-
Notifications
You must be signed in to change notification settings - Fork 27
/
train_lstm.py
259 lines (238 loc) · 12 KB
/
train_lstm.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
import os
import sys
import numpy as np
import lasagne as nn
import theano
import theano.tensor as T
import utils as u
import models as m
import config as c
def main(save_to='params',
dataset = 'mm',
kl_loss='true', # use kl-div in z-space instead of mse
diffs = 'false',
seq_length = 30,
num_epochs=1,
lstm_n_hid=1024,
max_per_epoch=-1
):
kl_loss = kl_loss.lower() == 'true'
diffs = diffs.lower() == 'true'
# set up functions for data pre-processing and model training
input_var = T.tensor4('inputs')
# different experimental setup for moving mnist vs pulp fiction dataests
if dataset == 'pf':
img_size = 64
cae_weights = c.pf_cae_params
cae_specstr = c.pf_cae_specstr
split_layer = 'conv7'
inpvar = T.tensor4('input')
net = m.build_cae(inpvar, specstr=cae_specstr, shape=(img_size, img_size))
convs_from_img,_ = m.encoder_decoder(cae_weights, specstr=cae_specstr,
layersplit=split_layer, shape=(img_size, img_size), poolinv=True)
laydict = dict((l.name, l) for l in nn.layers.get_all_layers(net))
zdec_in_shape = nn.layers.get_output_shape(laydict[split_layer])
deconv_weights = c.pf_deconv_params
vae_weights = c.pf_vae_params
img_from_convs = m.deconvoluter(deconv_weights, specstr=cae_specstr, shape=zdec_in_shape)
L=2
vae_n_hid = 1500
binary = False
z_dim = 256
l_tup = l_z_mu, l_z_ls, l_x_mu_list, l_x_ls_list, l_x_list, l_x = \
m.build_vae(input_var, L=L, binary=binary, z_dim=z_dim, n_hid=vae_n_hid,
shape=(zdec_in_shape[2], zdec_in_shape[3]), channels=zdec_in_shape[1])
u.load_params(l_x, vae_weights)
datafile = 'data/pf.hdf5'
frame_skip=3 # every 3rd frame in sequence
z_decode_layer = l_x_mu_list[0]
pixel_shift = 0.5
samples_per_image = 4
tr_batch_size = 16 # must be a multiple of samples_per_image
elif dataset == 'mm':
img_size = 64
cvae_weights = c.mm_cvae_params
L=2
vae_n_hid = 1024
binary = True
z_dim = 32
zdec_in_shape = (None, 1, img_size, img_size)
l_tup = l_z_mu, l_z_ls, l_x_mu_list, l_x_ls_list, l_x_list, l_x = \
m.build_vcae(input_var, L=L, z_dim=z_dim, n_hid=vae_n_hid, binary=binary,
shape=(zdec_in_shape[2], zdec_in_shape[3]), channels=zdec_in_shape[1])
u.load_params(l_x, cvae_weights)
datafile = 'data/moving_mnist.hdf5'
frame_skip=1
w,h=img_size,img_size # of raw input image in the hdf5 file
z_decode_layer = l_x_list[0]
pixel_shift = 0
samples_per_image = 1
tr_batch_size = 128 # must be a multiple of samples_per_image
# functions for moving to/from image or conv-space, and z-space
z_mat = T.matrix('z')
zenc = theano.function([input_var], nn.layers.get_output(l_z_mu, deterministic=True))
zdec = theano.function([z_mat], nn.layers.get_output(z_decode_layer, {l_z_mu:z_mat},
deterministic=True).reshape((-1, zdec_in_shape[1]) + zdec_in_shape[2:]))
zenc_ls = theano.function([input_var], nn.layers.get_output(l_z_ls, deterministic=True))
# functions for encoding sequences of z's
print 'compiling functions'
z_var = T.tensor3('z_in')
z_ls_var = T.tensor3('z_ls_in')
tgt_mu_var = T.tensor3('z_tgt')
tgt_ls_var = T.tensor3('z_ls_tgt')
learning_rate = theano.shared(nn.utils.floatX(1e-4))
# separate function definitions if we are using MSE and predicting only z, or KL divergence
# and predicting both mean and sigma of z
if kl_loss:
def kl(p_mu, p_sigma, q_mu, q_sigma):
return 0.5 * T.sum(T.sqr(p_sigma)/T.sqr(q_sigma) + T.sqr(q_mu - p_mu)/T.sqr(q_sigma)
- 1 + 2*T.log(q_sigma) - 2*T.log(p_sigma))
