def get_sequence(rangeOfSequence, mode):
#modes available (1,2,3,4)
#1 categorical input, categorical output
#2 categorical input, continous output
#3 continous input, categorical output
#4 continous input, continous output
c = rangeOfSequence
data = [[i, j] for i in range(1, c + 1) for j in range(i, c + 1)]
data2 = [[j, i] for i in range(1, c + 1) for j in range(i, c + 1)]
#the -1 serves as an 'end of sentence' indicator
target_in = [[[0]] + [[k] for k in range(d[0], d[1] + 1)] + [[-1]]
for d in data] #teacher forcing
target_out = [[[k] for k in range(d[0], d[1] + 1)] + [[-1]] + [[0]]
for d in data] #output
target_in2 = [[[0]] + [[k] for k in range(d[0], d[1] - 1, -1)] + [[-1]]
for d in data2] #for the other way around
target_out2 = [[[k] for k in range(d[0], d[1] - 1, -1)] + [[-1]] + [[0]]
for d in data2]
#combine
target_in = target_in + target_in2
target_out = target_out + target_out2
target_in = pad(target_in, padding='post')
target_out = pad(target_out, padding='post')
data = [[[i], [j]] for i in range(1, c + 1) for j in range(i, c + 1)]
data2 = [[[j], [i]] for i in range(1, c + 1) for j in range(i, c + 1)]
data = data + data2
data = np.array(data, dtype=float)
target_out = np.array(target_out, dtype=float)
target_in = np.array(target_in, dtype=float)
if mode == 4:
return data, target_in, target_out
dim1 = target_in.shape[0]
dim2 = target_in.shape[1]
d = to_categorical([data], num_classes=c + 2)
d = d.reshape(dim1, 2, c + 2)
if mode == 2:
return d, target_in, target_out
t_in = to_categorical([target_in], num_classes=c + 2)
t_in = t_in.reshape(dim1, dim2, c + 2)
t_out = to_categorical([target_out], num_classes=c + 2)
t_out = t_out.reshape(dim1, dim2, c + 2)
if mode == 3:
return data, t_in, t_out
return d, t_in, t_out #mode 1
def pad_sequences(train_X, train_y):
# Used to find the max array size.
tempMax = -1
# Keep and store the max length of a sentence as it goes through train_X.
for s in train_X:
if (len(s) > tempMax):
tempMax = len(s)
# Define the MAX_LENGTH of a sentence.
MAX_LENGTH = tempMax
# Pad train_X and train_y.
train_X = pad(train_X, MAX_LENGTH, 'int32', 'post', 'pre', 0.0)
train_y = pad(train_y, MAX_LENGTH, 'int32', 'post', 'pre', 0.0)
# Return the numpy arrays and MAX-LENGTH.
return train_X, train_y, MAX_LENGTH
def process_text(text,
to_pad=False,
max_len=None,
tok=None,
save_name=None,
num_word=None):
"""tips :
tok is tokenizer:
pass tok as none while processing trainig data
text :
should have line breaks and in each line there should be a '\t'
also it should be a String
NAME : Caption formatted
return :
tok: tokenizer,
text_dict:
"""
names = [i.split('\t')[0] for i in text.split('\n')]
descs = [i.split('\t')[1] for i in text.split('\n')]
clean_descs = clean_text(descs)
clean_descs = [i.split(' ') for i in clean_descs]
if tok == None:
tok = Tokenizer(num_words=num_word)
tok.fit_on_texts(clean_descs)
desc_seqs = tok.texts_to_sequences(clean_descs)
text_dict = dict()
if to_pad:
if max_len == None:
max_len = max([len(i) for i in clean_descs])
else:
desc_seqs = pad(desc_seqs, maxlen=max_len, padding='post', value=0)
for i in range(len(names)):
if names[i] in text_dict:
(text_dict[names[i]]).append(desc_seqs[i])
else:
text_dict[names[i]] = [desc_seqs[i]]
if not save_name == None:
with open(save_name, 'wb') as fil:
pickle.dump(obj=text_dict, file=fil)
with open(save_name + 'tokenizer', 'wb') as fil:
pickle.dump(obj=tok, file=fil)
return tok, text_dict
def __getitem__(self, i):
pg, tg = 'post', 'post'
target = [self.target[i]]
question = str(self.text[i])
quest_ids = self.tokenizer.encode(question.strip())
attention_mask_idx = len(quest_ids) - 1
