byte-level BPE
A base vocabulary that includes all possible base characters can be quite large if *e.g.* all unicode characters are
considered as base characters. To have a better base vocabulary, [GPT-2](https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf) uses bytes
as the base vocabulary, which is a clever trick to force the base vocabulary to be of size 256 while ensuring that
every base character is included in the vocabulary. With some additional rules to deal with punctuation, the GPT2's
tokenizer can tokenize every text without the need for the <unk> symbol. [GPT-2](model_doc/gpt) has a vocabulary
size of 50,257, which corresponds to the 256 bytes base tokens, a special end-of-text token and the symbols learned
with 50,000 merges.
<a id='wordpiece'></a>
### WordPiece
WordPiece is the subword tokenization algorithm used for [BERT](model_doc/bert), [DistilBERT](model_doc/distilbert), and [Electra](model_doc/electra). The algorithm was outlined in [Japanese and Korean
Voice Search (Schuster et al., 2012)](https://static.googleusercontent.com/media/research.google.com/ja//pubs/archive/37842.pdf) and is very similar to
BPE. WordPiece first initializes the vocabulary to include every character present in the training data and
progressively learns a given number of merge rules. In contrast to BPE, WordPiece does not choose the most frequent
symbol pair, but the one that maximizes the likelihood of the training data once added to the vocabulary.
So what does this mean exactly? Referring to the previous example, maximizing the likelihood of the training data is
equivalent to finding the symbol pair, whose probability divided by the probabilities of its first symbol followed by
its second symbol is the greatest among all symbol pairs. *E.g.* `"u"`, followed by `"g"` would have only been
merged if the probability of `"ug"` divided by `"u"`, `"g"` would have been greater than for any other symbol
pair. Intuitively, WordPiece is slightly different to BPE in that it evaluates what it _loses_ by merging two symbols
to ensure it's _worth it_.
<a id='unigram'></a>
### Unigram
Unigram is a subword tokenization algorithm introduced in [Subword Regularization: Improving Neural Network Translation
Models with Multiple Subword Candidates (Kudo, 2018)](https://arxiv.org/pdf/1804.10959.pdf). In contrast to BPE or
WordPiece, Unigram initializes its base vocabulary to a large number of symbols and progressively trims down each
symbol to obtain a smaller vocabulary. The base vocabulary could for instance correspond to all pre-tokenized words and
the most common substrings. Unigram is not used directly for any of the models in the transformers, but it's used in
conjunction with [SentencePiece](#sentencepiece).
At each training step, the Unigram algorithm defines a loss (often defined as the log-likelihood) over the training
data given the current vocabulary and a unigram language model. Then, for each symbol in the vocabulary, the algorithm
computes how much the overall loss would increase if the symbol was to be removed from the vocabulary. Unigram then
removes p (with p usually being 10% or 20%) percent of the symbols whose loss increase is the lowest, *i.e.* those
symbols that least affect the overall loss over the training data. This process is repeated until the vocabulary has
reached the desired size. The Unigram algorithm always keeps the base characters so that any word can be tokenized.
Because Unigram is not based on merge rules (in contrast to BPE and WordPiece), the algorithm has several ways of
tokenizing new text after training. As an example, if a trained Unigram tokenizer exhibits the vocabulary:
```
["b", "g", "h", "n", "p", "s", "u", "ug", "un", "hug"],
```
`"hugs"` could be tokenized both as `["hug", "s"]`, `["h", "ug", "s"]` or `["h", "u", "g", "s"]`. So which one
to choose? Unigram saves the probability of each token in the training corpus on top of saving the vocabulary so that
the probability of each possible tokenization can be computed after training. The algorithm simply picks the most
likely tokenization in practice, but also offers the possibility to sample a possible tokenization according to their
probabilities.
Those probabilities are defined by the loss the tokenizer is trained on. Assuming that the training data consists of
the words \\(x_{1}, \dots, x_{N}\\) and that the set of all possible tokenizations for a word \\(x_{i}\\) is
defined as \\(S(x_{i})\\), then the overall loss is defined as
$$\mathcal{L} = -\sum_{i=1}^{N} \log \left ( \sum_{x \in S(x_{i})} p(x) \right )$$
<a id='sentencepiece'></a>
### SentencePiece
All tokenization algorithms described so far have the same problem: It is assumed that the input text uses spaces to
separate words. However, not all languages use spaces to separate words. One possible solution is to use language
specific pre-tokenizers, *e.g.* [XLM](model_doc/xlm) uses a specific Chinese, Japanese, and Thai pre-tokenizer.
To solve this problem more generally, [SentencePiece: A simple and language independent subword tokenizer and
detokenizer for Neural Text Processing (Kudo et al., 2018)](https://arxiv.org/pdf/1808.06226.pdf) treats the input
as a raw input stream, thus including the space in the set of characters to use. It then uses the BPE or unigram
algorithm to construct the appropriate vocabulary.
The [`XLNetTokenizer`] uses SentencePiece for example, which is also why in the example earlier the
`"▁"` character was included in the vocabulary. Decoding with SentencePiece is very easy since all tokens can just be
concatenated and `"▁"` is replaced by a space.
All transformers models in the library that use SentencePiece use it in combination with unigram. Examples of models
using SentencePiece are [ALBERT](model_doc/albert), [XLNet](model_doc/xlnet), [Marian](model_doc/marian), and [T5](model_doc/t5).
|
transformers/docs/source/en/tokenizer_summary.md/0
|
{
"file_path": "transformers/docs/source/en/tokenizer_summary.md",
"repo_id": "transformers",
"token_count": 4940
}
| 288
|
<!--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.
-->
# Compartir modelos personalizados
La biblioteca 🤗 Transformers está diseñada para ser fácilmente ampliable. Cada modelo está completamente codificado
sin abstracción en una subcarpeta determinada del repositorio, por lo que puedes copiar fácilmente un archivo del modelo
y ajustarlo según tus necesidades.
Si estás escribiendo un modelo completamente nuevo, podría ser más fácil comenzar desde cero. En este tutorial, te mostraremos
cómo escribir un modelo personalizado y su configuración para que pueda usarse dentro de Transformers, y cómo puedes compartirlo
con la comunidad (con el código en el que se basa) para que cualquiera pueda usarlo, incluso si no está presente en la biblioteca
🤗 Transformers.
Ilustraremos todo esto con un modelo ResNet, envolviendo la clase ResNet de la [biblioteca timm](https://github.com/rwightman/pytorch-image-models) en un [`PreTrainedModel`].
## Escribir una configuración personalizada
Antes de adentrarnos en el modelo, primero escribamos su configuración. La configuración de un modelo es un objeto que
contendrá toda la información necesaria para construir el modelo. Como veremos en la siguiente sección, el modelo solo puede
tomar un `config` para ser inicializado, por lo que realmente necesitamos que ese objeto esté lo más completo posible.
En nuestro ejemplo, tomaremos un par de argumentos de la clase ResNet que tal vez queramos modificar. Las diferentes
configuraciones nos darán los diferentes tipos de ResNet que son posibles. Luego simplemente almacenamos esos argumentos
después de verificar la validez de algunos de ellos.
```python
from transformers import PretrainedConfig
from typing import List
class ResnetConfig(PretrainedConfig):
model_type = "resnet"
def __init__(
self,
block_type="bottleneck",
layers: List[int] = [3, 4, 6, 3],
num_classes: int = 1000,
input_channels: int = 3,
cardinality: int = 1,
base_width: int = 64,
stem_width: int = 64,
stem_type: str = "",
avg_down: bool = False,
**kwargs,
):
if block_type not in ["basic", "bottleneck"]:
raise ValueError(f"`block_type` must be 'basic' or bottleneck', got {block_type}.")
if stem_type not in ["", "deep", "deep-tiered"]:
raise ValueError(f"`stem_type` must be '', 'deep' or 'deep-tiered', got {stem_type}.")
self.block_type = block_type
self.layers = layers
self.num_classes = num_classes
self.input_channels = input_channels
self.cardinality = cardinality
self.base_width = base_width
self.stem_width = stem_width
self.stem_type = stem_type
self.avg_down = avg_down
super().__init__(**kwargs)
```
Las tres cosas importantes que debes recordar al escribir tu propia configuración son las siguientes:
- tienes que heredar de `PretrainedConfig`,
- el `__init__` de tu `PretrainedConfig` debe aceptar cualquier `kwargs`,
- esos `kwargs` deben pasarse a la superclase `__init__`.
La herencia es para asegurarte de obtener toda la funcionalidad de la biblioteca 🤗 Transformers, mientras que las otras dos
restricciones provienen del hecho de que una `PretrainedConfig` tiene más campos que los que estás configurando. Al recargar una
`config` con el método `from_pretrained`, esos campos deben ser aceptados por tu `config` y luego enviados a la superclase.
Definir un `model_type` para tu configuración (en este caso `model_type="resnet"`) no es obligatorio, a menos que quieras
registrar tu modelo con las clases automáticas (ver la última sección).
Una vez hecho esto, puedes crear y guardar fácilmente tu configuración como lo harías con cualquier otra configuración de un
modelo de la biblioteca. Así es como podemos crear una configuración resnet50d y guardarla:
```py
resnet50d_config = ResnetConfig(block_type="bottleneck", stem_width=32, stem_type="deep", avg_down=True)
resnet50d_config.save_pretrained("custom-resnet")
```
Esto guardará un archivo llamado `config.json` dentro de la carpeta `custom-resnet`. Luego puedes volver a cargar tu configuración
con el método `from_pretrained`:
```py
resnet50d_config = ResnetConfig.from_pretrained("custom-resnet")
```
También puedes usar cualquier otro método de la clase [`PretrainedConfig`], como [`~PretrainedConfig.push_to_hub`], para cargar
directamente tu configuración en el Hub.
## Escribir un modelo personalizado
Ahora que tenemos nuestra configuración de ResNet, podemos seguir escribiendo el modelo. En realidad escribiremos dos: una que
extrae las características ocultas de un grupo de imágenes (como [`BertModel`]) y una que es adecuada para clasificación de
imagenes (como [`BertForSequenceClassification`]).
Como mencionamos antes, solo escribiremos un envoltura (_wrapper_) libre del modelo para simplificar este ejemplo. Lo único que debemos
hacer antes de escribir esta clase es un mapeo entre los tipos de bloques y las clases de bloques reales. Luego se define el
modelo desde la configuración pasando todo a la clase `ResNet`:
```py
from transformers import PreTrainedModel
from timm.models.resnet import BasicBlock, Bottleneck, ResNet
from .configuration_resnet import ResnetConfig
BLOCK_MAPPING = {"basic": BasicBlock, "bottleneck": Bottleneck}
class ResnetModel(PreTrainedModel):
config_class = ResnetConfig
def __init__(self, config):
super().__init__(config)
block_layer = BLOCK_MAPPING[config.block_type]
self.model = ResNet(
block_layer,
config.layers,
num_classes=config.num_classes,
in_chans=config.input_channels,
cardinality=config.cardinality,
base_width=config.base_width,
stem_width=config.stem_width,
stem_type=config.stem_type,
avg_down=config.avg_down,
)
def forward(self, tensor):
return self.model.forward_features(tensor)
```
Para el modelo que clasificará las imágenes, solo cambiamos el método de avance (es decir, el método `forward`):
```py
import torch
class ResnetModelForImageClassification(PreTrainedModel):
config_class = ResnetConfig
def __init__(self, config):
super().__init__(config)
block_layer = BLOCK_MAPPING[config.block_type]
self.model = ResNet(
block_layer,
config.layers,
num_classes=config.num_classes,
in_chans=config.input_channels,
cardinality=config.cardinality,
base_width=config.base_width,
stem_width=config.stem_width,
stem_type=config.stem_type,
avg_down=config.avg_down,
)
def forward(self, tensor, labels=None):
logits = self.model(tensor)
if labels is not None:
loss = torch.nn.functional.cross_entropy(logits, labels)
return {"loss": loss, "logits": logits}
return {"logits": logits}
```
En ambos casos, observa cómo heredamos de `PreTrainedModel` y llamamos a la inicialización de la superclase con `config`
(un poco como cuando escribes `torch.nn.Module`). La línea que establece `config_class` no es obligatoria, a menos
que quieras registrar tu modelo con las clases automáticas (consulta la última sección).
<Tip>
Si tu modelo es muy similar a un modelo dentro de la biblioteca, puedes reutilizar la misma configuración de ese modelo.
</Tip>
Puedes hacer que tu modelo devuelva lo que quieras, pero devolver un diccionario como lo hicimos para
`ResnetModelForImageClassification`, con el `loss` incluido cuando se pasan las etiquetas, hará que tu modelo se pueda
usar directamente dentro de la clase [`Trainer`]. Usar otro formato de salida está bien, siempre y cuando estés planeando usar
tu propio bucle de entrenamiento u otra biblioteca para el entrenamiento.
Ahora que tenemos nuestra clase, vamos a crear un modelo:
```py
resnet50d = ResnetModelForImageClassification(resnet50d_config)
```
Nuevamente, puedes usar cualquiera de los métodos de [`PreTrainedModel`], como [`~PreTrainedModel.save_pretrained`] o
[`~PreTrainedModel.push_to_hub`]. Usaremos el segundo en la siguiente sección y veremos cómo pasar los pesos del modelo
con el código de nuestro modelo. Pero primero, carguemos algunos pesos previamente entrenados dentro de nuestro modelo.
En tu caso de uso, probablemente estarás entrenando tu modelo personalizado con tus propios datos. Para ir rápido en este
tutorial, usaremos la versión preentrenada de resnet50d. Dado que nuestro modelo es solo un envoltorio alrededor del resnet50d
original, será fácil transferir esos pesos:
```py
import timm
pretrained_model = timm.create_model("resnet50d", pretrained=True)
resnet50d.model.load_state_dict(pretrained_model.state_dict())
```
Ahora veamos cómo asegurarnos de que cuando hacemos [`~PreTrainedModel.save_pretrained`] o [`~PreTrainedModel.push_to_hub`],
se guarda el código del modelo.
## Enviar el código al _Hub_
<Tip warning={true}>
Esta _API_ es experimental y puede tener algunos cambios leves en las próximas versiones.
</Tip>
Primero, asegúrate de que tu modelo esté completamente definido en un archivo `.py`. Puedes basarte en importaciones
relativas a otros archivos, siempre que todos los archivos estén en el mismo directorio (aún no admitimos submódulos
para esta característica). Para nuestro ejemplo, definiremos un archivo `modeling_resnet.py` y un archivo
`configuration_resnet.py` en una carpeta del directorio de trabajo actual llamado `resnet_model`. El archivo de configuración
contiene el código de `ResnetConfig` y el archivo del modelo contiene el código de `ResnetModel` y
`ResnetModelForImageClassification`.
```
.
└── resnet_model
├── __init__.py
├── configuration_resnet.py
└── modeling_resnet.py
```
El `__init__.py` puede estar vacío, solo está ahí para que Python detecte que `resnet_model` se puede usar como un módulo.
<Tip warning={true}>
Si copias archivos del modelo desde la biblioteca, deberás reemplazar todas las importaciones relativas en la parte superior
del archivo para importarlos desde el paquete `transformers`.
</Tip>
Ten en cuenta que puedes reutilizar (o subclasificar) una configuración o modelo existente.
Para compartir tu modelo con la comunidad, sigue estos pasos: primero importa el modelo y la configuración de ResNet desde
los archivos recién creados:
```py
from resnet_model.configuration_resnet import ResnetConfig
from resnet_model.modeling_resnet import ResnetModel, ResnetModelForImageClassification
```
Luego, debes decirle a la biblioteca que deseas copiar el código de esos objetos cuando usas el método `save_pretrained`
y registrarlos correctamente con una determinada clase automática (especialmente para modelos), simplemente ejecuta:
```py
ResnetConfig.register_for_auto_class()
ResnetModel.register_for_auto_class("AutoModel")
ResnetModelForImageClassification.register_for_auto_class("AutoModelForImageClassification")
```
Ten en cuenta que no es necesario especificar una clase automática para la configuración (solo hay una clase automática
para ellos, [`AutoConfig`]), pero es diferente para los modelos. Tu modelo personalizado podría ser adecuado para muchas
tareas diferentes, por lo que debes especificar cuál de las clases automáticas es la correcta para tu modelo.
A continuación, vamos a crear la configuración y los modelos como lo hicimos antes:
```py
resnet50d_config = ResnetConfig(block_type="bottleneck", stem_width=32, stem_type="deep", avg_down=True)
resnet50d = ResnetModelForImageClassification(resnet50d_config)
pretrained_model = timm.create_model("resnet50d", pretrained=True)
resnet50d.model.load_state_dict(pretrained_model.state_dict())
```
Ahora, para enviar el modelo al Hub, asegúrate de haber iniciado sesión. Ejecuta en tu terminal:
```bash
huggingface-cli login
```
o desde un _notebook_:
```py
from huggingface_hub import notebook_login
notebook_login()
```
Luego puedes ingresar a tu propio espacio (o una organización de la que seas miembro) de esta manera:
```py
resnet50d.push_to_hub("custom-resnet50d")
```
Además de los pesos del modelo y la configuración en formato json, esto también copió los archivos `.py` del modelo y la
configuración en la carpeta `custom-resnet50d` y subió el resultado al Hub. Puedes verificar el resultado en este
[repositorio de modelos](https://huggingface.co/sgugger/custom-resnet50d).
Consulta el tutorial sobre cómo [compartir modelos](model_sharing) para obtener más información sobre el método para subir modelos al Hub.
## Usar un modelo con código personalizado
Puedes usar cualquier configuración, modelo o _tokenizador_ con archivos de código personalizado en tu repositorio con las
clases automáticas y el método `from_pretrained`. Todos los archivos y códigos cargados en el Hub se analizan en busca de
malware (consulta la documentación de [seguridad del Hub](https://huggingface.co/docs/hub/security#malware-scanning) para
obtener más información), pero aún debes revisar el código del modelo y el autor para evitar la ejecución de código malicioso
en tu computadora. Configura `trust_remote_code=True` para usar un modelo con código personalizado:
```py
from transformers import AutoModelForImageClassification
model = AutoModelForImageClassification.from_pretrained("sgugger/custom-resnet50d", trust_remote_code=True)
```
También se recomienda encarecidamente pasar un _hash_ de confirmación como una "revisión" para asegurarte de que el autor
de los modelos no actualizó el código con algunas líneas nuevas maliciosas (a menos que confíes plenamente en los autores
de los modelos).
```py
commit_hash = "ed94a7c6247d8aedce4647f00f20de6875b5b292"
model = AutoModelForImageClassification.from_pretrained(
"sgugger/custom-resnet50d", trust_remote_code=True, revision=commit_hash
)
```
Ten en cuenta que al navegar por el historial de confirmaciones del repositorio del modelo en Hub, hay un botón para copiar
fácilmente el hash de confirmación de cualquier _commit_.
## Registrar un model con código personalizado a las clases automáticas
Si estás escribiendo una biblioteca que amplía 🤗 Transformers, es posible que quieras ampliar las clases automáticas para
incluir tu propio modelo. Esto es diferente de enviar el código al Hub en el sentido de que los usuarios necesitarán importar
tu biblioteca para obtener los modelos personalizados (al contrario de descargar automáticamente el código del modelo desde Hub).
Siempre que tu configuración tenga un atributo `model_type` que sea diferente de los tipos de modelos existentes, y que tus
clases modelo tengan los atributos `config_class` correctos, puedes agregarlos a las clases automáticas de la siguiente manera:
```py
from transformers import AutoConfig, AutoModel, AutoModelForImageClassification
AutoConfig.register("resnet", ResnetConfig)
AutoModel.register(ResnetConfig, ResnetModel)
AutoModelForImageClassification.register(ResnetConfig, ResnetModelForImageClassification)
```
Ten en cuenta que el primer argumento utilizado al registrar tu configuración personalizada en [`AutoConfig`] debe coincidir
con el `model_type` de tu configuración personalizada, y el primer argumento utilizado al registrar tus modelos personalizados
en cualquier clase del modelo automático debe coincidir con el `config_class ` de esos modelos.
|
transformers/docs/source/es/custom_models.md/0
|
{
"file_path": "transformers/docs/source/es/custom_models.md",
"repo_id": "transformers",
"token_count": 5985
}
| 289
|
<!--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.
-->
# Preprocesamiento
[[open-in-colab]]
Antes de que puedas utilizar los datos en un modelo, debes procesarlos en un formato aceptable para el modelo. Un modelo no entiende el texto en bruto, las imágenes o el audio. Estas entradas necesitan ser convertidas en números y ensambladas en tensores. En este tutorial, podrás:
* Preprocesar los datos textuales con un tokenizador.
* Preprocesar datos de imagen o audio con un extractor de características.
* Preprocesar datos para una tarea multimodal con un procesador.
## NLP
<Youtube id="Yffk5aydLzg"/>
La principal herramienta para procesar datos textuales es un [tokenizador](main_classes/tokenizer). Un tokenizador comienza dividiendo el texto en *tokens* según un conjunto de reglas. Los tokens se convierten en números, que se utilizan para construir tensores como entrada a un modelo. El tokenizador también añade cualquier entrada adicional que requiera el modelo.
<Tip>
Si tienes previsto utilizar un modelo pre-entrenado, es importante que utilices el tokenizador pre-entrenado asociado. Esto te asegura que el texto se divide de la misma manera que el corpus de pre-entrenamiento y utiliza el mismo índice de tokens correspondiente (usualmente referido como el *vocab*) durante el pre-entrenamiento.
</Tip>
Comienza rápidamente cargando un tokenizador pre-entrenado con la clase [`AutoTokenizer`]. Esto descarga el *vocab* utilizado cuando un modelo es pre-entrenado.
### Tokenizar
Carga un tokenizador pre-entrenado con [`AutoTokenizer.from_pretrained`]:
```py
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-cased")
```
A continuación, pasa tu frase al tokenizador:
```py
>>> encoded_input = tokenizer("Do not meddle in the affairs of wizards, for they are subtle and quick to anger.")
>>> print(encoded_input)
{'input_ids': [101, 2079, 2025, 19960, 10362, 1999, 1996, 3821, 1997, 16657, 1010, 2005, 2027, 2024, 11259, 1998, 4248, 2000, 4963, 1012, 102],
'token_type_ids': [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}
```
El tokenizador devuelve un diccionario con tres ítems importantes:
* [input_ids](glossary#input-ids) son los índices correspondientes a cada token de la frase.
* [attention_mask](glossary#attention-mask) indica si un token debe ser atendido o no.
* [token_type_ids](glossary#token-type-ids) identifica a qué secuencia pertenece un token cuando hay más de una secuencia.
Tu puedes decodificar el `input_ids` para devolver la entrada original:
```py
>>> tokenizer.decode(encoded_input["input_ids"])
'[CLS] Do not meddle in the affairs of wizards, for they are subtle and quick to anger. [SEP]'
```
Como puedes ver, el tokenizador ha añadido dos tokens especiales - `CLS` y `SEP` (clasificador y separador) - a la frase. No todos los modelos necesitan
tokens especiales, pero si lo llegas a necesitar, el tokenizador los añadirá automáticamente.
Si hay varias frases que quieres preprocesar, pasa las frases como una lista al tokenizador:
```py
>>> batch_sentences = [
... "But what about second breakfast?",
... "Don't think he knows about second breakfast, Pip.",
... "What about elevensies?",
... ]
>>> encoded_inputs = tokenizer(batch_sentences)
>>> print(encoded_inputs)
{'input_ids': [[101, 1252, 1184, 1164, 1248, 6462, 136, 102],
[101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102],
[101, 1327, 1164, 5450, 23434, 136, 102]],
'token_type_ids': [[0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]],
'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1],
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
[1, 1, 1, 1, 1, 1, 1]]}
```
### Pad
Esto nos lleva a un tema importante. Cuando se procesa un batch de frases, no siempre tienen la misma longitud. Esto es un problema porque los tensores que se introducen en el modelo deben tener una forma uniforme. El pad es una estrategia para asegurar que los tensores sean rectangulares añadiendo un "padding token" especial a las oraciones con menos tokens.
Establece el parámetro `padding` en `True` aplicando el pad a las secuencias más cortas del batch para que coincidan con la secuencia más larga:
```py
>>> batch_sentences = [
... "But what about second breakfast?",
... "Don't think he knows about second breakfast, Pip.",
... "What about elevensies?",
... ]
>>> encoded_input = tokenizer(batch_sentences, padding=True)
>>> print(encoded_input)
{'input_ids': [[101, 1252, 1184, 1164, 1248, 6462, 136, 102, 0, 0, 0, 0, 0, 0, 0],
[101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102],
[101, 1327, 1164, 5450, 23434, 136, 102, 0, 0, 0, 0, 0, 0, 0, 0]],
'token_type_ids': [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],
'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 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, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]]}
```
Observa que el tokenizador ha aplicado el pad a la primera y la tercera frase con un "0" porque son más cortas.
### Truncamiento
En el otro extremo del espectro, a veces una secuencia puede ser demasiado larga para un modelo. En este caso, tendrás que truncar la secuencia a una longitud más corta.
Establece el parámetro `truncation` a `True` para truncar una secuencia a la longitud máxima aceptada por el modelo:
```py
>>> batch_sentences = [
... "But what about second breakfast?",
... "Don't think he knows about second breakfast, Pip.",
... "What about elevensies?",
... ]
>>> encoded_input = tokenizer(batch_sentences, padding=True, truncation=True)
>>> print(encoded_input)
{'input_ids': [[101, 1252, 1184, 1164, 1248, 6462, 136, 102, 0, 0, 0, 0, 0, 0, 0],
[101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102],
[101, 1327, 1164, 5450, 23434, 136, 102, 0, 0, 0, 0, 0, 0, 0, 0]],
'token_type_ids': [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],
'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 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, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]]}
```
### Construye tensores
Finalmente, si quieres que el tokenizador devuelva los tensores reales que se introducen en el modelo.
Establece el parámetro `return_tensors` como `pt` para PyTorch, o `tf` para TensorFlow:
```py
>>> batch_sentences = [
... "But what about second breakfast?",
... "Don't think he knows about second breakfast, Pip.",
... "What about elevensies?",
... ]
>>> encoded_input = tokenizer(batch, padding=True, truncation=True, return_tensors="pt")
>>> print(encoded_input)
{'input_ids': tensor([[ 101, 153, 7719, 21490, 1122, 1114, 9582, 1623, 102],
[ 101, 5226, 1122, 9649, 1199, 2610, 1236, 102, 0]]),
'token_type_ids': tensor([[0, 0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0, 0]]),
'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1],
[1, 1, 1, 1, 1, 1, 1, 1, 0]])}
===PT-TF-SPLIT===
>>> batch_sentences = [
... "But what about second breakfast?",
... "Don't think he knows about second breakfast, Pip.",
... "What about elevensies?",
... ]
>>> encoded_input = tokenizer(batch, padding=True, truncation=True, return_tensors="tf")
>>> print(encoded_input)
{'input_ids': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
array([[ 101, 153, 7719, 21490, 1122, 1114, 9582, 1623, 102],
[ 101, 5226, 1122, 9649, 1199, 2610, 1236, 102, 0]],
dtype=int32)>,
'token_type_ids': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
array([[0, 0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0, 0]], dtype=int32)>,
'attention_mask': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
array([[1, 1, 1, 1, 1, 1, 1, 1, 1],
[1, 1, 1, 1, 1, 1, 1, 1, 0]], dtype=int32)>}
```
## Audio
Las entradas de audio se preprocesan de forma diferente a las entradas textuales, pero el objetivo final es el mismo: crear secuencias numéricas que el modelo pueda entender. Un [extractor de características](main_classes/feature_extractor) (o feature extractor en inglés) está diseñado para extraer características de datos provenientes de imágenes o audio sin procesar y convertirlos en tensores. Antes de empezar, instala 🤗 Datasets para cargar un dataset de audio para experimentar:
```bash
pip install datasets
```
Carga la tarea de detección de palabras clave del benchmark [SUPERB](https://huggingface.co/datasets/superb) (consulta el [tutorial 🤗 Dataset](https://huggingface.co/docs/datasets/load_hub) para que obtengas más detalles sobre cómo cargar un dataset):
```py
>>> from datasets import load_dataset, Audio
>>> dataset = load_dataset("superb", "ks")
```
Accede al primer elemento de la columna `audio` para echar un vistazo a la entrada. Al llamar a la columna `audio` se cargará y volverá a muestrear automáticamente el archivo de audio:
```py
>>> dataset["train"][0]["audio"]
{'array': array([ 0. , 0. , 0. , ..., -0.00592041,
-0.00405884, -0.00253296], dtype=float32),
'path': '/root/.cache/huggingface/datasets/downloads/extracted/05734a36d88019a09725c20cc024e1c4e7982e37d7d55c0c1ca1742ea1cdd47f/_background_noise_/doing_the_dishes.wav',
'sampling_rate': 16000}
```
Esto devuelve tres elementos:
* `array` es la señal de voz cargada - y potencialmente remuestreada - como un array 1D.
* `path` apunta a la ubicación del archivo de audio.
* `sampling_rate` se refiere a cuántos puntos de datos de la señal de voz se miden por segundo.
### Resample
Para este tutorial, se utilizará el modelo [Wav2Vec2](https://huggingface.co/facebook/wav2vec2-base). Como puedes ver en la model card, el modelo Wav2Vec2 está pre-entrenado en audio de voz muestreado a 16kHz. Es importante que la tasa de muestreo de tus datos de audio coincida con la tasa de muestreo del dataset utilizado para pre-entrenar el modelo. Si la tasa de muestreo de tus datos no es la misma, deberás volver a muestrear tus datos de audio.
Por ejemplo, carga el dataset [LJ Speech](https://huggingface.co/datasets/lj_speech) que tiene una tasa de muestreo de 22050kHz. Para utilizar el modelo Wav2Vec2 con este dataset, reduce la tasa de muestreo a 16kHz:
```py
>>> lj_speech = load_dataset("lj_speech", split="train")
>>> lj_speech[0]["audio"]
{'array': array([-7.3242188e-04, -7.6293945e-04, -6.4086914e-04, ...,
7.3242188e-04, 2.1362305e-04, 6.1035156e-05], dtype=float32),
'path': '/root/.cache/huggingface/datasets/downloads/extracted/917ece08c95cf0c4115e45294e3cd0dee724a1165b7fc11798369308a465bd26/LJSpeech-1.1/wavs/LJ001-0001.wav',
'sampling_rate': 22050}
```
1. Usa el método 🤗 Datasets' [`cast_column`](https://huggingface.co/docs/datasets/package_reference/main_classes#datasets.Dataset.cast_column) para reducir la tasa de muestreo a 16kHz:
```py
>>> lj_speech = lj_speech.cast_column("audio", Audio(sampling_rate=16_000))
```
2. Carga el archivo de audio:
```py
>>> lj_speech[0]["audio"]
{'array': array([-0.00064146, -0.00074657, -0.00068768, ..., 0.00068341,
0.00014045, 0. ], dtype=float32),
'path': '/root/.cache/huggingface/datasets/downloads/extracted/917ece08c95cf0c4115e45294e3cd0dee724a1165b7fc11798369308a465bd26/LJSpeech-1.1/wavs/LJ001-0001.wav',
'sampling_rate': 16000}
```
Como puedes ver, el `sampling_rate` se ha reducido a 16kHz. Ahora que sabes cómo funciona el resampling, volvamos a nuestro ejemplo anterior con el dataset SUPERB.
### Extractor de características
El siguiente paso es cargar un extractor de características para normalizar y aplicar el pad a la entrada. Cuando se aplica padding a los datos textuales, se añade un "0" para las secuencias más cortas. La misma idea se aplica a los datos de audio y el extractor de características de audio añadirá un "0" - interpretado como silencio - al "array".
Carga el extractor de características con [`AutoFeatureExtractor.from_pretrained`]:
```py
>>> from transformers import AutoFeatureExtractor
>>> feature_extractor = AutoFeatureExtractor.from_pretrained("facebook/wav2vec2-base")
```
Pasa el `array` de audio al extractor de características. También te recomendamos añadir el argumento `sampling_rate` en el extractor de características para poder depurar mejor los errores silenciosos que puedan producirse.
```py
>>> audio_input = [dataset["train"][0]["audio"]["array"]]
>>> feature_extractor(audio_input, sampling_rate=16000)
{'input_values': [array([ 0.00045439, 0.00045439, 0.00045439, ..., -0.1578519 , -0.10807519, -0.06727459], dtype=float32)]}
```
### Pad y truncamiento
Al igual que el tokenizador, puedes aplicar padding o truncamiento para manejar secuencias variables en un batch. Fíjate en la longitud de la secuencia de estas dos muestras de audio:
```py
>>> dataset["train"][0]["audio"]["array"].shape
(1522930,)
>>> dataset["train"][1]["audio"]["array"].shape
(988891,)
```
Como puedes ver, el `sampling_rate` se ha reducido a 16kHz.
```py
>>> def preprocess_function(examples):
... audio_arrays = [x["array"] for x in examples["audio"]]
... inputs = feature_extractor(
... audio_arrays,
... sampling_rate=16000,
... padding=True,
... max_length=1000000,
... truncation=True,
... )
... return inputs
```
Aplica la función a los primeros ejemplos del dataset:
```py
>>> processed_dataset = preprocess_function(dataset["train"][:5])
```
Ahora echa un vistazo a las longitudes de las muestras procesadas:
```py
>>> processed_dataset["input_values"][0].shape
(1000000,)
>>> processed_dataset["input_values"][1].shape
(1000000,)
```
Las longitudes de las dos primeras muestras coinciden ahora con la longitud máxima especificada.
## Visión
También se utiliza un extractor de características para procesar imágenes para tareas de visión por computadora. Una vez más, el objetivo es convertir la imagen en bruto en un batch de tensores como entrada.
Vamos a cargar el dataset [food101](https://huggingface.co/datasets/food101) para este tutorial. Usa el parámetro 🤗 Datasets `split` para cargar solo una pequeña muestra de la división de entrenamiento ya que el dataset es bastante grande:
```py
>>> from datasets import load_dataset
>>> dataset = load_dataset("food101", split="train[:100]")
```
A continuación, observa la imagen con la función 🤗 Datasets [`Image`](https://huggingface.co/docs/datasets/package_reference/main_classes?highlight=image#datasets.Image):
```py
>>> dataset[0]["image"]
```
### Extractor de características
Carga el extractor de características con [`AutoFeatureExtractor.from_pretrained`]:
```py
>>> from transformers import AutoFeatureExtractor
>>> feature_extractor = AutoFeatureExtractor.from_pretrained("google/vit-base-patch16-224")
```
### Aumento de Datos
Para las tareas de visión por computadora es común añadir algún tipo de aumento de datos (o data augmentation) a las imágenes como parte del preprocesamiento. Puedes añadir el método de aumento de datos con cualquier librería que quieras, pero en este tutorial utilizarás el módulo [`transforms`](https://pytorch.org/vision/stable/transforms.html) de torchvision.
1. Normaliza la imagen y utiliza [`Compose`](https://pytorch.org/vision/master/generated/torchvision.transforms.Compose.html) para encadenar algunas transformaciones - [`RandomResizedCrop`](https://pytorch.org/vision/main/generated/torchvision.transforms.RandomResizedCrop.html) y [`ColorJitter`](https://pytorch.org/vision/main/generated/torchvision.transforms.ColorJitter.html) - juntas:
```py
>>> from torchvision.transforms import Compose, Normalize, RandomResizedCrop, ColorJitter, ToTensor
>>> normalize = Normalize(mean=feature_extractor.image_mean, std=feature_extractor.image_std)
>>> _transforms = Compose(
... [RandomResizedCrop(feature_extractor.size), ColorJitter(brightness=0.5, hue=0.5), ToTensor(), normalize]
... )
```
2. El modelo acepta [`pixel_values`](model_doc/visionencoderdecoder#transformers.VisionEncoderDecoderModel.forward.pixel_values) como entrada. Este valor es generado por el extractor de características. Crea una función que genere `pixel_values` a partir de las transformaciones:
```py
>>> def transforms(examples):
... examples["pixel_values"] = [_transforms(image.convert("RGB")) for image in examples["image"]]
... return examples
```
3. A continuación, utiliza 🤗 Datasets [`set_transform`](https://huggingface.co/docs/datasets/process#format-transform) para aplicar las transformaciones sobre la marcha:
```py
>>> dataset.set_transform(transforms)
```
4. Ahora, cuando accedes a la imagen, observarás que el extractor de características ha añadido a la entrada del modelo `pixel_values`:
```py
>>> dataset[0]["image"]
{'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=384x512 at 0x7F1A7B0630D0>,
'label': 6,
'pixel_values': tensor([[[ 0.0353, 0.0745, 0.1216, ..., -0.9922, -0.9922, -0.9922],
[-0.0196, 0.0667, 0.1294, ..., -0.9765, -0.9843, -0.9922],
[ 0.0196, 0.0824, 0.1137, ..., -0.9765, -0.9686, -0.8667],
...,
[ 0.0275, 0.0745, 0.0510, ..., -0.1137, -0.1216, -0.0824],
[ 0.0667, 0.0824, 0.0667, ..., -0.0588, -0.0745, -0.0980],
[ 0.0353, 0.0353, 0.0431, ..., -0.0039, -0.0039, -0.0588]],
[[ 0.2078, 0.2471, 0.2863, ..., -0.9451, -0.9373, -0.9451],
[ 0.1608, 0.2471, 0.3098, ..., -0.9373, -0.9451, -0.9373],
[ 0.2078, 0.2706, 0.3020, ..., -0.9608, -0.9373, -0.8275],
...,
[-0.0353, 0.0118, -0.0039, ..., -0.2392, -0.2471, -0.2078],
[ 0.0196, 0.0353, 0.0196, ..., -0.1843, -0.2000, -0.2235],
[-0.0118, -0.0039, -0.0039, ..., -0.0980, -0.0980, -0.1529]],
[[ 0.3961, 0.4431, 0.4980, ..., -0.9216, -0.9137, -0.9216],
[ 0.3569, 0.4510, 0.5216, ..., -0.9059, -0.9137, -0.9137],
[ 0.4118, 0.4745, 0.5216, ..., -0.9137, -0.8902, -0.7804],
...,
[-0.2314, -0.1922, -0.2078, ..., -0.4196, -0.4275, -0.3882],
[-0.1843, -0.1686, -0.2000, ..., -0.3647, -0.3804, -0.4039],
[-0.1922, -0.1922, -0.1922, ..., -0.2941, -0.2863, -0.3412]]])}
```
Este es el aspecto de la imagen después de preprocesarla. Como era de esperar por las transformaciones aplicadas, la imagen ha sido recortada aleatoriamente y sus propiedades de color son diferentes.
```py
>>> import numpy as np
>>> import matplotlib.pyplot as plt
>>> img = dataset[0]["pixel_values"]
>>> plt.imshow(img.permute(1, 2, 0))
```
## Multimodal
Para las tareas multimodales utilizarás una combinación de todo lo que has aprendido hasta ahora y aplicarás tus habilidades a una tarea de reconocimiento automático de voz (ASR). Esto significa que necesitarás un:
* Extractor de características para preprocesar los datos de audio.
* Un tokenizador para procesar el texto.
Volvamos al dataset [LJ Speech](https://huggingface.co/datasets/lj_speech):
```py
>>> from datasets import load_dataset
>>> lj_speech = load_dataset("lj_speech", split="train")
```
Suponiendo que te interesan principalmente las columnas `audio` y `texto`, elimina las demás columnas:
```py
>>> lj_speech = lj_speech.map(remove_columns=["file", "id", "normalized_text"])
```
Ahora echa un vistazo a las columnas `audio` y `texto`:
```py
>>> lj_speech[0]["audio"]
{'array': array([-7.3242188e-04, -7.6293945e-04, -6.4086914e-04, ...,
7.3242188e-04, 2.1362305e-04, 6.1035156e-05], dtype=float32),
'path': '/root/.cache/huggingface/datasets/downloads/extracted/917ece08c95cf0c4115e45294e3cd0dee724a1165b7fc11798369308a465bd26/LJSpeech-1.1/wavs/LJ001-0001.wav',
'sampling_rate': 22050}
>>> lj_speech[0]["text"]
'Printing, in the only sense with which we are at present concerned, differs from most if not from all the arts and crafts represented in the Exhibition'
```
Recuerda la sección anterior sobre el procesamiento de datos de audio, siempre debes [volver a muestrear](preprocessing#audio) la tasa de muestreo de tus datos de audio para que coincida con la tasa de muestreo del dataset utilizado para preentrenar un modelo:
```py
>>> lj_speech = lj_speech.cast_column("audio", Audio(sampling_rate=16_000))
```
### Processor
Un processor combina un extractor de características y un tokenizador. Cargue un procesador con [`AutoProcessor.from_pretrained`]:
```py
>>> from transformers import AutoProcessor
>>> processor = AutoProcessor.from_pretrained("facebook/wav2vec2-base-960h")
```
1. Crea una función para procesar los datos de audio en `input_values`, y tokeniza el texto en `labels`. Estas son las entradas del modelo:
```py
>>> def prepare_dataset(example):
... audio = example["audio"]
... example.update(processor(audio=audio["array"], text=example["text"], sampling_rate=16000))
... return example
```
2. Aplica la función `prepare_dataset` a una muestra:
```py
>>> prepare_dataset(lj_speech[0])
```
Observa que el método processor ha añadido `input_values` y `labels`. La tasa de muestreo también se ha reducido correctamente a 16kHz.
Genial, ahora deberías ser capaz de preprocesar datos para cualquier modalidad e incluso combinar diferentes modalidades. En el siguiente tutorial, aprenderás aplicar fine tuning a un modelo en tus datos recién preprocesados.
## Todo lo que siempre quisiste saber sobre el padding y el truncamiento
Hemos visto los comandos que funcionarán para la mayoría de los casos (hacer pad a tu batch teniendo en cuenta la longitud de la frase máxima y
truncar a la longitud máxima que el modelo puede aceptar). Sin embargo, la API admite más estrategias si las necesitas. Los
tres argumentos que necesitas conocer para ello son `padding`, `truncation` y `max_length`.
- `padding` controla el aplicarme padding al texto. Puede ser un booleano o una cadena que debe ser:
- `True` o `'longest'` para aplicar el pad hasta la secuencia más larga del batch (no apliques el padding si sólo le proporcionas
una sola secuencia).
- `'max_length'` para aplicar el pad hasta la longitud especificada por el argumento `max_length` o la longitud máxima aceptada
por el modelo si no le proporcionas `longitud_máxima` (`longitud_máxima=None`). Si sólo le proporcionas una única secuencia
se le aplicará el padding.
`False` o `'do_not_pad'` para no aplicar pad a las secuencias. Como hemos visto antes, este es el comportamiento por
defecto.
- `truncation` controla el truncamiento. Puede ser un booleano o una string que debe ser:
- `True` o `'longest_first'` truncan hasta la longitud máxima especificada por el argumento `max_length` o
la longitud máxima aceptada por el modelo si no le proporcionas `max_length` (`max_length=None`). Esto
truncará token por token, eliminando un token de la secuencia más larga del par hasta alcanzar la longitud
adecuada.
- `'only_second'` trunca hasta la longitud máxima especificada por el argumento `max_length` o la
longitud máxima aceptada por el modelo si no le proporcionas `max_length` (`max_length=None`). Esto sólo truncará
la segunda frase de un par si le proporcionas un par de secuencias (o un batch de pares de secuencias).
- `'only_first'` trunca hasta la longitud máxima especificada por el argumento `max_length` o la longitud máxima
aceptada por el modelo si no se proporciona `max_length` (`max_length=None`). Esto sólo truncará
la primera frase de un par si se proporciona un par de secuencias (o un lote de pares de secuencias).
- `False` o `'do_not_truncate'` para no truncar las secuencias. Como hemos visto antes, este es el comportamiento
por defecto.
- `max_length` para controlar la longitud del padding/truncamiento. Puede ser un número entero o `None`, en cuyo caso
será por defecto la longitud máxima que el modelo puede aceptar. Si el modelo no tiene una longitud máxima de entrada específica, el
padding/truncamiento a `longitud_máxima` se desactiva.
A continuación te mostramos en una tabla que resume la forma recomendada de configurar el padding y el truncamiento. Si utilizas un par de secuencias de entrada en
algunos de los siguientes ejemplos, puedes sustituir `truncation=True` por una `STRATEGY` seleccionada en
`['only_first', 'only_second', 'longest_first']`, es decir, `truncation='only_second'` o `truncation= 'longest_first'` para controlar cómo se truncan ambas secuencias del par como se ha detallado anteriormente.
| Truncation | Padding | Instrucciones |
|--------------------------------------|-----------------------------------|---------------------------------------------------------------------------------------------|
| no truncation | no padding | `tokenizer(batch_sentences)` |
| | padding secuencia max del batch | `tokenizer(batch_sentences, padding=True)` or |
| | | `tokenizer(batch_sentences, padding='longest')` |
| | padding long max de input model | `tokenizer(batch_sentences, padding='max_length')` |
| | padding a una long especifica | `tokenizer(batch_sentences, padding='max_length', max_length=42)` |
| truncation long max del input model | no padding | `tokenizer(batch_sentences, truncation=True)` or |
| | | `tokenizer(batch_sentences, truncation=STRATEGY)` |
| | padding secuencia max del batch | `tokenizer(batch_sentences, padding=True, truncation=True)` or |
| | | `tokenizer(batch_sentences, padding=True, truncation=STRATEGY)` |
| | padding long max de input model | `tokenizer(batch_sentences, padding='max_length', truncation=True)` or |
| | | `tokenizer(batch_sentences, padding='max_length', truncation=STRATEGY)` |
| | padding a una long especifica | Not possible |
| truncation a una long especifica | no padding | `tokenizer(batch_sentences, truncation=True, max_length=42)` or |
| | | `tokenizer(batch_sentences, truncation=STRATEGY, max_length=42)` |
| | padding secuencia max del batch | `tokenizer(batch_sentences, padding=True, truncation=True, max_length=42)` or |
| | | `tokenizer(batch_sentences, padding=True, truncation=STRATEGY, max_length=42)` |
| | padding long max de input model | Not possible |
| | padding a una long especifica | `tokenizer(batch_sentences, padding='max_length', truncation=True, max_length=42)` or |
| | | `tokenizer(batch_sentences, padding='max_length', truncation=STRATEGY, max_length=42)` |
|
transformers/docs/source/es/preprocessing.md/0
|
{
"file_path": "transformers/docs/source/es/preprocessing.md",
"repo_id": "transformers",
"token_count": 13019
}
| 290
|
<!--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.
-->
# El Trainer
El [`Trainer`] es un bucle completo de entrenamiento y evaluación para modelos de PyTorch implementado en la biblioteca Transformers. Solo necesitas pasarle las piezas necesarias para el entrenamiento (modelo, tokenizador, conjunto de datos, función de evaluación, hiperparámetros de entrenamiento, etc.), y la clase [`Trainer`] se encarga del resto. Esto facilita comenzar a entrenar más rápido sin tener que escribir manualmente tu propio bucle de entrenamiento. Pero al mismo tiempo, [`Trainer`] es muy personalizable y ofrece una gran cantidad de opciones de entrenamiento para que puedas adaptarlo a tus necesidades exactas de entrenamiento.
<Tip>
Además de la clase [`Trainer`], Transformers también proporciona una clase [`Seq2SeqTrainer`] para tareas de secuencia a secuencia como traducción o resumen. También está la clase [~trl.SFTTrainer] de la biblioteca [TRL](https://hf.co/docs/trl) que envuelve la clase [`Trainer`] y está optimizada para entrenar modelos de lenguaje como Llama-2 y Mistral con técnicas autoregresivas. [`~trl.SFTTrainer`] también admite funciones como el empaquetado de secuencias, LoRA, cuantización y DeepSpeed para escalar eficientemente a cualquier tamaño de modelo.
<br>
Siéntete libre de consultar [la referencia de API](./main_classes/trainer) para estas otras clases tipo [`Trainer`] para aprender más sobre cuándo usar cada una. En general, [`Trainer`] es la opción más versátil y es apropiada para una amplia gama de tareas. [`Seq2SeqTrainer`] está diseñado para tareas de secuencia a secuencia y [`~trl.SFTTrainer`] está diseñado para entrenar modelos de lenguaje.
</Tip>
Antes de comenzar, asegúrate de tener instalado [Accelerate](https://hf.co/docs/accelerate), una biblioteca para habilitar y ejecutar el entrenamiento de PyTorch en entornos distribuidos.
```bash
pip install accelerate
# upgrade
pip install accelerate --upgrade
```
Esta guía proporciona una visión general de la clase [`Trainer`].
## Uso básico
[`Trainer`] incluye todo el código que encontrarías en un bucle de entrenamiento básico:
1. Realiza un paso de entrenamiento para calcular la pérdida
2. Calcula los gradientes con el método [~accelerate.Accelerator.backward]
3. Actualiza los pesos basados en los gradientes
4. Repite este proceso hasta alcanzar un número predeterminado de épocas
La clase [`Trainer`] abstrae todo este código para que no tengas que preocuparte por escribir manualmente un bucle de entrenamiento cada vez o si estás empezando con PyTorch y el entrenamiento. Solo necesitas proporcionar los componentes esenciales requeridos para el entrenamiento, como un modelo y un conjunto de datos, y la clase [`Trainer`] maneja todo lo demás.
Si deseas especificar opciones de entrenamiento o hiperparámetros, puedes encontrarlos en la clase [`TrainingArguments`]. Por ejemplo, vamos a definir dónde guardar el modelo en output_dir y subir el modelo al Hub después del entrenamiento con `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,
)
```
Pase `training_args` al [`Trainer`] con un modelo, un conjunto de datos o algo para preprocesar el conjunto de datos (dependiendo en el tipo de datos pueda ser un tokenizer, extractor de caracteristicas o procesor del imagen), un recopilador de datos y una función para calcular las métricas que desea rastrear durante el entrenamiento.
Finalmente, llame [`~Trainer.train`] para empezar entrenamiento!
```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()
```
### Los puntos de control
La clase [`Trainer`] guarda los puntos de control del modelo en el directorio especificado en el parámetro `output_dir` de [`TrainingArguments`]. Encontrarás los puntos de control guardados en una subcarpeta checkpoint-000 donde los números al final corresponden al paso de entrenamiento. Guardar puntos de control es útil para reanudar el entrenamiento más tarde.
```py
# resume from latest checkpoint
trainer.train(resume_from_checkpoint=True)
# resume from specific checkpoint saved in output directory
trainer.train(resume_from_checkpoint="your-model/checkpoint-1000")
```
Puedes guardar tus puntos de control (por defecto, el estado del optimizador no se guarda) en el Hub configurando `push_to_hub=True` en [`TrainingArguments`] para confirmar y enviarlos. Otras opciones para decidir cómo se guardan tus puntos de control están configuradas en el parámetro [`hub_strategy`](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.TrainingArguments.hub_strategy):
* hub_strategy="checkpoint" envía el último punto de control a una subcarpeta llamada "last-checkpoint" desde la cual puedes reanudar el entrenamiento.
* hub_strategy="all_checkpoints" envía todos los puntos de control al directorio definido en `output_dir` (verás un punto de control por carpeta en tu repositorio de modelos).
Cuando reanudas el entrenamiento desde un punto de control, el [`Trainer`] intenta mantener los estados de los generadores de números aleatorios (RNG) de Python, NumPy y PyTorch iguales a como estaban cuando se guardó el punto de control. Pero debido a que PyTorch tiene varias configuraciones predeterminadas no determinísticas, no se garantiza que los estados de RNG sean los mismos. Si deseas habilitar la plena determinismo, echa un vistazo a la guía ["Controlling sources of randomness"](https://pytorch.org/docs/stable/notes/randomness#controlling-sources-of-randomness) para aprender qué puedes habilitar para hacer que tu entrenamiento sea completamente determinista. Sin embargo, ten en cuenta que al hacer ciertas configuraciones deterministas, el entrenamiento puede ser más lento.
## Personaliza el Trainer
Si bien la clase [`Trainer`] está diseñada para ser accesible y fácil de usar, también ofrece mucha capacidad de personalización para usuarios más aventureros. Muchos de los métodos del [`Trainer`] pueden ser subclasificados y sobrescritos para admitir la funcionalidad que deseas, sin tener que reescribir todo el bucle de entrenamiento desde cero para adaptarlo. Estos métodos incluyen:
* [~Trainer.get_train_dataloader] crea un entrenamiento de DataLoader
* [~Trainer.get_eval_dataloader] crea una evaluación DataLoader
* [~Trainer.get_test_dataloader] crea una prueba de DataLoader
* [~Trainer.log] anota la información de los objetos varios que observa el entrenamiento
* [~Trainer.create_optimizer_and_scheduler] crea un optimizador y la tasa programada de aprendizaje si no lo pasaron en __init__; estos pueden ser personalizados independientes con [~Trainer.create_optimizer] y [~Trainer.create_scheduler] respectivamente
* [~Trainer.compute_loss] computa la pérdida en lote con las aportes del entrenamiento
* [~Trainer.training_step] realiza el paso del entrenamiento
* [~Trainer.prediction_step] realiza la predicción y paso de prueba
* [~Trainer.evaluate] evalua el modelo y da las metricas evaluativas
* [~Trainer.predict] hace las predicciones (con las metricas si hay etiquetas disponibles) en lote de prueba
Por ejemplo, si deseas personalizar el método [`~Trainer.compute_loss`] para usar una pérdida ponderada en su lugar, puedes hacerlo de la siguiente manera:
```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
```
### Callbacks
Otra opción para personalizar el [`Trainer`] es utilizar [callbacks](callbacks). Los callbacks *no cambian nada* en el bucle de entrenamiento. Inspeccionan el estado del bucle de entrenamiento y luego ejecutan alguna acción (detención anticipada, registro de resultados, etc.) según el estado. En otras palabras, un callback no puede usarse para implementar algo como una función de pérdida personalizada y necesitarás subclasificar y sobrescribir el método [`~Trainer.compute_loss`] para eso.
Por ejemplo, si deseas agregar un callback de detención anticipada al bucle de entrenamiento después de 10 pasos.
```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 {}
```
Luego, pásalo al parámetro `callback` del [`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()],
)
```
## Logging
<Tip>
Comprueba el API referencia [logging](./main_classes/logging) para mas información sobre los niveles differentes de logging.
</Tip>
El [`Trainer`] está configurado a `logging.INFO` de forma predeterminada el cual informa errores, advertencias y otra información basica. Un [`Trainer`] réplica - en entornos distributos - está configurado a `logging.WARNING` el cual solamente informa errores y advertencias. Puedes cambiar el nivel de logging con los parametros [`log_level`](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.TrainingArguments.log_level) y [`log_level_replica`](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.TrainingArguments.log_level_replica) en [`TrainingArguments`].
Para configurar el nivel de registro para cada nodo, usa el parámetro [`log_on_each_node`](https://huggingface.co/docs/transformers/main/en/main_classes/trainer#transformers.TrainingArguments.log_on_each_node) para determinar si deseas utilizar el nivel de registro en cada nodo o solo en el nodo principal.
<Tip>
[`Trainer`] establece el nivel de registro por separado para cada nodo en el método [`Trainer.init`], por lo que es posible que desees considerar establecer esto antes si estás utilizando otras funcionalidades de Transformers antes de crear el objeto [`Trainer`].
</Tip>
Por ejemplo, para establecer que tu código principal y los módulos utilicen el mismo nivel de registro según cada nodo:
```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(...)
```
<hfoptions id="logging">
<hfoption id="single node">
Usa diferentes combinaciones de `log_level` y `log_level_replica` para configurar qué se registra en cada uno de los nodos.
```bash
my_app.py ... --log_level warning --log_level_replica error
```
</hfoption>
<hfoption id="multi-node">
Agrega el parámetro `log_on_each_node 0` para entornos multi-nodo.
```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) es una técnica que puede mejorar el rendimiento al agregar ruido a los vectores de incrustación durante el entrenamiento. Para habilitarlo en [`Trainer`], establece el parámetro `neftune_noise_alpha` en [`TrainingArguments`] para controlar cuánto ruido se agrega.
```py
from transformers import TrainingArguments, Trainer
training_args = TrainingArguments(..., neftune_noise_alpha=0.1)
trainer = Trainer(..., args=training_args)
```
NEFTune se desactiva después del entrenamiento para restaurar la capa de incrustación original y evitar cualquier comportamiento inesperado.
## Accelerate y Trainer
La clase [`Trainer`] está impulsada por [Accelerate](https://hf.co/docs/accelerate), una biblioteca para entrenar fácilmente modelos de PyTorch en entornos distribuidos con soporte para integraciones como [FullyShardedDataParallel (FSDP)](https://pytorch.org/blog/introducing-pytorch-fully-sharded-data-parallel-api/) y [DeepSpeed](https://www.deepspeed.ai/).
<Tip>
Aprende más sobre las estrategias de fragmentación FSDP, descarga de CPU y más con el [`Trainer`] en la guía [Paralela de Datos Completamente Fragmentados](fsdp).
</Tip>
Para usar Accelerate con [`Trainer`], ejecuta el comando [`accelerate.config`](https://huggingface.co/docs/accelerate/package_reference/cli#accelerate-config) para configurar el entrenamiento para tu entorno de entrenamiento. Este comando crea un `config_file.yaml` que se utilizará cuando inicies tu script de entrenamiento. Por ejemplo, algunas configuraciones de ejemplo que puedes configurar son:
<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>
El comando [`accelerate_launch`](https://huggingface.co/docs/accelerate/package_reference/cli#accelerate-launch) es la forma recomendada de lanzar tu script de entrenamiento en un sistema distribuido con Accelerate y [`Trainer`] con los parámetros especificados en `config_file.yaml`. Este archivo se guarda en la carpeta de caché de Accelerate y se carga automáticamente cuando ejecutas `accelerate_launch`.
Por ejemplo, para ejecutar el script de entrenamiento [`run_glue.py`](https://github.com/huggingface/transformers/blob/f4db565b695582891e43a5e042e5d318e28f20b8/examples/pytorch/text-classification/run_glue.py#L4) con la configuración de FSDP:
```bash
accelerate launch \
./examples/pytorch/text-classification/run_glue.py \
--model_name_or_path 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
```
También puedes especificar los parámetros del archivo config_file.yaml directamente en la línea de comandos:
```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 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
```
Consulta el tutorial [Lanzamiento de tus scripts con Accelerate](https://huggingface.co/docs/accelerate/basic_tutorials/launch) para obtener más información sobre `accelerate_launch` y las configuraciones personalizadas.
|
transformers/docs/source/es/trainer.md/0
|
{
"file_path": "transformers/docs/source/es/trainer.md",
"repo_id": "transformers",
"token_count": 7327
}
| 291
|
<!--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.
-->
# Come aggiungere un modello a 🤗 Transformers?
Aggiungere un nuovo modello é spesso difficile e richiede una profonda conoscenza della libreria 🤗 Transformers e anche
della repository originale del modello. A Hugging Face cerchiamo di dare alla community sempre piú poteri per aggiungere
modelli independentemente. Quindi, per alcuni nuovi modelli che la community vuole aggiungere a 🤗 Transformers, abbiamo
creato una specifica *call-for-model-addition* che spiega passo dopo passo come aggiungere il modello richiesto. Con
questo *call-for-model-addition* vogliamo insegnare a volenterosi e esperti collaboratori della community come implementare
un modello in 🤗 Transformers.
Se questo é qualcosa che può interessarvi, siete liberi di controllare l'attuale “calls-for-model-addition” [qui](https://github.com/huggingface/transformers/tree/main/templates/adding_a_new_model/open_model_proposals/README.md)
e contattarci.
Se il modello sarà selezionato, allora potrete lavorare insieme a un membro di Hugging Face per integrare il modello in 🤗
Transformers. Così facendo, ci guadagnerai in una comprensione totale, sia teorica che pratica, del modello proposto. Inoltre,
sarai l'artefice di un importante contributo open-source a 🤗 Transformers. Durante l'implementazione avrai l'opportunità di:
- ottenere più comprensione delle best practices in open-source
- capire i principi di design di una della librerie NLP più popolari
- capire come efficientemente testare complessi modelli NLP
- capire come integrare utilit Python come `black`, `ruff`, `make fix-copies` in una libreria per garantire sempre di avere un codice leggibile e pulito
Siamo anche contenti se vuoi aggiungere un modello che non può essere trovato nella cartella “calls-for-model-addition”.
Le seguenti sezioni spiegano in dettaglio come aggiungere un nuovo modello. Può anche essere molto utile controllare modelli
già aggiunti [qui](https://github.com/huggingface/transformers/pulls?q=is%3Apr+label%3A%22PR+for+Model+Addition%22+is%3Aclosed),
per capire se richiamano il modello che vorreste aggiungere.
Per cominciare, vediamo una panoramica general della libreria Transformers.
## Panoramica generale su 🤗 Transformers
Prima di tutto, vediamo in generale 🤗 Transformers. 🤗 Transformers é una libreria molto strutturata, quindi
puà essere che a volte ci sia un disaccordo con alcune filosofie della libreria o scelte di design. Dalla nostra esperienza,
tuttavia, abbiamo trovato che le scelte fondamentali di design della libreria sono cruciali per usare 🤗 Transformers efficacemente
su larga scala, mantenendo i costi a un livello accettabile.
Un buon primo punto di partenza per capire al meglio la libreria é leggere la [documentazione sulla nostra filosofia](filosofia)
Da qui, ci sono alcune scelte sul modo di lavorare che cerchiamo di applicare a tutti i modelli:
- La composizione é generalmente favorita sulla sovra-astrazione
- Duplicare il codice non é sempre male, soprattutto se migliora notevolmente la leggibilità e accessibilità del modello
- Tutti i files creati per il nuovo modello devono il piu possibile "compatti". Questo vuol dire che quando qualcuno leggerá il codice
di uno specifico modello, potrá vedere solo il corrispettivo file `modeling_....py` senza avere multiple dipendenze.
La cosa piú importante, é che consideriamo la libreria non solo un mezzo per dare un prodotto, *per esempio* dare la possibilità
di usare BERT per inferenza, ma é anche il prodotto reale che noi vogliamo migliorare sempre più. Quindi, quando aggiungi
un modello, non sei solo la persona che userà il modello, ma rappresenti anche tutti coloro che leggeranno,
cercheranno di capire e modificare il tuo modello.
Tenendo questi principi in mente, immergiamoci nel design generale della libreria.
### Panoramica sui modelli
Per aggiungere con successo un modello, é importante capire l'interazione tra il tuo modello e la sua configurazione,
[`PreTrainedModel`], e [`PretrainedConfig`]. Per dare un esempio, chiameremo il modello da aggiungere a 🤗 Transformers
`BrandNewBert`.
Diamo un'occhiata:
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers_overview.png"/>
Come potete vedere, ci basiamo sull'ereditarietà in 🤗 Transformers, tenendo però il livello di astrazione a un minimo
assoluto. Non ci sono mai più di due livelli di astrazione per ogni modello nella libreria. `BrandNewBertModel` eredita
da `BrandNewBertPreTrainedModel` che, a sua volta, eredita da [`PreTrainedModel`] - semplice no?
Come regola generale, vogliamo essere sicuri che un nuovo modello dipenda solo da [`PreTrainedModel`]. Le funzionalità
importanti che sono automaticamente conferite a ogni nuovo modello sono [`~PreTrainedModel.from_pretrained`]
e [`~PreTrainedModel.save_pretrained`], che sono usate per serializzazione e deserializzazione. Tutte le altre importanti
funzionalità, come ad esempio `BrandNewBertModel.forward` devono essere definite completamente nel nuovo script
`modeling_brand_new_bert.py`. Inoltre, vogliamo essere sicuri che un modello con uno specifico head layer, come
`BrandNewBertForMaskedLM` non erediti da `BrandNewBertModel`, ma piuttosto usi `BrandNewBertModel`
come componente che può essere chiamata nel passaggio forward per mantenere il livello di astrazione basso. Ogni
nuovo modello richieste una classe di configurazione, chiamata `BrandNewBertConfig`. Questa configurazione é sempre
mantenuta come un attributo in [`PreTrainedModel`], e quindi può essere accessibile tramite l'attributo `config`
per tutte le classi che ereditano da `BrandNewBertPreTrainedModel`:
```python
model = BrandNewBertModel.from_pretrained("brandy/brand_new_bert")
model.config # il modello ha accesso al suo config
```
Analogamente al modello, la configurazione eredita le funzionalità base di serializzazione e deserializzazione da
[`PretrainedConfig`]. É da notare che la configurazione e il modello sono sempre serializzati in due formati differenti -
il modello é serializzato in un file *pytorch_model.bin* mentre la configurazione con *config.json*. Chiamando
[`~PreTrainedModel.save_pretrained`] automaticamente chiamerà [`~PretrainedConfig.save_pretrained`], cosicché sia il
modello che la configurazione siano salvati.
