forked from PantoMatrix/PantoMatrix
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathae_trainer.py
164 lines (152 loc) · 7.89 KB
/
ae_trainer.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
# Copyright (c) HuaWei, Inc. and its affiliates.
# Train script for audio2pose
import os
import time
import csv
import sys
import warnings
import random
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
import numpy as np
import time
import pprint
from loguru import logger
from utils import config, logger_tools, other_tools
from dataloaders import data_tools
from dataloaders.build_vocab import Vocab
from optimizers.optim_factory import create_optimizer
from optimizers.scheduler_factory import create_scheduler
from optimizers.loss_factory import get_loss_func
from train import BaseTrainer
class CustomTrainer(BaseTrainer):
def __init__(self, args):
super().__init__(args)
self.g_name = args.g_name
self.pose_length = args.pose_length
self.loss_meters = {
'rec_val': other_tools.AverageMeter('rec_val'),
'vel_val': other_tools.AverageMeter('vel_val'),
'kl_val': other_tools.AverageMeter('kl_val'),
'all': other_tools.AverageMeter('all'),
'rec_l1': other_tools.AverageMeter('rec_l1'),
'vel_l1': other_tools.AverageMeter('vel_l1'),
'kl_loss': other_tools.AverageMeter('kl_loss'),
#'acceleration_loss': other_tools.AverageMeter('acceleration_loss'),
}
self.best_epochs = {
'rec_val': [np.inf, 0],
'vel_val': [np.inf, 0],
'kl_val': [np.inf, 0],
}
self.rec_loss = torch.nn.L1Loss(reduction='none')
self.vel_loss = torch.nn.MSELoss(reduction='none')
self.variational_encoding = args.variational_encoding
self.rec_weight = args.rec_weight
self.vel_weight = args.vel_weight
def train(self, epoch):
self.model.train()
its_len = len(self.train_loader)
t_start = time.time()
for its, dict_data in enumerate(self.train_loader):
tar_pose = dict_data["pose"]
t_data = time.time() - t_start
tar_pose = tar_pose.cuda()
t_data = time.time() - t_start
self.opt.zero_grad()
poses_feat, pose_mu, pose_logvar, recon_data = \
self.model(None, tar_pose, variational_encoding=self.variational_encoding)
recon_loss = self.rec_loss(recon_data, tar_pose) # 128*34*123
recon_loss = torch.mean(recon_loss, dim=(1, 2)) # 128
self.loss_meters['rec_l1'].update(torch.sum(recon_loss).item()*self.rec_weight)
recon_loss = torch.sum(recon_loss*self.rec_weight)
# rec vel loss
if self.vel_weight > 0: # use pose diff
target_diff = tar_pose[:, 1:] - tar_pose[:, :-1]
recon_diff = recon_data[:, 1:] - recon_data[:, :-1]
vel_rec_loss = torch.mean(self.vel_loss(recon_diff, target_diff), dim=(1, 2))
self.loss_meters['vel_l1'].update(torch.sum(vel_rec_loss).item()*self.vel_weight)
recon_loss += (torch.sum(vel_rec_loss)*self.vel_weight)
# KLD
if self.variational_encoding:
KLD = -0.5 * torch.sum(1 + pose_logvar - pose_mu.pow(2) - pose_logvar.exp())
if epoch < 10:
KLD_weight = 0
else:
KLD_weight = min(1.0, (epoch - 10) * 0.05) * 0.01
loss = recon_loss + KLD_weight * KLD
self.loss_meters['kl_loss'].update(KLD_weight * KLD.item())
else:
loss = recon_loss
self.loss_meters['all'].update(loss.item())
if self.grad_norm != 0 and "LSTM" in self.g_name: torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.grad_norm)
# logger.warning(total_norm)
loss.backward()
self.opt.step()
t_train = time.time() - t_start - t_data
t_start = time.time()
mem_cost = torch.cuda.memory_cached() / 1E9
lr_g = self.opt.param_groups[0]['lr']
# --------------------------- recording ---------------------------------- #
if its % self.log_period == 0:
self.recording(epoch, its, its_len, self.loss_meters, lr_g, 0, t_data, t_train, mem_cost)
#if its == 1:break
self.opt_s.step(epoch)
def val(self, epoch):
self.model.eval()
with torch.no_grad():
its_len = len(self.val_loader)
for its, dict_data in enumerate(self.val_loader):
tar_pose = dict_data["pose"]
tar_pose = tar_pose.cuda()
poses_feat, pose_mu, pose_logvar, recon_data = \
self.model(None, tar_pose, variational_encoding=self.variational_encoding)
if self.vel_weight > 0: # use pose diff
target_diff = tar_pose[:, 1:] - tar_pose[:, :-1]
recon_diff = recon_data[:, 1:] - recon_data[:, :-1]
vel_rec_loss = torch.mean(self.vel_loss(recon_diff, target_diff), dim=(0, 1, 2))
self.loss_meters['vel_val'].update(vel_rec_loss.item())
#print(recon_data.shape, tar_pose.shape)
recon_loss = F.l1_loss(recon_data, tar_pose, reduction='none')
recon_loss = torch.mean(recon_loss, dim=(0, 1, 2))
self.loss_meters['rec_val'].update(recon_loss.item())
self.val_recording(epoch, self.loss_meters)
def test(self, epoch):
results_save_path = self.checkpoint_path + f"/{epoch}/"
start_time = time.time()
total_length = 0
test_seq_list = os.listdir(self.test_demo)
test_seq_list.sort()
self.model.eval()
with torch.no_grad():
if not os.path.exists(results_save_path):
os.makedirs(results_save_path)
for its, dict_data in enumerate(self.test_loader):
tar_pose = dict_data["pose"]
tar_pose = tar_pose.cuda() # no mean
if "LSTM" in self.g_name or "multi_length" in self.notes:
poses_feat, pose_mu, pose_logvar, recon_data = \
self.model(**dict(pre_poses=None, poses=tar_pose, variational_encoding=self.variational_encoding))
out_final = (recon_data.cpu().numpy().reshape(-1, self.pose_dims) * self.std_pose) + self.mean_pose
else:
for i in range(tar_pose.shape[1]//(self.pose_length)):
tar_pose_new = tar_pose[:,i*(self.pose_length):i*(self.pose_length)+self.pose_length,:]
poses_feat, pose_mu, pose_logvar, recon_data = \
self.model(**dict(pre_poses=None, poses=tar_pose_new, variational_encoding=self.variational_encoding,))
out_sub = (recon_data.cpu().numpy().reshape(-1, self.pose_dims) * self.std_pose) + self.mean_pose
if i != 0:
out_final = np.concatenate((out_final,out_sub), 0)
else:
out_final = out_sub
total_length += out_final.shape[0]
with open(f"{results_save_path}result_raw_{test_seq_list[its]}", 'w+') as f_real:
for line_id in range(out_final.shape[0]): #,args.pre_frames, args.pose_length
line_data = np.array2string(out_final[line_id], max_line_width=np.inf, precision=6, suppress_small=False, separator=' ')
f_real.write(line_data[1:-2]+'\n')
data_tools.result2target_vis(self.pose_version, results_save_path, results_save_path, self.test_demo, False)
end_time = time.time() - start_time
logger.info(f"total inference time: {int(end_time)} s for {int(total_length/self.pose_fps)} s motion")