-
Notifications
You must be signed in to change notification settings - Fork 7
/
train_ddp.py
257 lines (211 loc) · 9.84 KB
/
train_ddp.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
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
import os
from myutils.dataloader import UWS_Dataset_Retinex, UWS_Dataset_Retinex_test
from torchvision.utils import save_image,make_grid
import torchvision.transforms as transforms
from torch.utils.data import DataLoader
from torch.autograd import Variable
from PIL import Image
import torch
from pytorch_fid_test.src.pytorch_fid.fid_score import *
from time import time
from tqdm import *
from option import opt
import time
from model.FSU2 import *
from loss.perceptual import PerceptualLoss2,PerceptualLoss
from loss.DCLoss import DCLoss_Two
from tensorboardX import SummaryWriter
from option import *
from metrics import *
from myutils.utils import *
import shutil
import glob
from ptflops import get_model_complexity_info
#DDP
from apex import amp
import torch.distributed as dist
from apex.parallel import DistributedDataParallel as DDP
from apex.fp16_utils import *
from loss.retinex_loss import *
warnings.filterwarnings("ignore")
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1"
# python3 -m torch.distributed.launch --master_port 42563 --nproc_per_node 2 train_ddp.py --resume=True
#ddp
torch.cuda.set_device(opt.local_rank)
dist.init_process_group(backend='nccl')
opt.shutil = True
if opt.local_rank == 0:
if opt.shutil == True:
#cover the previous tensorboardX file
file_path_runs = glob.glob(os.path.join(opt.tensorboardX_path,f'*{opt.model_name}'))
for i in file_path_runs:
shutil.rmtree(i)
writer = SummaryWriter(os.path.join(opt.tensorboardX_path,opt.model_name))
def init_seeds(seed=0, cuda_deterministic=False):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if cuda_deterministic: # slower, more reproducible
cudnn.deterministic = True
cudnn.benchmark = False
else: # faster, less reproducible
cudnn.deterministic = False
cudnn.benchmark = True
rank = torch.distributed.get_rank()
init_seeds(41 + rank)
def train(netG_A2B,loader_train,loader_test,test_data,test_depth):
start_epoch=0
min_fid = 1000
vgg_loss =PerceptualLoss().cuda()
dc_loss = DCLoss_Two(15).cuda()
lrc = torch.nn.L1Loss()
retinex_loss = Retinex_loss1().cuda()
flops_t, params_t = get_model_complexity_info(netG_A2B, (7, opt.crop_size, opt.crop_size), as_strings=True, print_per_layer_stat=True)
print(f"net flops:{flops_t} parameters:{params_t}")
optimizer_G1 = torch.optim.AdamW(netG_A2B.parameters(), lr=opt.lr,betas=[0.9,0.999],weight_decay=0.000001)
netG_A2B,optimizer_G1= amp.initialize(netG_A2B,optimizer_G1,opt_level='O1')
netG_A2B = DDP(netG_A2B,delay_allreduce=True)
lr_scheduler_G = torch.optim.lr_scheduler.LambdaLR(optimizer_G1,
lr_lambda=LambdaLR(opt.n_epochs, opt.epoch, opt.decay_epoch).step)
if opt.resume and os.path.exists(opt.model_dir+'utils_'+opt.model_name+'.pk'):
ckp=torch.load(opt.model_dir+'utils_'+opt.model_name+'.pk')
start_epoch=ckp['epoch']
min_fid = ckp['min_fid']
step = ckp['step']
if opt.local_rank == 0:
print(f'resume from {opt.model_dir}')
print(f'*******min FID{min_fid:.4f}*******')
G1 = torch.load(opt.model_dir+f'net_render-epoch{start_epoch:02d}-fid{min_fid:.2f}.pk')
from collections import OrderedDict
new_state_dict = OrderedDict()
for k, v in G1['G1'].items():
name = k
new_state_dict[name] = v
netG_A2B.load_state_dict(new_state_dict)
new_state_dict = OrderedDict()
optimizer_G1.load_state_dict(G1['optimizer_weightG1'])
else :
step =0
if opt.local_rank == 0:
netG_A2B.apply(weights_init_normal)
start_epoch = 0
print('train from scratch *** ')
#**************train****************#
step = step
for epoch in range(start_epoch+1, opt.n_epochs+1):
netG_A2B.train()
loop = tqdm(enumerate(loader_train),total=len(loader_train))
loader_train.sampler.set_epoch(epoch)
for i, batch in loop:
step += 1
(data,target,depth_map,A_map) = batch
data=data.cuda();target=target.cuda();depth_map=depth_map.cuda();A_map=A_map.cuda()
x = torch.cat([target,depth_map,A_map],1)
I_0 = netG_A2B(x)
reti_loss_l1 = retinex_loss(data,I_0)
