def encode_article(self):
"""
Method to embed sentences into sentence vectors.
:return: List of encoded sentences
"""
sentences_to_encode = []
encoded_article = [None]
sentence_list = [0]
sent_count = 0
for sentence in self.article:
sent_count += len(sentence)
sentence_list.append(sent_count)
for sent in sentence:
sentences_to_encode.append(sent) # Split into sentences
print('Loading models...'),
model = skipthoughts.LoadModel().load_model()
encoder = skipthoughts.Encoder(model)
print('Encoding sentences...'),
encoded_sentences = encoder.encode(sentences_to_encode, verbose=False)
print('Done')
for i in range(len(self.article)):
begin = sentence_list[i]
end = sentence_list[i + 1]
encoded_article[i] = encoded_sentences[begin:end]
return encoded_article
def load_inputs(path, limit=-1):
'''
Returns captions,resnet_embeddings,vectors
'''
# load captions and resnet embeddings
split = pickle.load(open(path, "rb"))
ret = None
if limit != -1:
split = split[:limit]
if len(split[0]) == 2:
# encode captions via skipthought, if there are not emebddings already include
model = skipthoughts.load_model()
encoder = skipthoughts.Encoder(model)
captions, resnet_embeddings = [s[0] for s in split
], [s[1].flatten() for s in split]
vectors = encoder.encode(captions)
ret = captions, resnet_embeddings, vectors
elif len(split[0]) == 3:
captions, resnet_embeddings, vectors = [s[0] for s in split], [
s[1].flatten() for s in split
], [s[2].flatten() for s in split]
vectors = np.array(vectors)
ret = captions, resnet_embeddings, vectors
else:
raise "Input pickled vectors should be a list of two tuples (caption,resnet) or\
three-tuples (caption,resnet,embedding)"
return ret
def skipthought_encode(sentences):
"""
Obtains sentence embeddings for each sentence in the text
"""
# print('Loading pre-trained models...')
model = skipthoughts.load_model()
encoder = skipthoughts.Encoder(model)
# print('Encoding sentences...')
encoded = encoder.encode(sentences)
return encoded
def load_input(path):
input_images = load_input_images(path)
input_sentences, labels = load_input_sentences()
inputs = list(zip(input_images, input_sentences, labels))
random.shuffle(inputs)
input_images, input_sentences, labels = zip(*inputs)
model = skipthoughts.load_model()
encoder = skipthoughts.Encoder(model)
input_sentences = encoder.encode(input_sentences)
return input_images, input_sentences, labels
def transform_sentences(sentences):
"""
Builds sentence embeddings using the Skip-thoughts model.
:param _df: Input data frame with column of sentences.
:return: Torch matrix of embeddings, size 1024.
"""
model = skipthoughts.load_model()
encoder = skipthoughts.Encoder(model)
#sentences = _df['sentText'].tolist()
_sent_embs = encoder.encode(sentences, verbose=True)
_sent_tensors = [torch.from_numpy(j) for j in _sent_embs]
return torch.stack(_sent_tensors)
def encodeSentences(reviews):
enc_reviews = [None] * len(reviews)
cum_sum_sentences = [0]
sent_count = 0
for review in reviews:
sent_count += len(review)
cum_sum_sentences.append(sent_count)
print('Loading pretrained model...')
model = skipthoughts.load_model()
encoder = skipthoughts.Encoder(model)
print('Encoding sentences...')
enc_sentences = encoder.encode(reviews, verbose=False)
return enc_sentences
def get_pretrained_encodings(pretrained=False):
'''
Get encodings using the pre-trained models.
'''
word_set = set()
dict_f = open(os.path.join(DATA_PATH, 'word2vec/dict.txt'), 'r')
for line in dict_f:
word_set.add(line.strip())
dict_f.close()
