-
Notifications
You must be signed in to change notification settings - Fork 0
/
adain_train.py
49 lines (40 loc) · 1.62 KB
/
adain_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
from torch.utils.data import DataLoader
from dataprocess.StyleTransferDataset import StyleTransferDataset
from resources.utilities import *
from Model.AdaIN import AdaIN
import pprint
# DEBUG ONLY
def rename(directory):
import os
for i, filename in enumerate(sorted(os.listdir(directory))):
if filename.endswith(".jpg") or filename.endswith(".png"):
ext = os.path.splitext(filename)[1]
os.rename(os.path.join(directory, filename), os.path.join(directory, str(i)+ext))
# FOR TEST PURPOSES
if __name__ == "__main__":
# TRAIN
transformed_dataset = StyleTransferDataset(
CONTENT_PATH,
STYLE_PATH,
transform=transforms.Compose([
transforms.Resize(256),
transforms.RandomCrop(224),
transforms.ToTensor()]))
dataloader = DataLoader(transformed_dataset, batch_size=16,
shuffle=True)
adain = AdaIN()
pprint.pprint(adain.encoder)
pprint.pprint(adain.decoder)
# Save the random weights for reuse
torch.save(adain.decoder.state_dict(), "decoder_random.pth")
# Train with different style weights to see the difference
import os
if not os.path.exists(RESULTS_PATH):
os.mkdir(RESULTS_PATH)
style_weights = [10e2]
for sw in style_weights:
if not os.path.exists(os.path.join(RESULTS_PATH, "SW_{0}".format(sw))):
os.mkdir(os.path.join(RESULTS_PATH, "SW_{0}".format(sw)))
adain.train(dataloader=dataloader, style_weight=sw, epochs=25)
# Reset decoder to starting weights
adain.decoder.load_state_dict(torch.load("decoder_random.pth"))