File size: 10,537 Bytes
fbd9366 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 | """
Functions that work on nested structures of torch.Tensor or numpy array
"""
from typing import Any, Dict, List, Optional, Union
import numpy as np
import torch
import tree
from ..data_structure.tree_utils import (
copy_non_leaf,
is_sequence,
tree_assign_at_path,
tree_value_at_path,
)
from .functional_utils import make_recursive_func
def is_array_tensor(obj):
return isinstance(obj, (np.ndarray, torch.Tensor))
def is_numpy(obj):
return isinstance(obj, np.ndarray)
def is_tensor(obj):
return torch.is_tensor(obj)
def any_stack(xs: List, *, dim: int = 0):
"""
Works for both torch Tensor and numpy array
"""
def _any_stack_helper(*xs):
x = xs[0]
if isinstance(x, np.ndarray):
return np.stack(xs, axis=dim)
elif torch.is_tensor(x):
return torch.stack(xs, dim=dim)
elif isinstance(x, float):
# special treatment for float, defaults to float32
return np.array(xs, dtype=np.float32)
else:
return np.array(xs)
return tree.map_structure(_any_stack_helper, *xs)
def any_concat(xs: List, *, dim: int = 0):
"""
Works for both torch Tensor and numpy array
"""
def _any_concat_helper(*xs):
x = xs[0]
if isinstance(x, np.ndarray):
return np.concatenate(xs, axis=dim)
elif torch.is_tensor(x):
return torch.cat(xs, dim=dim)
elif isinstance(x, float):
# special treatment for float, defaults to float32
return np.array(xs, dtype=np.float32)
else:
return np.array(xs)
return tree.map_structure(_any_concat_helper, *xs)
def any_chunk(x, chunks: int, *, dim: int = 0, strict: bool = True) -> List[Any]:
"""
Works for both torch Tensor and numpy array
Returns:
list of chunked nested structures
"""
assert chunks >= 1
x_copies = [copy_non_leaf(x) for _ in range(chunks)]
def _any_chunk_helper(path, x):
if is_array_tensor(x):
if isinstance(x, np.ndarray):
chunked_values = np.split(x, chunks, axis=dim)
else:
chunked_values = torch.chunk(x, chunks, dim=dim)
if path:
for xc, chunked in zip(x_copies, chunked_values):
tree_assign_at_path(xc, path, chunked)
else: # top-level, no nested path
for i, chunked in enumerate(chunked_values):
x_copies[i] = chunked
else:
if strict:
raise NotImplementedError(f"Cannot chunk type {type(x)}")
else:
return
tree.map_structure_with_path(_any_chunk_helper, x)
return x_copies
def chunk_seq(arr, chunks: int, check_divide=True):
"""
Args:
check_divide: True to force arr must divide n
"""
k, m = divmod(len(arr), chunks)
if check_divide and m != 0:
raise ValueError(f"Array len {len(arr)} does not divide chunks {chunks}")
return (arr[i * k + min(i, m) : (i + 1) * k + min(i + 1, m)] for i in range(chunks))
@make_recursive_func
def any_zeros_like(x: Union[Dict, np.ndarray, torch.Tensor, int, float, np.number]):
"""Returns a zero-filled object of the same (d)type and shape as the input.
The difference between this and `np.zeros_like()` is that this works well
with `np.number`, `int`, `float`, and `jax.numpy.DeviceArray` objects without
converting them to `np.ndarray`s.
Args:
x: The object to replace with 0s.
Returns:
A zero-filed object of the same (d)type and shape as the input.
"""
if isinstance(x, (int, float, np.number)):
return type(x)(0)
elif is_tensor(x):
return torch.zeros_like(x)
elif is_numpy(x):
return np.zeros_like(x)
else:
raise ValueError(
f"Input ({type(x)}) must be either a numpy array, a tensor, an int, or a float."
)
@make_recursive_func
def any_ones_like(x: Union[Dict, np.ndarray, torch.Tensor, int, float, np.number]):
"""Returns a one-filled object of the same (d)type and shape as the input.
The difference between this and `np.ones_like()` is that this works well
with `np.number`, `int`, `float`, and `jax.numpy.DeviceArray` objects without
converting them to `np.ndarray`s.
Args:
x: The object to replace with 1s.
Returns:
A one-filed object of the same (d)type and shape as the input.
"""
if isinstance(x, (int, float, np.number)):
return type(x)(1)
elif is_tensor(x):
return torch.ones_like(x)
elif is_numpy(x):
return np.ones_like(x)
else:
raise ValueError(
f"Input ({type(x)}) must be either a numpy array, a tensor, an int, or a float."
