-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtrain_velocity_network.py
185 lines (153 loc) · 7.41 KB
/
train_velocity_network.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
import argparse
import torch
from torch.utils.data import DataLoader
import models.motion_gan as gan_models
import models.motion_vae as vae_models
import utils.paramUtil as paramUtil
from trainer.vae_trainer import *
from models.networks import *
from dataProcessing import dataset
from utils.plot_script import plot_loss, print_current_loss
from options.train_vae_options import TrainOptions
import os
def get_cate_one_hot(categories):
classes_to_generate = np.array(categories).reshape((-1,))
# dim (num_samples, dim_category)
one_hot = np.zeros((categories.shape[0], opt.dim_category), dtype=np.float32)
one_hot[np.arange(categories.shape[0]), classes_to_generate] = 1
# dim (num_samples, dim_category)
one_hot_motion = torch.from_numpy(one_hot).to(device).requires_grad_(False)
return one_hot_motion, classes_to_generate
def save_network(network, save_path, save_name):
if not os.path.exists(save_path):
os.makedirs(save_path)
save_path = os.path.join(save_path, save_name)
torch.save(network.state_dict(), save_path)
def load_network(network, save_path, save_name):
save_path = os.path.join(save_path, save_name)
params = torch.load(save_path)
network.load_state_dict(params)
if __name__ == "__main__":
parser = TrainOptions()
opt = parser.parse()
device = torch.device("cuda:" + str(opt.gpu_id) if torch.cuda.is_available() else "cpu")
opt.save_root = os.path.join(opt.checkpoints_dir, opt.dataset_type, opt.name)
opt.model_path = os.path.join(opt.save_root, 'model')
if not os.path.exists(opt.model_path):
os.makedirs(opt.model_path)
dataset_path = ""
joints_num = 0
input_size = 72
data = None
if opt.dataset_type == "humanact13":
dataset_path = "./dataset/humanact13"
input_size = 72
joints_num = 24
data = dataset.MotionFolderDatasetHumanAct13(dataset_path, opt, False, True)
elif opt.dataset_type == "mocap":
dataset_path = "./dataset/mocap/mocap_3djoints/"
clip_path = './dataset/mocap/pose_clip.csv'
input_size = 60
joints_num = 20
raw_offsets = paramUtil.mocap_raw_offsets
kinematic_chain = paramUtil.mocap_kinematic_chain
data = dataset.MotionFolderDatasetMocap(clip_path, dataset_path, opt)
elif opt.dataset_type == "ntu_rgbd_vibe":
file_prefix = "./dataset"
motion_desc_file = "ntu_vibe_list.txt"
joints_num = 18
input_size = 54
labels = paramUtil.ntu_action_labels
raw_offsets = paramUtil.vibe_raw_offsets
kinematic_chain = paramUtil.vibe_kinematic_chain
data = dataset.MotionFolderDatasetNtuVIBE(file_prefix, motion_desc_file, labels, opt, joints_num=joints_num,
offset=True, extract_joints=paramUtil.kinect_vibe_extract_joints)
else:
raise NotImplementedError('This dataset is unregonized!!!')
opt.dim_category = len(data.labels)
opt.input_size = input_size * 2 + opt.dim_category
opt.output_size = 3
if opt.use_vel_S:
motion_dataset = dataset.PairFrameDataset(data)
motion_loader = DataLoader(motion_dataset, batch_size=opt.batch_size, drop_last=True, num_workers=2,
shuffle=True)
model = VelocityNetwork_Sim(opt.input_size, opt.output_size, opt.hidden_size)
else:
motion_dataset = dataset.MotionDataset(data, opt)
motion_loader = DataLoader(motion_dataset, batch_size=opt.batch_size, drop_last=True, num_workers=2,
shuffle=True)
model = VelocityNetwork(opt.input_size, opt.output_size, opt.hidden_size, 1, opt.batch_size, device)
mse = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=0.0002, betas=(0.9, 0.999), weight_decay=0.00001)
if opt.is_continue:
load_network(model, opt.model_path, 'latest.tar')
model.to(device)
model.train()
def __init_log__():
log = OrderedDict()
log['total_loss'] = []
return log
total_steps = 1
loss_log = __init_log__()
start_time = time.time()
niter_per_epo = len(motion_loader)
Tensor = torch.cuda.FloatTensor if torch.cuda.is_available() else torch.Tensor
pc_model = sum(param.numel() for param in model.parameters())
print(model)
print("Total parameters of prior net: {}".format(pc_model))
print("# training dataset size {0}".format(len(motion_dataset)))
print("# number of iterations per epoch {0}".format(niter_per_epo))
for epoch in range(opt.epoch_size):
for niter, batch in enumerate(motion_loader):
optimizer.zero_grad()
if opt.use_vel_S:
# data1 (batch_size, num_joints * 3)
(data1, data2, labels) = batch
data1 = Tensor(data1.size()).copy_(data1)
data2 = Tensor(data2.size()).copy_(data2)
# amplify the error by 10 times
ground = (data2[..., :3] - data1[..., :3]) * 10
cate_ones, _ = get_cate_one_hot(labels)
data1 = data1 - data1[..., :3].repeat(1, joints_num)
data2 = data2 - data2[..., :3].repeat(1, joints_num)
inputs = torch.cat((cate_ones, data1, data2), dim=1)
inputs = Tensor(inputs.size()).copy_(inputs).detach()
ground = Tensor(ground.size()).copy_(ground).detach()
output = model(inputs)
total_loss = mse(output, ground)
loss_log['total_loss'].append(total_loss.item())
else:
data, labels = batch
# data (batch_size, motion_len, num_joints*3)
data = Tensor(data.size()).copy_(data)
# ground (batch_size, motion_len - 1, 3)
ground = (data[:, 1:, :3] - data[:, :-1, :3]) * 10
data = data - data[..., :3].repeat(1, 1, joints_num)
data1 = data[:, 1:, :]
data2 = data[:, :-1, :]
# cate_ones (batch_size, cate_dim)
cate_ones, _ = get_cate_one_hot(labels)
# cate_ones (batch_size, motion_len-1, cate_dim)
cate_ones = cate_ones.unsqueeze(1).repeat(1, data1.shape[1], 1)
inputs = torch.cat((cate_ones, data1, data2), dim=-1)
inputs = Tensor(inputs.size()).copy_(inputs).detach()
ground = Tensor(ground.size()).copy_(ground).detach()
model.init_hidden()
total_loss = 0
for i in range(inputs.shape[1]):
h_in = inputs[:, i, :]
h_out = model(h_in)
total_loss += mse(h_out, ground[:, i, :])
loss_log['total_loss'].append(total_loss.item() / inputs.shape[1])
total_loss.backward()
optimizer.step()
if total_steps % opt.print_every == 0:
mean_loss = OrderedDict()
for k, v in loss_log.items():
mean_loss[k] = sum(loss_log[k][-1 * opt.print_every:]) / opt.print_every
print_current_loss(start_time, total_steps, opt.epoch_size * niter_per_epo, mean_loss, epoch, niter)
if total_steps % opt.save_latest == 0:
save_network(model, opt.model_path, 'latest.tar')
total_steps += 1
save_network(model, opt.model_path, 'latest.tar')
plot_loss(loss_log, os.path.join(opt.save_root, "loss_curve.png"), intervals=opt.plot_every)