forked from asampat3090/arctic-captions
-
Notifications
You must be signed in to change notification settings - Fork 0
/
generate_caps.py
180 lines (151 loc) · 6 KB
/
generate_caps.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
"""
Sampling script for attention models
Works on CPU with support for multi-process
"""
import argparse
import numpy
import cPickle as pkl
from util import load_params, init_tparams
from capgen import build_sampler, gen_sample, \
init_params, \
get_dataset \
from multiprocessing import Process, Queue
# Helper functions
import sys
import pdb
class ForkedPdb(pdb.Pdb):
"""A Pdb subclass that may be used
from a forked multiprocessing child
"""
def interaction(self, *args, **kwargs):
_stdin = sys.stdin
try:
sys.stdin = file('/dev/stdin')
pdb.Pdb.interaction(self, *args, **kwargs)
finally:
sys.stdin = _stdin
# single instance of a sampling process
def gen_model(queue, rqueue, pid, model, options, k, normalize, word_idict, sampling):
import theano
from theano import tensor
from theano.sandbox.rng_mrg import MRG_RandomStreams as RandomStreams
trng = RandomStreams(1234)
# this is zero indicate we are not using dropout in the graph
use_noise = theano.shared(numpy.float32(0.), name='use_noise')
# get the parameters
params = init_params(options)
params = load_params(model, params)
tparams = init_tparams(params)
# build the sampling computational graph
# see capgen.py for more detailed explanations
f_init, f_next = build_sampler(tparams, options, use_noise, trng, sampling=sampling)
def _gencap(cc0):
sample, score = gen_sample(tparams, f_init, f_next, cc0, options,
trng=trng, k=k, maxlen=200, stochastic=False)
# adjust for length bias
if normalize:
lengths = numpy.array([len(s) for s in sample])
score = score / lengths
sidx = numpy.argmin(score)
return sample[sidx]
while True:
req = queue.get()
# exit signal
if req is None:
break
idx, context = req[0], req[1]
print "Processing example %d in process # %d" % (idx, pid)
seq = _gencap(context)
print seq
rqueue.put((idx, seq))
print "Added example %d to the result queue" % idx
print "gen_model process w/ pid %d has returned..." % pid
return
def main(model, saveto, k=5, normalize=False, zero_pad=False, n_process=5, datasets='dev,test', sampling=False, pkl_name=None):
# load model model_options
if pkl_name is None:
pkl_name = model
with open('%s.pkl'% pkl_name, 'rb') as f:
options = pkl.load(f)
# fetch data, skip ones we aren't using to save time
load_data, prepare_data = get_dataset(options['dataset'])
_, valid, test, worddict = load_data(path='./data', load_train=False, load_dev=True if 'dev' in datasets else False,
load_test=True if 'test' in datasets else False)
# <eos> means end of sequence (aka periods), UNK means unknown
word_idict = dict()
for kk, vv in worddict.iteritems():
word_idict[vv] = kk
word_idict[0] = '<eos>'
word_idict[1] = 'UNK'
# create processes
queue = Queue()
rqueue = Queue()
processes = [None] * n_process
for midx in xrange(n_process):
processes[midx] = Process(target=gen_model,
args=(queue,rqueue,midx,model,options,k,normalize,word_idict, sampling))
processes[midx].start()
# index -> words
def _seqs2words(caps):
capsw = []
for cc in caps:
ww = []
for w in cc:
if w == 0:
break
ww.append(word_idict[w])
capsw.append(' '.join(ww))
return capsw
# unsparsify, reshape, and queue
def _send_jobs(contexts):
for idx, ctx in enumerate(contexts):
cc = ctx.todense().reshape([14*14,512])
if zero_pad:
cc0 = numpy.zeros((cc.shape[0]+1, cc.shape[1])).astype('float32')
cc0[:-1,:] = cc
else:
cc0 = cc
queue.put((idx, cc0))
# retrieve caption from process
def _retrieve_jobs(n_samples):
caps = [None] * n_samples
for idx in xrange(n_samples):
resp = rqueue.get()
caps[resp[0]] = resp[1]
if numpy.mod(idx, 10) == 0:
print 'Sample ', (idx+1), '/', n_samples, ' Done'
return caps
ds = datasets.strip().split(',')
# send all the features for the various datasets
for dd in ds:
if dd == 'dev':
print 'Development Set...',
_send_jobs(valid[1])
caps = _seqs2words(_retrieve_jobs(valid[1].shape[0]))
import pdb; pdb.set_trace()
with open(saveto+'.dev.txt', 'w') as f:
print >>f, '\n'.join(caps)
print 'Done'
if dd == 'test':
print 'Test Set...',
_send_jobs(test[1])
caps = _seqs2words(_retrieve_jobs(test[1].shape[0]))
with open(saveto+'.test.txt', 'w') as f:
print >>f, '\n'.join(caps)
print 'Done'
# end processes
for midx in xrange(n_process):
queue.put(None)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('-k', type=int, default=1)
parser.add_argument('-sampling', action="store_true", default=False) # this only matters for hard attention
parser.add_argument('-p', type=int, default=5, help="number of processes to use")
parser.add_argument('-n', action="store_true", default=False)
parser.add_argument('-z', action="store_true", default=False)
parser.add_argument('-d', type=str, default='dev,test')
parser.add_argument('-pkl_name', type=str, default=None, help="name of pickle file (without the .pkl)")
parser.add_argument('model', type=str)
parser.add_argument('saveto', type=str)
args = parser.parse_args()
main(args.model, args.saveto, k=args.k, zero_pad=args.z, pkl_name=args.pkl_name, n_process=args.p, normalize=args.n, datasets=args.d, sampling=args.sampling)