-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathsingle_pixel_experiments.py
173 lines (144 loc) · 8.14 KB
/
single_pixel_experiments.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
import os
import sys
import torchbearer
from torchbearer.callbacks import EarlyStopping, on_end_epoch, CSVLogger
from torchvision import transforms
from torchvision.models import ResNet18_Weights, resnet18, ResNet101_Weights, resnet101
from torchvision.models.vgg import vgg16, VGG16_Weights
from torch.utils.data import DataLoader
from datasets.singlepixel import MalhotraSinglePixelDataset, SinglePixelDataset, PixelClusterDataset
from utils.modelfitting import fit_model, evaluate_model, set_seed
import argparse
import torch.nn as nn
import torch
def load_data(data_dir, size, seed, batch_size, num_workers, pos_sd, col_sd, augment, presize, dataset_cls="MalhotraSinglePixelDataset"):
tf = transforms.Compose([transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
if augment:
train_tf = transforms.Compose([transforms.ToTensor(),
transforms.RandomHorizontalFlip(),
transforms.RandomAffine(degrees=0, translate=(0.1, 0.1)),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
else:
train_tf = tf
shared_args_ds = {'download': True, 'seed': seed, 'size': size, 'position_sd': pos_sd, 'colour_sd': col_sd,
'presize': presize}
train_ds = dataset_cls(data_dir, train=True, transform=train_tf, **shared_args_ds)
same_ds = dataset_cls(data_dir, train=False, condition='same', transform=tf, **shared_args_ds)
diff_ds = dataset_cls(data_dir, train=False, condition='diff', transform=tf, **shared_args_ds)
nopix_ds = dataset_cls(data_dir, train=False, condition='nopix', transform=tf, **shared_args_ds)
train_nopix_ds = dataset_cls(data_dir, train=True, condition='nopix', transform=train_tf, **shared_args_ds)
shared_args_loader = {'pin_memory': True, 'batch_size': batch_size, 'num_workers': num_workers}
return {'train': DataLoader(train_ds, shuffle=True, **shared_args_loader),
'val_same': DataLoader(same_ds, shuffle=False, **shared_args_loader),
'val_diff': DataLoader(diff_ds, shuffle=False, **shared_args_loader),
'val_nopix': DataLoader(nopix_ds, shuffle=False, **shared_args_loader),
'train_nopix': DataLoader(train_nopix_ds, shuffle=False, **shared_args_loader)}
def val_callback(datasets):
@on_end_epoch
def cb(state):
trial = torchbearer.Trial(state[torchbearer.MODEL], state[torchbearer.OPTIMIZER], state[torchbearer.CRITERION],
metrics=state[torchbearer.METRIC_LIST].metric_list, verbose=2)
for name, ds in datasets.items():
if name.startswith('val_'):
res = trial.with_val_generator(ds).to(state[torchbearer.DEVICE]).evaluate()
for k, v in res.items():
if k.startswith('val_'):
nk = k.replace('val_', f'{name}_')
state[torchbearer.METRICS][nk] = v
return cb
if __name__ == '__main__':
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--size', type=int, default='224', choices=[32, 224])
parser.add_argument('--model', type=str, default='vgg16', choices=['vgg16', 'resnet18', 'resnet18_3x3', 'resnet101'])
parser.add_argument('--pretrained', action="store_true", default=False)
parser.add_argument('--freeze', action="store_true", default=False)
parser.add_argument('--data-seed', type=int, default=0)
parser.add_argument('--seed', type=int, default=0)
parser.add_argument('--log-dir', type=str, default=None)
parser.add_argument('--data-dir', type=str, default="/scratch/jsh2/datasets")
parser.add_argument('--batch-size', type=int, default=256)
parser.add_argument('--max-epochs', type=int, default=150)
parser.add_argument('--checkpoint-period', type=int, default=10)
