def perplexity(self, ngrams, word_dict, Im=None, context=5):
"""
Compute the perplexity of ngrams from net
"""
ll = 0
N = 0
x = T.matrix('x', dtype='int32')
im = T.matrix('im')
forward_T = theano.function([x, im], self.forward(x, im))
for i, ng in enumerate(ngrams):
instances = lm_tools.model_inputs([ng], word_dict)
if Im != None:
ll += self.compute_ll(instances.astype(np.int32), np.tile(Im[i], (len(ng), 1)).astype(theano.config.floatX), forward_T)
else:
ll += self.compute_ll(instances)
N += len(instances)
return np.power(2, (-1.0 / N) * ll)
def perplexity(self, ngrams, word_dict, Im=None, context=5):
"""
Compute the perplexity of ngrams from net
"""
ll = 0
N = 0
x = T.matrix('x', dtype='int32')
im = T.matrix('im')
forward_T = theano.function([x, im], self.forward(x, im))
for i, ng in enumerate(ngrams):
instances = lm_tools.model_inputs([ng], word_dict)
if Im != None:
ll += self.compute_ll(
instances.astype(np.int32),
np.tile(Im[i], (len(ng), 1)).astype(theano.config.floatX),
forward_T)
else:
ll += self.compute_ll(instances)
N += len(instances)
return np.power(2, (-1.0 / N) * ll)
def process(context=5):
"""
Main process function
"""
# Load images
print 'Loading images...'
(trainIM, devIM, testIM) = load_features_npy()
# Load sentences
print 'Loading sentences...'
d = load_sentences()
# Load image ids
print 'Loading image ids...'
(dx_train, dx_dev) = image_ids()
# Load splits
print 'Loading splits...'
(train_sp, dev_sp, test_sp) = load_splits()
# Load captions
print 'Loading captions...'
train = construct_captions(d, train_sp)
dev = construct_captions(d, dev_sp)
test = construct_captions(d, test_sp)
# Tokenize
(train_tokens, topwords) = tokenize(train, context=context)
dev_tokens = tokenize(dev, context=context, topwords=topwords)[0]
test_tokens = tokenize(test, context=context, topwords=topwords)[0]
# Index words and create vocabulary
print 'Creating vocabulary...'
(word_dict, index_dict) = index_words(train_tokens + dev_tokens)
# Compute n-grams
print 'Computing n-grams...'
train_ngrams = lm_tools.get_ngrams(train_tokens, context=context)
dev_ngrams = lm_tools.get_ngrams(dev_tokens, context=context)
test_ngrams = lm_tools.get_ngrams(test_tokens, context=context)
# Compute sparse label matrix
print 'Computing labels...'
train_labels = compute_labels(train_ngrams, word_dict, context=context)
dev_labels = compute_labels(dev_ngrams, word_dict, context=context)
# Compute model instances
print 'Computing model instances...'
(train_instances, train_index) = lm_tools.model_inputs(train_ngrams,
word_dict,
context=context,
include_last=False,
include_index=True)
(dev_instances, dev_index) = lm_tools.model_inputs(dev_ngrams,
word_dict,
context=context,
include_last=False,
include_index=True)
(test_instances, test_index) = lm_tools.model_inputs(test_ngrams,
word_dict,
context=context,
include_last=False,
include_index=True)
# Save everything into dictionaries
print 'Packing up...'
z = {}
z['text'] = train
z['tokens'] = train_tokens
z['word_dict'] = word_dict
z['index_dict'] = index_dict
z['ngrams'] = train_ngrams
z['labels'] = train_labels
z['instances'] = train_instances
z['IM'] = trainIM
z['index'] = train_index
z['context'] = context
zd = {}
zd['text'] = dev
zd['tokens'] = dev_tokens
zd['ngrams'] = dev_ngrams
zd['labels'] = dev_labels
zd['instances'] = dev_instances
zd['IM'] = devIM
zd['index'] = dev_index
zd['context'] = context
zt = {}
zt['text'] = test
zt['tokens'] = test_tokens
zt['ngrams'] = test_ngrams
zt['instances'] = test_instances
zt['IM'] = testIM
zt['index'] = test_index
zt['context'] = context
return (z, zd, zt)
def process():
"""
Specify the following:
"""
##################################
train_captions = os.getcwd() + '/mnlm/engine/iaprtc12/train_captions.txt'
train_images = os.getcwd() + '/mnlm/engine/iaprtc12/train_hidden7.txt'
test_captions = os.getcwd() + '/mnlm/engine/iaprtc12/test_captions.txt'
test_images = os.getcwd() + '/mnlm/engine/iaprtc12/test_hidden7.txt'
context = 5
##################################
# Load captions
print 'Loading captions...'