reti_loss_l1 = 0.5*reti_loss_l1
vg_loss = 0.5*(vgg_loss(data,I_0))
t_loss = 0.5*(dc_loss(data[:,1:],I_0[:,1:]))
lrcloss = 0.6*(lrc(data,I_0))
loss =t_loss+vg_loss+lrcloss+reti_loss_l1
optimizer_G1.zero_grad()
with amp.scale_loss(loss,optimizer_G1)as scale_losss:
scale_losss.backward()
optimizer_G1.step()
if opt.local_rank == 0:
writer.add_scalar('epoch',epoch,step)
writer.add_scalar('loss',loss.item(),step)
writer.add_scalar('dc_loss',t_loss.item(),step)
writer.add_scalar('lrcloss',lrcloss.item(),step)
writer.add_scalar('reti_loss_pe',reti_loss_l1.item(),step)
writer.add_scalar('vgloss',vg_loss.item(),step)
if i % 2 == 0 :
print(f'\n|loss : {loss.item():.5f} |epoch :{epoch}/{opt.n_epochs} |step :{i*2}/{len(loader_train)*2}')
torch.distributed.barrier()
# test
if opt.local_rank == 0:
if (epoch) % opt.test_epoch ==0 :
netG_A2B.eval()
with torch.no_grad():
len1,fid=test(netG_A2B,loader_test,test_data,test_depth,epoch)
print(f'\nepoch :{epoch} |fid:{fid:.4f} |len:{len1:.4f}')
writer.add_scalar('fid',fid,step)
# save every epoch model
file_path = glob.glob(opt.model_dir+r'net*')
if file_path :
os.remove(file_path[0])
torch.save({
'G1':netG_A2B.state_dict(),
'optimizer_weightG1': optimizer_G1.state_dict(),
},opt.model_dir+f'net_render-epoch{epoch:02d}-fid{fid:.2f}.pk')
# save the min_fid model
if (fid < min_fid):
min_fid = fid
torch.save({
'min_fid':min_fid,
'epoch':epoch,
'step':step
},opt.model_dir+'utils_'+opt.model_name+'.pk')
file_path = glob.glob(opt.model_dir+r'sota*')
if file_path :
os.remove(file_path[0])
torch.save({
'G1':netG_A2B.state_dict(),
'optimizer_weightG1': optimizer_G1.state_dict(),
},opt.model_dir+f'sota_render-epoch{epoch:02d}-fid{min_fid:.2f}.pk')
print(f'\n model saved at step :{epoch}| min_fid:{min_fid:.4f}')
# save fixed epoch model
if (epoch % 40)== 0 and (epoch>=80):
torch.save({
'G1':netG_A2B.state_dict(),
'optimizer_weightG1': optimizer_G1.state_dict(),
},opt.model_dir+f'epoch_render-epoch{epoch:02d}-fid{fid:.2f}.pk')
torch.distributed.barrier()
# Update learning rates
lr_scheduler_G.step()
writer.close()
def test(G1,loader_test,test_data,test_depth,epoch):
G1.eval()
torch.cuda.empty_cache()
loop = tqdm(enumerate(loader_test),total=len(loader_test))
for i ,(A_map,data) in loop:
A_map=A_map.cuda()
#print(A_map.shape)
gt = test_data[i]
img_name = gt.split('/')[-1].split('.')[0]
gt = Image.open(gt).convert("RGB")
gt = transforms.functional.to_tensor(gt)
gt = transforms.Resize([256,256])(gt)
gt = gt.unsqueeze(0)
gt= gt.cuda()
#print('gt',gt)
depth_map = test_depth[i]
depth_map = Image.open(depth_map).convert("L")
depth_map = transforms.functional.to_tensor(depth_map)
depth_map = transforms.Resize([256,256])(depth_map)
#print(depth_map)
depth_map = depth_map.unsqueeze(0)
depth_map= depth_map.cuda()
x = torch.cat([gt,depth_map,A_map],1)
g1_output = G1(x)
save_image(g1_output,os.path.join(opt.train_save_path,'%s.png'%(img_name)),normalize=False)
#calculate fid
fid = calculate_fid_given_paths([opt.train_save_path,opt.fid_gt_path],50,'cuda:0',2048,1)
#print(fid)
return len(loader_test),fid
if __name__ == "__main__":
os.makedirs(opt.train_save_path,exist_ok=True)
train_sampler = torch.utils.data.distributed.DistributedSampler(dataset=UWS_Dataset_Retinex(opt.underwater_path,train=True,size=opt.crop_size,dcp=True),shuffle=True,num_replicas=2,rank=opt.local_rank)
UW_train_loader = DataLoader(dataset=UWS_Dataset_Retinex(opt.underwater_path,train=True,size=opt.crop_size,dcp=False),batch_size=opt.bs,num_workers=0,sampler =train_sampler)
UW_test_loader = DataLoader(dataset=UWS_Dataset_Retinex_test(opt.underwater_path,train=False,size=opt.crop_size_test,dcp=False),batch_size=1,shuffle=False,num_workers=0)
#clean image
test_lir=os.listdir(opt.clean_img_path)
test_lir.sort()
test_data =[os.path.join(opt.clean_img_path,test_img) for test_img in test_lir]
#clean image depthmap
test_lir_depth=os.listdir(opt.depth_img_path)
test_lir_depth.sort()
test_depth =[os.path.join(opt.depth_img_path,test_img) for test_img in test_lir_depth]
#MHBnet
model = Generator().cuda()
train(model,UW_train_loader,UW_test_loader,test_data,test_depth)