-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
executable file
·459 lines (432 loc) · 24.1 KB
/
main.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
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
#!/usr/bin/env python
from __future__ import division
import sys
sys.path.append('/home/gchrupala/repos/Passage')
sys.path.append('/home/gchrupala/repos/neuraltalk')
from passage.layers import Embedding, SimpleRecurrent, LstmRecurrent, GatedRecurrent #, Dense
from layers import *
from passage.costs import MeanSquaredError
from imaginet import *
from passage.preprocessing import Tokenizer, tokenize
import passage.utils
import passage.updates
from passage.iterators import SortedPadded
import imagernn.data_provider as dp
import cPickle
from scipy.spatial.distance import cosine, cdist
import numpy
import os.path
import argparse
import random
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.linear_model import Ridge
from sklearn.preprocessing import StandardScaler
import json
import gzip
def main():
parser = argparse.ArgumentParser(
description='Learn to rank images according to similarity to \
caption meaning')
parser.add_argument('--predict', dest='predict',
action='store_true', help='Run in prediction mode')
parser.add_argument('--paraphrase', dest='paraphrase',
action='store_true', help='Run in paraphrasing mode')
parser.add_argument('--paraphrase_state', dest='paraphrase_state', default='hidden_multi',
help='Which state to use for paraphrase retrieval (hidden_multi, hidden_vis, hidden_text, output_vis)')
parser.add_argument('--extract_embeddings', dest='extract_embeddings',
action='store_true',
help='Extract embeddings from trained model')
parser.add_argument('--project_words', dest='project_words',
action='store_true', help='Project words from vocabulary to visual space')
parser.add_argument('--model', dest='model', default='model.dat.gz',
help='Path to write model to')
parser.add_argument('--model_type', dest='model_type', default='simple',
help='Type of model: (linear, simple, shared_embeddings, shared_all)')
parser.add_argument('--character', dest='character', action='store_true',
help='Character-level model')
parser.add_argument('--zero_shot', dest='zero_shot', action='store_true',
help='Disable visual signal for sentences containing words in zero_shot.pkl.gz')
parser.add_argument('--tokenizer', dest='tokenizer', default='tok.pkl.gz',
help='Path to write tokenizer to')
parser.add_argument('--init_model', dest='init_model', default=None,
help='Initialize model weights with model from given path')
parser.add_argument('--init_tokenizer', dest='init_tokenizer', default=None,
help='Use tokenizer from given path')
parser.add_argument('--iter_predict', type=int,
help='Model after that many iterations will be used to predict')
parser.add_argument('--scramble', action='store_true',
help='Scramble words in a test sentence')
parser.add_argument('--distance', default='cosine',
help='Distance metric to rank images')
parser.add_argument('--dataset', dest='dataset', default='flickr8k',
help='Dataset: flick8k, flickr30k, coco')
parser.add_argument('--hidden_size', dest='hidden_size', type=int, default=256,
help='size of the hidden layer')
parser.add_argument('--embedding_size', dest='embedding_size', type=int, default=None,
help='size of (word) embeddings')
parser.add_argument('--hidden_type', default='gru',
help='recurrent layer type: gru, lstm')
parser.add_argument('--activation', default='tanh',
help='activation of the hidden layer units')
parser.add_argument('--out_activation', default='linear',
help='Activation of output units')
parser.add_argument('--cost', default='MeanSquaredError',
help='Image prediction cost function')
parser.add_argument('--scaler', dest='scaler', default='none',
help='Method to scale targets (none, standard)')
parser.add_argument('--rate', dest='rate', type=float, default=0.0002,
help='Learning rate')
parser.add_argument('--clipnorm', dest='clipnorm', type=float, default=0.0,
help='Gradients with norm larger than clipnorm will be scaled')
parser.add_argument('--alpha', dest='alpha', type=float, default=0.0,