# Getting the data.
with open(os.path.join(DATA_PATH, 'amazon_food/train_data.pkl'), 'r') as f:
train_data = cPickle.load(f)
train_preprocessed = preprocess(train_data[0], word_set)
with open(os.path.join(DATA_PATH, 'amazon_food/test_data.pkl'), 'r') as f:
test_data = cPickle.load(f)
test_preprocessed = preprocess(test_data[0], word_set)
if pretrained:
model = skipthoughts.load_model()
encoder = skipthoughts.Encoder(model)
test_save_path = os.path.join(
DATA_PATH,
'amazon_food/skip_thought_vecs/skip_thought_vecs_test_pretrained.npy'
)
train_save_path = os.path.join(
DATA_PATH,
'amazon_food/skip_thought_vecs/skip_thought_vecs_train_pretrained.npy'
)
print('Encoding training vectors')
train_vectors = encoder.encode(train_preprocessed)
print('Encoding test vectors')
test_vectors = encoder.encode(test_preprocessed)
else:
model = tools.load_model(None)
test_save_path = os.path.join(
DATA_PATH,
'amazon_food/skip_thought_vecs/skip_thought_vecs_test_bi.npy')
train_save_path = os.path.join(
DATA_PATH,
'amazon_food/skip_thought_vecs/skip_thought_vecs_train_bi.npy')
print('Encoding training vectors')
train_vectors = tools.encode(model, train_preprocessed)
print('Encoding test vectors')
test_vectors = tools.encode(model, test_preprocessed)
np.save(train_save_path, train_vectors)
np.save(test_save_path, test_vectors)
def test_data(text_path):
encoder = skipthoughts.Encoder(skipthoughts.load_model())
testing_text = open(text_path).read().strip().split('\n')
ids = []
texts = []
for line in testing_text:
line = line.split(',')
id = line[0]
text = line[1]
for i in range(5):
ids.append((id, i))
texts.append(text)
captions = encoder.encode(texts)
return ids, captions
def generate_and_save_word_embeddings_for_sentences_text(
input_file,
embeddings_output_path,
embeddings_id_output_file,
for_testing=False,
is_labeled=True):
""" Generate the embeddings and save them in a different file
Parameters:
-----------
input_file: string
path to the file to load the data from
embeddings_output_file_path: string
path to the directory to save the embeddings in
embeddings_id_output_file: string
path to the file to save the embeddings id in
"""
list_of_stories = load_and_process_text_data(input_file, for_testing,
is_labeled)
model = skipthoughts.load_model()
encoder = skipthoughts.Encoder(model)
fout_id = open(embeddings_id_output_file, "wb")
for story in list_of_stories:
print(story)
if not for_testing:
embeddings = encoder.encode(
story.get_story_with_right_ending_as_list())
else:
embeddings = encoder.encode(
story.get_story_with_both_endings_as_list())
output_file = open(embeddings_output_path + story.id, "wb")
for embed in embeddings:
b = bytes()
b = b.join((struct.pack('f', e) for e in embed))
output_file.write(b)
print(story.id)
output_file.close()
fout_id.write(story.id)
fout_id.write("\n")
fout_id.close()
def main():
"""
Executes the entire pipeline of the code
:return: void
"""
gt = getGroundTruth()
model_sum, gt_sum = [], []
#print("Fetching encoder model...", end=" ")
#enc_model = SentenceTransformer('bert-base-nli-mean-tokens')
model = skipthoughts.load_model()
encoder = skipthoughts.Encoder(model)
#print("Done")
for full_text, catch_phrases in gt:
# Embed each sentence
#sentence_embeddings = enc_model.encode(full_text)
encoded = encoder.encode(full_text)
# Cluster each embedding
cluster_n = 11
#clusters = cluster(sentence_embeddings, minimum_samples=cluster_n)
clusters = cluster(encoded, minimum_samples=cluster_n)
centroids = []
for idx in range(cluster_n):
centroid_id = np.where(clusters.labels_ == idx)[0]
centroids.append(np.mean(centroid_id))
# Select representative cluster
closest, _ = pairwise_distances_argmin_min(clusters.cluster_centers_,
encoded)
ordering = sorted(range(cluster_n), key=lambda k: centroids[k])
print(ordering)
summary = ' '.join([full_text[closest[idx]]
for idx in ordering]).replace('\n', ' ')
model_sum.append(summary)
print([(full_text[closest[idx]], closest[idx]) for idx in ordering])
print(summary)
print(len(catch_phrases))
print(".".join(catch_phrases))
gt_sum.append(".".join(catch_phrases))
break
def skipthought_encode(emails):
"""
Obtains sentence embeddings for each sentence in the emails
"""
enc_emails = [None] * len(emails)
cum_sum_sentences = [0]
sent_count = 0
for email in emails:
sent_count += len(email)
cum_sum_sentences.append(sent_count)
all_sentences = [sent for email in emails for sent in email]
print('Loading pre-trained models...')
model = skipthoughts.load_model()
encoder = skipthoughts.Encoder(model)
print('Encoding sentences...')
enc_sentences = encoder.encode(all_sentences, verbose=False)
for i in range(len(emails)):
begin = cum_sum_sentences[i]
end = cum_sum_sentences[i + 1]
enc_emails[i] = enc_sentences[begin:end]
return enc_emails
def summarize(self, sentences):
# Getting sentence embeddings
encoder = skipthoughts.Encoder(self.model)
vectors = encoder.encode(sentences, verbose=False)
print('Sentences have been encoded...')
