-
Notifications
You must be signed in to change notification settings - Fork 3
/
train_nf.py
278 lines (243 loc) · 10.7 KB
/
train_nf.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
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
import os
import sys
import pickle
import argparse
import time
from torch import optim
from torch.utils.tensorboard import SummaryWriter
import matplotlib.pyplot as plt
sys.path.append(os.getcwd())
from utils import *
from motion_pred.utils.config import Config
from motion_pred.utils.dataset_h36m import DatasetH36M
from motion_pred.utils.dataset_humaneva import DatasetHumanEva
from models.motion_pred_ours import *
from utils import util
def loss_function(prior_lkh, log_det_jacobian):
loss_p = -prior_lkh.mean()
loss_jac = - log_det_jacobian.mean()
loss_r = loss_p + loss_jac
return loss_r, np.array([loss_r.item(), loss_p.item(), loss_jac.item()])
def train(epoch):
model.train()
t_s = time.time()
train_losses = 0
train_grad = 0
total_num_sample = 0
loss_names = ['LKH', 'log_p(z)', 'log_det']
generator = dataset.sampling_generator(num_samples=cfg.num_vae_data_sample, batch_size=cfg.batch_size)
prior = torch.distributions.Normal(torch.tensor(0, dtype=dtype, device=device),
torch.tensor(1, dtype=dtype, device=device))
for traj_np in generator:
with torch.no_grad():
traj_np = traj_np[:, 0]
traj_np[:, 0] = 0
traj_np = util.absolute2relative(traj_np, parents=dataset.skeleton.parents())
traj = tensor(traj_np, device=device, dtype=dtype) # .permute(0, 2, 1).contiguous()
bs, nj, _ = traj.shape
x = traj.reshape([bs, -1])
z, log_det_jacobian = model(x)
prior_likelihood = prior.log_prob(z).sum(dim=1)
loss, losses = loss_function(prior_likelihood, log_det_jacobian)
optimizer.zero_grad()
loss.backward()
# grad_norm = torch.nn.utils.clip_grad_norm_(list(model.parameters()), max_norm=10000)
grad_norm = 0
train_grad += grad_norm
optimizer.step()
train_losses += losses
total_num_sample += 1
del loss, z, prior_likelihood, log_det_jacobian
scheduler.step()
# dt = time.time() - t_s
train_losses /= total_num_sample
lr = optimizer.param_groups[0]['lr']
losses_str = ' '.join(['{}: {:.4f}'.format(x, y) for x, y in zip(loss_names, train_losses)])
# average cost of log time 20s
tb_logger.add_scalar('train_grad', train_grad / total_num_sample, epoch)
for name, loss in zip(loss_names, train_losses):
tb_logger.add_scalars('vae_' + name, {'train': loss}, epoch)
logger.info('====> Epoch: {} Time: {:.2f} {} lr: {:.5f}'.format(epoch, time.time() - t_s, losses_str, lr))
def val(epoch):
model.eval()
t_s = time.time()
train_losses = 0
total_num_sample = 0
loss_names = ['LKH', 'log_p(z)', 'log_det']
generator = dataset.sampling_generator(num_samples=cfg.num_vae_data_sample // 2, batch_size=cfg.batch_size)
prior = torch.distributions.Normal(torch.tensor(0, dtype=dtype, device=device),
torch.tensor(1, dtype=dtype, device=device))
loginfos = None
with torch.no_grad():
xx = []
for traj_np in generator:
traj_np = traj_np[:, 0]
traj_np[:, 0] = 0
traj_np = util.absolute2relative(traj_np, parents=dataset.skeleton.parents())
traj = tensor(traj_np, device=device, dtype=dtype) # .permute(0, 2, 1).contiguous()
bs, nj, _ = traj.shape
x = traj.reshape([bs, -1])
