-
Notifications
You must be signed in to change notification settings - Fork 0
/
char_rnn_theano_uncleaned.py
245 lines (199 loc) · 6.91 KB
/
char_rnn_theano_uncleaned.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
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation
from keras.layers.embeddings import Embedding
from keras.layers.recurrent import LSTM
from copy import deepcopy
from numpy import array as npa
import numpy as np
from numpy.random import RandomState
from theano import tensor as T
import string
import sys
from jbpickle import pickle, unpickle
import os
SAVEDIR = './out/'
rng = RandomState() # will reseed after param selection
# random number string for book keeping
pool = string.ascii_uppercase + string.digits
RUNID = ''.join([str(pool[rng.randint(len(pool))]) for _ in xrange(8)])
def mkdir(directory):
if not os.path.exists(directory):
os.makedirs(directory)
SAVEDIR = SAVEDIR + RUNID + '/'
WDIR = SAVEDIR + 'W/'
mkdir(SAVEDIR)
mkdir(WDIR)
PRINTTO = SAVEDIR + 'generated_output.txt'
SEQLEN = rng.randint(20, 150)
SHUF_DATA = rng.randint(2)
NLAYER = rng.randint(1, 4)
NHID = [rng.randint(50, 500) for _ in xrange(NLAYER)]
PDROP = [rng.rand() for _ in xrange(NLAYER)]
BATCHSIZE = rng.randint(50, 200)
LRINIT = rng.uniform(0.001, 0.1)
RNGSEED = rng.randint(4525348)
rng = RandomState(RNGSEED)
runparams = {'SEQLEN': SEQLEN,
'SHUF_DATA': SHUF_DATA,
'NLAYER': NLAYER,
'NHID': NHID,
'PDROP': PDROP,
'BATCHSIZE': BATCHSIZE,
'LRINIT': LRINIT,
'RNGSEED': RNGSEED}
pickle(runparams, SAVEDIR+'runparams.pkl')
print 'RUNID: ' + RUNID
for k, v in runparams.items():
print k + ': ' + str(v)
# save params for book keeping
with open(PRINTTO, 'w+') as f:
for k,v in runparams.items():
f.write('\n')
f.write('%s: %s' % (k,v))
f.write('\n\n')
NB_EPOCH = 1
NEPOCH = 50
LRDECAY = 0.97
LRDECAYAFTER = 10
FNAME = 'input.txt'
PTRAIN = 0.7
PVAL = 0.15
PTEST = 0.15
# x will be text. y will be same thing shifted by 1
def load_text(fname):
lochars = []
chars = {}
ichar = -1
with open(fname) as f:
while True:
c = f.read(1)
if not c: break # exit if file over
if c not in chars:
ichar += 1
chars[c] = ichar
lochars.append(c)
return (lochars, chars)
def lochars_to_mats(text, char2i):
ntot = len(text)
nchar = len(char2i)
X = np.zeros([ntot, nchar])
for ic,c in enumerate(text):
X[ic, char2i[c]] = 1.
y = deepcopy(X)
y = np.vstack([y[-1], y[:-1]]) # wraps around to beginning, i guess
return (X, y)
def seqify(X, y, seqlen):
ntot, nchar = X.shape
nseq = ntot / seqlen
clip = ntot % seqlen # remove dangling text
if clip != 0:
X = X[:-clip] # make so can have even seqlens
y = y[:-clip] # make so can have even seqlens
Xseq = np.zeros([nseq, seqlen, nchar])
# yseq = np.zeros([nseq, seqlen, nchar])
yseq = np.zeros([nseq, nchar])
for iseq in xrange(nseq):
Xseq[iseq] = X[seqlen*iseq:seqlen*(iseq+1)]
if iseq==nseq-1: yseq[iseq] = X[0]
else: yseq[iseq] = X[seqlen*(iseq+1)]
# yseq[iseq] = y[seqlen*iseq:seqlen*(iseq+1)]
return Xseq, yseq
def trainvaltest_split(Xseq, yseq, p_train, p_val, p_test=None, shuffle=False):
ncase, seqlen, nbatch = Xseq.shape
if not p_test: p_test = 1.-(p_train+p_val)
# can be less if don't want to use full dataset
assert npa([p_train, p_val, p_test]).sum() <= 1.
