-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_1.py
144 lines (132 loc) · 6.46 KB
/
train_1.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
import os
import cv2
import sys
sys.path.append(r'/home/jason/RIFE')
import math
import time
import torch
import torch.distributed as dist
import numpy as np
import random
import argparse
from model.RIFE import Model
from dataset import *
from torch.utils.data import DataLoader, Dataset
from torch.utils.tensorboard import SummaryWriter
from torch.utils.data.distributed import DistributedSampler
device = torch.device("cuda")
log_path = '/home/jason/RIFE/ECCV2022-RIFE/train_log'
def get_learning_rate(step):
if step < 2000:
mul = step / 2000.
return 3e-4 * mul
else:
mul = np.cos((step - 2000) / (args.epoch * args.step_per_epoch - 2000.) * math.pi) * 0.5 + 0.5
return (3e-4 - 3e-6) * mul + 3e-6
def flow2rgb(flow_map_np):
h, w, _ = flow_map_np.shape
rgb_map = np.ones((h, w, 3)).astype(np.float32)
normalized_flow_map = flow_map_np / (np.abs(flow_map_np).max())
rgb_map[:, :, 0] += normalized_flow_map[:, :, 0]
rgb_map[:, :, 1] -= 0.5 * (normalized_flow_map[:, :, 0] + normalized_flow_map[:, :, 1])
rgb_map[:, :, 2] += normalized_flow_map[:, :, 1]
return rgb_map.clip(0, 1)
def train(model):
writer = SummaryWriter('train')
writer_val = SummaryWriter('validate')
step=0
nr_eval = 0
dataset = VimeoDataset('train')
train_data = DataLoader(dataset,batch_size=args.batch_size,pin_memory=True,drop_last=True)
args.step_per_epoch = train_data.__len__()
dataset_val = VimeoDataset('validation')
val_data = DataLoader(dataset_val,batch_size=16,pin_memory=True)
print('training')
time_stamp = time.time()
for epoch in range(args.epoch):
for i, data in enumerate(train_data):
data_time_interval = time.time() - time_stamp
time_stamp = time.time()
data_gpu,timestep = data
data_gpu = data_gpu.to(device,non_blocking= True)/255.
timestep = timestep.to(device,non_blocking= True)
imgs = data_gpu[:,:6] #取第二个维度的前6个元素
gt = data_gpu[:,6:9] #取第二个维度的
learning_rate = get_learning_rate(step)
pred,info = model.update(imgs,gt,learning_rate,training=True)
train_time_interval = time.time() - time_stamp
time_stamp = time.time()
if step % 200==1:
writer.add_scalar('learning_rate', learning_rate, step)
writer.add_scalar('loss/l1', info['loss_l1'], step)
writer.add_scalar('loss/tea', info['loss_tea'], step)
writer.add_scalar('loss/distill', info['loss_distill'], step)
if step % 1000 ==1:
gt = (gt.permute(0,2,3,1).detach().cpu().numpy()*255).astype('uint8')
#.detach()取出中间结果tensor，并不影响梯度的反向传播或计算图 .astype()表示强行转换为uint8类型
mask = (torch.cat((info['mask'], info['mask_tea']), 3).permute(0, 2, 3,1).detach().cpu().numpy() * 255).astype('uint8')
pred = (pred.permute(0, 2, 3, 1).detach().cpu().numpy() * 255).astype('uint8')
merged_img = (info['merged_tea'].permute(0, 2, 3, 1).detach().cpu().numpy() * 255).astype('uint8')
flow0 = info['flow'].permute(0, 2, 3, 1).detach().cpu().numpy()
flow1 = info['flow_tea'].permute(0, 2, 3, 1).detach().cpu().numpy()
for i in range(5):
imgs = np.concatenate((merged_img[i], pred[i], gt[i]), 1)[:, :, ::-1]
writer.add_image(str(i) + '/img', imgs, step, dataformats='HWC')
writer.add_image(str(i) + '/flow', np.concatenate((flow2rgb(flow0[i]), flow2rgb(flow1[i])), 1), step, dataformats='HWC')
writer.add_image(str(i) + '/mask', mask[i], step, dataformats='HWC')
writer.flush()
print('epoch:{} {}/{} time:{:.2f} loss_l1:{}'.format(epoch,i,args.step_per_epoch,data_time_interval,train_time_interval,info['loss_l1']))
step += 1
nr_eval +=1
model.save_model(log_path)
def evaluate(model,val_data,nr_eval,writer_val):
loss_l1_list = []
loss_distill_list = []
loss_tea_list = []
psnr_list = []
psnr_list_teacher = []
time_stamp = time.time()
for i,data in enumerate(val_data):
data_gpu,time_step = data
data_gpu = data_gpu.to(device,non_blocking=True)/255.
imgs = data_gpu[:,:6]
gt = data_gpu[:,6:9]
with torch.no_grad():
pred,info = model.update(imgs,gt,training=False)
merged_img = info['merged_tea']
loss_l1_list.append(info['loss_l1'].cpu().numpy())
loss_tea_list.append(info['loss_tea'].cpu().numpy())
loss_distill_list.append(info['loss_distill'].cpu().numpy())
for j in range(gt.shape[0]):
psnr = -10 * math.log10(torch.mean((gt[j] - pred[j]) * (gt[j] - pred[j])).cpu().data)
psnr_list.append(psnr)
psnr = -10 * math.log10(torch.mean((merged_img[j] - gt[j]) * (merged_img[j] - gt[j])).cpu().data)
psnr_list_teacher.append(psnr)
gt = (gt.permute(0, 2, 3, 1).cpu().numpy() * 255).astype('uint8')
pred = (pred.permute(0, 2, 3, 1).cpu().numpy() * 255).astype('uint8')
merged_img = (merged_img.permute(0, 2, 3, 1).cpu().numpy() * 255).astype('uint8')
flow0 = info['flow'].permute(0, 2, 3, 1).cpu().numpy()
flow1 = info['flow_tea'].permute(0, 2, 3, 1).cpu().numpy()
if i == 0:
for j in range(10):
imgs = np.concatenate((merged_img[j], pred[j], gt[j]), 1)[:, :, ::-1]
writer_val.add_image(str(j) + '/img', imgs.copy(), nr_eval, dataformats='HWC')
writer_val.add_image(str(j) + '/flow', flow2rgb(flow0[j][:, :, ::-1]), nr_eval, dataformats='HWC')
eval_time_interval = time.time() - time_stamp
writer_val.add_scalar('psnr', np.array(psnr_list).mean(), nr_eval)
writer_val.add_scalar('psnr_teacher', np.array(psnr_list_teacher).mean(), nr_eval)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--epoch',default =200,type=int)
parser.add_argument('--batch_size',default=8,type=int,help='minibatch size')
parser.add_argument('--local_rank',default=0,type= int,help='local rank')
torch.cuda.set_device(0)
args = parser.parse_args()
seed =1234
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.benchmark =True
model = Model()
train(model)