-
Notifications
You must be signed in to change notification settings - Fork 0
/
image.py
179 lines (161 loc) · 7.03 KB
/
image.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
import argparse
import os
import sys
import time
if (slurm_submit_dir := os.environ.get('SLURM_SUBMIT_DIR', default=None)) is not None:
sys.path.append(os.environ['SLURM_SUBMIT_DIR'])
import torch
from eval.models.image import UnitaryCCT, SinusoidalCCT, AbsoluteCCT, UnitarySeqCCT
from eval.models.image.augmentations import CIFAR10Policy
from unitaryPE.nn.schedule import make_schedule
from torchvision.datasets import CIFAR10
from torchvision import transforms
from torch.utils.data import DataLoader
from torch.optim import AdamW
from torch.optim.lr_scheduler import LambdaLR
from typing import Literal
def run(
model: Literal['Unitary', 'Sinusoidal', 'Absolute', 'UnitarySeq'],
num_epochs: int,
num_layers: int,
num_heads: int,
dim: int,
mlp_ratio: int,
dataset: Literal['cifar10', 'mnist', 'cifar100'],
data_dir: str,
batch_size: int,
seed: int = 42,
store_path: str | None = None):
start_time = time.time()
match dataset:
case 'cifar10':
augmentations = [CIFAR10Policy(),
transforms.RandomCrop((32, 32), padding=4),
transforms.RandomHorizontalFlip()]
transformations = [transforms.ToTensor(),
transforms.Normalize(mean=[0.4914, 0.4822, 0.4465], std=[0.2470, 0.2435, 0.2616])]
train_set = CIFAR10(data_dir, train=True, download=True,
transform=transforms.Compose([*augmentations, *transformations]))
test_set = CIFAR10(data_dir, train=False, download=True,
transform=transforms.Compose(transformations))
in_channels, num_classes, image_size = 3, 10, 10
case _:
raise NotImplementedError
train_dl = DataLoader(train_set, batch_size=batch_size, shuffle=True, num_workers=4)
test_dl = DataLoader(test_set, batch_size=batch_size, shuffle=False, num_workers=4)
torch.manual_seed(seed)
match model:
case 'Unitary':
model = UnitaryCCT(
dim=dim,
num_heads=num_heads,
num_layers=num_layers,
kernel_size=(3, 3),
in_channels=in_channels,
num_classes=num_classes,
mlp_ratio=mlp_ratio).cuda()
case 'Sinusoidal':
model = SinusoidalCCT(
dim=dim,
num_heads=num_heads,
num_layers=num_layers,
kernel_size=(3, 3),
in_channels=in_channels,
num_classes=num_classes,
mlp_ratio=mlp_ratio).cuda()
case 'Absolute':
model = AbsoluteCCT(
dim=dim,
num_heads=num_heads,
num_layers=num_layers,
kernel_size=(3, 3),
in_channels=in_channels,
num_classes=num_classes,
mlp_ratio=mlp_ratio,
num_embeddings=image_size).cuda()
case 'UnitarySeq':
model = UnitarySeqCCT(
dim=dim,
num_heads=num_heads,
num_layers=num_layers,
kernel_size=(3, 3),
in_channels=in_channels,
num_classes=num_classes,
mlp_ratio=mlp_ratio).cuda()
case _:
raise ValueError
steps_per_epoch = len(train_dl)
optim = AdamW(model.parameters(), lr=1, weight_decay=3e-2)
scheduler = LambdaLR(
optimizer=optim,
lr_lambda=make_schedule(
warmup_steps=steps_per_epoch * 5,
warmdown_steps=steps_per_epoch * (num_epochs - 5),
total_steps=steps_per_epoch * num_epochs,
min_lr=1e-9,
max_lr=5e-4,
init_lr=1e-7))
best_epoch, best_dev_acc = None, -1e10
for epoch in range(num_epochs):
model.train()
print(f'{epoch}')
epoch_loss, batch_correct, total_correct, rma = 0, 0, 0, 0
for batch_input, target in train_dl:
batch_input = batch_input.cuda()
target = target.cuda()
preds = model.forward(batch_input)
loss = torch.nn.functional.cross_entropy(preds, target, label_smoothing=0.1)
loss.backward()
optim.step()
scheduler.step()
optim.zero_grad()
epoch_loss += loss.item()
batch_correct = (preds.argmax(dim=-1) == target).sum()
total_correct += batch_correct
print(f'Train loss {epoch_loss}')
print(f'Accuracy (m) {total_correct/len(train_set)}')
model.eval()
epoch_loss, total_correct = 0, 0
with torch.no_grad():
for batch_input, target in test_dl:
batch_input = batch_input.cuda()
target = target.cuda()
preds = model.forward(batch_input)
loss = torch.nn.functional.cross_entropy(preds, target)
epoch_loss += loss.item()
total_correct += (preds.argmax(dim=-1) == target).sum()
print(f'Test loss {epoch_loss}')
print(f'Dev acc (token) {(dev_acc := total_correct / len(test_set))}')
if dev_acc > best_dev_acc and store_path is not None:
best_epoch, best_dev_acc = epoch, dev_acc
torch.save(model.state_dict(), store_path)
sys.stdout.flush()
duration = time.time() - start_time
print(f'Training took {duration} seconds. Best epoch was {best_epoch}')
sys.stdout.flush()
def parse_args():
parser = argparse.ArgumentParser(description='Run a single training iteration')
parser.add_argument('--model', type=str, required=True, choices=['Unitary', 'Sinusoidal', 'Absolute', 'UnitarySeq'], help='Type of model to use')
parser.add_argument('--data_dir', type=str, required=True, help='Where is the data located')
parser.add_argument('--num_epochs', type=int, default=300, help='Number of training epochs')
parser.add_argument('--num_layers', type=int, default=7, help='Number of layers for the model')
parser.add_argument('--dim', type=int, default=256, help='Dimension of the model')
parser.add_argument('--num_heads', type=int, default=4, help='Number of attention heads')
parser.add_argument('--store_path', type=str, default=None, help='If/where to store the trained model')
parser.add_argument('--seed', type=int, default=42, help='The id of the current repetition')
parser.add_argument('--batch_size', type=int, default=128, help='Batch size to train with')
parser.add_argument('--mlp_ratio', type=int, default=2, help='How big the intermediate FF dimension is')
return parser.parse_args()
if __name__ == '__main__':
args = parse_args()
run(model=args.model,
num_heads=args.num_heads,
num_epochs=args.num_epochs,
num_layers=args.num_layers,
dim=args.dim,
seed=args.seed,
dataset='cifar10',
store_path=args.store_path,
batch_size=args.batch_size,
mlp_ratio=args.mlp_ratio,
data_dir=args.data_dir)