### Stile per il codice
Quando codifichi un nuovo modello, tieni presente che Transformers ha una sua struttura di fondo come libreria, perciò
ci sono alcuni fatti da considerare su come scrivere un codice :-)
1. Il forward pass del tuo modello dev'essere scritto completamente nel file del modello, mentre dev'essere indipendente
da altri modelli nella libreria. Se vuoi riutilizzare un blocco di codice da un altro modello, copia e incolla il codice con un commento `# Copied from` in cima al codice (guarda [qui](https://github.com/huggingface/transformers/blob/v4.17.0/src/transformers/models/roberta/modeling_roberta.py#L160)
per un ottimo esempio).
2. Il codice dev'essere interamente comprensibile, anche da persone che non parlano in inglese. Questo significa che le
variabili devono avere un nome descrittivo e bisogna evitare abbreviazioni. Per esempio, `activation` é molto meglio
che `act`. Le variabili con una lettera sono da evitare fortemente, almeno che non sia per un indce in un for loop.
3. Generamente é meglio avere un codice esplicito e piú lungo che un codice corto e magico.
4. Evita di subclassare `nn.Sequential` in Pytorch, puoi subclassare `nn.Module` e scrivere il forward pass, cosicché
chiunque può effettuare debug sul tuo codice, aggiungendo print o breaking points.
5. La tua function-signature dev'essere type-annoted. Per il resto, é meglio preferire variabili con un nome accettabile
piuttosto che annotazioni per aumentare la comprensione e leggibilità del codice.
### Panoramica sui tokenizers
Questa sezione sarà creata al piu presto :-(
## Aggiungere un modello a 🤗 Transformers passo dopo passo
Ci sono differenti modi per aggiungere un modello a Hugging Face. Qui trovi una lista di blog posts da parte della community su come aggiungere un modello:
1. [Aggiungere GPT2](https://medium.com/huggingface/from-tensorflow-to-pytorch-265f40ef2a28) scritto da [Thomas](https://huggingface.co/thomwolf)
2. [Aggiungere WMT19 MT](https://huggingface.co/blog/porting-fsmt) scritto da [Stas](https://huggingface.co/stas)
Per esperienza, possiamo dirti che quando si aggiunge un modello é meglio tenere a mente le seguenti considerazioni:
- Non sfondare una porta giá aperta! La maggior parte del codice che aggiungerai per un nuovo modello 🤗 Transformers
esiste già da qualche parte in 🤗 Transformers. Prendi un po' di tempo per trovare codici simili in modelli e tokenizers esistenti e fare un copia-incolla. Ricorda che [grep](https://www.gnu.org/software/grep/) e [rg](https://github.com/BurntSushi/ripgrep) sono tuoi buoni amici. Inoltre, ricorda che puó essere molto probabile che il tokenizer per il tuo modello sia basato sull'implementazione di un altro modello, e il codice del tuo modello stesso su un altro ancora. *Per esempio* il modello FSMT é basato su BART, mentre il tokenizer di FSMT é basato su XLM.
- Ricorda che qui é piu una sfida ingegneristica che scientifica. Spendi piú tempo per create un efficiente ambiente di debugging piuttosto che cercare di capire tutti gli aspetti teorici dell'articolo del modello.
- Chiedi aiuto se sei in panne! I modelli sono la parte principale di 🤗 Transformers, perciò qui a Hugging Face siamo più che contenti di aiutarti in ogni passo per aggiungere il tuo modello. Non esitare a chiedere se vedi che non riesci a progredire.
Di seguito, diamo una ricetta generale per aiutare a portare un modello in 🤗 Transformers.
La lista seguente é un sommario di tutto quello che é stato fatto per aggiungere un modello, e può essere usata come To-Do List:
- 1. ☐ (Opzionale) Capire gli aspetti teorici del modello
- 2. ☐ Preparare l'ambiente dev per transformers
- 3. ☐ Preparare l'ambiente debugging della repository originale
- 4. ☐ Create uno script che gestisca con successo il forward pass usando la repository originale e checkpoint
- 5. ☐ Aggiungere con successo lo scheletro del modello a Transformers
- 6. ☐ Convertire i checkpoint original a Transformers checkpoint
- 7. ☐ Effettuare con successo la forward pass in Transformers, di modo che dia un output identico al checkpoint originale
- 8. ☐ Finire i tests per il modello in Transformers
- 9. ☐ Aggiungere con successo Tokenizer in Transformers
- 10. ☐ Testare e provare gli integration tests da capo a fine
- 11. ☐ Completare i docs
- 12. ☐ Caricare i moedl weights all'hub
- 13. ☐ Sottomettere una pull request
- 14. ☐ (Opzionale) Aggiungere un notebook con una demo
Per cominciare di solito consigliamo `BrandNewBert`, partendo dalla teoria, di modo da avere una buona comprensione della teoria generale. TUttavia, se preferisci imparare l'aspetto teorico del modello mentre *lavori* sul modello é ok immergersi direttamente nel codice di `BrandNewBert`. Questa opzione puó essere buona se le tue skills ingegneristiche sono meglio che quelle teoriche, o se il paper `BrandNewBert` ti dá problemi, o se semplicemente ti piace programmare piú che leggere articoli scientifici.
### 1. (Opzionale) Aspetti teorici di BrandNewBert
Allora con calma, prendi un po' di tempo per leggere l'articolo su *BrandNewBert* . Sicuramente, alcune sezioni dell'articolo sono molto complesse, ma non preoccuparti! L'obiettivo non é avere una compresione immensa della teoria alla base, ma estrarre le informazioni necessarie per re-implementare con successo il modello in 🤗 Transformers. Quindi, non impazzire sugli aspetti teorici, ma piuttosto focalizzati su quelli pratici, ossia:
- Che tipo di modello é *brand_new_bert*? É solo un encoder in stile BERT? O tipo decoder come GPT2? O encoder e decoder stile BART? Dai un'occhiata a [model_summary](model_summary) se non sei famigliare con le differenze tra questi modelli
- Quali sono le applicazioni di *brand_new_bert*? Classificazione di testo? Generazione di testo? O per tasks del genere seq2seq?
- Quali sono le nuove aggiunte al modello che lo rendono diverso da BERT/GPT-2/BART?
- Quali modelli estistenti in [🤗 Transformers models](https://huggingface.co/transformers/#contents) sono molto simili a *brand_new_bert*?
- Che tipo di tokenizer si usa in questo caso? Un sentencepiece tokenizer? O un word piece tokenizer? Il tokenizer é lo stesso di BERT o BART?
Una volta che senti che hai avuto una bella overview dell'architettura del modello, puoi scrivere senza problemi al team di Hugging Face per ogni domanda che tu hai. Questo puó includere domande sull'architettura del modello, o sull'attention layer, etc. Saremo molto felici di aiutarti :)
### 2. Prepare il tuo ambiente
1. Forka la [repository](https://github.com/huggingface/transformers) cliccando sul tasto ‘Fork' nella pagina della repository. Questo crea una copia del codice nel tuo account GitHub
2. Clona il tuo fork `transfomers` sul tuo dico locale, e aggiungi la repository base come remota:
```bash
git clone https://github.com/[your Github handle]/transformers.git
cd transformers
git remote add upstream https://github.com/huggingface/transformers.git
```
3. Crea un ambiente di sviluppo, per esempio tramite questo comando:
```bash
python -m venv .env
source .env/bin/activate
pip install -e ".[dev]"
```
quindi torna alla directory principale:
```bash
cd ..
```
4. Attenzione, raccomandiamo di aggiungere la versione di PyTorch di *brand_new_bert* a Transfomers. Per installare PyTorch, basta seguire queste istruzioni https://pytorch.org/get-started/locally/.
**Nota bene:** Non c'é bisogno di installare o avere installato CUDA. Il nuovo modello può funzionare senza problemi su una CPU.
5. Per trasferire *brand_new_bert* To port *brand_new_bert* avrai bisogno anche accesso alla sua repository originale:
```bash
git clone https://github.com/org_that_created_brand_new_bert_org/brand_new_bert.git
cd brand_new_bert
pip install -e .
```
Ok, ora hai un ambiente di sviluppo per portare *brand_new_bert* in 🤗 Transformers.
### 3.-4. Provare un pretrained checkpoint usando la repo originale
Per cominciare, comincerai a lavorare sulla repo originale di *brand_new_bert*. Come spesso accade, l'implementazione originale é molto sullo stile "ricerca". Questo significa che a volte la documentazione non é al top, magari manca qualche cosa e il codice puó essere difficile da capire. Tuttavia, questa é e dev'essere la motivazione per reimplementare *brand_new_bert*. In Hugging Face, uno degli obiettivi principali é di *mettere le persone sulle spalle dei giganti*, il che si traduce, in questo contesto, di prendere un modello funzionante e riscriverlo e renderlo il piú possibile **accessibile, user-friendly, e leggibile**. Questa é la top motivazione per re-implementare modelli in 🤗 Transformers - cercare di creare nuove complesse tecnologie NLP accessibili a **chiunque**.
Riuscire a far girare il modello pretrained originale dalla repository ufficiale é spesso il passo **piu arduo**. Dalla nostra esperienza, é molto importante spendere un p' di tempo per diventare familiari con il codice base originale. Come test, prova a capire i seguenti punti:
- Dove si trovano i pretrained weights?
- Come caricare i pretrained weights nel modello corrispondente?
- Come girare un tokenizer independentemente dal modello?
- Prova a tracciare un singolo forward pass, cosicché potrai sapere che classi e funzioni sono richieste per un semplice forward pass. Di solito, dovrai reimplementare queste funzioni e basta
- Prova a localizzare i componenti importanti del modello: Dove si trova la classe del modello? Ci sono sotto classi nel modello *per esempio* EngoderModel, DecoderMOdel? Dove si trova il self-attention layer? Ci sono molteplici differenti layer di attention, *per esempio * *self-attention*, *cross-attention*...?
- Come puoi fare debug sul modello nell'ambiente originale della repo? Devi aggiungere dei *print* o puoi usare *ipdb* come debugger interattivo, o vabene anche un IDE efficiente per debug come PyCharm?
É molto importante che prima di cominciare a trasferire il modello nuovo tu spenda tempo a fare debug del codice originale in maniera **efficiente**! Inoltre, ricorda che tutta la library é open-soruce, quindi non temere di aprire issue o fare una pull request nella repo originale. Tutti coloro che mantengono la repository saranno piú che felici di avere qualcuno che guarda e gioca con i loro codici!
A questo punto, sta a te decidere quale ambiente per debug vuoi usare. Noi consilgiamo di evitare setup con GPU, che potrebbero costare assai, lavorare su una CPU puó essere un ottimo punto di partenza per indagare la repository originale e per cominciare a scrivere il codice per 🤗 Transformers. Solo alla fine, quando il modello é stato portato con successo in 🤗 Transformers, allora si potrá verificare il suo funzionamento su GPU.
In generale ci sono due possibili ambienti di debug per il testare il modello originale:
- [Jupyter notebooks](https://jupyter.org/) / [google colab](https://colab.research.google.com/notebooks/intro.ipynb)
- Scripts locali in Python
Il vantaggio dei Jupyter notebooks é la possibilità di eseguire cella per cella, il che può essere utile per decomporre tutte le componenti logiche, cosi da a vere un ciclo di debug più rapido, siccome si possono salvare i risultati da steps intermedi. Inoltre, i notebooks spesso sono molto facili da condividere con altri contributors, il che può essere molto utile se vuoi chiedere aiuto al team di Hugging Face. Se sei famigliare con Jupyter notebooks allora racommandiamo di lavorare in questa maniera.
Ovviamente se non siete abituati a lavorare con i notebook, questo può essere uno svantaggio nell'usare questa tecnologia, sprecando un sacco di tempo per setup e portare tutto al nuovo ambiente, siccome non potreste neanche usare dei tools di debug come `ipdb`.
Per ogni pratica code-base, é sempre meglio come primo step caricare un **piccolo** checkpoint pretrained e cercare di riprodurre un singolo forward pass usando un vettore fittizio di IDs fatti da numeri interi. Un esempio per uno script simile, in pseudocodice é:
```python
model = BrandNewBertModel.load_pretrained_checkpoint("/path/to/checkpoint/")
input_ids = [0, 4, 5, 2, 3, 7, 9] # vector of input ids
original_output = model.predict(input_ids)
```
Per quanto riguarda la strategia di debugging, si può scegliere tra:
- Decomporre il modello originario in piccole componenenti e testare ognuna di esse
- Decomporre il modello originario nel *tokenizer* originale e nel *modello* originale, testare un forward pass su questi,
e usare dei print statement o breakpoints intermedi per verificare
Ancora una volta, siete liberi di scegliere quale strategia sia ottimale per voi. Spesso una strategia é piu
avvantaggiosa di un'altra, ma tutto dipende dall'code-base originario.
Se il code-base vi permette di decomporre il modello in piccole sub-componenenti, *per esempio* se il code-base
originario può essere facilmente testato in eager mode, allora vale la pena effettuare un debugging di questo genere.
Ricordate che ci sono dei vantaggi nel decidere di prendere la strada piu impegnativa sin da subito:
- negli stage piu finali, quando bisognerà comparare il modello originario all'implementazione in Hugging Face, potrete verificare
automaticamente ogni componente, individualmente, di modo che ci sia una corrispondenza 1:1
- avrete l'opportunità di decomporre un problema molto grande in piccoli passi, così da strutturare meglio il vostro lavoro
- separare il modello in componenti logiche vi aiuterà ad avere un'ottima overview sul design del modello, quindi una migliore
comprensione del modello stesso
- verso gli stage finali i test fatti componente per componente vi aiuterà ad essere sicuri di non andare avanti e indietro
nell'implementazione, così da continuare la modifica del codice senza interruzione
Un ottimo esempio di come questo può essere fatto é dato da [Lysandre](https://gist.github.com/LysandreJik/db4c948f6b4483960de5cbac598ad4ed)
per il modello ELECTRA
Tuttavia, se il code-base originale é molto complesso o le componenti intermedie possono essere testate solo in tramite
compilazione, potrebbe richiedere parecchio tempo o addirittura essere impossibile separare il modello in piccole sotto-componenti.
Un buon esempio é [MeshTensorFlow di T5](https://github.com/tensorflow/mesh/tree/master/mesh_tensorflow). Questa libreria
é molto complessa e non offre un metodo semplice di decomposizione in sotto-componenti. Per simili librerie, potrete fare
affidamento ai print statements.
In ogni caso, indipendentemente da quale strategia scegliete, la procedura raccomandata é di cominciare a fare debug dal
primo layer al layer finale.
É consigliato recuperare gli output dai layers, tramite print o sotto-componenti, nel seguente ordine:
1. Recuperare gli IDs di input dati al modello
2. Recuperare i word embeddings
3. Recuperare l'input del primo Transformer layer
4. Recuperare l'output del primo Transformer layer
5. Recuperare l'output dei seguenti `n - 1` Transformer layers
6. Recuperare l'output dell'intero BrandNewBert Model
Gli IDs in input dovrebbero essere un arrary di interi, *per esempio* `input_ids = [0, 4, 4, 3, 2, 4, 1, 7, 19]`
Gli output dei seguenti layer di solito dovrebbero essere degli array di float multi-dimensionali come questo:
```
[[
[-0.1465, -0.6501, 0.1993, ..., 0.1451, 0.3430, 0.6024],
[-0.4417, -0.5920, 0.3450, ..., -0.3062, 0.6182, 0.7132],
[-0.5009, -0.7122, 0.4548, ..., -0.3662, 0.6091, 0.7648],
...,
[-0.5613, -0.6332, 0.4324, ..., -0.3792, 0.7372, 0.9288],
[-0.5416, -0.6345, 0.4180, ..., -0.3564, 0.6992, 0.9191],
[-0.5334, -0.6403, 0.4271, ..., -0.3339, 0.6533, 0.8694]]],
```
Ci aspettiamo che ogni modello aggiunto a 🤗 Transformers passi con successo un paio di test d'integrazione. Questo
significa che il modello originale e la sua implementazione in 🤗 Transformers abbiano lo stesso output con una precisione
di 0.001! Siccome é normale che lo stesso esatto modello, scritto in librerie diverse, possa dare output leggermente
diversi, la tolleranza accettata é 1e-3 (0.001). Ricordate che i due modelli devono dare output quasi identici. Dunque,
é molto conveniente comparare gli output intermedi di 🤗 Transformers molteplici volte con gli output intermedi del
modello originale di *brand_new_bert*. Di seguito vi diamo alcuni consigli per avere un ambiente di debug il piu efficiente
possibile:
- Trovate la migliore strategia per fare debug dei risultati intermedi. Per esempio, é la repository originale scritta in PyTorch?
Se si, molto probabilmente dovrete dedicare un po' di tempo per scrivere degli script piu lunghi, così da decomporre il
modello originale in piccole sotto-componenti, in modo da poter recuperare i valori intermedi. Oppure, la repo originale
é scritta in Tensorflow 1? Se é così dovrete fare affidamento ai print di Tensorflow [tf.print](https://www.tensorflow.org/api_docs/python/tf/print)
per avere i valori intermedi. Altro caso, la repo é scritta in Jax? Allora assicuratevi che il modello non sia in **jit**
quanto testate il foward pass, *per esempio* controllate [questo link](https://github.com/google/jax/issues/196).
- Usate i più piccoli pretrained checkpoint che potete trovare. Piu piccolo é il checkpoint, piu velocemente sarà il vostro
ciclo di debug. Non é efficiente avere un pretrained model così gigante che per il forward pass impieghi piu di 10 secondi.
Nel caso in cui i checkpoints siano molto grandi, e non si possa trovare di meglio, allora é buona consuetudine ricorrere
a fare un dummy model nel nuovo ambiente, con weights inizializzati random e salvare quei weights per comprare la versione 🤗 Transformers
con il vostro modello
- Accertatevi di usare la via piu semplice per chiamare il forward pass nella repo originale. Sarebbe opportuno trovare
la funzione originaria che chiami **solo** un singolo forward pass, *per esempio* questa funzione spesso viene chiamata
`predict`, `evaluate`, `forward` o `__call__`. Siate sicuri di non fare debug su una funzione che chiami `forward` molteplici
volte, *per esempio* per generare testo, come `autoregressive_sample`, `generate`.
- Cercate di separare la tokenization dal forward pass del modello. Se la repo originaria mostra esempio dove potete dare
come input una stringa, provate a cercare dove nella forward call la stringa viene cambiata in input ids e cominciate il
debug da questo punto. Questo vi garantisce un ottimo punto di partenza per scrivere un piccolo script personale dove dare
gli input al modello, anziche delle stringhe in input.
- Assicuratevi che il debugging **non** sia in training mode. Spesso questo potra il modello a dare degli output random, per
via dei molteplici dropout layers. Assicuratevi che il forward pass nell'ambiente di debug sia **deterministico**, cosicche
i dropout non siano usati. Alternativamente, potete usare *transformers.utils.set_seed* se la vecchia e nuova implementazione
sono nello stesso framework.
La seguente sezione vi da ulteriori dettagli e accorgimenti su come potete fare tutto questo per *brand_new_bert*.
### 5.-14. Trasferire BrandNewBert in 🤗 Transformers
Allora cominciamo ad aggiungere un nuovo codice in 🤗 Transformers. Andate nel vostro fork clone di 🤗 Transformers:
```bash
cd transformers
```
Nel caso speciale in cui stiate aggiungendo un modello, la cui architettura sia identica a una di un modello già esistente,
dovrete solo aggiugnere uno script di conversione, come descritto [qui](#write-a-conversion-script).
In questo caso, potete riutilizzare l'intera architettura del modello gia esistente.
Se questo non é il caso, cominciamo con il generare un nuovo modello. Ti consigliamo di utilizzare il seguente script per aggiungere un modello a partire da
un modello esistente:
```bash
transformers-cli add-new-model-like
```
Ti verrà richiesto con un questionario di compilare le informazioni di base del tuo modello.
**Aprire una Pull Request in main huggingface/transformers repo**
Prime di cominciare ad adattare il codice automaticamente generato, aprite una nuova PR come "Work in progress (WIP)",
*per esempio* "[WIP] Aggiungere *brand_new_bert*", cosicché il team di Hugging Face possa lavorare al vostro fianco nell'
integrare il modello in 🤗 Transformers.
Questi sarebbero gli step generali da seguire:
1. Creare un branch dal main branch con un nome descrittivo
```bash
git checkout -b add_brand_new_bert
```
2. Commit del codice automaticamente generato
```bash
git add .
git commit
```
3. Fare fetch e rebase del main esistente
```bash
git fetch upstream
git rebase upstream/main
```
4. Push dei cambiamenti al proprio account:
```bash
git push -u origin a-descriptive-name-for-my-changes
```
5. Una volte che siete soddisfatti dei nuovi cambiamenti, andate sulla webpage del vostro fork su GitHub. Cliccate "Pull request".
Assiuratevi di aggiungere alcuni membri di Hugging Face come reviewers, nel riguardo alla destra della pagina della PR, cosicche il team
Hugging Face verrà notificato anche per i futuri cambiamenti.
6. Cambiare la PR a draft, cliccando su "Convert to draft" alla destra della pagina della PR
Da quel punto in poi, ricordate di fare commit di ogni progresso e cambiamento, cosicche venga mostrato nella PR. Inoltre,
ricordatevi di tenere aggiornato il vostro lavoro con il main esistente:
```bash
git fetch upstream
git merge upstream/main
```
In generale, tutte le domande che avrete riguardo al modello o l'implementazione dovranno essere fatte nella vostra PR
e discusse/risolte nella PR stessa. In questa maniera, il team di Hugging Face sarà sempre notificato quando farete commit
di un nuovo codice o se avrete qualche domanda. É molto utile indicare al team di Hugging Face il codice a cui fate riferimento
nella domanda, cosicche il team potra facilmente capire il problema o la domanda.
Per fare questo andate sulla tab "Files changed", dove potrete vedere tutti i vostri cambiamenti al codice, andate sulla linea
dove volete chiedere una domanda, e cliccate sul simbolo "+" per aggiungere un commento. Ogni volta che una domanda o problema
é stato risolto, cliccate sul bottone "Resolve".
In questa stessa maniera, Hugging Face aprirà domande o commenti nel rivedere il vostro codice. Mi raccomando, chiedete più
domande possibili nella pagina della vostra PR. Se avete domande molto generali, non molto utili per il pubblico, siete liberi
di chiedere al team Hugging Face direttamente su slack o email.
**5. Adattare i codici per brand_new_bert**
Per prima cosa, ci focalizzeremo sul modello e non sui tokenizer. Tutto il codice relative dovrebbe trovarsi in
`src/transformers/models/brand_new_bert/modeling_brand_new_bert.py` e
`src/transformers/models/brand_new_bert/configuration_brand_new_bert.py`.
Ora potete finalmente cominciare il codice :). Il codice generato in
`src/transformers/models/brand_new_bert/modeling_brand_new_bert.py` avrà sia la stessa architettura di BERT se é un
modello encoder-only o BART se é encoder-decoder. A questo punto, ricordatevi cio che avete imparato all'inizio, riguardo
agli aspetti teorici del modello: *In che maniera il modello che sto implmementando é diverso da BERT o BART?*. Implementare
questi cambi spesso vuol dire cambiare il layer *self-attention*, l'ordine dei layer di normalizzazione e così via...
Ancora una volta ripetiamo, é molto utile vedere architetture simili di modelli gia esistenti in Transformers per avere
un'idea migliore su come implementare il modello.
**Notate** che a questo punto non dovete avere subito un codice tutto corretto o pulito. Piuttosto, é consigliato cominciare con un
codice poco pulito, con copia-incolla del codice originale in `src/transformers/models/brand_new_bert/modeling_brand_new_bert.py`
fino a che non avrete tutto il codice necessario. In base alla nostra esperienza, é molto meglio aggiungere una prima bozza
del codice richiesto e poi correggere e migliorare iterativamente. L'unica cosa essenziale che deve funzionare qui é la seguente
instanza:
```python
from transformers import BrandNewBertModel, BrandNewBertConfig
model = BrandNewBertModel(BrandNewBertConfig())
```
Questo comando creerà un modello con i parametri di default definiti in `BrandNewBergConfig()` e weights random. Questo garantisce
che `init()` di tutte le componenti funzioni correttamente.
**6. Scrivere uno script di conversione**
Il prossimo step é scrivere uno script per convertire il checkpoint che avete usato per fare debug su *brand_new_berts* nella
repo originale in un checkpoint per la nuova implementazione di *brand_new_bert* in 🤗 Transformers. Non é consigliato scrivere
lo script di conversione da zero, ma piuttosto cercate e guardate script gia esistenti in 🤗 Transformers, così da trovarne
uno simile al vostro modello. Di solito basta fare una copia di uno script gia esistente e adattarlo al vostro caso.
Non esistate a chiedre al team di Hugging Face a riguardo.
- Se state convertendo un modello da TensorFlow a PyTorch, un ottimo inizio é vedere [questo script di conversione per BERT](https://github.com/huggingface/transformers/blob/7acfa95afb8194f8f9c1f4d2c6028224dbed35a2/src/transformers/models/bert/modeling_bert.py#L91)
- Se state convertendo un modello da PyTorch a PyTorch, [lo script di conversione di BART può esservi utile](https://github.com/huggingface/transformers/blob/main/src/transformers/models/bart/convert_bart_original_pytorch_checkpoint_to_pytorch.py)
Qui di seguito spiegheremo come i modelli PyTorch salvano i weights per ogni layer e come i nomi dei layer sono definiti. In PyTorch,
il nomde del layer é definito dal nome della class attribute che date al layer. Definiamo un modello dummy in PyTorch,
chiamato `SimpleModel`:
```python
from torch import nn
class SimpleModel(nn.Module):
def __init__(self):
super().__init__()
self.dense = nn.Linear(10, 10)
self.intermediate = nn.Linear(10, 10)
self.layer_norm = nn.LayerNorm(10)
```
Ora possiamo creare un'instanza di questa definizione di modo da inizializzare a random weights: `dense`, `intermediate`, `layer_norm`.
Possiamo usare print per vedere l'architettura del modello:
```python
model = SimpleModel()
print(model)
```
Da cui si ottiene:
```
SimpleModel(
(dense): Linear(in_features=10, out_features=10, bias=True)
(intermediate): Linear(in_features=10, out_features=10, bias=True)
(layer_norm): LayerNorm((10,), eps=1e-05, elementwise_affine=True)
)
```
Si può vedere come i nomi dei layers siano definiti dal nome della class attribute in PyTorch. I valori dei weights di uno
specifico layer possono essere visualizzati:
```python
print(model.dense.weight.data)
```
ad esempio:
```
tensor([[-0.0818, 0.2207, -0.0749, -0.0030, 0.0045, -0.1569, -0.1598, 0.0212,
-0.2077, 0.2157],
[ 0.1044, 0.0201, 0.0990, 0.2482, 0.3116, 0.2509, 0.2866, -0.2190,
0.2166, -0.0212],
[-0.2000, 0.1107, -0.1999, -0.3119, 0.1559, 0.0993, 0.1776, -0.1950,
-0.1023, -0.0447],
[-0.0888, -0.1092, 0.2281, 0.0336, 0.1817, -0.0115, 0.2096, 0.1415,
-0.1876, -0.2467],
[ 0.2208, -0.2352, -0.1426, -0.2636, -0.2889, -0.2061, -0.2849, -0.0465,
0.2577, 0.0402],
[ 0.1502, 0.2465, 0.2566, 0.0693, 0.2352, -0.0530, 0.1859, -0.0604,
0.2132, 0.1680],
[ 0.1733, -0.2407, -0.1721, 0.1484, 0.0358, -0.0633, -0.0721, -0.0090,
0.2707, -0.2509],
[-0.1173, 0.1561, 0.2945, 0.0595, -0.1996, 0.2988, -0.0802, 0.0407,
0.1829, -0.1568],
[-0.1164, -0.2228, -0.0403, 0.0428, 0.1339, 0.0047, 0.1967, 0.2923,
0.0333, -0.0536],
[-0.1492, -0.1616, 0.1057, 0.1950, -0.2807, -0.2710, -0.1586, 0.0739,
0.2220, 0.2358]]).
```
Nello script di conversione, dovreste riempire quei valori di inizializzazione random con gli stessi weights del corrispondente
layer nel checkpoint. *Per esempio*
```python
# retrieve matching layer weights, e.g. by
# recursive algorithm
layer_name = "dense"
pretrained_weight = array_of_dense_layer
model_pointer = getattr(model, "dense")
model_pointer.weight.data = torch.from_numpy(pretrained_weight)
```
Così facendo, dovete verificare che ogni inizializzazione random di un peso del modello PyTorch e il suo corrispondente peso nel pretrained checkpoint
siano esattamente gli stessi e uguali in **dimensione/shape e nome**. Per fare questo, é **necessario** aggiungere un `assert`
per la dimensione/shape e nome:
```python
assert (
model_pointer.weight.shape == pretrained_weight.shape
), f"Pointer shape of random weight {model_pointer.shape} and array shape of checkpoint weight {pretrained_weight.shape} mismatched"
```
Inoltre, dovrete fare il print sia dei nomi che dei weights per essere sicuri che siano gli stessi:
```python
logger.info(f"Initialize PyTorch weight {layer_name} from {pretrained_weight.name}")
```
Se la dimensione o il nome non sono uguali, probabilmente avete sbagliato ad assegnare il peso nel checkpoint o nel layer costrutture di
🤗 Transformers.
Una dimensione sbagliata può essere dovuta ad un errore nei parameteri in `BrandNewBertConfig()`. Tuttavia, può essere anche
che l'implementazione del layer in PyTorch richieda di fare una transposizione della matrice dei weights.
Infine, controllate **tutti** che tutti i weights inizializzati e fate print di tutti i weights del checkpoint che non sono stati
usati per l'inizializzazione, di modo da essere sicuri che il modello sia correttamente convertito. É normale che ci siano
errori nel test di conversione, fai per un errore in `BrandNewBertConfig()`, o un errore nell'architettura in 🤗 Transformers,
o un bug in `init()`.
Questo step dev'essere fatto tramite iterazioni fino a che non si raggiungano gli stessi valori per i weights. Una volta che
il checkpoint é stato correttamente caricato in 🤗 Transformers, potete salvare il modello in una cartella di vostra scelta
`/path/to/converted/checkpoint/folder` che contenga sia
`pytorch_model.bin` che `config.json`:
```python
model.save_pretrained("/path/to/converted/checkpoint/folder")
```
**7. Implementare il forward pass**
Una volta che i weights pretrained sono stati correttamente caricati in 🤗 Transformers, dovrete assicurarvi che il forward pass
sia correttamente implementato. [Qui](#3-4-provare-un-pretrained-checkpoint-usando-la-repo-originale), avete give creato e provato
uno script che testi il forward pass del modello usando la repo originaria. Ora dovrete fare lo stesso con uno script analogo
usando l'implementazione in 🤗 Transformers anziché l'originale. Piu o meno lo script dovrebbe essere:
```python
model = BrandNewBertModel.from_pretrained("/path/to/converted/checkpoint/folder")
input_ids = [0, 4, 4, 3, 2, 4, 1, 7, 19]
output = model(input_ids).last_hidden_states
```
Di solito l'output da 🤗 Transformers non é uguale uguale all'output originario, sopratto la prima volta. Non vi abbattete -
é normale! Prima di tutto assicuratevi che non ci siano errori o che non vengano segnalati degli errori nella forward pass.
Spesso capita che ci siano dimensioni sbagliate o data type sbagliati, *ad esempio* `torch.long` anziche `torch.float32`.
Non esistate a chiedere al team Hugging Face!
Nella parte finale assicuratevi che l'implementazione 🤗 Transformers funzioni correttamente cosi da testare che gli output
siano equivalenti a una precisione di `1e-3`. Controllate che `outputs.shape` siano le stesse tra 🤗 Transformers e l'implementazione
originaria. Poi, controllate che i valori in output siano identici. Questa é sicuramente la parte più difficile, qui una serie
di errori comuni quando gli output non sono uguali:
- Alcuni layers non sono stati aggiunti, *ad esempio* un *activation* layer non é stato aggiunto, o ci si é scordati di una connessione
- La matrice del word embedding non é stata ripareggiata
- Ci sono degli embeddings posizionali sbagliati perché l'implementazione originaria ha un offset
- Il dropout é in azione durante il forward pass. Per sistemare questo errore controllate che *model.training = False* e che
il dropout non sia stato attivato nel forward pass, * per esempio * passate *self.training* a [PyTorch's functional dropout](https://pytorch.org/docs/stable/nn.functional.html?highlight=dropout#torch.nn.functional.dropout)
La miglior maniera per sistemare il problema é di vedere all'implementazione originaria del forward pass e in 🤗 Transformers
fianco a fianco e vedere se ci sono delle differenze. In teoria, con debug e print degli output intermedie di entrambe le
implementazioni nel forward pass nell'esatta posizione del network dovrebbe aiutarvi a vedere dove ci sono differenze tra
i due frameworks. Come prima mossa controllate che `input_ids` siano identici in entrambi gli scripts. Da lì andate fino
all'ultimo layer. Potrete notare una differenza tra le due implementazioni a quel punto.
Una volta che lo stesso output é stato ragguingi, verificate gli output con `torch.allclose(original_output, output, atol=1e-3)`.
A questo punto se é tutto a posto: complimenti! Le parti seguenti saranno una passeggiata 😊.
**8. Aggiungere i test necessari per il modello**
A questo punto avete aggiunto con successo il vostro nuovo modello. Tuttavia, é molto probabile che il modello non sia
del tutto ok con il design richiesto. Per essere sicuri che l'implementazione sia consona e compatibile con 🤗 Transformers é
necessario implementare dei tests. Il Cookiecutter dovrebbe fornire automaticamente dei file per test per il vostro modello,
di solito nella folder `tests/test_modeling_brand_new_bert.py`. Provate questo per verificare l'ok nei test piu comuni:
```bash
pytest tests/test_modeling_brand_new_bert.py
```
Una volta sistemati i test comuni, bisogna assicurarsi che il vostro lavoro sia correttamente testato cosicchè:
- a) La community puo capire in maniera semplice il vostro lavoro controllando tests specifici del modello *brand_new_bert*,
- b) Implementazioni future del vostro modello non rompano alcune feature importante del modello.
Per prima cosa agguingete dei test d'integrazione. Questi sono essenziali perche fanno la stessa funzione degli scripts di
debug usati precedentemente. Un template per questi tests esiste gia nel Cookiecutter ed é sotto il nome di `BrandNewBertModelIntegrationTests`,
voi dovrete solo completarlo. Una volta che questi tests sono OK, provate:
```bash
RUN_SLOW=1 pytest -sv tests/test_modeling_brand_new_bert.py::BrandNewBertModelIntegrationTests
```
<Tip>
Nel caso siate su Windows, sostituite `RUN_SLOW=1` con `SET RUN_SLOW=1`
</Tip>
Di seguito, tutte le features che sono utili e necessarire per *brand_new_bert* devono essere testate in test separati,
contenuti in `BrandNewBertModelTester`/ `BrandNewBertModelTest`. spesso la gente si scorda questi test, ma ricordate che sono utili per:
- Aiuta gli utenti a capire il vostro codice meglio, richiamando l'attenzione su queste nuove features
- Developers e contributors futuri potranno velocemente testare nuove implementazioni del modello testanto questi casi speciali.
**9. Implementare il tokenizer**
A questo punto avremo bisogno un tokenizer per *brand_new_bert*. Di solito il tokenizer é uguale ad altri modelli in 🤗 Transformers.
É importante che troviate il file con il tokenizer originale e che lo carichiate in 🤗 Transformers.
Per controllare che il tokenizer funzioni in modo corretto, create uno script nella repo originaria che riceva come input
una stringa e ritorni gli `input_ids`. Piu o meno questo potrebbe essere il codice:
```python
input_str = "This is a long example input string containing special characters .$?-, numbers 2872 234 12 and words."
model = BrandNewBertModel.load_pretrained_checkpoint("/path/to/checkpoint/")
input_ids = model.tokenize(input_str)
```
Potrebbe richiedere un po' di tempo, ma guardate ancora alla repo originaria per trovare la funzione corretta del tokenizer.
A volte capita di dover riscrivere il tokenizer nella repo originaria, di modo da avere come output gli `input_ids`.
A quel punto uno script analogo é necessario in 🤗 Transformers:
```python
from transformers import BrandNewBertTokenizer
input_str = "This is a long example input string containing special characters .$?-, numbers 2872 234 12 and words."
tokenizer = BrandNewBertTokenizer.from_pretrained("/path/to/tokenizer/folder/")
input_ids = tokenizer(input_str).input_ids
```
Una volta che `input_ids` sono uguali, bisogna aggiungere un test per il tokenizer.
Il file test per tokenizer di *brand_new_brand* dovrebbe avere un paio di hard-coded test d'integrazione.
**10. Test end-to-end**
Ora che avete il tokenizer, dovrete aggiungere dei test d'integrazione per l'intero workflow in `tests/test_modeling_brand_new_bert.py` in 🤗 Transformer.
Questi test devono mostrare che un significante campione text-to-text funzioni come ci si aspetta nell'implementazione di 🤗 Transformers.
*Per esempio* potreste usare dei source-to-target-translation, o un sommario di un articolo, o un domanda-risposta e cosi via.
Se nessuno dei checkpoints é stato ultra parametrizzato per task simili, allora i tests per il modello sono piu che sufficienti.
Nello step finale dovete assicurarvi che il modello sia totalmente funzionale, e consigliamo anche di provare a testare su GPU.
Puo succedere che ci si scordi un `.to(self.device)` ad esempio. Se non avete accesso a GPU, il team Hugging Face puo provvedere
a testare questo aspetto per voi.
**11. Aggiungere una Docstring**
Siete quasi alla fine! L'ultima cosa rimasta é avere una bella docstring e una pagina doc. Il Cookiecutter dovrebbe provvedere già
un template chiamato `docs/source/model_doc/brand_new_bert.rst`, che dovrete compilare. La prima cosa che un utente farà
per usare il vostro modello sarà dare una bella lettura al doc. Quindi proponete una documentazione chiara e concisa. É molto
utile per la community avere anche delle *Tips* per mostrare come il modello puo' essere usato. Non esitate a chiedere a Hugging Face
riguardo alle docstirng.
Quindi, assicuratevi che la docstring sia stata aggiunta a `src/transformers/models/brand_new_bert/modeling_brand_new_bert.py`.
Assicuratevi che la docstring sia corretta e che includa tutti i necessari input e output. Abbiamo una guida dettagliata per
scrivere la documentazione e docstring.
**Rifattorizzare il codice**
Perfetto! Ora che abbiamo tutto per *brand_new_bert* controllate che lo stile del codice sia ok:
```bash
make style
```
E che il codice passi i quality check:
```bash
make quality
```
A volte capita che manchino delle informazioninella docstring o alcuni nomi sbagliati, questo farà fallire i tests sopra.
Ripetiamo: chiedete pure a Hugging Face, saremo lieti di aiutarvi.
Per ultimo, fare del refactoring del codice una volta che é stato creato.
Avete finito con il codice, congratulazioni! 🎉 Siete fantasticiiiiiii! 😎
**12. Caricare il modello sul model hub**
In questa ultima parte dovrete convertire e caricare il modello, con tutti i checkpoints, nel model hub e aggiungere una
model card per ogni checkpoint caricato. Leggete la nostra guida [Model sharing and uploading Page](model_sharing) per
avere familiarità con l'hub. Di solito in questa parte lavorate a fianco di Hugging face per decidere un nome che sia ok
per ogni checkpoint, per ottenere i permessi necessari per caricare il modello nell'organizzazione dell'autore di *brand_new_bert*.
Il metodo `push_to_hub`, presente in tutti i modelli `transformers`, é una maniera rapida e indolore per caricare il vostro checkpoint sull'hub:
```python
brand_new_bert.push_to_hub(
repo_path_or_name="brand_new_bert",
# Uncomment the following line to push to an organization
# organization="<ORGANIZATION>",
commit_message="Add model",
use_temp_dir=True,
)
```
Vale la pena spendere un po' di tempo per creare una model card ad-hoc per ogni checkpoint. Le model cards dovrebbero
suggerire le caratteristiche specifiche del checkpoint, *per esempio* su che dataset il checkpoint é stato pretrained o fine-tuned.
O che su che genere di task il modello lavoro? E anche buona pratica includere del codice su come usare il modello correttamente.
**13. (Opzionale) Aggiungere un notebook**
É molto utile aggiungere un notebook, che dimostri in dettaglio come *brand_new_bert* si utilizzi per fare inferenza e/o
fine-tuned su specifiche task. Non é una cosa obbligatoria da avere nella vostra PR, ma é molto utile per la community.
**14. Sottomettere la PR**
L'ultimissimo step! Ovvero il merge della PR nel main. Di solito il team Hugging face a questo punto vi avrà gia aiutato,
ma é ok prendere un po' di tempo per pulire la descirzione e commenti nel codice.
### Condividete il vostro lavoro!!
É ora tempo di prendere un po' di credito dalla communità per il vostro lavoro! Caricare e implementare un nuovo modello
é un grandissimo contributo per Transformers e l'intera community NLP. Il codice e la conversione dei modelli pre-trained sara
sicuramente utilizzato da centinaia o migliaia di sviluppatori e ricercatori. Siate fieri e orgogliosi di condividere il vostro
traguardo con l'intera community :)
** Avete create un altro modello che é super facile da usare per tutti quanti nella community! 🤯**
|
transformers/docs/source/it/add_new_model.md/0
|
{
"file_path": "transformers/docs/source/it/add_new_model.md",
"repo_id": "transformers",
"token_count": 17225
}
| 292
|
<!--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.
-->
# Inferenza Efficiente su GPU Multiple
Questo documento contiene informazioni su come fare inferenza in maniera efficiente su GPU multiple.
<Tip>
Nota: Un setup con GPU multiple può utilizzare la maggior parte delle strategie descritte nella [sezione con GPU singola](./perf_infer_gpu_one). Tuttavia, è necessario conoscere delle tecniche semplici che possono essere utilizzate per un risultato migliore.
</Tip>
## `BetterTransformer` per inferenza più rapida
Abbiamo recentemente integrato `BetterTransformer` per inferenza più rapida su multi-GPU per modelli su testo, immagini e audio. Controlla il documento con queste integrazioni [qui](https://huggingface.co/docs/optimum/bettertransformer/overview) per maggiori dettagli.
|
transformers/docs/source/it/perf_infer_gpu_many.md/0
|
{
"file_path": "transformers/docs/source/it/perf_infer_gpu_many.md",
"repo_id": "transformers",
"token_count": 420
}
| 293
|
<!--
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
⚠️ このファイルはMarkdown形式ですが、特定の文法が含まれており、通常のMarkdownビューアーでは正しく表示されない場合があります。
-->
# How to add a model to 🤗 Transformers?
🤗 Transformersライブラリは、コミュニティの貢献者のおかげで新しいモデルを提供できることがよくあります。
しかし、これは難しいプロジェクトであり、🤗 Transformersライブラリと実装するモデルについての深い知識が必要です。
Hugging Faceでは、コミュニティの多くの人々に積極的にモデルを追加する力を与えようと努力しており、
このガイドをまとめて、PyTorchモデルを追加するプロセスを説明します（[PyTorchがインストールされていることを確認してください](https://pytorch.org/get-started/locally/)）。
この過程で、以下のことを学びます：
- オープンソースのベストプラクティスに関する洞察
- 最も人気のある深層学習ライブラリの設計原則を理解する
- 大規模なモデルを効率的にテストする方法を学ぶ
- `black`、`ruff`、および`make fix-copies`などのPythonユーティリティを統合して、クリーンで読みやすいコードを確保する方法を学ぶ
Hugging Faceチームのメンバーがサポートを提供するので、一人ぼっちになることはありません。 🤗 ❤️
さあ、始めましょう！🤗 Transformersで見たいモデルについての[New model addition](https://github.com/huggingface/transformers/issues/new?assignees=&labels=New+model&template=new-model-addition.yml)のイシューを開いてください。
特定のモデルを提供することに特にこだわりがない場合、[New model label](https://github.com/huggingface/transformers/labels/New%20model)で未割り当てのモデルリクエストがあるかどうかを確認して、それに取り組むことができます。
新しいモデルリクエストを開いたら、最初のステップは🤗 Transformersをよく理解することです！
## General overview of 🤗 Transformers
まず、🤗 Transformersの一般的な概要を把握する必要があります。🤗 Transformersは非常に意見が分かれるライブラリですので、
ライブラリの哲学や設計選択について同意できない可能性があります。ただし、私たちの経験から、ライブラリの基本的な設計選択と哲学は、
🤗 Transformersを効率的にスケーリングし、適切なレベルで保守コストを抑えるために不可欠です。
ライブラリの理解を深めるための良い出発点は、[哲学のドキュメント](philosophy)を読むことです。
私たちの作業方法の結果、すべてのモデルに適用しようとするいくつかの選択肢があります：
- 一般的に、抽象化よりも構成が優先されます。
- コードの重複は、読みやすさやアクセス可能性を大幅に向上させる場合、必ずしも悪いわけではありません。
- モデルファイルはできるだけ自己完結的であるべきで、特定のモデルのコードを読む際には、理想的には該当する`modeling_....py`ファイルのみを見る必要があります。
私たちの意見では、このライブラリのコードは単なる製品を提供する手段だけでなく、*例えば、推論のためにBERTを使用する能力*などの製品そのもの.
### Overview of models
モデルを正常に追加するためには、モデルとその設定、[`PreTrainedModel`]、および[`PretrainedConfig`]の相互作用を理解することが重要です。
例示的な目的で、🤗 Transformersに追加するモデルを「BrandNewBert」と呼びます。
以下をご覧ください：
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers_overview.png"/>
ご覧のように、🤗 Transformersでは継承を使用していますが、抽象化のレベルを最小限に保っています。
ライブラリ内のどのモデルにも、抽象化のレベルが2つを超えることはありません。
`BrandNewBertModel` は `BrandNewBertPreTrainedModel` を継承し、さらに[`PreTrainedModel`]を継承しています。
これだけです。
一般的なルールとして、新しいモデルは[`PreTrainedModel`]にのみ依存するようにしたいと考えています。
すべての新しいモデルに自動的に提供される重要な機能は、[`~PreTrainedModel.from_pretrained`]および
[`~PreTrainedModel.save_pretrained`]です。
これらはシリアライゼーションとデシリアライゼーションに使用されます。
`BrandNewBertModel.forward`などの他の重要な機能は、新しい「modeling_brand_new_bert.py」スクリプトで完全に定義されるべきです。
次に、特定のヘッドレイヤーを持つモデル（たとえば `BrandNewBertForMaskedLM` ）が `BrandNewBertModel` を継承するのではなく、
抽象化のレベルを低く保つために、そのフォワードパスで `BrandNewBertModel` を呼び出すコンポーネントとして使用されるようにしたいと考えています。
新しいモデルには常に `BrandNewBertConfig` という設定クラスが必要です。この設定は常に[`PreTrainedModel`]の属性として保存され、
したがって、`BrandNewBertPreTrainedModel`から継承するすべてのクラスで`config`属性を介してアクセスできます。
```python
model = BrandNewBertModel.from_pretrained("brandy/brand_new_bert")
model.config # model has access to its config
```
モデルと同様に、設定は[`PretrainedConfig`]から基本的なシリアル化および逆シリアル化の機能を継承しています。注意すべきは、設定とモデルは常に2つの異なる形式にシリアル化されることです - モデルは*pytorch_model.bin*ファイルに、設定は*config.json*ファイルにシリアル化されます。[`~PreTrainedModel.save_pretrained`]を呼び出すと、自動的に[`~PretrainedConfig.save_pretrained`]も呼び出され、モデルと設定の両方が保存されます。
### Code style
新しいモデルをコーディングする際には、Transformersは意見があるライブラリであり、コードの書き方に関していくつかの独自の考え方があります :-)
1. モデルのフォワードパスはモデリングファイルに完全に記述され、ライブラリ内の他のモデルとは完全に独立している必要があります。他のモデルからブロックを再利用したい場合、コードをコピーしてトップに`# Copied from`コメントを付けて貼り付けます（良い例は[こちら](https://github.com/huggingface/transformers/blob/v4.17.0/src/transformers/models/roberta/modeling_roberta.py#L160)、コピーに関する詳細なドキュメンテーションは[ここ](pr_checks#check-copies)を参照してください）。
2. コードは完全に理解可能でなければなりません。これは記述的な変数名を選択し、省略形を避けるべきであることを意味します。例えば、`act`ではなく`activation`が好まれます。1文字の変数名は、forループ内のインデックスでない限り、強く非推奨です。
3. より一般的に、魔法のような短いコードよりも長くて明示的なコードを好みます。
4. PyTorchでは`nn.Sequential`をサブクラス化せずに、`nn.Module`をサブクラス化し、フォワードパスを記述し、コードを使用する他の人が簡単にデバッグできるようにします。プリントステートメントやブレークポイントを追加してデバッグできるようにします。
5. 関数のシグネチャは型アノテーションを付けるべきです。その他の部分に関しては、型アノテーションよりも良い変数名が読みやすく理解しやすいことがあります。
### Overview of tokenizers
まだ完了していません :-( このセクションは近日中に追加されます！
## Step-by-step recipe to add a model to 🤗 Transformers
モデルを追加する方法は人それぞれ異なるため、他のコントリビューターが🤗 Transformersにモデルを追加する際の要約を確認することが非常に役立つ場合があります。以下は、他のコントリビューターが🤗 Transformersにモデルをポートする際のコミュニティブログ投稿のリストです。
1. [GPT2モデルのポーティング](https://medium.com/huggingface/from-tensorflow-to-pytorch-265f40ef2a28) by [Thomas](https://huggingface.co/thomwolf)
2. [WMT19 MTモデルのポーティング](https://huggingface.co/blog/porting-fsmt) by [Stas](https://huggingface.co/stas)
経験から言えることは、モデルを追加する際に最も重要なことは次のようになります：
- 車輪の再発明をしないでください！新しい🤗 Transformersモデルのために追加するコードのほとんどはすでに🤗 Transformers内のどこかに存在しています。類似した既存のモデルやトークナイザを見つけるために、いくつかの時間をかけて探すことが重要です。[grep](https://www.gnu.org/software/grep/)と[rg](https://github.com/BurntSushi/ripgrep)はあなたの友達です。モデルのトークナイザは1つのモデル実装に基づいているかもしれませんが、モデルのモデリングコードは別の実装に基づいていることがあることに注意してください。例えば、FSMTのモデリングコードはBARTに基づいており、FSMTのトークナイザコードはXLMに基づいています。
- これは科学的な課題よりもエンジニアリングの課題です。モデルの論文の理論的な側面をすべて理解しようとするよりも、効率的なデバッグ環境を作成するために時間を費やすべきです。
- 行き詰まった場合は助けを求めてください！モデルは🤗 Transformersのコアコンポーネントであり、Hugging Faceではモデルを追加するための各ステップでお手伝いするのを喜んでいます。進行がないことに気付いた場合は、進展していないことを気にしないでください。
以下では、🤗 Transformersにモデルをポートする際に最も役立つと考えられる一般的なレシピを提供しようとしています。
次のリストは、モデルを追加するために行う必要があるすべてのことの要約であり、To-Doリストとして使用できます：
- ☐ （オプション）モデルの理論的な側面を理解しました
- ☐ 🤗 Transformersの開発環境を準備しました
- ☐ オリジナルのリポジトリのデバッグ環境をセットアップしました
- ☐ `forward()` パスをオリジナルのリポジトリとチェックポイントで正常に実行するスクリプトを作成しました
- ☐ モデルの骨格を🤗 Transformersに正常に追加しました
- ☐ オリジナルのチェックポイントを🤗 Transformersのチェックポイントに正常に変換しました
- ☐ 🤗 Transformersで実行される `forward()` パスを正常に実行し、オリジナルのチェックポイントと同一の出力を得ました
- ☐ 🤗 Transformersでのモデルテストを完了しました
- ☐ 🤗 Transformersにトークナイザを正常に追加しました
- ☐ エンドツーエンドの統合テストを実行しました
- ☐ ドキュメントを完成させました
- ☐ モデルのウェイトをHubにアップロードしました
- ☐ プルリクエストを提出しました
- ☐ （オプション）デモノートブックを追加しました
まず、通常、`BrandNewBert`の理論的な理解を深めることをお勧めします。
ただし、もしモデルの理論的な側面を「実務中に理解する」方が好ましい場合、`BrandNewBert`のコードベースに直接アクセスするのも問題ありません。
このオプションは、エンジニアリングのスキルが理論的なスキルよりも優れている場合、
`BrandNewBert`の論文を理解するのに苦労している場合、または科学的な論文を読むよりもプログラミングを楽しんでいる場合に適しています。
### 1. (Optional) Theoretical aspects of BrandNewBert
BrandNewBertの論文がある場合、その説明を読むための時間を取るべきです。論文の中には理解が難しい部分があるかもしれません。
その場合でも心配しないでください。目標は論文の深い理論的理解を得ることではなく、
🤗 Transformersでモデルを効果的に再実装するために必要な情報を抽出することです。
ただし、理論的な側面にあまり多くの時間をかける必要はありません。代わりに、実践的な側面に焦点を当てましょう。具体的には次の点です：
- *brand_new_bert*はどの種類のモデルですか？ BERTのようなエンコーダーのみのモデルですか？ GPT2のようなデコーダーのみのモデルですか？ BARTのようなエンコーダー-デコーダーモデルですか？
[model_summary](model_summary)を参照して、これらの違いについて詳しく知りたい場合があります。
- *brand_new_bert*の応用分野は何ですか？ テキスト分類ですか？ テキスト生成ですか？ Seq2Seqタスク、例えば要約ですか？
- モデルをBERT/GPT-2/BARTとは異なるものにする新しい機能は何ですか？
- 既存の[🤗 Transformersモデル](https://huggingface.co/transformers/#contents)の中で*brand_new_bert*に最も似ているモデルはどれですか？
- 使用されているトークナイザの種類は何ですか？ SentencePieceトークナイザですか？ WordPieceトークナイザですか？ BERTやBARTで使用されているトークナイザと同じですか？
モデルのアーキテクチャの良い概要を得たと感じたら、Hugging Faceチームに質問を送ることができます。
これにはモデルのアーキテクチャ、注意層などに関する質問が含まれるかもしれません。
私たちは喜んでお手伝いします。
### 2. Next prepare your environment
1. リポジトリのページで「Fork」ボタンをクリックして、[リポジトリ](https://github.com/huggingface/transformers)をフォークします。
これにより、コードのコピーがGitHubユーザーアカウントの下に作成されます。
2. ローカルディスクにある`transformers`フォークをクローンし、ベースリポジトリをリモートとして追加します：
```bash
git clone https://github.com/[your Github handle]/transformers.git
cd transformers
git remote add upstream https://github.com/huggingface/transformers.git
```
```bash
python -m venv .env
source .env/bin/activate
pip install -e ".[dev]"
```
3. 開発環境をセットアップするために、次のコマンドを実行してください：
```bash
python -m venv .env
source .env/bin/activate
pip install -e ".[dev]"
```
お使いのOSに応じて、およびTransformersのオプションの依存関係の数が増えているため、このコマンドでエラーが発生する可能性があります。
その場合は、作業しているDeep Learningフレームワーク（PyTorch、TensorFlow、および/またはFlax）をインストールし、次の手順を実行してください：
```bash
pip install -e ".[quality]"
```
これはほとんどのユースケースには十分であるはずです。その後、親ディレクトリに戻ることができます。
```bash
cd ..
```
4. Transformersに*brand_new_bert*のPyTorchバージョンを追加することをお勧めします。PyTorchをインストールするには、
https://pytorch.org/get-started/locally/ の指示に従ってください。
**注意:** CUDAをインストールする必要はありません。新しいモデルをCPUで動作させることで十分です。
5. *brand_new_bert*を移植するには、元のリポジトリへのアクセスも必要です。
```bash
git clone https://github.com/org_that_created_brand_new_bert_org/brand_new_bert.git
cd brand_new_bert
pip install -e .
