-
Notifications
You must be signed in to change notification settings - Fork 1
/
train-fabric.py
188 lines (169 loc) · 7.48 KB
/
train-fabric.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
#!/usr/bin/env python
# coding: utf-8
import os
import comet_ml
import torch
import torch.nn as nn
import torchvision.transforms as tf
from torchvision.utils import make_grid,save_image
from PIL import Image
import torch.autograd as autograd
from networks import Discriminator,Generator2
from loss_network import LossNetwork,LossNetwork2
from loss import content_style_loss,adv_loss_d,adv_loss_g,gaze_loss_d,gaze_loss_g,reconstruction_loss
from PIL import Image
import numpy as np
#import lightning as L
import yaml
from munch import DefaultMunch
from tqdm import tqdm
from comet_ml.integration.pytorch import log_model
from random import randint
import socket
from data import PreProcessData,TrainDataset,TestDataset
from images import recover_image,save_images,prepare_images
# The images files have the form "ID_2m_0P_xV_yH_z.jpg" where ID is the ID of the person, 2m is fixed, 0P means head pose of 0 degrees (only head pose used in this notebook)
# x is the vertical orientation, y is the horizontal orientation and z is either L for left or R for right eye (note that the right eye patch was flipped horizontally).
# In training the images are grouped as follows:
# For a given person and a given eye (R or L) all orientations are grouped together. One element of the data set is of the form
# imgs_r,angles_r,labels,imgs_t,angles_g where imgs_r is considered the "real" image with orientation angles_r, or x_r in the paper,
# imgs_t with orientation angles_g is the image of the same person with different orientation (could be the same image since we go through a double loop) and the label is the ID of the person
torch.set_float32_matmul_precision('medium')
with open('config.yaml') as f:
config = yaml.safe_load(f)
config = DefaultMunch.fromDict(config)
print(config)
try:
os.makedirs(os.getcwd()+"/"+config.debug_path, exist_ok=True)
except:
print("cannot create directory ("+os.getcwd()+"/"+config.debug_path+")")
exit()
if config.use_comet == 'online':
try:
socket.gethostbyname('google.com')
except:
config.use_comet='offline'
if config.use_comet is not None:
if config.use_comet=='offline':
print("using comet offline")
experiment = comet_ml.OfflineExperiment(project_name=config.comet_project, workspace=config.comet_workspace,
auto_metric_logging=False, auto_output_logging=False)
else:
experiment = comet_ml.Experiment(project_name=config.comet_project, workspace=config.comet_workspace,
auto_metric_logging=False, auto_output_logging=False)
experiment.log_parameters(config)
else:
print("don't use comet")
data=PreProcessData(config.data_path)
transform=tf.Compose([tf.ToTensor(),tf.Resize((64,64),antialias=True)])
train_dataset=TrainDataset(*data.training_data(),transform=transform)
test_dataset=TestDataset(*data.testing_data(),transform=transform)
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=config.batch_size, shuffle=True)
test_loader = torch.utils.data.DataLoader(test_dataset, batch_size=config.test_batch_size, shuffle=False)
device='cuda' if torch.cuda.is_available() else 'cpu'
if os.path.isfile('discriminator.pth') and config.resume_training:
discriminator=torch.load('discriminator.pth')
print('loaded discriminator')
else:
discriminator=Discriminator()
print('created discriminator')
if os.path.isfile('generator.pth') and config.resume_training:
generator=torch.load('generator.pth')
