-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_VRAE.py
163 lines (119 loc) · 4.56 KB
/
train_VRAE.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
#%%
import time
import math
import sys
import argparse
import cPickle as pickle
import copy
import os
import six
import numpy as np
from chainer import cuda, Variable, FunctionSet, optimizers
import chainer.functions as F
from VRAE import VRAE, make_initial_state
import dataset
parser = argparse.ArgumentParser()
parser.add_argument('--data_path', type=str, default="dataset")
parser.add_argument('--output_dir', type=str, default="model")
parser.add_argument('--dataset', type=str, default="midi")
parser.add_argument('--init_from', type=str, default="")
parser.add_argument('--clip_grads', type=int, default=5)
parser.add_argument('--gpu', type=int, default=-1)
args = parser.parse_args()
if not os.path.exists(args.output_dir):
os.mkdir(args.output_dir)
if args.dataset == 'midi':
midi = dataset.load_midi_data('%s/midi/sample.mid' % args.data_path)
train_x = midi[:120].astype(np.float32)
n_x = train_x.shape[1]
n_hidden = [500]
n_z = 2
n_y = n_x
frames = train_x.shape[0]
n_batch = 6
seq_length = frames / n_batch
split_x = np.vsplit(train_x, n_batch)
n_epochs = 500
continuous = False
if args.dataset == 'bvh':
frames, frame_time, motion_data = dataset.load_bvh_data("%s/bvh/sample.bvh")
max_motion = np.max(motion_data, axis=0)
min_motion = np.min(motion_data, axis=0)
norm_motion_data = (motion_data - min_motion) / (max_motion - min_motion)
train_x = norm_motion_data
train_y = norm_motion_data
n_x = train_x.shape[1]
n_hidden = [250]
n_z = 10
n_y = n_x
n_online= 10
n_batch = train_x.shape[0] / n_online
if train_x.shape[0] % n_online != 0:
reduced_sample = train_x.shape[0] % n_online
train_x = train_x[:train_x.shape[0] - reduced_sample]
n_epochs = 500
continuous = True
n_hidden_recog = n_hidden
n_hidden_gen = n_hidden
n_layers_recog = len(n_hidden_recog)
n_layers_gen = len(n_hidden_gen)
layers = {}
# Recognition model.
rec_layer_sizes = [(train_x.shape[1], n_hidden_recog[0])]
rec_layer_sizes += zip(n_hidden_recog[:-1], n_hidden_recog[1:])
rec_layer_sizes += [(n_hidden_recog[-1], n_z)]
layers['recog_in_h'] = F.Linear(train_x.shape[1], n_hidden_recog[0], nobias=True)
layers['recog_h_h'] = F.Linear(n_hidden_recog[0], n_hidden_recog[0])
layers['recog_mean'] = F.Linear(n_hidden_recog[-1], n_z)
layers['recog_log_sigma'] = F.Linear(n_hidden_recog[-1], n_z)
# Generating model.
gen_layer_sizes = [(n_z, n_hidden_gen[0])]
gen_layer_sizes += zip(n_hidden_gen[:-1], n_hidden_gen[1:])
gen_layer_sizes += [(n_hidden_gen[-1], train_x.shape[1])]
layers['z'] = F.Linear(n_z, n_hidden_gen[0])
layers['gen_in_h'] = F.Linear(train_x.shape[1], n_hidden_gen[0], nobias=True)
layers['gen_h_h'] = F.Linear(n_hidden_gen[0], n_hidden_gen[0])
layers['output'] = F.Linear(n_hidden_gen[-1], train_x.shape[1])
if args.init_from == "":
model = VRAE(**layers)
else:
model = pickle.load(open(args.init_from))
# state pattern
state_pattern = ['recog_h', 'gen_h']
if args.gpu >= 0:
cuda.init(args.gpu)
model.to_gpu()
# use Adam
optimizer = optimizers.Adam()
optimizer.setup(model.collect_parameters())
total_losses = np.zeros(n_epochs, dtype=np.float32)
for epoch in xrange(1, n_epochs + 1):
print('epoch', epoch)
t1 = time.time()
total_rec_loss = 0.0
total_kl_loss = 0.0
total_loss = 0.0
outputs = np.zeros(train_x.shape, dtype=np.float32)
# state = make_initial_state(n_hidden_recog[0], state_pattern)
for i in xrange(n_batch):
state = make_initial_state(n_hidden_recog[0], state_pattern)
x_batch = split_x[i]
if args.gpu >= 0:
x_batch = cuda.to_gpu(x_batch)
output, rec_loss, kl_loss, state = model.forward_one_step(x_batch, state, continuous, nonlinear_q='tanh', nonlinear_p='tanh', output_f = 'sigmoid', gpu=-1)
outputs[i*seq_length:(i+1)*seq_length, :] = output
loss = rec_loss + kl_loss
total_loss += loss
total_rec_loss += rec_loss
total_losses[epoch-1] = total_loss.data
optimizer.zero_grads()
loss.backward()
loss.unchain_backward()
optimizer.clip_grads(args.clip_grads)
optimizer.update()
saved_output = outputs
print "{}/{}, train_loss = {}, total_rec_loss = {}, time = {}".format(epoch, n_epochs, total_loss.data, total_rec_loss.data, time.time()-t1)
if epoch % 100 == 0:
model_path = "%s/VRAE_%s_%d.pkl" % (args.output_dir, args.dataset, epoch)
with open(model_path, "w") as f:
pickle.dump(copy.deepcopy(model).to_cpu(), f)