```
*brand_new_bert*を🤗 Transformersにポートするための開発環境を設定しました。
### 3.-4. Run a pretrained checkpoint using the original repository
最初に、オリジナルの*brand_new_bert*リポジトリで作業します。通常、オリジナルの実装は非常に「研究的」であり、ドキュメンテーションが不足していたり、コードが理解しにくいことがあります。しかし、これが*brand_new_bert*を再実装する動機となるべきです。Hugging Faceでは、主要な目標の1つが、動作するモデルを取り、それをできるだけ**アクセス可能でユーザーフレンドリーで美しい**ものに書き直すことです。これは、🤗 Transformersにモデルを再実装する最も重要な動機です - 複雑な新しいNLP技術を**誰にでも**アクセス可能にしようとする試みです。
まず、オリジナルのリポジトリに入り込むことから始めるべきです。
公式の事前学習済みモデルをオリジナルのリポジトリで正常に実行することは、通常、**最も困難な**ステップです。
私たちの経験から、オリジナルのコードベースに慣れるのに時間をかけることが非常に重要です。以下のことを理解する必要があります：
- 事前学習済みの重みをどこで見つけるか？
- 対応するモデルに事前学習済みの重みをロードする方法は？
- モデルから独立してトークナイザを実行する方法は？
- 1つのフォワードパスを追跡して、単純なフォワードパスに必要なクラスと関数がわかるようにします。通常、これらの関数だけを再実装する必要があります。
- モデルの重要なコンポーネントを特定できること：モデルのクラスはどこにありますか？モデルのサブクラス、*例* EncoderModel、DecoderModelがありますか？自己注意レイヤーはどこにありますか？複数の異なる注意レイヤー、*例* *自己注意*、*クロスアテンション*などが存在しますか？
- オリジナルのリポジトリの環境でモデルをデバッグする方法は？*print*ステートメントを追加する必要があるか、*ipdb*のような対話型デバッガを使用できるか、PyCharmのような効率的なIDEを使用してモデルをデバッグする必要がありますか？
重要なのは、ポーティングプロセスを開始する前に、オリジナルのリポジトリでコードを**効率的に**デバッグできることです！また、これはオープンソースライブラリで作業していることを覚えておいてください。オリジナルのリポジトリでコードを調べる誰かを歓迎するために、問題をオープンにしたり、プルリクエストを送信したりすることをためらわないでください。このリポジトリのメンテナーは、彼らのコードを調べてくれる人に対して非常に喜んでいる可能性が高いです！
この段階では、オリジナルのモデルのデバッグにどのような環境と戦略を使用するかは、あなた次第です。最初にオリジナルのリポジトリに関するコードをデバッグできることが非常に重要です。また、GPU環境をセットアップすることはお勧めしません。まず、CPU上で作業し、モデルがすでに🤗 Transformersに正常にポートされていることを確認します。最後に、モデルがGPU上でも期待通りに動作するかどうかを検証する必要があります。
一般的に、オリジナルのモデルを実行するための2つのデバッグ環境があります：
- [Jupyter notebooks](https://jupyter.org/) / [google colab](https://colab.research.google.com/notebooks/intro.ipynb)
- ローカルなPythonスクリプト。
Jupyterノートブックは、セルごとに実行できるため、論理的なコンポーネントをより分割し、中間結果を保存できるため、デバッグサイクルが速くなるという利点があります。また、ノートブックは他の共同作業者と簡単に共有できることが多く、Hugging Faceチームに助けを求める場合に非常に役立つ場合があります。Jupyterノートブックに精通している場合、それ
```python
model = BrandNewBertModel.load_pretrained_checkpoint("/path/to/checkpoint/")
input_ids = [0, 4, 5, 2, 3, 7, 9] # vector of input ids
original_output = model.predict(input_ids)
```
デバッグ戦略については、通常、いくつかの選択肢があります：
- 元のモデルを多くの小さなテスト可能なコンポーネントに分解し、それぞれに対して前方パスを実行して検証します
- 元のモデルを元のトークナイザと元のモデルにのみ分解し、それらに対して前方パスを実行し、検証のために中間のプリントステートメントまたはブレークポイントを使用します
再度、どの戦略を選択するかはあなた次第です。元のコードベースに依存することが多く、元のコードベースに応じて一方または他方が有利なことがあります。
元のコードベースがモデルを小さなサブコンポーネントに分解できる場合、*例えば*元のコードベースが簡単にイーガーモードで実行できる場合、それを行う価値が通常あります。最初からより難しい方法を選択することにはいくつかの重要な利点があります：
- 後で元のモデルを🤗 Transformersの実装と比較する際に、各コンポーネントが対応する🤗 Transformers実装のコンポーネントと一致することを自動的に検証できるため、視覚的な比較に依存せずに済みます
- 大きな問題を小さな問題に分解する、つまり個々のコンポーネントのみをポーティングする問題に分割するのに役立ち、作業を構造化するのに役立ちます
- モデルを論理的な意味のあるコンポーネントに分割することで、モデルの設計をよりよく理解しやすくし、モデルをよりよく理解するのに役立ちます
- 後で、コンポーネントごとのテストを行うことで、コードを変更し続ける際にリグレッションが発生しないことを確認するのに役立ちます
[Lysandreの](https://gist.github.com/LysandreJik/db4c948f6b4483960de5cbac598ad4ed) ELECTRAの統合チェックは、これがどのように行われるかの良い例です。
ただし、元のコードベースが非常に複雑で、中間コンポーネントをコンパイルモードで実行することしか許可しない場合、モデルを小さなテスト可能なサブコンポーネントに分解することが時間がかかりすぎるか、不可能であることがあります。
良い例は[T5のMeshTensorFlow](https://github.com/tensorflow/mesh/tree/master/mesh_tensorflow)ライブラリであり、非常に複雑でモデルをサブコンポーネントに分解する簡単な方法を提供しないことがあります。このようなライブラリでは、通常、プリントステートメントを検証することに依存します。
どの戦略を選択しても、推奨される手順は通常同じで、最初のレイヤーからデバッグを開始し、最後のレイヤーからデバッグを行うべきです。
通常、以下の順序で次のレイヤーからの出力を取得することをお勧めします：
1. モデルに渡された入力IDを取得する
2. 単語の埋め込みを取得する
3. 最初のTransformerレイヤーの入力を取得する
4. 最初のTransformerレイヤーの出力を取得する
5. 次のn - 1つのTransformerレイヤーの出力を取得する
6. BrandNewBertモデル全体の出力を取得する
入力IDは整数の配列である必要があり、*例：* `input_ids = [0, 4, 4, 3, 2, 4, 1, 7, 19]` のようになります。
以下のレイヤーの出力は多次元の浮動小数点配列であることが多く、次のようになることがあります：
```
[[
[-0.1465, -0.6501, 0.1993, ..., 0.1451, 0.3430, 0.6024],
[-0.4417, -0.5920, 0.3450, ..., -0.3062, 0.6182, 0.7132],
[-0.5009, -0.7122, 0.4548, ..., -0.3662, 0.6091, 0.7648],
...,
[-0.5613, -0.6332, 0.4324, ..., -0.3792, 0.7372, 0.9288],
[-0.5416, -0.6345, 0.4180, ..., -0.3564, 0.6992, 0.9191],
[-0.5334, -0.6403, 0.4271, ..., -0.3339, 0.6533, 0.8694]]],
```
🤗 Transformersに追加されるすべてのモデルは、統合テストを数回合格することが期待されており、元のモデルと🤗 Transformersで再実装されたバージョンが、0.001の精度までまったく同じ出力を提供する必要があります。
異なるライブラリフレームワークで同じモデルを書いた場合、わずかに異なる出力を返すことが正常であるため、誤差許容値として1e-3（0.001）を受け入れています。モデルがほぼ同じ出力を返すだけでは不十分で、ほぼ同一である必要があります。そのため、🤗 Transformersバージョンの中間出力を元の*brand_new_bert*の実装の中間出力と複数回にわたって比較することになるでしょう。その際、元のリポジトリの**効率的な**デバッグ環境が非常に重要です。以下は、デバッグ環境をできるだけ効率的にするためのアドバイスです。
- 中間結果をデバッグする最適な方法を見つける。元のリポジトリはPyTorchで書かれていますか？その場合、元のモデルをより小さなサブコンポーネントに分解して中間値を取得する長いスクリプトを書くことがおそらく適切です。元のリポジトリがTensorflow 1で書かれている場合、[tf.print](https://www.tensorflow.org/api_docs/python/tf/print)などのTensorFlowのプリント操作を使用して中間値を出力する必要があるかもしれません。元のリポジトリがJaxで書かれている場合、フォワードパスの実行時にモデルが**jittedされていない**ことを確認してください。例：[このリンク](https://github.com/google/jax/issues/196)をチェック。
- 使用可能な最小の事前学習済みチェックポイントを使用します。チェックポイントが小さいほど、デバッグサイクルが速くなります。事前学習済みモデルがフォワードパスに10秒以上かかる場合、効率的ではありません。非常に大きなチェックポイントしか利用できない場合、新しい環境でランダムに初期化されたウェイトを持つダミーモデルを作成し、それらのウェイトを🤗 Transformersバージョンのモデルと比較する方が良いかもしれません。
- 元のリポジトリでフォワードパスを呼び出す最も簡単な方法を使用していることを確認してください。理想的には、元のリポジトリで**単一のフォワードパス**を呼び出す関数を見つけたいです。これは通常「predict」、「evaluate」、「forward」、「__call__」と呼ばれます。複数回「forward」を呼び出す関数をデバッグしたくありません。例：テキストを生成するために「autoregressive_sample」、「generate」と呼ばれる関数。
- トークナイゼーションとモデルの「フォワード」パスを分離しようとしてください。元のリポジトリが入力文字列を入力する必要がある例を示す場合、フォワードコール内で文字列入力が入力IDに変更される場所を特定し、このポイントから開始します。これは、スクリプトを自分で書くか、入力文字列ではなく入力IDを直接入力できるように元のコードを変更する必要があるかもしれません。
- デバッグセットアップ内のモデルがトレーニングモードではないことを確認してください。トレーニングモードでは、モデル内の複数のドロップアウトレイヤーのためにランダムな出力が生成されることがあります。デバッグ環境のフォワードパスが**決定論的**であることを確認し、ドロップアウトレイヤーが使用されないようにします。または、新しい実装が同じフレームワーク内にある場合、*transformers.utils.set_seed*を使用してください。
以下のセクションでは、*brand_new_bert*についてこれを具体的にどのように行うかについての詳細/ヒントを提供します。
### 5.-14. Port BrandNewBert to 🤗 Transformers
次に、ついに新しいコードを🤗 Transformersに追加できます。🤗 Transformersのフォークのクローンに移動してください：
```bash
cd transformers
```
特別なケースとして、既存のモデルと完全に一致するアーキテクチャのモデルを追加する場合、
[このセクション](#write-a-conversion-script)で説明されているように、変換スクリプトを追加するだけで済みます。
この場合、既存のモデルの完全なモデルアーキテクチャを再利用できます。
それ以外の場合は、新しいモデルの生成を開始しましょう。 次のスクリプトを使用して、以下から始まるモデルを追加することをお勧めします。
既存のモデル:
```bash
transformers-cli add-new-model-like
```
モデルの基本情報を入力するためのアンケートが表示されます。
**主要な huggingface/transformers リポジトリでプルリクエストを開く**
自動生成されたコードを適応し始める前に、🤗 Transformers に「作業中（WIP）」プルリクエストを開くタイミングです。
例：「[WIP] *brand_new_bert* を追加」などです。
これにより、ユーザーと Hugging Face チームが🤗 Transformers にモデルを統合する作業を並行して行うことができます。
以下の手順を実行してください：
1. メインブランチから分かりやすい名前のブランチを作成します。
```bash
git checkout -b add_brand_new_bert
```
2. 自動生成されたコードをコミットしてください:
```bash
git add .
git commit
```
3. 現在の main ブランチにフェッチしてリベース
```bash
git fetch upstream
git rebase upstream/main
```
4. 変更をあなたのアカウントにプッシュするには、次のコマンドを使用します：
```bash
git push -u origin a-descriptive-name-for-my-changes
```
5. 満足したら、GitHub上のフォークのウェブページに移動します。[プルリクエスト]をクリックします。将来の変更に備えて、Hugging Face チームのメンバーのGitHubハンドルをレビュアーとして追加してください。
6. GitHubのプルリクエストウェブページの右側にある「ドラフトに変換」をクリックして、PRをドラフトに変更します。
以下では、進捗があった場合は常に作業をコミットし、プッシュしてプルリクエストに表示されるようにしてください。さらに、定期的にメインからの最新の変更を取り込むために、次のように行うことを忘れないでください：
```bash
git fetch upstream
git merge upstream/main
```
一般的に、モデルや実装に関する質問はPull Request (PR) で行い、PR内で議論し、解決します。
これにより、Hugging Face チームは新しいコードをコミットする際や質問がある場合に常に通知を受けることができます。
質問や問題が解決された際に、問題や質問が理解されやすいように、Hugging Face チームにコードを指摘することが非常に役立ちます。
このためには、「Files changed」タブに移動してすべての変更を表示し、質問したい行に移動して「+」シンボルをクリックしてコメントを追加します。
質問や問題が解決された場合は、作成されたコメントの「Resolve」ボタンをクリックできます。
同様に、Hugging Face チームはコードをレビューする際にコメントを開きます。
PR上でのほとんどの質問はGitHub上で行うことをお勧めします。
一般的な質問に関しては、公にはあまり役立たない質問については、SlackやメールでHugging Face チームに連絡することもできます。
**5. 生成されたモデルコードを"brand_new_bert"に適応させる**
最初に、モデル自体に焦点を当て、トークナイザには気にしないでください。
関連するコードは、生成されたファイル`src/transformers/models/brand_new_bert/modeling_brand_new_bert.py`および`src/transformers/models/brand_new_bert/configuration_brand_new_bert.py`で見つかるはずです。
さて、ついにコーディングを始めることができます :smile:。
`src/transformers/models/brand_new_bert/modeling_brand_new_bert.py`にある生成されたコードは、エンコーダーのみのモデルであればBERTと同じアーキテクチャを持っているか、エンコーダー-デコーダーモデルであればBARTと同じアーキテクチャを持っているはずです。
この段階では、モデルの理論的な側面について学んだことを思い出すべきです。つまり、「このモデルはBERTまたはBARTとどのように異なるのか？」ということです。
これらの変更を実装しますが、これは通常、セルフアテンションレイヤー、正規化レイヤーの順序などを変更することを意味します。
再び、あなたのモデルがどのように実装されるべきかをより良く理解するために、Transformers内に既存のモデルの類似アーキテクチャを見ることが役立つことがあります。
この時点では、コードが完全に正確またはクリーンである必要はありません。
むしろ、まずは必要なコードの最初の*クリーンでない*コピー＆ペーストバージョンを
`src/transformers/models/brand_new_bert/modeling_brand_new_bert.py`に追加し、必要なコードがすべて追加されていると感じるまで改善/修正を反復的に行うことがお勧めです。
私たちの経験から、必要なコードの最初のバージョンを迅速に追加し、次のセクションで説明する変換スクリプトを使用してコードを繰り返し改善/修正する方が効率的であることが多いです。
この時点で動作する必要があるのは、🤗 Transformersの"brand_new_bert"の実装をインスタンス化できることだけです。つまり、以下のコマンドが機能する必要があります：
```python
from transformers import BrandNewBertModel, BrandNewBertConfig
model = BrandNewBertModel(BrandNewBertConfig())
```
上記のコマンドは、`BrandNewBertConfig()` で定義されたデフォルトパラメータに従ってモデルを作成し、
すべてのコンポーネントの `init()` メソッドが正常に動作することを確認します。
すべてのランダムな初期化は、`BrandnewBertPreTrainedModel` クラスの `_init_weights` メソッドで行う必要があります。
このメソッドは、設定変数に依存するすべてのリーフモジュールを初期化する必要があります。以下は、BERT の `_init_weights` メソッドの例です：
```py
def _init_weights(self, module):
"""Initialize the weights"""
if isinstance(module, nn.Linear):
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.Embedding):
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
elif isinstance(module, nn.LayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
```
特定のモジュールに特別な初期化が必要な場合、カスタムスキームをさらに持つことができます。たとえば、
`Wav2Vec2ForPreTraining`では、最後の2つの線形層には通常のPyTorchの`nn.Linear`の初期化が必要ですが、
他のすべての層は上記のような初期化を使用する必要があります。これは以下のようにコーディングされています：
```py
def _init_weights(self, module):
"""Initialize the weights"""
if isinstance(module, Wav2Vec2ForPreTraining):
module.project_hid.reset_parameters()
module.project_q.reset_parameters()
module.project_hid._is_hf_initialized = True
module.project_q._is_hf_initialized = True
elif isinstance(module, nn.Linear):
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
if module.bias is not None:
module.bias.data.zero_()
```
`_is_hf_initialized`フラグは、サブモジュールを一度だけ初期化することを確実にするために内部で使用されます。
`module.project_q`と`module.project_hid`のためにそれを`True`に設定することで、
カスタム初期化が後で上書きされないようにし、`_init_weights`関数がそれらに適用されないようにします。
**6. 変換スクリプトを書く**
次に、*brand_new_bert* の元のリポジトリでデバッグに使用したチェックポイントを、新しく作成した 🤗 Transformers 実装の *brand_new_bert* と互換性のあるチェックポイントに変換できる変換スクリプトを書く必要があります。
変換スクリプトをゼロから書くことはお勧めされませんが、代わりに 🤗 Transformers で既に存在する類似のモデルを同じフレームワークで変換したスクリプトを調べることが良いでしょう。
通常、既存の変換スクリプトをコピーして、自分のユースケースにわずかに適応させることで十分です。
Hugging Face チームに既存のモデルに類似した変換スクリプトを教えてもらうことも躊躇しないでください。
- TensorFlowからPyTorchにモデルを移植している場合、良い出発点はBERTの変換スクリプトかもしれません [here](https://github.com/huggingface/transformers/blob/7acfa95afb8194f8f9c1f4d2c6028224dbed35a2/src/transformers/models/bert/modeling_bert.py#L91)
- PyTorchからPyTorchにモデルを移植している場合、良い出発点はBARTの変換スクリプトかもしれません [here](https://github.com/huggingface/transformers/blob/main/src/transformers/models/bart/convert_bart_original_pytorch_checkpoint_to_pytorch.py)
以下では、PyTorchモデルが層の重みをどのように保存し、層の名前を定義するかについて簡単に説明します。
PyTorchでは、層の名前は層に与えるクラス属性の名前によって定義されます。
PyTorchで `SimpleModel` というダミーモデルを定義しましょう：
```python
from torch import nn
class SimpleModel(nn.Module):
def __init__(self):
super().__init__()
self.dense = nn.Linear(10, 10)
self.intermediate = nn.Linear(10, 10)
self.layer_norm = nn.LayerNorm(10)
```
これで、このモデル定義のインスタンスを作成し、`dense`、`intermediate`、`layer_norm`のすべての重みをランダムな重みで埋めたモデルを作成できます。モデルのアーキテクチャを確認するために、モデルを印刷してみましょう。
```python
model = SimpleModel()
print(model)
```
これは以下を出力します：
```
SimpleModel(
(dense): Linear(in_features=10, out_features=10, bias=True)
(intermediate): Linear(in_features=10, out_features=10, bias=True)
(layer_norm): LayerNorm((10,), eps=1e-05, elementwise_affine=True)
)
```
層の名前はPyTorchのクラス属性の名前によって定義されています。特定の層の重み値を出力することができます：
```python
print(model.dense.weight.data)
```
ランダムに初期化された重みを確認するために
```
tensor([[-0.0818, 0.2207, -0.0749, -0.0030, 0.0045, -0.1569, -0.1598, 0.0212,
-0.2077, 0.2157],
[ 0.1044, 0.0201, 0.0990, 0.2482, 0.3116, 0.2509, 0.2866, -0.2190,
0.2166, -0.0212],
[-0.2000, 0.1107, -0.1999, -0.3119, 0.1559, 0.0993, 0.1776, -0.1950,
-0.1023, -0.0447],
[-0.0888, -0.1092, 0.2281, 0.0336, 0.1817, -0.0115, 0.2096, 0.1415,
-0.1876, -0.2467],
[ 0.2208, -0.2352, -0.1426, -0.2636, -0.2889, -0.2061, -0.2849, -0.0465,
0.2577, 0.0402],
[ 0.1502, 0.2465, 0.2566, 0.0693, 0.2352, -0.0530, 0.1859, -0.0604,
0.2132, 0.1680],
[ 0.1733, -0.2407, -0.1721, 0.1484, 0.0358, -0.0633, -0.0721, -0.0090,
0.2707, -0.2509],
[-0.1173, 0.1561, 0.2945, 0.0595, -0.1996, 0.2988, -0.0802, 0.0407,
0.1829, -0.1568],
[-0.1164, -0.2228, -0.0403, 0.0428, 0.1339, 0.0047, 0.1967, 0.2923,
0.0333, -0.0536],
[-0.1492, -0.1616, 0.1057, 0.1950, -0.2807, -0.2710, -0.1586, 0.0739,
0.2220, 0.2358]]).
```
スクリプト内の変換スクリプトでは、ランダムに初期化された重みを、対応するチェックポイント内の正確な重みで埋める必要があります。例えば、以下のように翻訳します：
```python
# retrieve matching layer weights, e.g. by
# recursive algorithm
layer_name = "dense"
pretrained_weight = array_of_dense_layer
model_pointer = getattr(model, "dense")
model_pointer.weight.data = torch.from_numpy(pretrained_weight)
```
PyTorchモデルの各ランダム初期化された重みと対応する事前学習済みチェックポイントの重みが
**形状と名前の両方**で正確に一致することを確認する必要があります。
これを行うために、形状に対するassertステートメントを追加し、チェックポイントの重みの名前を出力することが
**必要不可欠**です。例えば、次のようなステートメントを追加する必要があります：
```python
assert (
model_pointer.weight.shape == pretrained_weight.shape
), f"Pointer shape of random weight {model_pointer.shape} and array shape of checkpoint weight {pretrained_weight.shape} mismatched"
```
また、両方の重みの名前を印刷して、一致していることを確認する必要があります。例えば、次のようにします：
```python
logger.info(f"Initialize PyTorch weight {layer_name} from {pretrained_weight.name}")
```
もし形状または名前のいずれかが一致しない場合、おそらく誤って🤗 Transformersの実装に初期化されたレイヤーに間違ったチェックポイントの重みを割り当ててしまった可能性があります。
誤った形状は、おそらく`BrandNewBertConfig()`での設定パラメーターが、変換したいチェックポイントで使用されたものと正確に一致しないためです。
ただし、PyTorchのレイヤーの実装によっては、重みを事前に転置する必要がある場合もあります。
最後に、**すべて**の必要な重みが初期化されていることを確認し、初期化に使用されなかったすべてのチェックポイントの重みを表示して、モデルが正しく変換されていることを確認してください。
変換トライアルが誤った形状ステートメントまたは誤った名前割り当てで失敗するのは完全に正常です。
これはおそらく、`BrandNewBertConfig()`で誤ったパラメーターを使用したか、🤗 Transformersの実装に誤ったアーキテクチャがあるか、🤗 Transformersの実装の1つのコンポーネントの`init()`関数にバグがあるか、チェックポイントの重みの1つを転置する必要があるためです。
このステップは、以前のステップと繰り返すべきです。すべてのチェックポイントの重みが正しく🤗 Transformersモデルに読み込まれるまで繰り返すべきです。
🤗 Transformers実装に正しくチェックポイントを読み込んだ後、選択したフォルダーにモデルを保存できます `/path/to/converted/checkpoint/folder`。このフォルダには`pytorch_model.bin`ファイルと`config.json`ファイルの両方が含まれるはずです。
```python
model.save_pretrained("/path/to/converted/checkpoint/folder")
```
**7. 順伝播（forward pass）の実装**
🤗 Transformers実装で事前学習済みの重みを正しく読み込んだ後、順伝播が正しく実装されていることを確認する必要があります。[元のリポジトリを理解する](#3-4-run-a-pretrained-checkpoint-using-the-original-repository)で、元のリポジトリを使用してモデルの順伝播を実行するスクリプトをすでに作成しました。今度は、元のリポジトリの代わりに🤗 Transformers実装を使用して類似のスクリプトを作成する必要があります。以下のようになります：
```python
model = BrandNewBertModel.from_pretrained("/path/to/converted/checkpoint/folder")
input_ids = [0, 4, 4, 3, 2, 4, 1, 7, 19]
output = model(input_ids).last_hidden_states
```
🤗 Transformersの実装と元のモデルの実装が最初の実行で完全に同じ出力を提供しないか、
フォワードパスでエラーが発生する可能性が非常に高いです。失望しないでください - これは予想されていることです！
まず、フォワードパスがエラーをスローしないことを確認する必要があります。
間違った次元が使用され、*次元の不一致*エラーや、誤ったデータ型オブジェクトが使用されることがよくあります。
例えば、`torch.long`ではなく`torch.float32`が使用されます。特定のエラーを解決できない場合は、
Hugging Faceチームに助けを求めることを躊躇しないでください。
🤗 Transformers実装が正しく機能することを確認する最終的な部分は、出力が`1e-3`の精度で同等であることを確認することです。
まず、出力の形状が同一であること、つまりスクリプトの🤗 Transformers実装と元の実装の両方で`outputs.shape`が同じ値を生成する必要があります。
次に、出力値が同一であることを確認する必要があります。
これは新しいモデルを追加する際の最も難しい部分の1つです。
出力が同一でない理由の一般的な間違いは以下の通りです。
- 一部のレイヤーが追加されていない、つまり*活性化*レイヤーが追加されていないか、リザバル接続が忘れられている
- 単語埋め込み行列が結ばれていない
- オリジナルの実装がオフセットを使用しているため、誤った位置埋め込みが使用されている
- フォワードパス中にドロップアウトが適用されています。これを修正するには、*model.trainingがFalse*であることを確認し、フォワードパス中に誤ってドロップアウトレイヤーがアクティブ化されないようにします。
*つまり* [PyTorchのfunctional dropout](https://pytorch.org/docs/stable/nn.functional.html?highlight=dropout#torch.nn.functional.dropout)に*model.training*を渡します。
問題を修正する最良の方法は、通常、元の実装と🤗 Transformers実装のフォワードパスを並べて表示し、違いがあるかどうかを確認することです。
理想的には、フォワードパスの両方の実装の中間出力をデバッグ/プリントアウトして、🤗 Transformers実装が元の実装と異なる出力を示すネットワーク内の正確な位置を見つけることができます。
最初に、両方のスクリプトのハードコーディングされた`input_ids`が同一であることを確認します。
次に、`input_ids`の最初の変換（通常、単語埋め込み）の出力が同一であることを確認します。
その後、ネットワークの最後のレイヤーまで作業を進めます。
いずれかの時点で、2つの実装間で違いがあることに気付くはずで、それにより🤗 Transformers実装のバグの場所が特定されます。
経験上、元の実装と🤗 Transformers実装のフォワードパスの同じ位置に多くのプリントステートメントを追加し、
中間プレゼンテーションで同じ値を示すプリントステートメントを段階的に削除するのがシンプルかつ効果的な方法です。
両方の実装が同じ出力を生成することに自信を持っている場合、`torch.allclose(original_output, output, atol=1e-3)`を使用して出力を確認すると、最も難しい部分が完了します！
おめでとうございます - 完了する作業は簡単なものになるはずです 😊。
**8. 必要なすべてのモデルテストを追加**
この時点で、新しいモデルが正常に追加されました。
ただし、モデルがまだ必要な設計に完全に準拠していない可能性が非常に高いです。
🤗 Transformersと完全に互換性があることを確認するために、すべての一般的なテストがパスする必要があります。
Cookiecutterはおそらくモデル用のテストファイルを自動的に追加しているはずで、おそらく同じディレクトリに`tests/models/brand_new_bert/test_modeling_brand_new_bert.py`として存在します。
このテストファイルを実行して、すべての一般的なテストがパスすることを確認してください：
```bash
pytest tests/models/brand_new_bert/test_modeling_brand_new_bert.py
```
すべての一般的なテストを修正したら、今度は実行したすべての素晴らしい作業が適切にテストされていることを確認することが非常に重要です。これにより、
- a) コミュニティは*brand_new_bert*の特定のテストを見ることで、あなたの作業を簡単に理解できます。
- b) モデルへの将来の変更がモデルの重要な機能を壊さないようにすることができます。
まず、統合テストを追加する必要があります。これらの統合テストは、基本的にはデバッグスクリプトと同じことを行います。これらのモデルテストのテンプレートはCookiecutterによって既に追加されており、「BrandNewBertModelIntegrationTests」と呼ばれています。このテストを記入するだけです。これらのテストが合格していることを確認するには、次のコマンドを実行します。
```bash
RUN_SLOW=1 pytest -sv tests/models/brand_new_bert/test_modeling_brand_new_bert.py::BrandNewBertModelIntegrationTests
```
<Tip>
Windowsを使用している場合、`RUN_SLOW=1`を`SET RUN_SLOW=1`に置き換えてください。
</Tip>
次に、*brand_new_bert*に特有のすべての特徴は、別個のテスト内で追加されるべきです。
`BrandNewBertModelTester`/`BrandNewBertModelTest`の下に。この部分はよく忘れられますが、2つの点で非常に役立ちます：
- モデルの追加中に獲得した知識をコミュニティに伝え、*brand_new_bert*の特別な機能がどのように動作するかを示すことによって、知識の共有を支援します。
- 将来の貢献者は、これらの特別なテストを実行することでモデルへの変更を迅速にテストできます。
**9. トークナイザの実装**
次に、*brand_new_bert*のトークナイザを追加する必要があります。通常、トークナイザは🤗 Transformersの既存のトークナイザと同等か非常に似ています。
トークナイザが正しく動作することを確認するためには、まず、元のリポジトリ内で文字列を入力し、`input_ids`を返すスクリプトを作成することをお勧めします。
このスクリプトは、次のように見えるかもしれません（疑似コードで示します）：
```python
input_str = "This is a long example input string containing special characters .$?-, numbers 2872 234 12 and words."
model = BrandNewBertModel.load_pretrained_checkpoint("/path/to/checkpoint/")
input_ids = model.tokenize(input_str)
```
オリジナルのリポジトリを詳しく調査し、正しいトークナイザの関数を見つける必要があるかもしれません。
または、オリジナルのリポジトリのクローンを変更して、`input_ids`だけを出力するようにする必要があるかもしれません。
オリジナルのリポジトリを使用した機能的なトークナイゼーションスクリプトを作成した後、
🤗 Transformers向けの類似したスクリプトを作成する必要があります。
以下のように見えるべきです：
```python
from transformers import BrandNewBertTokenizer
input_str = "This is a long example input string containing special characters .$?-, numbers 2872 234 12 and words."
tokenizer = BrandNewBertTokenizer.from_pretrained("/path/to/tokenizer/folder/")
input_ids = tokenizer(input_str).input_ids
```
`input_ids`が同じ値を生成した場合、最終ステップとしてトークナイザのテストファイルも追加するべきです。
*brand_new_bert*のモデルングテストファイルと同様に、*brand_new_bert*のトークナイズテストファイルには、いくつかのハードコードされた統合テストが含まれるべきです。
**10. エンドツーエンド統合テストの実行**
トークナイザを追加した後、`🤗 Transformers`内の`tests/models/brand_new_bert/test_modeling_brand_new_bert.py`に
モデルとトークナイザの両方を使用するいくつかのエンドツーエンド統合テストも追加する必要があります。
このようなテストは、🤗 Transformersの実装が期待どおりに機能することを示すべきです。
意味のあるテキスト対テキストのサンプルが含まれます。有用なテキスト対テキストのサンプルには、ソースからターゲットへの翻訳ペア、記事から要約へのペア、質問から回答へのペアなどが含まれます。
ポートされたチェックポイントがダウンストリームタスクでファインチューニングされていない場合、モデルのテストに依存するだけで十分です。
モデルが完全に機能していることを確認するために、すべてのテストをGPU上で実行することもお勧めします。
モデルの内部テンソルに`.to(self.device)`ステートメントを追加するのを忘れる可能性があるため、そのようなテストではエラーが表示されることがあります。
GPUにアクセスできない場合、Hugging Faceチームが代わりにこれらのテストを実行できます。
**11. ドキュメントの追加**
これで、*brand_new_bert*の必要なすべての機能が追加されました - ほぼ完了です！残りの追加すべきことは、良いドキュメントとドキュメントページです。
Cookiecutterが`docs/source/model_doc/brand_new_bert.md`というテンプレートファイルを追加しているはずで、これを記入する必要があります。
モデルのユーザーは通常、モデルを使用する前にまずこのページを見ます。したがって、ドキュメンテーションは理解しやすく簡潔である必要があります。
モデルの使用方法を示すためにいくつかの*Tips*を追加することはコミュニティにとって非常に役立ちます。ドキュメンテーションに関しては、Hugging Faceチームに問い合わせることをためらわないでください。
次に、`src/transformers/models/brand_new_bert/modeling_brand_new_bert.py`に追加されたドキュメンテーション文字列が正しいこと、およびすべての必要な入力および出力を含んでいることを確認してください。
ドキュメンテーションの書き方とドキュメンテーション文字列のフォーマットについて詳細なガイドが[こちら](writing-documentation)にあります。
ドキュメンテーションは通常、コミュニティとモデルの最初の接触点であるため、コードと同じくらい注意深く扱うべきであることを常に念頭に置いてください。
**コードのリファクタリング**
素晴らしい、これで*brand_new_bert*に必要なすべてのコードが追加されました。
この時点で、次のようなポテンシャルなコードスタイルの誤りを訂正するために以下を実行する必要があります：
```bash
make style
```
あなたのコーディングスタイルが品質チェックをパスすることを確認してください:
```bash
make quality
```
🤗 Transformersの非常に厳格なデザインテストには、まだ合格していない可能性があるいくつかの他のテストが存在するかもしれません。
これは、ドキュメント文字列に情報が不足しているか、名前が間違っていることが原因であることが多いです。Hugging Faceチームは、ここで詰まっている場合には必ず助けてくれるでしょう。
最後に、コードが正しく機能することを確認した後、コードをリファクタリングするのは常に良いアイデアです。
すべてのテストがパスした今、追加したコードを再度確認してリファクタリングを行うのは良いタイミングです。
これでコーディングの部分は完了しました、おめでとうございます！ 🎉 あなたは素晴らしいです！ 😎
**12. モデルをモデルハブにアップロード**
最後のパートでは、すべてのチェックポイントをモデルハブに変換してアップロードし、各アップロードしたモデルチェックポイントにモデルカードを追加する必要があります。
モデルハブの機能について詳しくは、[Model sharing and uploading Page](model_sharing)を読んで理解できます。
ここでは、*brand_new_bert*の著者組織の下にモデルをアップロードできるように必要なアクセス権を取得するために、Hugging Faceチームと協力する必要があります。
`transformers`のすべてのモデルに存在する`push_to_hub`メソッドは、チェックポイントをハブにプッシュする迅速かつ効率的な方法です。
以下に、少しのコードスニペットを示します：
```python
brand_new_bert.push_to_hub("brand_new_bert")
# Uncomment the following line to push to an organization.
# brand_new_bert.push_to_hub("<organization>/brand_new_bert")
```
各チェックポイントに適切なモデルカードを作成する価値があります。モデルカードは、この特定のチェックポイントの特性をハイライトするべきです。例えば、このチェックポイントはどのデータセットで事前学習/ファインチューニングされたか、どのような下流タスクでモデルを使用すべきかを示すべきです。また、モデルの正しい使用方法に関するコードも含めるべきです。
**13.（オプション）ノートブックの追加**
*brand_new_bert*を推論または下流タスクのファインチューニングにどのように詳細に使用できるかを示すノートブックを追加することは非常に役立ちます。これはあなたのPRをマージするために必須ではありませんが、コミュニティにとって非常に有用です。
**14. 完成したPRの提出**
プログラミングが完了したら、最後のステップに移動し、PRをメインブランチにマージしましょう。通常、Hugging Faceチームはこの時点で既にあなたをサポートしているはずですが、PRに良い説明を追加し、コードにコメントを追加して、レビュアーに特定の設計の選択肢を指摘したい場合はコメントを追加することも価値があります。
### Share your work!!
さあ、コミュニティからあなたの作業に対する評価を得る時が来ました！モデルの追加を完了することは、TransformersおよびNLPコミュニティにとって重要な貢献です。あなたのコードとポートされた事前学習済みモデルは、何百人、何千人という開発者や研究者によって確実に使用されるでしょう。あなたの仕事に誇りを持ち、コミュニティとあなたの成果を共有しましょう。
**あなたはコミュニティの誰でも簡単にアクセスできる別のモデルを作成しました！ 🤯**
|
transformers/docs/source/ja/add_new_model.md/0
|
{
"file_path": "transformers/docs/source/ja/add_new_model.md",
"repo_id": "transformers",
"token_count": 27745
}
| 294
|
<!---
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): `HUGGINGFACE_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": 4809
}
| 295
|
<!--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.
-->
# AltCLIP
## 概要
AltCLIPモデルは、「[AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679v2)」という論文でZhongzhi Chen、Guang Liu、Bo-Wen Zhang、Fulong Ye、Qinghong Yang、Ledell Wuによって提案されました。AltCLIP（CLIPの言語エンコーダーの代替）は、様々な画像-テキストペアおよびテキスト-テキストペアでトレーニングされたニューラルネットワークです。CLIPのテキストエンコーダーを事前学習済みの多言語テキストエンコーダーXLM-Rに置き換えることで、ほぼ全てのタスクでCLIPに非常に近い性能を得られ、オリジナルのCLIPの能力を多言語理解などに拡張しました。
論文の要旨は以下の通りです：
*この研究では、強力なバイリンガルマルチモーダル表現モデルを訓練するための概念的に単純で効果的な方法を提案します。OpenAIによってリリースされたマルチモーダル表現モデルCLIPから開始し、そのテキストエンコーダを事前学習済みの多言語テキストエンコーダXLM-Rに交換し、教師学習と対照学習からなる2段階のトレーニングスキーマを用いて言語と画像の表現を整合させました。幅広いタスクの評価を通じて、我々の方法を検証します。ImageNet-CN、Flicker30k-CN、COCO-CNを含む多くのタスクで新たな最先端の性能を達成しました。さらに、ほぼすべてのタスクでCLIPに非常に近い性能を得ており、これはCLIPのテキストエンコーダを変更するだけで、多言語理解などの拡張を実現できることを示唆しています。*
このモデルは[jongjyh](https://huggingface.co/jongjyh)により提供されました。
## 使用上のヒントと使用例
AltCLIPの使用方法はCLIPに非常に似ています。CLIPとの違いはテキストエンコーダーにあります。私たちはカジュアルアテンションではなく双方向アテンションを使用し、XLM-Rの[CLS]トークンをテキスト埋め込みを表すものとして取ることに留意してください。
AltCLIPはマルチモーダルな視覚言語モデルです。これは画像とテキストの類似度や、ゼロショット画像分類に使用できます。AltCLIPはViTのようなTransformerを使用して視覚的特徴を、双方向言語モデルを使用してテキスト特徴を取得します。テキストと視覚の両方の特徴は、同一の次元を持つ潜在空間に射影されます。射影された画像とテキスト特徴間のドット積が類似度スコアとして使用されます。
Transformerエンコーダーに画像を与えるには、各画像を固定サイズの重複しないパッチの系列に分割し、それらを線形に埋め込みます。画像全体を表現するための[CLS]トークンが追加されます。著者は絶対位置埋め込みも追加し、結果として得られるベクトルの系列を標準的なTransformerエンコーダーに供給します。[`CLIPImageProcessor`]を使用して、モデルのために画像のサイズ変更（または拡大縮小）と正規化を行うことができます。
[`AltCLIPProcessor`]は、テキストのエンコードと画像の前処理を両方行うために、[`CLIPImageProcessor`]と[`XLMRobertaTokenizer`]を単一のインスタンスにラップします。以下の例は、[`AltCLIPProcessor`]と[`AltCLIPModel`]を使用して画像-テキスト類似スコアを取得する方法を示しています。
```python
>>> from PIL import Image
>>> import requests
>>> from transformers import AltCLIPModel, AltCLIPProcessor
>>> model = AltCLIPModel.from_pretrained("BAAI/AltCLIP")
>>> processor = AltCLIPProcessor.from_pretrained("BAAI/AltCLIP")
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> inputs = processor(text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pt", padding=True)
>>> outputs = model(**inputs)
>>> logits_per_image = outputs.logits_per_image # this is the image-text similarity score
>>> probs = logits_per_image.softmax(dim=1) # we can take the softmax to get the label probabilities
```
<Tip>
このモデルは`CLIPModel`をベースにしており、オリジナルの[CLIP](clip)と同じように使用してください。
</Tip>
## AltCLIPConfig
[[autodoc]] AltCLIPConfig
- from_text_vision_configs
## AltCLIPTextConfig
[[autodoc]] AltCLIPTextConfig
## AltCLIPVisionConfig
[[autodoc]] AltCLIPVisionConfig
## AltCLIPProcessor
[[autodoc]] AltCLIPProcessor
## AltCLIPModel
[[autodoc]] AltCLIPModel
- forward
- get_text_features
- get_image_features
## AltCLIPTextModel
[[autodoc]] AltCLIPTextModel
- forward
## AltCLIPVisionModel
[[autodoc]] AltCLIPVisionModel
- forward
|
transformers/docs/source/ja/model_doc/altclip.md/0
|
{
"file_path": "transformers/docs/source/ja/model_doc/altclip.md",
"repo_id": "transformers",
"token_count": 2400
}
| 296
|
<!--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.
-->
# Big Transfer (BiT)
## Overview
BiT モデルは、Alexander Kolesnikov、Lucas Beyer、Xiaohua Zhai、Joan Puigcerver、Jessica Yung、Sylvain Gelly によって [Big Transfer (BiT): General Visual Representation Learning](https://arxiv.org/abs/1912.11370) で提案されました。ニール・ホールズビー。
BiT は、[ResNet](resnet) のようなアーキテクチャ (具体的には ResNetv2) の事前トレーニングをスケールアップするための簡単なレシピです。この方法により、転移学習が大幅に改善されます。
論文の要約は次のとおりです。
*事前トレーニングされた表現の転送により、サンプル効率が向上し、視覚用のディープ ニューラル ネットワークをトレーニングする際のハイパーパラメーター調整が簡素化されます。大規模な教師ありデータセットでの事前トレーニングと、ターゲット タスクでのモデルの微調整のパラダイムを再検討します。私たちは事前トレーニングをスケールアップし、Big Transfer (BiT) と呼ぶシンプルなレシピを提案します。いくつかの慎重に選択されたコンポーネントを組み合わせ、シンプルなヒューリスティックを使用して転送することにより、20 を超えるデータセットで優れたパフォーマンスを実現します。 BiT は、クラスごとに 1 つのサンプルから合計 100 万のサンプルまで、驚くほど広範囲のデータ領域にわたって良好にパフォーマンスを発揮します。 BiT は、ILSVRC-2012 で 87.5%、CIFAR-10 で 99.4%、19 タスクの Visual Task Adaptation Benchmark (VTAB) で 76.3% のトップ 1 精度を達成しました。小規模なデータセットでは、BiT は ILSVRC-2012 (クラスあたり 10 例) で 76.8%、CIFAR-10 (クラスあたり 10 例) で 97.0% を達成しました。高い転写性能を実現する主要成分を詳細に分析※。
## Usage tips
- BiT モデルは、アーキテクチャの点で ResNetv2 と同等ですが、次の点が異なります: 1) すべてのバッチ正規化層が [グループ正規化](https://arxiv.org/abs/1803.08494) に置き換えられます。
2) [重みの標準化](https://arxiv.org/abs/1903.10520) は畳み込み層に使用されます。著者らは、両方の組み合わせが大きなバッチサイズでのトレーニングに役立ち、重要な効果があることを示しています。
転移学習への影響。
このモデルは、[nielsr](https://huggingface.co/nielsr) によって提供されました。
元のコードは [こちら](https://github.com/google-research/big_transfer) にあります。
## Resources
BiT を始めるのに役立つ公式 Hugging Face およびコミュニティ (🌎 で示されている) リソースのリスト。
<PipelineTag pipeline="image-classification"/>
- [`BitForImageClassification`] は、この [サンプル スクリプト](https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-classification) および [ノートブック](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/image_classification.ipynb)。
- 参照: [画像分類タスク ガイド](../tasks/image_classification)
ここに含めるリソースの送信に興味がある場合は、お気軽にプル リクエストを開いてください。審査させていただきます。リソースは、既存のリソースを複製するのではなく、何か新しいものを示すことが理想的です。
## BitConfig
[[autodoc]] BitConfig
## BitImageProcessor
[[autodoc]] BitImageProcessor
- preprocess
## BitModel
[[autodoc]] BitModel
- forward
## BitForImageClassification
[[autodoc]] BitForImageClassification
- forward
|
transformers/docs/source/ja/model_doc/bit.md/0
|
{
"file_path": "transformers/docs/source/ja/model_doc/bit.md",
"repo_id": "transformers",
"token_count": 1858
}
| 297
|
<!--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.
-->
# CLVP
## Overview
CLVP (Contrastive Language-Voice Pretrained Transformer) モデルは、James Betker によって [Better speech synthesis through scaling](https://arxiv.org/abs/2305.07243) で提案されました。
論文の要約は次のとおりです。
*近年、画像生成の分野は自己回帰変換器と DDPM の応用によって革命を起こしています。これらのアプローチは、画像生成のプロセスを段階的な確率的プロセスとしてモデル化し、大量のコンピューティングとデータを活用して画像の分布を学習します。パフォーマンスを向上させるこの方法論は、画像に限定される必要はありません。この論文では、画像生成ドメインの進歩を音声合成に適用する方法について説明します。その結果、表現力豊かなマルチ音声テキスト読み上げシステムである TorToise が誕生しました。
このモデルは [Susnato Dhar](https://huggingface.co/susnato) によって提供されました。
元のコードは [ここ](https://github.com/neonbjb/tortoise-tts) にあります。
## Usage tips
1. CLVP は Tortoise TTS モデルの不可欠な部分です。
2. CLVP を使用して、生成されたさまざまな音声候補を提供されたテキストと比較することができ、最良の音声トークンが拡散モデルに転送されます。
3. Tortoise の使用には、[`ClvpModelForConditionalGeneration.generate()`] メソッドの使用を強くお勧めします。
4. 16 kHz を期待する他のオーディオ モデルとは対照的に、CLVP モデルはオーディオが 22.05 kHz でサンプリングされることを期待していることに注意してください。
## Brief Explanation:
- [`ClvpTokenizer`] はテキスト入力をトークン化し、[`ClvpFeatureExtractor`] は目的のオーディオからログ メル スペクトログラムを抽出します。
- [`ClvpConditioningEncoder`] は、これらのテキスト トークンとオーディオ表現を取得し、テキストとオーディオに基づいて条件付けされた埋め込みに変換します。
- [`ClvpForCausalLM`] は、これらの埋め込みを使用して複数の音声候補を生成します。
- 各音声候補は音声エンコーダ ([`ClvpEncoder`]) を通過してベクトル表現に変換され、テキスト エンコーダ ([`ClvpEncoder`]) はテキスト トークンを同じ潜在空間に変換します。
- 最後に、各音声ベクトルをテキスト ベクトルと比較して、どの音声ベクトルがテキスト ベクトルに最も類似しているかを確認します。
- [`ClvpModelForConditionalGeneration.generate()`] は、上記のすべてのロジックを 1 つのメソッドに圧縮します。
例 ：
```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))
>>> sample = ds[0]["audio"]
>>> # 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=sample["array"], sampling_rate=sample["sampling_rate"], text=text, return_tensors="pt")
>>> generated_output = model.generate(**processor_output)
```
## ClvpConfig
[[autodoc]] ClvpConfig
- from_sub_model_configs
## ClvpEncoderConfig
[[autodoc]] ClvpEncoderConfig
## ClvpDecoderConfig
[[autodoc]] ClvpDecoderConfig
## ClvpTokenizer
[[autodoc]] ClvpTokenizer
- save_vocabulary
## ClvpFeatureExtractor
[[autodoc]] ClvpFeatureExtractor
- __call__
## ClvpProcessor
[[autodoc]] ClvpProcessor
- __call__
- decode
- batch_decode
## ClvpModelForConditionalGeneration
[[autodoc]] ClvpModelForConditionalGeneration
- forward
- generate
- get_text_features
- get_speech_features
## ClvpForCausalLM
[[autodoc]] ClvpForCausalLM
## ClvpModel
[[autodoc]] ClvpModel
## ClvpEncoder
[[autodoc]] ClvpEncoder
## ClvpDecoder
[[autodoc]] ClvpDecoder
|
transformers/docs/source/ja/model_doc/clvp.md/0
|
{
"file_path": "transformers/docs/source/ja/model_doc/clvp.md",
"repo_id": "transformers",
"token_count": 2038
}
| 298
|
<!--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.
-->
# DeiT
## Overview
DeiT モデルは、Hugo Touvron、Matthieu Cord、Matthijs Douze、Francisco Massa、Alexandre
Sablayrolles, Hervé Jégou.によって [Training data-efficient image Transformers & distillation through attention](https://arxiv.org/abs/2012.12877) で提案されました。
サブレイロール、エルヴェ・ジェグー。 [Dosovitskiy et al., 2020](https://arxiv.org/abs/2010.11929) で紹介された [Vision Transformer (ViT)](vit) は、既存の畳み込みニューラルと同等、またはそれを上回るパフォーマンスを発揮できることを示しました。
Transformer エンコーダ (BERT のような) を使用したネットワーク。ただし、その論文で紹介された ViT モデルには、次のトレーニングが必要でした。
外部データを使用して、数週間にわたる高価なインフラストラクチャ。 DeiT (データ効率の高い画像変換器) はさらに優れています
画像分類用に効率的にトレーニングされたトランスフォーマーにより、必要なデータとコンピューティング リソースがはるかに少なくなります。
オリジナルの ViT モデルとの比較。
論文の要約は次のとおりです。
*最近、純粋に注意に基づくニューラル ネットワークが、画像などの画像理解タスクに対処できることが示されました。
分類。ただし、これらのビジュアル トランスフォーマーは、
インフラストラクチャが高価であるため、その採用が制限されています。この作業では、コンボリューションフリーの競争力のあるゲームを作成します。
Imagenet のみでトレーニングしてトランスフォーマーを作成します。 1 台のコンピューターで 3 日以内にトレーニングを行います。私たちの基準となるビジョン
トランス (86M パラメータ) は、外部なしで ImageNet 上で 83.1% (単一クロップ評価) のトップ 1 の精度を達成します。
データ。さらに重要なのは、トランスフォーマーに特有の教師と生徒の戦略を導入することです。蒸留に依存している
学生が注意を払って教師から学ぶことを保証するトークン。私たちはこのトークンベースに興味を示します
特に convnet を教師として使用する場合。これにより、convnet と競合する結果を報告できるようになります。
Imagenet (最大 85.2% の精度が得られます) と他のタスクに転送するときの両方で。私たちはコードを共有し、
モデル。*
このモデルは、[nielsr](https://huggingface.co/nielsr) によって提供されました。このモデルの TensorFlow バージョンは、[amyeroberts](https://huggingface.co/amyeroberts) によって追加されました。
## Usage tips
- ViT と比較して、DeiT モデルはいわゆる蒸留トークンを使用して教師から効果的に学習します (これは、
DeiT 論文は、ResNet のようなモデルです)。蒸留トークンは、バックプロパゲーションを通じて、と対話することによって学習されます。
セルフアテンション層を介したクラス ([CLS]) とパッチ トークン。
- 抽出されたモデルを微調整するには 2 つの方法があります。(1) 上部に予測ヘッドを配置するだけの古典的な方法。
クラス トークンの最終的な非表示状態を抽出し、蒸留シグナルを使用しない、または (2) 両方の
予測ヘッドはクラス トークンの上と蒸留トークンの上にあります。その場合、[CLS] 予測は
head は、head の予測とグラウンド トゥルース ラベル間の通常のクロスエントロピーを使用してトレーニングされます。
蒸留予測ヘッドは、硬蒸留 (予測と予測の間のクロスエントロピー) を使用してトレーニングされます。
蒸留ヘッドと教師が予測したラベル）。推論時に、平均予測を取得します。
最終的な予測として両頭の間で。 (2) は「蒸留による微調整」とも呼ばれます。
下流のデータセットですでに微調整されている教師。モデル的には (1) に相当します。
[`DeiTForImageClassification`] と (2) に対応します。
[`DeiTForImageClassificationWithTeacher`]。
- 著者らは (2) についてもソフト蒸留を試みたことに注意してください (この場合、蒸留予測ヘッドは
教師のソフトマックス出力に一致するように KL ダイバージェンスを使用してトレーニングしました）が、ハード蒸留が最良の結果をもたらしました。
- リリースされたすべてのチェックポイントは、ImageNet-1k のみで事前トレーニングおよび微調整されました。外部データは使用されませんでした。これは
JFT-300M データセット/Imagenet-21k などの外部データを使用した元の ViT モデルとは対照的です。
事前トレーニング。
- DeiT の作者は、より効率的にトレーニングされた ViT モデルもリリースしました。これは、直接プラグインできます。
[`ViTModel`] または [`ViTForImageClassification`]。データなどのテクニック
はるかに大規模なデータセットでのトレーニングをシミュレートするために、拡張、最適化、正則化が使用されました。
(ただし、事前トレーニングには ImageNet-1k のみを使用します)。 4 つのバリエーション (3 つの異なるサイズ) が利用可能です。
*facebook/deit-tiny-patch16-224*、*facebook/deit-small-patch16-224*、*facebook/deit-base-patch16-224* および
*facebook/deit-base-patch16-384*。以下を行うには [`DeiTImageProcessor`] を使用する必要があることに注意してください。
モデル用の画像を準備します。
## Resources
DeiT を始めるのに役立つ公式 Hugging Face およびコミュニティ (🌎 で示されている) リソースのリスト。
<PipelineTag pipeline="image-classification"/>
- [`DeiTForImageClassification`] は、この [サンプル スクリプト](https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-classification) および [ノートブック](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/image_classification.ipynb)。
- 参照: [画像分類タスク ガイド](../tasks/image_classification)
それに加えて:
- [`DeiTForMaskedImageModeling`] は、この [サンプル スクリプト](https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-pretraining) でサポートされています。
ここに含めるリソースの送信に興味がある場合は、お気軽にプル リクエストを開いてください。審査させていただきます。リソースは、既存のリソースを複製するのではなく、何か新しいものを示すことが理想的です。
## DeiTConfig
[[autodoc]] DeiTConfig
## DeiTFeatureExtractor
[[autodoc]] DeiTFeatureExtractor
- __call__
## DeiTImageProcessor
[[autodoc]] DeiTImageProcessor
- preprocess
<frameworkcontent>
<pt>
## DeiTModel
[[autodoc]] DeiTModel
- forward
## DeiTForMaskedImageModeling
[[autodoc]] DeiTForMaskedImageModeling
- forward
## DeiTForImageClassification
[[autodoc]] DeiTForImageClassification
- forward
## DeiTForImageClassificationWithTeacher
[[autodoc]] DeiTForImageClassificationWithTeacher
- forward
</pt>
<tf>
## TFDeiTModel
[[autodoc]] TFDeiTModel
- call
## TFDeiTForMaskedImageModeling
[[autodoc]] TFDeiTForMaskedImageModeling
- call
## TFDeiTForImageClassification
[[autodoc]] TFDeiTForImageClassification
- call
## TFDeiTForImageClassificationWithTeacher
[[autodoc]] TFDeiTForImageClassificationWithTeacher
- call
</tf>
</frameworkcontent>
|
transformers/docs/source/ja/model_doc/deit.md/0
|
{
"file_path": "transformers/docs/source/ja/model_doc/deit.md",
"repo_id": "transformers",
"token_count": 3682
}
| 299
|
<!--
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.
⚠️ このファイルはMarkdown形式ですが、Hugging Faceのドキュメントビルダー向けに特定の構文を含んでいるため、
通常のMarkdownビューアーで正しく表示されないことに注意してください。
-->
# Quick tour
[[open-in-colab]]
🤗 Transformersを使い始めましょう！ 開発者であろうと、日常的なユーザーであろうと、このクイックツアーは
初めて始めるのを支援し、[`pipeline`]を使った推論方法、[AutoClass](./model_doc/auto)で事前学習済みモデルとプリプロセッサをロードする方法、
そしてPyTorchまたはTensorFlowで素早くモデルをトレーニングする方法を示します。 初心者の場合、ここで紹介されたコンセプトの詳細な説明を提供する
チュートリアルまたは[コース](https://huggingface.co/course/chapter1/1)を次に参照することをお勧めします。
始める前に、必要なライブラリがすべてインストールされていることを確認してください：
```bash
!pip install transformers datasets evaluate accelerate
```
あなたはまた、好きな機械学習フレームワークをインストールする必要があります:
<frameworkcontent>
<pt>
```bash
pip install torch
```
</pt>
<tf>
```bash
pip install tensorflow
```
</tf>
</frameworkcontent>
## Pipeline
<Youtube id="tiZFewofSLM"/>
[`pipeline`] は、事前学習済みモデルを推論に最も簡単で高速な方法です。
[`pipeline`] を使用することで、さまざまなモダリティにわたる多くのタスクに対して即座に使用できます。
いくつかのタスクは以下の表に示されています：
<Tip>
使用可能なタスクの完全な一覧については、[pipeline API リファレンス](./main_classes/pipelines)を確認してください。
</Tip>
| **タスク** | **説明** | **モダリティ** | **パイプライン識別子** |
|------------------------------|--------------------------------------------------------------------------------------------------------------|-----------------|-----------------------------------------------|
| テキスト分類 | テキストのシーケンスにラベルを割り当てる | NLP | pipeline(task="sentiment-analysis") |
| テキスト生成 | プロンプトを指定してテキストを生成する | NLP | pipeline(task="text-generation") |
| 要約 | テキストまたはドキュメントの要約を生成する | NLP | pipeline(task="summarization") |
| 画像分類 | 画像にラベルを割り当てる | コンピュータビジョン | pipeline(task="image-classification") |
| 画像セグメンテーション | 画像の各個別のピクセルにラベルを割り当てる（セマンティック、パノプティック、およびインスタンスセグメンテーションをサポート） | コンピュータビジョン | pipeline(task="image-segmentation") |
| オブジェクト検出 | 画像内のオブジェクトの境界ボックスとクラスを予測する | コンピュータビジョン | pipeline(task="object-detection") |
| オーディオ分類 | オーディオデータにラベルを割り当てる | オーディオ | pipeline(task="audio-classification") |
| 自動音声認識 | 音声をテキストに変換する | オーディオ | pipeline(task="automatic-speech-recognition") |
| ビジュアルクエスチョン応答 | 画像と質問が与えられた場合に、画像に関する質問に回答する | マルチモーダル | pipeline(task="vqa") |
| ドキュメントクエスチョン応答 | ドキュメントと質問が与えられた場合に、ドキュメントに関する質問に回答する | マルチモーダル | pipeline(task="document-question-answering") |
| 画像キャプショニング | 与えられた画像にキャプションを生成する | マルチモーダル | pipeline(task="image-to-text") |
まず、[`pipeline`] のインスタンスを作成し、使用したいタスクを指定します。
このガイドでは、センチメント分析のために [`pipeline`] を使用する例を示します：
```python
>>> from transformers import pipeline
>>> classifier = pipeline("sentiment-analysis")
```
[`pipeline`]は、感情分析のためのデフォルトの[事前学習済みモデル](https://huggingface.co/distilbert/distilbert-base-uncased-finetuned-sst-2-english)とトークナイザをダウンロードしてキャッシュし、使用できるようになります。
これで、`classifier`を対象のテキストに使用できます：
```python
>>> classifier("私たちは🤗 Transformersライブラリをお見せできてとても嬉しいです。")
[{'label': 'POSITIVE', 'score': 0.9998}]
```
複数の入力がある場合は、[`pipeline`]に入力をリストとして渡して、辞書のリストを返します：
```py
>>> results = classifier(["🤗 Transformersライブラリをご紹介できて非常に嬉しいです。", "嫌いにならないでほしいです。"])
>>> for result in results:
... print(f"label: {result['label']}, スコア: {round(result['score'], 4)}")
label: POSITIVE, スコア: 0.9998
label: NEGATIVE, スコア: 0.5309
```
[`pipeline`]は、任意のタスクに対してデータセット全体を繰り返し処理することもできます。この例では、自動音声認識をタスクとして選びましょう：
```python
>>> import torch
>>> from transformers import pipeline
>>> speech_recognizer = pipeline("automatic-speech-recognition", model="facebook/wav2vec2-base-960h")
```
オーディオデータセットをロードします（詳細については🤗 Datasets [クイックスタート](https://huggingface.co/docs/datasets/quickstart#audio)を参照してください）。
たとえば、[MInDS-14](https://huggingface.co/datasets/PolyAI/minds14)データセットをロードします：
```python
>>> from datasets import load_dataset, Audio
>>> dataset = load_dataset("PolyAI/minds14", name="en-US", split="train") # doctest: +IGNORE_RESULT
```
データセットのサンプリングレートが[`facebook/wav2vec2-base-960h`](https://huggingface.co/facebook/wav2vec2-base-960h)がトレーニングされたサンプリングレートと一致することを確認してください：
```py
>>> dataset = dataset.cast_column("audio", Audio(sampling_rate=speech_recognizer.feature_extractor.sampling_rate))
```
"audio"列を呼び出すと、オーディオファイルは自動的にロードされ、リサンプリングされます。最初の4つのサンプルから生の波形配列を抽出し、それをパイプラインにリストとして渡します。
```py
>>> result = speech_recognizer(dataset[:4]["audio"])
>>> print([d["text"] for d in result])
['I WOULD LIKE TO SET UP A JOINT ACCOUNT WITH MY PARTNER HOW DO I PROCEED WITH DOING THAT', "FONDERING HOW I'D SET UP A JOIN TO HELL T WITH MY WIFE AND WHERE THE AP MIGHT BE", "I I'D LIKE TOY SET UP A JOINT ACCOUNT WITH MY PARTNER I'M NOT SEEING THE OPTION TO DO IT ON THE APSO I CALLED IN TO GET SOME HELP CAN I JUST DO IT OVER THE PHONE WITH YOU AND GIVE YOU THE INFORMATION OR SHOULD I DO IT IN THE AP AN I'M MISSING SOMETHING UQUETTE HAD PREFERRED TO JUST DO IT OVER THE PHONE OF POSSIBLE THINGS", 'HOW DO I FURN A JOINA COUT']
```
大規模なデータセットで、入力が大きい場合（音声や画像など）、すべての入力をメモリに読み込む代わりに、リストではなくジェネレータを渡すことがお勧めです。詳細については[パイプラインAPIリファレンス](./main_classes/pipelines)を参照してください。
### Use another model and tokenizer in the pipeline
[`pipeline`]は[Hub](https://huggingface.co/models)からの任意のモデルを収容でき、他のユースケースに[`pipeline`]を適応させることが容易です。たとえば、フランス語のテキストを処理できるモデルが必要な場合、Hubのタグを使用して適切なモデルをフィルタリングできます。トップのフィルタリングされた結果は、フランス語のテキストに使用できる感情分析用に調整された多言語の[BERTモデル](https://huggingface.co/nlptown/bert-base-multilingual-uncased-sentiment)を返します：
```py
>>> model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
```
<frameworkcontent>
<pt>
[`AutoModelForSequenceClassification`]と[`AutoTokenizer`]を使用して事前学習済みモデルとそれに関連するトークナイザをロードします（次のセクションで`AutoClass`について詳しく説明します）：
```python
>>> from transformers import AutoTokenizer, AutoModelForSequenceClassification
>>> model = AutoModelForSequenceClassification.from_pretrained(model_name)
>>> tokenizer = AutoTokenizer.from_pretrained(model_name)
```
</pt>
<tf>
以下のコードは、[`TFAutoModelForSequenceClassification`]および[`AutoTokenizer`]を使用して、事前学習済みモデルとその関連するトークナイザをロードする方法を示しています（`TFAutoClass`については次のセクションで詳しく説明します）：
```python
>>> from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
>>> model = TFAutoModelForSequenceClassification.from_pretrained(model_name)
>>> tokenizer = AutoTokenizer.from_pretrained(model_name)
```
</tf>
</frameworkcontent>
指定したモデルとトークナイザを[`pipeline`]に設定し、今度はフランス語のテキストに`classifier`を適用できます：
```py
>>> classifier = pipeline("sentiment-analysis", model=model, tokenizer=tokenizer)
>>> classifier("Nous sommes très heureux de vous présenter la bibliothèque 🤗 Transformers.")