ntrain, nval, ntest = [int(p * ncase) for p in [p_train, p_val, p_test]]
if shuffle:
neworder = rng.permutation(ncase)
Xseq = Xseq[neworder]
yseq = yseq[neworder]
return {'train': {'X': Xseq[:ntrain], 'y': yseq[:ntrain]},
'val': {'X': Xseq[ntrain:ntrain+nval], 'y': yseq[ntrain:ntrain+nval]},
'test': {'X': Xseq[ntrain+nval:ntrain+nval+ntest], 'y': yseq[ntrain+nval:ntrain+nval+ntest]}
}
# for generation
def onehot_to_char(onehot, i2char):
return i2char[onehot.argmax()]
def char_to_onehot(char, char2i):
onehot = np.zeros(len(char2i))
onehot[char2i[char]] = 1.
return onehot
def softmax(arr, temp):
assert temp >= 0
if temp == 0:
out = np.zeros_like(arr)
out[arr.argmax()] = 1.
else:
out = np.exp(arr/temp)
out /= out.sum()
return out
def sample(arr, temp):
return rng.choice(np.arange(len(arr)), p=softmax(arr, temp))
print 'prepping data'
lochars, char2i = load_text(FNAME)
i2char = {v: k for k, v in char2i.items()}
nchar = len(char2i)
X, y = lochars_to_mats(lochars, char2i)
Xseq, yseq = seqify(X, y, SEQLEN)
tvt = trainvaltest_split(Xseq, yseq, PTRAIN, PVAL, PTEST, SHUF_DATA)
if not NB_EPOCH:
NB_EPOCH = tvt['train']['X'].shape[0] / BATCHSIZE
# DEFINE MODEL
model = Sequential()
# add stacked lstm layers
for ilayer in xrange(NLAYER):
if ilayer==0: NIN = nchar
else: NIN = NHID[ilayer-1]
RET_SEQ = ilayer != NLAYER-1
model.add(LSTM(NIN, NHID[ilayer],\
return_sequences=RET_SEQ,\
activation='tanh',\
inner_activation='hard_sigmoid'))
if PDROP[ilayer] != 0:
model.add(Dropout(PDROP[ilayer]))
# readout layer
model.add(Dense(NHID[-1], nchar))
# def logsoftmax(x):
# return T.log(T.nnet.softmax(x))
model.add(Activation('softmax'))
print 'compling model...'
model.compile(loss='categorical_crossentropy', optimizer='rmsprop')
for ie in xrange(NEPOCH):
print ''.join(['epoch ', str(ie)])
if ie < LRDECAYAFTER:
model.optimizer.lr = LRINIT
else:
model.optimizer.lr *= LRDECAY
trainscores = model.fit(tvt['train']['X'], tvt['train']['y'],\
batch_size=BATCHSIZE,
nb_epoch=NB_EPOCH)
wname = ''.join([WDIR, 'epoch_', str(ie), '.hdf5'])
model.save_weights(wname)
valscore = model.evaluate(tvt['val']['X'], tvt['val']['y'],\
batch_size=BATCHSIZE)
# print out for human eval
outlen = 600 # num char to predict out
istart = rng.randint(tvt['test']['X'].shape[0]-outlen-1)
for temp in np.linspace(0., 4., 9):
print()
print('----- temp:', temp)
print '----- epoch:' + str(ie)
starter = X[istart:istart+SEQLEN]
generated = ''.join([onehot_to_char(oh, i2char) for oh in starter])
inittxt = deepcopy(generated)
print('----- Generating with seed: "' + generated + '"')
sys.stdout.write(generated)
for ig in xrange(outlen):
preds = model.predict(starter[None,:], verbose=0)[0]
next_ichar = sample(preds, temp)
next_char = i2char[next_ichar]
generated += next_char
starter = np.vstack([starter[1:], char_to_onehot(next_char, char2i)])
# print to cmd
# sys.stdout.write(next_char)
# sys.stdout.flush()
with open(PRINTTO, 'a') as f:
f.write('\ngenerated with seed \'%s\n\'' % inittxt)
f.write('\n\nepoch %d, temp=%d:\n\n' % (ie, temp))
f.write(generated)