lstm, _ = m.Z_VLSTM(z_var, z_ls_var, z_dim=z_dim, nhid=lstm_n_hid, training=True)
z_mu_expr, z_ls_expr = nn.layers.get_output([lstm['output_mu'], lstm['output_ls']])
z_mu_expr_det, z_ls_expr_det = nn.layers.get_output([lstm['output_mu'],
lstm['output_ls']], deterministic=True)
loss = kl(tgt_mu_var, T.exp(tgt_ls_var), z_mu_expr, T.exp(z_ls_expr))
te_loss = kl(tgt_mu_var, T.exp(tgt_ls_var), z_mu_expr_det, T.exp(z_ls_expr_det))
params = nn.layers.get_all_params(lstm['output'], trainable=True)
updates = nn.updates.adam(loss, params, learning_rate=learning_rate)
train_fn = theano.function([z_var, z_ls_var, tgt_mu_var, tgt_ls_var], loss,
updates=updates)
test_fn = theano.function([z_var, z_ls_var, tgt_mu_var, tgt_ls_var], te_loss)
else:
lstm, _ = m.Z_LSTM(z_var, z_dim=z_dim, nhid=lstm_n_hid, training=True)
loss = nn.objectives.squared_error(nn.layers.get_output(lstm['output']),
tgt_mu_var).mean()
te_loss = nn.objectives.squared_error(nn.layers.get_output(lstm['output'],
deterministic=True), tgt_mu_var).mean()
params = nn.layers.get_all_params(lstm['output'], trainable=True)
updates = nn.updates.adam(loss, params, learning_rate=learning_rate)
train_fn = theano.function([z_var, tgt_mu_var], loss, updates=updates)
test_fn = theano.function([z_var, tgt_mu_var], te_loss)
if dataset == 'pf':
z_from_img = lambda x: zenc(convs_from_img(x))
z_ls_from_img = lambda x: zenc_ls(convs_from_img(x))
img_from_z = lambda z: img_from_convs(zdec(z))
elif dataset == 'mm':
z_from_img = zenc
z_ls_from_img = zenc_ls
img_from_z = zdec
# training loop
print('training for {} epochs'.format(num_epochs))
nbatch = (seq_length+1) * tr_batch_size * frame_skip / samples_per_image
data = u.DataH5PyStreamer(datafile, batch_size=nbatch)
# for taking arrays of uint8 (non square) and converting them to batches of sequences
def transform_data(ims_batch, center=False):
imb = u.raw_to_floatX(ims_batch, pixel_shift=pixel_shift,
center=center)[np.random.randint(frame_skip)::frame_skip]
zbatch = np.zeros((tr_batch_size, seq_length+1, z_dim), dtype=theano.config.floatX)
zsigbatch = np.zeros((tr_batch_size, seq_length+1, z_dim), dtype=theano.config.floatX)
for i in xrange(samples_per_image):
chunk = tr_batch_size/samples_per_image
if diffs:
zf = z_from_img(imb).reshape((chunk, seq_length+1, -1))
zbatch[i*chunk:(i+1)*chunk, 1:] = zf[:,1:] - zf[:,:-1]
if kl_loss:
zls = z_ls_from_img(imb).reshape((chunk, seq_length+1, -1))
zsigbatch[i*chunk:(i+1)*chunk, 1:] = zls[:,1:] - zls[:,:-1]
else:
zbatch[i*chunk:(i+1)*chunk] = z_from_img(imb).reshape((chunk, seq_length+1, -1))
if kl_loss:
zsigbatch[i*chunk:(i+1)*chunk] = z_ls_from_img(imb).reshape((chunk,
seq_length+1, -1))
if kl_loss:
return zbatch[:,:-1,:], zsigbatch[:,:-1,:], zbatch[:,1:,:], zsigbatch[:,1:,:]
return zbatch[:,:-1,:], zbatch[:,1:,:]
# we need sequences of images, so we do not shuffle data during trainin
hist = u.train_with_hdf5(data, num_epochs=num_epochs, train_fn=train_fn, test_fn=test_fn,
train_shuffle=False,
max_per_epoch=max_per_epoch,
tr_transform=lambda x: transform_data(x[0], center=False),
te_transform=lambda x: transform_data(x[0], center=True))
hist = np.asarray(hist)
u.save_params(lstm['output'], os.path.join(save_to, 'lstm_{}.npz'.format(hist[-1,-1])))
# build functions to sample from LSTM
# separate cell_init and hid_init from the other learned model parameters
all_param_values = nn.layers.get_all_param_values(lstm['output'])
init_indices = [i for i,p in enumerate(nn.layers.get_all_params(lstm['output']))