parser.add_argument('--resume', type=str, default=None)
parser.add_argument('--num-workers', type=int, default=8)
parser.add_argument('--position-sd', type=float, default=1.0)
parser.add_argument('--colour-sd', type=float, default=1.0)
parser.add_argument('--momentum', type=float, default=0.9)
parser.add_argument('--weight-decay', type=float, default=0)
parser.add_argument('--augment', action="store_true", default=False)
parser.add_argument('--post-size', action="store_true", default=False)
parser.add_argument('--early-stop', action="store_true", default=False)
parser.add_argument("--dataset-class", type=str, default="MalhotraSinglePixelDataset",
choices=["MalhotraSinglePixelDataset", "SinglePixelDataset", "PixelClusterDataset"],)
args = parser.parse_args()
set_seed(args.seed)
if args.model == 'vgg16':
if args.pretrained:
weights = VGG16_Weights.IMAGENET1K_V1
net = vgg16(weights=weights)
net.classifier[-1] = nn.Linear(net.classifier[-1].in_features, 10)
if args.freeze:
for param in net.features:
param.requires_grad = False
else:
net = vgg16(num_classes=10)
elif args.model == 'resnet18':
if args.pretrained:
weights = ResNet18_Weights.IMAGENET1K_V1
net = resnet18(weights=weights)
if args.freeze:
for param in net:
param.requires_grad = False
net.fc = nn.Linear(net.fc.in_features, 10)
else:
net = resnet18(num_classes=10)
elif args.model == 'resnet18_3x3':
if args.pretrained:
raise Exception("Not Implemented")
else:
net = resnet18(num_classes=10)
net.conv1 = nn.Conv2d(3, 64,
kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
elif args.model == 'resnet101':
if args.pretrained:
weights = ResNet101_Weights.IMAGENET1K_V1
net = resnet101(weights=weights)
if args.freeze:
for param in net:
param.requires_grad = False
net.fc = nn.Linear(net.fc.in_features, 10)
else:
net = resnet101(num_classes=10)
else:
sys.exit(f"Unsupported model {args.model}")
loss = nn.CrossEntropyLoss()
lr = 1e-5 if args.pretrained else 1e-3
opt = torch.optim.SGD(filter(lambda p: p.requires_grad, net.parameters()), lr=lr, momentum=args.momentum,
weight_decay=args.weight_decay)
data = load_data(args.data_dir, args.size, args.data_seed, args.batch_size, args.num_workers,
args.position_sd, args.colour_sd, args.augment, not args.post_size,
dataset_cls=globals()[args.dataset_class])
ds = f"-{args.dataset_class}" if args.dataset_class != 'MalhotraSinglePixelDataset' else ''
run_id = f"{args.model}-{args.size}{'-pre' if args.pretrained else ''}-seed_{args.seed}" + \
f"-dataseed_{args.data_seed}-pos_sd_{args.position_sd}-col_sd_{args.colour_sd}" + \
f"-momentum_{args.momentum}-wd_{args.weight_decay}{'-frozen' if args.freeze else ''}" + \
f"{'-augment' if args.augment else ''}{'-post-size' if args.post_size else ''}" + \
f"{ds}"
if args.early_stop:
extra_callbacks = [EarlyStopping(monitor='acc', patience=5, mode='max')]
else:
extra_callbacks = []
if args.log_dir:
args.log_dir = f"{args.log_dir}/{run_id}"
models_dir = f"{args.log_dir}/models"
os.makedirs(models_dir, exist_ok=True)
model_file = models_dir + "/weights_{epoch:03d}.pt"
extra_callbacks.append(CSVLogger(f"{args.log_dir}/log.csv"))
torch.save({"model": net.state_dict()}, f"{models_dir}/weights_init.pt")
else:
model_file = None
pre_extra_callbacks = [val_callback(data)]
fit_model(net, loss, opt, data['train'], None, epochs=args.max_epochs, device='auto',
verbose=2, acc='acc', model_file=model_file, run_id=run_id, log_dir=args.log_dir, resume=args.resume,
pre_extra_callbacks=pre_extra_callbacks, extra_callbacks=extra_callbacks, period=args.checkpoint_period)