-
Notifications
You must be signed in to change notification settings - Fork 13
/
Copy pathhelpers.py
205 lines (172 loc) · 7.23 KB
/
helpers.py
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
import coinpp.conversion as conversion
import coinpp.models as models
import data.audio as audio
import data.era5 as era5
import data.fastmri as fastmri
import data.image as image
import data.vimeo90k as vimeo90k
import torchvision
import yaml
from pathlib import Path
def get_dataset_root(dataset_name: str):
"""Returns path to data based on dataset_paths.yaml file."""
with open(r"data/dataset_paths.yaml") as f:
dataset_paths = yaml.safe_load(f)
return Path(dataset_paths[dataset_name])
def dataset_name_to_dims(dataset_name):
"""Returns appropriate dim_in and dim_out for dataset."""
if dataset_name == "mnist":
dim_in, dim_out = 2, 1
if dataset_name in ("cifar10", "kodak", "vimeo90k"):
dim_in, dim_out = 2, 3
if dataset_name == "fastmri":
dim_in, dim_out = 3, 1
if dataset_name == "era5":
dim_in = 3
dim_out = 1
if dataset_name == "librispeech":
dim_in = 1
dim_out = 1
return dim_in, dim_out
def get_datasets_and_converter(args, force_no_random_crop=False):
"""Returns train and test datasets as well as appropriate data converters.
Args:
args: Arguments parsed from input.
force_no_random_crop (bool): If True, forces datasets to not use random
crops (which is the default for the training set when using
patching). This is useful after the model is trained when we store
modulations.
"""
# Extract input and output dimensions of function rep
dim_in, dim_out = dataset_name_to_dims(args.train_dataset)
# When using patching, perform random crops equal to patch size on training
# dataset
use_patching = hasattr(args, "patch_shape") and args.patch_shape != [-1]
if use_patching:
if dim_in == 2:
random_crop = torchvision.transforms.RandomCrop(args.patch_shape)
if "mnist" in (args.train_dataset, args.test_dataset):
converter = conversion.Converter("image")
if args.train_dataset == "mnist":
train_dataset = image.MNIST(
root=get_dataset_root("mnist"),
train=True,
transform=random_crop if use_patching else None,
)
if args.test_dataset == "mnist":
test_dataset = image.MNIST(root=get_dataset_root("mnist"), train=False)
if "cifar10" in (args.train_dataset, args.test_dataset):
converter = conversion.Converter("image")
if args.train_dataset == "cifar10":
transforms = [torchvision.transforms.ToTensor()]
if use_patching and not force_no_random_crop:
transforms.append(random_crop)
train_dataset = image.CIFAR10(
root=get_dataset_root("cifar10"),
train=True,
transform=torchvision.transforms.Compose(transforms),
)
if args.test_dataset == "cifar10":
test_dataset = image.CIFAR10(
root=get_dataset_root("cifar10"),
train=False,
transform=torchvision.transforms.ToTensor(),
)
if "kodak" in (args.train_dataset, args.test_dataset):
converter = conversion.Converter("image")
if args.train_dataset == "kodak":
transforms = [torchvision.transforms.ToTensor()]
if use_patching and not force_no_random_crop:
transforms.append(random_crop)
train_dataset = image.Kodak(
root=get_dataset_root("kodak"),
download=True,
transform=torchvision.transforms.Compose(transforms),
)
if args.test_dataset == "kodak":
test_dataset = image.Kodak(
root=get_dataset_root("kodak"),
download=True,
transform=torchvision.transforms.ToTensor(),
)
if "fastmri" in (args.train_dataset, args.test_dataset):
converter = conversion.Converter("mri")
if args.train_dataset == "fastmri":
train_dataset = fastmri.FastMRI(
root=get_dataset_root("fastmri"),
split="train",
challenge="multicoil",
patch_shape=args.patch_shape
if (use_patching and not force_no_random_crop)
else -1,
)
if args.test_dataset == "fastmri":
test_dataset = fastmri.FastMRI(
root=get_dataset_root("fastmri"),
split="val",
challenge="multicoil",
)
if "era5" in (args.train_dataset, args.test_dataset):
converter = conversion.Converter("era5")
if args.train_dataset == "era5":
train_dataset = era5.ERA5(root=get_dataset_root("era5"), split="train")
if args.test_dataset == "era5":
test_dataset = era5.ERA5(root=get_dataset_root("era5"), split="test")
if "vimeo90k" in (args.train_dataset, args.test_dataset):
converter = conversion.Converter("image")
if args.train_dataset == "vimeo90k":
transforms = [torchvision.transforms.ToTensor()]
if use_patching and not force_no_random_crop:
transforms.append(random_crop)
# As Vimeo90k dataset is large, do not download by default. If you wish to
# download this dataset, manually set download=True.
train_dataset = vimeo90k.Vimeo90k(
root=get_dataset_root("vimeo90k"),
train=True,
download=False,
transform=torchvision.transforms.Compose(transforms),
)
if args.test_dataset == "vimeo90k":
test_dataset = vimeo90k.Vimeo90k(
root=get_dataset_root("vimeo90k"),
train=False,
download=False,
transform=torchvision.transforms.ToTensor(),
)
if "librispeech" in (args.train_dataset, args.test_dataset):
converter = conversion.Converter("audio")
# We use first 3 seconds of each audio sample
if args.train_dataset == "librispeech":
train_dataset = audio.LIBRISPEECH(
root=get_dataset_root("librispeech"),
url="train-clean-100",
patch_shape=args.patch_shape[0]
if (use_patching and not force_no_random_crop)
else -1,
num_secs=3,
download=True,
)
if args.test_dataset == "librispeech":
test_dataset = audio.LIBRISPEECH(
root=get_dataset_root("librispeech"),
url="test-clean",
num_secs=3,
download=True,
)
return train_dataset, test_dataset, converter
def get_model(args):
dim_in, dim_out = dataset_name_to_dims(args.train_dataset)
return models.ModulatedSiren(
dim_in=dim_in,
dim_hidden=args.dim_hidden,
dim_out=dim_out,
num_layers=args.num_layers,
w0=args.w0,
w0_initial=args.w0,
modulate_scale=args.modulate_scale,
modulate_shift=args.modulate_shift,
use_latent=args.use_latent,
latent_dim=args.latent_dim,
modulation_net_dim_hidden=args.modulation_net_dim_hidden,
modulation_net_num_layers=args.modulation_net_num_layers,
).to(args.device)