)
@make_recursive_func
def any_zero_(x: Union[Dict, np.ndarray, torch.Tensor]):
"""
Apply in-place zero-out to a tensor, i.e. x.zero_()
"""
if is_tensor(x):
x.zero_()
elif is_numpy(x):
x.fill(0)
else:
raise ValueError(f"Input ({type(x)}) must be either a numpy array or a tensor")
@make_recursive_func
def any_fill_(x: Union[Dict, np.ndarray, torch.Tensor], value):
"""
Apply in-place zero-out to a tensor, i.e. x.zero_()
"""
if is_tensor(x):
x.fill_(value)
elif is_numpy(x):
x.fill(value)
else:
raise ValueError(f"Input ({type(x)}) must be either a numpy array or a tensor")
def get_batch_size(x, strict: bool = False) -> int:
"""
Args:
x: can be any arbitrary nested structure of np array and torch tensor
strict: True to check all batch sizes are the same
"""
def _get_batch_size(x):
if isinstance(x, np.ndarray):
return x.shape[0]
elif torch.is_tensor(x):
return x.size(0)
else:
return len(x)
xs = tree.flatten(x)
if strict:
batch_sizes = [_get_batch_size(x) for x in xs]
assert all(
b == batch_sizes[0] for b in batch_sizes
), f"batch sizes must all be the same in nested structure: {batch_sizes}"
return batch_sizes[0]
else:
return _get_batch_size(xs[0])
@make_recursive_func
def add_batch_dim(x):
if is_numpy(x):
return np.expand_dims(x, axis=0)
elif is_tensor(x):
return x.unsqueeze(0)
else:
raise NotImplementedError(f"Unsupported data structure: {type(x)}")
@make_recursive_func
def remove_batch_dim(x):
if is_numpy(x):
return np.squeeze(x, axis=0)
elif is_tensor(x):
return x.squeeze(0)
else:
raise NotImplementedError(f"Unsupported data structure: {type(x)}")
@make_recursive_func
def any_to_primitive(x):
if isinstance(x, (np.ndarray, np.number, torch.Tensor)):
return x.tolist()
else:
return x
@make_recursive_func
def any_get_shape(x):
if is_numpy(x):
return tuple(x.shape)
elif is_tensor(x):
return tuple(x.size())
else:
raise NotImplementedError(f"Unsupported data structure: {type(x)}")
@make_recursive_func
def any_mean(x, dim: Optional[int] = None, keepdim: bool = False):
if is_numpy(x):
return np.mean(x, axis=dim, keepdims=keepdim)
elif is_tensor(x):
return torch.mean(x, dim=dim, keepdim=keepdim)
else:
raise NotImplementedError(f"Unsupported data structure: {type(x)}")
@make_recursive_func
def any_variance(x, dim: Optional[int] = None, keepdim: bool = False, unbiased: bool = False):
if is_numpy(x):
return np.var(x, axis=dim, keepdims=keepdim, ddof=1 if unbiased else 0)
elif is_tensor(x):
return torch.var(x, dim=dim, keepdim=keepdim, unbiased=unbiased)
else:
raise NotImplementedError(f"Unsupported data structure: {type(x)}")
@make_recursive_func
def any_describe_str(x, shape_only=False):
"""
Describe type, shape, device, data type (of np array/tensor)
Very useful for debugging
"""
t = type(x)
tname = type(x).__name__
if is_numpy(x):
shape = list(x.shape)
if x.size == 1:
if shape_only:
return f"np scalar: {x.item()} {shape}"
else:
return f"np scalar: {x.item()} {shape} {x.dtype}"
else:
if shape_only:
return f"np: {shape}"
else:
return f"np: {shape} {x.dtype}"
elif is_tensor(x):
shape = list(x.size())
if x.numel() == 1:
if shape_only:
return f"torch scalar: {x.item()} {shape}"
else:
return f"torch scalar: {x.item()} {shape} {x.dtype} {x.device}"
else:
if shape_only:
return f"torch: {shape}"
else:
return f"torch: {shape} {x.dtype} {x.device}"
elif is_sequence(x):
return f"{tname}[{len(x)}]"
elif isinstance(x, str):
return x
elif x is None:
return "None"
elif np.issubdtype(t, np.number) or np.issubdtype(t, np.bool_):
return f"{tname}: {x}"
else:
return f"{tname}"
def any_describe(x, msg="", *, shape_only=False):
# from omlet.utils import yaml_dumps
from pprint import pprint
if isinstance(x, str) and msg != "":
x, msg = msg, x
if msg:
msg += ": "
print(msg, end="")
pprint(any_describe_str(x, shape_only=shape_only))
@make_recursive_func
def any_slice(x, slice):
"""
Args:
slice: you can use np.s_[...] to return the slice object
"""
if is_array_tensor(x):
return x[slice]
else:
return x
def any_assign(x, assign_value, slice):
"""
Recursive version of x[slice] = assign_value
If structures of x and assign_value do not match, we will respect `assign_value`
E.g. x = {'a': ..., 'b': ...}, assign_value = {'a': ...}, then 'b' will not change
Use np.s_[...] to get advanced slicing
"""
def _any_assign_helper(path, v):
y = tree_value_at_path(x, path)
y[slice] = v
tree.map_structure_with_path(_any_assign_helper, assign_value)
@make_recursive_func
def any_transpose_first_two_axes(x):
"""
util to convert between (L, B, ...) and (B, L, ...)
"""
if is_numpy(x):
return np.swapaxes(x, 0, 1)
elif is_tensor(x):
return torch.swapaxes(x, 0, 1)
else:
raise ValueError(f"Input ({type(x)}) must be either a numpy array or a tensor.")
|