help='Interpolation parameter for LM cost vs image cost')
parser.add_argument('--ridge_alpha', dest='ridge_alpha', type=float, default=1.0,
help='Regularization for linear regression model')
parser.add_argument('--non_interpolated', dest='non_interpolated', action='store_true',
help='Use non-interpolated cost')
parser.add_argument('--iterations', dest='iterations', type=int, default=10,
help='Number of training iterations')
parser.add_argument('--word_freq_threshold', dest='word_freq_threshold', type=int, default=10,
help='Map words below this threshold to UNK')
parser.add_argument('--shuffle', dest='shuffle', action='store_true',
help='Shuffle training data')
parser.add_argument('--random_seed', dest='random_seed', default=None, type=int,
help='Random seed')
parser.add_argument('--snapshot_freq', dest='snapshot_freq', type=int, default=5,
help='How many iterations to save model')
parser.add_argument('--batch_size', dest='batch_size', type=int, default=64,
help='Batch size')
args = parser.parse_args()
if args.random_seed is not None:
numpy.random.seed(args.random_seed)
if args.project_words:
project_words(args)
elif args.predict and args.model_type == 'linear':
test_linear(args)
elif args.predict and args.model_type != 'linear':
test(args)
elif args.extract_embeddings:
extract_embeddings(args)
elif args.model_type == 'linear':
train_linear(args)
else:
train(args)
def train_linear(args):
p = dp.getDataProvider(args.dataset)
data = list(p.iterImageSentencePair(split='train'))
texts = [ pair['sentence']['raw'] for pair in data ]
images = [ pair['image']['feat'] for pair in data ]
analyzer = 'char' if args.character else 'word'
vectorizer = CountVectorizer(min_df=args.word_freq_threshold, analyzer=analyzer, lowercase=True,
ngram_range=(1,1))
X = vectorizer.fit_transform(texts)
scaler = StandardScaler() if args.scaler == 'standard' else NoScaler()
sys.stderr.write("BOW computed\n")
Y = scaler.fit_transform(numpy.array(images))
model = Ridge(solver='lsqr', alpha=args.ridge_alpha)
sys.stderr.write("Starting training\n")
model.fit(X,Y)
sys.stderr.write("Saving model\n")
cPickle.dump(model, gzip.open('model.dat.gz','w'))
cPickle.dump(vectorizer, gzip.open('vec.pkl.gz','w'))
cPickle.dump(scaler, gzip.open('vec.pkl.gz', 'w'))
def test_linear(args):
if args.random_seed is not None:
numpy.random.seed(args.random_seed)
D = Cdist()
model = cPickle.load(gzip.open('model.dat.gz'))
vectorizer = cPickle.load(gzip.open('vec.pkl.gz'))
scaler = cPickle.load(gzip.open('scaler.pkl.gz'))
real_stdout = sys.stdout
with open('/dev/null', 'w') as f:
sys.stdout = f
d = dp.getDataProvider(args.dataset)
sys.stdout = real_stdout
pairs = list(d.iterImageSentencePair(split='val'))
texts = [ pair['sentence']['raw'] for pair in pairs ]
images = list(d.iterImages(split='val')) # With pairs we'd get duplicate images!
X = vectorizer.transform(texts)
Y_pred = numpy.asarray(model.predict(X), dtype='float32') # candidates are identical to Y_pred
if args.paraphrase:
#distances = D.cosine_distance(Y_pred, Y_pred)
distances = cdist(Y_pred, Y_pred, metric='cosine')
N = 0
score = 0.0
for j,row in enumerate(distances):
imgid = pairs[j]['sentence']['imgid']
sentid = pairs[j]['sentence']['sentid']
best = numpy.argsort(row)
top4 = sum([ imgid == pairs[b]['sentence']['imgid'] for b
in best[0:5] if sentid != pairs[b]['sentence']['sentid'] ][0:4]) # exclude self
score = score + top4/4.0
N = N+1
print args.iter_predict, N, score/N
else:
Y = numpy.array([ image['feat'] for image in images], dtype='float32')
distances = D.cosine_distance(Y_pred, Y)
errors = 0
N = 0
for j,row in enumerate(distances):
imgid = pairs[j]['sentence']['imgid']
best = numpy.argsort(row)
top5 = [ images[b]['imgid'] for b in best[:5] ]
N = N+1
if imgid not in top5:
errors = errors + 1
print errors, N, errors/N
def train(args):
zero_words = cPickle.load(gzip.open("zero_shot.pkl.gz")) if args.zero_shot else set()
def maybe_zero(s, i):
overlap = set(tokenize(s)).intersection(zero_words)
if args.zero_shot and len(overlap) > 0:
return numpy.zeros(i.shape)
else:
return i