[{'label': '5 stars', 'score': 0.7273}]
```
もし、あなたのユースケースに適したモデルが見つからない場合、事前学習済みモデルをあなたのデータでファインチューニングする必要があります。
ファインチューニングの方法については、[ファインチューニングのチュートリアル](./training)をご覧ください。
最後に、ファインチューニングした事前学習済みモデルを共有し、コミュニティと共有ハブで共有することを検討してください。これにより、機械学習を民主化する手助けができます！ 🤗
## AutoClass
<Youtube id="AhChOFRegn4"/>
[`AutoModelForSequenceClassification`] および [`AutoTokenizer`] クラスは、上記で使用した [`pipeline`] を駆動するために協力して動作します。
[AutoClass](./model_doc/auto) は、事前学習済みモデルのアーキテクチャをその名前またはパスから自動的に取得するショートカットです。
適切な `AutoClass` を選択し、それに関連する前処理クラスを選択するだけで済みます。
前のセクションからの例に戻り、`AutoClass` を使用して [`pipeline`] の結果を再現する方法を見てみましょう。
### AutoTokenizer
トークナイザはテキストをモデルの入力として使用できる数値の配列に前処理する役割を果たします。
トークナイゼーションプロセスには、単語をどのように分割するかや、単語をどのレベルで分割するかといった多くのルールがあります
（トークナイゼーションについての詳細は [トークナイザサマリー](./tokenizer_summary) をご覧ください）。
最も重要なことは、モデルが事前学習済みになったときと同じトークナイゼーションルールを使用するために、同じモデル名でトークナイザをインスタンス化する必要があることです。
[`AutoTokenizer`] を使用してトークナイザをロードします：
```python
>>> from transformers import AutoTokenizer
>>> model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
>>> tokenizer = AutoTokenizer.from_pretrained(model_name)
```
Pass your text to the tokenizer:
```python
>>> encoding = tokenizer("私たちは🤗 Transformersライブラリをお見せできてとても嬉しいです。")
>>> print(encoding)
{'input_ids': [101, 11312, 10320, 12495, 19308, 10114, 11391, 10855, 10103, 100, 58263, 13299, 119, 102],
'token_type_ids': [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}
```
トークナイザは、次の情報を含む辞書を返します：
- [input_ids](./glossary#input-ids): トークンの数値表現。
- [attention_mask](.glossary#attention-mask): どのトークンにアテンションを向けるかを示します。
トークナイザはまた、入力のリストを受け入れ、一様な長さのバッチを返すためにテキストをパディングおよび切り詰めることができます。
<frameworkcontent>
<pt>
```py
>>> pt_batch = tokenizer(
... ["🤗 Transformersライブラリをお見せできて非常に嬉しいです。", "嫌いではないことを願っています。"],
... padding=True,
... truncation=True,
... max_length=512,
... return_tensors="pt",
... )
```
</pt>
<tf>
```py
>>> tf_batch = tokenizer(
... ["We are very happy to show you the 🤗 Transformers library.", "We hope you don't hate it."],
... padding=True,
... truncation=True,
... max_length=512,
... return_tensors="tf",
... )
```
</tf>
</frameworkcontent>
<Tip>
[前処理](./preprocessing)チュートリアルをご覧いただき、トークナイゼーションの詳細や、[`AutoImageProcessor`]、[`AutoFeatureExtractor`]、[`AutoProcessor`]を使用して画像、オーディオ、およびマルチモーダル入力を前処理する方法について詳しく説明されているページもご覧ください。
</Tip>
### AutoModel
<frameworkcontent>
<pt>
🤗 Transformersは事前学習済みインスタンスを簡単に統一的にロードする方法を提供します。
これは、[`AutoTokenizer`]をロードするのと同じように[`AutoModel`]をロードできることを意味します。
タスクに適した[`AutoModel`]を選択する以外の違いはありません。
テキスト（またはシーケンス）分類の場合、[`AutoModelForSequenceClassification`]をロードする必要があります：
```py
>>> from transformers import AutoModelForSequenceClassification
>>> model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
>>> pt_model = AutoModelForSequenceClassification.from_pretrained(model_name)
```
<Tip>
[`AutoModel`]クラスでサポートされているタスクに関する詳細については、[タスクの概要](./task_summary)を参照してください。
</Tip>
今、前処理済みのバッチを直接モデルに渡します。辞書を展開するだけで、`**`を追加する必要があります：
```python
>>> pt_outputs = pt_model(**pt_batch)
```
モデルは、`logits`属性に最終的なアクティベーションを出力します。 `logits`にsoftmax関数を適用して確率を取得します：
```py
>>> from torch import nn
>>> pt_predictions = nn.functional.softmax(pt_outputs.logits, dim=-1)
>>> print(pt_predictions)
tensor([[0.0021, 0.0018, 0.0115, 0.2121, 0.7725],
[0.2084, 0.1826, 0.1969, 0.1755, 0.2365]], grad_fn=<SoftmaxBackward0>)
```
</pt>
<tf>
🤗 Transformersは事前学習済みインスタンスをロードするためのシンプルで統一された方法を提供します。
これは、[`TFAutoModel`]を[`AutoTokenizer`]をロードするのと同じようにロードできることを意味します。
唯一の違いは、タスクに適した[`TFAutoModel`]を選択することです。
テキスト（またはシーケンス）分類の場合、[`TFAutoModelForSequenceClassification`]をロードする必要があります：
```py
>>> from transformers import TFAutoModelForSequenceClassification
>>> model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
>>> tf_model = TFAutoModelForSequenceClassification.from_pretrained(model_name)
```
<Tip>
詳細については、[`AutoModel`]クラスでサポートされているタスクに関する情報は、[タスクの概要](./task_summary)を参照してください。
</Tip>
次に、前処理済みのバッチを直接モデルに渡します。テンソルをそのまま渡すことができます：
```python
>>> tf_outputs = tf_model(tf_batch)
```
モデルは`logits`属性に最終的なアクティベーションを出力します。`logits`にソフトマックス関数を適用して確率を取得します：
```python
>>> import tensorflow as tf
>>> tf_predictions = tf.nn.softmax(tf_outputs.logits, axis=-1)
>>> tf_predictions # doctest: +IGNORE_RESULT
```
</tf>
</frameworkcontent>
<Tip>
🤗 Transformersのすべてのモデル（PyTorchまたはTensorFlow）は、最終的な活性化関数（softmaxなど）*前*のテンソルを出力します。
最終的な活性化関数は、しばしば損失と結合されているためです。モデルの出力は特別なデータクラスであり、その属性はIDEで自動補完されます。
モデルの出力は、タプルまたは辞書のように動作します（整数、スライス、または文字列でインデックスを付けることができます）。
この場合、Noneである属性は無視されます。
</Tip>
### Save a Model
<frameworkcontent>
<pt>
モデルをファインチューニングしたら、[`PreTrainedModel.save_pretrained`]を使用してトークナイザと共に保存できます：
```py
>>> pt_save_directory = "./pt_save_pretrained"
>>> tokenizer.save_pretrained(pt_save_directory) # doctest: +IGNORE_RESULT
>>> pt_model.save_pretrained(pt_save_directory)
```
再びモデルを使用する準備ができたら、[`PreTrainedModel.from_pretrained`]を使用して再度ロードします：
```py
>>> pt_model = AutoModelForSequenceClassification.from_pretrained("./pt_save_pretrained")
```
</pt>
<tf>
モデルをファインチューニングしたら、そのトークナイザを使用してモデルを保存できます。[`TFPreTrainedModel.save_pretrained`]を使用します：
```py
>>> tf_save_directory = "./tf_save_pretrained"
>>> tokenizer.save_pretrained(tf_save_directory) # doctest: +IGNORE_RESULT
>>> tf_model.save_pretrained(tf_save_directory)
```
モデルを再度使用する準備ができたら、[`TFPreTrainedModel.from_pretrained`]を使用して再度ロードします：
```py
>>> tf_model = TFAutoModelForSequenceClassification.from_pretrained("./tf_save_pretrained")
```
</tf>
</frameworkcontent>
🤗 Transformersの特に素晴らしい機能の一つは、モデルを保存し、それをPyTorchモデルまたはTensorFlowモデルとして再ロードできることです。 `from_pt`または`from_tf`パラメータを使用してモデルをフレームワーク間で変換できます：
<frameworkcontent>
<pt>
```py
>>> from transformers import AutoModel
>>> tokenizer = AutoTokenizer.from_pretrained(tf_save_directory)
>>> pt_model = AutoModelForSequenceClassification.from_pretrained(tf_save_directory, from_tf=True)
```
</pt>
<tf>
```py
>>> from transformers import TFAutoModel
>>> tokenizer = AutoTokenizer.from_pretrained(pt_save_directory)
>>> tf_model = TFAutoModelForSequenceClassification.from_pretrained(pt_save_directory, from_pt=True)
```
</tf>
</frameworkcontent>
## Custom model builds
モデルを構築方法を変更するには、モデルの設定クラスを変更できます。設定はモデルの属性を指定します。例えば、隠れ層の数やアテンションヘッドの数などがこれに含まれます。カスタム設定クラスからモデルを初期化する際には、ゼロから始めます。モデルの属性はランダムに初期化され、有意義な結果を得るためにモデルをトレーニングする必要があります。
最初に[`AutoConfig`]をインポートし、変更したい事前学習済みモデルをロードします。[`AutoConfig.from_pretrained`]内で、変更したい属性（例：アテンションヘッドの数）を指定できます：
```python
>>> from transformers import AutoConfig
>>> my_config = AutoConfig.from_pretrained("distilbert/distilbert-base-uncased", n_heads=12)
```
<frameworkcontent>
<pt>
[`AutoModel.from_config`]を使用してカスタム設定からモデルを作成します：
```python
>>> from transformers import AutoModel
>>> my_model = AutoModel.from_config(my_config)
```
</pt>
<tf>
カスタム構成からモデルを作成するには、[`TFAutoModel.from_config`]を使用します：
```py
>>> from transformers import TFAutoModel
>>> my_model = TFAutoModel.from_config(my_config)
```
</tf>
</frameworkcontent>
[カスタムアーキテクチャを作成](./create_a_model)ガイドを参照して、カスタム構成の詳細情報を確認してください。
## Trainer - PyTorch向けの最適化されたトレーニングループ
すべてのモデルは標準の[`torch.nn.Module`](https://pytorch.org/docs/stable/nn.html#torch.nn.Module)であるため、通常のトレーニングループで使用できます。
独自のトレーニングループを作成できますが、🤗 TransformersはPyTorch向けに[`Trainer`]クラスを提供しており、基本的なトレーニングループに加えて、
分散トレーニング、混合精度などの機能の追加を行っています。
タスクに応じて、通常は[`Trainer`]に以下のパラメータを渡します：
1. [`PreTrainedModel`]または[`torch.nn.Module`](https://pytorch.org/docs/stable/nn.html#torch.nn.Module)から始めます：
```py
>>> from transformers import AutoModelForSequenceClassification
>>> model = AutoModelForSequenceClassification.from_pretrained("distilbert/distilbert-base-uncased")
```
2. [`TrainingArguments`]には、変更できるモデルのハイパーパラメータが含まれており、学習率、バッチサイズ、トレーニングエポック数などが変更できます。指定しない場合、デフォルト値が使用されます：
```py
>>> from transformers import TrainingArguments
>>> training_args = TrainingArguments(
... output_dir="path/to/save/folder/",
... learning_rate=2e-5,
... per_device_train_batch_size=8,
... per_device_eval_batch_size=8,
... num_train_epochs=2,
... )
```
3. トークナイザ、画像プロセッサ、特徴量抽出器、またはプロセッサのような前処理クラスをロードします：
```py
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("distilbert/distilbert-base-uncased")
```
4. データセットをロードする:
```py
>>> from datasets import load_dataset
>>> dataset = load_dataset("rotten_tomatoes") # doctest: +IGNORE_RESULT
```
5. データセットをトークン化するための関数を作成します：
```python
>>> def tokenize_dataset(dataset):
... return tokenizer(dataset["text"])
```
その後、[`~datasets.Dataset.map`]を使用してデータセット全体に適用します：
```python
>>> dataset = dataset.map(tokenize_dataset, batched=True)
```
6. データセットからの例のバッチを作成するための [`DataCollatorWithPadding`]：
```py
>>> from transformers import DataCollatorWithPadding
>>> data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
```
次に、これらのクラスを[`Trainer`]にまとめます：
```python
>>> 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,
... ) # doctest: +SKIP
```
訓練を開始する準備ができたら、[`~Trainer.train`]を呼び出してトレーニングを開始します：
```py
>>> trainer.train() # doctest: +SKIP
```
<Tip>
翻訳や要約など、シーケンス間モデルを使用するタスクには、代わりに[`Seq2SeqTrainer`]と[`Seq2SeqTrainingArguments`]クラスを使用してください。
</Tip>
[`Trainer`]内のメソッドをサブクラス化することで、トレーニングループの動作をカスタマイズできます。これにより、損失関数、オプティマイザ、スケジューラなどの機能をカスタマイズできます。サブクラス化できるメソッドの一覧については、[`Trainer`]リファレンスをご覧ください。
トレーニングループをカスタマイズする別の方法は、[Callbacks](./main_classes/callback)を使用することです。コールバックを使用して他のライブラリと統合し、トレーニングループを監視して進捗状況を報告したり、トレーニングを早期に停止したりできます。コールバックはトレーニングループ自体には何も変更を加えません。損失関数などのカスタマイズを行う場合は、[`Trainer`]をサブクラス化する必要があります。
## Train with TensorFlow
すべてのモデルは標準の[`tf.keras.Model`](https://www.tensorflow.org/api_docs/python/tf/keras/Model)であるため、[Keras](https://keras.io/) APIを使用してTensorFlowでトレーニングできます。
🤗 Transformersは、データセットを`tf.data.Dataset`として簡単にロードできるようにする[`~TFPreTrainedModel.prepare_tf_dataset`]メソッドを提供しており、Kerasの[`compile`](https://keras.io/api/models/model_training_apis/#compile-method)および[`fit`](https://keras.io/api/models/model_training_apis/#fit-method)メソッドを使用してトレーニングをすぐに開始できます。
1. [`TFPreTrainedModel`]または[`tf.keras.Model`](https://www.tensorflow.org/api_docs/python/tf/keras/Model)から始めます：
```py
>>> from transformers import TFAutoModelForSequenceClassification
>>> model = TFAutoModelForSequenceClassification.from_pretrained("distilbert/distilbert-base-uncased")
```
2. トークナイザ、画像プロセッサ、特徴量抽出器、またはプロセッサのような前処理クラスをロードします：
```py
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("distilbert/distilbert-base-uncased")
```
3. データセットをトークナイズするための関数を作成します：
```python
>>> def tokenize_dataset(dataset):
... return tokenizer(dataset["text"]) # doctest: +SKIP
```
4. [`~datasets.Dataset.map`]を使用してデータセット全体にトークナイザを適用し、データセットとトークナイザを[`~TFPreTrainedModel.prepare_tf_dataset`]に渡します。バッチサイズを変更し、データセットをシャッフルすることもできます。
```python
>>> dataset = dataset.map(tokenize_dataset) # doctest: +SKIP
>>> tf_dataset = model.prepare_tf_dataset(
... dataset["train"], batch_size=16, shuffle=True, tokenizer=tokenizer
... ) # doctest: +SKIP
```
5. 準備ができたら、`compile`と`fit`を呼び出してトレーニングを開始できます。 Transformersモデルはすべてデフォルトのタスクに関連する損失関数を持っているため、指定しない限り、損失関数を指定する必要はありません。
```python
>>> from tensorflow.keras.optimizers import Adam
>>> model.compile(optimizer=Adam(3e-5)) # 損失関数の引数は不要です！
>>> model.fit(tf
```
## What's next?
🤗 Transformersのクイックツアーを完了したら、ガイドをチェックして、カスタムモデルの作成、タスクのためのファインチューニング、スクリプトを使用したモデルのトレーニングなど、より具体的なことを学ぶことができます。🤗 Transformersのコアコンセプトについてもっと詳しく知りたい場合は、コンセプチュアルガイドを読んでみてください！
|
transformers/docs/source/ja/quicktour.md/0
|
{
"file_path": "transformers/docs/source/ja/quicktour.md",
"repo_id": "transformers",
"token_count": 13066
}
| 300
|
<!--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.
-->
# Object detection
[[open-in-colab]]
オブジェクト検出は、画像内のインスタンス (人間、建物、車など) を検出するコンピューター ビジョン タスクです。物体検出モデルは画像を入力および出力として受け取ります
検出されたオブジェクトの境界ボックスと関連するラベルの座標。画像には複数のオブジェクトを含めることができます。
それぞれに独自の境界ボックスとラベルがあり (例: 車と建物を持つことができます)、各オブジェクトは
画像のさまざまな部分に存在する必要があります (たとえば、画像には複数の車が含まれている可能性があります)。
このタスクは、歩行者、道路標識、信号機などを検出するために自動運転で一般的に使用されます。
他のアプリケーションには、画像内のオブジェクトのカウント、画像検索などが含まれます。
このガイドでは、次の方法を学習します。
1. Finetune [DETR](https://huggingface.co/docs/transformers/model_doc/detr)、畳み込みアルゴリズムを組み合わせたモデル
[CPPE-5](https://huggingface.co/datasets/cppe-5) 上のエンコーダー/デコーダー トランスフォーマーを備えたバックボーン
データセット。
2. 微調整したモデルを推論に使用します。
<Tip>
このタスクと互換性のあるすべてのアーキテクチャとチェックポイントを確認するには、[タスクページ](https://huggingface.co/tasks/object-detection) を確認することをお勧めします。
</Tip>
始める前に、必要なライブラリがすべてインストールされていることを確認してください。
```bash
pip install -q datasets transformers evaluate timm albumentations
```
🤗 データセットを使用して Hugging Face Hub からデータセットをロードし、🤗 トランスフォーマーを使用してモデルをトレーニングします。
データを増強するための`albumentations`。 `timm` は現在、DETR モデルの畳み込みバックボーンをロードするために必要です。
モデルをコミュニティと共有することをお勧めします。 Hugging Face アカウントにログインして、ハブにアップロードします。
プロンプトが表示されたら、トークンを入力してログインします。
```py
>>> from huggingface_hub import notebook_login
>>> notebook_login()
```
## Load the CPPE-5 dataset
[CPPE-5 データセット](https://huggingface.co/datasets/cppe-5) には、次の画像が含まれています。
新型コロナウイルス感染症のパンデミックにおける医療用個人保護具 (PPE) を識別する注釈。
データセットをロードすることから始めます。
```py
>>> from datasets import load_dataset
>>> cppe5 = load_dataset("cppe-5")
>>> cppe5
DatasetDict({
train: Dataset({
features: ['image_id', 'image', 'width', 'height', 'objects'],
num_rows: 1000
})
test: Dataset({
features: ['image_id', 'image', 'width', 'height', 'objects'],
num_rows: 29
})
})
```
このデータセットには、1000 枚の画像を含むトレーニング セットと 29 枚の画像を含むテスト セットがすでに付属していることがわかります。
データに慣れるために、例がどのようなものかを調べてください。
```py
>>> cppe5["train"][0]
{'image_id': 15,
'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=943x663 at 0x7F9EC9E77C10>,
'width': 943,
'height': 663,
'objects': {'id': [114, 115, 116, 117],
'area': [3796, 1596, 152768, 81002],
'bbox': [[302.0, 109.0, 73.0, 52.0],
[810.0, 100.0, 57.0, 28.0],
[160.0, 31.0, 248.0, 616.0],
[741.0, 68.0, 202.0, 401.0]],
'category': [4, 4, 0, 0]}}
```
データセット内の例には次のフィールドがあります。
- `image_id`: サンプルの画像ID
- `image`: 画像を含む `PIL.Image.Image` オブジェクト
- `width`: 画像の幅
- `height`: 画像の高さ
- `objects`: 画像内のオブジェクトの境界ボックスのメタデータを含む辞書:
- `id`: アノテーションID
- `area`: 境界ボックスの領域
- `bbox`: オブジェクトの境界ボックス ([COCO 形式](https://albumentations.ai/docs/getting_started/bounding_boxes_augmentation/#coco) )
- `category`: オブジェクトのカテゴリー。可能な値には、`Coverall (0)`、`Face_Shield (1)`、`Gloves (2)`、`Goggles (3)`、および `Mask (4)` が含まれます。
`bbox`フィールドが COCO 形式に従っていることに気づくかもしれません。これは DETR モデルが予期する形式です。
ただし、「オブジェクト」内のフィールドのグループ化は、DETR が必要とする注釈形式とは異なります。あなたはするであろう
このデータをトレーニングに使用する前に、いくつかの前処理変換を適用する必要があります。
データをさらに深く理解するには、データセット内の例を視覚化します。
```py
>>> import numpy as np
>>> import os
>>> from PIL import Image, ImageDraw
>>> image = cppe5["train"][0]["image"]
>>> annotations = cppe5["train"][0]["objects"]
>>> draw = ImageDraw.Draw(image)
>>> categories = cppe5["train"].features["objects"].feature["category"].names
>>> id2label = {index: x for index, x in enumerate(categories, start=0)}
>>> label2id = {v: k for k, v in id2label.items()}
>>> for i in range(len(annotations["id"])):
... box = annotations["bbox"][i]
... class_idx = annotations["category"][i]
... x, y, w, h = tuple(box)
... draw.rectangle((x, y, x + w, y + h), outline="red", width=1)
... draw.text((x, y), id2label[class_idx], fill="white")
>>> image
```
<div class="flex justify-center">
<img src="https://i.imgur.com/TdaqPJO.png" alt="CPPE-5 Image Example"/>
</div>
関連付けられたラベルを使用して境界ボックスを視覚化するには、データセットのメタデータからラベルを取得します。
`category`フィールド。
また、ラベル ID をラベル クラスにマッピングする辞書 (`id2label`) やその逆 (`label2id`) を作成することもできます。
これらは、後でモデルをセットアップするときに使用できます。これらのマップを含めると、共有した場合に他の人がモデルを再利用できるようになります。
ハグフェイスハブに取り付けます。
データに慣れるための最後のステップとして、潜在的な問題がないかデータを調査します。データセットに関する一般的な問題の 1 つは、
オブジェクト検出は、画像の端を越えて「伸びる」境界ボックスです。このような「暴走」境界ボックスは、
トレーニング中にエラーが発生するため、この段階で対処する必要があります。このデータセットには、この問題に関する例がいくつかあります。
このガイドでは内容をわかりやすくするために、これらの画像をデータから削除します。
```py
>>> remove_idx = [590, 821, 822, 875, 876, 878, 879]
>>> keep = [i for i in range(len(cppe5["train"])) if i not in remove_idx]
>>> cppe5["train"] = cppe5["train"].select(keep)
```
## Preprocess the data
モデルを微調整するには、事前トレーニングされたモデルに使用されるアプローチと正確に一致するように、使用する予定のデータを前処理する必要があります。
[`AutoImageProcessor`] は、画像データを処理して `pixel_values`、`pixel_mask`、および
DETR モデルをトレーニングできる「ラベル」。画像プロセッサには、心配する必要のないいくつかの属性があります。
- `image_mean = [0.485, 0.456, 0.406 ]`
- `image_std = [0.229, 0.224, 0.225]`
これらは、モデルの事前トレーニング中に画像を正規化するために使用される平均と標準偏差です。これらの価値観は非常に重要です
事前にトレーニングされた画像モデルを推論または微調整するときに複製します。
微調整するモデルと同じチェックポイントからイメージ プロセッサをインスタンス化します。
```py
>>> from transformers import AutoImageProcessor
>>> checkpoint = "facebook/detr-resnet-50"
>>> image_processor = AutoImageProcessor.from_pretrained(checkpoint)
```
画像を`image_processor`に渡す前に、2 つの前処理変換をデータセットに適用します。
- 画像の拡張
- DETR の期待に応えるための注釈の再フォーマット
まず、モデルがトレーニング データにオーバーフィットしないようにするために、任意のデータ拡張ライブラリを使用して画像拡張を適用できます。ここでは[Albumentations](https://albumentations.ai/docs/)を使用します...
このライブラリは、変換が画像に影響を与え、それに応じて境界ボックスを更新することを保証します。
🤗 データセット ライブラリのドキュメントには、詳細な [物体検出用に画像を拡張する方法に関するガイド](https://huggingface.co/docs/datasets/object_detection) が記載されています。
例としてまったく同じデータセットを使用しています。ここでも同じアプローチを適用し、各画像のサイズを (480, 480) に変更します。
水平に反転して明るくします。
```py
>>> import albumentations
>>> import numpy as np
>>> import torch
>>> transform = albumentations.Compose(
... [
... albumentations.Resize(480, 480),
... albumentations.HorizontalFlip(p=1.0),
... albumentations.RandomBrightnessContrast(p=1.0),
... ],
... bbox_params=albumentations.BboxParams(format="coco", label_fields=["category"]),
... )
```
`image_processor` は、注釈が次の形式であることを期待します: `{'image_id': int, 'annotations': List[Dict]}`,
ここで、各辞書は COCO オブジェクトの注釈です。 1 つの例として、注釈を再フォーマットする関数を追加してみましょう。
```py
>>> def formatted_anns(image_id, category, area, bbox):
... annotations = []
... for i in range(0, len(category)):
... new_ann = {
... "image_id": image_id,
... "category_id": category[i],
... "isCrowd": 0,
... "area": area[i],
... "bbox": list(bbox[i]),
... }
... annotations.append(new_ann)
... return annotations
```
これで、画像と注釈の変換を組み合わせてサンプルのバッチで使用できるようになりました。
```py
>>> # transforming a batch
>>> def transform_aug_ann(examples):
... image_ids = examples["image_id"]
... images, bboxes, area, categories = [], [], [], []
... for image, objects in zip(examples["image"], examples["objects"]):
... image = np.array(image.convert("RGB"))[:, :, ::-1]
... out = transform(image=image, bboxes=objects["bbox"], category=objects["category"])
... area.append(objects["area"])
... images.append(out["image"])
... bboxes.append(out["bboxes"])
... categories.append(out["category"])
... targets = [
... {"image_id": id_, "annotations": formatted_anns(id_, cat_, ar_, box_)}
... for id_, cat_, ar_, box_ in zip(image_ids, categories, area, bboxes)
... ]
... return image_processor(images=images, annotations=targets, return_tensors="pt")
```
🤗 Datasets [`~datasets.Dataset.with_transform`] メソッドを使用して、この前処理関数をデータセット全体に適用します。この方法が適用されるのは、
データセットの要素を読み込むときに、その場で変換します。
この時点で、データセットの例が変換後にどのようになるかを確認できます。テンソルが表示されるはずです
`pixel_values`、テンソルと `pixel_mask`、および `labels` を使用します。
```py
>>> cppe5["train"] = cppe5["train"].with_transform(transform_aug_ann)
>>> cppe5["train"][15]
{'pixel_values': tensor([[[ 0.9132, 0.9132, 0.9132, ..., -1.9809, -1.9809, -1.9809],
[ 0.9132, 0.9132, 0.9132, ..., -1.9809, -1.9809, -1.9809],
[ 0.9132, 0.9132, 0.9132, ..., -1.9638, -1.9638, -1.9638],
...,
[-1.5699, -1.5699, -1.5699, ..., -1.9980, -1.9980, -1.9980],
[-1.5528, -1.5528, -1.5528, ..., -1.9980, -1.9809, -1.9809],
[-1.5528, -1.5528, -1.5528, ..., -1.9980, -1.9809, -1.9809]],
[[ 1.3081, 1.3081, 1.3081, ..., -1.8431, -1.8431, -1.8431],
[ 1.3081, 1.3081, 1.3081, ..., -1.8431, -1.8431, -1.8431],
[ 1.3081, 1.3081, 1.3081, ..., -1.8256, -1.8256, -1.8256],
...,
[-1.3179, -1.3179, -1.3179, ..., -1.8606, -1.8606, -1.8606],
[-1.3004, -1.3004, -1.3004, ..., -1.8606, -1.8431, -1.8431],
[-1.3004, -1.3004, -1.3004, ..., -1.8606, -1.8431, -1.8431]],
[[ 1.4200, 1.4200, 1.4200, ..., -1.6476, -1.6476, -1.6476],
[ 1.4200, 1.4200, 1.4200, ..., -1.6476, -1.6476, -1.6476],
[ 1.4200, 1.4200, 1.4200, ..., -1.6302, -1.6302, -1.6302],
...,
[-1.0201, -1.0201, -1.0201, ..., -1.5604, -1.5604, -1.5604],
[-1.0027, -1.0027, -1.0027, ..., -1.5604, -1.5430, -1.5430],
[-1.0027, -1.0027, -1.0027, ..., -1.5604, -1.5430, -1.5430]]]),
'pixel_mask': tensor([[1, 1, 1, ..., 1, 1, 1],
[1, 1, 1, ..., 1, 1, 1],
[1, 1, 1, ..., 1, 1, 1],
...,
[1, 1, 1, ..., 1, 1, 1],
[1, 1, 1, ..., 1, 1, 1],
[1, 1, 1, ..., 1, 1, 1]]),
'labels': {'size': tensor([800, 800]), 'image_id': tensor([756]), 'class_labels': tensor([4]), 'boxes': tensor([[0.7340, 0.6986, 0.3414, 0.5944]]), 'area': tensor([519544.4375]), 'iscrowd': tensor([0]), 'orig_size': tensor([480, 480])}}
```
個々の画像を正常に拡張し、それらの注釈を準備しました。ただし、前処理はそうではありません。
まだ完成しています。最後のステップでは、画像をバッチ処理するためのカスタム `collate_fn` を作成します。
画像 (現在は `pixel_values`) をバッチ内の最大の画像にパディングし、対応する `pixel_mask` を作成します
どのピクセルが実数 (1) で、どのピクセルがパディング (0) であるかを示します。
```py
>>> def collate_fn(batch):
... pixel_values = [item["pixel_values"] for item in batch]
... encoding = image_processor.pad(pixel_values, return_tensors="pt")
... labels = [item["labels"] for item in batch]
... batch = {}
... batch["pixel_values"] = encoding["pixel_values"]
... batch["pixel_mask"] = encoding["pixel_mask"]
... batch["labels"] = labels
... return batch
```
## Training the DETR model
前のセクションで重労働のほとんどを完了したので、モデルをトレーニングする準備が整いました。
このデータセット内の画像は、サイズを変更した後でも依然として非常に大きいです。これは、このモデルを微調整すると、
少なくとも 1 つの GPU が必要です。
トレーニングには次の手順が含まれます。
1. 前処理と同じチェックポイントを使用して、[`AutoModelForObjectDetection`] でモデルを読み込みます。
2. [`TrainingArguments`] でトレーニング ハイパーパラメータを定義します。
3. トレーニング引数をモデル、データセット、画像プロセッサ、データ照合器とともに [`Trainer`] に渡します。
4. [`~Trainer.train`] を呼び出してモデルを微調整します。
前処理に使用したのと同じチェックポイントからモデルをロードするときは、必ず`label2id`を渡してください。
および `id2label` マップは、以前にデータセットのメタデータから作成したものです。さらに、`ignore_mismatched_sizes=True`を指定して、既存の分類頭部を新しい分類頭部に置き換えます。
```py
>>> from transformers import AutoModelForObjectDetection
>>> model = AutoModelForObjectDetection.from_pretrained(
... checkpoint,
... id2label=id2label,
... label2id=label2id,
... ignore_mismatched_sizes=True,
... )
```
[`TrainingArguments`] で、`output_dir` を使用してモデルの保存場所を指定し、必要に応じてハイパーパラメーターを構成します。
画像列が削除されるため、未使用の列を削除しないことが重要です。画像列がないと、
`pixel_values` を作成できません。このため、`remove_unused_columns`を`False`に設定します。
ハブにプッシュしてモデルを共有したい場合は、`push_to_hub` を `True` に設定します (Hugging にサインインする必要があります)
顔に向かってモデルをアップロードします）。
```py
>>> from transformers import TrainingArguments
>>> training_args = TrainingArguments(
... output_dir="detr-resnet-50_finetuned_cppe5",
... per_device_train_batch_size=8,
... num_train_epochs=10,
... fp16=True,
... save_steps=200,
... logging_steps=50,
... learning_rate=1e-5,
... weight_decay=1e-4,
... save_total_limit=2,
... remove_unused_columns=False,
... push_to_hub=True,
... )
```
最後に、すべてをまとめて、[`~transformers.Trainer.train`] を呼び出します。
```py
>>> from transformers import Trainer
>>> trainer = Trainer(
... model=model,
... args=training_args,
... data_collator=collate_fn,
... train_dataset=cppe5["train"],
... tokenizer=image_processor,
... )
>>> trainer.train()
```
`training_args`で`push_to_hub`を`True`に設定した場合、トレーニング チェックポイントは
ハグフェイスハブ。トレーニングが完了したら、[`~transformers.Trainer.push_to_hub`] メソッドを呼び出して、最終モデルもハブにプッシュします。
```py
>>> trainer.push_to_hub()
```
## Evaluate
物体検出モデルは通常、一連の <a href="https://cocodataset.org/#detection-eval">COCO スタイルの指標</a>を使用して評価されます。
既存のメトリクス実装のいずれかを使用できますが、ここでは`torchvision`のメトリクス実装を使用して最終的なメトリクスを評価します。
ハブにプッシュしたモデル。
`torchvision`エバリュエーターを使用するには、グラウンド トゥルース COCO データセットを準備する必要があります。 COCO データセットを構築するための API
データを特定の形式で保存する必要があるため、最初に画像と注釈をディスクに保存する必要があります。と同じように
トレーニング用にデータを準備するとき、`cppe5["test"]` からの注釈をフォーマットする必要があります。ただし、画像
そのままでいるべきです。
評価ステップには少し作業が必要ですが、大きく 3 つのステップに分けることができます。
まず、`cppe5["test"]` セットを準備します。注釈をフォーマットし、データをディスクに保存します。
```py
>>> import json
>>> # format annotations the same as for training, no need for data augmentation
>>> def val_formatted_anns(image_id, objects):
... annotations = []
... for i in range(0, len(objects["id"])):
... new_ann = {
... "id": objects["id"][i],
... "category_id": objects["category"][i],
... "iscrowd": 0,
... "image_id": image_id,
... "area": objects["area"][i],
... "bbox": objects["bbox"][i],
... }
... annotations.append(new_ann)
... return annotations
>>> # Save images and annotations into the files torchvision.datasets.CocoDetection expects
>>> def save_cppe5_annotation_file_images(cppe5):
... output_json = {}
... path_output_cppe5 = f"{os.getcwd()}/cppe5/"
... if not os.path.exists(path_output_cppe5):
... os.makedirs(path_output_cppe5)
... path_anno = os.path.join(path_output_cppe5, "cppe5_ann.json")
... categories_json = [{"supercategory": "none", "id": id, "name": id2label[id]} for id in id2label]
... output_json["images"] = []
... output_json["annotations"] = []
... for example in cppe5:
... ann = val_formatted_anns(example["image_id"], example["objects"])
... output_json["images"].append(
... {
... "id": example["image_id"],
... "width": example["image"].width,
... "height": example["image"].height,
... "file_name": f"{example['image_id']}.png",
... }
... )
... output_json["annotations"].extend(ann)
... output_json["categories"] = categories_json
... with open(path_anno, "w") as file:
... json.dump(output_json, file, ensure_ascii=False, indent=4)
... for im, img_id in zip(cppe5["image"], cppe5["image_id"]):
... path_img = os.path.join(path_output_cppe5, f"{img_id}.png")
... im.save(path_img)
... return path_output_cppe5, path_anno
```
次に、`cocoevaluator`で利用できる`CocoDetection`クラスのインスタンスを用意します。
```py
>>> import torchvision
>>> class CocoDetection(torchvision.datasets.CocoDetection):
... def __init__(self, img_folder, image_processor, ann_file):
... super().__init__(img_folder, ann_file)
... self.image_processor = image_processor
... def __getitem__(self, idx):
... # read in PIL image and target in COCO format
... img, target = super(CocoDetection, self).__getitem__(idx)
... # preprocess image and target: converting target to DETR format,
... # resizing + normalization of both image and target)
... image_id = self.ids[idx]
... target = {"image_id": image_id, "annotations": target}
... encoding = self.image_processor(images=img, annotations=target, return_tensors="pt")
... pixel_values = encoding["pixel_values"].squeeze() # remove batch dimension
... target = encoding["labels"][0] # remove batch dimension
... return {"pixel_values": pixel_values, "labels": target}
>>> im_processor = AutoImageProcessor.from_pretrained("devonho/detr-resnet-50_finetuned_cppe5")
>>> path_output_cppe5, path_anno = save_cppe5_annotation_file_images(cppe5["test"])
>>> test_ds_coco_format = CocoDetection(path_output_cppe5, im_processor, path_anno)
```
最後に、メトリクスをロードして評価を実行します。
```py
>>> import evaluate
>>> from tqdm import tqdm
>>> model = AutoModelForObjectDetection.from_pretrained("devonho/detr-resnet-50_finetuned_cppe5")
>>> module = evaluate.load("ybelkada/cocoevaluate", coco=test_ds_coco_format.coco)
>>> val_dataloader = torch.utils.data.DataLoader(
... test_ds_coco_format, batch_size=8, shuffle=False, num_workers=4, collate_fn=collate_fn
... )
>>> with torch.no_grad():
... for idx, batch in enumerate(tqdm(val_dataloader)):
... pixel_values = batch["pixel_values"]
... pixel_mask = batch["pixel_mask"]
... labels = [
... {k: v for k, v in t.items()} for t in batch["labels"]
... ] # these are in DETR format, resized + normalized
... # forward pass
... outputs = model(pixel_values=pixel_values, pixel_mask=pixel_mask)
... orig_target_sizes = torch.stack([target["orig_size"] for target in labels], dim=0)
... results = im_processor.post_process(outputs, orig_target_sizes) # convert outputs of model to Pascal VOC format (xmin, ymin, xmax, ymax)
... module.add(prediction=results, reference=labels)
... del batch
>>> results = module.compute()
>>> print(results)
Accumulating evaluation results...
DONE (t=0.08s).
IoU metric: bbox
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.352
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.681
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.292
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.168
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.208
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.429
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.274
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.484
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.501
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.191
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.323
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.590
```
これらの結果は、[`~transformers.TrainingArguments`] のハイパーパラメータを調整することでさらに改善できます。試してごらん！
## Inference
DETR モデルを微調整して評価し、Hugging Face Hub にアップロードしたので、それを推論に使用できます。
推論用に微調整されたモデルを試す最も簡単な方法は、それを [`pipeline`] で使用することです。パイプラインをインスタンス化する
モデルを使用してオブジェクトを検出し、それに画像を渡します。
```py
>>> from transformers import pipeline
>>> import requests
>>> url = "https://i.imgur.com/2lnWoly.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> obj_detector = pipeline("object-detection", model="devonho/detr-resnet-50_finetuned_cppe5")
>>> obj_detector(image)
```
必要に応じて、パイプラインの結果を手動で複製することもできます。
```py
>>> image_processor = AutoImageProcessor.from_pretrained("devonho/detr-resnet-50_finetuned_cppe5")
>>> model = AutoModelForObjectDetection.from_pretrained("devonho/detr-resnet-50_finetuned_cppe5")
>>> with torch.no_grad():
... inputs = image_processor(images=image, return_tensors="pt")
... outputs = model(**inputs)
... target_sizes = torch.tensor([image.size[::-1]])
... results = image_processor.post_process_object_detection(outputs, threshold=0.5, target_sizes=target_sizes)[0]
>>> for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
... box = [round(i, 2) for i in box.tolist()]
... print(
... f"Detected {model.config.id2label[label.item()]} with confidence "
... f"{round(score.item(), 3)} at location {box}"
... )
Detected Coverall with confidence 0.566 at location [1215.32, 147.38, 4401.81, 3227.08]
Detected Mask with confidence 0.584 at location [2449.06, 823.19, 3256.43, 1413.9]
```
結果をプロットしてみましょう:
```py
>>> draw = ImageDraw.Draw(image)
>>> for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
... box = [round(i, 2) for i in box.tolist()]
... x, y, x2, y2 = tuple(box)
... draw.rectangle((x, y, x2, y2), outline="red", width=1)
... draw.text((x, y), model.config.id2label[label.item()], fill="white")
>>> image
```
<div class="flex justify-center">
<img src="https://i.imgur.com/4QZnf9A.png" alt="Object detection result on a new image"/>
</div>
|
transformers/docs/source/ja/tasks/object_detection.md/0
|
{
"file_path": "transformers/docs/source/ja/tasks/object_detection.md",
"repo_id": "transformers",
"token_count": 12693
}
| 301
|
<!--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 TFLite
[TensorFlow Lite](https://www.tensorflow.org/lite/guide)は、モバイルフォン、組み込みシステム、およびモノのインターネット（IoT）デバイスなど、リソースに制約のあるデバイスに機械学習モデルを展開するための軽量なフレームワークです。TFLiteは、計算能力、メモリ、および電力消費が限られているこれらのデバイス上でモデルを効率的に最適化して実行するために設計されています。
TensorFlow Liteモデルは、`.tflite`ファイル拡張子で識別される特別な効率的なポータブル形式で表されます。
🤗 Optimumは、🤗 TransformersモデルをTFLiteにエクスポートするための機能を`exporters.tflite`モジュールを介して提供しています。サポートされているモデルアーキテクチャのリストについては、[🤗 Optimumのドキュメント](https://huggingface.co/docs/optimum/exporters/tflite/overview)をご参照ください。
モデルをTFLiteにエクスポートするには、必要な依存関係をインストールしてください：
```bash
pip install optimum[exporters-tf]
```
すべての利用可能な引数を確認するには、[🤗 Optimumドキュメント](https://huggingface.co/docs/optimum/main/en/exporters/tflite/usage_guides/export_a_model)を参照するか、コマンドラインでヘルプを表示してください：
```bash
optimum-cli export tflite --help
```
🤗 Hubからモデルのチェックポイントをエクスポートするには、例えば `google-bert/bert-base-uncased` を使用する場合、次のコマンドを実行します：
```bash
optimum-cli export tflite --model google-bert/bert-base-uncased --sequence_length 128 bert_tflite/
```
進行状況を示すログが表示され、生成された `model.tflite` が保存された場所も表示されるはずです：
```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
```
上記の例は🤗 Hubからチェックポイントをエクスポートする方法を示しています。ローカルモデルをエクスポートする場合、まずモデルの重みファイルとトークナイザファイルを同じディレクトリ（`local_path`）に保存したことを確認してください。CLIを使用する場合、🤗 Hubのチェックポイント名の代わりに`model`引数に`local_path`を渡します。
|
transformers/docs/source/ja/tflite.md/0
|
{
"file_path": "transformers/docs/source/ja/tflite.md",
"repo_id": "transformers",
"token_count": 1416
}
| 302
|
<!--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.
-->
# 대규모 언어 모델의 속도 및 메모리 최적화 [[optimizing-llms-for-speed-and-memory]]
[[open-in-colab]]
GPT3/4, [Falcon](https://huggingface.co/tiiuae/falcon-40b), [Llama](https://huggingface.co/meta-llama/Llama-2-70b-hf)와 같은 대규모 언어 모델의 인간 중심 과제를 해결하는 능력이 빠르게 발전하고 있으며, 현대 지식 기반 산업에서 필수 도구로 자리잡고 있습니다. 그러나 이러한 모델을 실제 과제에 배포하는 것은 여전히 어려운 과제입니다.
- 인간과 비슷한 텍스트 이해 및 생성 능력을 보이기 위해, 현재 대규모 언어 모델은 수십억 개의 매개변수로 구성되어야 합니다 (참조: [Kaplan et al](https://arxiv.org/abs/2001.08361), [Wei et. al](https://arxiv.org/abs/2206.07682)). 이는 추론을 위한 메모리 요구를 크게 증가시킵니다.
- 많은 실제 과제에서 대규모 언어 모델은 방대한 맥락 정보를 제공받아야 합니다. 이는 모델이 추론 과정에서 매우 긴 입력 시퀀스를 처리할 수 있어야 한다는 것을 뜻합니다.
이러한 과제의 핵심은 대규모 언어 모델의 계산 및 메모리 활용 능력을 증대시키는 데 있습니다. 특히 방대한 입력 시퀀스를 처리할 때 이러한 능력이 중요합니다.
이 가이드에서는 효율적인 대규모 언어 모델 배포를 위한 효과적인 기법들을 살펴보겠습니다.
1. **낮은 정밀도:** 연구에 따르면, [8비트와 4비트](./main_classes/quantization.md)와 같이 낮은 수치 정밀도로 작동하면 모델 성능의 큰 저하 없이 계산상의 이점을 얻을 수 있습니다.
2. **플래시 어텐션:** 플래시 어텐션은 메모리 효율성을 높일 뿐만 아니라 최적화된 GPU 메모리 활용을 통해 효율성을 향상시키는 어텐션 알고리즘의 변형입니다.
3. **아키텍처 혁신:** 추론 시 대규모 언어 모델은 주로 동일한 방식(긴 입력 맥락을 가진 자기회귀 텍스트 생성 방식)으로 배포되는데, 더 효율적인 추론을 가능하게 하는 특화된 모델 아키텍처가 제안되었습니다. 이러한 모델 아키텍처의 가장 중요한 발전으로는 [Alibi](https://arxiv.org/abs/2108.12409), [Rotary embeddings](https://arxiv.org/abs/2104.09864), [Multi-Query Attention (MQA)](https://arxiv.org/abs/1911.02150), [Grouped-Query-Attention (GQA)]((https://arxiv.org/abs/2305.13245))이 있습니다.
이 가이드에서는 텐서의 관점에서 자기회귀 생성에 대한 분석을 제공합니다. 낮은 정밀도를 채택하는 것의 장단점을 논의하고, 최신 어텐션 알고리즘을 포괄적으로 탐구하며, 향상된 대규모 언어 모델 아키텍처에 대해 논합니다. 이 과정에서 각 기능의 개선 사항을 보여주는 실용적인 예제를 확인합니다.
## 1. 낮은 정밀도 [[1-lower-precision]]
대규모 언어 모델을 가중치 행렬과 벡터의 집합으로 보고, 텍스트 입력을 벡터의 시퀀스로 본다면, 대규모 언어 모델의 메모리 요구사항을 가장 잘 이해할 수 있습니다. 이어지는 내용에서 *가중치*는 모델의 모든 가중치 행렬과 벡터를 의미합니다.
이 가이드를 작성하는 시점의 대규모 언어 모델은 최소 몇십억 개의 매개변수로 구성되어 있습니다. 각 매개변수는 `4.5689`와 같은 십진수로 이루어져 있으며, 보통 [float32](https://en.wikipedia.org/wiki/Single-precision_floating-point_format), [bfloat16](https://en.wikipedia.org/wiki/Bfloat16_floating-point_format) 또는 [float16](https://en.wikipedia.org/wiki/Half-precision_floating-point_format) 형식으로 저장됩니다. 이를 통해 대규모 언어 모델을 메모리에 로드하는 데 필요한 메모리의 요구사항을 쉽게 계산할 수 있습니다:
> *X * 10억 개의 매개변수를 가진 모델의 가중치를 로드하려면 float32 정밀도에서 대략 4 * X GB의 VRAM이 필요합니다.*
요즘에는 모델이 float32 정밀도로 훈련되는 경우는 드물고, 일반적으로 bfloat16 정밀도나 가끔 float16 정밀도로 훈련됩니다. 따라서 경험적으로 알아낸 법칙은 다음과 같습니다:
> *X * 10억 개의 매개변수를 가진 모델의 가중치를 로드하려면 bfloat16/float16 정밀도에서 대략 2 * X GB의 VRAM이 필요합니다.*
짧은 텍스트 입력(1024 토큰 미만)의 경우, 추론을 위한 메모리 요구 사항의 대부분은 가중치를 로드하는 데 필요한 메모리 요구 사항입니다. 따라서 지금은 추론을 위한 메모리 요구 사항이 모델의 가중치를 GPU VRAM에 로드하는 데 필요한 메모리 요구 사항과 같다고 가정합시다.
모델을 bfloat16으로 로드하는 데 대략 얼마나 많은 VRAM이 필요한지 몇 가지 예를 들어보겠습니다:
- **GPT3**는 2 \* 175 GB = **350 GB** VRAM이 필요합니다.
- [**Bloom**](https://huggingface.co/bigscience/bloom)은 2 \* 176 GB = **352 GB** VRAM이 필요합니다.
- [**Llama-2-70b**](https://huggingface.co/meta-llama/Llama-2-70b-hf)는 2 \* 70 GB = **140 GB** VRAM이 필요합니다.
- [**Falcon-40b**](https://huggingface.co/tiiuae/falcon-40b)는 2 \* 40 GB = **80 GB** VRAM이 필요합니다.
- [**MPT-30b**](https://huggingface.co/mosaicml/mpt-30b)는 2 * 30 GB = **60 GB** VRAM이 필요합니다.
- [**bigcode/starcoder**](https://huggingface.co/bigcode/starcoder)는 2 * 15.5 GB = **31 GB** VRAM이 필요합니다.
이 문서를 작성하는 시점에서, 현재 시장에서 가장 큰 GPU 칩은 80GB의 VRAM을 제공하는 A100과 H100입니다. 앞서 언급된 대부분의 모델들은 로드하기 위해서는 최소 80GB 이상의 용량을 필요로 하며, 따라서 [텐서 병렬 처리](https://huggingface.co/docs/transformers/perf_train_gpu_many#tensor-parallelism) 및/또는 [파이프라인 병렬 처리](https://huggingface.co/docs/transformers/perf_train_gpu_many#naive-model-parallelism-vertical-and-pipeline-parallelism)를 반드시 필요로 합니다.
🤗 Transformers는 텐서 병렬 처리를 바로 지원하지 않습니다. 이는 모델 아키텍처가 특정 방식으로 작성되어야 하기 때문입니다. 텐서 병렬 처리를 지원하는 방식으로 모델을 작성하는 데 관심이 있다면 [the text-generation-inference library](https://github.com/huggingface/text-generation-inference/tree/main/server/text_generation_server/models/custom_modeling)를 참조해 보시기 바랍니다.
기본적인 파이프라인 병렬 처리는 바로 지원됩니다. 이를 위해 단순히 모델을 `device="auto"`로 로드하면 [여기](https://huggingface.co/docs/accelerate/v0.22.0/en/concept_guides/big_model_inference)에 설명된 대로 사용 가능한 GPU에 모델의 서로 다른 레이어를 자동으로 배치합니다. 이것은 매우 효과적이긴 하지만 이러한 기본 파이프라인 병렬 처리는 GPU 유휴 문제를 해결하지 못한다는 점을 유의해야 합니다. 더 발전된 파이프라인 병렬 처리가 필요하며, 이에 대한 설명은 [여기](https://huggingface.co/docs/transformers/en/perf_train_gpu_many#naive-model-parallelism-vertical-and-pipeline-parallelism)에서 확인할 수 있습니다.
80GB A100 GPU 8개를 가진 노드에 접근할 수 있다면, BLOOM을 다음과 같이 로드할 수 있습니다.
```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)
```
`device_map="auto"`를 사용하면 모든 사용 가능한 GPU에 어텐션 레이어가 고르게 분산됩니다.
이 가이드에서는 [bigcode/octocoder](https://huggingface.co/bigcode/octocoder)를 사용할 것입니다. 이 모델은 단일 40GB A100 GPU 장치에서 실행할 수 있습니다. 앞으로 적용할 모든 메모리 및 속도 최적화는 모델 또는 텐서 병렬 처리를 필요로 하는 다른 모델에도 동일하게 적용될 수 있습니다.
모델이 bfloat16 정밀도로 로드되기 때문에, 위의 경험적으로 알아낸 법칙을 사용하면 `bigcode/octocoder`를 사용하여 추론을 실행하기 위한 메모리 요구 사항이 약 31GB VRAM일 것으로 예상됩니다. 한 번 시도해 보겠습니다.
먼저 모델과 토크나이저를 로드한 다음, 둘 다 Transformers의 [파이프라인](https://huggingface.co/docs/transformers/main_classes/pipelines) 객체에 전달합니다.
```python
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
import torch
model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", torch_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
```
**출력**:
```
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
```
좋습니다. 이제 결과를 직접 사용하여 바이트를 기가바이트로 변환할 수 있습니다.
```python
def bytes_to_giga_bytes(bytes):
return bytes / 1024 / 1024 / 1024
```
[`torch.cuda.max_memory_allocated`](https://pytorch.org/docs/stable/generated/torch.cuda.max_memory_allocated.html)를 호출하여 최대 GPU 메모리 할당을 측정해 보겠습니다.
```python
bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
```
**출력**:
```bash
29.0260648727417
```
대략적으로 계산한 결과와 거의 일치합니다! 바이트에서 킬로바이트로 변환할 때 1000이 아닌 1024로 곱해야 하므로 숫자가 정확하지 않은 것을 알 수 있습니다. 따라서 대략적으로 계산할 때 공식은 "최대 X GB"으로 이해할 수 있습니다. 만약 우리가 모델을 float32 정밀도로 실행하려고 했다면 더 큰 크기인 64GB의 VRAM이 필요했을 것입니다.
> 거의 모든 모델이 요즘 bfloat16으로 학습되므로, [GPU가 bfloat16을 지원](https://discuss.pytorch.org/t/bfloat16-native-support/117155/5)한다면 모델을 float32 정밀도로 실행할 이유가 없습니다. float32로 돌리는 모델은 학습할 때 사용했던 정밀도보다 더 나은 추론 결과를 제공하지 않습니다.
모델 가중치가 어떤 정밀도 형식으로 Hub에 저장되어 있는지 확실하지 않은 경우, HuggingFace Hub에서 해당 체크포인트 config의 `"torch_dtype"`을 확인하면 됩니다, *예*를 들어 [여기](https://huggingface.co/meta-llama/Llama-2-7b-hf/blob/6fdf2e60f86ff2481f2241aaee459f85b5b0bbb9/config.json#L21)를 확인하세요. 모델을 `from_pretrained(..., torch_dtype=...)`로 로드할 때는 config에 명시된 정밀도 유형과 동일한 정밀도로 설정하는 것이 권장됩니다. 단, 원래 유형이 float32인 경우 추론을 위해 `float16` 또는 `bfloat16`을 둘 다 사용할 수 있습니다.
이제 `flush(...)` 함수를 정의하여 모든 메모리를 해제하고, GPU 메모리의 최대 할당량을 정확하게 측정하도록 합시다.
```python
del pipe
del model
import gc
import torch
def flush():
gc.collect()
torch.cuda.empty_cache()
torch.cuda.reset_peak_memory_stats()
```
다음 실험을 위해 바로 호출해 봅시다.
```python
flush()
```
최근 버전의 accelerate 라이브러리에서는 `release_memory()`라는 유틸리티 메소드도 사용할 수 있습니다.
```python
from accelerate.utils import release_memory
# ...
release_memory(model)
```
만약 GPU에 32GB의 VRAM이 없다면 어떻게 될까요? 모델 가중치를 성능에 큰 손실 없이 8비트 또는 4비트로 양자화할 수 있다는 것이 밝혀졌습니다(참고: [Dettmers et al.](https://arxiv.org/abs/2208.07339)). 최근의 [GPTQ 논문](https://arxiv.org/abs/2210.17323) 에서는 모델을 3비트 또는 2비트로 양자화해도 성능 손실이 허용 가능한 수준임을 보여주었습니다🤯.
너무 자세한 내용은 다루지 않고 설명하자면, 양자화는 가중치의 정밀도를 줄이면서 모델의 추론 결과를 가능한 한 정확하게(즉, bfloat16과 최대한 가깝게) 유지하려고 합니다. 양자화는 특히 텍스트 생성에 잘 작동하는데, 이는 우리가 *가장 가능성 있는 다음 토큰 집합*을 선택하는 것에 초점을 두고 있기 때문이며, 다음 토큰의 *logit* 분포값을 정확하게 예측할 필요는 없기 때문입니다. 핵심은 다음 토큰 *logit* 분포가 대략적으로 동일하게 유지되어 `argmax` 또는 `topk` 연산이 동일한 결과를 제공하는 것입니다.
다양한 양자화 기법이 존재하지만, 자세히 다루지는 않을 것입니다. 일반적으로 모든 양자화 기법은 다음과 같이 작동합니다:
- 1. 모든 가중치를 목표 정밀도로 양자화합니다.
- 2. 양자화된 가중치를 로드하고, bfloat16 정밀도의 입력 벡터 시퀀스를 모델에 전달합니다.
- 3. 가중치를 동적으로 bfloat16으로 반대로 양자화(dequantize)하여 입력 벡터와 함께 bfloat16 정밀도로 계산을 수행합니다.
간단히 말해서, *입력-가중치 행렬* 곱셈은, \\( X \\)가 *입력*, \\( W \\)가 가중치 행렬, \\( Y \\)가 출력인 경우 다음과 같습니다:
$$ Y = X * W $$
위 공식이 다음과 같이 변경됩니다
$$ Y = X * \text{dequantize}(W) $$
모든 행렬 곱셈에 대해 위와 같이 수행됩니다. 입력이 네트워크 그래프를 통과하면서 모든 가중치 행렬에 대해 역양자화(dequantization)와 재양자화(re-quantization)가 순차적으로 수행됩니다.
따라서, 양자화된 가중치를 사용할 때 추론 시간이 감소하지 **않고** 오히려 증가하는 경우가 많습니다. 이제 이론은 충분하니 실제로 시도해 봅시다! Transformers를 사용하여 가중치를 양자화하려면 [`bitsandbytes`](https://github.com/TimDettmers/bitsandbytes) 라이브러리가 설치되어 있는지 확인해야 합니다.
```bash
!pip install bitsandbytes
```
그런 다음 `from_pretrained`에 `load_in_8bit=True` 플래그를 추가하여 8비트 양자화로 모델을 로드할 수 있습니다.
```python
model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", load_in_8bit=True, pad_token_id=0)
```
이제 예제를 다시 실행하고 메모리 사용량을 측정해 봅시다.
```python
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
result = pipe(prompt, max_new_tokens=60)[0]["generated_text"][len(prompt):]
result
```
**출력**:
```
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
```
좋습니다. 정확도 손실 없이 이전과 동일한 결과를 얻고 있습니다! 이번에는 사용된 메모리 양을 확인해 봅시다.
```python
bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
```
**출력**:
```
15.219234466552734
```
훨씬 적네요! 메모리 사용량이 15GB를 조금 넘는 수준으로 줄어들어 4090과 같은 소비자용 GPU에서도 이 모델을 실행할 수 있습니다. 메모리 효율성에서 매우 큰 향상을 보이고 있으며 모델 출력의 품질 저하도 거의 없습니다. 그러나 추론 중에 약간의 속도 저하가 발생한 것을 확인할 수 있습니다.
모델을 삭제하고 메모리를 다시 초기화합니다.
```python
del model
del pipe
```
```python
flush()
```
이제 4비트 양자화가 제공하는 최대 GPU 메모리 사용량을 확인해 봅시다. 4비트로 모델을 양자화하려면 이전과 동일한 API를 사용하되 이번에는 `load_in_8bit=True` 대신 `load_in_4bit=True`를 전달하면 됩니다.
```python
model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", load_in_4bit=True, low_cpu_mem_usage=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
```
**출력**:
```
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
```
바로 전 코드 스니펫에서 `python`만 누락되고, 이 전과 거의 동일한 출력 텍스트를 보고 있습니다. 이제 얼마나 많은 메모리가 필요했는지 확인해 봅시다.
```python
bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
```
**출력**:
```
9.543574333190918
```
9.5GB밖에 되지 않습니다! 150억 개 이상의 파라미터를 가진 모델인 것을 감안하면 매우 적은 양입니다.
여기서는 모델의 정확도 저하가 거의 없음을 확인할 수 있지만, 실제로는 4비트 양자화를 8비트 양자화나 `bfloat16`를 사용한 추론 결과와 비교하면 결과가 다를 수 있습니다. 사용자가 직접 시도해 보는 것이 좋겠습니다.
또한 4비트 양자화에 사용된 더 공격적인 양자화 방법으로 인해 추론 시 \\( \text{quantize} \\)와 \\( \text{dequantize} \\) 과정이 더 오래 걸리므로 여기서도 8비트 양자화와 비교하여 추론 속도가 약간 느려졌음을 유의하세요.
```python
del model
del pipe
```
```python
flush()
```
전체적으로 OctoCoder를 8비트 정밀도로 실행하면 필요한 GPU VRAM이 32GB에서 15GB로 줄어들었고, 4비트 정밀도로 모델을 실행하면 필요한 GPU VRAM이 9GB로 더 줄어드는 것을 확인했습니다.
4비트 양자화는 RTX3090, V100, T4와 같은 GPU에서 모델을 실행할 수 있게 해주며, 이는 대부분의 사람들이 접근할 수 있는 GPU입니다.
양자화에 대한 더 많은 정보를 확인하고 4비트보다 더 적은 GPU VRAM 메모리로 모델을 양자화하거나, 더 많은 양자화 관련 정보를 보려면 [`AutoGPTQ`](https://huggingface.co/docs/transformers/main/en/main_classes/quantization#autogptq-integration%60) 구현을 참조하는 것을 추천합니다.
> 결론적으로, 모델 양자화는 향상된 메모리 효율성과 모델 정확성 간의 균형을 맞추는 것이며, 경우에 따라 추론 시간에도 영향을 미칠 수 있습니다.
실제 사례에서 GPU 메모리가 충분하다면, 양자화를 고려할 필요가 없습니다. 그러나 많은 GPU는 양자화 없이 대규모 언어 모델을 실행할 수 없으며, 이 경우 4비트 및 8비트 양자화가 매우 유용한 도구입니다.
사용과 관련한 더 자세한 정보는 [트랜스포머 양자화 문서](https://huggingface.co/docs/transformers/main_classes/quantization#general-usage)를 참고하는 것을 강력히 추천합니다. 다음으로, 더 나은 알고리즘과 개선된 모델 아키텍처를 사용하여 계산 및 메모리 효율성을 향상시키는 방법을 살펴보겠습니다.
## 2. 플래시 어텐션 [[2-flash-attention]]
오늘날의 최고 성능을 자랑하는 대규모 언어 모델은 대체로 피드포워드 레이어(feed-forward layer), 활성화 레이어(activation layer), 레이어 정규화 레이어(layer normalization layer), 그리고 가장 중요한 셀프 어텐션 레이어(self-attention layer)로 구성된 아키텍처를 공유하고 있습니다.
셀프 어텐션 레이어는 입력 토큰 간의 문맥적 관계를 이해할 수 있게 해 주기 때문에 대규모 언어 모델의 핵심 요소입니다.
하지만 셀프 어텐션 레이어의 최대 GPU 메모리 소비는 입력 토큰의 수(이하 \\( N \\)으로 표기)와 함께 계산 및 메모리 복잡성이 *2차적*으로 증가합니다. 입력 시퀀스가 짧은 경우(최대 1000개)에는 크게 눈에 띄지 않지만, 더 긴 입력 시퀀스(약 16000개)에서는 심각한 문제가 됩니다.
자세히 한 번 들여다 봅시다. 길이 \\( N \\)의 입력 \\( \mathbf{X} \\)에 대한 셀프 어텐션 레이어의 출력 \\( \mathbf{O} \\)을 계산하는 공식은 다음과 같습니다:
$$ \textbf{O} = \text{Attn}(\mathbf{X}) = \mathbf{V} \times \text{Softmax}(\mathbf{QK}^T) \text{ with } \mathbf{Q} = \mathbf{W}_q \mathbf{X}, \mathbf{V} = \mathbf{W}_v \mathbf{X}, \mathbf{K} = \mathbf{W}_k \mathbf{X} $$
\\( \mathbf{X} = (\mathbf{x}1, ... \mathbf{x}{N}) \\)는 어텐션 레이어의 입력 시퀀스입니다. 프로젝션 \\( \mathbf{Q} \\)와 \\( \mathbf{K} \\)는 각각 \\( N \\)개의 벡터로 구성되며, 그 결과 \\( \mathbf{QK}^T \\)의 크기는 \\( N^2 \\)가 됩니다.
대규모 언어 모델은 일반적으로 여러 개의 어텐션 헤드를 가지고 있어 여러 개의 셀프 어텐션 계산을 병렬로 수행합니다. 대규모 언어 모델이 40개의 어텐션 헤드를 가지고 bfloat16 정밀도로 실행된다고 가정하면, \\( \mathbf{QK^T} \\) 행렬을 저장하는 데 필요한 메모리를 \\( 40 * 2 * N^2 \\) 바이트로 계산할 수 있습니다. \\( N=1000 \\)일 때는 약 50MB의 VRAM만 필요하지만, \\( N=16000 \\)일 때는 19GB의 VRAM이 필요하며, \\( N=100,000 \\)일 때는 \\( \mathbf{QK^T} \\) 행렬을 저장하기 위해 거의 1TB의 VRAM이 필요합니다.
요약하자면, 기본 셀프 어텐션 알고리즘은 큰 입력 컨텍스트에 대해 매우 과도한 메모리 사용을 요구하게 됩니다.
대규모 언어 모델의 텍스트 이해 및 생성 능력이 개선되면서 점점 더 복잡한 작업에 사용되고 있습니다. 한때 몇 문장의 번역이나 요약을 처리하던 모델이 이제는 전체 페이지를 처리해야 하게 되면서 광범위한 입력 길이를 처리할 수 있는 능력이 요구되고 있습니다.
어떻게 하면 큰 입력 길이에 대한 과도한 메모리 요구를 없앨 수 있을까요? \\( QK^T \\) 행렬을 제거하는 새로운 셀프 어텐션 메커니즘을 계산하는 방법이 필요합니다. [Tri Dao et al.](https://arxiv.org/abs/2205.14135)은 바로 이러한 새로운 알고리즘을 개발하였고, 그것이 **플래시 어텐션(Flash Attention)**입니다.
간단히 말해, 플래시 어텐션은 \\(\mathbf{V} \times \text{Softmax}(\mathbf{QK}^T\\)) 계산을 분할하는데, 여러 번의 소프트맥스 계산을 반복하면서 작은 청크 단위로 출력을 계산합니다:
$$ \textbf{O}_i \leftarrow s^a_{ij} * \textbf{O}_i + s^b_{ij} * \mathbf{V}_{j} \times \text{Softmax}(\mathbf{QK}^T_{i,j}) \text{ for multiple } i, j \text{ iterations} $$
여기서 \\( s^a_{ij} \\)와 \\( s^b_{ij} \\)는 각 \\( i \\)와 \\( j \\)에 대해 계산되는 소프트맥스 정규화 통계량입니다.
플래시 어텐션의 전체 알고리즘은 더 복잡하며, 본 가이드의 범위를 벗어나기 때문에 크게 단순화하였습니다. 여러분은 잘 작성된 [Flash Attention paper](https://arxiv.org/abs/2205.14135) 논문을 참조하여 더 자세한 내용을 확인해 보시기 바랍니다.
주요 요점은 다음과 같습니다:
> 소프트맥스 정규화 통계량과 몇 가지 스마트한 수학적 방법을 사용함으로써, 플래시 어텐션은 기본 셀프 어텐션 레이어와 **숫자적으로 동일한** 출력을 제공하고 메모리 비용은 \\( N \\)에 따라 선형적으로만 증가합니다.
공식을 보면, 플래시 어텐션이 더 많은 계산을 필요로 하기 때문에 기본 셀프 어텐션 공식보다 훨씬 느릴 것이라고 생각할 수 있습니다. 실제로 플래시 어텐션은 소프트맥스 정규화 통계량을 지속적으로 다시 계산해야 하기 때문에 일반 어텐션보다 더 많은 FLOP이 필요합니다. (더 자세한 내용은 [논문](https://arxiv.org/abs/2205.14135)을 참조하세요)
> 그러나 플래시 어텐션은 기본 어텐션보다 추론 속도가 훨씬 빠릅니다. 이는 GPU의 느리고 고대역폭 메모리(VRAM)의 사용량을 크게 줄이고 대신 빠른 온칩 메모리(SRAM)에 집중할 수 있기 때문입니다.
본질적으로, 플래시 어텐션의 모든 중간 단계의 쓰기 및 읽기 작업은 느린 VRAM 메모리에 접근하지 않고 빠른 *온칩* SRAM 메모리를 사용하여 출력 벡터 \\( \mathbf{O} \\)를 계산할 수 있도록 합니다.
현실적으로 플래시 어텐션이 사용 가능한 경우 이를 **사용하지 않을** 이유는 전혀 없습니다. 이 알고리즘은 수학적으로 동일한 출력을 제공하며, 더 빠르고 메모리 효율적입니다.
실제 예를 살펴보겠습니다.
우리의 OctoCoder 모델은 이제 *시스템 프롬프트*가 포함된 훨씬 더 긴 입력 프롬프트를 받게 됩니다. 시스템 프롬프트는 대규모 언어 모델을 사용자의 작업에 맞춘 더 나은 어시스턴트로 유도하는 데 사용됩니다. 다음 예제에서는 OctoCoder를 더 나은 코딩 어시스턴트로 만들기 위한 시스템 프롬프트를 사용합니다.
```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
-----
"""
```
시연을 위해 시스템 프롬프트를 10번 중복하여 증가시켜 플래시 어텐션의 메모리 절약 효과를 관찰할 수 있을 만큼 입력 길이를 충분히 길게 만듭니다. 원래의 텍스트 프롬프트를 다음과 같이 추가합니다. `"Question: Please write a function in Python that transforms bytes to Giga bytes.\n\nAnswer: Here"`
```python
long_prompt = 10 * system_prompt + prompt
```
모델을 다시 bfloat16 정밀도로 인스턴스화합니다.
```python
model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", torch_dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("bigcode/octocoder")
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
```
이제 플래시 어텐션을 *사용하지 않고* 이전과 동일하게 모델을 실행하여 최대 GPU 메모리 요구량과 추론 시간을 측정해 봅시다.
```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
```
**출력**:
```
Generated in 10.96854019165039 seconds.
Sure. Here is a function that does that.\n\ndef bytes_to_giga(bytes):\n return bytes / 1024 / 1024 / 1024\n\nAnswer: Sure. Here is a function that does that.\n\ndef
````
이전과 동일한 출력을 얻고 있지만, 이번에는 모델이 답변을 여러 번 반복하여 60개의 토큰이 잘릴 때까지 계속됩니다. 시연을 위해 시스템 프롬프트를 10번 반복했기 때문에 모델이 스스로 반복하도록 유도한 결과입니다. 이는 놀라운 일이 아닙니다.
**참고** 실제 응용에서는 시스템 프롬프트를 10번 반복할 필요가 없습니다. 한 번만 사용하면 충분합니다!
최대 GPU 메모리 요구량을 측정해 봅시다.