train = load_captions(train_captions)
test = load_captions(test_captions)
# Tokenize the data
print 'Tokenizing...'
train_tokens = tokenize(train, context=context)
test_tokens = tokenize(test, context=context)
# Index words and create vocabulary
print 'Creating vocabulary...'
(word_dict, index_dict) = index_words(train_tokens) # build a map between words and indexes
# Compute n-grams
print 'Computing n-grams...'
train_ngrams = lm_tools.get_ngrams(train_tokens, context=context) # compute all tuple of len context+1, group in caption
test_ngrams = lm_tools.get_ngrams(test_tokens, context=context)
# Compute sparse label matrix
print 'Computing labels...'
labels = compute_labels(train_ngrams, word_dict, context=context)
# Compute model instances
print 'Computing model instances...'
(train_instances, train_index) = lm_tools.model_inputs(train_ngrams, word_dict,
context=context, include_last=False, include_index=True)
(test_instances, test_index) = lm_tools.model_inputs(test_ngrams, word_dict,
context=context, include_last=False, include_index=True)
# Load image features
print 'Loading image features...'
trainIM = load_convfeatures(train_images)
testIM = load_convfeatures(test_images)
# Save everything into dictionaries
print 'Packing up...'
z = {}
z['text'] = train
z['tokens'] = train_tokens
z['word_dict'] = word_dict
z['index_dict'] = index_dict
z['ngrams'] = train_ngrams
z['labels'] = labels
z['instances'] = train_instances
z['IM'] = trainIM
z['index'] = train_index
z['context'] = context
zt = {}
zt['text'] = test
zt['tokens'] = test_tokens
zt['ngrams'] = test_ngrams
zt['instances'] = test_instances
zt['IM'] = testIM
zt['index'] = test_index
zt['context'] = context
return (z, zt)
def process():
"""
Specify the following:
"""
##################################
train_captions = 'iaprtc12/train_captions.txt'
train_images = 'iaprtc12/train_hidden7.txt'
test_captions = 'iaprtc12/test_captions.txt'
test_images = 'iaprtc12/test_hidden7.txt'
context = 5
##################################
# Load captions
print 'Loading captions...'
train = load_captions(train_captions)
test = load_captions(test_captions)
# Tokenize the data
print 'Tokenizing...'
train_tokens = tokenize(train, context=context)
test_tokens = tokenize(test, context=context)
# Index words and create vocabulary
print 'Creating vocabulary...'
(word_dict, index_dict) = index_words(train_tokens)
# Compute n-grams
print 'Computing n-grams...'
train_ngrams = lm_tools.get_ngrams(train_tokens, context=context)
test_ngrams = lm_tools.get_ngrams(test_tokens, context=context)
# Compute sparse label matrix
print 'Computing labels...'
labels = compute_labels(train_ngrams, word_dict, context=context)
# Compute model instances
print 'Computing model instances...'
(train_instances, train_index) = lm_tools.model_inputs(train_ngrams,
word_dict,
context=context,
include_last=False,
include_index=True)
(test_instances, test_index) = lm_tools.model_inputs(test_ngrams,
word_dict,
context=context,
include_last=False,
include_index=True)
# Load image features
print 'Loading image features...'
trainIM = load_convfeatures(train_images)
testIM = load_convfeatures(test_images)
# Save everything into dictionaries
print 'Packing up...'
z = {}
z['text'] = train
z['tokens'] = train_tokens
z['word_dict'] = word_dict
z['index_dict'] = index_dict
z['ngrams'] = train_ngrams
z['labels'] = labels
z['instances'] = train_instances
z['IM'] = trainIM
z['index'] = train_index
z['context'] = context
zt = {}
zt['text'] = test
zt['tokens'] = test_tokens
zt['ngrams'] = test_ngrams
zt['instances'] = test_instances
zt['IM'] = testIM
zt['index'] = test_index
zt['context'] = context
return (z, zt)