dataset = args.dataset
tok_path = args.tokenizer
model_path = args.model
d = dp.getDataProvider(dataset)
pairs = list(d.iterImageSentencePair(split='train'))
if args.shuffle:
numpy.random.shuffle(pairs)
output_size = len(pairs[0]['image']['feat'])
embedding_size = args.embedding_size if args.embedding_size is not None else args.hidden_size
tokenizer = cPickle.load(gzip.open(args.init_tokenizer)) \
if args.init_tokenizer else Tokenizer(min_df=args.word_freq_threshold, character=args.character)
sentences, images = zip(*[ (pair['sentence']['raw'], maybe_zero(pair['sentence']['raw'],pair['image']['feat']))
for pair in pairs ])
scaler = StandardScaler() if args.scaler == 'standard' else NoScaler()
images = scaler.fit_transform(images)
tokens = [ [tokenizer.encoder['PAD']] + sent + [tokenizer.encoder['END'] ]
for sent in tokenizer.fit_transform(sentences) ]
tokens_inp = [ token[:-1] for token in tokens ]
tokens_out = [ token[1:] for token in tokens ]
cPickle.dump(tokenizer, gzip.open(tok_path, 'w'))
cPickle.dump(scaler, gzip.open('scaler.pkl.gz','w'))
# Validation data
valid_pairs = list(d.iterImageSentencePair(split='val'))
valid_sents, valid_images = zip(*[ (pair['sentence']['raw'], pair['image']['feat'])
for pair in valid_pairs ])
valid_images = scaler.transform(valid_images)
valid_tokens = [ [ tokenizer.encoder['PAD'] ] + sent + [tokenizer.encoder['END'] ]
for sent in tokenizer.transform(valid_sents) ]
valid_tokens_inp = [ token[:-1] for token in valid_tokens ]
valid_tokens_out = [ token[1:] for token in valid_tokens ]
valid = (valid_tokens_inp, valid_tokens_out, valid_images)
updater = passage.updates.Adam(lr=args.rate, clipnorm=args.clipnorm)
if args.cost == 'MeanSquaredError':
z_cost = MeanSquaredError
elif args.cost == 'CosineDistance':
z_cost = CosineDistance
else:
raise ValueError("Unknown cost")
if args.hidden_type == 'gru':
Recurrent = GatedRecurrent
elif args.hidden_type == 'lstm':
Recurrent = LstmRecurrent
else:
Recurrent = GatedRecurrent
# if args.init_model is not None:
# model_init = cPickle.load(open(args.init_model))
# def values(ps):
# return [ p.get_value() for p in ps ]
# # FIXME enable this for shared only embeddings
# layers = [ Embedding(size=args.hidden_size, n_features=tokenizer.n_features,
# weights=values(model_init.layers[0].params)),
# Recurrent(seq_output=True, size=args.hidden_size, activation=args.activation,
# weights=values(model_init.layers[1].params)),
# Combined(left=Dense(size=tokenizer.n_features, activation='softmax', reshape=True,
# weights=values(model_init.layers[2].left.params)),
# right=Dense(size=output_size, activation=args.out_activation,
# weights=values(model_init.layers[2].right.params))
# ) ]
# else:
# FIXME implement proper pretraining FIXME
interpolated = True if not args.non_interpolated else False
if args.model_type in ['add', 'mult', 'matrix']:
if args.model_type == 'add':
layer0 = Direct(size=embedding_size, n_features=tokenizer.n_features, op=Add)
elif args.model_type == 'mult':
layer0 = Direct(size=embedding_size, n_features=tokenizer.n_features, op=Mult)
elif args.model_type == 'matrix':
sqrt_size = embedding_size ** 0.5
if not sqrt_size.is_integer():
raise ValueError("Sqrt of embedding_size not integral for matrix model")
layer0 = Direct(size=embedding_size, n_features=tokenizer.n_features, op=MatrixMult)
layers = [ layer0, Dense(size=output_size, activation=args.out_activation, reshape=False) ]
valid = (valid_tokens_inp, valid_images)
model = RNN(layers=layers, updater=updater, cost=z_cost,
iterator=SortedPadded(shuffle=False), verbose=1)
model.fit(tokens_inp, images, n_epochs=args.iterations, batch_size=args.batch_size, len_filter=None,
snapshot_freq=args.snapshot_freq, path=model_path, valid=valid)
elif args.model_type == 'simple':
layers = [ Embedding(size=embedding_size, n_features=tokenizer.n_features),
Recurrent(seq_output=False, size=args.hidden_size, activation=args.activation),
Dense(size=output_size, activation=args.out_activation, reshape=False)
]
valid = (valid_tokens_inp, valid_images)
model = RNN(layers=layers, updater=updater, cost=z_cost,