```python
bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
```
**출력**:
```bash
37.668193340301514
```
보시다시피 최대 GPU 메모리 요구량이 처음보다 상당히 높아졌습니다. 이는 주로 입력 시퀀스가 길어졌기 때문입니다. 또한 생성 시간이 이제 1분을 넘어갑니다.
다음 실험을 위해 `flush()`를 호출하여 GPU 메모리를 초기화합니다.
```python
flush()
```
비교를 위해, 동일한 기능을 실행하되 플래시 어텐션을 활성화해 보겠습니다.
이를 위해 모델을 [BetterTransformer](https://huggingface.co/docs/optimum/bettertransformer/overview)로 변환하고, 이를 통해 PyTorch의 [SDPA self-attention](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention)을 활성화하면 플래시 어텐션을 사용할 수 있습니다.
```python
model.to_bettertransformer()
```
이제 이전과 동일한 코드 스니펫을 실행하면, 내부적으로 Transformers가 플래시 어텐션을 사용할 것입니다.
```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
```
**출력**:
```
Generated in 3.0211617946624756 seconds.
Sure. Here is a function that does that.\n\ndef bytes_to_giga(bytes):\n return bytes / 1024 / 1024 / 1024\n\nAnswer: Sure. Here is a function that does that.\n\ndef
```
이전과 동일한 결과를 얻었지만, 플래시 어텐션 덕분에 매우 큰 속도 향상을 관찰할 수 있습니다.
메모리 소비량을 마지막으로 한 번 더 측정해 봅시다.
```python
bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
```
**출력**:
```
32.617331981658936
```
그리고 우리는 처음에 보았던 GPU 메모리 요구량인 29GB로 돌아왔습니다.
플래시 어텐션을 사용하여 매우 긴 입력 시퀀스를 전달할 때 처음에 짧은 입력 시퀀스를 전달했을 때와 비교하여 약 100MB 정도의 GPU 메모리를 더 사용한다는 것을 관찰할 수 있습니다.
```py
flush()
```
플래시 어텐션 사용에 대한 자세한 정보는 [이 문서 페이지](https://huggingface.co/docs/transformers/en/perf_infer_gpu_one#flashattention-2)를 참조해 주세요.
## 3. 아키텍처 혁신 [[3-architectural-innovations]]
지금까지 우리는 계산 및 메모리 효율성을 개선하기 위해 다음을 살펴보았습니다:
- 가중치를 낮은 정밀도 형식으로 변환
- 셀프 어텐션 알고리즘을 보다 더 메모리 및 계산 효율적인 버전으로 교체
이제 긴 텍스트 입력이 필요한 작업에 가장 효과적이고 효율적인 대규모 언어 모델 아키텍처로 변경하는 방법을 살펴보겠습니다. 작업의 예시는 다음과 같습니다:
- 검색 증강 질의 응답
- 요약
- 채팅
*채팅*을 위해서는 대규모 언어 모델이 긴 텍스트 입력을 처리하는 것뿐만 아니라 사용자와 어시스턴트 간의 대화도 효율적으로 처리할 수 있어야 합니다(예: ChatGPT).
한번 학습된 후에는 대규모 언어 모델의 기본 아키텍처를 변경하기 어렵기 때문에, 대규모 언어 모델의 작업에 대한 고려를 미리 하고 이에 따라 모델의 아키텍처를 최적화하는 것이 중요합니다. 긴 입력 시퀀스에 대해 메모리 또는 성능의 병목 현상을 빠르게 발생시키는 모델 아키텍처의 중요한 두 가지 구성 요소가 있습니다.
- 위치 임베딩
- 키-값 캐시
각 구성 요소를 더 자세히 살펴보겠습니다.
### 3.1 대규모 언어 모델의 위치 임베딩 개선 [[31-improving-positional-embeddings-of-llms]]
셀프 어텐션은 각 토큰을 서로의 토큰과 연관시킵니다.
예를 들어, 텍스트 입력 시퀀스 *"Hello", "I", "love", "you"*의 \\( \text{Softmax}(\mathbf{QK}^T) \\) 행렬은 다음과 같을 수 있습니다:
각 단어 토큰은 다른 모든 단어 토큰에 주의를 기울이는 확률 질량을 부여받아 모든 다른 단어 토큰과 관계를 맺게 됩니다. 예를 들어, 단어 *"love"*는 단어 *"Hello"*에 5%, *"I"*에 30%, 그리고 자신에게 65%의 주의를 기울입니다.
셀프 어텐션 기반 대규모 언어 모델이 위치 임베딩이 없는 경우 텍스트 입력의 위치를 이해하는 데 큰 어려움을 겪을 것입니다. 이는 \\( \mathbf{QK}^T \\)에 의해 계산된 확률 점수가 상대적 위치 거리에 상관없이 각 단어 토큰을 다른 모든 단어 토큰과 \\( O(1) \\) 계산으로 연관시키기 때문입니다. 따라서 위치 임베딩이 없는 대규모 언어 모델은 각 토큰이 다른 모든 토큰과 동일한 거리에 있는 것으로 나타나기 때문에, *"Hello I love you"*와 *"You love I hello"*를 구분하는 것이 매우 어렵습니다.
대규모 언어 모델이 문장의 순서를 이해하려면 추가적인 *단서*가 필요하며, 이는 일반적으로 *위치 인코딩* (또는 *위치 임베딩*이라고도 함)의 형태로 적용됩니다.
위치 인코딩은 각 토큰의 위치를 숫자 표현으로 인코딩하여 대규모 언어 모델이 문장의 순서를 더 잘 이해할 수 있도록 도와줍니다.
[*Attention Is All You Need*](https://arxiv.org/abs/1706.03762) 논문의 저자들은 사인 함수 기반의 위치 임베딩 \\( \mathbf{P} = \mathbf{p}_1, \ldots, \mathbf{p}_N \\)을 도입했습니다. 각 벡터 \\( \mathbf{p}_i \\)는 위치 \\( i \\)의 사인 함수로 계산됩니다. 위치 인코딩은 입력 시퀀스 벡터에 단순히 더해져 \\( \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 \\) 모델이 문장 순서를 더 잘 학습할 수 있도록 합니다.
고정된 위치 임베딩 대신 [Devlin et al.](https://arxiv.org/abs/1810.04805)과 같은 다른 연구자들은 학습된 위치 인코딩을 사용했습니다. 이 경우 위치 임베딩 \\( \mathbf{P} \\)은 학습 중에 사용됩니다.
사인 함수 및 학습된 위치 임베딩은 문장 순서를 대규모 언어 모델에 인코딩하는 주요 방법이었지만, 이러한 위치 인코딩과 관련된 몇 가지 문제가 발견되었습니다:
1. 사인 함수와 학습된 위치 임베딩은 모두 절대 위치 임베딩으로, 각 위치 ID \\( 0, \ldots, N \\)에 대해 고유한 임베딩을 인코딩합니다. [Huang et al.](https://arxiv.org/abs/2009.13658) 및 [Su et al.](https://arxiv.org/abs/2104.09864)의 연구에 따르면, 절대 위치 임베딩은 긴 텍스트 입력에 대해 대규모 언어 모델 성능이 저하됩니다. 긴 텍스트 입력의 경우, 모델이 절대 위치 대신 입력 토큰 간의 상대적 위치 거리를 학습하는 것이 유리합니다.
2. 학습된 위치 임베딩을 사용할 때, 대규모 언어 모델은 고정된 입력 길이 \\( N \\)으로 학습되어야 하므로, 학습된 입력 길이보다 더 긴 입력 길이에 대해 추론하는 것이 어렵습니다.
최근에는 위에서 언급한 문제를 해결할 수 있는 상대적 위치 임베딩이 더 인기를 끌고 있습니다. 특히 다음과 같은 방법들이 주목받고 있습니다:
- [Rotary Position Embedding (RoPE)](https://arxiv.org/abs/2104.09864)
- [ALiBi](https://arxiv.org/abs/2108.12409)
*RoPE*와 *ALiBi*는 모두 셀프 어텐션 알고리즘 내에서 직접적으로 문장 순서를 모델에게 알려주는 것이 최선이라고 주장합니다. 이는 단어 토큰이 서로 관계를 맺는 곳이기 때문입니다. 구체적으로, 문장 순서를 \\( \mathbf{QK}^T \\) 계산을 수정하는 방식으로 알려주어야 한다는 것입니다.
너무 많은 세부 사항을 다루지 않고, *RoPE*는 위치 정보를 쿼리-키 쌍에 인코딩할 수 있다고 지적합니다. 예를 들어, 각 벡터 \\( \mathbf{q}_i \\)와 \\( \mathbf{x}_j \\)를 각각 \\( \theta * i \\)와 \\( \theta * j \\)의 각도로 회전시킴으로써 다음과 같이 표현할 수 있습니다:
$$ \mathbf{\hat{q}}_i^T \mathbf{\hat{x}}_j = \mathbf{{q}}_i^T \mathbf{R}_{\theta, i -j} \mathbf{{x}}_j. $$
여기서 \\( \mathbf{R}_{\theta, i - j} \\)는 회전 행렬을 나타냅니다. \\( \theta \\)는 훈련 중에 *학습되지 않으며*, 대신 학습 중 최대 입력 시퀀스 길이에 따라 사전 정의된 값으로 설정됩니다.
> 이렇게 함으로써 \\( \mathbf{q}_i \\)와 \\( \mathbf{q}_j \\) 간의 확률 점수는 \\( i \ne j \\)인 경우에만 영향을 받으며, 각 벡터의 특정 위치 \\( i \\)와 \\( j \\)와는 상관없이 오직 상대적 거리 \\( i - j \\)에만 의존하게 됩니다.
*RoPE*는 현재 여러 중요한 대규모 언어 모델이 사용되고 있습니다. 예를 들면:
- [**Falcon**](https://huggingface.co/tiiuae/falcon-40b)
- [**Llama**](https://arxiv.org/abs/2302.13971)
- [**PaLM**](https://arxiv.org/abs/2204.02311)
대안으로, *ALiBi*는 훨씬 더 간단한 상대적 위치 인코딩 방식을 제안합니다. 입력 토큰 간의 상대적 거리를 음수인 정수로서 사전 정의된 값 `m`으로 스케일링하여 \\( \mathbf{QK}^T \\) 행렬의 각 쿼리-키 항목에 소프트맥스 계산 직전에 추가합니다.
[ALiBi](https://arxiv.org/abs/2108.12409) 논문에서 보여주듯이, 이 간단한 상대적 위치 인코딩은 매우 긴 텍스트 입력 시퀀스에서도 모델이 높은 성능을 유지할 수 있게 합니다.
*ALiBi*는 현재 여러 중요한 대규모 언어 모델 모델이 사용하고 있습니다. 예를 들면:
- [**MPT**](https://huggingface.co/mosaicml/mpt-30b)
- [**BLOOM**](https://huggingface.co/bigscience/bloom)
*RoPE*와 *ALiBi* 위치 인코딩은 모두 학습 중에 보지 못한 입력 길이에 대해 확장할 수 있으며, *ALiBi*가 *RoPE*보다 더 잘 확장되는 것으로 나타났습니다. *ALiBi*의 경우, 하삼각 위치 행렬의 값을 입력 시퀀스 길이에 맞추어 증가시키기만 하면 됩니다. *RoPE*의 경우, 학습 중에 사용된 동일한 \\( \theta \\)를 유지하면 학습 중에 보지 못한 매우 긴 텍스트 입력을 전달할 때 성능이 저하됩니다(참고: [Press et al.](https://arxiv.org/abs/2108.12409)). 그러나 커뮤니티는 \\( \theta \\)를 조정하는 몇 가지 효과적인 트릭을 찾아냈으며, 이를 통해 *RoPE* 위치 임베딩이 확장된 텍스트 입력 시퀀스에서도 잘 작동할 수 있게 되었습니다(참고: [here](https://github.com/huggingface/transformers/pull/24653)).
> RoPE와 ALiBi는 모두 훈련 중에 *학습되지 않는* 상대적 위치 임베딩으로 다음과 같은 직관에 기반합니다:
- 텍스트 입력에 대한 위치 단서는 셀프 어텐션 레이어의 \\( QK^T \\) 행렬에 직접 제공되어야 합니다.
- 대규모 언어 모델은 일정한 *상대적* 거리 위치 인코딩을 서로 학습하도록 유도되어야 합니다.
- 텍스트 입력 토큰 간의 거리가 멀어질수록, 그들의 쿼리-값 확률은 낮아져야 합니다. RoPE와 ALiBi는 서로 멀리 떨어진 토큰의 쿼리-키 확률을 낮춥니다. RoPE는 쿼리-키 벡터 간의 각도를 증가시켜 벡터 곱을 감소시키는 방식으로, ALiBi는 벡터 곱에 큰 음수를 추가하는 방식으로 이 작업을 수행합니다.
결론적으로, 큰 텍스트 입력을 처리해야 하는 작업에 배포될 예정인 대규모 언어 모델은 RoPE와 ALiBi와 같은 상대적 위치 임베딩으로 훈련하는 것이 더 좋습니다. 또한 RoPE와 ALiBi를 사용하여 훈련된 대규모 언어 모델이 고정 길이 \\( N_1 = 2048 \\)에서만 훈련되었더라도 위치 임베딩을 외삽하여 \\( N_1 \\)보다 훨씬 큰 텍스트 입력 \\( N_2 = 8192 > N_1 \\)로 실습에서 사용할 수 있음을 유의하세요.
### 3.2 키-값 캐시 [[32-the-key-value-cache]]
대규모 언어 모델을 이용한 자기회귀 텍스트 생성은 입력 시퀀스를 반복적으로 넣고, 다음 토큰을 샘플링하며, 그 다음 토큰을 입력 시퀀스에 추가하고, 대규모 언어 모델이 생성을 완료했다는 토큰을 생성할 때까지 이를 계속 수행하는 방식으로 작동합니다.
자기회귀 생성이 어떻게 작동하는지에 대한 시각적 설명을 보려면 [Transformer's Generate Text Tutorial](https://huggingface.co/docs/transformers/llm_tutorial#generate-text)을 참조하세요.
자기회귀 생성이 실제로 어떻게 작동하는지 보여주는 간단한 코드 스니펫을 실행해 보겠습니다. 여기서는 `torch.argmax`를 통해 가장 가능성이 높은 다음 토큰을 가져올 것입니다.
```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
```
**출력**:
```
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']
```
보시다시피 샘플링된 토큰에 의해 텍스트 입력 토큰을 매번 증가시킵니다.
매우 예외적인 경우를 제외하고, 대규모 언어 모델은 [인과적인 언어 모델링 목표](https://huggingface.co/docs/transformers/tasks/language_modeling#causal-language-modeling)를 사용하여 학습되므로 어텐션 점수의 상삼각 행렬을 마스킹합니다. 이것이 위의 두 다이어그램에서 어텐션 점수가 비어 있는 이유입니다 (즉, 0 확률을 가짐). 인과 언어 모델링에 대한 빠른 요약은 [*Illustrated Self Attention 블로그*](https://jalammar.github.io/illustrated-gpt2/#part-2-illustrated-self-attention)를 참조할 수 있습니다.
결과적으로, 토큰은 *절대* 이전 토큰에 의존하지 않습니다. 더 구체적으로는 \\( \mathbf{q}_i \\) 벡터가 \\( j > i \\)인 경우 어떤 키, 값 벡터 \\( \mathbf{k}_j, \mathbf{v}j \\)와도 연관되지 않습니다. 대신 \\( \mathbf{q}i \\)는 이전의 키-값 벡터 \\( \mathbf{k}{m < i}, \mathbf{v}{m < i} \text{ , for } m \in {0, \ldots i - 1} \\)에만 주의를 기울입니다. 불필요한 계산을 줄이기 위해 각 층의 키-값 벡터를 모든 이전 시간 단계에 대해 캐시할 수 있습니다.
다음으로, 대규모 언어 모델이 각 포워드 패스마다 키-값 캐시를 검색하고 전달하여 이를 활용하도록 합니다.
Transformers에서는 `forward` 호출에 `use_cache` 플래그를 전달하여 키-값 캐시를 검색한 다음 현재 토큰과 함께 전달할 수 있습니다.
```python
past_key_values = None # past_key_values 는 키-값 캐시를 의미
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 형태: [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
```
**출력**:
```
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']
```
키-값 캐시를 사용할 때, 텍스트 입력 토큰의 길이는 *증가하지 않고* 단일 입력 벡터로 유지되는 것을 볼 수 있습니다. 반면에 키-값 캐시의 길이는 각 디코딩 단계마다 하나씩 증가합니다.
> 키-값 캐시를 사용하면 \\( \mathbf{QK}^T \\)가 본질적으로 \\( \mathbf{q}_c\mathbf{K}^T \\)로 줄어드는데, 여기서 \\( \mathbf{q}_c \\)는 현재 전달된 입력 토큰의 쿼리 프로젝션으로, *항상* 단일 벡터입니다.
키-값 캐시를 사용하는 것에는 두 가지 장점이 있습니다:
- 전체 \\( \mathbf{QK}^T \\) 행렬을 계산하는 것과 비교하여 계산 효율성이 크게 향상됩니다. 이는 추론 속도의 증가로 이어집니다.
- 생성된 토큰 수에 따라 필요한 최대 메모리가 이차적으로 증가하지 않고, 선형적으로만 증가합니다.
> 더 긴 입력 시퀀스에 대해 동일한 결과와 큰 속도 향상을 가져오기 때문에 키-값 캐시를 *항상* 사용해야 합니다. Transformers는 텍스트 파이프라인이나 [`generate` 메서드](https://huggingface.co/docs/transformers/main_classes/text_generation)를 사용할 때 기본적으로 키-값 캐시를 활성화합니다.
<Tip warning={true}>
참고로, 키-값 캐시를 사용할 것을 권장하지만, 이를 사용할 때 LLM 출력이 약간 다를 수 있습니다. 이것은 행렬 곱셈 커널 자체의 특성 때문입니다 -- 더 자세한 내용은 [여기](https://github.com/huggingface/transformers/issues/25420#issuecomment-1775317535)에서 읽어볼 수 있습니다.
</Tip>
#### 3.2.1 멀티 라운드 대화 [[321-multi-round-conversation]]
키-값 캐시는 여러 번의 자기회귀 디코딩이 필요한 채팅과 같은 애플리케이션에 특히 유용합니다. 예제를 살펴보겠습니다.
```
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
```
이 채팅에서 대규모 언어 모델은 두 번의 자기회귀 디코딩을 실행합니다:
1. 첫 번째로, 키-값 캐시는 비어 있고 입력 프롬프트는 `"User: How many people live in France?"`입니다. 모델은 자기회귀적으로 `"Roughly 75 million people live in France"`라는 텍스트를 생성하며 디코딩 단계마다 키-값 캐시를 증가시킵니다.
2. 두 번째로, 입력 프롬프트는 `"User: How many people live in France? \n Assistant: Roughly 75 million people live in France \n User: And how many in Germany?"`입니다. 캐시 덕분에 첫 번째 두 문장에 대한 모든 키-값 벡터는 이미 계산되어 있습니다. 따라서 입력 프롬프트는 `"User: And how many in Germany?"`로만 구성됩니다. 줄어든 입력 프롬프트를 처리하는 동안 계산된 키-값 벡터가 첫 번째 디코딩의 키-값 캐시에 연결됩니다. 두 번째 어시스턴트의 답변인 `"Germany has ca. 81 million inhabitants"`는 `"User: How many people live in France? \n Assistant: Roughly 75 million people live in France \n User: And how many are in Germany?"`의 인코딩된 키-값 벡터로 구성된 키-값 캐시를 사용하여 자기회귀적으로 생성됩니다.
여기서 두 가지를 주목해야 합니다:
1. 대규모 언어 모델이 대화의 모든 이전 문맥을 이해할 수 있도록 모든 문맥을 유지하는 것이 채팅에 배포된 대규모 언어 모델에서는 매우 중요합니다. 예를 들어, 위의 예에서 대규모 언어 모델은 사용자가 `"And how many are in Germany"`라고 물을 때 인구를 언급하고 있음을 이해해야 합니다.
2. 키-값 캐시는 채팅에서 매우 유용합니다. 이는 인코딩된 채팅 기록을 처음부터 다시 인코딩할 필요 없이 계속해서 확장할 수 있게 해주기 때문입니다(예: 인코더-디코더 아키텍처를 사용할 때와 같은 경우).
`transformers`에서 `generate` 호출은 기본적으로 `use_cache=True`와 함께 `return_dict_in_generate=True`를 전달하면 `past_key_values`를 반환합니다. 이는 아직 `pipeline` 인터페이스를 통해서는 사용할 수 없습니다.
```python
# 일반적인 생성
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]
# 리턴된 `past_key_values`를 파이프라인화하여 다음 대화 라운드를 가속화
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):]
```
**출력**:
```
is a modified version of the function that returns Mega bytes instead.
def bytes_to_megabytes(bytes):
return bytes / 1024 / 1024
Answer: The function takes a number of bytes as input and returns the number of
```
훌륭합니다. 어텐션 층의 동일한 키와 값을 다시 계산하는 데 추가 시간이 소요되지 않습니다! 그러나 한 가지 문제가 있습니다. \\( \mathbf{QK}^T \\) 행렬에 필요한 최대 메모리는 크게 줄어들지만, 긴 입력 시퀀스나 다회차 채팅의 경우 키-값 캐시를 메모리에 보관하는 것이 매우 메모리 집약적이 될 수 있습니다. 키-값 캐시는 모든 자기 어텐션 층과 모든 어텐션 헤드에 대해 이전 입력 벡터 \\( \mathbf{x}_i \text{, for } i \in {1, \ldots, c - 1} \\)의 키-값 벡터를 저장해야 한다는 점을 기억하세요.
이전에 사용한 대규모 언어 모델 `bigcode/octocoder`에 대해 키-값 캐시에 저장해야 하는 부동 소수점 값의 수를 계산해 봅시다.
부동 소수점 값의 수는 시퀀스 길이의 두 배의 어텐션 헤드 수, 어텐션 헤드 차원, 레이어 수를 곱한 값입니다.
가상의 입력 시퀀스 길이 16000에서 대규모 언어 모델에 대해 이를 계산하면 다음과 같습니다.
```python
config = model.config
2 * 16_000 * config.n_layer * config.n_head * config.n_embd // config.n_head
```
**출력**:
```
7864320000
```
대략 80억 개의 부동 소수점 값입니다! `float16` 정밀도로 80억 개의 부동 소수점 값을 저장하는 데는 약 15GB의 RAM이 필요하며, 이는 모델 가중치 자체의 절반 정도입니다.
연구자들은 키-값 캐시를 저장하는 데 필요한 메모리 비용을 크게 줄일 수 있는 두 가지 방법을 제안했으며, 이는 다음 절에서 살펴보겠습니다.
#### 3.2.2 멀티 쿼리 어텐션 (MQA) [[322-multi-query-attention-mqa]]
[멀티 쿼리 어텐션 (MQA)](https://arxiv.org/abs/1911.02150)은 Noam Shazeer의 *Fast Transformer Decoding: One Write-Head is All You Need* 논문에서 제안되었습니다. 제목에서 알 수 있듯이, Noam은 `n_head` 키-값 프로젝션 가중치 대신, 모든 어텐션 헤드에서 공유되는 단일 헤드-값 프로젝션 가중치를 사용할 수 있으며, 이를 통해 모델 성능이 크게 저하되지 않는다는 것을 발견했습니다.
> 단일 헤드-값 프로젝션 가중치를 사용함으로써, 키-값 벡터 \\( \mathbf{k}_i, \mathbf{v}_i \\)는 모든 어텐션 헤드에서 동일해야 하며, 이는 캐시에 `n_head` 개 대신 하나의 키-값 프로젝션 쌍만 저장하면 된다는 것을 의미합니다.
대부분의 대규모 언어 모델이 20에서 100 사이의 어텐션 헤드를 사용하기 때문에, MQA는 키-값 캐시의 메모리 소비를 크게 줄입니다. 이 노트북에서 사용된 대규모 언어 모델의 경우, 입력 시퀀스 길이 16000에서 필요한 메모리 소비를 15GB에서 400MB 미만으로 줄일 수 있습니다.
메모리 절감 외에도, MQA는 계산 효율성도 향상시킵니다. 다음과 같이 설명합니다.
자기회귀 디코딩에서는 큰 키-값 벡터를 다시 로드하고, 현재 키-값 벡터 쌍과 연결한 후 \\( \mathbf{q}_c\mathbf{K}^T \\) 계산에 매 단계마다 입력해야 합니다. 자기회귀 디코딩의 경우, 지속적인 재로드에 필요한 메모리 대역폭이 심각한 시간 병목 현상을 가져올 수 있습니다. 키-값 벡터의 크기를 줄이면 접근해야 하는 메모리 양이 줄어들어 메모리 대역폭 병목 현상이 감소합니다. 자세한 내용은 [Noam의 논문](https://arxiv.org/abs/1911.02150)을 참조하세요.
여기서 이해해야 할 중요한 부분은 키-값 어텐션 헤드 수를 1로 줄이는 것이 키-값 캐시를 사용할 때만 의미가 있다는 것입니다. 키-값 캐시 없이 단일 포워드 패스에 대한 모델의 최대 메모리 소비는 변경되지 않으며, 각 어텐션 헤드는 여전히 고유한 쿼리 벡터를 가지므로 각 어텐션 헤드는 여전히 다른 \\( \mathbf{QK}^T \\) 행렬을 가집니다.
MQA는 커뮤니티에서 널리 채택되어 현재 가장 인기 있는 많은 대규모 언어 모델에서 사용되고 있습니다.
- [**Falcon**](https://huggingface.co/tiiuae/falcon-40b)
- [**PaLM**](https://arxiv.org/abs/2204.02311)
- [**MPT**](https://huggingface.co/mosaicml/mpt-30b)
- [**BLOOM**](https://huggingface.co/bigscience/bloom)
또한, 이 노트북에서 사용된 체크포인트 `bigcode/octocoder`는 MQA를 사용합니다.
#### 3.2.3 그룹 쿼리 어텐션 (GQA) [[323-grouped-query-attention-gqa]]
[그룹 쿼리 어텐션 (GQA)](https://arxiv.org/abs/2305.13245)은 Google의 Ainslie 등의 연구진들에 의해 제안되었습니다. 그들은 MQA를 사용하는 것이 종종 일반적인 멀티 키-값 헤드 프로젝션을 사용하는 것보다 품질 저하를 가져올 수 있다는 것을 발견했습니다. 이 논문은 쿼리 헤드 프로젝션 가중치의 수를 너무 극단적으로 줄이는 대신, 더 많은 모델 성능을 유지할 수 있다고 주장합니다. 단일 키-값 프로젝션 가중치 대신, `n < n_head` 키-값 프로젝션 가중치를 사용해야 합니다. `n_head`보다 훨씬 작은 `n`값, 예를 들어 2, 4 또는 8을 선택하면, MQA의 거의 모든 메모리 및 속도 이점을 유지하면서 모델 용량을 덜 희생하고 따라서 성능 저하를 줄일 수 있습니다.
또한, GQA의 저자들은 기존 모델 체크포인트를 원래 사전 학습 계산의 5% 정도의 적은 양으로 GQA 아키텍처로 *업트레이닝*할 수 있음을 발견했습니다. 원래 사전 학습 계산의 5%가 여전히 엄청난 양일 수 있지만, GQA *업트레이닝*은 기존 체크포인트가 더 긴 입력 시퀀스에서도 유용하도록 합니다.
GQA는 최근에 제안되었기 때문에 이 노트북을 작성할 당시에는 채택이 덜 되었습니다.
GQA의 가장 주목할 만한 적용 사례는 [Llama-v2](https://huggingface.co/meta-llama/Llama-2-70b-hf)입니다.
> 결론적으로, 대규모 언어 모델이 자기회귀 디코딩으로 배포되면서 채팅과 같이 큰 입력 시퀀스를 가진 작업을 처리해야 하는 경우 GQA 또는 MQA를 사용하는 것이 강력히 권장됩니다.
## 결론 [[conclusion]]
연구 커뮤니티는 점점 더 큰 대규모 언어 모델의 추론 시간을 가속화하기 위한 새로운 기발한 방법들을 끊임없이 찾아내고 있습니다. 예를 들어, [추측 디코딩](https://arxiv.org/abs/2211.17192)이라는 유망한 연구 방향이 있습니다. 여기서 "쉬운 토큰"은 더 작고 빠른 언어 모델에 의해 생성되고, "어려운 토큰"만 대규모 언어 모델 자체에 의해 생성됩니다. 자세한 내용은 이 노트북의 범위를 벗어나지만, [멋진 블로그 포스트](https://huggingface.co/blog/assisted-generation)에서 읽어볼 수 있습니다.
GPT3/4, Llama-2-70b, Claude, PaLM과 같은 거대한 대규모 언어 모델이 [Hugging Face Chat](https://huggingface.co/chat/) 또는 ChatGPT와 같은 채팅 인터페이스에서 빠르게 실행될 수 있는 이유는 위에서 언급한 정밀도, 알고리즘, 아키텍처의 개선 덕분입니다. 앞으로 GPU, TPU 등과 같은 가속기는 점점 더 빨라지고 더 많은 메모리를 사용할 것입니다. 따라서 가장 좋은 알고리즘과 아키텍처를 사용하여 최고의 효율을 얻는 것이 중요합니다 🤗
|
transformers/docs/source/ko/llm_tutorial_optimization.md/0
|
{
"file_path": "transformers/docs/source/ko/llm_tutorial_optimization.md",
"repo_id": "transformers",
"token_count": 42195
}
| 303
|
<!--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://arxiv.org/abs/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://arxiv.org/abs/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://arxiv.org/abs/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/microsoft/DeepSpeed)
- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)
- [Varuna](https://github.com/microsoft/varuna)
- [SageMaker](https://arxiv.org/abs/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/microsoft/DeepSpeed) - DeepSpeed는 더욱 효율적인 DP인 ZeRO-DP라고도 부릅니다.
- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)
- [Varuna](https://github.com/microsoft/varuna)
- [SageMaker](https://arxiv.org/abs/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://arxiv.org/abs/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://arxiv.org/abs/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": 28499
}
| 304
|
<!--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.
-->
# 객관식 문제[[multiple-choice]]
[[open-in-colab]]
객관식 과제는 문맥과 함께 여러 개의 후보 답변이 제공되고 모델이 정답을 선택하도록 학습된다는 점을 제외하면 질의응답과 유사합니다.
진행하는 방법은 아래와 같습니다:
1. [SWAG](https://huggingface.co/datasets/swag) 데이터 세트의 'regular' 구성으로 [BERT](https://huggingface.co/google-bert/bert-base-uncased)를 미세 조정하여 여러 옵션과 일부 컨텍스트가 주어졌을 때 가장 적합한 답을 선택합니다.
2. 추론에 미세 조정된 모델을 사용합니다.
시작하기 전에 필요한 라이브러리가 모두 설치되어 있는지 확인하세요:
```bash
pip install transformers datasets evaluate
```
모델을 업로드하고 커뮤니티와 공유할 수 있도록 허깅페이스 계정에 로그인하는 것이 좋습니다. 메시지가 표시되면 토큰을 입력하여 로그인합니다:
```py
>>> from huggingface_hub import notebook_login
>>> notebook_login()
```
## SWAG 데이터 세트 가져오기[[load-swag-dataset]]
먼저 🤗 Datasets 라이브러리에서 SWAG 데이터셋의 '일반' 구성을 가져옵니다:
```py
>>> from datasets import load_dataset
>>> swag = load_dataset("swag", "regular")
```
이제 데이터를 살펴봅니다:
```py
>>> swag["train"][0]
{'ending0': 'passes by walking down the street playing their instruments.',
'ending1': 'has heard approaching them.',
'ending2': "arrives and they're outside dancing and asleep.",
'ending3': 'turns the lead singer watches the performance.',
'fold-ind': '3416',
'gold-source': 'gold',
'label': 0,
'sent1': 'Members of the procession walk down the street holding small horn brass instruments.',
'sent2': 'A drum line',
'startphrase': 'Members of the procession walk down the street holding small horn brass instruments. A drum line',
'video-id': 'anetv_jkn6uvmqwh4'}
```
여기에는 많은 필드가 있는 것처럼 보이지만 실제로는 매우 간단합니다:
- `sent1` 및 `sent2`: 이 필드는 문장이 어떻게 시작되는지 보여주며, 이 두 필드를 합치면 `시작 구절(startphrase)` 필드가 됩니다.
- `종료 구절(ending)`: 문장이 어떻게 끝날 수 있는지에 대한 가능한 종료 구절를 제시하지만 그 중 하나만 정답입니다.
- `레이블(label)`: 올바른 문장 종료 구절을 식별합니다.
## 전처리[[preprocess]]
다음 단계는 문장의 시작과 네 가지 가능한 구절을 처리하기 위해 BERT 토크나이저를 불러옵니다:
```py
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-uncased")
```
생성하려는 전처리 함수는 다음과 같아야 합니다:
1. `sent1` 필드를 네 개 복사한 다음 각각을 `sent2`와 결합하여 문장이 시작되는 방식을 재현합니다.
2. `sent2`를 네 가지 가능한 문장 구절 각각과 결합합니다.
3. 이 두 목록을 토큰화할 수 있도록 평탄화(flatten)하고, 각 예제에 해당하는 `input_ids`, `attention_mask` 및 `labels` 필드를 갖도록 다차원화(unflatten) 합니다.
```py
>>> ending_names = ["ending0", "ending1", "ending2", "ending3"]
>>> def preprocess_function(examples):
... first_sentences = [[context] * 4 for context in examples["sent1"]]
... question_headers = examples["sent2"]
... second_sentences = [
... [f"{header} {examples[end][i]}" for end in ending_names] for i, header in enumerate(question_headers)
... ]
... first_sentences = sum(first_sentences, [])
... second_sentences = sum(second_sentences, [])
... tokenized_examples = tokenizer(first_sentences, second_sentences, truncation=True)
... return {k: [v[i : i + 4] for i in range(0, len(v), 4)] for k, v in tokenized_examples.items()}
```
전체 데이터 집합에 전처리 기능을 적용하려면 🤗 Datasets [`~datasets.Dataset.map`] 메소드를 사용합니다. `batched=True`를 설정하여 데이터 집합의 여러 요소를 한 번에 처리하면 `map` 함수의 속도를 높일 수 있습니다:
```py
tokenized_swag = swag.map(preprocess_function, batched=True)
```
🤗 Transformers에는 객관식용 데이터 콜레이터가 없으므로 예제 배치를 만들려면 [`DataCollatorWithPadding`]을 조정해야 합니다. 데이터 정렬 중에 전체 데이터 집합을 최대 길이로 패딩하는 대신 배치 중 가장 긴 길이로 문장을 *동적 패딩*하는 것이 더 효율적입니다.
`DataCollatorForMultipleChoice`는 모든 모델 입력을 평탄화하고 패딩을 적용하며 그 결과를 결과를 다차원화합니다:
<frameworkcontent>
<pt>
```py
>>> from dataclasses import dataclass
>>> from transformers.tokenization_utils_base import PreTrainedTokenizerBase, PaddingStrategy
>>> from typing import Optional, Union
>>> import torch
>>> @dataclass
... class DataCollatorForMultipleChoice:
... """
... Data collator that will dynamically pad the inputs for multiple choice received.
... """
... tokenizer: PreTrainedTokenizerBase
... padding: Union[bool, str, PaddingStrategy] = True
... max_length: Optional[int] = None
... pad_to_multiple_of: Optional[int] = None
... def __call__(self, features):
... label_name = "label" if "label" in features[0].keys() 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, [])
... 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="pt",
... )
... 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
```
</pt>
<tf>
```py
>>> from dataclasses import dataclass
>>> from transformers.tokenization_utils_base import PreTrainedTokenizerBase, PaddingStrategy
>>> from typing import Optional, Union
>>> import tensorflow as tf
>>> @dataclass
... class DataCollatorForMultipleChoice:
... """
... Data collator that will dynamically pad the inputs for multiple choice received.
... """
... tokenizer: PreTrainedTokenizerBase
... padding: Union[bool, str, PaddingStrategy] = True
... max_length: Optional[int] = None
... pad_to_multiple_of: Optional[int] = None
... def __call__(self, features):
... label_name = "label" if "label" in features[0].keys() 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, [])
... 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
```
</tf>
</frameworkcontent>
## 평가 하기[[evaluate]]
훈련 중에 메트릭을 포함하면 모델의 성능을 평가하는 데 도움이 되는 경우가 많습니다. 🤗[Evaluate](https://huggingface.co/docs/evaluate/index) 라이브러리를 사용하여 평가 방법을 빠르게 가져올 수 있습니다. 이 작업에서는 [accuracy](https://huggingface.co/spaces/evaluate-metric/accuracy) 지표를 가져옵니다(🤗 Evaluate [둘러보기](https://huggingface.co/docs/evaluate/a_quick_tour)를 참조하여 지표를 가져오고 계산하는 방법에 대해 자세히 알아보세요):
```py
>>> import evaluate
>>> accuracy = evaluate.load("accuracy")
```
그리고 예측과 레이블을 [`~evaluate.EvaluationModule.compute`]에 전달하여 정확도를 계산하는 함수를 만듭니다:
```py
>>> import numpy as np
>>> def compute_metrics(eval_pred):
... predictions, labels = eval_pred
... predictions = np.argmax(predictions, axis=1)
... return accuracy.compute(predictions=predictions, references=labels)
```
이제 `compute_metrics` 함수를 사용할 준비가 되었으며, 훈련을 설정할 때 이 함수로 돌아가게 됩니다.
## 훈련 하기[[train]]
<frameworkcontent>
<pt>
<Tip>
[`Trainer`]로 모델을 미세 조정하는 데 익숙하지 않다면 기본 튜토리얼 [여기](../training#train-with-pytorch-trainer)를 살펴보세요!
</Tip>
이제 모델 훈련을 시작할 준비가 되었습니다! [`AutoModelForMultipleChoice`]로 BERT를 로드합니다:
```py
>>> from transformers import AutoModelForMultipleChoice, TrainingArguments, Trainer
>>> model = AutoModelForMultipleChoice.from_pretrained("google-bert/bert-base-uncased")
```
이제 세 단계만 남았습니다:
1. 훈련 하이퍼파라미터를 [`TrainingArguments`]에 정의합니다. 유일한 필수 매개변수는 모델을 저장할 위치를 지정하는 `output_dir`입니다. `push_to_hub=True`를 설정하여 이 모델을 허브에 푸시합니다(모델을 업로드하려면 허깅 페이스에 로그인해야 합니다). 각 에폭이 끝날 때마다 [`Trainer`]가 정확도를 평가하고 훈련 체크포인트를 저장합니다.
2. 모델, 데이터 세트, 토크나이저, 데이터 콜레이터, `compute_metrics` 함수와 함께 훈련 인자를 [`Trainer`]에 전달합니다.
3. [`~Trainer.train`]을 사용하여 모델을 미세 조정합니다.
```py
>>> training_args = TrainingArguments(
... output_dir="my_awesome_swag_model",
... eval_strategy="epoch",
... save_strategy="epoch",
... load_best_model_at_end=True,
... learning_rate=5e-5,
... per_device_train_batch_size=16,
... per_device_eval_batch_size=16,
... num_train_epochs=3,
... weight_decay=0.01,
... push_to_hub=True,
... )
>>> trainer = Trainer(
... model=model,
... args=training_args,
... train_dataset=tokenized_swag["train"],
... eval_dataset=tokenized_swag["validation"],
... tokenizer=tokenizer,
... data_collator=DataCollatorForMultipleChoice(tokenizer=tokenizer),
... compute_metrics=compute_metrics,
... )
>>> trainer.train()
```
훈련이 완료되면 모든 사람이 모델을 사용할 수 있도록 [`~transformers.Trainer.push_to_hub`] 메소드를 사용하여 모델을 허브에 공유하세요:
```py
>>> trainer.push_to_hub()
```
</pt>
<tf>
<Tip>
Keras로 모델을 미세 조정하는 데 익숙하지 않다면 기본 튜토리얼 [여기](../training#train-a-tensorflow-model-with-keras)를 살펴보시기 바랍니다!
</Tip>
TensorFlow에서 모델을 미세 조정하려면 최적화 함수, 학습률 스케쥴 및 몇 가지 학습 하이퍼파라미터를 설정하는 것부터 시작하세요:
```py
>>> from transformers import create_optimizer
>>> batch_size = 16
>>> num_train_epochs = 2
>>> total_train_steps = (len(tokenized_swag["train"]) // batch_size) * num_train_epochs
>>> optimizer, schedule = create_optimizer(init_lr=5e-5, num_warmup_steps=0, num_train_steps=total_train_steps)
```
그리고 [`TFAutoModelForMultipleChoice`]로 BERT를 가져올 수 있습니다:
```py
>>> from transformers import TFAutoModelForMultipleChoice
>>> model = TFAutoModelForMultipleChoice.from_pretrained("google-bert/bert-base-uncased")
```
[`~transformers.TFPreTrainedModel.prepare_tf_dataset`]을 사용하여 데이터 세트를 `tf.data.Dataset` 형식으로 변환합니다:
```py
>>> data_collator = DataCollatorForMultipleChoice(tokenizer=tokenizer)
>>> tf_train_set = model.prepare_tf_dataset(
... tokenized_swag["train"],
... shuffle=True,
... batch_size=batch_size,
... collate_fn=data_collator,
... )
>>> tf_validation_set = model.prepare_tf_dataset(
... tokenized_swag["validation"],
... shuffle=False,
... batch_size=batch_size,
... collate_fn=data_collator,
... )
```
[`compile`](https://keras.io/api/models/model_training_apis/#compile-method)을 사용하여 훈련 모델을 구성합니다:
```py
>>> model.compile(optimizer=optimizer)
```
훈련을 시작하기 전에 설정해야 할 마지막 두 가지는 예측의 정확도를 계산하고 모델을 허브로 푸시하는 방법을 제공하는 것입니다. 이 두 가지 작업은 모두 [Keras 콜백](../main_classes/keras_callbacks)을 사용하여 수행할 수 있습니다.
`compute_metrics`함수를 [`~transformers.KerasMetricCallback`]에 전달하세요:
```py
>>> from transformers.keras_callbacks import KerasMetricCallback
>>> metric_callback = KerasMetricCallback(metric_fn=compute_metrics, eval_dataset=tf_validation_set)
```
모델과 토크나이저를 업로드할 위치를 [`~transformers.PushToHubCallback`]에서 지정하세요:
```py
>>> from transformers.keras_callbacks import PushToHubCallback
>>> push_to_hub_callback = PushToHubCallback(
... output_dir="my_awesome_model",
... tokenizer=tokenizer,
... )
```
그리고 콜백을 함께 묶습니다:
```py
>>> callbacks = [metric_callback, push_to_hub_callback]
```
이제 모델 훈련을 시작합니다! 훈련 및 검증 데이터 세트, 에폭 수, 콜백을 사용하여 [`fit`](https://keras.io/api/models/model_training_apis/#fit-method)을 호출하고 모델을 미세 조정합니다:
```py
>>> model.fit(x=tf_train_set, validation_data=tf_validation_set, epochs=2, callbacks=callbacks)
```
훈련이 완료되면 모델이 자동으로 허브에 업로드되어 누구나 사용할 수 있습니다!
</tf>
</frameworkcontent>
<Tip>
객관식 모델을 미세 조정하는 방법에 대한 보다 심층적인 예는 아래 문서를 참조하세요.
[PyTorch notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/multiple_choice.ipynb)
또는 [TensorFlow notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/multiple_choice-tf.ipynb).
</Tip>
## 추론 하기[[inference]]
이제 모델을 미세 조정했으니 추론에 사용할 수 있습니다!
텍스트와 두 개의 후보 답안을 작성합니다:
```py
>>> prompt = "France has a bread law, Le Décret Pain, with strict rules on what is allowed in a traditional baguette."
>>> candidate1 = "The law does not apply to croissants and brioche."
>>> candidate2 = "The law applies to baguettes."
```
<frameworkcontent>
<pt>
각 프롬프트와 후보 답변 쌍을 토큰화하여 PyTorch 텐서를 반환합니다. 또한 `labels`을 생성해야 합니다:
```py
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("my_awesome_swag_model")
>>> inputs = tokenizer([[prompt, candidate1], [prompt, candidate2]], return_tensors="pt", padding=True)
>>> labels = torch.tensor(0).unsqueeze(0)
```
입력과 레이블을 모델에 전달하고 `logits`을 반환합니다:
```py
>>> from transformers import AutoModelForMultipleChoice
>>> model = AutoModelForMultipleChoice.from_pretrained("my_awesome_swag_model")
>>> outputs = model(**{k: v.unsqueeze(0) for k, v in inputs.items()}, labels=labels)
>>> logits = outputs.logits
```
가장 높은 확률을 가진 클래스를 가져옵니다:
```py
>>> predicted_class = logits.argmax().item()
>>> predicted_class
'0'
```
</pt>
<tf>
각 프롬프트와 후보 답안 쌍을 토큰화하여 텐서플로 텐서를 반환합니다:
```py
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("my_awesome_swag_model")
>>> inputs = tokenizer([[prompt, candidate1], [prompt, candidate2]], return_tensors="tf", padding=True)
```
모델에 입력을 전달하고 `logits`를 반환합니다:
```py
>>> from transformers import TFAutoModelForMultipleChoice
>>> model = TFAutoModelForMultipleChoice.from_pretrained("my_awesome_swag_model")
>>> inputs = {k: tf.expand_dims(v, 0) for k, v in inputs.items()}
>>> outputs = model(inputs)
>>> logits = outputs.logits
```
가장 높은 확률을 가진 클래스를 가져옵니다:
```py
>>> predicted_class = int(tf.math.argmax(logits, axis=-1)[0])
>>> predicted_class
'0'
```
</tf>
</frameworkcontent>
|
transformers/docs/source/ko/tasks/multiple_choice.md/0
|
{
"file_path": "transformers/docs/source/ko/tasks/multiple_choice.md",
"repo_id": "transformers",
"token_count": 9539
}
| 305
|
<!--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.
-->
# TFLite로 내보내기[[export-to-tflite]]
[TensorFlow Lite](https://www.tensorflow.org/lite/guide)는 자원이 제한된 휴대폰, 임베디드 시스템, 사물인터넷(IoT) 기기에서
기계학습 모델을 배포하기 위한 경량 프레임워크입니다.
TFLite는 연산 능력, 메모리, 전력 소비가 제한된 기기에서 모델을 효율적으로 최적화하고 실행하기 위해
설계되었습니다.
TensorFlow Lite 모델은 `.tflite` 파일 확장자로 식별되는 특수하고 효율적인 휴대용 포맷으로 표현됩니다.
🤗 Optimum은 `exporters.tflite` 모듈로 🤗 Transformers 모델을 TFLite로 내보내는 기능을 제공합니다.
지원되는 모델 아키텍처 목록은 [🤗 Optimum 문서](https://huggingface.co/docs/optimum/exporters/tflite/overview)를 참고하세요.
모델을 TFLite로 내보내려면, 필요한 종속성을 설치하세요:
```bash
pip install optimum[exporters-tf]
```
모든 사용 가능한 인수를 확인하려면, [🤗 Optimum 문서](https://huggingface.co/docs/optimum/main/en/exporters/tflite/usage_guides/export_a_model)를 참고하거나
터미널에서 도움말을 살펴보세요:
```bash
optimum-cli export tflite --help
```
예를 들어 🤗 Hub에서의 `google-bert/bert-base-uncased` 모델 체크포인트를 내보내려면, 다음 명령을 실행하세요:
```bash
optimum-cli export tflite --model google-bert/bert-base-uncased --sequence_length 128 bert_tflite/
```
다음과 같이 진행 상황을 나타내는 로그와 결과물인 `model.tflite`가 저장된 위치를 보여주는 로그가 표시됩니다:
```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
```
위 예제는 🤗 Hub에서의 체크포인트를 내보내는 방법을 보여줍니다.
로컬 모델을 내보낸다면, 먼저 모델 가중치와 토크나이저 파일이 모두 같은 디렉터리( `local_path` )에 저장됐는지 확인하세요.
CLI를 사용할 때, 🤗 Hub에서의 체크포인트 이름 대신 `model` 인수에 `local_path`를 전달하면 됩니다.
|
transformers/docs/source/ko/tflite.md/0
|
{
"file_path": "transformers/docs/source/ko/tflite.md",
"repo_id": "transformers",
"token_count": 1852
}
| 306
|
<!--版权2023年HuggingFace团队保留所有权利。
根据Apache许可证第2.0版（“许可证”）许可；除非符合许可证，否则您不得使用此文件。您可以在以下网址获取许可证的副本：
http://www.apache.org/licenses/LICENSE-2.0
除非适用法律要求或书面同意，否则按“按原样”分发的软件，无论是明示还是暗示的，都没有任何担保或条件。请参阅许可证以了解特定语言下的权限和限制。
⚠️ 请注意，本文件虽然使用Markdown编写，但包含了特定的语法，适用于我们的doc-builder（类似于MDX），可能无法在您的Markdown查看器中正常渲染。
-->
# 🤗 加速分布式训练
随着模型变得越来越大，并行性已经成为在有限硬件上训练更大模型和加速训练速度的策略，增加了数个数量级。在Hugging Face，我们创建了[🤗 加速](https://huggingface.co/docs/accelerate)库，以帮助用户在任何类型的分布式设置上轻松训练🤗 Transformers模型，无论是在一台机器上的多个GPU还是在多个机器上的多个GPU。在本教程中，了解如何自定义您的原生PyTorch训练循环，以启用分布式环境中的训练。
## 设置
通过安装🤗 加速开始:
```bash
pip install accelerate
```
然后导入并创建[`~accelerate.Accelerator`]对象。[`~accelerate.Accelerator`]将自动检测您的分布式设置类型，并初始化所有必要的训练组件。您不需要显式地将模型放在设备上。
```py
>>> from accelerate import Accelerator
>>> accelerator = Accelerator()
```
## 准备加速
下一步是将所有相关的训练对象传递给[`~accelerate.Accelerator.prepare`]方法。这包括您的训练和评估DataLoader、一个模型和一个优化器:
```py
>>> train_dataloader, eval_dataloader, model, optimizer = accelerator.prepare(
... train_dataloader, eval_dataloader, model, optimizer
... )
```
## 反向传播
最后一步是用🤗 加速的[`~accelerate.Accelerator.backward`]方法替换训练循环中的典型`loss.backward()`:
```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)
```
如您在下面的代码中所见，您只需要添加四行额外的代码到您的训练循环中即可启用分布式训练！
```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)
```
## 训练
在添加了相关代码行后，可以在脚本或笔记本（如Colaboratory）中启动训练。
### 用脚本训练
如果您从脚本中运行训练，请运行以下命令以创建和保存配置文件:
```bash
accelerate config
```
然后使用以下命令启动训练:
```bash
accelerate launch train.py
```
### 用笔记本训练
🤗 加速还可以在笔记本中运行，如果您计划使用Colaboratory的TPU，则可在其中运行。将负责训练的所有代码包装在一个函数中，并将其传递给[`~accelerate.notebook_launcher`]:
```py
>>> from accelerate import notebook_launcher
>>> notebook_launcher(training_function)
```
有关🤗 加速及其丰富功能的更多信息，请参阅[文档](https://huggingface.co/docs/accelerate)。
|
transformers/docs/source/zh/accelerate.md/0
|
{
"file_path": "transformers/docs/source/zh/accelerate.md",
"repo_id": "transformers",
"token_count": 2552
}
| 307
|
<!--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.
-->
# 用于生成的工具
此页面列出了所有由 [`~generation.GenerationMixin.generate`]。
## 生成输出
[`~generation.GenerationMixin.generate`] 的输出是 [`~utils.ModelOutput`] 的一个子类的实例。这个输出是一种包含 [`~generation.GenerationMixin.generate`] 返回的所有信息数据结构，但也可以作为元组或字典使用。
这里是一个例子：
```python
from transformers import GPT2Tokenizer, GPT2LMHeadModel
tokenizer = GPT2Tokenizer.from_pretrained("openai-community/gpt2")
model = GPT2LMHeadModel.from_pretrained("openai-community/gpt2")
inputs = tokenizer("Hello, my dog is cute and ", return_tensors="pt")
generation_output = model.generate(**inputs, return_dict_in_generate=True, output_scores=True)
```
`generation_output` 的对象是 [`~generation.GenerateDecoderOnlyOutput`] 的一个实例，从该类的文档中我们可以看到，这意味着它具有以下属性：
- `sequences`: 生成的tokens序列
- `scores`（可选）: 每个生成步骤的语言建模头的预测分数
- `hidden_states`（可选）: 每个生成步骤模型的hidden states
- `attentions`（可选）: 每个生成步骤模型的注意力权重
在这里，由于我们传递了 `output_scores=True`，我们具有 `scores` 属性。但我们没有 `hidden_states` 和 `attentions`，因为没有传递 `output_hidden_states=True` 或 `output_attentions=True`。
您可以像通常一样访问每个属性，如果该属性未被模型返回，则将获得 `None`。例如，在这里 `generation_output.scores` 是语言建模头的所有生成预测分数，而 `generation_output.attentions` 为 `None`。
当我们将 `generation_output` 对象用作元组时，它只保留非 `None` 值的属性。例如，在这里它有两个元素，`loss` 然后是 `logits`，所以
```python
generation_output[:2]
```
将返回元组`(generation_output.sequences, generation_output.scores)`。
当我们将`generation_output`对象用作字典时，它只保留非`None`的属性。例如，它有两个键，分别是`sequences`和`scores`。
我们在此记录所有输出类型。
### PyTorch
[[autodoc]] generation.GenerateDecoderOnlyOutput
[[autodoc]] generation.GenerateEncoderDecoderOutput
[[autodoc]] generation.GenerateBeamDecoderOnlyOutput
[[autodoc]] generation.GenerateBeamEncoderDecoderOutput
### TensorFlow
[[autodoc]] generation.TFGreedySearchEncoderDecoderOutput
[[autodoc]] generation.TFGreedySearchDecoderOnlyOutput
[[autodoc]] generation.TFSampleEncoderDecoderOutput
[[autodoc]] generation.TFSampleDecoderOnlyOutput
[[autodoc]] generation.TFBeamSearchEncoderDecoderOutput
[[autodoc]] generation.TFBeamSearchDecoderOnlyOutput
[[autodoc]] generation.TFBeamSampleEncoderDecoderOutput
[[autodoc]] generation.TFBeamSampleDecoderOnlyOutput
[[autodoc]] generation.TFContrastiveSearchEncoderDecoderOutput
[[autodoc]] generation.TFContrastiveSearchDecoderOnlyOutput
### FLAX
[[autodoc]] generation.FlaxSampleOutput
[[autodoc]] generation.FlaxGreedySearchOutput
[[autodoc]] generation.FlaxBeamSearchOutput
## LogitsProcessor
[`LogitsProcessor`] 可以用于修改语言模型头的预测分数以进行生成
### PyTorch
[[autodoc]] AlternatingCodebooksLogitsProcessor
- __call__
[[autodoc]] ClassifierFreeGuidanceLogitsProcessor
- __call__
[[autodoc]] EncoderNoRepeatNGramLogitsProcessor
- __call__
[[autodoc]] EncoderRepetitionPenaltyLogitsProcessor
- __call__
[[autodoc]] EpsilonLogitsWarper
- __call__
[[autodoc]] EtaLogitsWarper
- __call__
[[autodoc]] ExponentialDecayLengthPenalty
- __call__
[[autodoc]] ForcedBOSTokenLogitsProcessor
- __call__
[[autodoc]] ForcedEOSTokenLogitsProcessor
- __call__
[[autodoc]] HammingDiversityLogitsProcessor
- __call__
[[autodoc]] InfNanRemoveLogitsProcessor
- __call__
[[autodoc]] LogitNormalization
- __call__
[[autodoc]] LogitsProcessor
- __call__
[[autodoc]] LogitsProcessorList
- __call__
[[autodoc]] MinLengthLogitsProcessor
- __call__
[[autodoc]] MinNewTokensLengthLogitsProcessor
- __call__
[[autodoc]] NoBadWordsLogitsProcessor
- __call__
[[autodoc]] NoRepeatNGramLogitsProcessor
- __call__
[[autodoc]] PrefixConstrainedLogitsProcessor
- __call__
[[autodoc]] RepetitionPenaltyLogitsProcessor
- __call__
[[autodoc]] SequenceBiasLogitsProcessor
- __call__
[[autodoc]] SuppressTokensAtBeginLogitsProcessor
- __call__
[[autodoc]] SuppressTokensLogitsProcessor
- __call__
[[autodoc]] TemperatureLogitsWarper
- __call__
[[autodoc]] TopKLogitsWarper
- __call__
[[autodoc]] TopPLogitsWarper
- __call__
[[autodoc]] TypicalLogitsWarper
- __call__
[[autodoc]] UnbatchedClassifierFreeGuidanceLogitsProcessor
- __call__
[[autodoc]] WhisperTimeStampLogitsProcessor
- __call__
### TensorFlow
[[autodoc]] TFForcedBOSTokenLogitsProcessor
- __call__
[[autodoc]] TFForcedEOSTokenLogitsProcessor
- __call__
[[autodoc]] TFForceTokensLogitsProcessor
- __call__
[[autodoc]] TFLogitsProcessor
- __call__
[[autodoc]] TFLogitsProcessorList
- __call__
[[autodoc]] TFLogitsWarper
- __call__
[[autodoc]] TFMinLengthLogitsProcessor
- __call__
[[autodoc]] TFNoBadWordsLogitsProcessor
- __call__
[[autodoc]] TFNoRepeatNGramLogitsProcessor
- __call__
[[autodoc]] TFRepetitionPenaltyLogitsProcessor
- __call__
[[autodoc]] TFSuppressTokensAtBeginLogitsProcessor
- __call__
[[autodoc]] TFSuppressTokensLogitsProcessor
- __call__
[[autodoc]] TFTemperatureLogitsWarper
- __call__
[[autodoc]] TFTopKLogitsWarper
- __call__
[[autodoc]] TFTopPLogitsWarper
- __call__
### FLAX
[[autodoc]] FlaxForcedBOSTokenLogitsProcessor
- __call__
[[autodoc]] FlaxForcedEOSTokenLogitsProcessor
- __call__
[[autodoc]] FlaxForceTokensLogitsProcessor
- __call__
[[autodoc]] FlaxLogitsProcessor
- __call__
[[autodoc]] FlaxLogitsProcessorList
- __call__
[[autodoc]] FlaxLogitsWarper
- __call__
[[autodoc]] FlaxMinLengthLogitsProcessor
- __call__
[[autodoc]] FlaxSuppressTokensAtBeginLogitsProcessor
- __call__
[[autodoc]] FlaxSuppressTokensLogitsProcessor
- __call__
[[autodoc]] FlaxTemperatureLogitsWarper
- __call__
[[autodoc]] FlaxTopKLogitsWarper
- __call__
[[autodoc]] FlaxTopPLogitsWarper
- __call__
[[autodoc]] FlaxWhisperTimeStampLogitsProcessor
- __call__
## StoppingCriteria
可以使用[`StoppingCriteria`]来更改停止生成的时间（除了EOS token以外的方法）。请注意，这仅适用于我们的PyTorch实现。
[[autodoc]] StoppingCriteria
- __call__
[[autodoc]] StoppingCriteriaList
- __call__
[[autodoc]] MaxLengthCriteria
- __call__
[[autodoc]] MaxTimeCriteria
- __call__
## Constraints
可以使用[`Constraint`]来强制生成结果包含输出中的特定tokens或序列。请注意，这仅适用于我们的PyTorch实现。
[[autodoc]] Constraint
[[autodoc]] PhrasalConstraint
[[autodoc]] DisjunctiveConstraint
[[autodoc]] ConstraintListState
## BeamSearch
[[autodoc]] BeamScorer
- process
- finalize
[[autodoc]] BeamSearchScorer
- process
- finalize
[[autodoc]] ConstrainedBeamSearchScorer
- process
- finalize
## Streamers
[[autodoc]] TextStreamer
[[autodoc]] TextIteratorStreamer
|
transformers/docs/source/zh/internal/generation_utils.md/0
|
{
"file_path": "transformers/docs/source/zh/internal/generation_utils.md",
"repo_id": "transformers",
"token_count": 3447
}
| 308
|
<!--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.
-->
# Logging
🤗 Transformers拥有一个集中式的日志系统，因此您可以轻松设置库输出的日志详细程度。
当前库的默认日志详细程度为`WARNING`。
要更改日志详细程度，只需使用其中一个直接的setter。例如，以下是如何将日志详细程度更改为INFO级别的方法：
```python
import transformers
transformers.logging.set_verbosity_info()
```
您还可以使用环境变量`TRANSFORMERS_VERBOSITY`来覆盖默认的日志详细程度。您可以将其设置为以下级别之一：`debug`、`info`、`warning`、`error`、`critical`。例如：
```bash
TRANSFORMERS_VERBOSITY=error ./myprogram.py
```
此外，通过将环境变量`TRANSFORMERS_NO_ADVISORY_WARNINGS`设置为`true`（如*1*），可以禁用一些`warnings`。这将禁用[`logger.warning_advice`]记录的任何警告。例如：
```bash
TRANSFORMERS_NO_ADVISORY_WARNINGS=1 ./myprogram.py
```
以下是如何在您自己的模块或脚本中使用与库相同的logger的示例：
```python
from transformers.utils import logging
logging.set_verbosity_info()
logger = logging.get_logger("transformers")
logger.info("INFO")
logger.warning("WARN")
```
此日志模块的所有方法都在下面进行了记录，主要的方法包括 [`logging.get_verbosity`] 用于获取logger当前输出日志详细程度的级别和 [`logging.set_verbosity`] 用于将详细程度设置为您选择的级别。按照顺序（从最不详细到最详细），这些级别（及其相应的整数值）为：
- `transformers.logging.CRITICAL` 或 `transformers.logging.FATAL`（整数值，50）：仅报告最关键的errors。
- `transformers.logging.ERROR`（整数值，40）：仅报告errors。
- `transformers.logging.WARNING` 或 `transformers.logging.WARN`（整数值，30）：仅报告error和warnings。这是库使用的默认级别。
- `transformers.logging.INFO`（整数值，20）：报告error、warnings和基本信息。
- `transformers.logging.DEBUG`（整数值，10）：报告所有信息。
默认情况下，将在模型下载期间显示`tqdm`进度条。[`logging.disable_progress_bar`] 和 [`logging.enable_progress_bar`] 可用于禁止或启用此行为。
## `logging` vs `warnings`
Python有两个经常一起使用的日志系统：如上所述的`logging`，和对特定buckets中的警告进行进一步分类的`warnings`，例如，`FutureWarning`用于输出已经被弃用的功能或路径，`DeprecationWarning`用于指示即将被弃用的内容。
我们在`transformers`库中同时使用这两个系统。我们利用并调整了`logging`的`captureWarning`方法，以便通过上面的详细程度setters来管理这些警告消息。
对于库的开发人员，这意味着什么呢？我们应该遵循以下启发法则：
- 库的开发人员和依赖于`transformers`的库应优先使用`warnings`
- `logging`应该用于在日常项目中经常使用它的用户
以下是`captureWarnings`方法的参考。
[[autodoc]] logging.captureWarnings
## Base setters
[[autodoc]] logging.set_verbosity_error
[[autodoc]] logging.set_verbosity_warning
[[autodoc]] logging.set_verbosity_info
[[autodoc]] logging.set_verbosity_debug
## Other functions
[[autodoc]] logging.get_verbosity
[[autodoc]] logging.set_verbosity
[[autodoc]] logging.get_logger
[[autodoc]] logging.enable_default_handler
[[autodoc]] logging.disable_default_handler
[[autodoc]] logging.enable_explicit_format
[[autodoc]] logging.reset_format
[[autodoc]] logging.enable_progress_bar
[[autodoc]] logging.disable_progress_bar
|
transformers/docs/source/zh/main_classes/logging.md/0
|
{
"file_path": "transformers/docs/source/zh/main_classes/logging.md",
"repo_id": "transformers",
"token_count": 2154
}
| 309
|
<!---
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.
