File size: 9,482 Bytes
2ce7b1a |
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 |
"""Implementation of basic benchmark datasets used in S4 experiments: MNIST, CIFAR10 and Speech Commands."""
import numpy as np
import torch
import torchvision
from einops.layers.torch import Rearrange
from .base import default_data_path, ImageResolutionSequenceDataset, ResolutionSequenceDataset, SequenceDataset
from ..utils import permutations
class MNIST(SequenceDataset):
_name_ = "mnist"
d_input = 1
d_output = 10
l_output = 0
L = 784
@property
def init_defaults(self):
return {
"permute": True,
"val_split": 0.1,
"seed": 42, # For train/val split
}
def setup(self):
self.data_dir = self.data_dir or default_data_path / self._name_
transform_list = [
torchvision.transforms.ToTensor(),
torchvision.transforms.Lambda(lambda x: x.view(self.d_input, self.L).t()),
] # (L, d_input)
if self.permute:
# below is another permutation that other works have used
# permute = np.random.RandomState(92916)
# permutation = torch.LongTensor(permute.permutation(784))
permutation = permutations.bitreversal_permutation(self.L)
transform_list.append(
torchvision.transforms.Lambda(lambda x: x[permutation])
)
# TODO does MNIST need normalization?
# torchvision.transforms.Normalize((0.1307,), (0.3081,)) # normalize inputs
transform = torchvision.transforms.Compose(transform_list)
self.dataset_train = torchvision.datasets.MNIST(
self.data_dir,
train=True,
download=True,
transform=transform,
)
self.dataset_test = torchvision.datasets.MNIST(
self.data_dir,
train=False,
transform=transform,
)
self.split_train_val(self.val_split)
def __str__(self):
return f"{'p' if self.permute else 's'}{self._name_}"
class CIFAR10(ImageResolutionSequenceDataset):
_name_ = "cifar"
d_output = 10
l_output = 0
@property
def init_defaults(self):
return {
"permute": None,
"grayscale": False,
"tokenize": False, # if grayscale, tokenize into discrete byte inputs
"augment": False,
"cutout": False,
"rescale": None,
"random_erasing": False,
"val_split": 0.1,
"seed": 42, # For validation split
}
@property
def d_input(self):
if self.grayscale:
if self.tokenize:
return 256
else:
return 1
else:
assert not self.tokenize
return 3
def setup(self):
img_size = 32
if self.rescale:
img_size //= self.rescale
if self.grayscale:
preprocessors = [
torchvision.transforms.Grayscale(),
torchvision.transforms.ToTensor(),
]
permutations_list = [
torchvision.transforms.Lambda(
lambda x: x.view(1, img_size * img_size).t()
) # (L, d_input)
]
if self.tokenize:
preprocessors.append(
torchvision.transforms.Lambda(lambda x: (x * 255).long())
)
permutations_list.append(Rearrange("l 1 -> l"))
else:
preprocessors.append(
torchvision.transforms.Normalize(
mean=122.6 / 255.0, std=61.0 / 255.0
)
)
else:
preprocessors = [
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(
(0.4914, 0.4822, 0.4465), (0.247, 0.243, 0.261)
),
]
permutations_list = [
torchvision.transforms.Lambda(
Rearrange("z h w -> (h w) z", z=3, h=img_size, w=img_size)
) # (L, d_input)
]
# Permutations and reshaping
if self.permute == "br":
permutation = permutations.bitreversal_permutation(img_size * img_size)
print("bit reversal", permutation)
permutations_list.append(torchvision.transforms.Lambda(lambda x: x[permutation]))
elif self.permute == "snake":
permutation = permutations.snake_permutation(img_size, img_size)
print("snake", permutation)
permutations_list.append(torchvision.transforms.Lambda(lambda x: x[permutation]))
elif self.permute == "hilbert":
permutation = permutations.hilbert_permutation(img_size)
print("hilbert", permutation)
permutations_list.append(torchvision.transforms.Lambda(lambda x: x[permutation]))
elif self.permute == "transpose":
permutation = permutations.transpose_permutation(img_size, img_size)
transform = torchvision.transforms.Lambda(
lambda x: torch.cat([x, x[permutation]], dim=-1)
)
permutations_list.append(transform)
elif self.permute == "2d": # h, w, c
permutation = torchvision.transforms.Lambda(
Rearrange("(h w) c -> h w c", h=img_size, w=img_size)
)
permutations_list.append(permutation)
elif self.permute == "2d_transpose": # c, h, w
permutation = torchvision.transforms.Lambda(
Rearrange("(h w) c -> c h w", h=img_size, w=img_size)
)
permutations_list.append(permutation)
# Augmentation
if self.augment:
augmentations = [
torchvision.transforms.RandomCrop(
img_size, padding=4, padding_mode="symmetric"
),
torchvision.transforms.RandomHorizontalFlip(),
]
post_augmentations = []
if self.cutout:
raise NotImplementedError("Cutout not currently supported.")