iterator=SortedPadded(shuffle=False), verbose=1)
model.fit(tokens_inp, images, n_epochs=args.iterations, batch_size=args.batch_size, len_filter=None,
snapshot_freq=args.snapshot_freq, path=model_path, valid=valid)
# FIXME need validation
elif args.model_type == 'deep-simple':
layers = [ Embedding(size=embedding_size, n_features=tokenizer.n_features),
Recurrent(seq_output=True, size=args.hidden_size, activation=args.activation),
Recurrent(seq_output=False, size=args.hidden_size, activation=args.activation),
Dense(size=output_size, activation=args.out_activation, reshape=False)
]
valid = (valid_tokens_inp, valid_images)
model = RNN(layers=layers, updater=updater, cost=z_cost,
iterator=SortedPadded(shuffle=False), verbose=1)
model.fit(tokens_inp, images, n_epochs=args.iterations, batch_size=args.batch_size, len_filter=None,
snapshot_freq=args.snapshot_freq, path=model_path, valid=valid)
# FIXME need validation
elif args.model_type == 'shared_all':
if args.zero_shot:
raise NotImplementedError # FIXME zero_shot not implemented
layers = [ Embedding(size=embedding_size, n_features=tokenizer.n_features),
Recurrent(seq_output=True, size=args.hidden_size, activation=args.activation),
Combined(left=Dense(size=tokenizer.n_features, activation='softmax', reshape=True),
right=Dense(size=output_size, activation=args.out_activation, reshape=False)) ]
model = ForkedRNN(layers=layers, updater=updater, cost_y=CategoricalCrossEntropySwapped,
cost_z=z_cost, alpha=args.alpha, size_y=tokenizer.n_features,
verbose=1, interpolated=interpolated)
model.fit(tokens_inp, tokens_out, images, n_epochs=args.iterations, batch_size=args.batch_size,
snapshot_freq=args.snapshot_freq, path=model_path, valid=valid)
elif args.model_type == 'shared_embeddings':
layers = [ Embedding(size=embedding_size, n_features=tokenizer.n_features),
Combined(left=Stacked([Recurrent(seq_output=True, size=args.hidden_size, activation=args.activation),
Dense(size=tokenizer.n_features, activation='softmax', reshape=True)]),
left_type='id',
right=Stacked([Recurrent(seq_output=False, size=args.hidden_size, activation=args.activation),
Dense(size=output_size, activation=args.out_activation, reshape=False)]),
right_type='id')
]
model = ForkedRNN(layers=layers, updater=updater, cost_y=CategoricalCrossEntropySwapped,
cost_z=z_cost, alpha=args.alpha, size_y=tokenizer.n_features,
verbose=1, interpolated=interpolated, zero_shot=args.zero_shot)
model.fit(tokens_inp, tokens_out, images, n_epochs=args.iterations, batch_size=args.batch_size,
snapshot_freq=args.snapshot_freq, path=model_path, valid=valid)
cPickle.dump(model, gzip.open(model_path,"w"))
def test(args):
if args.random_seed is not None:
numpy.random.seed(args.random_seed)
def scramble(words):
ixs = range(len(words))
random.shuffle(ixs)
return [ words[ix] for ix in ixs ]
testInfo = {'argv': sys.argv,
'dataset': args.dataset,
'scramble': args.scramble,
'model_type': args.model_type,
'alpha': args.alpha,
'iter_predict': args.iter_predict,
'task': 'paraphrase' if args.paraphrase else 'image',
'items': []}
D = Cdist()
dataset = args.dataset
suffix = '' if args.iter_predict is None else ".{0}".format(args.iter_predict)
model = cPickle.load(gzip.open('model.dat.gz' + suffix))
tokenizer = cPickle.load(gzip.open('tok.pkl.gz'))
scaler = cPickle.load(gzip.open('scaler.pkl.gz'))
real_stdout = sys.stdout
with open('/dev/null', 'w') as f:
sys.stdout = f
d = dp.getDataProvider(args.dataset)
sys.stdout = real_stdout
pairs = list(d.iterImageSentencePair(split='val'))
inputs = [ scramble(s) if args.scramble else s for s in tokenizer.transform([ pair['sentence']['raw'] for pair in pairs]) ]
if args.paraphrase:
candidates = tokenizer.transform([ pair['sentence']['raw'] for pair in pairs]) # No scrambling of candidates
if args.paraphrase_state == 'output_vis':
preds = model.predict(inputs)
candidates_pred = model.predict(candidates)
elif args.paraphrase_state == 'hidden_text':
preds, _ = predict_h(model, inputs)
candidates_pred, _ = predict_h(model, candidates)
elif args.paraphrase_state == 'hidden_vis' and hasattr(model.layers[1], 'left'):