-->
# 性能与可扩展性
训练大型transformer模型并将其部署到生产环境会面临各种挑战。
在训练过程中，模型可能需要比可用的GPU内存更多的资源，或者表现出较慢的训练速度。在部署阶段，模型可能在生产环境中难以处理所需的吞吐量。
本文档旨在帮助您克服这些挑战，并找到适合您使用场景的最佳设置。教程分为训练和推理部分，因为每个部分都有不同的挑战和解决方案。在每个部分中，您将找到针对不同硬件配置的单独指南，例如单GPU与多GPU用于训练或CPU与GPU用于推理。
将此文档作为您的起点，进一步导航到与您的情况匹配的方法。
## 训练
高效训练大型transformer模型需要使用加速器硬件，如GPU或TPU。最常见的情况是您只有一个GPU。您应用于单个GPU上提高训练效率的方法可以扩展到其他设置，如多个GPU。然而，也有一些特定于多GPU或CPU训练的技术。我们在单独的部分中介绍它们。
* [在单个GPU上进行高效训练的方法和工具](perf_train_gpu_one)：从这里开始学习常见的方法，可以帮助优化GPU内存利用率、加快训练速度或两者兼备。
* [多GPU训练部分](perf_train_gpu_many)：探索此部分以了解适用于多GPU设置的进一步优化方法，例如数据并行、张量并行和流水线并行。
* [CPU训练部分](perf_train_cpu)：了解在CPU上的混合精度训练。
* [在多个CPU上进行高效训练](perf_train_cpu_many)：了解分布式CPU训练。
* [使用TensorFlow在TPU上进行训练](perf_train_tpu_tf)：如果您对TPU还不熟悉，请参考此部分，了解有关在TPU上进行训练和使用XLA的建议性介绍。
* [自定义硬件进行训练](perf_hardware)：在构建自己的深度学习机器时查找技巧和窍门。
* [使用Trainer API进行超参数搜索](hpo_train)
## 推理
在生产环境中对大型模型进行高效推理可能与训练它们一样具有挑战性。在接下来的部分中，我们将详细介绍如何在CPU和单/多GPU设置上进行推理的步骤。
* [在单个CPU上进行推理](perf_infer_cpu)
* [在单个GPU上进行推理](perf_infer_gpu_one)
* [多GPU推理](perf_infer_gpu_one)
* [TensorFlow模型的XLA集成](tf_xla)
## 训练和推理
在这里，您将找到适用于训练模型或使用它进行推理的技巧、窍门和技巧。
* [实例化大型模型](big_models)
* [解决性能问题](debugging)
## 贡献
这份文档还远远没有完成，还有很多需要添加的内容，所以如果你有补充或更正的内容，请毫不犹豫地提交一个PR（Pull Request），或者如果你不确定，可以创建一个Issue，我们可以在那里讨论细节。
在做出贡献时，如果A比B更好，请尽量包含可重复的基准测试和(或)该信息来源的链接（除非它直接来自您）。
|
transformers/docs/source/zh/performance.md/0
|
{
"file_path": "transformers/docs/source/zh/performance.md",
"repo_id": "transformers",
"token_count": 2220
}
| 310
|
# Using the `diff_converter` linter
`pip install libcst` is a must!
# `sh examples/diff-conversion/convert_examples.sh` to get the converted outputs
The diff converter is a new `linter` specific to `transformers`. It allows us to unpack inheritance in python to convert a modular `diff` file like `diff_gemma.py` into a `single model single file`.
Examples of possible usage are available in the `examples/diff-conversion`, or `diff_gemma` for a full model usage.
`python utils/diff_model_converter.py --files_to_parse "/Users/arthurzucker/Work/transformers/examples/diff-conversion/diff_my_new_model2.py"`
## How it works
We use the `libcst` parser to produce an AST representation of the `diff_xxx.py` file. For any imports that are made from `transformers.models.modeling_xxxx` we parse the source code of that module, and build a class dependency mapping, which allows us to unpack the difference dependencies.
The code from the `diff` file and the class dependency mapping are "merged" to produce the single model single file.
We use ruff to automatically remove the potential duplicate imports.
## Why we use libcst instead of the native AST?
AST is super powerful, but it does not keep the `docstring`, `comment` or code formatting. Thus we decided to go with `libcst`
|
transformers/examples/diff-conversion/README.md/0
|
{
"file_path": "transformers/examples/diff-conversion/README.md",
"repo_id": "transformers",
"token_count": 358
}
| 311
|
#!/usr/bin/env python
# coding=utf-8
# 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.
"""
Pre-training/Fine-tuning the library models for causal language modeling (GPT, GPT-2, CTRL, ...) on a text file or a dataset.
Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
https://huggingface.co/models?filter=text-generation
"""
# You can also adapt this script on your own causal language modeling task. Pointers for this are left as comments.
import json
import logging
import math
import os
import sys
import time
from dataclasses import asdict, dataclass, field
from enum import Enum
from itertools import chain
from pathlib import Path
from typing import Callable, Optional
import datasets
import jax
import jax.numpy as jnp
import numpy as np
import optax
from datasets import Dataset, load_dataset
from flax import jax_utils, traverse_util
from flax.jax_utils import pad_shard_unpad, unreplicate
from flax.training import train_state
from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
from huggingface_hub import HfApi
from tqdm import tqdm
import transformers
from transformers import (
CONFIG_MAPPING,
FLAX_MODEL_FOR_CAUSAL_LM_MAPPING,
AutoConfig,
AutoTokenizer,
FlaxAutoModelForCausalLM,
HfArgumentParser,
is_tensorboard_available,
set_seed,
)
from transformers.testing_utils import CaptureLogger
from transformers.utils import send_example_telemetry
logger = logging.getLogger(__name__)
MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_CAUSAL_LM_MAPPING.keys())
MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
@dataclass
class TrainingArguments:
output_dir: str = field(
metadata={"help": "The output directory where the model predictions and checkpoints will be written."},
)
overwrite_output_dir: bool = field(
default=False,
metadata={
"help": (
"Overwrite the content of the output directory. "
"Use this to continue training if output_dir points to a checkpoint directory."
)
},
)
do_train: bool = field(default=False, metadata={"help": "Whether to run training."})
do_eval: bool = field(default=False, metadata={"help": "Whether to run eval on the dev set."})
per_device_train_batch_size: int = field(
default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for training."}
)
per_device_eval_batch_size: int = field(
default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for evaluation."}
)
learning_rate: float = field(default=5e-5, metadata={"help": "The initial learning rate for AdamW."})
weight_decay: float = field(default=0.0, metadata={"help": "Weight decay for AdamW if we apply some."})
adam_beta1: float = field(default=0.9, metadata={"help": "Beta1 for AdamW optimizer"})
adam_beta2: float = field(default=0.999, metadata={"help": "Beta2 for AdamW optimizer"})
adam_epsilon: float = field(default=1e-8, metadata={"help": "Epsilon for AdamW optimizer."})
adafactor: bool = field(default=False, metadata={"help": "Whether or not to replace AdamW by Adafactor."})
num_train_epochs: float = field(default=3.0, metadata={"help": "Total number of training epochs to perform."})
warmup_steps: int = field(default=0, metadata={"help": "Linear warmup over warmup_steps."})
logging_steps: int = field(default=500, metadata={"help": "Log every X updates steps."})
save_steps: int = field(default=500, metadata={"help": "Save checkpoint every X updates steps."})
eval_steps: int = field(default=None, metadata={"help": "Run an evaluation every X steps."})
seed: int = field(default=42, metadata={"help": "Random seed that will be set at the beginning of training."})
push_to_hub: bool = field(
default=False, metadata={"help": "Whether or not to upload the trained model to the model hub after training."}
)
hub_model_id: str = field(
default=None, metadata={"help": "The name of the repository to keep in sync with the local `output_dir`."}
)
hub_token: str = field(default=None, metadata={"help": "The token to use to push to the Model Hub."})
def __post_init__(self):
if self.output_dir is not None:
self.output_dir = os.path.expanduser(self.output_dir)
def to_dict(self):
"""
Serializes this instance while replace `Enum` by their values (for JSON serialization support). It obfuscates
the token values by removing their value.
"""
d = asdict(self)
for k, v in d.items():
if isinstance(v, Enum):
d[k] = v.value
if isinstance(v, list) and len(v) > 0 and isinstance(v[0], Enum):
d[k] = [x.value for x in v]
if k.endswith("_token"):
d[k] = f"<{k.upper()}>"
return d
@dataclass
class ModelArguments:
"""
Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
"""
model_name_or_path: Optional[str] = field(
default=None,
metadata={
"help": (
"The model checkpoint for weights initialization. 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: 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 s3"}
)
use_fast_tokenizer: bool = field(
default=True,
metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
)
dtype: Optional[str] = field(
default="float32",
metadata={
"help": (
"Floating-point format in which the model weights should be initialized and trained. Choose one of"
" `[float32, float16, bfloat16]`."
)
},
)
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 `huggingface-cli 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.
"""
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 text file)."})
validation_file: Optional[str] = field(
default=None,
metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
)
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."
)
},
)
validation_split_percentage: Optional[int] = field(
default=5,
metadata={
"help": "The percentage of the train set used as validation set in case there's no validation split"
},
)
block_size: Optional[int] = field(
default=None,
metadata={
"help": (
"Optional input sequence length after tokenization. "
"The training dataset will be truncated in block of this size for training. "
"Default to the model max input length for single sentence inputs (take into account special tokens)."
)
},
)
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."},
)
keep_linebreaks: bool = field(
default=True, metadata={"help": "Whether to keep line breaks when using TXT files or not."}
)
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]
if extension not in ["csv", "json", "txt"]:
raise ValueError("train_file` should be a csv, json or text file.")
if self.validation_file is not None:
extension = self.validation_file.split(".")[-1]
if extension not in ["csv", "json", "txt"]:
raise ValueError("`validation_file` should be a csv, json or text file.")
class TrainState(train_state.TrainState):
dropout_rng: jnp.ndarray
def replicate(self):
return jax_utils.replicate(self).replace(dropout_rng=shard_prng_key(self.dropout_rng))
def data_loader(rng: jax.random.PRNGKey, dataset: Dataset, batch_size: int, shuffle: bool = False, drop_last=True):
"""
Returns batches of size `batch_size` from `dataset`. If `drop_last` is set to `False`, the final batch may be incomplete,
and range in size from 1 to `batch_size`. Shuffle batches if `shuffle` is `True`.
"""
if shuffle:
batch_idx = jax.random.permutation(rng, len(dataset))
batch_idx = np.asarray(batch_idx)
else:
batch_idx = np.arange(len(dataset))
if drop_last:
steps_per_epoch = len(dataset) // batch_size
batch_idx = batch_idx[: steps_per_epoch * batch_size] # Skip incomplete batch.
batch_idx = batch_idx.reshape((steps_per_epoch, batch_size))
else:
steps_per_epoch = math.ceil(len(dataset) / batch_size)
batch_idx = np.array_split(batch_idx, steps_per_epoch)
for idx in batch_idx:
batch = dataset[idx]
batch = {k: np.array(v) for k, v in batch.items()}
yield batch
def write_train_metric(summary_writer, train_metrics, train_time, step):
summary_writer.scalar("train_time", train_time, step)
train_metrics = get_metrics(train_metrics)
for key, vals in train_metrics.items():
tag = f"train_{key}"
for i, val in enumerate(vals):
summary_writer.scalar(tag, val, step - len(vals) + i + 1)
def write_eval_metric(summary_writer, eval_metrics, step):
for metric_name, value in eval_metrics.items():
summary_writer.scalar(f"eval_{metric_name}", value, step)
def create_learning_rate_fn(
train_ds_size: int, train_batch_size: int, num_train_epochs: int, num_warmup_steps: int, learning_rate: float
) -> Callable[[int], jnp.ndarray]:
"""Returns a linear warmup, linear_decay learning rate function."""
steps_per_epoch = train_ds_size // train_batch_size
num_train_steps = steps_per_epoch * num_train_epochs
warmup_fn = optax.linear_schedule(init_value=0.0, end_value=learning_rate, transition_steps=num_warmup_steps)
decay_fn = optax.linear_schedule(
init_value=learning_rate, end_value=0, transition_steps=num_train_steps - num_warmup_steps
)
schedule_fn = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps])
return schedule_fn
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_clm", model_args, data_args, framework="flax")
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome."
)
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
# Setup logging, we only want one process per machine to log things on the screen.
logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
if jax.process_index() == 0:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# Set the verbosity to info of the Transformers logger (on main process only):
logger.info(f"Training/evaluation parameters {training_args}")
# Set seed before initializing model.
set_seed(training_args.seed)
# Handle the repository creation
if training_args.push_to_hub:
# Retrieve of infer repo_name
repo_name = training_args.hub_model_id
if repo_name is None:
repo_name = Path(training_args.output_dir).absolute().name
# Create repo and retrieve repo_id
api = HfApi()
repo_id = api.create_repo(repo_name, exist_ok=True, token=training_args.hub_token).repo_id
# 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 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 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantees 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.
dataset = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir,
keep_in_memory=False,
token=model_args.token,
num_proc=data_args.preprocessing_num_workers,
trust_remote_code=model_args.trust_remote_code,
)
if "validation" not in dataset.keys():
dataset["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[:{data_args.validation_split_percentage}%]",
cache_dir=model_args.cache_dir,
token=model_args.token,
num_proc=data_args.preprocessing_num_workers,
trust_remote_code=model_args.trust_remote_code,
)
dataset["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[{data_args.validation_split_percentage}%:]",
cache_dir=model_args.cache_dir,
token=model_args.token,
num_proc=data_args.preprocessing_num_workers,
trust_remote_code=model_args.trust_remote_code,
)
else:
data_files = {}
dataset_args = {}
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]
if extension == "txt":
extension = "text"
dataset_args["keep_linebreaks"] = data_args.keep_linebreaks
dataset = load_dataset(
extension,
data_files=data_files,
cache_dir=model_args.cache_dir,
**dataset_args,
token=model_args.token,
num_proc=data_args.preprocessing_num_workers,
)
if "validation" not in dataset.keys():
dataset["validation"] = load_dataset(
extension,
data_files=data_files,
split=f"train[:{data_args.validation_split_percentage}%]",
cache_dir=model_args.cache_dir,
**dataset_args,
token=model_args.token,
num_proc=data_args.preprocessing_num_workers,
)
dataset["train"] = load_dataset(
extension,
data_files=data_files,
split=f"train[{data_args.validation_split_percentage}%:]",
cache_dir=model_args.cache_dir,
**dataset_args,
token=model_args.token,
num_proc=data_args.preprocessing_num_workers,
)
# 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.
# Load pretrained model and tokenizer
# 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,
cache_dir=model_args.cache_dir,
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,
cache_dir=model_args.cache_dir,
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.")
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
else:
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 model_args.model_name_or_path:
model = FlaxAutoModelForCausalLM.from_pretrained(
model_args.model_name_or_path,
config=config,
seed=training_args.seed,
dtype=getattr(jnp, model_args.dtype),
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
else:
model = FlaxAutoModelForCausalLM.from_config(
config,
seed=training_args.seed,
dtype=getattr(jnp, model_args.dtype),
trust_remote_code=model_args.trust_remote_code,
)
# Preprocessing the datasets.
# First we tokenize all the texts.
if training_args.do_train:
column_names = dataset["train"].column_names
else:
column_names = dataset["validation"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
# since this will be pickled to avoid _LazyModule error in Hasher force logger loading before tokenize_function
tok_logger = transformers.utils.logging.get_logger("transformers.tokenization_utils_base")
def tokenize_function(examples):
with CaptureLogger(tok_logger) as cl:
output = tokenizer(examples[text_column_name])
# clm input could be much much longer than block_size
if "Token indices sequence length is longer than the" in cl.out:
tok_logger.warning(
"^^^^^^^^^^^^^^^^ Please ignore the warning above - this long input will be chunked into smaller bits"
" before being passed to the model."
)
return output
tokenized_datasets = dataset.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if data_args.block_size is None:
block_size = tokenizer.model_max_length
if block_size > config.max_position_embeddings:
logger.warning(
f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
f"Using block_size={min(1024, config.max_position_embeddings)} instead. You can change that default value by passing --block_size xxx."
)
block_size = min(1024, config.max_position_embeddings)
else:
if data_args.block_size > tokenizer.model_max_length:
logger.warning(
f"The block_size passed ({data_args.block_size}) is larger than the maximum length for the model "
f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
)
block_size = min(data_args.block_size, tokenizer.model_max_length)
# Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
def group_texts(examples):
# Concatenate all texts.
concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
total_length = len(concatenated_examples[list(examples.keys())[0]])
# We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
# customize this part to your needs.
if total_length >= block_size:
total_length = (total_length // block_size) * block_size
# Split by chunks of max_len.
result = {
k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
for k, t in concatenated_examples.items()
}
result["labels"] = result["input_ids"].copy()
return result
# Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
# for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower
# to preprocess.
#
# To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
# https://huggingface.co/docs/datasets/process#map
lm_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_train:
if "train" not in tokenized_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = lm_datasets["train"]
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 training_args.do_eval:
if "validation" not in tokenized_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = lm_datasets["validation"]
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))
# Enable tensorboard only on the master node
has_tensorboard = is_tensorboard_available()
if has_tensorboard and jax.process_index() == 0:
try:
from flax.metrics.tensorboard import SummaryWriter
summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
except ImportError as ie:
has_tensorboard = False
logger.warning(
f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
)
else:
logger.warning(
"Unable to display metrics through TensorBoard because the package is not installed: "
"Please run pip install tensorboard to enable."
)
# Initialize our training
rng = jax.random.PRNGKey(training_args.seed)
rng, dropout_rng = jax.random.split(rng)
# Store some constant
num_epochs = int(training_args.num_train_epochs)
train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
per_device_eval_batch_size = int(training_args.per_device_eval_batch_size)
eval_batch_size = per_device_eval_batch_size * jax.device_count()
steps_per_epoch = len(train_dataset) // train_batch_size
total_train_steps = steps_per_epoch * num_epochs
# Create learning rate schedule
linear_decay_lr_schedule_fn = create_learning_rate_fn(
len(train_dataset),
train_batch_size,
training_args.num_train_epochs,
training_args.warmup_steps,
training_args.learning_rate,
)
# We use Optax's "masking" functionality to not apply weight decay
# to bias and LayerNorm scale parameters. decay_mask_fn returns a
# mask boolean with the same structure as the parameters.
# The mask is True for parameters that should be decayed.
def decay_mask_fn(params):
flat_params = traverse_util.flatten_dict(params)
# find out all LayerNorm parameters
layer_norm_candidates = ["layernorm", "layer_norm", "ln"]
layer_norm_named_params = {
layer[-2:]
for layer_norm_name in layer_norm_candidates
for layer in flat_params.keys()
if layer_norm_name in "".join(layer).lower()
}
flat_mask = {path: (path[-1] != "bias" and path[-2:] not in layer_norm_named_params) for path in flat_params}
return traverse_util.unflatten_dict(flat_mask)
# create adam optimizer
if training_args.adafactor:
# We use the default parameters here to initialize adafactor,
# For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
optimizer = optax.adafactor(
learning_rate=linear_decay_lr_schedule_fn,
)
else:
optimizer = optax.adamw(
learning_rate=linear_decay_lr_schedule_fn,
b1=training_args.adam_beta1,
b2=training_args.adam_beta2,
eps=training_args.adam_epsilon,
weight_decay=training_args.weight_decay,
mask=decay_mask_fn,
)
# Setup train state
state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer, dropout_rng=dropout_rng)
def loss_fn(logits, labels):
shift_logits = logits[..., :-1, :]
shift_labels = labels[..., 1:]
loss = optax.softmax_cross_entropy(shift_logits, onehot(shift_labels, shift_logits.shape[-1]))
return loss.mean()
# Define gradient update step fn
def train_step(state, batch):
dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
def compute_loss(params):
labels = batch.pop("labels")
logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
loss = loss_fn(logits, labels)
return loss
grad_fn = jax.value_and_grad(compute_loss)
loss, grad = grad_fn(state.params)
grad = jax.lax.pmean(grad, "batch")
new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng)
metrics = {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}
metrics = jax.lax.pmean(metrics, axis_name="batch")
return new_state, metrics
# Define eval fn
def eval_step(params, batch):
labels = batch.pop("labels")
logits = model(**batch, params=params, train=False)[0]
loss = loss_fn(logits, labels)
# summarize metrics
metrics = {"loss": loss}
metrics = jax.lax.pmean(metrics, axis_name="batch")
return metrics
# Create parallel version of the train and eval step
p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
p_eval_step = jax.pmap(eval_step, "batch")
# Replicate the train state on each device
state = state.replicate()
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {num_epochs}")
logger.info(f" Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
logger.info(f" Total train batch size (w. parallel & distributed) = {train_batch_size}")
logger.info(f" Total optimization steps = {total_train_steps}")
train_time = 0
train_metrics = []
epochs = tqdm(range(num_epochs), desc="Epoch ... ", position=0)
for epoch in epochs:
# ======================== Training ================================
train_start = time.time()
# Create sampling rng
rng, input_rng = jax.random.split(rng)
# Generate an epoch by shuffling sampling indices from the train dataset
train_loader = data_loader(input_rng, train_dataset, train_batch_size, shuffle=True)
steps_per_epoch = len(train_dataset) // train_batch_size
# train
for step in tqdm(range(steps_per_epoch), desc="Training...", position=1, leave=False):
batch = next(train_loader)
batch = shard(batch)
state, train_metric = p_train_step(state, batch)
train_metrics.append(train_metric)
cur_step = epoch * (len(train_dataset) // train_batch_size) + step
if cur_step % training_args.logging_steps == 0 and cur_step > 0:
# Save metrics
train_metric = unreplicate(train_metric)
train_time += time.time() - train_start
if has_tensorboard and jax.process_index() == 0:
write_train_metric(summary_writer, train_metrics, train_time, cur_step)
epochs.write(
f"Step... ({cur_step} | Loss: {train_metric['loss'].mean()}, Learning Rate:"
f" {train_metric['learning_rate'].mean()})"
)
train_metrics = []
if cur_step % training_args.eval_steps == 0 and cur_step > 0:
# ======================== Evaluating ==============================
eval_metrics = []
eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size, drop_last=False)
eval_steps = math.ceil(len(eval_dataset) / eval_batch_size)
for _ in tqdm(range(eval_steps), desc="Evaluating...", position=2, leave=False):
# Model forward
batch = next(eval_loader)
metrics = pad_shard_unpad(p_eval_step, static_return=True)(
state.params, batch, min_device_batch=per_device_eval_batch_size
)
eval_metrics.append(metrics)
# normalize eval metrics
eval_metrics = get_metrics(eval_metrics)
eval_metrics = jax.tree_util.tree_map(jnp.mean, eval_metrics)
try:
eval_metrics["perplexity"] = math.exp(eval_metrics["loss"])
except OverflowError:
eval_metrics["perplexity"] = float("inf")
# Print metrics and update progress bar
desc = (
f"Step... ({cur_step} | Eval Loss: {eval_metrics['loss']} | Eval Perplexity:"
f" {eval_metrics['perplexity']})"
)
epochs.write(desc)
epochs.desc = desc
# Save metrics
if has_tensorboard and jax.process_index() == 0:
write_eval_metric(summary_writer, eval_metrics, cur_step)
if cur_step % training_args.save_steps == 0 and cur_step > 0:
# save checkpoint after each epoch and push checkpoint to the hub
if jax.process_index() == 0:
params = jax.device_get(unreplicate(state.params))
model.save_pretrained(training_args.output_dir, params=params)
tokenizer.save_pretrained(training_args.output_dir)
if training_args.push_to_hub:
api.upload_folder(
commit_message=f"Saving weights and logs of step {cur_step}",
folder_path=training_args.output_dir,
repo_id=repo_id,
repo_type="model",
token=training_args.hub_token,
)
# Eval after training
if training_args.do_eval:
eval_metrics = []
eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size, drop_last=False)
eval_steps = math.ceil(len(eval_dataset) / eval_batch_size)
for _ in tqdm(range(eval_steps), desc="Evaluating...", position=2, leave=False):
# Model forward
batch = next(eval_loader)
metrics = pad_shard_unpad(p_eval_step, static_return=True)(
state.params, batch, min_device_batch=per_device_eval_batch_size
)
eval_metrics.append(metrics)
# normalize eval metrics
eval_metrics = get_metrics(eval_metrics)
eval_metrics = jax.tree_util.tree_map(lambda x: jnp.mean(x).item(), eval_metrics)
try:
eval_metrics["perplexity"] = math.exp(eval_metrics["loss"])
except OverflowError:
eval_metrics["perplexity"] = float("inf")
if jax.process_index() == 0:
eval_metrics = {f"eval_{metric_name}": value for metric_name, value in eval_metrics.items()}
path = os.path.join(training_args.output_dir, "eval_results.json")
with open(path, "w") as f:
json.dump(eval_metrics, f, indent=4, sort_keys=True)
if __name__ == "__main__":
main()
|
transformers/examples/flax/language-modeling/run_clm_flax.py/0
|
{
"file_path": "transformers/examples/flax/language-modeling/run_clm_flax.py",
"repo_id": "transformers",
"token_count": 15979
}
| 312
|
import os
import sys
sys.path.insert(1, os.path.dirname(os.path.realpath(__file__)))
|
transformers/examples/legacy/seq2seq/__init__.py/0
|
{
"file_path": "transformers/examples/legacy/seq2seq/__init__.py",
"repo_id": "transformers",
"token_count": 34
}
| 313
|
# 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 fire
from utils import calculate_rouge, save_json
def calculate_rouge_path(pred_path, tgt_path, save_path=None, **kwargs):
"""Kwargs will be passed to calculate_rouge"""
pred_lns = [x.strip() for x in open(pred_path).readlines()]
tgt_lns = [x.strip() for x in open(tgt_path).readlines()][: len(pred_lns)]
metrics = calculate_rouge(pred_lns, tgt_lns, **kwargs)
if save_path is not None:
save_json(metrics, save_path, indent=None)
return metrics # these print nicely
if __name__ == "__main__":
fire.Fire(calculate_rouge_path)
|
transformers/examples/legacy/seq2seq/rouge_cli.py/0
|
{
"file_path": "transformers/examples/legacy/seq2seq/rouge_cli.py",
"repo_id": "transformers",
"token_count": 385
}
| 314
|
# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION. 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.
"""Named entity recognition fine-tuning: utilities to work with CoNLL-2003 task."""
import logging
import os
from dataclasses import dataclass
from enum import Enum
from typing import List, Optional, Union
from filelock import FileLock
from transformers import PreTrainedTokenizer, is_tf_available, is_torch_available
logger = logging.getLogger(__name__)
@dataclass
class InputExample:
"""
A single training/test example for token classification.
Args:
guid: Unique id for the example.
words: list. The words of the sequence.
labels: (Optional) list. The labels for each word of the sequence. This should be
specified for train and dev examples, but not for test examples.
"""
guid: str
words: List[str]
labels: Optional[List[str]]
@dataclass
class InputFeatures:
"""
A single set of features of data.
Property names are the same names as the corresponding inputs to a model.
"""
input_ids: List[int]
attention_mask: List[int]
token_type_ids: Optional[List[int]] = None
label_ids: Optional[List[int]] = None
class Split(Enum):
train = "train"
dev = "dev"
test = "test"
class TokenClassificationTask:
@staticmethod
def read_examples_from_file(data_dir, mode: Union[Split, str]) -> List[InputExample]:
raise NotImplementedError
@staticmethod
def get_labels(path: str) -> List[str]:
raise NotImplementedError
@staticmethod
def convert_examples_to_features(
examples: List[InputExample],
label_list: List[str],
max_seq_length: int,
tokenizer: PreTrainedTokenizer,
cls_token_at_end=False,
cls_token="[CLS]",
cls_token_segment_id=1,
sep_token="[SEP]",
sep_token_extra=False,
pad_on_left=False,
pad_token=0,
pad_token_segment_id=0,
pad_token_label_id=-100,
sequence_a_segment_id=0,
mask_padding_with_zero=True,
) -> List[InputFeatures]:
"""Loads a data file into a list of `InputFeatures`
`cls_token_at_end` define the location of the CLS token:
- False (Default, BERT/XLM pattern): [CLS] + A + [SEP] + B + [SEP]
- True (XLNet/GPT pattern): A + [SEP] + B + [SEP] + [CLS]
`cls_token_segment_id` define the segment id associated to the CLS token (0 for BERT, 2 for XLNet)
"""
# TODO clean up all this to leverage built-in features of tokenizers
label_map = {label: i for i, label in enumerate(label_list)}
features = []
for ex_index, example in enumerate(examples):
if ex_index % 10_000 == 0:
logger.info("Writing example %d of %d", ex_index, len(examples))
tokens = []
label_ids = []
for word, label in zip(example.words, example.labels):
word_tokens = tokenizer.tokenize(word)
# google-bert/bert-base-multilingual-cased sometimes output "nothing ([]) when calling tokenize with just a space.
if len(word_tokens) > 0:
tokens.extend(word_tokens)
# Use the real label id for the first token of the word, and padding ids for the remaining tokens
label_ids.extend([label_map[label]] + [pad_token_label_id] * (len(word_tokens) - 1))
# Account for [CLS] and [SEP] with "- 2" and with "- 3" for RoBERTa.
special_tokens_count = tokenizer.num_special_tokens_to_add()
if len(tokens) > max_seq_length - special_tokens_count:
tokens = tokens[: (max_seq_length - special_tokens_count)]
label_ids = label_ids[: (max_seq_length - special_tokens_count)]
# The convention in BERT is:
# (a) For sequence pairs:
# tokens: [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
# type_ids: 0 0 0 0 0 0 0 0 1 1 1 1 1 1
# (b) For single sequences:
# tokens: [CLS] the dog is hairy . [SEP]
# type_ids: 0 0 0 0 0 0 0
#
# Where "type_ids" are used to indicate whether this is the first
# sequence or the second sequence. The embedding vectors for `type=0` and
# `type=1` were learned during pre-training and are added to the wordpiece
# embedding vector (and position vector). This is not *strictly* necessary
# since the [SEP] token unambiguously separates the sequences, but it makes
# it easier for the model to learn the concept of sequences.
#
# For classification tasks, the first vector (corresponding to [CLS]) is
# used as the "sentence vector". Note that this only makes sense because
# the entire model is fine-tuned.
tokens += [sep_token]
label_ids += [pad_token_label_id]
if sep_token_extra:
# roberta uses an extra separator b/w pairs of sentences
tokens += [sep_token]
label_ids += [pad_token_label_id]
segment_ids = [sequence_a_segment_id] * len(tokens)
if cls_token_at_end:
tokens += [cls_token]
label_ids += [pad_token_label_id]
segment_ids += [cls_token_segment_id]
else:
tokens = [cls_token] + tokens
label_ids = [pad_token_label_id] + label_ids
segment_ids = [cls_token_segment_id] + segment_ids
input_ids = tokenizer.convert_tokens_to_ids(tokens)
# The mask has 1 for real tokens and 0 for padding tokens. Only real
# tokens are attended to.
input_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
# Zero-pad up to the sequence length.
padding_length = max_seq_length - len(input_ids)
if pad_on_left:
input_ids = ([pad_token] * padding_length) + input_ids
input_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + input_mask
segment_ids = ([pad_token_segment_id] * padding_length) + segment_ids
label_ids = ([pad_token_label_id] * padding_length) + label_ids
else:
input_ids += [pad_token] * padding_length
input_mask += [0 if mask_padding_with_zero else 1] * padding_length
segment_ids += [pad_token_segment_id] * padding_length
label_ids += [pad_token_label_id] * padding_length
assert len(input_ids) == max_seq_length
assert len(input_mask) == max_seq_length
assert len(segment_ids) == max_seq_length
assert len(label_ids) == max_seq_length
if ex_index < 5:
logger.info("*** Example ***")
logger.info("guid: %s", example.guid)
logger.info("tokens: %s", " ".join([str(x) for x in tokens]))
logger.info("input_ids: %s", " ".join([str(x) for x in input_ids]))
logger.info("input_mask: %s", " ".join([str(x) for x in input_mask]))
logger.info("segment_ids: %s", " ".join([str(x) for x in segment_ids]))
logger.info("label_ids: %s", " ".join([str(x) for x in label_ids]))
if "token_type_ids" not in tokenizer.model_input_names:
segment_ids = None
features.append(
InputFeatures(
input_ids=input_ids, attention_mask=input_mask, token_type_ids=segment_ids, label_ids=label_ids
)
)
return features
if is_torch_available():
import torch
from torch import nn
from torch.utils.data import Dataset
class TokenClassificationDataset(Dataset):
"""
This will be superseded by a framework-agnostic approach
soon.
"""
features: List[InputFeatures]
pad_token_label_id: int = nn.CrossEntropyLoss().ignore_index
# Use cross entropy ignore_index as padding label id so that only
# real label ids contribute to the loss later.
def __init__(
self,
token_classification_task: TokenClassificationTask,
data_dir: str,
tokenizer: PreTrainedTokenizer,
labels: List[str],
model_type: str,
max_seq_length: Optional[int] = None,
overwrite_cache=False,
mode: Split = Split.train,
):
# Load data features from cache or dataset file
cached_features_file = os.path.join(
data_dir,
"cached_{}_{}_{}".format(mode.value, tokenizer.__class__.__name__, str(max_seq_length)),
)
# Make sure only the first process in distributed training processes the dataset,
# and the others will use the cache.
lock_path = cached_features_file + ".lock"
with FileLock(lock_path):
if os.path.exists(cached_features_file) and not overwrite_cache:
logger.info(f"Loading features from cached file {cached_features_file}")
self.features = torch.load(cached_features_file)
else:
logger.info(f"Creating features from dataset file at {data_dir}")
examples = token_classification_task.read_examples_from_file(data_dir, mode)
# TODO clean up all this to leverage built-in features of tokenizers
self.features = token_classification_task.convert_examples_to_features(
examples,
labels,
max_seq_length,
tokenizer,
cls_token_at_end=bool(model_type in ["xlnet"]),
# xlnet has a cls token at the end
cls_token=tokenizer.cls_token,
cls_token_segment_id=2 if model_type in ["xlnet"] else 0,
sep_token=tokenizer.sep_token,
sep_token_extra=False,
# roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
pad_on_left=bool(tokenizer.padding_side == "left"),
pad_token=tokenizer.pad_token_id,
pad_token_segment_id=tokenizer.pad_token_type_id,
pad_token_label_id=self.pad_token_label_id,
)
logger.info(f"Saving features into cached file {cached_features_file}")
torch.save(self.features, cached_features_file)
def __len__(self):
return len(self.features)
def __getitem__(self, i) -> InputFeatures:
return self.features[i]
if is_tf_available():
import tensorflow as tf
class TFTokenClassificationDataset:
"""
This will be superseded by a framework-agnostic approach
soon.
"""
features: List[InputFeatures]
pad_token_label_id: int = -100
# Use cross entropy ignore_index as padding label id so that only
# real label ids contribute to the loss later.
def __init__(
self,
token_classification_task: TokenClassificationTask,
data_dir: str,
tokenizer: PreTrainedTokenizer,
labels: List[str],
model_type: str,
max_seq_length: Optional[int] = None,
overwrite_cache=False,
mode: Split = Split.train,
):
examples = token_classification_task.read_examples_from_file(data_dir, mode)
# TODO clean up all this to leverage built-in features of tokenizers
self.features = token_classification_task.convert_examples_to_features(
examples,
labels,
max_seq_length,
tokenizer,
cls_token_at_end=bool(model_type in ["xlnet"]),
# xlnet has a cls token at the end
cls_token=tokenizer.cls_token,
cls_token_segment_id=2 if model_type in ["xlnet"] else 0,
sep_token=tokenizer.sep_token,
sep_token_extra=False,
# roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
pad_on_left=bool(tokenizer.padding_side == "left"),
pad_token=tokenizer.pad_token_id,
pad_token_segment_id=tokenizer.pad_token_type_id,
pad_token_label_id=self.pad_token_label_id,
)
def gen():
for ex in self.features:
if ex.token_type_ids is None:
yield (
{"input_ids": ex.input_ids, "attention_mask": ex.attention_mask},
ex.label_ids,
)
else:
yield (
{
"input_ids": ex.input_ids,
"attention_mask": ex.attention_mask,
"token_type_ids": ex.token_type_ids,
},
ex.label_ids,
)
if "token_type_ids" not in tokenizer.model_input_names:
self.dataset = tf.data.Dataset.from_generator(
gen,
({"input_ids": tf.int32, "attention_mask": tf.int32}, tf.int64),
(
{"input_ids": tf.TensorShape([None]), "attention_mask": tf.TensorShape([None])},
tf.TensorShape([None]),
),
)
else:
self.dataset = tf.data.Dataset.from_generator(
gen,
({"input_ids": tf.int32, "attention_mask": tf.int32, "token_type_ids": tf.int32}, tf.int64),
(
{
"input_ids": tf.TensorShape([None]),
"attention_mask": tf.TensorShape([None]),
"token_type_ids": tf.TensorShape([None]),
},
tf.TensorShape([None]),
),
)
def get_dataset(self):
self.dataset = self.dataset.apply(tf.data.experimental.assert_cardinality(len(self.features)))
return self.dataset
def __len__(self):
return len(self.features)
def __getitem__(self, i) -> InputFeatures:
return self.features[i]
|
transformers/examples/legacy/token-classification/utils_ner.py/0
|
{
"file_path": "transformers/examples/legacy/token-classification/utils_ner.py",
"repo_id": "transformers",
"token_count": 7661
}
| 315
|
#!/usr/bin/env python
# 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
import logging
import os
import sys
from dataclasses import dataclass, field
from typing import Optional
import torch
from datasets import load_dataset
from torchvision.transforms import Compose, Lambda, Normalize, RandomHorizontalFlip, RandomResizedCrop, ToTensor
from torchvision.transforms.functional import InterpolationMode
import transformers
from transformers import (
HfArgumentParser,
Trainer,
TrainingArguments,
ViTImageProcessor,
ViTMAEConfig,
ViTMAEForPreTraining,
)
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 🤗 ViT model as an MAE (masked autoencoder), as proposed in https://arxiv.org/abs/2111.06377."""
logger = logging.getLogger(__name__)
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
check_min_version("4.45.0.dev0")
require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/image-pretraining/requirements.txt")
@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)."}
)
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_column_name: Optional[str] = field(
default=None, metadata={"help": "The column name of the images in the files."}
)
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."}
)
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. Don't set if you want to train a model from scratch."
)
},
)
config_name: Optional[str] = field(
default=None, metadata={"help": "Pretrained config name or path if not the same as model_name_or_path"}
)
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 the pretrained models downloaded from s3"}
)
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 `huggingface-cli login` (stored in `~/.huggingface`)."
)
},
)
mask_ratio: float = field(
default=0.75, metadata={"help": "The ratio of the number of masked tokens in the input sequence."}
)
norm_pix_loss: bool = field(
default=True, metadata={"help": "Whether or not to train with normalized pixel values as target."}
)
@dataclass
class CustomTrainingArguments(TrainingArguments):
base_learning_rate: float = field(
default=1e-3, metadata={"help": "Base learning rate: absolute_lr = base_lr * total_batch_size / 256."}
)
def collate_fn(examples):
pixel_values = torch.stack([example["pixel_values"] for example in examples])
return {"pixel_values": pixel_values}
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, CustomTrainingArguments))
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_mae", 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=data_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.keys() 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"]
# Load pretrained model and image processor
#
# 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,
}
if model_args.config_name:
config = ViTMAEConfig.from_pretrained(model_args.config_name, **config_kwargs)
elif model_args.model_name_or_path:
config = ViTMAEConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
else:
config = ViTMAEConfig()
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}")
# adapt config
config.update(
{
"mask_ratio": model_args.mask_ratio,
"norm_pix_loss": model_args.norm_pix_loss,
}
)
# create image processor
if model_args.image_processor_name:
image_processor = ViTImageProcessor.from_pretrained(model_args.image_processor_name, **config_kwargs)
elif model_args.model_name_or_path:
image_processor = ViTImageProcessor.from_pretrained(model_args.model_name_or_path, **config_kwargs)
else:
image_processor = ViTImageProcessor()
# create model
if model_args.model_name_or_path:
model = ViTMAEForPreTraining.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,
)
else:
logger.info("Training new model from scratch")
model = ViTMAEForPreTraining(config)
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 MAE paper
# source: https://github.com/facebookresearch/mae/blob/main/main_pretrain.py
if "shortest_edge" in image_processor.size:
size = image_processor.size["shortest_edge"]
else:
size = (image_processor.size["height"], image_processor.size["width"])
transforms = Compose(
[
Lambda(lambda img: img.convert("RGB") if img.mode != "RGB" else img),
RandomResizedCrop(size, scale=(0.2, 1.0), interpolation=InterpolationMode.BICUBIC),
RandomHorizontalFlip(),
ToTensor(),
Normalize(mean=image_processor.image_mean, std=image_processor.image_std),
]
)
def preprocess_images(examples):
"""Preprocess a batch of images by applying transforms."""
examples["pixel_values"] = [transforms(image) for image in 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)
# Compute absolute learning rate
total_train_batch_size = (
training_args.train_batch_size * training_args.gradient_accumulation_steps * training_args.world_size
)
if training_args.base_learning_rate is not None:
training_args.learning_rate = training_args.base_learning_rate * total_train_batch_size / 256
# 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,
tokenizer=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 = {
"tasks": "masked-auto-encoding",
"dataset": data_args.dataset_name,
"tags": ["masked-auto-encoding"],
}
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
def _mp_fn(index):
# For xla_spawn (TPUs)
main()
if __name__ == "__main__":
main()
|
transformers/examples/pytorch/image-pretraining/run_mae.py/0
|
{
"file_path": "transformers/examples/pytorch/image-pretraining/run_mae.py",
"repo_id": "transformers",
"token_count": 6331
}
| 316
|
<!---
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.
-->
# Multiple Choice
## Fine-tuning on SWAG with the Trainer
`run_swag` allows you to fine-tune any model from our [hub](https://huggingface.co/models) (as long as its architecture as a `ForMultipleChoice` version in the library) on the SWAG dataset or your own csv/jsonlines files as long as they are structured the same way. To make it works on another dataset, you will need to tweak the `preprocess_function` inside the script.
```bash
python examples/multiple-choice/run_swag.py \
--model_name_or_path FacebookAI/roberta-base \
--do_train \
--do_eval \
--learning_rate 5e-5 \
--num_train_epochs 3 \
--output_dir /tmp/swag_base \
--per_device_eval_batch_size=16 \
--per_device_train_batch_size=16 \
--overwrite_output
```
Training with the defined hyper-parameters yields the following results:
```
***** Eval results *****
eval_acc = 0.8338998300509847
eval_loss = 0.44457291918821606
```
## With Accelerate
Based on the script [run_swag_no_trainer.py](https://github.com/huggingface/transformers/blob/main/examples/pytorch/multiple-choice/run_swag_no_trainer.py).
Like `run_swag.py`, this script allows you to fine-tune any of the models on the [hub](https://huggingface.co/models) (as long as its architecture as a `ForMultipleChoice` version in the library) on
the SWAG dataset or your own data in a csv or a JSON file. The main difference is that this
script exposes the bare training loop, to allow you to quickly experiment and add any customization you would like.
It offers less options than the script with `Trainer` (but you can easily change the options for the optimizer
or the dataloaders directly in the script) but still run in a distributed setup, on TPU and supports mixed precision by
the mean of the [🤗 `Accelerate`](https://github.com/huggingface/accelerate) library. You can use the script normally
after installing it:
```bash
pip install git+https://github.com/huggingface/accelerate
```
then
```bash
export DATASET_NAME=swag
python run_swag_no_trainer.py \
--model_name_or_path google-bert/bert-base-cased \
--dataset_name $DATASET_NAME \
--max_seq_length 128 \
--per_device_train_batch_size 32 \
--learning_rate 2e-5 \
--num_train_epochs 3 \
--output_dir /tmp/$DATASET_NAME/
```
You can then use your usual launchers to run in it in a distributed environment, but the easiest way is to run
```bash
accelerate config
```
and reply to the questions asked. Then
```bash
accelerate test
```
that will check everything is ready for training. Finally, you can launch training with
```bash
export DATASET_NAME=swag
accelerate launch run_swag_no_trainer.py \
--model_name_or_path google-bert/bert-base-cased \
--dataset_name $DATASET_NAME \
--max_seq_length 128 \
--per_device_train_batch_size 32 \
--learning_rate 2e-5 \
--num_train_epochs 3 \
--output_dir /tmp/$DATASET_NAME/
```
This command is the same and will work for:
- a CPU-only setup
- a setup with one GPU
- a distributed training with several GPUs (single or multi node)
- a training on TPUs
Note that this library is in alpha release so your feedback is more than welcome if you encounter any problem using it.
|
transformers/examples/pytorch/multiple-choice/README.md/0
|
{
"file_path": "transformers/examples/pytorch/multiple-choice/README.md",
"repo_id": "transformers",
"token_count": 1170
}
| 317
|
#!/usr/bin/env python
# coding=utf-8
# 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.
"""
Fine-tuning the library's seq2seq models for question answering using the 🤗 Seq2SeqTrainer.
"""
# You can also adapt this script on your own question answering task. Pointers for this are left as comments.
import logging
import os
import sys
from dataclasses import dataclass, field
from typing import List, Optional, Tuple
import datasets
import evaluate
import numpy as np
from datasets import load_dataset
from trainer_seq2seq_qa import QuestionAnsweringSeq2SeqTrainer
import transformers
from transformers import (
AutoConfig,
AutoModelForSeq2SeqLM,
AutoTokenizer,
DataCollatorForSeq2Seq,
HfArgumentParser,
Seq2SeqTrainingArguments,
set_seed,
)
from transformers.trainer_utils import EvalLoopOutput, EvalPrediction, get_last_checkpoint
from transformers.utils import check_min_version, send_example_telemetry
from transformers.utils.versions import require_version
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
check_min_version("4.45.0.dev0")
require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/question-answering/requirements.txt")
logger = logging.getLogger(__name__)
@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": "Path to directory to store the pretrained models downloaded from huggingface.co"},
)
use_fast_tokenizer: bool = field(
default=True,
metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
)
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 `huggingface-cli 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.
"""
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)."}
)
context_column: Optional[str] = field(
default="context",
metadata={"help": "The name of the column in the datasets containing the contexts (for question answering)."},
)
question_column: Optional[str] = field(
default="question",
metadata={"help": "The name of the column in the datasets containing the questions (for question answering)."},
)
answer_column: Optional[str] = field(
default="answers",
metadata={"help": "The name of the column in the datasets containing the answers (for question answering)."},
)
train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
validation_file: Optional[str] = field(
default=None,
metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
)
test_file: Optional[str] = field(
default=None,
metadata={"help": "An optional input test data file to evaluate the perplexity on (a text 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_seq_length: int = field(
default=384,
metadata={
"help": (
"The maximum total input sequence length after tokenization. Sequences longer "
"than this will be truncated, sequences shorter will be padded."
)
},
)
max_answer_length: int = field(
default=30,
metadata={
"help": (
"The maximum length of an answer that can be generated. This is needed because the start "
"and end predictions are not conditioned on one another."
)
},
)
val_max_answer_length: Optional[int] = field(
default=None,
metadata={
"help": (
"The maximum total sequence length for validation target text after tokenization. Sequences longer "
"than this will be truncated, sequences shorter will be padded. Will default to `max_answer_length`. "
"This argument is also used to override the ``max_length`` param of ``model.generate``, which is used "
"during ``evaluate`` and ``predict``."
)
},
)
pad_to_max_length: bool = field(
default=True,
metadata={
"help": (
"Whether to pad all samples to `max_seq_length`. If False, will pad the samples dynamically when"
" batching to the maximum length in the batch (which can be faster on GPU but will be slower on 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."
)
},
)
version_2_with_negative: bool = field(
default=False, metadata={"help": "If true, some of the examples do not have an answer."}
)
null_score_diff_threshold: float = field(
default=0.0,
metadata={
"help": (
"The threshold used to select the null answer: if the best answer has a score that is less than "
"the score of the null answer minus this threshold, the null answer is selected for this example. "
"Only useful when `version_2_with_negative=True`."
)
},
)
doc_stride: int = field(
default=128,
metadata={"help": "When splitting up a long document into chunks, how much stride to take between chunks."},
)
n_best_size: int = field(
default=20,
metadata={"help": "The total number of n-best predictions to generate when looking for an answer."},
)
num_beams: Optional[int] = field(
default=None,
metadata={
"help": (
"Number of beams to use for evaluation. This argument will be passed to ``model.generate``, "
"which is used during ``evaluate`` and ``predict``."
)
},
)
ignore_pad_token_for_loss: bool = field(
default=True,
metadata={
"help": "Whether to ignore the tokens corresponding to padded labels in the loss computation or not."
},
)
def __post_init__(self):
if (
self.dataset_name is None
and self.train_file is None
and self.validation_file is None
and self.test_file is None
):
raise ValueError("Need either a dataset name or a training/validation file/test_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."
if self.test_file is not None:
extension = self.test_file.split(".")[-1]
assert extension in ["csv", "json"], "`test_file` should be a csv or a json file."
if self.val_max_answer_length is None:
self.val_max_answer_length = self.max_answer_length
question_answering_column_name_mapping = {
"squad_v2": ("question", "context", "answer"),
}
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, Seq2SeqTrainingArguments))
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_seq2seq_qa", 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)
datasets.utils.logging.set_verbosity(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."
)
# Set seed before initializing model.
set_seed(training_args.seed)
# 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 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 'text' or the first column if no column called
# 'text' 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,
cache_dir=model_args.cache_dir,
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]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
raw_datasets = load_dataset(
extension,
data_files=data_files,
field="data",
cache_dir=model_args.cache_dir,
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.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
model = AutoModelForSeq2SeqLM.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,
)
# 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.
embedding_size = model.get_input_embeddings().weight.shape[0]
if len(tokenizer) > embedding_size:
model.resize_token_embeddings(len(tokenizer))
if model.config.decoder_start_token_id is None:
raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
# Preprocessing the datasets.
# We need to generate and tokenize inputs and targets.
if training_args.do_train:
column_names = raw_datasets["train"].column_names
elif training_args.do_eval:
column_names = raw_datasets["validation"].column_names
elif training_args.do_predict:
column_names = raw_datasets["test"].column_names
else:
logger.info("There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`.")
return
# Get the column names for input/target.
dataset_columns = question_answering_column_name_mapping.get(data_args.dataset_name, None)
if data_args.question_column is None:
question_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
else:
question_column = data_args.question_column
if question_column not in column_names:
raise ValueError(
f"--question_column' value '{data_args.question_column}' needs to be one of: {', '.join(column_names)}"
)
if data_args.context_column is None:
context_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
else:
context_column = data_args.context_column
if context_column not in column_names:
raise ValueError(
f"--context_column' value '{data_args.context_column}' needs to be one of: {', '.join(column_names)}"
)
if data_args.answer_column is None:
answer_column = dataset_columns[2] if dataset_columns is not None else column_names[2]
else:
answer_column = data_args.answer_column
if answer_column not in column_names:
raise ValueError(
f"--answer_column' value '{data_args.answer_column}' needs to be one of: {', '.join(column_names)}"
)
# Temporarily set max_answer_length for training.
max_answer_length = data_args.max_answer_length
padding = "max_length" if data_args.pad_to_max_length else False
if training_args.label_smoothing_factor > 0 and not hasattr(model, "prepare_decoder_input_ids_from_labels"):
logger.warning(
"label_smoothing is enabled but the `prepare_decoder_input_ids_from_labels` method is not defined for "
f"`{model.__class__.__name__}`. This will lead to loss being calculated twice and will take up more memory"
)
if data_args.max_seq_length > tokenizer.model_max_length:
logger.warning(
f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the "
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
def preprocess_squad_batch(
examples,
question_column: str,
context_column: str,
answer_column: str,
) -> Tuple[List[str], List[str]]:
questions = examples[question_column]
contexts = examples[context_column]
answers = examples[answer_column]
def generate_input(_question, _context):
return " ".join(["question:", _question.lstrip(), "context:", _context.lstrip()])
inputs = [generate_input(question, context) for question, context in zip(questions, contexts)]
targets = [answer["text"][0] if len(answer["text"]) > 0 else "" for answer in answers]
return inputs, targets
def preprocess_function(examples):
inputs, targets = preprocess_squad_batch(examples, question_column, context_column, answer_column)
model_inputs = tokenizer(inputs, max_length=max_seq_length, padding=padding, truncation=True)
# Tokenize targets with text_target=...
labels = tokenizer(text_target=targets, max_length=max_answer_length, padding=padding, truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [
[(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels["input_ids"]
]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
# Validation preprocessing
def preprocess_validation_function(examples):
inputs, targets = preprocess_squad_batch(examples, question_column, context_column, answer_column)
model_inputs = tokenizer(
inputs,
max_length=max_seq_length,
padding=padding,
truncation=True,
return_overflowing_tokens=True,
return_offsets_mapping=True,
)
# Tokenize targets with the `text_target` keyword argument
labels = tokenizer(text_target=targets, max_length=max_answer_length, padding=padding, truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [
[(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels["input_ids"]
]
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = model_inputs.pop("overflow_to_sample_mapping")
# For evaluation, we will need to convert our predictions to substrings of the context, so we keep the
# corresponding example_id and we will store the offset mappings.
model_inputs["example_id"] = []
# Augment the overflowing tokens to the labels
labels_out = []
for i in range(len(model_inputs["input_ids"])):
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
model_inputs["example_id"].append(examples["id"][sample_index])
labels_out.append(labels["input_ids"][sample_index])
model_inputs["labels"] = labels_out
return model_inputs
if training_args.do_train:
if "train" not in raw_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = raw_datasets["train"]
if data_args.max_train_samples is not None:
# We will select sample from whole data if argument is specified
max_train_samples = min(len(train_dataset), data_args.max_train_samples)
train_dataset = train_dataset.select(range(max_train_samples))
# Create train feature from dataset
with training_args.main_process_first(desc="train dataset map pre-processing"):
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on train dataset",
)
if data_args.max_train_samples is not None:
# Number of samples might increase during Feature Creation, We select only specified max samples
max_train_samples = min(len(train_dataset), data_args.max_train_samples)
train_dataset = train_dataset.select(range(max_train_samples))
if training_args.do_eval:
if "validation" not in raw_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_examples = raw_datasets["validation"]
if data_args.max_eval_samples is not None:
# We will select sample from whole data
max_eval_samples = min(len(eval_examples), data_args.max_eval_samples)
eval_examples = eval_examples.select(range(max_eval_samples))
# Validation Feature Creation
with training_args.main_process_first(desc="validation dataset map pre-processing"):
eval_dataset = eval_examples.map(
preprocess_validation_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on validation dataset",
)
if data_args.max_eval_samples is not None:
# During Feature creation dataset samples might increase, we will select required samples again
max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
eval_dataset = eval_dataset.select(range(max_eval_samples))
if training_args.do_predict:
if "test" not in raw_datasets:
raise ValueError("--do_predict requires a test dataset")
predict_examples = raw_datasets["test"]
if data_args.max_predict_samples is not None:
# We will select sample from whole data
predict_examples = predict_examples.select(range(data_args.max_predict_samples))
# Predict Feature Creation
with training_args.main_process_first(desc="prediction dataset map pre-processing"):
predict_dataset = predict_examples.map(
preprocess_validation_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on prediction dataset",
)
if data_args.max_predict_samples is not None:
# During Feature creation dataset samples might increase, we will select required samples again
max_predict_samples = min(len(predict_dataset), data_args.max_predict_samples)
predict_dataset = predict_dataset.select(range(max_predict_samples))
# Data collator
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if training_args.fp16 else None,
)
metric = evaluate.load(
"squad_v2" if data_args.version_2_with_negative else "squad", cache_dir=model_args.cache_dir
)
def compute_metrics(p: EvalPrediction):
return metric.compute(predictions=p.predictions, references=p.label_ids)
# Post-processing:
def post_processing_function(
examples: datasets.Dataset, features: datasets.Dataset, outputs: EvalLoopOutput, stage="eval"
):
# Decode the predicted tokens.
preds = outputs.predictions
if isinstance(preds, tuple):
preds = preds[0]
# Replace -100s used for padding as we can't decode them
preds = np.where(preds != -100, preds, tokenizer.pad_token_id)
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
# Build a map example to its corresponding features.
example_id_to_index = {k: i for i, k in enumerate(examples["id"])}
feature_per_example = {example_id_to_index[feature["example_id"]]: i for i, feature in enumerate(features)}
predictions = {}
# Let's loop over all the examples!
for example_index, example in enumerate(examples):
# This is the index of the feature associated to the current example.
feature_index = feature_per_example[example_index]
predictions[example["id"]] = decoded_preds[feature_index]
# Format the result to the format the metric expects.
if data_args.version_2_with_negative:
formatted_predictions = [
{"id": k, "prediction_text": v, "no_answer_probability": 0.0} for k, v in predictions.items()
]
else:
formatted_predictions = [{"id": k, "prediction_text": v} for k, v in predictions.items()]
references = [{"id": ex["id"], "answers": ex[answer_column]} for ex in examples]
return EvalPrediction(predictions=formatted_predictions, label_ids=references)
# Initialize our Trainer
trainer = QuestionAnsweringSeq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
eval_examples=eval_examples if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics if training_args.predict_with_generate else None,
post_process_function=post_processing_function,
)
# 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() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
max_train_samples = (
data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)
)
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
max_length = (
training_args.generation_max_length
if training_args.generation_max_length is not None
else data_args.val_max_answer_length
)
num_beams = data_args.num_beams if data_args.num_beams is not None else training_args.generation_num_beams
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(max_length=max_length, num_beams=num_beams, metric_key_prefix="eval")
max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(eval_dataset)
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Prediction
if training_args.do_predict:
logger.info("*** Predict ***")
results = trainer.predict(predict_dataset, predict_examples)
metrics = results.metrics
max_predict_samples = (
data_args.max_predict_samples if data_args.max_predict_samples is not None else len(predict_dataset)
)
metrics["predict_samples"] = min(max_predict_samples, len(predict_dataset))
trainer.log_metrics("predict", metrics)
trainer.save_metrics("predict", metrics)
if training_args.push_to_hub:
kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "question-answering"}
if data_args.dataset_name is not None:
kwargs["dataset_tags"] = data_args.dataset_name
if data_args.dataset_config_name is not None:
kwargs["dataset_args"] = data_args.dataset_config_name
kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
else:
kwargs["dataset"] = data_args.dataset_name
trainer.push_to_hub(**kwargs)
def _mp_fn(index):
# For xla_spawn (TPUs)
main()
if __name__ == "__main__":
main()
|
transformers/examples/pytorch/question-answering/run_seq2seq_qa.py/0
|
{
"file_path": "transformers/examples/pytorch/question-answering/run_seq2seq_qa.py",
"repo_id": "transformers",
"token_count": 13280
}
| 318
|
<!---
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.
-->
## Summarization
This directory contains examples for finetuning and evaluating transformers on summarization tasks.
Please tag @patil-suraj with any issues/unexpected behaviors, or send a PR!
For deprecated `bertabs` instructions, see [`bertabs/README.md`](https://github.com/huggingface/transformers/blob/main/examples/research_projects/bertabs/README.md).
For the old `finetune_trainer.py` and related utils, see [`examples/legacy/seq2seq`](https://github.com/huggingface/transformers/blob/main/examples/legacy/seq2seq).
### Supported Architectures
- `BartForConditionalGeneration`
- `FSMTForConditionalGeneration` (translation only)
- `MBartForConditionalGeneration`
- `MarianMTModel`
- `PegasusForConditionalGeneration`
- `T5ForConditionalGeneration`
- `MT5ForConditionalGeneration`
`run_summarization.py` is a lightweight example of how to download and preprocess a dataset from the [🤗 Datasets](https://github.com/huggingface/datasets) library or use your own files (jsonlines or csv), then fine-tune one of the architectures above on it.
For custom datasets in `jsonlines` format please see: https://huggingface.co/docs/datasets/loading_datasets#json-files
and you also will find examples of these below.
## With Trainer
Here is an example on a summarization task:
```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
```
Only T5 models `google-t5/t5-small`, `google-t5/t5-base`, `google-t5/t5-large`, `google-t5/t5-3b` and `google-t5/t5-11b` must use an additional argument: `--source_prefix "summarize: "`.
We used CNN/DailyMail dataset in this example as `google-t5/t5-small` was trained on it and one can get good scores even when pre-training with a very small sample.
Extreme Summarization (XSum) Dataset is another commonly used dataset for the task of summarization. To use it replace `--dataset_name cnn_dailymail --dataset_config "3.0.0"` with `--dataset_name xsum`.
And here is how you would use it on your own files, after adjusting the values for the arguments
`--train_file`, `--validation_file`, `--text_column` and `--summary_column` to match your setup:
```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 \
--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
```
The task of summarization supports custom CSV and JSONLINES formats.
#### Custom CSV Files
If it's a csv file the training and validation files should have a column for the inputs texts and a column for the summaries.
If the csv file has just two columns as in the following example:
```csv
text,summary
"I'm sitting here in a boring room. It's just another rainy Sunday afternoon. I'm wasting my time I got nothing to do. I'm hanging around I'm waiting for you. But nothing ever happens. And I wonder","I'm sitting in a room where I'm waiting for something to happen"
"I see trees so green, red roses too. I see them bloom for me and you. And I think to myself what a wonderful world. I see skies so blue and clouds so white. The bright blessed day, the dark sacred night. And I think to myself what a wonderful world.","I'm a gardener and I'm a big fan of flowers."
"Christmas time is here. Happiness and cheer. Fun for all that children call. Their favorite time of the year. Snowflakes in the air. Carols everywhere. Olden times and ancient rhymes. Of love and dreams to share","It's that time of year again."
```
The first column is assumed to be for `text` and the second is for summary.
If the csv file has multiple columns, you can then specify the names of the columns to use:
```bash
--text_column text_column_name \
--summary_column summary_column_name \
```
For example if the columns were:
```csv
id,date,text,summary
```
and you wanted to select only `text` and `summary`, then you'd pass these additional arguments:
```bash
--text_column text \
--summary_column summary \
```
#### Custom JSONLINES Files
The second supported format is jsonlines. Here is an example of a jsonlines custom data file.
```json
{"text": "I'm sitting here in a boring room. It's just another rainy Sunday afternoon. I'm wasting my time I got nothing to do. I'm hanging around I'm waiting for you. But nothing ever happens. And I wonder", "summary": "I'm sitting in a room where I'm waiting for something to happen"}
{"text": "I see trees so green, red roses too. I see them bloom for me and you. And I think to myself what a wonderful world. I see skies so blue and clouds so white. The bright blessed day, the dark sacred night. And I think to myself what a wonderful world.", "summary": "I'm a gardener and I'm a big fan of flowers."}
{"text": "Christmas time is here. Happiness and cheer. Fun for all that children call. Their favorite time of the year. Snowflakes in the air. Carols everywhere. Olden times and ancient rhymes. Of love and dreams to share", "summary": "It's that time of year again."}
```
Same as with the CSV files, by default the first value will be used as the text record and the second as the summary record. Therefore you can use any key names for the entries, in this example `text` and `summary` were used.
And as with the CSV files, you can specify which values to select from the file, by explicitly specifying the corresponding key names. In our example this again would be:
```bash
--text_column text \
--summary_column summary \
```
## With Accelerate
Based on the script [`run_summarization_no_trainer.py`](https://github.com/huggingface/transformers/blob/main/examples/pytorch/summarization/run_summarization_no_trainer.py).
Like `run_summarization.py`, this script allows you to fine-tune any of the models supported on a
summarization task, the main difference is that this
script exposes the bare training loop, to allow you to quickly experiment and add any customization you would like.
