-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
129 lines (106 loc) · 4.61 KB
/
train.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
import torch
import argparse
import losses as l
import torch.nn as nn
from data import LoadData
from simclr import SimCLR
import torch.nn.functional as F
from pl_bolts.models.self_supervised.simclr.transforms import (
SimCLRTrainDataTransform as DefaultTrain,
SimCLREvalDataTransform as DefaultEval,
)
parser = argparse.ArgumentParser()
parser.add_argument("model")
parser.add_argument("data_aug")
parser.add_argument("batch_size", type=int)
parser.add_argument("epochs", type=int)
parser.add_argument("save_as", type=int)
args = parser.parse_args()
class FineTune(nn.Module):
"""Projection module for SimCLR (Pytorch Lightning implementation)"""
def __init__(self, model):
super().__init__()
self.model = model
self.linear = nn.Linear(2048, 5)
def forward(self, x):
return self.linear(self.model(x))
def train(model, data_aug, batch_size, epochs, save_as):
if model == "simclr"
if data_aug == "default":
data = LoadData([DefaultTrain(100), DefaultEval(100)]).generate_split_dataloader()
else:
data = LoadData([LoadData.random_masking_transform()]).generate_split_dataloader()
else:
data = LoadData([LoadData.default_transform()]).generate_dataloader()
masked_test_data = LoadData([LoadData.default_transform()], "LFW_masked").generate_dataloader()
if model == "simclr":
model = torch.jit.script(SimCLR(batch_size, len(data(1, "train")), epochs=epochs))
optimizer, scheduler = model.configure_optimizer()
loss_fn = l. NT_Xent(batch_size)
mode = "SSL"
elif "simclr" in model:
model = torch.jit.script(SimCLR(32, len(loader(1, "train")), epochs=epochs))
model.load_state_dict(torch.load(model))
model = FineTune(model)
mode = " "
else: pass
train = data(batch_size, "train")
val = data(batch_size, "val")
test = data(batch_size, "test")
masked_test = data(batch_size)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = model.to(device)
for epoch in range(epochs):
print(f"Epoch: {epoch}")
loss_l, acc_l = [], []
for (data, labels) in train:
optimizer.zero_grad()
if mode == "SSL":
logits = [model(i.to(device)) for i in data]
loss = loss_fn(*logits)
else:
logits = model(data.to(device))
loss = loss_fn(logits, labels)
loss.backward()
optimizer.step()
loss_l.append(loss.item())
if mode != "SSL":
acc_l.append(get_acc(logits, l))
wandb.log({"train_loss" : torch.mean(torch.tensor(loss_l)), "epoch" : epoch})
if mode != "SSL":
wandb.log({"train_acc" : torch.mean(torch.tensor(acc_l)), "epoch" : epoch})
vloss_l, vacc_l = [], []
for (data, labels) in val:
if mode == "SSL":
logits = [model(i.to(device)) for i in data]
loss = loss_fn(*logits)
else:
logits = model(data.to(device))
loss = loss_fn(logits, labels)
vloss_l.append(loss.item())
if mode != "SSL":
vacc_l.append(get_acc(logits, l))
wandb.log({"val_loss" : torch.mean(torch.tensor(vloss_l)), "epoch" : epoch})
if mode != "SSL":
wandb.log({"val_acc" : torch.mean(torch.tensor(vacc_l)), "epoch" : epoch})
scheduler.step()
if mode != "SSL":
tloss_l, tacc_l = [], []
for (data, labels) in test:
logits = model(data.to(device))
loss = loss_fn(logits, labels)
tloss_l.append(loss.item())
tacc_l.append(get_acc(logits, l))
wandb.log({"unmasked_test_loss" : torch.mean(torch.tensor(tloss_l))})
wandb.log({"unmasked_test_loss" : torch.mean(torch.tensor(tacc_l))})
tloss_l, tacc_l = [], []
for (data, labels) in masked_test:
logits = model(data.to(device))
loss = loss_fn(logits, labels)
tloss_l.append(loss.item())
tacc_l.append(get_acc(logits, l))
wandb.log({"unmasked_test_loss" : torch.mean(torch.tensor(tloss_l))})
wandb.log({"unmasked_test_loss" : torch.mean(torch.tensor(tacc_l))})
torch.save(model.state_dict(), f"{save_as}.pt")
if __name__ == "__main__":
train(model, data_aug, batch_size, epochs, save_as)