nikunjkdtechnoland
commited on
Commit
•
575c901
1
Parent(s):
518eb4f
trainer add
Browse files- trainer.py +155 -0
trainer.py
ADDED
@@ -0,0 +1,155 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import torch
|
3 |
+
import torch.nn as nn
|
4 |
+
from torch import autograd
|
5 |
+
from model.networks import Generator, LocalDis, GlobalDis
|
6 |
+
|
7 |
+
|
8 |
+
from utils.tools import get_model_list, local_patch, spatial_discounting_mask
|
9 |
+
from utils.logger import get_logger
|
10 |
+
|
11 |
+
logger = get_logger()
|
12 |
+
|
13 |
+
|
14 |
+
class Trainer(nn.Module):
|
15 |
+
def __init__(self, config):
|
16 |
+
super(Trainer, self).__init__()
|
17 |
+
self.config = config
|
18 |
+
self.use_cuda = self.config['cuda']
|
19 |
+
self.device_ids = self.config['gpu_ids']
|
20 |
+
|
21 |
+
self.netG = Generator(self.config['netG'], self.use_cuda, self.device_ids)
|
22 |
+
self.localD = LocalDis(self.config['netD'], self.use_cuda, self.device_ids)
|
23 |
+
self.globalD = GlobalDis(self.config['netD'], self.use_cuda, self.device_ids)
|
24 |
+
|
25 |
+
self.optimizer_g = torch.optim.Adam(self.netG.parameters(), lr=self.config['lr'],
|
26 |
+
betas=(self.config['beta1'], self.config['beta2']))
|
27 |
+
d_params = list(self.localD.parameters()) + list(self.globalD.parameters())
|
28 |
+
self.optimizer_d = torch.optim.Adam(d_params, lr=config['lr'],
|
29 |
+
betas=(self.config['beta1'], self.config['beta2']))
|
30 |
+
if self.use_cuda:
|
31 |
+
self.netG.to(self.device_ids[0])
|
32 |
+
self.localD.to(self.device_ids[0])
|
33 |
+
self.globalD.to(self.device_ids[0])
|
34 |
+
|
35 |
+
def forward(self, x, bboxes, masks, ground_truth, compute_loss_g=False):
|
36 |
+
self.train()
|
37 |
+
l1_loss = nn.L1Loss()
|
38 |
+
losses = {}
|
39 |
+
|
40 |
+
x1, x2, offset_flow = self.netG(x, masks)
|
41 |
+
local_patch_gt = local_patch(ground_truth, bboxes)
|
42 |
+
x1_inpaint = x1 * masks + x * (1. - masks)
|
43 |
+
x2_inpaint = x2 * masks + x * (1. - masks)
|
44 |
+
local_patch_x1_inpaint = local_patch(x1_inpaint, bboxes)
|
45 |
+
local_patch_x2_inpaint = local_patch(x2_inpaint, bboxes)
|
46 |
+
|
47 |
+
# D part
|
48 |
+
# wgan d loss
|
49 |
+
local_patch_real_pred, local_patch_fake_pred = self.dis_forward(
|
50 |
+
self.localD, local_patch_gt, local_patch_x2_inpaint.detach())
|
51 |
+
global_real_pred, global_fake_pred = self.dis_forward(
|
52 |
+
self.globalD, ground_truth, x2_inpaint.detach())
|
53 |
+
losses['wgan_d'] = torch.mean(local_patch_fake_pred - local_patch_real_pred) + \
|
54 |
+
torch.mean(global_fake_pred - global_real_pred) * self.config['global_wgan_loss_alpha']
|
55 |
+
# gradients penalty loss
|
56 |
+
local_penalty = self.calc_gradient_penalty(
|
57 |
+
self.localD, local_patch_gt, local_patch_x2_inpaint.detach())
|
58 |
+
global_penalty = self.calc_gradient_penalty(self.globalD, ground_truth, x2_inpaint.detach())
|
59 |
+
losses['wgan_gp'] = local_penalty + global_penalty
|
60 |
+
|
61 |
+
# G part
|
62 |
+
if compute_loss_g:
|
63 |
+
sd_mask = spatial_discounting_mask(self.config)
|
64 |
+
losses['l1'] = l1_loss(local_patch_x1_inpaint * sd_mask, local_patch_gt * sd_mask) * \
|
65 |
+
self.config['coarse_l1_alpha'] + \
|
66 |
+
l1_loss(local_patch_x2_inpaint * sd_mask, local_patch_gt * sd_mask)
|
67 |
+
losses['ae'] = l1_loss(x1 * (1. - masks), ground_truth * (1. - masks)) * \
|
68 |
+
self.config['coarse_l1_alpha'] + \
|
69 |
+
l1_loss(x2 * (1. - masks), ground_truth * (1. - masks))
|
70 |
+
|
71 |
+
# wgan g loss
|
72 |
+
local_patch_real_pred, local_patch_fake_pred = self.dis_forward(
|
73 |
+
self.localD, local_patch_gt, local_patch_x2_inpaint)
|
74 |
+
global_real_pred, global_fake_pred = self.dis_forward(
|
75 |