It offers less options than the script with `Trainer` (for instance you can easily change the options for the optimizer
or the dataloaders directly in the script) but still run in a distributed setup, on TPU and supports mixed precision by
the mean of the [🤗 `Accelerate`](https://github.com/huggingface/accelerate) library. You can use the script normally
after installing it:
```bash
pip install git+https://github.com/huggingface/accelerate
```
then
```bash
python 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
```
You can then use your usual launchers to run in it in a distributed environment, but the easiest way is to run
```bash
accelerate config
```
and reply to the questions asked. Then
```bash
accelerate test
```
that will check everything is ready for training. Finally, you can launch training with
```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
```
This command is the same and will work for:
- a CPU-only setup
- a setup with one GPU
- a distributed training with several GPUs (single or multi node)
- a training on TPUs
Note that this library is in alpha release so your feedback is more than welcome if you encounter any problem using it.
|
transformers/examples/pytorch/summarization/README.md/0
|
{
"file_path": "transformers/examples/pytorch/summarization/README.md",
"repo_id": "transformers",
"token_count": 2636
}
| 319
|
<!---
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.
-->
# Token classification
## PyTorch version
Fine-tuning the library models for token classification task such as Named Entity Recognition (NER), Parts-of-speech
tagging (POS) or phrase extraction (CHUNKS). The main scrip `run_ner.py` leverages the 🤗 Datasets library and the Trainer API. You can easily
customize it to your needs if you need extra processing on your datasets.
It will either run on a datasets hosted on our [hub](https://huggingface.co/datasets) or with your own text files for
training and validation, you might just need to add some tweaks in the data preprocessing.
### Using your own data
If you use your own data, the script expects the following format of the data -
```bash
{
"chunk_tags": [11, 12, 12, 21, 13, 11, 11, 21, 13, 11, 12, 13, 11, 21, 22, 11, 12, 17, 11, 21, 17, 11, 12, 12, 21, 22, 22, 13, 11, 0],
"id": "0",
"ner_tags": [0, 3, 4, 0, 0, 0, 0, 0, 0, 7, 0, 0, 0, 0, 0, 7, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
"pos_tags": [12, 22, 22, 38, 15, 22, 28, 38, 15, 16, 21, 35, 24, 35, 37, 16, 21, 15, 24, 41, 15, 16, 21, 21, 20, 37, 40, 35, 21, 7],
"tokens": ["The", "European", "Commission", "said", "on", "Thursday", "it", "disagreed", "with", "German", "advice", "to", "consumers", "to", "shun", "British", "lamb", "until", "scientists", "determine", "whether", "mad", "cow", "disease", "can", "be", "transmitted", "to", "sheep", "."]
}
```
The following example fine-tunes BERT on CoNLL-2003:
```bash
python run_ner.py \
--model_name_or_path google-bert/bert-base-uncased \
--dataset_name conll2003 \
--output_dir /tmp/test-ner \
--do_train \
--do_eval
```
or just can just run the bash script `run.sh`.
To run on your own training and validation files, use the following command:
```bash
python run_ner.py \
--model_name_or_path google-bert/bert-base-uncased \
--train_file path_to_train_file \
--validation_file path_to_validation_file \
--output_dir /tmp/test-ner \
--do_train \
--do_eval
```
**Note:** This script only works with models that have a fast tokenizer (backed by the 🤗 Tokenizers library) as it
uses special features of those tokenizers. You can check if your favorite model has a fast tokenizer in
[this table](https://huggingface.co/transformers/index.html#supported-frameworks), if it doesn't you can still use the old version
of the script.
> 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.
## Old version of the script
You can find the old version of the PyTorch script [here](https://github.com/huggingface/transformers/blob/main/examples/legacy/token-classification/run_ner.py).
## Pytorch version, no Trainer
Based on the script [run_ner_no_trainer.py](https://github.com/huggingface/transformers/blob/main/examples/pytorch/token-classification/run_ner_no_trainer.py).
Like `run_ner.py`, this script allows you to fine-tune any of the models on the [hub](https://huggingface.co/models) on a
token classification task, either NER, POS or CHUNKS tasks or your own data in a csv or a JSON file. The main difference is that this
script exposes the bare training loop, to allow you to quickly experiment and add any customization you would like.
It offers less options than the script with `Trainer` (for instance you can easily change the options for the optimizer
or the dataloaders directly in the script) but still run in a distributed setup, on TPU and supports mixed precision by
the mean of the [🤗 `Accelerate`](https://github.com/huggingface/accelerate) library. You can use the script normally
after installing it:
```bash
pip install git+https://github.com/huggingface/accelerate
```
then
```bash
export TASK_NAME=ner
python run_ner_no_trainer.py \
--model_name_or_path google-bert/bert-base-cased \
--dataset_name conll2003 \
--task_name $TASK_NAME \
--max_length 128 \
--per_device_train_batch_size 32 \
--learning_rate 2e-5 \
--num_train_epochs 3 \
--output_dir /tmp/$TASK_NAME/
```
You can then use your usual launchers to run in it in a distributed environment, but the easiest way is to run
```bash
accelerate config
```
and reply to the questions asked. Then
```bash
accelerate test
```
that will check everything is ready for training. Finally, you can launch training with
```bash
export TASK_NAME=ner
accelerate launch run_ner_no_trainer.py \
--model_name_or_path google-bert/bert-base-cased \
--dataset_name conll2003 \
--task_name $TASK_NAME \
--max_length 128 \
--per_device_train_batch_size 32 \
--learning_rate 2e-5 \
--num_train_epochs 3 \
--output_dir /tmp/$TASK_NAME/
```
This command is the same and will work for:
- a CPU-only setup
- a setup with one GPU
- a distributed training with several GPUs (single or multi node)
- a training on TPUs
Note that this library is in alpha release so your feedback is more than welcome if you encounter any problem using it.
|
transformers/examples/pytorch/token-classification/README.md/0
|
{
"file_path": "transformers/examples/pytorch/token-classification/README.md",
"repo_id": "transformers",
"token_count": 1813
}
| 320
|
# Patience-based Early Exit
Patience-based Early Exit (PABEE) is a plug-and-play inference method for pretrained language models.
We have already implemented it on BERT and ALBERT. Basically, you can make your LM faster and more robust with PABEE. It can even improve the performance of ALBERT on GLUE. The only sacrifice is that the batch size can only be 1.
Learn more in the paper ["BERT Loses Patience: Fast and Robust Inference with Early Exit"](https://arxiv.org/abs/2006.04152) and the official [GitHub repo](https://github.com/JetRunner/PABEE).
## Training
You can fine-tune a pretrained language model (you can choose from BERT and ALBERT) and train the internal classifiers by:
```bash
export GLUE_DIR=/path/to/glue_data
export TASK_NAME=MRPC
python ./run_glue_with_pabee.py \
--model_type albert \
--model_name_or_path google-bert/bert-base-uncased/albert/albert-base-v2 \
--task_name $TASK_NAME \
--do_train \
--do_eval \
--do_lower_case \
--data_dir "$GLUE_DIR/$TASK_NAME" \
--max_seq_length 128 \
--per_gpu_train_batch_size 32 \
--per_gpu_eval_batch_size 32 \
--learning_rate 2e-5 \
--save_steps 50 \
--logging_steps 50 \
--num_train_epochs 5 \
--output_dir /path/to/save/ \
--evaluate_during_training
```
## Inference
You can inference with different patience settings by:
```bash
export GLUE_DIR=/path/to/glue_data
export TASK_NAME=MRPC
python ./run_glue_with_pabee.py \
--model_type albert \
--model_name_or_path /path/to/save/ \
--task_name $TASK_NAME \
--do_eval \
--do_lower_case \
--data_dir "$GLUE_DIR/$TASK_NAME" \
--max_seq_length 128 \
--per_gpu_eval_batch_size 1 \
--learning_rate 2e-5 \
--logging_steps 50 \
--num_train_epochs 15 \
--output_dir /path/to/save/ \
--eval_all_checkpoints \
--patience 3,4,5,6,7,8
```
where `patience` can be a list of patience settings, separated by a comma. It will help determine which patience works best.
When evaluating on a regression task (STS-B), you may add `--regression_threshold 0.1` to define the regression threshold.
## Results
On the GLUE dev set:
| Model | \#Param | Speed | CoLA | MNLI | MRPC | QNLI | QQP | RTE | SST\-2 | STS\-B |
|--------------|---------|--------|-------|-------|-------|-------|-------|-------|--------|--------|
| ALBERT\-base | 12M | | 58\.9 | 84\.6 | 89\.5 | 91\.7 | 89\.6 | 78\.6 | 92\.8 | 89\.5 |
| \+PABEE | 12M | 1\.57x | 61\.2 | 85\.1 | 90\.0 | 91\.8 | 89\.6 | 80\.1 | 93\.0 | 90\.1 |
| Model | \#Param | Speed\-up | MNLI | SST\-2 | STS\-B |
|---------------|---------|-----------|-------|--------|--------|
| BERT\-base | 108M | | 84\.5 | 92\.1 | 88\.9 |
| \+PABEE | 108M | 1\.62x | 83\.6 | 92\.0 | 88\.7 |
| ALBERT\-large | 18M | | 86\.4 | 94\.9 | 90\.4 |
| \+PABEE | 18M | 2\.42x | 86\.8 | 95\.2 | 90\.6 |
## Citation
If you find this resource useful, please consider citing the following paper:
```bibtex
@misc{zhou2020bert,
title={BERT Loses Patience: Fast and Robust Inference with Early Exit},
author={Wangchunshu Zhou and Canwen Xu and Tao Ge and Julian McAuley and Ke Xu and Furu Wei},
year={2020},
eprint={2006.04152},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
```
|
transformers/examples/research_projects/bert-loses-patience/README.md/0
|
{
"file_path": "transformers/examples/research_projects/bert-loses-patience/README.md",
"repo_id": "transformers",
"token_count": 1329
}
| 321
|
# coding=utf-8
# Copyright 2019-present, the HuggingFace Inc. team and Facebook, Inc.
#
# 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.
"""The distiller to distil the student.
Adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
"""
import math
import os
import time
import psutil
import torch
from grouped_batch_sampler import GroupedBatchSampler, create_lengths_groups
from lm_seqs_dataset import LmSeqsDataset
from torch import nn
from torch.optim import AdamW
from torch.utils.data import BatchSampler, DataLoader, RandomSampler
from torch.utils.data.distributed import DistributedSampler
from tqdm import tqdm
from transformers import get_linear_schedule_with_warmup
from utils import logger
try:
from torch.utils.tensorboard import SummaryWriter
except ImportError:
from tensorboardX import SummaryWriter
class Distiller:
def __init__(
self, params: dict, dataset: LmSeqsDataset, token_probs: torch.tensor, student: nn.Module, teacher: nn.Module
):
logger.info("Initializing Distiller")
self.params = params
self.dump_path = params.dump_path
self.multi_gpu = params.multi_gpu
self.fp16 = params.fp16
self.student = student
self.teacher = teacher
self.student_config = student.config
self.vocab_size = student.config.vocab_size
if params.n_gpu <= 1:
sampler = RandomSampler(dataset)
else:
sampler = DistributedSampler(dataset)
if params.group_by_size:
groups = create_lengths_groups(lengths=dataset.lengths, k=params.max_model_input_size)
sampler = GroupedBatchSampler(sampler=sampler, group_ids=groups, batch_size=params.batch_size)
else:
sampler = BatchSampler(sampler=sampler, batch_size=params.batch_size, drop_last=False)
self.dataloader = DataLoader(dataset=dataset, batch_sampler=sampler, collate_fn=dataset.batch_sequences)
self.temperature = params.temperature
assert self.temperature > 0.0
self.alpha_ce = params.alpha_ce
self.alpha_mlm = params.alpha_mlm
self.alpha_clm = params.alpha_clm
self.alpha_mse = params.alpha_mse
self.alpha_cos = params.alpha_cos
self.mlm = params.mlm
if self.mlm:
logger.info("Using MLM loss for LM step.")
self.mlm_mask_prop = params.mlm_mask_prop
assert 0.0 <= self.mlm_mask_prop <= 1.0
assert params.word_mask + params.word_keep + params.word_rand == 1.0
self.pred_probs = torch.FloatTensor([params.word_mask, params.word_keep, params.word_rand])
self.pred_probs = self.pred_probs.to(f"cuda:{params.local_rank}") if params.n_gpu > 0 else self.pred_probs
self.token_probs = token_probs.to(f"cuda:{params.local_rank}") if params.n_gpu > 0 else token_probs
if self.fp16:
self.pred_probs = self.pred_probs.half()
self.token_probs = self.token_probs.half()
else:
logger.info("Using CLM loss for LM step.")
self.epoch = 0
self.n_iter = 0
self.n_total_iter = 0
self.n_sequences_epoch = 0
self.total_loss_epoch = 0
self.last_loss = 0
self.last_loss_ce = 0
self.last_loss_mlm = 0
self.last_loss_clm = 0
if self.alpha_mse > 0.0:
self.last_loss_mse = 0
if self.alpha_cos > 0.0:
self.last_loss_cos = 0
self.last_log = 0
self.ce_loss_fct = nn.KLDivLoss(reduction="batchmean")
self.lm_loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
if self.alpha_mse > 0.0:
self.mse_loss_fct = nn.MSELoss(reduction="sum")
if self.alpha_cos > 0.0:
self.cosine_loss_fct = nn.CosineEmbeddingLoss(reduction="mean")
logger.info("--- Initializing model optimizer")
assert params.gradient_accumulation_steps >= 1
self.num_steps_epoch = len(self.dataloader)
num_train_optimization_steps = (
int(self.num_steps_epoch / params.gradient_accumulation_steps * params.n_epoch) + 1
)
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in student.named_parameters() if not any(nd in n for nd in no_decay) and p.requires_grad
],
"weight_decay": params.weight_decay,
},
{
"params": [
p for n, p in student.named_parameters() if any(nd in n for nd in no_decay) and p.requires_grad
],
"weight_decay": 0.0,
},
]
logger.info(
"------ Number of trainable parameters (student): %i"
% sum([p.numel() for p in self.student.parameters() if p.requires_grad])
)
logger.info("------ Number of parameters (student): %i" % sum([p.numel() for p in self.student.parameters()]))
self.optimizer = AdamW(
optimizer_grouped_parameters, lr=params.learning_rate, eps=params.adam_epsilon, betas=(0.9, 0.98)
)
warmup_steps = math.ceil(num_train_optimization_steps * params.warmup_prop)
self.scheduler = get_linear_schedule_with_warmup(
self.optimizer, num_warmup_steps=warmup_steps, num_training_steps=num_train_optimization_steps
)
if self.fp16:
try:
from apex import amp
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
logger.info(f"Using fp16 training: {self.params.fp16_opt_level} level")
self.student, self.optimizer = amp.initialize(
self.student, self.optimizer, opt_level=self.params.fp16_opt_level
)
self.teacher = self.teacher.half()
if self.multi_gpu:
if self.fp16:
from apex.parallel import DistributedDataParallel
logger.info("Using apex.parallel.DistributedDataParallel for distributed training.")
self.student = DistributedDataParallel(self.student)
else:
from torch.nn.parallel import DistributedDataParallel
logger.info("Using nn.parallel.DistributedDataParallel for distributed training.")
self.student = DistributedDataParallel(
self.student,
device_ids=[params.local_rank],
output_device=params.local_rank,
find_unused_parameters=True,
)
self.is_master = params.is_master
if self.is_master:
logger.info("--- Initializing Tensorboard")
self.tensorboard = SummaryWriter(log_dir=os.path.join(self.dump_path, "log", "train"))
self.tensorboard.add_text(tag="config/training", text_string=str(self.params), global_step=0)
self.tensorboard.add_text(tag="config/student", text_string=str(self.student_config), global_step=0)
def prepare_batch_mlm(self, batch):
"""
Prepare the batch: from the token_ids and the lengths, compute the attention mask and the masked label for MLM.
Input:
------
batch: `Tuple`
token_ids: `torch.tensor(bs, seq_length)` - The token ids for each of the sequence. It is padded.
lengths: `torch.tensor(bs)` - The lengths of each of the sequences in the batch.
Output:
-------
token_ids: `torch.tensor(bs, seq_length)` - The token ids after the modifications for MLM.
attn_mask: `torch.tensor(bs, seq_length)` - The attention mask for the self-attention.
mlm_labels: `torch.tensor(bs, seq_length)` - The masked language modeling labels. There is a -100 where there is nothing to predict.
"""
token_ids, lengths = batch
token_ids, lengths = self.round_batch(x=token_ids, lengths=lengths)
assert token_ids.size(0) == lengths.size(0)
attn_mask = torch.arange(token_ids.size(1), dtype=torch.long, device=lengths.device) < lengths[:, None]
bs, max_seq_len = token_ids.size()
mlm_labels = token_ids.new(token_ids.size()).copy_(token_ids)
x_prob = self.token_probs[token_ids.flatten()]
n_tgt = math.ceil(self.mlm_mask_prop * lengths.sum().item())
tgt_ids = torch.multinomial(x_prob / x_prob.sum(), n_tgt, replacement=False)
pred_mask = torch.zeros(
bs * max_seq_len, dtype=torch.bool, device=token_ids.device
) # previously `dtype=torch.uint8`, cf pytorch 1.2.0 compatibility
pred_mask[tgt_ids] = 1
pred_mask = pred_mask.view(bs, max_seq_len)
pred_mask[token_ids == self.params.special_tok_ids["pad_token"]] = 0
# mask a number of words == 0 [8] (faster with fp16)
if self.fp16:
n1 = pred_mask.sum().item()
if n1 > 8:
pred_mask = pred_mask.view(-1)
n2 = max(n1 % 8, 8 * (n1 // 8))
if n2 != n1:
pred_mask[torch.nonzero(pred_mask).view(-1)[: n1 - n2]] = 0
pred_mask = pred_mask.view(bs, max_seq_len)
assert pred_mask.sum().item() % 8 == 0, pred_mask.sum().item()
_token_ids_real = token_ids[pred_mask]
_token_ids_rand = _token_ids_real.clone().random_(self.vocab_size)
_token_ids_mask = _token_ids_real.clone().fill_(self.params.special_tok_ids["mask_token"])
probs = torch.multinomial(self.pred_probs, len(_token_ids_real), replacement=True)
_token_ids = (
_token_ids_mask * (probs == 0).long()
+ _token_ids_real * (probs == 1).long()
+ _token_ids_rand * (probs == 2).long()
)
token_ids = token_ids.masked_scatter(pred_mask, _token_ids)
mlm_labels[~pred_mask] = -100 # previously `mlm_labels[1-pred_mask] = -1`, cf pytorch 1.2.0 compatibility
# sanity checks
assert 0 <= token_ids.min() <= token_ids.max() < self.vocab_size
return token_ids, attn_mask, mlm_labels
def prepare_batch_clm(self, batch):
"""
Prepare the batch: from the token_ids and the lengths, compute the attention mask and the labels for CLM.
Input:
------
batch: `Tuple`
token_ids: `torch.tensor(bs, seq_length)` - The token ids for each of the sequence. It is padded.
lengths: `torch.tensor(bs)` - The lengths of each of the sequences in the batch.
Output:
-------
token_ids: `torch.tensor(bs, seq_length)` - The token ids after the modifications for MLM.
attn_mask: `torch.tensor(bs, seq_length)` - The attention mask for the self-attention.
clm_labels: `torch.tensor(bs, seq_length)` - The causal language modeling labels. There is a -100 where there is nothing to predict.
"""
token_ids, lengths = batch
token_ids, lengths = self.round_batch(x=token_ids, lengths=lengths)
assert token_ids.size(0) == lengths.size(0)
attn_mask = torch.arange(token_ids.size(1), dtype=torch.long, device=lengths.device) < lengths[:, None]
clm_labels = token_ids.new(token_ids.size()).copy_(token_ids)
clm_labels[~attn_mask] = -100 # previously `clm_labels[1-attn_mask] = -1`, cf pytorch 1.2.0 compatibility
# sanity checks
assert 0 <= token_ids.min() <= token_ids.max() < self.vocab_size
return token_ids, attn_mask, clm_labels
def round_batch(self, x: torch.tensor, lengths: torch.tensor):
"""
For float16 only.
Sub-sample sentences in a batch, and add padding, so that each dimension is a multiple of 8.
Input:
------
x: `torch.tensor(bs, seq_length)` - The token ids.
lengths: `torch.tensor(bs, seq_length)` - The lengths of each of the sequence in the batch.
Output:
-------
x: `torch.tensor(new_bs, new_seq_length)` - The updated token ids.
lengths: `torch.tensor(new_bs, new_seq_length)` - The updated lengths.
"""
if not self.fp16 or len(lengths) < 8:
return x, lengths
# number of sentences == 0 [8]
bs1 = len(lengths)
bs2 = 8 * (bs1 // 8)
assert bs2 > 0 and bs2 % 8 == 0
if bs1 != bs2:
idx = torch.randperm(bs1)[:bs2]
lengths = lengths[idx]
slen = lengths.max().item()
x = x[idx, :slen]
else:
idx = None
# sequence length == 0 [8]
ml1 = x.size(1)
if ml1 % 8 != 0:
pad = 8 - (ml1 % 8)
ml2 = ml1 + pad
if self.mlm:
pad_id = self.params.special_tok_ids["pad_token"]
else:
pad_id = self.params.special_tok_ids["unk_token"]
padding_tensor = torch.zeros(bs2, pad, dtype=torch.long, device=x.device).fill_(pad_id)
x = torch.cat([x, padding_tensor], 1)
assert x.size() == (bs2, ml2)
assert x.size(0) % 8 == 0
assert x.size(1) % 8 == 0
return x, lengths
def train(self):
"""
The real training loop.
"""
if self.is_master:
logger.info("Starting training")
self.last_log = time.time()
self.student.train()
self.teacher.eval()
for _ in range(self.params.n_epoch):
if self.is_master:
logger.info(f"--- Starting epoch {self.epoch}/{self.params.n_epoch-1}")
if self.multi_gpu:
torch.distributed.barrier()
iter_bar = tqdm(self.dataloader, desc="-Iter", disable=self.params.local_rank not in [-1, 0])
for batch in iter_bar:
if self.params.n_gpu > 0:
batch = tuple(t.to(f"cuda:{self.params.local_rank}") for t in batch)
if self.mlm:
token_ids, attn_mask, lm_labels = self.prepare_batch_mlm(batch=batch)
else:
token_ids, attn_mask, lm_labels = self.prepare_batch_clm(batch=batch)
self.step(input_ids=token_ids, attention_mask=attn_mask, lm_labels=lm_labels)
iter_bar.update()
iter_bar.set_postfix(
{"Last_loss": f"{self.last_loss:.2f}", "Avg_cum_loss": f"{self.total_loss_epoch/self.n_iter:.2f}"}
)
iter_bar.close()
if self.is_master:
logger.info(f"--- Ending epoch {self.epoch}/{self.params.n_epoch-1}")
self.end_epoch()
if self.is_master:
logger.info("Save very last checkpoint as `pytorch_model.bin`.")
self.save_checkpoint(checkpoint_name="pytorch_model.bin")
logger.info("Training is finished")
def step(self, input_ids: torch.tensor, attention_mask: torch.tensor, lm_labels: torch.tensor):
"""
One optimization step: forward of student AND teacher, backward on the loss (for gradient accumulation),
and possibly a parameter update (depending on the gradient accumulation).
Input:
------
input_ids: `torch.tensor(bs, seq_length)` - The token ids.
attention_mask: `torch.tensor(bs, seq_length)` - The attention mask for self attention.
lm_labels: `torch.tensor(bs, seq_length)` - The language modeling labels (mlm labels for MLM and clm labels for CLM).
"""
if self.mlm:
student_outputs = self.student(
input_ids=input_ids, attention_mask=attention_mask
) # (bs, seq_length, voc_size)
with torch.no_grad():
teacher_outputs = self.teacher(
input_ids=input_ids, attention_mask=attention_mask
) # (bs, seq_length, voc_size)
else:
student_outputs = self.student(input_ids=input_ids, attention_mask=None) # (bs, seq_length, voc_size)
with torch.no_grad():
teacher_outputs = self.teacher(input_ids=input_ids, attention_mask=None) # (bs, seq_length, voc_size)
s_logits, s_hidden_states = student_outputs["logits"], student_outputs["hidden_states"]
t_logits, t_hidden_states = teacher_outputs["logits"], teacher_outputs["hidden_states"]
assert s_logits.size() == t_logits.size()
# https://github.com/peterliht/knowledge-distillation-pytorch/blob/master/model/net.py#L100
# https://github.com/peterliht/knowledge-distillation-pytorch/issues/2
if self.params.restrict_ce_to_mask:
mask = (lm_labels > -1).unsqueeze(-1).expand_as(s_logits) # (bs, seq_length, voc_size)
else:
mask = attention_mask.unsqueeze(-1).expand_as(s_logits) # (bs, seq_length, voc_size)
s_logits_slct = torch.masked_select(s_logits, mask) # (bs * seq_length * voc_size) modulo the 1s in mask
s_logits_slct = s_logits_slct.view(-1, s_logits.size(-1)) # (bs * seq_length, voc_size) modulo the 1s in mask
t_logits_slct = torch.masked_select(t_logits, mask) # (bs * seq_length * voc_size) modulo the 1s in mask
t_logits_slct = t_logits_slct.view(-1, s_logits.size(-1)) # (bs * seq_length, voc_size) modulo the 1s in mask
assert t_logits_slct.size() == s_logits_slct.size()
loss_ce = (
self.ce_loss_fct(
nn.functional.log_softmax(s_logits_slct / self.temperature, dim=-1),
nn.functional.softmax(t_logits_slct / self.temperature, dim=-1),
)
* (self.temperature) ** 2
)
loss = self.alpha_ce * loss_ce
if self.alpha_mlm > 0.0:
loss_mlm = self.lm_loss_fct(s_logits.view(-1, s_logits.size(-1)), lm_labels.view(-1))
loss += self.alpha_mlm * loss_mlm
if self.alpha_clm > 0.0:
shift_logits = s_logits[..., :-1, :].contiguous()
shift_labels = lm_labels[..., 1:].contiguous()
loss_clm = self.lm_loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
loss += self.alpha_clm * loss_clm
if self.alpha_mse > 0.0:
loss_mse = self.mse_loss_fct(s_logits_slct, t_logits_slct) / s_logits_slct.size(
0
) # Reproducing batchmean reduction
loss += self.alpha_mse * loss_mse
if self.alpha_cos > 0.0:
s_hidden_states = s_hidden_states[-1] # (bs, seq_length, dim)
t_hidden_states = t_hidden_states[-1] # (bs, seq_length, dim)
mask = attention_mask.unsqueeze(-1).expand_as(s_hidden_states) # (bs, seq_length, dim)
assert s_hidden_states.size() == t_hidden_states.size()
dim = s_hidden_states.size(-1)
s_hidden_states_slct = torch.masked_select(s_hidden_states, mask) # (bs * seq_length * dim)
s_hidden_states_slct = s_hidden_states_slct.view(-1, dim) # (bs * seq_length, dim)
t_hidden_states_slct = torch.masked_select(t_hidden_states, mask) # (bs * seq_length * dim)
t_hidden_states_slct = t_hidden_states_slct.view(-1, dim) # (bs * seq_length, dim)
target = s_hidden_states_slct.new(s_hidden_states_slct.size(0)).fill_(1) # (bs * seq_length,)
loss_cos = self.cosine_loss_fct(s_hidden_states_slct, t_hidden_states_slct, target)
loss += self.alpha_cos * loss_cos
self.total_loss_epoch += loss.item()
self.last_loss = loss.item()
self.last_loss_ce = loss_ce.item()
if self.alpha_mlm > 0.0:
self.last_loss_mlm = loss_mlm.item()
if self.alpha_clm > 0.0:
self.last_loss_clm = loss_clm.item()
if self.alpha_mse > 0.0:
self.last_loss_mse = loss_mse.item()
if self.alpha_cos > 0.0:
self.last_loss_cos = loss_cos.item()
self.optimize(loss)
self.n_sequences_epoch += input_ids.size(0)
def optimize(self, loss):
"""
Normalization on the loss (gradient accumulation or distributed training), followed by
backward pass on the loss, possibly followed by a parameter update (depending on the gradient accumulation).
Also update the metrics for tensorboard.
"""
# Check for NaN
if (loss != loss).data.any():
logger.error("NaN detected")
exit()
if self.multi_gpu:
loss = loss.mean()
if self.params.gradient_accumulation_steps > 1:
loss = loss / self.params.gradient_accumulation_steps
if self.fp16:
from apex import amp
with amp.scale_loss(loss, self.optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
self.iter()
if self.n_iter % self.params.gradient_accumulation_steps == 0:
if self.fp16:
nn.utils.clip_grad_norm_(amp.master_params(self.optimizer), self.params.max_grad_norm)
else:
nn.utils.clip_grad_norm_(self.student.parameters(), self.params.max_grad_norm)
self.optimizer.step()
self.optimizer.zero_grad()
self.scheduler.step()
def iter(self):
"""
Update global counts, write to tensorboard and save checkpoint.
"""
self.n_iter += 1
self.n_total_iter += 1
if self.n_total_iter % self.params.log_interval == 0:
self.log_tensorboard()
self.last_log = time.time()
if self.n_total_iter % self.params.checkpoint_interval == 0:
self.save_checkpoint()
def log_tensorboard(self):
"""
Log into tensorboard. Only by the master process.
"""
if not self.is_master:
return
for param_name, param in self.student.named_parameters():
self.tensorboard.add_scalar(
tag="parameter_mean/" + param_name, scalar_value=param.data.mean(), global_step=self.n_total_iter
)
self.tensorboard.add_scalar(
tag="parameter_std/" + param_name, scalar_value=param.data.std(), global_step=self.n_total_iter
)
if param.grad is None:
continue
self.tensorboard.add_scalar(
tag="grad_mean/" + param_name, scalar_value=param.grad.data.mean(), global_step=self.n_total_iter
)
self.tensorboard.add_scalar(
tag="grad_std/" + param_name, scalar_value=param.grad.data.std(), global_step=self.n_total_iter
)
self.tensorboard.add_scalar(
tag="losses/cum_avg_loss_epoch",
scalar_value=self.total_loss_epoch / self.n_iter,
global_step=self.n_total_iter,
)
self.tensorboard.add_scalar(tag="losses/loss", scalar_value=self.last_loss, global_step=self.n_total_iter)
self.tensorboard.add_scalar(
tag="losses/loss_ce", scalar_value=self.last_loss_ce, global_step=self.n_total_iter
)
if self.alpha_mlm > 0.0:
self.tensorboard.add_scalar(
tag="losses/loss_mlm", scalar_value=self.last_loss_mlm, global_step=self.n_total_iter
)
if self.alpha_clm > 0.0:
self.tensorboard.add_scalar(
tag="losses/loss_clm", scalar_value=self.last_loss_clm, global_step=self.n_total_iter
)
if self.alpha_mse > 0.0:
self.tensorboard.add_scalar(
tag="losses/loss_mse", scalar_value=self.last_loss_mse, global_step=self.n_total_iter
)
if self.alpha_cos > 0.0:
self.tensorboard.add_scalar(
tag="losses/loss_cos", scalar_value=self.last_loss_cos, global_step=self.n_total_iter
)
self.tensorboard.add_scalar(
tag="learning_rate/lr", scalar_value=self.scheduler.get_lr()[0], global_step=self.n_total_iter
)
self.tensorboard.add_scalar(
tag="global/memory_usage",
scalar_value=psutil.virtual_memory()._asdict()["used"] / 1_000_000,
global_step=self.n_total_iter,
)
self.tensorboard.add_scalar(
tag="global/speed", scalar_value=time.time() - self.last_log, global_step=self.n_total_iter
)
def end_epoch(self):
"""
Finally arrived at the end of epoch (full pass on dataset).
Do some tensorboard logging and checkpoint saving.
"""
logger.info(f"{self.n_sequences_epoch} sequences have been trained during this epoch.")
if self.is_master:
self.save_checkpoint(checkpoint_name=f"model_epoch_{self.epoch}.pth")
self.tensorboard.add_scalar(
tag="epoch/loss", scalar_value=self.total_loss_epoch / self.n_iter, global_step=self.epoch
)
self.epoch += 1
self.n_sequences_epoch = 0
self.n_iter = 0
self.total_loss_epoch = 0
def save_checkpoint(self, checkpoint_name: str = "checkpoint.pth"):
"""
Save the current state. Only by the master process.
"""
if not self.is_master:
return
mdl_to_save = self.student.module if hasattr(self.student, "module") else self.student
mdl_to_save.config.save_pretrained(self.dump_path)
state_dict = mdl_to_save.state_dict()
torch.save(state_dict, os.path.join(self.dump_path, checkpoint_name))
|
transformers/examples/research_projects/distillation/distiller.py/0
|
{
"file_path": "transformers/examples/research_projects/distillation/distiller.py",
"repo_id": "transformers",
"token_count": 12492
}
| 322
|
<p align="center"> <img src="http://sayef.tech:8082/uploads/FSNER-LOGO-2.png" alt="FSNER LOGO"> </p>
<p align="center">
Implemented by <a href="https://huggingface.co/sayef"> sayef </a>.
</p>
## Overview
The FSNER model was proposed in [Example-Based Named Entity Recognition](https://arxiv.org/abs/2008.10570) by Morteza Ziyadi, Yuting Sun, Abhishek Goswami, Jade Huang, Weizhu Chen. To identify entity spans in a new domain, it uses a train-free few-shot learning approach inspired by question-answering.
## Abstract
----
> We present a novel approach to named entity recognition (NER) in the presence of scarce data that we call example-based NER. Our train-free few-shot learning approach takes inspiration from question-answering to identify entity spans in a new and unseen domain. In comparison with the current state-of-the-art, the proposed method performs significantly better, especially when using a low number of support examples.
## Model Training Details
-----
| identifier | epochs | datasets |
| ---------- |:----------:| :-----:|
| [sayef/fsner-bert-base-uncased](https://huggingface.co/sayef/fsner-bert-base-uncased) | 10 | ontonotes5, conll2003, wnut2017, and fin (Alvarado et al.). |
## Installation and Example Usage
------
You can use the FSNER model in 3 ways:
1. Install directly from PyPI: `pip install fsner` and import the model as shown in the code example below
or
2. Install from source: `python setup.py install` and import the model as shown in the code example below
or
3. Clone repo and change directory to `src` and import the model as shown in the code example below
```python
from fsner import FSNERModel, FSNERTokenizerUtils
model = FSNERModel("sayef/fsner-bert-base-uncased")
tokenizer = FSNERTokenizerUtils("sayef/fsner-bert-base-uncased")
# size of query and supports must be the same. If you want to find all the entitites in one particular query, just repeat the same query n times where n is equal to the number of supports (or entities).
query = [
'KWE 4000 can reach with a maximum speed from up to 450 P/min an accuracy from 50 mg',
'I would like to order a computer from eBay.',
]
# each list in supports are the examples of one entity type
# wrap entities around with [E] and [/E] in the examples
supports = [
[
'Horizontal flow wrapper [E] Pack 403 [/E] features the new retrofit-kit „paper-ON-form“',
'[E] Paloma Pick-and-Place-Roboter [/E] arranges the bakery products for the downstream tray-forming equipment',
'Finally, the new [E] Kliklok ACE [/E] carton former forms cartons and trays without the use of glue',
'We set up our pilot plant with the right [E] FibreForm® [/E] configuration to make prototypes for your marketing tests and package validation',
'The [E] CAR-T5 [/E] is a reliable, purely mechanically driven cartoning machine for versatile application fields'
],
[
"[E] Walmart [/E] is a leading e-commerce company",
"I recently ordered a book from [E] Amazon [/E]",
"I ordered this from [E] ShopClues [/E]",
"[E] Flipkart [/E] started it's journey from zero"
]
]
device = 'cpu'
W_query = tokenizer.tokenize(query).to(device)
W_supports = tokenizer.tokenize(supports).to(device)
start_prob, end_prob = model(W_query, W_supports)
output = tokenizer.extract_entity_from_scores(query, W_query, start_prob, end_prob, thresh=0.50)
print(output)
```
|
transformers/examples/research_projects/fsner/README.md/0
|
{
"file_path": "transformers/examples/research_projects/fsner/README.md",
"repo_id": "transformers",
"token_count": 1174
}
| 323
|
import json
import os
from dataclasses import dataclass
from functools import partial
from typing import Callable
import flax.linen as nn
import jax
import jax.numpy as jnp
import joblib
import optax
import wandb
from flax import jax_utils, struct, traverse_util
from flax.serialization import from_bytes, to_bytes
from flax.training import train_state
from flax.training.common_utils import shard
from tqdm.auto import tqdm
from transformers import BigBirdConfig, FlaxBigBirdForQuestionAnswering
from transformers.models.big_bird.modeling_flax_big_bird import FlaxBigBirdForQuestionAnsweringModule
class FlaxBigBirdForNaturalQuestionsModule(FlaxBigBirdForQuestionAnsweringModule):
"""
BigBirdForQuestionAnswering with CLS Head over the top for predicting category
This way we can load its weights with FlaxBigBirdForQuestionAnswering
"""
config: BigBirdConfig
dtype: jnp.dtype = jnp.float32
add_pooling_layer: bool = True
def setup(self):
super().setup()
self.cls = nn.Dense(5, dtype=self.dtype)
def __call__(self, *args, **kwargs):
outputs = super().__call__(*args, **kwargs)
cls_out = self.cls(outputs[2])
return outputs[:2] + (cls_out,)
class FlaxBigBirdForNaturalQuestions(FlaxBigBirdForQuestionAnswering):
module_class = FlaxBigBirdForNaturalQuestionsModule
def calculate_loss_for_nq(start_logits, start_labels, end_logits, end_labels, pooled_logits, pooler_labels):
def cross_entropy(logits, labels, reduction=None):
"""
Args:
logits: bsz, seqlen, vocab_size
labels: bsz, seqlen
"""
vocab_size = logits.shape[-1]
labels = (labels[..., None] == jnp.arange(vocab_size)[None]).astype("f4")
logits = jax.nn.log_softmax(logits, axis=-1)
loss = -jnp.sum(labels * logits, axis=-1)
if reduction is not None:
loss = reduction(loss)
return loss
cross_entropy = partial(cross_entropy, reduction=jnp.mean)
start_loss = cross_entropy(start_logits, start_labels)
end_loss = cross_entropy(end_logits, end_labels)
pooled_loss = cross_entropy(pooled_logits, pooler_labels)
return (start_loss + end_loss + pooled_loss) / 3
@dataclass
class Args:
model_id: str = "google/bigbird-roberta-base"
logging_steps: int = 3000
save_steps: int = 10500
block_size: int = 128
num_random_blocks: int = 3
batch_size_per_device: int = 1
max_epochs: int = 5
# tx_args
lr: float = 3e-5
init_lr: float = 0.0
warmup_steps: int = 20000
weight_decay: float = 0.0095
save_dir: str = "bigbird-roberta-natural-questions"
base_dir: str = "training-expt"
tr_data_path: str = "data/nq-training.jsonl"
val_data_path: str = "data/nq-validation.jsonl"
def __post_init__(self):
os.makedirs(self.base_dir, exist_ok=True)
self.save_dir = os.path.join(self.base_dir, self.save_dir)
self.batch_size = self.batch_size_per_device * jax.device_count()
@dataclass
class DataCollator:
pad_id: int
max_length: int = 4096 # no dynamic padding on TPUs
def __call__(self, batch):
batch = self.collate_fn(batch)
batch = jax.tree_util.tree_map(shard, batch)
return batch
def collate_fn(self, features):
input_ids, attention_mask = self.fetch_inputs(features["input_ids"])
batch = {
"input_ids": jnp.array(input_ids, dtype=jnp.int32),
"attention_mask": jnp.array(attention_mask, dtype=jnp.int32),
"start_labels": jnp.array(features["start_token"], dtype=jnp.int32),
"end_labels": jnp.array(features["end_token"], dtype=jnp.int32),
"pooled_labels": jnp.array(features["category"], dtype=jnp.int32),
}
return batch
def fetch_inputs(self, input_ids: list):
inputs = [self._fetch_inputs(ids) for ids in input_ids]
return zip(*inputs)
def _fetch_inputs(self, input_ids: list):
attention_mask = [1 for _ in range(len(input_ids))]
while len(input_ids) < self.max_length:
input_ids.append(self.pad_id)
attention_mask.append(0)
return input_ids, attention_mask
def get_batched_dataset(dataset, batch_size, seed=None):
if seed is not None:
dataset = dataset.shuffle(seed=seed)
for i in range(len(dataset) // batch_size):
batch = dataset[i * batch_size : (i + 1) * batch_size]
yield dict(batch)
@partial(jax.pmap, axis_name="batch")
def train_step(state, drp_rng, **model_inputs):
def loss_fn(params):
start_labels = model_inputs.pop("start_labels")
end_labels = model_inputs.pop("end_labels")
pooled_labels = model_inputs.pop("pooled_labels")
outputs = state.apply_fn(**model_inputs, params=params, dropout_rng=drp_rng, train=True)
start_logits, end_logits, pooled_logits = outputs
return state.loss_fn(
start_logits,
start_labels,
end_logits,
end_labels,
pooled_logits,
pooled_labels,
)
drp_rng, new_drp_rng = jax.random.split(drp_rng)
grad_fn = jax.value_and_grad(loss_fn)
loss, grads = grad_fn(state.params)
metrics = jax.lax.pmean({"loss": loss}, axis_name="batch")
grads = jax.lax.pmean(grads, "batch")
state = state.apply_gradients(grads=grads)
return state, metrics, new_drp_rng
@partial(jax.pmap, axis_name="batch")
def val_step(state, **model_inputs):
start_labels = model_inputs.pop("start_labels")
end_labels = model_inputs.pop("end_labels")
pooled_labels = model_inputs.pop("pooled_labels")
outputs = state.apply_fn(**model_inputs, params=state.params, train=False)
start_logits, end_logits, pooled_logits = outputs
loss = state.loss_fn(start_logits, start_labels, end_logits, end_labels, pooled_logits, pooled_labels)
metrics = jax.lax.pmean({"loss": loss}, axis_name="batch")
return metrics
class TrainState(train_state.TrainState):
loss_fn: Callable = struct.field(pytree_node=False)
@dataclass
class Trainer:
args: Args
data_collator: Callable
train_step_fn: Callable
val_step_fn: Callable
model_save_fn: Callable
logger: wandb
scheduler_fn: Callable = None
def create_state(self, model, tx, num_train_steps, ckpt_dir=None):
params = model.params
state = TrainState.create(
apply_fn=model.__call__,
params=params,
tx=tx,
loss_fn=calculate_loss_for_nq,
)
if ckpt_dir is not None:
params, opt_state, step, args, data_collator = restore_checkpoint(ckpt_dir, state)
tx_args = {
"lr": args.lr,
"init_lr": args.init_lr,
"warmup_steps": args.warmup_steps,
"num_train_steps": num_train_steps,
"weight_decay": args.weight_decay,
}
tx, lr = build_tx(**tx_args)
state = train_state.TrainState(
step=step,
apply_fn=model.__call__,
params=params,
tx=tx,
opt_state=opt_state,
)
self.args = args
self.data_collator = data_collator
self.scheduler_fn = lr
model.params = params
state = jax_utils.replicate(state)
return state
def train(self, state, tr_dataset, val_dataset):
args = self.args
total = len(tr_dataset) // args.batch_size
rng = jax.random.PRNGKey(0)
drp_rng = jax.random.split(rng, jax.device_count())
for epoch in range(args.max_epochs):
running_loss = jnp.array(0, dtype=jnp.float32)
tr_dataloader = get_batched_dataset(tr_dataset, args.batch_size, seed=epoch)
i = 0
for batch in tqdm(tr_dataloader, total=total, desc=f"Running EPOCH-{epoch}"):
batch = self.data_collator(batch)
state, metrics, drp_rng = self.train_step_fn(state, drp_rng, **batch)
running_loss += jax_utils.unreplicate(metrics["loss"])
i += 1
if i % args.logging_steps == 0:
state_step = jax_utils.unreplicate(state.step)
tr_loss = running_loss.item() / i
lr = self.scheduler_fn(state_step - 1)
eval_loss = self.evaluate(state, val_dataset)
logging_dict = {
"step": state_step.item(),
"eval_loss": eval_loss.item(),
"tr_loss": tr_loss,
"lr": lr.item(),
}
tqdm.write(str(logging_dict))
self.logger.log(logging_dict, commit=True)
if i % args.save_steps == 0:
self.save_checkpoint(args.save_dir + f"-e{epoch}-s{i}", state=state)
def evaluate(self, state, dataset):
dataloader = get_batched_dataset(dataset, self.args.batch_size)
total = len(dataset) // self.args.batch_size
running_loss = jnp.array(0, dtype=jnp.float32)
i = 0
for batch in tqdm(dataloader, total=total, desc="Evaluating ... "):
batch = self.data_collator(batch)
metrics = self.val_step_fn(state, **batch)
running_loss += jax_utils.unreplicate(metrics["loss"])
i += 1
return running_loss / i
def save_checkpoint(self, save_dir, state):
state = jax_utils.unreplicate(state)
print(f"SAVING CHECKPOINT IN {save_dir}", end=" ... ")
self.model_save_fn(save_dir, params=state.params)
with open(os.path.join(save_dir, "opt_state.msgpack"), "wb") as f:
f.write(to_bytes(state.opt_state))
joblib.dump(self.args, os.path.join(save_dir, "args.joblib"))
joblib.dump(self.data_collator, os.path.join(save_dir, "data_collator.joblib"))
with open(os.path.join(save_dir, "training_state.json"), "w") as f:
json.dump({"step": state.step.item()}, f)
print("DONE")
def restore_checkpoint(save_dir, state):
print(f"RESTORING CHECKPOINT FROM {save_dir}", end=" ... ")
with open(os.path.join(save_dir, "flax_model.msgpack"), "rb") as f:
params = from_bytes(state.params, f.read())
with open(os.path.join(save_dir, "opt_state.msgpack"), "rb") as f:
opt_state = from_bytes(state.opt_state, f.read())
args = joblib.load(os.path.join(save_dir, "args.joblib"))
data_collator = joblib.load(os.path.join(save_dir, "data_collator.joblib"))
with open(os.path.join(save_dir, "training_state.json"), "r") as f:
training_state = json.load(f)
step = training_state["step"]
print("DONE")
return params, opt_state, step, args, data_collator
def scheduler_fn(lr, init_lr, warmup_steps, num_train_steps):
decay_steps = num_train_steps - warmup_steps
warmup_fn = optax.linear_schedule(init_value=init_lr, end_value=lr, transition_steps=warmup_steps)
decay_fn = optax.linear_schedule(init_value=lr, end_value=1e-7, transition_steps=decay_steps)
lr = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[warmup_steps])
return lr
def build_tx(lr, init_lr, warmup_steps, num_train_steps, weight_decay):
def weight_decay_mask(params):
params = traverse_util.flatten_dict(params)
mask = {k: (v[-1] != "bias" and v[-2:] != ("LayerNorm", "scale")) for k, v in params.items()}
return traverse_util.unflatten_dict(mask)
lr = scheduler_fn(lr, init_lr, warmup_steps, num_train_steps)
tx = optax.adamw(learning_rate=lr, weight_decay=weight_decay, mask=weight_decay_mask)
return tx, lr
|
transformers/examples/research_projects/jax-projects/big_bird/bigbird_flax.py/0
|
{
"file_path": "transformers/examples/research_projects/jax-projects/big_bird/bigbird_flax.py",
"repo_id": "transformers",
"token_count": 5480
}
| 324
|
# Wav2Vec2 Contrastive Loss PreTraining examples
The following example showcases how to pretrain a wav2vec2 model using the JAX/Flax backend.
Pretraining Wav2Vec2 is rather complex, so it is highly recommended to read the
[official paper](https://arxiv.org/abs/2006.11477).
JAX/Flax allows you to trace pure functions and compile them into efficient, fused accelerator code on both GPU and TPU.
Models written in JAX/Flax are **immutable** and updated in a purely functional
way which enables simple and efficient model parallelism.
`run_wav2vec2_pretrain_flax.py` is a lightweight example of how to download and preprocess a dataset from the 🤗 Datasets library or use your own files (jsonlines or csv), then pretrain the wav2vec2 architectures above on it.
For custom datasets in `jsonlines` format please see: [the Datasets documentation](https://huggingface.co/docs/datasets/loading_datasets#json-files) and you also will find examples of these below.
Let's start by creating a model repository to save the trained model and logs.
Here we call the model `"wav2vec2-base-robust"`, but you can change the model name as you like.
You can do this either directly on [huggingface.co](https://huggingface.co/new) (assuming that
you are logged in) or via the command line:
```bash
huggingface-cli repo create wav2vec2-base-robust
```
Next we clone the model repository to add the tokenizer and model files.
```bash
git clone https://huggingface.co/<your-username>/wav2vec2-base-robust
```
To ensure that all tensorboard traces will be uploaded correctly, we need to
track them. You can run the following command inside your model repo to do so.
```bash
cd wav2vec2-base-robust
git lfs track "*tfevents*"
```
Great, we have set up our model repository. During training, we will automatically
push the training logs and model weights to the repo.
Next, let's add a symbolic link to the `run_wav2vec2_pretrain_flax`.
```bash
export MODEL_DIR="./wav2vec2-base-robust"
ln -s ~/transformers/examples/research_projects/jax-projects/wav2vec2/run_wav2vec2_pretrain_flax.py ./
```
### Create the model configuration
Let's first create the model configuration and store it in the model repository.
Note that many training parameters can be set in the model configuration including
the configuration about the masking distribution (`mask_time_length`, `mask_time_prob`),
dropout (`attention_dropout`, ...), the trade-off between the contrastive loss and
the diversity loss, etc...
Mostly likely you will need to change these parameters depending on your use case.
Again, we highly recommend to read the [official paper](https://arxiv.org/abs/2006.11477)
to better understand which parameters can be set for pretraining.
For this example, we will be using a `"base"`-sized model of Wav2Vec2 with robust
layer norm and keep most of the default settings.
```python
model_dir="./wav2vec2-base-robust"
from transformers import Wav2Vec2Config
config = Wav2Vec2Config.from_pretrained(
"facebook/wav2vec2-base",
mask_time_length=10,
mask_time_prob=0.05,
diversity_loss_weight=0.1,
num_negatives=100,
do_stable_layer_norm=True,
feat_extract_norm="layer",
)
config.save_pretrained(model_dir)
```
### Create a feature extractor configuration
Before we can start the training, we need to define
a feature extractor that takes care of normalization, etc...
Here we can also re-use the feature extractor of [wav2vec2-base-960h](https://huggingface.co/facebook/wav2vec2-base) while making sure that padding is allowed.
```python
model_dir="./wav2vec2-base-robust"
from transformers import Wav2Vec2FeatureExtractor
config = Wav2Vec2FeatureExtractor.from_pretrained("facebook/wav2vec2-base", return_attention_mask=True)
config.save_pretrained(model_dir)
```
### Train the model
Finally, we can run the example script to train the model:
```bash
./run_wav2vec2_pretrain_flax.py \
--output_dir=${MODEL_DIR} \
--num_train_epochs="5" \
--per_device_train_batch_size="32" \
--per_device_eval_batch_size="32" \
--learning_rate="5e-4" \
--weight_decay="0.01" \
--warmup_steps="2000" \
--model_name_or_path=${MODEL_DIR} \
--dataset_name="librispeech_asr" \
--dataset_config_name="clean" \
--train_split_name="train.100" \
--preprocessing_num_workers="4" \
--max_duration_in_seconds="10.0" \
--adam_beta1="0.9" \
--adam_beta2="0.98" \
--pad_to_multiple_of="16384" \
--push_to_hub
```
Note that this script is not fully tested yet, so we cannot ensure that
the above script leads to satisfying results.
|
transformers/examples/research_projects/jax-projects/wav2vec2/README.md/0
|
{
"file_path": "transformers/examples/research_projects/jax-projects/wav2vec2/README.md",
"repo_id": "transformers",
"token_count": 1507
}
| 325
|
"""
coding=utf-8
Copyright 2018, Antonio Mendoza Hao Tan, Mohit Bansal
Adapted From Facebook Inc, Detectron2
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 sys
from typing import Tuple
import numpy as np
import torch
from PIL import Image
from torch import nn
from transformers.image_utils import PILImageResampling
from utils import img_tensorize
class ResizeShortestEdge:
def __init__(self, short_edge_length, max_size=sys.maxsize):
"""
Args:
short_edge_length (list[min, max])
max_size (int): maximum allowed longest edge length.
"""
self.interp_method = "bilinear"
self.max_size = max_size
self.short_edge_length = short_edge_length
def __call__(self, imgs):
img_augs = []
for img in imgs:
h, w = img.shape[:2]
# later: provide list and randomly choose index for resize
size = np.random.randint(self.short_edge_length[0], self.short_edge_length[1] + 1)
if size == 0:
return img
scale = size * 1.0 / min(h, w)
if h < w:
newh, neww = size, scale * w
else:
newh, neww = scale * h, size
if max(newh, neww) > self.max_size:
scale = self.max_size * 1.0 / max(newh, neww)
newh = newh * scale
neww = neww * scale
neww = int(neww + 0.5)
newh = int(newh + 0.5)
if img.dtype == np.uint8:
pil_image = Image.fromarray(img)
pil_image = pil_image.resize((neww, newh), PILImageResampling.BILINEAR)
img = np.asarray(pil_image)
else:
img = img.permute(2, 0, 1).unsqueeze(0) # 3, 0, 1) # hw(c) -> nchw
img = nn.functional.interpolate(
img, (newh, neww), mode=self.interp_method, align_corners=False
).squeeze(0)
img_augs.append(img)
return img_augs
class Preprocess:
def __init__(self, cfg):
self.aug = ResizeShortestEdge([cfg.INPUT.MIN_SIZE_TEST, cfg.INPUT.MIN_SIZE_TEST], cfg.INPUT.MAX_SIZE_TEST)
self.input_format = cfg.INPUT.FORMAT
self.size_divisibility = cfg.SIZE_DIVISIBILITY
self.pad_value = cfg.PAD_VALUE
self.max_image_size = cfg.INPUT.MAX_SIZE_TEST
self.device = cfg.MODEL.DEVICE
self.pixel_std = torch.tensor(cfg.MODEL.PIXEL_STD).to(self.device).view(len(cfg.MODEL.PIXEL_STD), 1, 1)
self.pixel_mean = torch.tensor(cfg.MODEL.PIXEL_MEAN).to(self.device).view(len(cfg.MODEL.PIXEL_STD), 1, 1)
self.normalizer = lambda x: (x - self.pixel_mean) / self.pixel_std
def pad(self, images):
max_size = tuple(max(s) for s in zip(*[img.shape for img in images]))
image_sizes = [im.shape[-2:] for im in images]
images = [
nn.functional.pad(
im,
[0, max_size[-1] - size[1], 0, max_size[-2] - size[0]],
value=self.pad_value,
)
for size, im in zip(image_sizes, images)
]
return torch.stack(images), torch.tensor(image_sizes)
def __call__(self, images, single_image=False):
with torch.no_grad():
if not isinstance(images, list):
images = [images]
if single_image:
assert len(images) == 1
for i in range(len(images)):
if isinstance(images[i], torch.Tensor):
images.insert(i, images.pop(i).to(self.device).float())
elif not isinstance(images[i], torch.Tensor):
images.insert(
i,
torch.as_tensor(img_tensorize(images.pop(i), input_format=self.input_format))
.to(self.device)
.float(),
)
# resize smallest edge
raw_sizes = torch.tensor([im.shape[:2] for im in images])
images = self.aug(images)
# transpose images and convert to torch tensors
# images = [torch.as_tensor(i.astype("float32")).permute(2, 0, 1).to(self.device) for i in images]
# now normalize before pad to avoid useless arithmetic
images = [self.normalizer(x) for x in images]
# now pad them to do the following operations
images, sizes = self.pad(images)
# Normalize
if self.size_divisibility > 0:
raise NotImplementedError()
# pad
scales_yx = torch.true_divide(raw_sizes, sizes)
if single_image:
return images[0], sizes[0], scales_yx[0]
else:
return images, sizes, scales_yx
def _scale_box(boxes, scale_yx):
boxes[:, 0::2] *= scale_yx[:, 1]
boxes[:, 1::2] *= scale_yx[:, 0]
return boxes
def _clip_box(tensor, box_size: Tuple[int, int]):
assert torch.isfinite(tensor).all(), "Box tensor contains infinite or NaN!"
h, w = box_size
tensor[:, 0].clamp_(min=0, max=w)
tensor[:, 1].clamp_(min=0, max=h)
tensor[:, 2].clamp_(min=0, max=w)
tensor[:, 3].clamp_(min=0, max=h)
|
transformers/examples/research_projects/lxmert/processing_image.py/0
|
{
"file_path": "transformers/examples/research_projects/lxmert/processing_image.py",
"repo_id": "transformers",
"token_count": 2801
}
| 326
|
# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION. 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.
"""Masked BERT model configuration. It replicates the class `~transformers.BertConfig`
and adapts it to the specificities of MaskedBert (`pruning_method`, `mask_init` and `mask_scale`."""
import logging
from transformers.configuration_utils import PretrainedConfig
logger = logging.getLogger(__name__)
class MaskedBertConfig(PretrainedConfig):
"""
A class replicating the `~transformers.BertConfig` with additional parameters for pruning/masking configuration.
"""
model_type = "masked_bert"
def __init__(
self,
vocab_size=30522,
hidden_size=768,
num_hidden_layers=12,
num_attention_heads=12,
intermediate_size=3072,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=512,
type_vocab_size=2,
initializer_range=0.02,
layer_norm_eps=1e-12,
pad_token_id=0,
pruning_method="topK",
mask_init="constant",
mask_scale=0.0,
**kwargs,
):
super().__init__(pad_token_id=pad_token_id, **kwargs)
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.hidden_act = hidden_act
self.intermediate_size = intermediate_size
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.type_vocab_size = type_vocab_size
self.initializer_range = initializer_range
self.layer_norm_eps = layer_norm_eps
self.pruning_method = pruning_method
self.mask_init = mask_init
self.mask_scale = mask_scale
|
transformers/examples/research_projects/movement-pruning/emmental/configuration_bert_masked.py/0
|
{
"file_path": "transformers/examples/research_projects/movement-pruning/emmental/configuration_bert_masked.py",
"repo_id": "transformers",
"token_count": 995
}
| 327
|
# coding=utf-8
# Copyright 2020 The Google Research Authors.
#
# 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.
"""
IMPORTANT:
This code was copied from
https://github.com/google-research/google-research/blob/master/performer/fast_self_attention/fast_self_attention.py on
6/11/2020. This is very new code, so it might be prone to change soon -> make sure to check the original code and
update accordingly
Core Fast Attention Module for Flax. Implementation of the approximate fast softmax and generalized attention mechanism
leveraging structured random feature maps [RFM] techniques and low rank decomposition of the attention matrix.
"""
# pylint: disable=invalid-name, missing-function-docstring, line-too-long
import abc
import functools
from collections.abc import Iterable # pylint: disable=g-importing-member
import jax
import jax.numpy as jnp
import numpy as onp
from absl import logging
from jax import lax, random
def nonnegative_softmax_kernel_feature_creator(
data, projection_matrix, attention_dims_t, batch_dims_t, precision, is_query, normalize_data=True, eps=0.0001
):
"""
Constructs nonnegative kernel features for fast softmax attention
Args:
data: input for which features are computes
projection_matrix: random matrix used to compute features
attention_dims_t: tuple of attention dimensions
batch_dims_t: tuple of batch dimensions
precision: precision parameter
is_query: predicate indicating whether input data corresponds to queries or
keys
normalize_data: predicate indicating whether data should be normalized,
eps: numerical stabilizer
Returns:
Random features for fast softmax attention.
"""
del attention_dims_t
if normalize_data:
# We have e^{qk^T/sqrt{d}} = e^{q_norm k_norm^T}, where
# w_norm = w * data_normalizer for w in {q,k}.
data_normalizer = 1.0 / (jnp.sqrt(jnp.sqrt(data.shape[-1])))
else:
data_normalizer = 1.0
ratio = 1.0 / jnp.sqrt(projection_matrix.shape[0])
data_mod_shape = data.shape[0 : len(batch_dims_t)] + projection_matrix.shape
data_thick_random_matrix = jnp.zeros(data_mod_shape) + projection_matrix
data_dash = lax.dot_general(
data_normalizer * data,
data_thick_random_matrix,
(((data.ndim - 1,), (data_thick_random_matrix.ndim - 1,)), (batch_dims_t, batch_dims_t)),
precision=precision,
)
diag_data = jnp.square(data)
diag_data = jnp.sum(diag_data, axis=data.ndim - 1)
diag_data = (diag_data / 2.0) * data_normalizer * data_normalizer
diag_data = jnp.expand_dims(diag_data, axis=data.ndim - 1)
if is_query:
last_dims_t = (len(data_dash.shape) - 1,)
data_dash = ratio * (
jnp.exp(data_dash - diag_data - jnp.max(data_dash, axis=last_dims_t, keepdims=True)) + eps
)
else:
data_dash = ratio * (jnp.exp(data_dash - diag_data - jnp.max(data_dash)) + eps)
return data_dash
def sincos_softmax_kernel_feature_creator(
data, projection_matrix, attention_dims_t, batch_dims_t, precision, normalize_data=True
):
"""
Constructs kernel sin-cos features for fast softmax attention
Args:
data: input for which features are computes
projection_matrix: random matrix used to compute features
attention_dims_t: tuple of attention dimensions
batch_dims_t: tuple of batch dimensions
precision: precision parameter
normalize_data: predicate indicating whether data should be normalized
Returns:
Random features for fast softmax attention.
"""
if normalize_data:
# We have: exp(qk^T/sqrt{d}) = exp(|q|^2/2sqrt{d}) * exp(|k|^2/2sqrt{d}) *
# exp(-(|q*c-k*c|^2)/2), where c = 1.0 / sqrt{sqrt{d}}.
data_normalizer = 1.0 / (jnp.sqrt(jnp.sqrt(data.shape[-1])))
else:
data_normalizer = 1.0
ratio = 1.0 / jnp.sqrt(projection_matrix.shape[0])
data_mod_shape = data.shape[0 : len(batch_dims_t)] + projection_matrix.shape
data_thick_random_matrix = jnp.zeros(data_mod_shape) + projection_matrix
data_dash = lax.dot_general(
data_normalizer * data,
data_thick_random_matrix,
(((data.ndim - 1,), (data_thick_random_matrix.ndim - 1,)), (batch_dims_t, batch_dims_t)),
precision=precision,
)
data_dash_cos = ratio * jnp.cos(data_dash)
data_dash_sin = ratio * jnp.sin(data_dash)
data_dash = jnp.concatenate((data_dash_cos, data_dash_sin), axis=-1)
# Constructing D_data and data^{'}
diag_data = jnp.square(data)
diag_data = jnp.sum(diag_data, axis=data.ndim - 1)
diag_data = (diag_data / 2.0) * data_normalizer * data_normalizer
diag_data = jnp.expand_dims(diag_data, axis=data.ndim - 1)
# Additional renormalization for numerical stability
data_renormalizer = jnp.max(diag_data, attention_dims_t, keepdims=True)
diag_data -= data_renormalizer
diag_data = jnp.exp(diag_data)
data_prime = data_dash * diag_data
return data_prime
def generalized_kernel_feature_creator(
data, projection_matrix, batch_dims_t, precision, kernel_fn, kernel_epsilon, normalize_data
):
"""
Constructs kernel features for fast generalized attention
Args:
data: input for which features are computes
projection_matrix: matrix used to compute features
batch_dims_t: tuple of batch dimensions
precision: precision parameter
kernel_fn: kernel function used
kernel_epsilon: additive positive term added to every feature for numerical
stability
normalize_data: predicate indicating whether data should be normalized
Returns:
Random features for fast generalized attention.