# Retrieving clusters
n_clusters = int(np.ceil(len(vectors)**0.5))
# n_clusters = int(np.ceil(SUMMARY_LENGTH))
kmeans = KMeans(n_clusters=n_clusters, random_state=0)
# print pca embeddings
self.print_embeddings(vectors)
kmeans.fit(vectors)
avg = []
for j in range(n_clusters):
idx = np.where(kmeans.labels_ == j)[0]
avg.append(np.mean(idx))
# Choosing sentences closest to cluster centers
closest, _ = pairwise_distances_argmin_min(kmeans.cluster_centers_,
vectors)
ordering = sorted(range(n_clusters), key=lambda k: avg[k])
# Returning summary
summary = ' '.join([sentences[closest[idx]] for idx in ordering])
return summary
def caption2Vectors(video_filename, video_captions):
captions = []
video_ctr = 0
#Prepare processed lsit of captions.
for caption in video_captions:
caption = caption.decode('utf-8').strip()
caption = caption.strip()
caption = caption.rstrip('\n')
captions.append(caption)
print "Total number of captions are: "
print len(captions)
print "Initializing Skipthoughts vectorizer."
model = skipthoughts.load_model()
encoder = skipthoughts.Encoder(model)
#first 2400 -> uni-skip model, last 2400 ->bi-skip model
print "Starting Vectorizer conversion."
vectors = []
print "Skip-thought vector size is :"
vectors = encoder.encode(captions)
print vectors.shape
#save captions as numpy files using corresponding filenames.
for caption in vectors:
file_name_video = video_filename[video_ctr]
output_file = output_folder + str(file_name_video) + ".npy"
with open(output_file, 'w') as f2:
np.save(f2, caption)
print str(file_name_video) + " extracted"
video_ctr += 1
import pickle as pkl
if config.sim == 'word_max' or config.sim == 'combine':
emb_word, emb_id = pkl.load(open(config.emb_path))
import sys
sys.path.insert(0, config.skipthoughts_path)
sys.path.insert(0, config.emb_path)
sys.path.insert(0, '../utils/dict_emb')
from dict_use import dict_use
dict_use = dict_use(config.dict_path)
sen2id = dict_use.sen2id
id2sen = dict_use.id2sen
if config.sim == 'skipthoughts' or config.sim == 'combine':
import skipthoughts
skip_model = skipthoughts.load_model()
skip_encoder = skipthoughts.Encoder(skip_model)
if config.sim == 'word_max' or config.sim == 'combine':
#id2freq=pkl.load(open('./data/id2freq.pkl'))
pass
def normalize(x, e=0.05):
tem = copy(x)
if max(tem) == 0:
tem += e
return tem / tem.sum()
def reverse_seq(input, sequence_length, target):
batch_size = input.shape[0]
num_steps = input.shape[1]
solution = sentence.split(' ')[-1]
solutions.append(int(solution))
sentence = readline()
return sentences, solutions
input_images = load_input_images('./dataset/input')
input_sentences, labels = load_input_sentences()
inputs = list(zip(input_images, input_sentences, labels))
random.shuffle(inputs)
input_images, input_sentences, labels = zip(*inputs)
model = skipthoughts.load_model()
encoder = skipthoughts.Encoder(model)
input_sentences = encoder.encode(input_sentences)
def load_input(path):
input_images = load_input_images(path)
input_sentences, labels = load_input_sentences()
inputs = list(zip(input_images, input_sentences, labels))
random.shuffle(inputs)
input_images, input_sentences, labels = zip(*inputs)
model = skipthoughts.load_model()
encoder = skipthoughts.Encoder(model)
input_sentences = encoder.encode(input_sentences)
def __init__(self):
self.encoder = skipthoughts.Encoder(skipthoughts.load_model())
def __init__(self, **kwargs):
self.model = skipthoughts.load_model()
self.encoder = skipthoughts.Encoder(self.model)
return return_arr
@np.vectorize
def decode(x):
return x.decode('utf8')
if __name__ == '__main__':
from argparse import ArgumentParser, ArgumentDefaultsHelpFormatter
parser = ArgumentParser(
description='Extract skipthought vectors',
formatter_class=ArgumentDefaultsHelpFormatter)
m = skipthoughts.Encoder(skipthoughts.load_model())
args = parser.parse_args()
vuamc = pd.read_csv('./data/vuamc.csv',
keep_default_na=False)
vuamc.min_context = decode(vuamc.min_context)
unique_ctx = vuamc.min_context.unique()
ctx_embs = infer_vector_skipthought(m, unique_ctx)
ctx_to_idx = {ctx: i for i, ctx in enumerate(unique_ctx)}
v_embs = np.zeros((vuamc.shape[0], HIDDEN_DIM_SIZE), dtype=np.float32)
s_embs = np.zeros((vuamc.shape[0], HIDDEN_DIM_SIZE), dtype=np.float32)
o_embs = np.zeros((vuamc.shape[0], HIDDEN_DIM_SIZE), dtype=np.float32)