# post_augmentations.append(Cutout(1, img_size // 2))
pass
if self.random_erasing:
# augmentations.append(RandomErasing())
pass
else:
augmentations, post_augmentations = [], []
transforms_train = (
augmentations + preprocessors + post_augmentations + permutations_list
)
transforms_eval = preprocessors + permutations_list
transform_train = torchvision.transforms.Compose(transforms_train)
transform_eval = torchvision.transforms.Compose(transforms_eval)
self.dataset_train = torchvision.datasets.CIFAR10(
f"{default_data_path}/{self._name_}",
train=True,
download=True,
transform=transform_train,
)
self.dataset_test = torchvision.datasets.CIFAR10(
f"{default_data_path}/{self._name_}", train=False, transform=transform_eval
)
if self.rescale:
print(f"Resizing all images to {img_size} x {img_size}.")
self.dataset_train.data = self.dataset_train.data.reshape((self.dataset_train.data.shape[0], 32 // self.rescale, self.rescale, 32 // self.rescale, self.rescale, 3)).max(4).max(2).astype(np.uint8)
self.dataset_test.data = self.dataset_test.data.reshape((self.dataset_test.data.shape[0], 32 // self.rescale, self.rescale, 32 // self.rescale, self.rescale, 3)).max(4).max(2).astype(np.uint8)
self.split_train_val(self.val_split)
def __str__(self):
return f"{'p' if self.permute else 's'}{self._name_}"
class SpeechCommands(ResolutionSequenceDataset):
_name_ = "sc"
@property
def init_defaults(self):
return {
"mfcc": False,
"dropped_rate": 0.0,
"length": 16000,
"all_classes": False,
}
@property
def d_input(self):
_d_input = 20 if self.mfcc else 1
_d_input += 1 if self.dropped_rate > 0.0 else 0
return _d_input
@property
def d_output(self):
return 10 if not self.all_classes else 35
@property
def l_output(self):
return 0
@property
def L(self):
return 161 if self.mfcc else self.length
def setup(self):
self.data_dir = self.data_dir or default_data_path # TODO make same logic as other classes
from s5.dataloaders.sc import _SpeechCommands
# TODO refactor with data_dir argument
self.dataset_train = _SpeechCommands(
partition="train",
length=self.L,
mfcc=self.mfcc,
sr=self.sr,
dropped_rate=self.dropped_rate,
path=self.data_dir,
all_classes=self.all_classes,
)
self.dataset_val = _SpeechCommands(
partition="val",
length=self.L,
mfcc=self.mfcc,
sr=self.sr,
dropped_rate=self.dropped_rate,
path=self.data_dir,
all_classes=self.all_classes,
)
self.dataset_test = _SpeechCommands(
partition="test",
length=self.L,
mfcc=self.mfcc,
sr=self.sr,
dropped_rate=self.dropped_rate,
path=self.data_dir,
all_classes=self.all_classes,
)
|