z, log_det_jacobian = model(x)
prior_likelihood = prior.log_prob(z).sum(dim=1)
loginf = {}
loginf[f'z'] = z.cpu().data.numpy()
loss, losses = loss_function(prior_likelihood, log_det_jacobian)
train_losses += losses
total_num_sample += 1
del loss, z, prior_likelihood, log_det_jacobian
loginfos = combine_dict(loginf, loginfos)
# dt = time.time() - t_s
train_losses /= total_num_sample
lr = optimizer.param_groups[0]['lr']
losses_str = ' '.join(['{}: {:.4f}'.format(x, y) for x, y in zip(loss_names, train_losses)])
# average cost of log time 20s
for name, loss in zip(loss_names, train_losses):
tb_logger.add_scalars('vae_' + name, {'val': loss}, epoch)
logger.info('====> Epoch: {} Val Time: {:.2f} {} lr: {:.5f}'.format(epoch, time.time() - t_s, losses_str, lr))
t_s = time.time()
generator = dataset_test.sampling_generator(num_samples=cfg.num_vae_data_sample // 2, batch_size=cfg.batch_size)
loginfos_test = None
with torch.no_grad():
xx = []
for traj_np in generator:
traj_np = traj_np[:, 0]
traj_np[:, 0] = 0
traj_np = util.absolute2relative(traj_np, parents=dataset.skeleton.parents())
traj = tensor(traj_np, device=device, dtype=dtype) # .permute(0, 2, 1).contiguous()
bs, nj, _ = traj.shape
x = traj.reshape([bs, -1])
z, log_det_jacobian = model(x)
prior_likelihood = prior.log_prob(z).sum(dim=1)
loginf = {}
loginf[f'z'] = z.cpu().data.numpy()
loss, losses = loss_function(prior_likelihood, log_det_jacobian)
train_losses += losses
total_num_sample += 1
del loss, z, prior_likelihood, log_det_jacobian
loginfos_test = combine_dict(loginf, loginfos_test)
# dt = time.time() - t_s
train_losses /= total_num_sample
lr = optimizer.param_groups[0]['lr']
losses_str = ' '.join(['{}: {:.4f}'.format(x, y) for x, y in zip(loss_names, train_losses)])
# average cost of log time 20s
for name, loss in zip(loss_names, train_losses):
tb_logger.add_scalars('vae_' + name, {'test': loss}, epoch)
logger.info('====> Epoch: {} Test Time: {:.2f} {} lr: {:.5f}'.format(epoch, time.time() - t_s, losses_str, lr))
t_s = time.time()
zz = loginfos['z']
zz = zz.reshape([zz.shape[0], -1])
bs, data_dim = zz.shape
zz_test = loginfos_test['z']
zz_test = zz_test.reshape([zz.shape[0], -1])
# bs, data_dim = zz_test.shape
for ii in range(data_dim):
fig = plt.figure()
ax1 = plt.subplot(121)
_ = plt.hist(zz[:, ii].reshape(-1), bins=100, density=True, alpha=0.5, color='b')
x = torch.from_numpy(np.arange(-5, 5, 0.01)).float().to(device)
y = prior.cdf(x)
x = x[1:]
y = (y[1:] - y[:-1]) * 100
plt.plot(x.cpu().data, y.cpu().data)
ax1.set_title(f'z_val_{ii}')
ax2 = plt.subplot(122)
_ = plt.hist(zz_test[:, ii].reshape(-1), bins=100, density=True, alpha=0.5, color='b')
x = torch.from_numpy(np.arange(-5, 5, 0.01)).float().to(device)
y = prior.cdf(x)
x = x[1:]
y = (y[1:] - y[:-1]) * 100
plt.plot(x.cpu().data, y.cpu().data)
ax2.set_title(f'z_test_{ii}')
tb_logger.add_figure(f'z_{ii}', fig, epoch)
plt.clf()
plt.cla()
plt.close(fig)
# plot covariance matrix
fig = plt.figure()
ax1 = plt.subplot(121)
# zz = zz.reshape([bs, -1])
zz = zz - zz.mean(axis=0)
cov = 1 / bs * np.abs(np.matmul(zz.transpose([1, 0]), zz))
std = np.sqrt(np.diag(cov))[:, None] * np.sqrt(np.diag(cov))[None, :]
corr = cov / (std + 1e-10) - np.eye(cov.shape[0])