"""
if normalize_data:
data_normalizer = 1.0 / (jnp.sqrt(jnp.sqrt(data.shape[-1])))
else:
data_normalizer = 1.0
if projection_matrix is None:
return kernel_fn(data_normalizer * data) + kernel_epsilon
else:
data_mod_shape = data.shape[0 : len(batch_dims_t)] + projection_matrix.shape
data_thick_random_matrix = jnp.zeros(data_mod_shape) + projection_matrix
data_dash = lax.dot_general(
data_normalizer * data,
data_thick_random_matrix,
(((data.ndim - 1,), (data_thick_random_matrix.ndim - 1,)), (batch_dims_t, batch_dims_t)),
precision=precision,
)
data_prime = kernel_fn(data_dash) + kernel_epsilon
return data_prime
def make_fast_softmax_attention(
qkv_dim,
renormalize_attention=True,
numerical_stabilizer=0.000001,
nb_features=256,
ortho_features=True,
ortho_scaling=0.0,
redraw_features=True,
unidirectional=False,
nonnegative_features=True,
lax_scan_unroll=1,
):
"""Construct a fast softmax attention method."""
logging.info(
"Fast softmax attention: %s features and orthogonal=%s, renormalize=%s",
nb_features,
ortho_features,
renormalize_attention,
)
if ortho_features:
matrix_creator = functools.partial(GaussianOrthogonalRandomMatrix, nb_features, qkv_dim, scaling=ortho_scaling)
else:
matrix_creator = functools.partial(GaussianUnstructuredRandomMatrix, nb_features, qkv_dim)
if nonnegative_features:
def kernel_feature_creator(
data, projection_matrix, attention_dims_t, batch_dims_t, precision, is_query, normalize_data=True
):
return nonnegative_softmax_kernel_feature_creator(
data,
projection_matrix,
attention_dims_t,
batch_dims_t,
precision,
is_query,
normalize_data,
numerical_stabilizer,
)
else:
def kernel_feature_creator(
data, projection_matrix, attention_dims_t, batch_dims_t, precision, is_query, normalize_data=True
):
del is_query
return sincos_softmax_kernel_feature_creator(
data, projection_matrix, attention_dims_t, batch_dims_t, precision, normalize_data
)
attention_fn = FastAttentionviaLowRankDecomposition(
matrix_creator,
kernel_feature_creator,
renormalize_attention=renormalize_attention,
numerical_stabilizer=numerical_stabilizer,
redraw_features=redraw_features,
unidirectional=unidirectional,
lax_scan_unroll=lax_scan_unroll,
).dot_product_attention
return attention_fn
def make_fast_generalized_attention(
qkv_dim,
renormalize_attention=True,
numerical_stabilizer=0.0,
nb_features=256,
features_type="deterministic",
kernel_fn=jax.nn.relu,
kernel_epsilon=0.001,
redraw_features=False,
unidirectional=False,
lax_scan_unroll=1,
):
"""Construct a fast generalized attention menthod."""
logging.info("Fast generalized attention.: %s features and renormalize=%s", nb_features, renormalize_attention)
if features_type == "ortho":
matrix_creator = functools.partial(GaussianOrthogonalRandomMatrix, nb_features, qkv_dim, scaling=False)
elif features_type == "iid":
matrix_creator = functools.partial(GaussianUnstructuredRandomMatrix, nb_features, qkv_dim)
elif features_type == "deterministic":
matrix_creator = None
else:
raise ValueError("Unknown feature value type")
def kernel_feature_creator(
data, projection_matrix, attention_dims_t, batch_dims_t, precision, is_query, normalize_data=False
):
del attention_dims_t
del is_query
return generalized_kernel_feature_creator(
data, projection_matrix, batch_dims_t, precision, kernel_fn, kernel_epsilon, normalize_data
)
attention_fn = FastAttentionviaLowRankDecomposition(
matrix_creator,
kernel_feature_creator,
renormalize_attention=renormalize_attention,
numerical_stabilizer=numerical_stabilizer,
redraw_features=redraw_features,
unidirectional=unidirectional,
lax_scan_unroll=lax_scan_unroll,
).dot_product_attention
return attention_fn
class RandomMatrix:
r"""
Abstract class providing a method for constructing 2D random arrays. Class is responsible for constructing 2D
random arrays.
"""
__metaclass__ = abc.ABCMeta
@abc.abstractmethod
def get_2d_array(self):
raise NotImplementedError("Abstract method")
class GaussianUnstructuredRandomMatrix(RandomMatrix):
def __init__(self, nb_rows, nb_columns, key):
self.nb_rows = nb_rows
self.nb_columns = nb_columns
self.key = key
def get_2d_array(self):
return random.normal(self.key, (self.nb_rows, self.nb_columns))
class GaussianOrthogonalRandomMatrix(RandomMatrix):
r"""
Class providing a method to create Gaussian orthogonal matrix. Class is responsible for constructing 2D Gaussian
orthogonal arrays.
"""
def __init__(self, nb_rows, nb_columns, key, scaling=0):
self.nb_rows = nb_rows
self.nb_columns = nb_columns
self.key = key
self.scaling = scaling
def get_2d_array(self):
nb_full_blocks = int(self.nb_rows / self.nb_columns)
block_list = []
rng = self.key
for _ in range(nb_full_blocks):
rng, rng_input = jax.random.split(rng)
unstructured_block = random.normal(rng_input, (self.nb_columns, self.nb_columns))
q, _ = jnp.linalg.qr(unstructured_block)
q = jnp.transpose(q)
block_list.append(q)
remaining_rows = self.nb_rows - nb_full_blocks * self.nb_columns
if remaining_rows > 0:
rng, rng_input = jax.random.split(rng)
unstructured_block = random.normal(rng_input, (self.nb_columns, self.nb_columns))
q, _ = jnp.linalg.qr(unstructured_block)
q = jnp.transpose(q)
block_list.append(q[0:remaining_rows])
final_matrix = jnp.vstack(block_list)
if self.scaling == 0:
multiplier = jnp.linalg.norm(random.normal(self.key, (self.nb_rows, self.nb_columns)), axis=1)
elif self.scaling == 1:
multiplier = jnp.sqrt(float(self.nb_columns)) * jnp.ones((self.nb_rows))
else:
raise ValueError("Scaling must be one of {0, 1}. Was %s" % self._scaling)
return jnp.matmul(jnp.diag(multiplier), final_matrix)
class FastAttention:
r"""
Abstract class providing a method for fast attention. Class is responsible for providing a method
<dot_product_attention> for fast approximate attention.
"""
__metaclass__ = abc.ABCMeta
@abc.abstractmethod
def dot_product_attention(
self,
query,
key,
value,
dtype=jnp.float32,
bias=None,
axis=None,
broadcast_dropout=True,
dropout_rng=None,
dropout_rate=0.0,
deterministic=False,
precision=None,
):
"""
Computes dot-product attention given query, key, and value. This is the core function for applying fast
approximate dot-product attention. It calculates the attention weights given query and key and combines the
values using the attention weights. This function supports multi-dimensional inputs
Args:
query: queries for calculating attention with shape of [batch_size, dim1,
dim2, ..., dimN, num_heads, mem_channels].
key: keys for calculating attention with shape of [batch_size, dim1, dim2,
..., dimN, num_heads, mem_channels].
value: values to be used in attention with shape of [batch_size, dim1,
dim2,..., dimN, num_heads, value_channels].
dtype: the dtype of the computation (default: float32)
bias: bias for the attention weights. This can be used for incorporating
autoregressive mask, padding mask, proximity bias.
axis: axises over which the attention is applied.
broadcast_dropout: bool: use a broadcasted dropout along batch dims.
dropout_rng: JAX PRNGKey: to be used for dropout.
dropout_rate: dropout rate.
deterministic: bool, deterministic or not (to apply dropout).
precision: numerical precision of the computation see `jax.lax.Precision`
for details
Returns:
Output of shape [bs, dim1, dim2, ..., dimN,, num_heads, value_channels].
"""
raise NotImplementedError("Abstract method")
def _numerator(z_slice_shape, precision, unroll=1):
def fwd(qs, ks, vs):
def body(p, qkv):
(q, k, v) = qkv
p += jnp.einsum("...m,...d->...md", k, v, precision=precision)
X_slice = jnp.einsum("...m,...md->...d", q, p, precision=precision)
return p, X_slice
init_value = jnp.zeros(z_slice_shape)
p, W = lax.scan(body, init_value, (qs, ks, vs), unroll=unroll)
return W, (p, qs, ks, vs)
def bwd(pqkv, W_ct):
def body(carry, qkv_xct):
p, p_ct = carry
q, k, v, x_ct = qkv_xct
q_ct = jnp.einsum("...d,...md->...m", x_ct, p, precision=precision)
p_ct += jnp.einsum("...d,...m->...md", x_ct, q, precision=precision)
k_ct = jnp.einsum("...md,...d->...m", p_ct, v, precision=precision)
v_ct = jnp.einsum("...md,...m->...d", p_ct, k, precision=precision)
p -= jnp.einsum("...m,...d->...md", k, v, precision=precision)
return (p, p_ct), (q_ct, k_ct, v_ct)
p, qs, ks, vs = pqkv
_, (qs_ct, ks_ct, vs_ct) = lax.scan(
body, (p, jnp.zeros_like(p)), (qs, ks, vs, W_ct), reverse=True, unroll=unroll
)
return qs_ct, ks_ct, vs_ct
@jax.custom_vjp
def _numerator_impl(qs, ks, vs):
W, _ = fwd(qs, ks, vs)
return W
_numerator_impl.defvjp(fwd, bwd)
return _numerator_impl
def _denominator(t_slice_shape, precision, unroll=1):
def fwd(qs, ks):
def body(p, qk):
q, k = qk
p += k
x = jnp.einsum("...m,...m->...", q, p, precision=precision)
return p, x
p = jnp.zeros(t_slice_shape)
p, R = lax.scan(body, p, (qs, ks), unroll=unroll)
return R, (qs, ks, p)
def bwd(qkp, R_ct):
def body(carry, qkx):
p, p_ct = carry
q, k, x_ct = qkx
q_ct = jnp.einsum("...,...m->...m", x_ct, p, precision=precision)
p_ct += jnp.einsum("...,...m->...m", x_ct, q, precision=precision)
k_ct = p_ct
p -= k
return (p, p_ct), (q_ct, k_ct)
qs, ks, p = qkp
_, (qs_ct, ks_ct) = lax.scan(body, (p, jnp.zeros_like(p)), (qs, ks, R_ct), reverse=True, unroll=unroll)
return (qs_ct, ks_ct)
@jax.custom_vjp
def _denominator_impl(qs, ks):
R, _ = fwd(qs, ks)
return R
_denominator_impl.defvjp(fwd, bwd)
return _denominator_impl
class FastAttentionviaLowRankDecomposition(FastAttention):
r"""
Class providing a method for fast attention via low rank decomposition. Class is responsible for providing a method
<dot_product_attention> for fast dot-product attention with the use of low rank decomposition (e.g. with random
feature maps).
"""
def __init__(
self,
matrix_creator,
kernel_feature_creator,
renormalize_attention,
numerical_stabilizer,
redraw_features,
unidirectional,
lax_scan_unroll=1,
): # For optimal GPU performance, set to 16.
rng = random.PRNGKey(0)
self.matrix_creator = matrix_creator
self.projection_matrix = self.draw_weights(rng)
self.kernel_feature_creator = kernel_feature_creator
self.renormalize_attention = renormalize_attention
self.numerical_stabilizer = numerical_stabilizer
self.redraw_features = redraw_features
self.unidirectional = unidirectional
self.lax_scan_unroll = lax_scan_unroll
def draw_weights(self, key):
if self.matrix_creator is None:
return None
matrixrng, _ = random.split(key)
projection_matrix = self.matrix_creator(key=matrixrng).get_2d_array()
return projection_matrix
def dot_product_attention(
self,
query,
key,
value,
dtype=jnp.float32,
bias=None,
axis=None,
broadcast_dropout=True,
dropout_rng=None,
dropout_rate=0.0,
deterministic=False,
precision=None,
):
assert key.shape[:-1] == value.shape[:-1]
assert query.shape[0:1] == key.shape[0:1] and query.shape[-1] == key.shape[-1]
if axis is None:
axis = tuple(range(1, key.ndim - 2))
if not isinstance(axis, Iterable):
axis = (axis,)
assert key.ndim == query.ndim
assert key.ndim == value.ndim
for ax in axis:
if not (query.ndim >= 3 and 1 <= ax < query.ndim - 2):
raise ValueError("Attention axis must be between the batch axis and the last-two axes.")
n = key.ndim
# Constructing projection tensor.
if self.redraw_features:
# TODO(kchoro): Get rid of the constant below.
query_seed = lax.convert_element_type(jnp.ceil(jnp.sum(query) * 10000000.0), jnp.int32)
rng = random.PRNGKey(query_seed)
self.projection_matrix = self.draw_weights(rng)
# batch_dims is <bs, <non-attention dims>, num_heads>
batch_dims = tuple(onp.delete(range(n), axis + (n - 1,)))
# q & k -> (bs, <non-attention dims>, num_heads, <attention dims>, channels)
qk_perm = batch_dims + axis + (n - 1,)
k_extra_perm = axis + batch_dims + (n - 1,)
key_extra = key.transpose(k_extra_perm)
key = key.transpose(qk_perm)
query = query.transpose(qk_perm)
# v -> (bs, <non-attention dims>, num_heads, <attention dims>, channels)
v_perm = batch_dims + axis + (n - 1,)
value = value.transpose(v_perm)
batch_dims_t = tuple(range(len(batch_dims)))
attention_dims_t = tuple(range(len(batch_dims), len(batch_dims) + len(axis)))
# Constructing tensors Q^{'} and K^{'}.
query_prime = self.kernel_feature_creator(
query, self.projection_matrix, attention_dims_t, batch_dims_t, precision, True
)
key_prime = self.kernel_feature_creator(
key, self.projection_matrix, attention_dims_t, batch_dims_t, precision, False
)
if self.unidirectional:
index = attention_dims_t[0]
z_slice_shape = key_prime.shape[0 : len(batch_dims_t)] + (key_prime.shape[-1],) + (value.shape[-1],)
numerator_fn = _numerator(z_slice_shape, precision, self.lax_scan_unroll)
W = numerator_fn(
jnp.moveaxis(query_prime, index, 0), jnp.moveaxis(key_prime, index, 0), jnp.moveaxis(value, index, 0)
)
# Constructing W = (Q^{'}(K^{'})^{T})_{masked}V
W = jnp.moveaxis(W, 0, index)
if not self.renormalize_attention:
# Unidirectional, not-normalized attention.
perm_inv = _invert_perm(qk_perm)
result = W.transpose(perm_inv)
return result
else:
# Unidirectional, normalized attention.
thick_all_ones = jnp.zeros(key.shape[0:-1]) + jnp.ones(key_extra.shape[0 : len(axis)])
index = attention_dims_t[0]
t_slice_shape = key_prime.shape[0 : len(batch_dims_t)] + (key_prime.shape[-1],)
denominator_fn = _denominator(t_slice_shape, precision, self.lax_scan_unroll)
R = denominator_fn(jnp.moveaxis(query_prime, index, 0), jnp.moveaxis(key_prime, index, 0))
R = jnp.moveaxis(R, 0, index)
else:
contract_query = tuple(range(len(batch_dims) + len(axis), len(batch_dims) + len(axis) + 1))
contract_z = tuple(range(len(batch_dims), len(batch_dims) + 1))
# Constructing Z = (K^{'})^{T}V
# Z (bs, <non-attention dims>, num_heads, channels_m, channels_v)
Z = lax.dot_general(
key_prime,
value,
((attention_dims_t, attention_dims_t), (batch_dims_t, batch_dims_t)),
precision=precision,
)
# Constructing W = Q^{'}Z = Q^{'}(K^{'})^{T}V
# q (bs, <non-attention dims>, num_heads, <attention dims>, channels_m)
# Z (bs, <non-attention dims>, num_heads, channels_m, channels_v)
# W (bs, <non-attention dims>, num_heads, <attention dims>, channels_v)
W = lax.dot_general(
query_prime, Z, ((contract_query, contract_z), (batch_dims_t, batch_dims_t)), precision=precision
)
if not self.renormalize_attention:
# Bidirectional, not-normalized attention.
perm_inv = _invert_perm(qk_perm)
result = W.transpose(perm_inv)
return result
else:
# Bidirectional, normalized attention.
thick_all_ones = jnp.zeros(key.shape[0:-1]) + jnp.ones(key_extra.shape[0 : len(axis)])
contract_key = tuple(range(len(batch_dims), len(batch_dims) + len(axis)))
contract_thick_all_ones = tuple(range(thick_all_ones.ndim - len(axis), thick_all_ones.ndim))
# Construct T = (K^{'})^{T} 1_L
# k (bs, <non-attention dims>, num_heads, <attention dims>, channels)
T = lax.dot_general(
key_prime,
thick_all_ones,
((contract_key, contract_thick_all_ones), (batch_dims_t, batch_dims_t)),
precision=precision,
)
# Construct partition function: R = Q^{'} T = Q^{'}(K^{'})^{T} 1_L
# q_p (bs, <non-attention dims>, num_heads, <attention dims>, channs_m)
# T (bs, <non-attention dims>, num_heads, channels_m)
R = lax.dot_general(
query_prime,
T,
(((query_prime.ndim - 1,), (T.ndim - 1,)), (batch_dims_t, range(0, len(T.shape) - 1))),
precision=precision,
)
R = R + 2 * self.numerical_stabilizer * (jnp.abs(R) <= self.numerical_stabilizer)
R = jnp.reciprocal(R)
R = jnp.expand_dims(R, len(R.shape))
# W (bs, <non-attention dims>, num_heads, <attention dims>, channels_v)
# R (bs, <non-attention dims>, num_heads, <attention dims>, extra_channel)
result = W * R
# back to (bs, dim1, dim2, ..., dimN, num_heads, channels)
perm_inv = _invert_perm(qk_perm)
result = result.transpose(perm_inv)
return result
def _invert_perm(perm):
perm_inv = [0] * len(perm)
for i, j in enumerate(perm):
perm_inv[j] = i
return tuple(perm_inv)
|
transformers/examples/research_projects/performer/modeling_flax_performer_utils.py/0
|
{
"file_path": "transformers/examples/research_projects/performer/modeling_flax_performer_utils.py",
"repo_id": "transformers",
"token_count": 11667
}
| 328
|
# coding=utf-8
# Copyright 2020 The HuggingFace Team All rights reserved.
# Copyright 2021 NVIDIA Corporation. 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 logging
import os
import quant_trainer
import torch
from torch.utils.data import DataLoader
from transformers import Trainer, is_torch_xla_available
from transformers.trainer_utils import PredictionOutput
logger = logging.getLogger(__name__)
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, quant_trainer_args=None, **kwargs):
super().__init__(*args, **kwargs)
self.eval_examples = eval_examples
self.post_process_function = post_process_function
self.quant_trainer_args = quant_trainer_args
self.calib_num = 128 # default number of calibration samples
def get_calib_dataloader(self, calib_dataset=None):
"""
Returns the calibration dataloader :class:`~torch.utils.data.DataLoader`.
Args:
calib_dataset (:obj:`torch.utils.data.Dataset`, `optional`)
"""
if calib_dataset is None and self.calib_dataset is None:
raise ValueError("Trainer: calibration requires an calib_dataset.")
calib_dataset = calib_dataset if calib_dataset is not None else self.calib_dataset
calib_dataset = self._remove_unused_columns(calib_dataset, description="Calibration")
return DataLoader(
calib_dataset,
batch_size=self.args.eval_batch_size,
collate_fn=self.data_collator,
drop_last=self.args.dataloader_drop_last,
num_workers=self.args.dataloader_num_workers,
pin_memory=self.args.dataloader_pin_memory,
shuffle=True,
)
def calibrate(self, calib_dataset=None):
calib_dataset = self.train_dataset if calib_dataset is None else calib_dataset
calib_dataloader = self.get_calib_dataloader(calib_dataset)
model = self.model
quant_trainer.configure_model(model, self.quant_trainer_args, calib=True)
model.eval()
quant_trainer.enable_calibration(model)
logger.info("***** Running calibration *****")
logger.info(f" Num examples = {self.calib_num}")
logger.info(f" Batch size = {calib_dataloader.batch_size}")
for step, inputs in enumerate(calib_dataloader):
# Prediction step
loss, logits, labels = self.prediction_step(model, inputs, prediction_loss_only=True)
if (step + 1) * calib_dataloader.batch_size >= self.calib_num:
break
quant_trainer.finish_calibration(model, self.quant_trainer_args)
self.model = model
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
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,
)
finally:
self.compute_metrics = compute_metrics
if self.post_process_function is not None and self.compute_metrics is not None:
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)
self.log(metrics)
else:
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
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,
)
finally:
self.compute_metrics = compute_metrics
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)
return PredictionOutput(predictions=predictions.predictions, label_ids=predictions.label_ids, metrics=metrics)
def save_onnx(self, output_dir="./"):
eval_dataset = self.eval_dataset
eval_dataloader = self.get_eval_dataloader(eval_dataset)
batch = next(iter(eval_dataloader))
# saving device - to make it consistent
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# convert to tuple
input_tuple = tuple(v.to(device) for k, v in batch.items())
logger.info("Converting model to be onnx compatible")
from pytorch_quantization.nn import TensorQuantizer
TensorQuantizer.use_fb_fake_quant = True
model = self.model.to(device)
model.eval()
model.float()
model_to_save = model.module if hasattr(model, "module") else model
quant_trainer.configure_model(model_to_save, self.quant_trainer_args)
output_model_file = os.path.join(output_dir, "model.onnx")
logger.info(f"exporting model to {output_model_file}")
axes = {0: "batch_size", 1: "seq_len"}
torch.onnx.export(
model_to_save,
input_tuple,
output_model_file,
export_params=True,
opset_version=13,
do_constant_folding=True,
input_names=["input_ids", "attention_mask", "token_type_ids"],
output_names=["output_start_logits", "output_end_logits"],
dynamic_axes={
"input_ids": axes,
"attention_mask": axes,
"token_type_ids": axes,
"output_start_logits": axes,
"output_end_logits": axes,
},
verbose=True,
)
logger.info("onnx export finished")
|
transformers/examples/research_projects/quantization-qdqbert/trainer_quant_qa.py/0
|
{
"file_path": "transformers/examples/research_projects/quantization-qdqbert/trainer_quant_qa.py",
"repo_id": "transformers",
"token_count": 3711
}
| 329
|
# coding=utf-8
# Copyright 2022 The Google Research Authors.
#
# 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.
"""Self-training for sequence classification."""
import argparse
import dataclasses
import json
import logging
import os
import shutil
from typing import List, Optional
import datasets
from accelerate import Accelerator
from datasets import load_dataset
from finetuning import finetune
from tqdm.auto import tqdm
import transformers
from transformers import AutoConfig, set_seed
from transformers.trainer_utils import IntervalStrategy
logger = logging.getLogger(__name__)
MODEL_BIN_FILE = "pytorch_model.bin"
@dataclasses.dataclass
class STModelArguments:
"""Arguments pertaining to which config/tokenizer/model we are going to fine-tune from."""
model_name_or_path: str = dataclasses.field(
metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models."}
)
cache_dir: Optional[str] = dataclasses.field(
default=None,
metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co."},
)
@dataclasses.dataclass
class STDataArguments:
"""Arguments pertaining to what data we are going to input our model for training and evaluation."""
train_file: str = dataclasses.field(metadata={"help": "A csv or a json file containing the training data."})
infer_file: str = dataclasses.field(metadata={"help": "A csv or a json file containing the data to predict on."})
eval_file: Optional[str] = dataclasses.field(
default=None, metadata={"help": "A csv or a json file containing the validation data."}
)
task_name: Optional[str] = dataclasses.field(
default=None,
metadata={"help": "The name of the task to train on."},
)
label_list: Optional[List[str]] = dataclasses.field(
default=None, metadata={"help": "The list of labels for the task."}
)
@dataclasses.dataclass
class STTrainingArguments:
"""Training arguments pertaining to the training loop itself."""
output_dir: str = dataclasses.field(
metadata={"help": "The output directory where the model predictions and checkpoints will be written."}
)
eval_metric: Optional[str] = dataclasses.field(
default="accuracy", metadata={"help": "The evaluation metric used for the task."}
)
eval_strategy: Optional[str] = dataclasses.field(
default="no",
metadata={
"help": 'The evaluation strategy to adopt during training. Possible values are: ["no", "step", "epoch]'
},
)
early_stopping_patience: Optional[int] = dataclasses.field(
default=10,
metadata={"help": "Number of evaluation calls with no improvement after which training will be stopped."},
)
early_stopping_threshold: Optional[float] = dataclasses.field(
default=0.0,
metadata={
"help": "How much the specified evaluation metric must improve to satisfy early stopping conditions."
},
)
do_filter_by_confidence: Optional[bool] = dataclasses.field(
default=False,
metadata={"help": "Whether to filter the pseudo-labeled data based on the confidence score."},
)
do_filter_by_val_performance: Optional[bool] = dataclasses.field(
default=False,
metadata={"help": "Whether to filter the pseudo-labeled data based on the validation performance."},
)
finetune_on_labeled_data: Optional[bool] = dataclasses.field(
default=False,
metadata={"help": "Whether to fine-tune on labeled data after pseudo training."},
)
confidence_threshold: Optional[float] = dataclasses.field(
default=0.0,
metadata={"help": "Confidence threshold for pseudo-labeled data filtering."},
)
max_selftrain_iterations: Optional[int] = dataclasses.field(
default=100,
metadata={"help": "Number of evaluation calls with no improvement after which training will be stopped."},
)
seed: Optional[int] = dataclasses.field(
default=None,
metadata={"help": "Random seed for initialization."},
)
def create_pseudo_labeled_data(args, infer_input, infer_output, eval_result, id2label, next_data_dir):
"""Create pseudeo labeled data for the next self-training iteration."""
dataset = datasets.concatenate_datasets([infer_input, infer_output], axis=1)
if args.do_filter_by_confidence:
dataset = dataset.filter(lambda example: example["probability"] > args.confidence_threshold)
if args.do_filter_by_val_performance:
assert eval_result >= 0.0 and eval_result <= 1.0
num_selected_rows = int(eval_result * len(dataset))
print(num_selected_rows)
dataset = dataset.sort("probability", reverse=True)
dataset = dataset.select(range(num_selected_rows))
dataset = dataset.remove_columns(["label", "probability"])
dataset = dataset.rename_column("prediction", "label")
dataset = dataset.map(lambda example: {"label": id2label[example["label"]]})
dataset = dataset.shuffle(seed=args.seed)
pseudo_labeled_data_file = os.path.join(next_data_dir, f"train_pseudo.{args.data_file_extension}")
if args.data_file_extension == "csv":
dataset.to_csv(pseudo_labeled_data_file, index=False)
else:
dataset.to_json(pseudo_labeled_data_file)
def selftrain(model_name_or_path, train_file, infer_file, output_dir, **kwargs):
"""Self-training a pre-trained model on a downstream task.
Args:
model_name_or_path: Path to pretrained model or model identifier from
huggingface.co/models.
train_file: A csv or a json file containing the training data.
infer_file: A csv or a json file containing the data to predict on.
output_dir: The output directory where the model predictions and checkpoints
will be written.
**kwargs: Dictionary of key/value pairs with which to update the
configuration object after loading. The values in kwargs of any keys which
are configuration attributes will be used to override the loaded values.
"""
# Initialize the accelerator. We will let the accelerator handle device
# placement for us.
accelerator = Accelerator()
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state)
# 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 if accelerator.is_local_main_process else logging.ERROR)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
model_args = STModelArguments(model_name_or_path=model_name_or_path)
data_args = STDataArguments(train_file=train_file, infer_file=infer_file)
training_args = STTrainingArguments(output_dir=output_dir)
args = argparse.Namespace()
for arg_class in (model_args, data_args, training_args):
for key, value in vars(arg_class).items():
setattr(args, key, value)
for key, value in kwargs.items():
if hasattr(args, key):
setattr(args, key, value)
# Sanity checks
data_files = {}
args.data_file_extension = None
# You need to provide the training data and the data to predict on
assert args.train_file is not None
assert args.infer_file is not None
data_files["train"] = args.train_file
data_files["infer"] = args.infer_file
if args.eval_strategy != IntervalStrategy.NO.value:
assert args.eval_file is not None
data_files["eval"] = args.eval_file
for key in data_files:
extension = data_files[key].split(".")[-1]
assert extension in ["csv", "json"], f"`{key}_file` should be a csv or a json file."
if args.data_file_extension is None:
args.data_file_extension = extension
else:
assert extension == args.data_file_extension, f"`{key}_file` should be a {args.data_file_extension} file`."
assert (
args.eval_metric in datasets.list_metrics()
), f"{args.eval_metric} not in the list of supported metrics {datasets.list_metrics()}."
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
logger.info("Creating the initial data directory for self-training...")
data_dir_format = f"{args.output_dir}/self-train_iter-{{}}".format
initial_data_dir = data_dir_format(0)
if accelerator.is_main_process:
if args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
os.makedirs(initial_data_dir, exist_ok=True)
accelerator.wait_for_everyone()
best_iteration = None
best_eval_result = None
early_stopping_patience_counter = 0
should_training_stop = False
# Show the progress bar
progress_bar = tqdm(range(args.max_selftrain_iterations), disable=not accelerator.is_local_main_process)
# Self-train
for iteration in range(0, int(args.max_selftrain_iterations)):
current_data_dir = data_dir_format(iteration)
assert os.path.exists(current_data_dir)
# Stage 1: initial fine-tuning for iteration = 0 or pseudo-training for
# iteration > 0
current_output_dir = os.path.join(current_data_dir, "stage-1")
arguments_dict = {
"accelerator": accelerator,
"model_name_or_path": args.model_name_or_path,
"cache_dir": args.cache_dir,
"do_train": True,
"train_file": data_files["train"] if iteration == 0 else data_files["train_pseudo"],
"do_eval": True if args.eval_file is not None else False,
"eval_file": data_files["eval"],
"do_predict": True,
"infer_file": data_files["infer"],
"task_name": args.task_name,
"label_list": args.label_list,
"output_dir": current_output_dir,
"eval_metric": args.eval_metric,
"eval_strategy": args.eval_strategy,
"early_stopping_patience": args.early_stopping_patience,
"early_stopping_threshold": args.early_stopping_threshold,
"seed": args.seed,
}
# Add additional training arguments
for key, value in kwargs.items():
if key not in arguments_dict and not hasattr(training_args, key):
arguments_dict.update({key: value})
model_bin_file_path = os.path.join(current_output_dir, "best-checkpoint", MODEL_BIN_FILE)
if os.path.exists(model_bin_file_path):
logger.info(
"Found existing model checkpoint at %s. Skipping self-training: iteration: %d, stage: 1.",
model_bin_file_path,
iteration,
)
else:
logger.info("***** Running self-training: iteration: %d, stage: 1 *****", iteration)
finetune(**arguments_dict)
accelerator.wait_for_everyone()
assert os.path.exists(model_bin_file_path)
logger.info("Self-training job completed: iteration: %d, stage: 1.", iteration)
if iteration > 0 and args.finetune_on_labeled_data:
# Stage 2 (optional): fine-tuning on the original labeled data
model_path = os.path.join(current_output_dir, "best-checkpoint")
current_output_dir = os.path.join(current_data_dir, "stage-2")
# Update arguments_dict
arguments_dict["model_name_or_path"] = model_path
arguments_dict["train_file"] = data_files["train"]
arguments_dict["output_dir"] = current_output_dir
model_bin_file_path = os.path.join(current_output_dir, "best-checkpoint", MODEL_BIN_FILE)
if os.path.exists(model_bin_file_path):
logger.info(
"Found existing model checkpoint at %s. Skipping self-training: iteration: %d, stage: 2.",
model_bin_file_path,
iteration,
)
else:
logger.info("***** Running self-training: iteration: %d, stage: 2 *****", iteration)
finetune(**arguments_dict)
accelerator.wait_for_everyone()
assert os.path.exists(model_bin_file_path)
logger.info("Self-training job completed: iteration: %d, stage: 2.", iteration)
new_iteration = iteration
next_data_dir = data_dir_format(iteration + 1)
config = AutoConfig.from_pretrained(os.path.join(current_output_dir, "best-checkpoint"))
id2label = config.id2label
eval_results_file = os.path.join(current_output_dir, "eval_results_best-checkpoint.json")
test_results_file = os.path.join(current_output_dir, "test_results_best-checkpoint.json")
assert os.path.exists(eval_results_file)
with open(eval_results_file, "r") as f:
eval_result = float(json.load(f)[args.eval_metric])
infer_output_file = os.path.join(current_output_dir, "infer_output_best-checkpoint.csv")
assert os.path.exists(infer_output_file)
# Loading the dataset from local csv or json files.
infer_input = load_dataset(args.data_file_extension, data_files={"data": data_files["infer"]})["data"]
infer_output = load_dataset("csv", data_files={"data": infer_output_file})["data"]
if accelerator.is_main_process:
os.makedirs(next_data_dir, exist_ok=True)
shutil.copy(eval_results_file, os.path.join(output_dir, f"eval_results_iter-{iteration}.json"))
if os.path.exists(test_results_file):
shutil.copy(eval_results_file, os.path.join(output_dir, f"test_results_iter-{iteration}.json"))
create_pseudo_labeled_data(args, infer_input, infer_output, eval_result, id2label, next_data_dir)
accelerator.wait_for_everyone()
data_files["train_pseudo"] = os.path.join(next_data_dir, f"train_pseudo.{args.data_file_extension}")
if args.eval_strategy != IntervalStrategy.NO.value:
new_eval_result = eval_result
if best_iteration is None:
best_iteration = new_iteration
best_eval_result = new_eval_result
else:
if new_eval_result - best_eval_result > args.early_stopping_threshold:
best_iteration = new_iteration
best_eval_result = new_eval_result
early_stopping_patience_counter = 0
else:
if new_eval_result == best_eval_result:
best_iteration = new_iteration
best_eval_result = new_eval_result
early_stopping_patience_counter += 1
if early_stopping_patience_counter >= args.early_stopping_patience:
should_training_stop = True
progress_bar.update(1)
if should_training_stop:
break
if best_iteration is not None:
# Save the best iteration
logger.info("Best iteration: %d", best_iteration)
logger.info("Best evaluation result: %s = %f", args.eval_metric, best_eval_result)
accelerator.wait_for_everyone()
if accelerator.is_main_process:
shutil.copy(
os.path.join(output_dir, f"eval_results_iter-{iteration}.json"),
os.path.join(output_dir, "eval_results_best-iteration.json"),
)
else:
# Assume that the last iteration is the best
logger.info("Best iteration: %d", args.max_selftrain_iterations - 1)
logger.info("Best evaluation result: %s = %f", args.eval_metric, eval_result)
accelerator.wait_for_everyone()
if accelerator.is_main_process:
shutil.copy(
os.path.join(output_dir, f"eval_results_iter-{args.max_selftrain_iterations - 1}.json"),
os.path.join(output_dir, "eval_results_best-iteration.json"),
)
|
transformers/examples/research_projects/self-training-text-classification/selftraining.py/0
|
{
"file_path": "transformers/examples/research_projects/self-training-text-classification/selftraining.py",
"repo_id": "transformers",
"token_count": 6788
}
| 330
|
# Add parent directory to python path to access lightning_base.py
export PYTHONPATH="../":"${PYTHONPATH}"
python finetune.py \
--data_dir=$CNN_DIR \
--learning_rate=3e-5 \
--train_batch_size=$BS \
--eval_batch_size=$BS \
--output_dir=$OUTPUT_DIR \
--max_source_length=512 \
--max_target_length=56 \
--val_check_interval=0.1 --n_val=200 \
--do_train --do_predict \
"$@"
|
transformers/examples/research_projects/seq2seq-distillation/finetune_t5.sh/0
|
{
"file_path": "transformers/examples/research_projects/seq2seq-distillation/finetune_t5.sh",
"repo_id": "transformers",
"token_count": 148
}
| 331
|
# coding=utf-8
# Copyright 2022 The Microsoft, The Google 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.
import dataclasses
import enum
import functools
import math
import re
# The following script is adapted from the script of TaPas.
# Original: https://github.com/google-research/tapas/master/wikisql_utils.py
from typing import Any, List
EMPTY_ANSWER = "none"
EMPTY_ANSWER_AGG = "none"
def _split_thousands(delimiter, value):
split = value.split(delimiter)
return len(split) > 1 and any((len(x) == 3 for x in split))
def convert_to_float(value):
"""Converts value to a float using a series of increasingly complex heuristics.
Args:
value: object that needs to be converted. Allowed types include
float/int/strings.
Returns:
A float interpretation of value.
Raises:
ValueError if the float conversion of value fails.
"""
if isinstance(value, float):
return value
if isinstance(value, int):
return float(value)
if not isinstance(value, str):
raise TypeError("Argument value is not a string. Can't parse it as float")
sanitized = value
try:
# Example: 1,000.7
if "." in sanitized and "," in sanitized:
return float(sanitized.replace(",", ""))
# 1,000
if "," in sanitized and _split_thousands(",", sanitized):
return float(sanitized.replace(",", ""))
# 5,5556
if "," in sanitized and sanitized.count(",") == 1 and not _split_thousands(",", sanitized):
return float(sanitized.replace(",", "."))
# 0.0.0.1
if sanitized.count(".") > 1:
return float(sanitized.replace(".", ""))
# 0,0,0,1
if sanitized.count(",") > 1:
return float(sanitized.replace(",", ""))
return float(sanitized)
except ValueError:
# Avoid adding the sanitized value in the error message.
raise ValueError("Unable to convert value to float")
def _normalize_float(answer):
if answer is None:
return None
try:
value = convert_to_float(answer)
if isinstance(value, float) and math.isnan(value):
return None
return value
except ValueError:
return answer.lower()
_TYPE_CONVERTER = {
"text": lambda x: x,
"real": convert_to_float,
}
class _Aggregation(enum.Enum):
"""Aggregations as defined by WikiSQL. Indexes match the data."""
NONE = 0
MAX = 1
MIN = 2
COUNT = 3
SUM = 4
AVERAGE = 5
class _Operator(enum.Enum):
"""The boolean operators used by WikiSQL. Indexes match the data."""
EQUALS = 0
GREATER = 1
LESSER = 2
@dataclasses.dataclass
class _Condition:
"""Represents an SQL where clauses (e.g A = "a" or B > 5)."""
column: str
operator: _Operator
cmp_value: Any
_TOKENIZER = re.compile(r"\w+|[^\w\s]+", re.UNICODE | re.MULTILINE | re.DOTALL)
def _normalize_for_match(x):
return list(_TOKENIZER.findall(x.lower()))
def _compare(operator, src, tgt):
if operator == _Operator.EQUALS:
return src == tgt
elif operator == _Operator.GREATER:
return src > tgt
elif operator == _Operator.LESSER:
return src < tgt
raise ValueError(f"Unknown operator: {operator}")
def _parse_value(table, column, cell_value):
"""Convert numeric values to floats and keeps everything else as string."""
types = table["types"]
return _TYPE_CONVERTER[types[column]](cell_value)
def _is_string(x):
return isinstance(x, str)
def _respect_conditions(table, row, conditions):
"""True if 'row' satisfies all 'conditions'."""
for cond in conditions:
table_value = row[cond.column]
cmp_value = _parse_value(table, cond.column, cond.cmp_value)
if _is_string(table_value) and _is_string(cmp_value):
table_value = _normalize_for_match(table_value)
cmp_value = _normalize_for_match(cmp_value)
if not isinstance(table_value, type(cmp_value)):
raise TypeError("Type difference {} != {}".format(type(table_value), type(cmp_value)))
if not _compare(cond.operator, table_value, cmp_value):
return False
return True
def _get_float_answer(table, answer_coordinates, aggregation_op):
"""Applies operation to produce reference float answer."""
if not answer_coordinates:
if aggregation_op == _Aggregation.COUNT:
return 0.0
else:
return EMPTY_ANSWER_AGG
# Count can support non numeric answers.
if aggregation_op == _Aggregation.COUNT:
return float(len(answer_coordinates))
# If we have just one answer, if float returns it or try a conversion.
values = [table["rows"][i][j] for (i, j) in answer_coordinates]
if len(answer_coordinates) == 1:
try:
return convert_to_float(values[0])
except ValueError as e:
if aggregation_op != _Aggregation.NONE:
raise e
if aggregation_op == _Aggregation.NONE:
return None
# Other aggregation only support numeric values. Bail out if we have strings.
if not all((isinstance(v, (int, float)) for v in values)):
return None
if aggregation_op == _Aggregation.SUM:
return float(sum(values))
elif aggregation_op == _Aggregation.AVERAGE:
return sum(values) / len(answer_coordinates)
else:
raise ValueError(f"Unknown aggregation: {aggregation_op}")
def _get_answer_coordinates(table, sql_query):
"""Retrieves references coordinates by executing SQL."""
# MAX and MIN are automatically supported by the model.
aggregation_op_index = sql_query["agg"]
if aggregation_op_index >= 3:
aggregation_op = _Aggregation(aggregation_op_index)
else:
aggregation_op = _Aggregation.NONE
target_column = sql_query["sel"]
conditions = [
_Condition(column, _Operator(operator), cmp_value)
for column, operator, cmp_value in zip(
sql_query["conds"]["column_index"], sql_query["conds"]["operator_index"], sql_query["conds"]["condition"]
)
]
indices = []
for row in range(len(table["rows"])):
if _respect_conditions(table, table["rows"][row], conditions):
indices.append((row, target_column))
if not indices:
return [], aggregation_op
if len(indices) == 1:
return indices, aggregation_op
# Parsing of MIN/MAX.
if aggregation_op_index in (1, 2):
operators = {2: min, 1: max}
values = [(table["rows"][i][j], index) for index, (i, j) in enumerate(indices)]
reduced = functools.reduce(operators[sql_query["agg"]], values)
ret = [indices[reduced[1]]]
return ret, _Aggregation.NONE
return indices, aggregation_op
def _get_answer_text(table, answer_coordinates, float_answer):
if float_answer is not None:
return [str(float_answer)]
return [str(table["real_rows"][r][c]) for r, c in answer_coordinates]
def retrieve_wikisql_query_answer_tapas(table, example) -> List:
answer_coordinates, aggregation_op = _get_answer_coordinates(table, example)
float_answer = _get_float_answer(table, answer_coordinates, aggregation_op)
answer_text = _get_answer_text(table, answer_coordinates, float_answer)
# keep the original data the same with TaPas
if len(answer_text) == 0:
answer_text = [EMPTY_ANSWER]
return answer_text
|
transformers/examples/research_projects/tapex/wikisql_utils.py/0
|
{
"file_path": "transformers/examples/research_projects/tapex/wikisql_utils.py",
"repo_id": "transformers",
"token_count": 3098
}
| 332
|
from datetime import datetime
import matplotlib.pyplot as plt
import torch
def freeze_module(module):
for param in module.parameters():
param.requires_grad = False
def get_device():
device = "cuda" if torch.cuda.is_available() else "cpu"
if torch.backends.mps.is_available() and torch.backends.mps.is_built():
device = "mps"
if device == "mps":
print(
"WARNING: MPS currently doesn't seem to work, and messes up backpropagation without any visible torch"
" errors. I recommend using CUDA on a colab notebook or CPU instead if you're facing inexplicable issues"
" with generations."
)
return device
def show_pil(img):
fig = plt.imshow(img)
fig.axes.get_xaxis().set_visible(False)
fig.axes.get_yaxis().set_visible(False)
plt.show()
def get_timestamp():
current_time = datetime.now()
timestamp = current_time.strftime("%H:%M:%S")
return timestamp
|
transformers/examples/research_projects/vqgan-clip/utils.py/0
|
{
"file_path": "transformers/examples/research_projects/vqgan-clip/utils.py",
"repo_id": "transformers",
"token_count": 379
}
| 333
|
#!/usr/bin/env python3
import logging
import sys
from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional, Union
import librosa
import torch
from datasets import DatasetDict, load_dataset
from packaging import version
from torch import nn
from transformers import (
HfArgumentParser,
Trainer,
TrainingArguments,
Wav2Vec2Config,
Wav2Vec2FeatureExtractor,
Wav2Vec2ForPreTraining,
is_apex_available,
trainer_utils,
)
from transformers.models.wav2vec2.modeling_wav2vec2 import _compute_mask_indices
if is_apex_available():
from apex import amp
if version.parse(version.parse(torch.__version__).base_version) >= version.parse("1.6"):
_is_native_amp_available = True
from torch.cuda.amp import autocast
logger = logging.getLogger(__name__)
@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"}
)
cache_dir: Optional[str] = field(
default=None,
metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
)
freeze_feature_extractor: Optional[bool] = field(
default=True, metadata={"help": "Whether to freeze the feature extractor layers of the model."}
)
verbose_logging: Optional[bool] = field(
default=False,
metadata={"help": "Whether to log verbose messages or not."},
)
max_gumbel_temperature: Optional[float] = field(
default=2.0, metadata={"help": "Maximum temperature for gumbel softmax."}
)
min_gumbel_temperature: Optional[float] = field(
default=0.5, metadata={"help": "Minimum temperature for gumbel softmax."}
)
gumbel_temperature_decay: Optional[float] = field(
default=0.999995, metadata={"help": "Decay of gumbel temperature during training."}
)
def configure_logger(model_args: ModelArguments, training_args: TrainingArguments):
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logging_level = logging.WARNING
if model_args.verbose_logging:
logging_level = logging.DEBUG
elif trainer_utils.is_main_process(training_args.local_rank):
logging_level = logging.INFO
logger.setLevel(logging_level)
@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: 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_split_name: Optional[str] = field(
default="train",
metadata={
"help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
},
)
validation_split_name: Optional[str] = field(
default="validation",
metadata={
"help": (
"The name of the validation data set split to use (via the datasets library). Defaults to 'validation'"
)
},
)
speech_file_column: Optional[str] = field(
default="file",
metadata={"help": "Column in the dataset that contains speech file path. Defaults to 'file'"},
)
overwrite_cache: bool = field(
default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
)
validation_split_percentage: Optional[int] = field(
default=1,
metadata={
"help": "The percentage of the train set used as validation set in case there's no validation split"
},
)
preprocessing_num_workers: Optional[int] = field(
default=None,
metadata={"help": "The number of processes to use for the preprocessing."},
)
max_duration_in_seconds: Optional[float] = field(
default=20.0, metadata={"help": "Filter audio files that are longer than `max_duration_in_seconds` seconds"}
)
@dataclass
class DataCollatorForWav2Vec2Pretraining:
"""
Data collator that will dynamically pad the inputs received and prepare masked indices
for self-supervised pretraining.
Args:
model (:class:`~transformers.Wav2Vec2ForPreTraining`):
The Wav2Vec2 model used for pretraining. The data collator needs to have access
to config and ``_get_feat_extract_output_lengths`` function for correct padding.
feature_extractor (:class:`~transformers.Wav2Vec2FeatureExtractor`):
The processor used for proccessing the data.
padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
among:
* :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
sequence if provided).
* :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
maximum acceptable input length for the model if that argument is not provided.
* :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
different lengths).
max_length (:obj:`int`, `optional`):
Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
pad_to_multiple_of (:obj:`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 >=
7.5 (Volta).
"""
model: Wav2Vec2ForPreTraining
feature_extractor: Wav2Vec2FeatureExtractor
padding: Union[bool, str] = "longest"
pad_to_multiple_of: Optional[int] = None
max_length: Optional[int] = None
def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
# reformat list to dict and set to pytorch format
batch = self.feature_extractor.pad(
features,
max_length=self.max_length,
padding=self.padding,
pad_to_multiple_of=self.pad_to_multiple_of,
return_tensors="pt",
)
mask_indices_seq_length = self.model._get_feat_extract_output_lengths(batch["input_values"].shape[-1])
batch_size = batch["input_values"].shape[0]
# make sure that no loss is computed on padded inputs
if batch["attention_mask"] is not None:
# compute real output lengths according to convolution formula
output_lengths = self.model._get_feat_extract_output_lengths(batch["attention_mask"].sum(-1)).to(
torch.long
)
attention_mask = torch.zeros(
(batch_size, mask_indices_seq_length), dtype=torch.long, device=batch["input_values"].device
)
# these two operations makes sure that all values
# before the output lengths indices are attended to
attention_mask[
(torch.arange(attention_mask.shape[0], device=batch["input_values"].device), output_lengths - 1)
] = 1
attention_mask = attention_mask.flip([-1]).cumsum(-1).flip([-1]).bool()
# sample randomly masked indices
batch["mask_time_indices"] = _compute_mask_indices(
(batch_size, mask_indices_seq_length),
self.model.config.mask_time_prob,
self.model.config.mask_time_length,
attention_mask=attention_mask,
min_masks=2,
)
return batch
class Wav2Vec2PreTrainer(Trainer):
"""
Subclassed :class:`~transformers.Trainer` for Wav2Vec2-like pretraining. Trainer can decay gumbel softmax temperature during training.
"""
def __init__(self, *args, max_gumbel_temp=1, min_gumbel_temp=0, gumbel_temp_decay=1.0, **kwargs):
super().__init__(*args, **kwargs)
self.num_update_step = 0
self.max_gumbel_temp = max_gumbel_temp
self.min_gumbel_temp = min_gumbel_temp
self.gumbel_temp_decay = gumbel_temp_decay
def training_step(self, model: nn.Module, inputs: Dict[str, Union[torch.Tensor, Any]]) -> torch.Tensor:
"""
Perform a training step on a batch of inputs.
Subclass and override to inject custom behavior.
Args:
model (:obj:`nn.Module`):
The model to train.
inputs (:obj:`Dict[str, Union[torch.Tensor, Any]]`):
The inputs and targets of the model.
The dictionary will be unpacked before being fed to the model. Most models expect the targets under the
argument :obj:`labels`. Check your model's documentation for all accepted arguments.
Return:
:obj:`torch.Tensor`: The tensor with training loss on this batch.
"""
model.train()
inputs = self._prepare_inputs(inputs)
if self.use_amp:
with autocast():
loss = self.compute_loss(model, inputs)
else:
loss = self.compute_loss(model, inputs)
if self.args.n_gpu > 1 or self.deepspeed:
if model.module.config.ctc_loss_reduction == "mean":
loss = loss.mean()
elif model.module.config.ctc_loss_reduction == "sum":
loss = loss.sum() / (inputs["mask_time_indices"]).sum()
else:
raise ValueError(f"{model.config.ctc_loss_reduction} is not valid. Choose one of ['mean', 'sum']")
if self.args.gradient_accumulation_steps > 1:
loss = loss / self.args.gradient_accumulation_steps
if self.use_amp:
self.scaler.scale(loss).backward()
elif self.use_apex:
with amp.scale_loss(loss, self.optimizer) as scaled_loss:
scaled_loss.backward()
elif self.deepspeed:
self.deepspeed.backward(loss)
else:
loss.backward()
self.num_update_step += 1
# make sure gumbel softmax temperature is decayed
if self.args.n_gpu > 1 or self.deepspeed:
model.module.set_gumbel_temperature(
max(self.max_gumbel_temp * self.gumbel_temp_decay**self.num_update_step, self.min_gumbel_temp)
)
else:
model.set_gumbel_temperature(
max(self.max_gumbel_temp * self.gumbel_temp_decay**self.num_update_step, self.min_gumbel_temp)
)
return loss.detach()
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))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
configure_logger(model_args, training_args)
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir)
if "validation" not in datasets.keys():
# make sure only "validation" and "train" keys remain"
datasets = DatasetDict()
datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"{data_args.train_split_name}[:{data_args.validation_split_percentage}%]",
cache_dir=model_args.cache_dir,
)
datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"{data_args.train_split_name}[{data_args.validation_split_percentage}%:]",
cache_dir=model_args.cache_dir,
)
else:
# make sure only "validation" and "train" keys remain"
datasets = DatasetDict()
datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split="validation",
cache_dir=model_args.cache_dir,
)
datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"{data_args.train_split_name}",
cache_dir=model_args.cache_dir,
)
# only normalized-inputs-training is supported
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir, do_normalize=True
)
def prepare_dataset(batch):
# check that all files have the correct sampling rate
batch["speech"], _ = librosa.load(batch[data_args.speech_file_column], sr=feature_extractor.sampling_rate)
return batch
# load audio files into numpy arrays
vectorized_datasets = datasets.map(
prepare_dataset, num_proc=data_args.preprocessing_num_workers, remove_columns=datasets["train"].column_names
)
# filter audio files that are too long
vectorized_datasets = vectorized_datasets.filter(
lambda data: len(data["speech"]) < int(data_args.max_duration_in_seconds * feature_extractor.sampling_rate)
)
def normalize(batch):
return feature_extractor(batch["speech"], sampling_rate=feature_extractor.sampling_rate)
# normalize and transform to `BatchFeatures`
vectorized_datasets = vectorized_datasets.map(
normalize,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
remove_columns=vectorized_datasets["train"].column_names,
)
# pretraining is only supported for "newer" stable layer norm architecture
# apply_spec_augment has to be True, mask_feature_prob has to be 0.0
config = Wav2Vec2Config.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
gradient_checkpointing=training_args.gradient_checkpointing,
)
if not config.do_stable_layer_norm or config.feat_extract_norm != "layer":
raise ValueError(
"PreTraining is only supported for ``config.do_stable_layer_norm=True`` and"
" ``config.feat_extract_norm='layer'"
)
model = Wav2Vec2ForPreTraining(config)
data_collator = DataCollatorForWav2Vec2Pretraining(model=model, feature_extractor=feature_extractor)
trainer = Wav2Vec2PreTrainer(
model=model,
data_collator=data_collator,
args=training_args,
train_dataset=vectorized_datasets["train"],
eval_dataset=vectorized_datasets["validation"],
tokenizer=feature_extractor,
max_gumbel_temp=model_args.max_gumbel_temperature,
min_gumbel_temp=model_args.min_gumbel_temperature,
gumbel_temp_decay=model_args.gumbel_temperature_decay,
)
trainer.train()
if __name__ == "__main__":
main()
|
transformers/examples/research_projects/wav2vec2/run_pretrain.py/0
|
{
"file_path": "transformers/examples/research_projects/wav2vec2/run_pretrain.py",
"repo_id": "transformers",
"token_count": 6513
}
| 334
|
#!/usr/bin/env python
# coding=utf-8
# Copyright The HuggingFace Team 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.
"""
Fine-tuning the library models for multiple choice.
"""
# You can also adapt this script on your own multiple choice task. Pointers for this are left as comments.
import json
import logging
import os
import sys
from dataclasses import dataclass, field
from itertools import chain
from pathlib import Path
from typing import Optional, Union
import datasets
import tensorflow as tf
from datasets import load_dataset
import transformers
from transformers import (
CONFIG_NAME,
TF2_WEIGHTS_NAME,
AutoConfig,
AutoTokenizer,
DefaultDataCollator,
HfArgumentParser,
PushToHubCallback,
TFAutoModelForMultipleChoice,
TFTrainingArguments,
create_optimizer,
set_seed,
)
from transformers.tokenization_utils_base import PreTrainedTokenizerBase
from transformers.utils import PaddingStrategy, check_min_version, send_example_telemetry
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
check_min_version("4.45.0.dev0")
logger = logging.getLogger(__name__)
# region Helper classes and functions
@dataclass
class DataCollatorForMultipleChoice:
"""
Data collator that will dynamically pad the inputs for multiple choice 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'`: Pad to the longest sequence in the batch (or no padding if only a single sequence
if 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*):
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 >=
7.5 (Volta).
"""
tokenizer: PreTrainedTokenizerBase
padding: Union[bool, str, PaddingStrategy] = True
max_length: Optional[int] = None
pad_to_multiple_of: Optional[int] = None
def __call__(self, features):
label_name = "label" if "label" in features[0].keys() 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 = 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="np",
)
# Un-flatten
batch = {k: tf.reshape(v, (batch_size, num_choices, -1)) for k, v in batch.items()}
# Add back labels
batch["labels"] = tf.convert_to_tensor(labels, dtype=tf.int64)
return batch
# endregion
# region 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"},
)
use_fast_tokenizer: bool = field(
default=True,
metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
)
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 `huggingface-cli login` (stored in `~/.huggingface`)."
)
},
)
trust_remote_code: bool = field(
default=False,
metadata={
"help": (
"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."
)
},
)
@dataclass
class DataTrainingArguments:
"""
Arguments pertaining to what data we are going to input our model for training and eval.
"""
train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
validation_file: Optional[str] = field(
default=None,
metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text 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_seq_length: Optional[int] = field(
default=None,
metadata={
"help": (
"The maximum total input sequence length after tokenization. If passed, sequences longer "
"than this will be truncated, sequences shorter will be padded."
)
},
)
pad_to_max_length: bool = field(
default=False,
metadata={
"help": (
"Whether to pad all samples to the 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."
)
},
)
def __post_init__(self):
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."
# endregion
def main():
# region Argument parsing
# 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, TFTrainingArguments))
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_swag", model_args, data_args, framework="tensorflow")
output_dir = Path(training_args.output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
# endregion
# region Logging
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)
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# endregion
# region Checkpoints
checkpoint = None
if len(os.listdir(training_args.output_dir)) > 0 and not training_args.overwrite_output_dir:
if (output_dir / CONFIG_NAME).is_file() and (output_dir / TF2_WEIGHTS_NAME).is_file():
checkpoint = output_dir
logger.info(
f"Checkpoint detected, resuming training from checkpoint in {training_args.output_dir}. To avoid this"
" behavior, change the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
else:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to continue regardless."
)
# endregion
# Set seed before initializing model.
set_seed(training_args.seed)
# region Load 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 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 'text' or the first column if no column called
# 'text' 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.train_file is not None or data_args.validation_file is not None:
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,
cache_dir=model_args.cache_dir,
token=model_args.token,
)
else:
# Downloading and loading the swag dataset from the hub.
raw_datasets = load_dataset(
"swag",
"regular",
cache_dir=model_args.cache_dir,
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.
# When using your own dataset or a different dataset from swag, you will probably need to change this.
ending_names = [f"ending{i}" for i in range(4)]
context_name = "sent1"
question_header_name = "sent2"
# endregion
# region Load model config and tokenizer
if checkpoint is not None:
config_path = training_args.output_dir
elif model_args.config_name:
config_path = model_args.config_name
else:
config_path = model_args.model_name_or_path
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
config_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
# endregion
# region Dataset preprocessing
if data_args.max_seq_length is None:
max_seq_length = tokenizer.model_max_length
if max_seq_length > 1024:
logger.warning(
f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
"Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
)
max_seq_length = 1024
else:
if data_args.max_seq_length > tokenizer.model_max_length:
logger.warning(
f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the "
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
def preprocess_function(examples):
first_sentences = [[context] * 4 for context in examples[context_name]]
question_headers = examples[question_header_name]
second_sentences = [
[f"{header} {examples[end][i]}" for end in ending_names] for i, header in enumerate(question_headers)
]
# Flatten out
first_sentences = list(chain(*first_sentences))
second_sentences = list(chain(*second_sentences))
# Tokenize
tokenized_examples = tokenizer(first_sentences, second_sentences, truncation=True, max_length=max_seq_length)
# Un-flatten
data = {k: [v[i : i + 4] for i in range(0, len(v), 4)] for k, v in tokenized_examples.items()}
return data
if training_args.do_train:
if "train" not in raw_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = raw_datasets["train"]
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))
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
if "validation" not in raw_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = raw_datasets["validation"]
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))
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
if data_args.pad_to_max_length:
data_collator = DefaultDataCollator(return_tensors="np")
else:
# custom class defined above, as HF has no data collator for multiple choice
data_collator = DataCollatorForMultipleChoice(tokenizer)
# endregion
with training_args.strategy.scope():
# region Build model
if checkpoint is None:
model_path = model_args.model_name_or_path
else:
model_path = checkpoint
model = TFAutoModelForMultipleChoice.from_pretrained(
model_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,
)
num_replicas = training_args.strategy.num_replicas_in_sync
total_train_batch_size = training_args.per_device_train_batch_size * num_replicas
total_eval_batch_size = training_args.per_device_eval_batch_size * num_replicas
if training_args.do_train:
num_train_steps = (len(train_dataset) // total_train_batch_size) * int(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,
)
else:
optimizer = "sgd" # Just write anything because we won't be using it
# 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, metrics=["accuracy"], jit_compile=training_args.xla)
# 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:
push_to_hub_model_id = f"{model_name}-finetuned-multiplechoice"
model_card_kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "multiple-choice"}
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
eval_metrics = None
if training_args.do_train:
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,
shuffle=True,
batch_size=total_train_batch_size,
collate_fn=data_collator,
).with_options(dataset_options)
if training_args.do_eval:
validation_data = model.prepare_tf_dataset(
eval_dataset,
shuffle=False,
batch_size=total_eval_batch_size,
collate_fn=data_collator,
drop_remainder=True,
).with_options(dataset_options)
else:
validation_data = None
history = model.fit(
tf_train_dataset,
validation_data=validation_data,
epochs=int(training_args.num_train_epochs),
callbacks=callbacks,
)
eval_metrics = {key: val[-1] for key, val in history.history.items()}
# endregion
# region Evaluation
if training_args.do_eval and not training_args.do_train:
dataset_options = tf.data.Options()
dataset_options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.OFF
# Do a standalone evaluation pass
tf_eval_dataset = model.prepare_tf_dataset(
eval_dataset,
shuffle=False,
batch_size=total_eval_batch_size,
collate_fn=data_collator,
drop_remainder=True,
).with_options(dataset_options)
eval_results = model.evaluate(tf_eval_dataset)
eval_metrics = {"val_loss": eval_results[0], "val_accuracy": eval_results[1]}
# endregion
if eval_metrics is not None and 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_metrics))
# region Push to hub
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)
# endregion
if __name__ == "__main__":
main()
|
transformers/examples/tensorflow/multiple-choice/run_swag.py/0
|
{
"file_path": "transformers/examples/tensorflow/multiple-choice/run_swag.py",
"repo_id": "transformers",
"token_count": 9932
}
| 335
|
<!---
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.
-->
# Translation example
This script shows an example of training a *translation* model with the 🤗 Transformers library.
For straightforward use-cases you may be able to use these scripts without modification, although we have also
included comments in the code to indicate areas that you may need to adapt to your own projects.
### Multi-GPU and TPU usage
By default, these scripts use a `MirroredStrategy` and will use multiple GPUs effectively if they are available. TPUs
can also be used by passing the name of the TPU resource with the `--tpu` argument.
### Example commands and caveats
MBart and some T5 models require special handling.
T5 models `google-t5/t5-small`, `google-t5/t5-base`, `google-t5/t5-large`, `google-t5/t5-3b` and `google-t5/t5-11b` must use an additional argument: `--source_prefix "translate {source_lang} to {target_lang}"`. For example:
```bash
python run_translation.py \
--model_name_or_path google-t5/t5-small \
--do_train \
--do_eval \
--source_lang en \
--target_lang ro \
--source_prefix "translate English to Romanian: " \
--dataset_name wmt16 \
--dataset_config_name ro-en \
--output_dir /tmp/tst-translation \
--per_device_train_batch_size=16 \
--per_device_eval_batch_size=16 \
--overwrite_output_dir
```
If you get a terrible BLEU score, make sure that you didn't forget to use the `--source_prefix` argument.
For the aforementioned group of T5 models it's important to remember that if you switch to a different language pair, make sure to adjust the source and target values in all 3 language-specific command line argument: `--source_lang`, `--target_lang` and `--source_prefix`.
MBart models require a different format for `--source_lang` and `--target_lang` values, e.g. instead of `en` it expects `en_XX`, for `ro` it expects `ro_RO`. The full MBart specification for language codes can be found [here](https://huggingface.co/facebook/mbart-large-cc25). For example:
```bash
python run_translation.py \
--model_name_or_path facebook/mbart-large-en-ro \
--do_train \
--do_eval \
--dataset_name wmt16 \
--dataset_config_name ro-en \
--source_lang en_XX \
--target_lang ro_RO \
--output_dir /tmp/tst-translation \
--per_device_train_batch_size=16 \
--per_device_eval_batch_size=16 \
--overwrite_output_dir
```
|
transformers/examples/tensorflow/translation/README.md/0
|
{
"file_path": "transformers/examples/tensorflow/translation/README.md",
"repo_id": "transformers",
"token_count": 933
}
| 336
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.