_, preds = predict_h(model, inputs)
_, candidates_pred = predict_h(model, candidates)
elif args.paraphrase_state == 'hidden_vis' and not hasattr(model.layers[1], 'left'):
preds = predict_h_simple(model, inputs)
candidates_pred = predict_h_simple(model, candidates)
elif args.paraphrase_state == 'hidden_multi':
preds = numpy.hstack(predict_h(model, inputs))
candidates_pred = numpy.hstack(predict_h(model, candidates))
else:
raise ValueError("Unknown state")
distances = D.cosine_distance(preds, candidates_pred)
#distances = cdist(preds, candidates_pred, metric='cosine')
N = 0
score = 0.0
imgids = numpy.array([ pair['sentence']['imgid'] for pair in pairs ])
sentids = numpy.array([ pair['sentence']['sentid'] for pair in pairs])
for j,row in enumerate(distances):
imgid = pairs[j]['sentence']['imgid']
sentid = pairs[j]['sentence']['sentid']
best = numpy.argsort(row)
rank = numpy.where((imgids[best] == imgid) * (sentids[best] != sentid))[0][0] + 1
top4 = [ pairs[b]['sentence']['imgid'] for b
in best[0:5] if sentid != pairs[b]['sentence']['sentid'] ][0:4] # exclude self
top4sent = [ pairs[b]['sentence']['sentid'] for b in best[0:5] if sentid != pairs[b]['sentence']['sentid'] ][0:4]
score = score + sum([i == imgid for i in top4 ])/4.0
N = N+1
itemInfo = {'sentid':sentid, 'imgid': imgid, 'score': sum([i == imgid for i in top4 ])/4.0,
'rank': rank, 'topn': top4 , 'topnsentid': top4sent,
'input': tokenizer.inverse_transform([inputs[j]])[0]}
testInfo['items'].append(itemInfo)
print args.iter_predict, N, score/N
else:
preds = model.predict(inputs)
images = list(d.iterImages(split='val'))
distances = D.cosine_distance(preds, scaler.transform([image['feat'] for image in images ]))
errors = 0
N = 0
imgids = numpy.array([ img['imgid'] for img in images ])
for j,row in enumerate(distances):
imgid = pairs[j]['sentence']['imgid']
sentid = pairs[j]['sentence']['sentid']
best = numpy.argsort(row)
rank = numpy.where(imgids[best] == imgid)[0][0] + 1
top5 = [ images[b]['imgid'] for b in best[:5] ]
N = N+1
if imgid not in top5:
errors = errors + 1
itemInfo = {'sentid':sentid, 'imgid': imgid, 'score': float(imgid in top5), 'rank': rank, 'topn': top5,
'input':tokenizer.inverse_transform([inputs[j]])[0] }
testInfo['items'].append(itemInfo)
print args.iter_predict, errors, N, errors/N
testInfoPath = 'testInfo-task={0}-scramble={1}-iter_predict={2}.json.gz'.format(testInfo['task'], testInfo['scramble'], testInfo['iter_predict'])
json.dump(testInfo, gzip.open(testInfoPath,'w'))
def project_words(args):
suffix = '' if args.iter_predict is None else ".{0}".format(args.iter_predict)
model = cPickle.load(gzip.open('model.dat.gz' + suffix))
tokenizer = cPickle.load(gzip.open('tok.pkl.gz'))
scaler = cPickle.load(gzip.open('scaler.pkl.gz'))
exclude = ['PAD','END','UNK']
words, indexes = zip(*[ (w,i) for (w,i) in tokenizer.encoder.iteritems() if w not in exclude ])
inputs = [ [tokenizer.encoder['PAD'], i, tokenizer.encoder['END']] for i in indexes ] # FIXME actually for training we don't have END
preds = scaler.inverse_transform(model.predict(inputs))
proj = dict((words[i], preds[i]) for i in range(0, len(words)))
cPickle.dump(proj, gzip.open("proj.pkl.gz" + suffix, "w"))
def extract_embeddings(args):
tokenizer = cPickle.load(gzip.open('tok.pkl.gz'))
#scaler = cPickle.load(open('scaler.pkl'))
suffix = '' if args.iter_predict is None else ".{0}".format(args.iter_predict)
model = cPickle.load(gzip.open('model.dat.gz' + suffix))
embeddings = model.layers[0].params[0].get_value()
table = dict((word, embeddings[i]) for i,word in tokenizer.decoder.iteritems()
if word not in ['END','PAD','UNK'] )
cPickle.dump(table, gzip.open('embeddings.pkl.gz' + suffix, 'w'))
class Cdist():
def __init__(self):
self.U = T.matrix('U')
self.V = T.matrix('V')
self.U_norm = self.U / self.U.norm(2, axis=1).reshape((self.U.shape[0], 1))
self.V_norm = self.V / self.V.norm(2, axis=1).reshape((self.V.shape[0], 1))
self.W = T.dot(self.U_norm, self.V_norm.T)
self.cosine = theano.function([self.U, self.V], self.W)
def cosine_distance(self, A, B):
return 1 - self.cosine(A, B)
main()