+
self.globalD, ground_truth, x2_inpaint)
|
76 |
+
losses['wgan_g'] = - torch.mean(local_patch_fake_pred) - \
|
77 |
+
torch.mean(global_fake_pred) * self.config['global_wgan_loss_alpha']
|
78 |
+
|
79 |
+
return losses, x2_inpaint, offset_flow
|
80 |
+
|
81 |
+
def dis_forward(self, netD, ground_truth, x_inpaint):
|
82 |
+
assert ground_truth.size() == x_inpaint.size()
|
83 |
+
batch_size = ground_truth.size(0)
|
84 |
+
batch_data = torch.cat([ground_truth, x_inpaint], dim=0)
|
85 |
+
batch_output = netD(batch_data)
|
86 |
+
real_pred, fake_pred = torch.split(batch_output, batch_size, dim=0)
|
87 |
+
|
88 |
+
return real_pred, fake_pred
|
89 |
+
|
90 |
+
# Calculate gradient penalty
|
91 |
+
def calc_gradient_penalty(self, netD, real_data, fake_data):
|
92 |
+
batch_size = real_data.size(0)
|
93 |
+
alpha = torch.rand(batch_size, 1, 1, 1)
|
94 |
+
alpha = alpha.expand_as(real_data)
|
95 |
+
if self.use_cuda:
|
96 |
+
alpha = alpha.cuda()
|
97 |
+
|
98 |
+
interpolates = alpha * real_data + (1 - alpha) * fake_data
|
99 |
+
interpolates = interpolates.requires_grad_().clone()
|
100 |
+
|
101 |
+
disc_interpolates = netD(interpolates)
|
102 |
+
grad_outputs = torch.ones(disc_interpolates.size())
|
103 |
+
|
104 |
+
if self.use_cuda:
|
105 |
+
grad_outputs = grad_outputs.cuda()
|
106 |
+
|
107 |
+
gradients = autograd.grad(outputs=disc_interpolates, inputs=interpolates,
|
108 |
+
grad_outputs=grad_outputs, create_graph=True,
|
109 |
+
retain_graph=True, only_inputs=True)[0]
|
110 |
+
|
111 |
+
gradients = gradients.view(batch_size, -1)
|
112 |
+
gradient_penalty = ((gradients.norm(2, dim=1) - 1) ** 2).mean()
|
113 |
+
|
114 |
+
return gradient_penalty
|
115 |
+
|
116 |
+
def inference(self, x, masks):
|
117 |
+
self.eval()
|
118 |
+
x1, x2, offset_flow = self.netG(x, masks)
|
119 |
+
# x1_inpaint = x1 * masks + x * (1. - masks)
|
120 |
+
x2_inpaint = x2 * masks + x * (1. - masks)
|
121 |
+
|
122 |
+
return x2_inpaint, offset_flow
|
123 |
+
|
124 |
+
def save_model(self, checkpoint_dir, iteration):
|
125 |
+
# Save generators, discriminators, and optimizers
|
126 |
+
gen_name = os.path.join(checkpoint_dir, 'gen_%08d.pt' % iteration)
|
127 |
+
dis_name = os.path.join(checkpoint_dir, 'dis_%08d.pt' % iteration)
|
128 |
+
opt_name = os.path.join(checkpoint_dir, 'optimizer.pt')
|
129 |
+
torch.save(self.netG.state_dict(), gen_name)
|
130 |
+
torch.save({'localD': self.localD.state_dict(),
|
131 |
+
'globalD': self.globalD.state_dict()}, dis_name)
|
132 |
+
torch.save({'gen': self.optimizer_g.state_dict(),
|
133 |
+
'dis': self.optimizer_d.state_dict()}, opt_name)
|
134 |
+
|
135 |
+
def resume(self, checkpoint_dir, iteration=0, test=False):
|
136 |
+
# Load generators
|
137 |
+
last_model_name = get_model_list(checkpoint_dir, "gen", iteration=iteration)
|
138 |
+
self.netG.load_state_dict(torch.load(last_model_name))
|
139 |
+
iteration = int(last_model_name[-11:-3])
|
140 |
+
|
141 |
+
if not test:
|
142 |
+
# Load discriminators
|
143 |
+
last_model_name = get_model_list(checkpoint_dir, "dis", iteration=iteration)
|
144 |
+
state_dict = torch.load(last_model_name)
|
145 |
+
self.localD.load_state_dict(state_dict['localD'])
|
146 |
+
self.globalD.load_state_dict(state_dict['globalD'])
|
147 |
+
# Load optimizers
|
148 |
+
state_dict = torch.load(os.path.join(checkpoint_dir, 'optimizer.pt'))
|
149 |
+
self.optimizer_d.load_state_dict(state_dict['dis'])
|
150 |
+
self.optimizer_g.load_state_dict(state_dict['gen'])
|
151 |
+
|
152 |
+
print("Resume from {} at iteration {}".format(checkpoint_dir, iteration))
|
153 |
+
logger.info("Resume from {} at iteration {}".format(checkpoint_dir, iteration))
|
154 |
+
|
155 |
+
return iteration
|