forked from maccam912/dnc-theano
/
dnc.py
executable file
·277 lines (243 loc) · 8.63 KB
/
dnc.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
import sys
from random import randint
import numpy as np
import theano as th
import theano.tensor as T
from theano.scan_module import until
import layers as lyr
from optimize import AdamSGD, VanillaSGD
th.config.floatX = 'float32'
#hyperparameters are in CAPS
SEQ_LEN = 32
INP_DIMS = 8
OUT_DIMS = 8
N_CELLS = 32 #number of memory cells
CELL_SIZE = 64 #size of each memory cell
N_READS = 4 #number of read heads
B_SIZE = 32 #unused for now
LR = 1e-3 #learn rate
EPS = 1e-6 #to avoid division by zero
g_params = {}
g_states = {}
g_optimizer = AdamSGD()
g_optimizer.lr = LR
fn_predict = None
#temporal hack for theano issue #5197 at github
#TODO: revert to normal implementation once the issue is fixed
if th.config.device[:3] == 'cpu':
def op_cumprod_hack(s_x__, axis_=None):
return T.extra_ops.cumprod(s_x_*0.99+0.01, axis=axis_)
else:
def op_cumprod_hack(s_x_, axis_=None):
#due to cumprod has only CPU implementation
return T.exp(T.extra_ops.cumsum(T.log(s_x_*0.99+0.01), axis=axis_))
def build_model():
global g_params, g_states, g_optimizer, fn_predict, fn_rst
ctrl_inp_size = CELL_SIZE*N_READS+INP_DIMS
itrface_size = CELL_SIZE*N_READS+3*CELL_SIZE+5*N_READS+3
ctrl_wm_size = OUT_DIMS+itrface_size
#states
lyr.set_current_params(g_states)
v_lstm_cell = lyr.get_variable('lstm_cell',(ctrl_wm_size,))
v_lstm_hid = lyr.get_variable('lstm_hid',(ctrl_wm_size,))
v_usage = lyr.get_variable('usage',(N_CELLS,))
v_preced = lyr.get_variable('preced',(N_CELLS,))
v_link = lyr.get_variable('link',(N_CELLS,N_CELLS))
v_mem = lyr.get_variable('mem',(N_CELLS, CELL_SIZE))
v_read_val = lyr.get_variable('r_val',(N_READS,CELL_SIZE))
v_read_wgt = lyr.get_variable('r_wgt',(N_READS,N_CELLS))
v_write_wgt = lyr.get_variable('w_wgt',(N_CELLS,))
#build the actual model
lyr.set_current_params(g_params)
def dnc_step(
s_x_,
s_lstm_cell_,
s_lstm_hid_,
s_usage_,
s_preced_,
s_link_,
s_mem_,
s_read_val_,
s_read_wgt_,
s_write_wgt_):
s_states_li_ = [
s_lstm_cell_,
s_lstm_hid_,
s_usage_,
s_preced_,
s_link_,
s_mem_,
s_read_val_,
s_read_wgt_,
s_write_wgt_]
s_inp = T.join(-1, s_x_, s_read_val_.flatten())
s_lstm_cell_tp1, s_lstm_hid_tp1 = lyr.lyr_lstm(
'ctrl',
s_inp, s_lstm_cell_, s_lstm_hid_,
ctrl_inp_size, ctrl_wm_size
)
s_out, s_itrface = T.split(
lyr.lyr_linear(
'ctrl_out', s_lstm_hid_tp1, ctrl_wm_size, ctrl_wm_size, bias_=None),
[OUT_DIMS,itrface_size],2, axis=-1)
splits_len = [
N_READS*CELL_SIZE, N_READS, CELL_SIZE, 1,
CELL_SIZE, CELL_SIZE, N_READS, 1, 1, 3*N_READS
]
s_keyr, s_strr, s_keyw, s_strw, \
s_ers, s_write, s_freeg, s_allocg, s_writeg, s_rmode = \
T.split(s_itrface, splits_len, 10, axis=-1)
s_keyr = T.reshape(s_keyr, (CELL_SIZE,N_READS))
s_strr = 1.+T.nnet.softplus(s_strr)
s_strw = 1.+T.nnet.softplus(s_strw[0])
s_ers = T.nnet.sigmoid(s_ers)
s_freeg = T.nnet.sigmoid(s_freeg)
s_allocg = T.nnet.sigmoid(s_allocg[0])
s_writeg = T.nnet.sigmoid(s_writeg[0])
s_rmode = T.nnet.softmax(T.reshape(s_rmode,(N_READS,3))).dimshuffle(1,0,'x')
s_mem_retention = T.prod(
1.-s_freeg.dimshuffle(0,'x')*s_read_wgt_, axis=0)
s_usage_tp1 = s_mem_retention*(
s_usage_+s_write_wgt_-s_usage_*s_write_wgt_)
s_usage_order = T.argsort(s_usage_tp1)
s_usage_order_inv = T.inverse_permutation(s_usage_order)
s_usage_tp1_sorted = s_usage_tp1[s_usage_order]
s_alloc_wgt = ((1.-s_usage_tp1_sorted)*(
T.join(
0,np.array([1.],dtype=th.config.floatX),
op_cumprod_hack(s_usage_tp1_sorted[:-1])
)))[s_usage_order_inv]
s_content_wgt_w = T.nnet.softmax(
s_strw*T.dot(s_mem_, s_keyw)/(