print('loaded generator')
else:
generator=Generator2()
print('created generator')
LR=config.lr
beta1=config.beta1
beta2=config.beta2
optimizer_g = torch.optim.Adam(generator.parameters(), LR,betas=(beta1, beta2))
optimizer_d = torch.optim.Adam(discriminator.parameters(), LR,betas=(beta1, beta2))
loss_network=LossNetwork2()
#loss_network=LossNetwork()
#loss_network=loss_network.to(device)
from lightning.fabric import Fabric
fabric=Fabric(accelerator='cuda',devices=1,precision="bf16-mixed")
fabric.launch()
generator,optimizer_g=fabric.setup(generator,optimizer_g)
discriminator,optimizer_d=fabric.setup(discriminator,optimizer_d)
loss_network=fabric.setup(loss_network)
def generator_step(generator,discriminator,loss_network,batch):
imgs_r, angles_r, _, imgs_t, angles_g=batch
optimizer_g.zero_grad()
generator.train()
discriminator.eval()
x_g=generator(imgs_r,angles_g)
x_recon=generator(x_g,angles_r)
with fabric.autocast():
loss_adv=adv_loss_g(discriminator,imgs_r,x_g)
loss2=content_style_loss(loss_network,x_g,imgs_t)
loss_p=loss2[0]+loss2[1]
loss_gg=gaze_loss_g(discriminator,x_g,angles_g)
loss_recon=reconstruction_loss(generator,imgs_r,x_recon)
loss=loss_adv+config.lambda_p*loss_p+config.lambda_gaze*loss_gg+config.lambda_recon*loss_recon
#loss.backward()
fabric.backward(loss)
optimizer_g.step()
return loss.item()
def discriminator_step(generator,discriminator,batch):
imgs_r, angles_r, _, _, angles_g=batch
optimizer_d.zero_grad()
generator.eval()
discriminator.train()
x_g=generator(imgs_r,angles_g)
with fabric.autocast():
loss1=adv_loss_d(discriminator,imgs_r,x_g)
loss2=gaze_loss_d(discriminator,imgs_r,angles_r)
loss=loss1+config.lambda_gaze*loss2
#loss.backward()
fabric.backward(loss)
optimizer_d.step()
return loss.item()
loop=tqdm(range(config.epochs))
#for epoch in loop:
for epoch in range(config.epochs):
#loop.set_description(f"Epoch [{epoch+1}/{config.epochs}]")
batch_count=0
loss_d,loss_g=0.,0.
for batch in tqdm(train_loader):
batch_count+=1
batch=[x.to(device) for x in batch]
l_d=discriminator_step(generator,discriminator,batch)
loss_d=0.9*loss_d+0.1*l_d
if batch_count%config.critic_iter_per_gen==0:
l_g=generator_step(generator,discriminator,loss_network,batch)
loss_g=0.9*loss_g+0.1*l_g
if epoch%config.image_save_freq==0:
try:
orig=next(test_iter)[0].to(device)
except:
test_iter=iter(test_loader)
orig=next(test_iter)[0].to(device)# extract the images
imgs=[torch.unbind(orig,0)]
for h in [-15,-10,-5,0,5,10,15]:
a=torch.tile(torch.tensor([h/15.,0.]),[config.test_batch_size,1])
a=a.to(device)
y=generator(orig,a).detach()
y=torch.unbind(y,0)
imgs.append(y)
imgs=prepare_images(imgs,config.test_batch_size)
filename=config.debug_path+"/{}_{}.png".format(epoch,batch_count)
img=make_grid(imgs,nrow=8)#nrow specifies the number of columns !!!
img=recover_image(img)
img=Image.fromarray(img)
img.save(filename)
#save_image(imgs,filename,nrow=8)# nrow specifies the number of columns !!!
if config.use_comet is not None:
experiment.log_image(img)
#loop.set_postfix(loss_d=loss_d,loss_g=loss_g)
if config.use_comet is not None:
metrics={'loss_d':l_d,'loss_g':l_g}
experiment.log_metrics(metrics, epoch=epoch)
if epoch%config.model_save_freq==0:
if config.use_comet is not None:
log_model(experiment,generator,"generator")
log_model(experiment,discriminator,"discriminator")
# torch.save(generator, './generator.pth')
# torch.save(discriminator, './discriminator.pth')
torch.save(generator.state_dict(), './generator.pth')
torch.save(discriminator.state_dict(), './discriminator.pth')