if 'init' in str(p)]
init_values = [all_param_values[i] for i in init_indices]
params_noinit = [p for i,p in enumerate(all_param_values) if i not in init_indices]
# build model without learnable init values, and load non-init parameters
if kl_loss:
lstm_sample, state_vars = m.Z_VLSTM(z_var, z_ls_var, z_dim=z_dim, nhid=lstm_n_hid,
training=False)
else:
lstm_sample, state_vars = m.Z_LSTM(z_var, z_dim=z_dim, nhid=lstm_n_hid, training=False)
nn.layers.set_all_param_values(lstm_sample['output'], params_noinit)
# extract layers representing thee hidden and cell states, and have sample_fn
# return their outputs
state_layers_keys = [k for k in lstm_sample.keys() if 'hidfinal' in k or 'cellfinal' in k]
state_layers_keys = sorted(state_layers_keys)
state_layers_keys = sorted(state_layers_keys, key = lambda x:int(x.split('_')[1]))
state_layers = [lstm_sample[s] for s in state_layers_keys]
if kl_loss:
sample_fn = theano.function([z_var, z_ls_var] + state_vars,
nn.layers.get_output([lstm['output_mu'], lstm['output_ls']] + state_layers,
deterministic=True))
else:
sample_fn = theano.function([z_var] + state_vars,
nn.layers.get_output([lstm['output']] + state_layers, deterministic=True))
from images2gif import writeGif
from PIL import Image
# sample approximately 30 different generated video sequences
te_stream = data.streamer(training=True, shuffled=False)
interval = data.ntrain / data.batch_size / 30
for idx,imb in enumerate(te_stream.get_epoch_iterator()):
if idx % interval != 0:
continue
z_tup = transform_data(imb[0], center=True)
seg_idx = np.random.randint(z_tup[0].shape[0])
if kl_loss:
z_in, z_ls_in = z_tup[0], z_tup[1]
z_last, z_ls_last = z_in[seg_idx:seg_idx+1], z_ls_in[seg_idx:seg_idx+1]
z_vars = [z_last, z_ls_last]
else:
z_in = z_tup[0]
z_last = z_in[seg_idx:seg_idx+1]
z_vars = [z_last]
images = []
state_values = [np.dot(np.ones((z_last.shape[0],1), dtype=theano.config.floatX), s)
for s in init_values]
output_list = sample_fn(*(z_vars + state_values))
# use whole sequence of predictions for output
z_pred = output_list[0]
state_values = output_list[2 if kl_loss else 1:]
rec = img_from_z(z_pred.reshape(-1, z_dim))
for k in xrange(rec.shape[0]):
images.append(Image.fromarray(u.get_picture_array(rec, index=k, shift=pixel_shift)))
k += 1
# slice prediction to feed into lstm
z_pred = z_pred[:,-1:,:]
if kl_loss:
z_ls_pred = output_list[1][:,-1:,:]
z_vars = [z_pred, z_ls_pred]
else:
z_vars = [z_pred]
for i in xrange(30): # predict 30 frames after the end of the priming video
output_list = sample_fn(*(z_vars + state_values))
z_pred = output_list[0]
state_values = output_list[2 if kl_loss else 1:]
rec = img_from_z(z_pred.reshape(-1, z_dim))
images.append(Image.fromarray(u.get_picture_array(rec, index=0, shift=pixel_shift)))
if kl_loss:
z_ls_pred = output_list[1]
z_vars = [z_pred, z_ls_pred]
else:
z_vars = [z_pred]
writeGif("sample_{}.gif".format(idx),images,duration=0.1,dither=0)
if __name__ == '__main__':
# make all arguments of main(...) command line arguments (with type inferred from
# the default value) - this doesn't work on bools so those are strings when
# passed into main.
import argparse, inspect
parser = argparse.ArgumentParser(description='Command line options')
ma = inspect.getargspec(main)
for arg_name,arg_type in zip(ma.args[-len(ma.defaults):],[type(de) for de in ma.defaults]):
parser.add_argument('--{}'.format(arg_name), type=arg_type, dest=arg_name)
args = parser.parse_args(sys.argv[1:])
main(**{k:v for (k,v) in vars(args).items() if v is not None})