T.sqrt(
EPS+T.sum(T.sqr(s_mem_),axis=-1)*T.sum(T.sqr(s_keyw))))
).flatten()
s_write_wgt_tp1 = s_writeg*(
s_allocg*s_alloc_wgt+(1.-s_allocg)*s_content_wgt_w)
s_mem_tp1 = s_mem_*(
1.-T.outer(s_write_wgt_tp1,s_ers))+T.outer(s_write_wgt_tp1,s_write)
s_preced_tp1 = (1.-T.sum(s_write_wgt_))*s_preced_ + s_write_wgt_tp1
s_link_tp1 = (
1.-s_write_wgt_tp1-s_write_wgt_tp1.dimshuffle(0,'x')
)*s_link_ + T.outer(s_write_wgt_tp1,s_preced_)
s_link_tp1 = s_link_tp1 * (1.-T.identity_like(s_link_tp1))#X
s_fwd = T.dot(s_read_wgt_, s_link_tp1.transpose())#X
s_bwd = T.dot(s_read_wgt_, s_link_tp1)#X
s_content_wgt_r= T.nnet.softmax(T.dot(s_mem_tp1, s_keyr)/(T.sqrt(
EPS+T.outer(
T.sum(T.sqr(s_mem_tp1),axis=-1),T.sum(T.sqr(s_keyr),axis=0)
)))).transpose()
s_read_wgt_tp1 = s_bwd*s_rmode[0]+s_content_wgt_r*s_rmode[1]+s_fwd*s_rmode[2]
s_read_val_tp1 = T.dot(s_read_wgt_tp1, s_mem_tp1)
s_y = s_out + lyr.lyr_linear(
'read_out',
s_read_val_tp1.flatten(),
CELL_SIZE*N_READS,OUT_DIMS,
bias_=None)
return [
s_y,
s_lstm_cell_tp1,
s_lstm_hid_tp1,
s_usage_tp1,
s_preced_tp1,
s_link_tp1,
s_mem_tp1,
s_read_val_tp1,
s_read_wgt_tp1,
s_write_wgt_tp1]
s_x_li = T.matrix()
s_y_target_li = T.matrix()
v_states_li = [
v_lstm_cell,
v_lstm_hid,
v_usage,
v_preced,
v_link,
v_mem,
v_read_val,
v_read_wgt,
v_write_wgt
]
s_outputs_li, _ = th.scan(
dnc_step,
sequences=[s_x_li],
outputs_info=[None]+v_states_li
)
s_y_li = s_outputs_li[0]
s_loss = T.mean(T.sqr(s_y_li - s_y_target_li))
new_states = [s[-1] for s in s_outputs_li[1:]]
print('Compiling ... ', end='')
sys.stdout.flush()
g_optimizer.compile(
[s_x_li, s_y_target_li],
s_loss,
list(g_params.values()),
updates_=list(zip(v_states_li, new_states)),
fetches_= s_loss
)
fn_predict = th.function([s_x_li], s_y_li)
fn_rst = th.function([], updates=[(v,T.zeros_like(v)) for v in v_states_li])
print('Done')
def predict(v_x_):
return fn_predict(v_x_)
def reset_states():
fn_rst()
def gen_episode(lenr_ =(2,4)):
global SEQ_LEN, INP_DIMS, OUT_DIMS
X = np.zeros((SEQ_LEN,INP_DIMS),dtype=th.config.floatX)
Y = np.zeros((SEQ_LEN,OUT_DIMS),dtype=th.config.floatX)
data_len = min(randint(*lenr_),(SEQ_LEN-3)//2)
tot_len = data_len*2+4
offset = randint(0,SEQ_LEN-1-tot_len)
m = np.random.binomial(1,0.5,(data_len,INP_DIMS))
coinflip = randint(0,1)
if coinflip:
X[offset:offset+1,INP_DIMS//2:] = 1.
X[1+offset:1+offset+data_len] = m
Y[offset+data_len+4:offset+4+data_len*2] = m
else:
X[offset:offset+1,:INP_DIMS//2] = 1.
X[1+offset:1+offset+data_len] = m
Y[offset+data_len+4:offset+4+data_len*2] = m[::-1]
return X, Y
def save_params():
global g_params, g_states
with open('params.pkl','wb') as f:
lyr.save_params(g_params, f)
lyr.save_params(g_states, f)
def load_params():
global g_params, g_states
with open('params.pkl','rb') as f:
lyr.load_params(g_params, f)
lyr.load_params(g_states, f)
def train(nsess_=100, nitr_=100, lenr_=(2,4)):
'''
Trains model for some sessions, parameters are automatically saved to file after each session.
Args:
nsess_: number of training episodes.
nitr_: number of iteration.
lenr_: signal length range of copy task, must be tuple.
'''
try:
for i in range(nsess_):
loss = 0.
for j in range(nitr_):
X,Y = gen_episode(lenr_)
loss += g_optimizer.fit(X,Y)
if str(loss)=='nan': break
sys.stdout.write('.')
sys.stdout.flush()
else:
loss /= nitr_
print('\nIter %d/%d loss: %f'%((i+1)*nitr_,nsess_*nitr_,loss))
save_params()
continue
break
else:
print('Training finished.')
return
print('\nIter %d/%d loss: ???'%(i*nitr_+j,nsess_*nitr_))
print('ERROR: Model crashed, stop.')
except KeyboardInterrupt:
print('User hit CTRL-C, stop.')
build_model()