plt.imshow(corr)
plt.colorbar()
ax1.set_title('z_val_corr')
ax2 = plt.subplot(122)
# zz = zz.reshape([bs, -1])
zz_test = zz_test - zz_test.mean(axis=0)
cov = 1 / bs * np.abs(np.matmul(zz_test.transpose([1, 0]), zz_test))
std = np.sqrt(np.diag(cov))[:, None] * np.sqrt(np.diag(cov))[None, :]
corr = cov / (std + 1e-10) - np.eye(cov.shape[0])
plt.imshow(corr)
plt.colorbar()
ax2.set_title('z_test_corr')
tb_logger.add_figure('z_corr', plt.gcf(), epoch)
plt.close(fig)
print(f'>>>>log time {time.time() - t_s:.3f}')
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--cfg', default='h36m_nf')
parser.add_argument('--mode', default='train')
parser.add_argument('--test', action='store_true', default=False)
parser.add_argument('--iter', type=int, default=0)
parser.add_argument('--seed', type=int, default=0)
parser.add_argument('--gpu_index', type=int, default=1)
parser.add_argument('--n_pre', type=int, default=10)
parser.add_argument('--n_his', type=int, default=5)
parser.add_argument('--trial', type=int, default=1)
parser.add_argument('--num_coupling_layer', type=int, default=6)
parser.add_argument('--nz', type=int, default=10)
args = parser.parse_args()
"""setup"""
np.random.seed(args.seed)
torch.manual_seed(args.seed)
dtype = torch.float32
torch.set_default_dtype(dtype)
device = torch.device('cuda', index=args.gpu_index) if torch.cuda.is_available() else torch.device('cpu')
if torch.cuda.is_available():
torch.cuda.set_device(args.gpu_index)
cfg = Config(f'{args.cfg}', test=args.test)
tb_logger = SummaryWriter(cfg.tb_dir) if args.mode == 'train' else None
logger = create_logger(os.path.join(cfg.log_dir, 'log.txt'))
"""parameter"""
mode = args.mode
nz = cfg.nz
t_his = cfg.t_his
t_pred = cfg.t_pred
cfg.n_his = args.n_his
cfg.n_pre = args.n_pre
cfg.num_coupling_layer = args.num_coupling_layer
cfg.nz = args.nz
"""data"""
dataset_cls = DatasetH36M if cfg.dataset == 'h36m' else DatasetHumanEva
dataset = dataset_cls('train', t_his, t_pred, actions='all', use_vel=cfg.use_vel)
dataset_test = dataset_cls('test', t_his, t_pred, actions='all', use_vel=cfg.use_vel)
if cfg.normalize_data:
dataset.normalize_data()
dataset_test.normalize_data(dataset.mean, dataset.std)
"""model"""
model = get_model(cfg, dataset.traj_dim, args.cfg)
model.float()
optimizer = optim.Adam(model.parameters(), lr=cfg.vae_lr, weight_decay=1e-3)
scheduler = get_scheduler(optimizer, policy='lambda', nepoch_fix=cfg.num_vae_epoch_fix, nepoch=cfg.num_vae_epoch)
logger.info(">>> total params: {:.5f}M".format(sum(p.numel() for p in list(model.parameters())) / 1000000.0))
if args.iter > 0:
cp_path = cfg.vae_model_path % args.iter
print('loading model from checkpoint: %s' % cp_path)
model_cp = pickle.load(open(cp_path, "rb"))
model.load_state_dict(model_cp['model_dict'])
if mode == 'train':
model.to(device)
overall_iter = 0
for i in range(args.iter, cfg.num_vae_epoch):
model.train()
train(i)
model.eval()
val(i)
# test(i)
if cfg.save_model_interval > 0 and (i + 1) % cfg.save_model_interval == 0:
with to_cpu(model):
cp_path = cfg.vae_model_path % (i + 1)
model_cp = {'model_dict': model.state_dict(), 'meta': {'std': dataset.std, 'mean': dataset.mean}}
pickle.dump(model_cp, open(cp_path, 'wb'))