if 0 not in quest_ids: quest_ids = 0 + quest_ids
quest_ids = pad([quest_ids], maxlen=MAXLEN, value=1, padding=pg, truncating=tg)
attention_mask = np.zeros(MAXLEN)
attention_mask[1:attention_mask_idx] = 1
attention_mask = attention_mask.reshape((1, -1))
if 2 not in quest_ids: quest_ids[-1], attention_mask[-1] = 2, 0
return FloatTensor(target), LongTensor(quest_ids), LongTensor(attention_mask)
def predict_sentiment(tweet):
pg, tg = 'post', 'post'
tweet_ids = tokenizer.encode(tweet.strip())
sent = {0: 'positive', 1: 'neutral', 2: 'negative'}
att_mask_idx = len(tweet_ids) - 1
if 0 not in tweet_ids: tweet_ids = 0 + tweet_ids
tweet_ids = pad([tweet_ids], maxlen=MAXLEN, value=1, padding=pg, truncating=tg)
att_mask = np.zeros(MAXLEN)
att_mask[1:att_mask_idx] = 1
att_mask = att_mask.reshape((1, -1))
if 2 not in tweet_ids: tweet_ids[-1], att_mask[-1] = 2, 0
tweet_ids, att_mask = torch.LongTensor(tweet_ids), torch.LongTensor(att_mask)
return sent[np.argmax(network.forward(tweet_ids.to(device), att_mask.to(device)).detach().cpu().numpy())]
def predict_insincerity(question):
pg, tg = 'post', 'post'
ins = {0: 'sincere', 1: 'insincere'}
quest_ids = tokenizer.encode(question.strip())
attention_mask_idx = len(quest_ids) - 1
if 0 not in quest_ids: quest_ids = 0 + quest_ids
quest_ids = pad([quest_ids], maxlen=MAXLEN, value=1, padding=pg, truncating=tg)
att_mask = np.zeros(MAXLEN)
att_mask[1:attention_mask_idx] = 1
att_mask = att_mask.reshape((1, -1))
if 2 not in quest_ids: quest_ids[-1], attention_mask[-1] = 2, 0
quest_ids, att_mask = torch.LongTensor(quest_ids), torch.LongTensor(att_mask)
output = network.forward(quest_ids.to(device), att_mask.to(device))
return ins[int(np.round(nn.Sigmoid()(output.detach().cpu()).item()))]
def __getitem__(self, i):
pg, tg = 'post', 'post'
tweet = str(self.text[i]).strip()
tweet_ids = self.tokenizer.encode(tweet)
attention_mask_idx = len(tweet_ids) - 1
if 0 not in tweet_ids: tweet_ids = 0 + tweet_ids
tweet_ids = pad([tweet_ids], maxlen=MAXLEN, value=1, padding=pg, truncating=tg)
attention_mask = np.zeros(MAXLEN)
attention_mask[1:attention_mask_idx] = 1
attention_mask = attention_mask.reshape((1, -1))
if 2 not in tweet_ids: tweet_ids[-1], attention_mask[-1] = 2, 0
sentiment = [self.sentiment_dict[self.sentiment[i]]]
sentiment = torch.FloatTensor(to_categorical(sentiment, num_classes=3))
return sentiment, torch.LongTensor(tweet_ids), torch.LongTensor(attention_mask)
def get_supporting_facts_training(X, Xq, word_idx, train_supporting_facts, trained_attention, max_hops=2):
supporting_sentences = [[0] for i in range(len(X))]
totalX = []
totalXq = []
totalY = []
enough_memories = [0 for i in range(len(X))]
allX = X[:]
selected = [[] for i in range(len(X))]
leftoversX = []
for i in range(0, max_hops):
print (bcolors.BOLD + "Entering hop " + str(i) + " ..." + bcolors.ENDC)
_, combinedXq, leftoverS, supporting_sentences, X, Xq, Y , found_supporting, leftover= supporting_facts_inc(None, Xq, word_idx,
train_supporting_facts,
supporting_sentences,
trained_attention,
enough_memories, allX, selected,
)
totalX.extend(X)
totalY.extend(Y)
totalXq.extend(Xq)
leftoversX.extend(leftover)
Xq = combinedXq
train_supporting_facts = leftoverS
print (bcolors.BOLD + "Found supporting facts " + str(len(selected)) + " ..." + bcolors.ENDC)
print(np.sum(enough_memories), "break")
if (np.sum(enough_memories) == len(enough_memories) or found_supporting == 0 ):
print (bcolors.BOLD + "breaking at hop " + str(i) + " ..." )
break
X = pad(totalX, maxlen=max(map(len, totalX)))
Xq = pad(totalXq, maxlen=max(map(len, totalXq)))
Y = pad(totalY, maxlen=max(map(len, totalY)))
leftoversX = pad(leftoversX, maxlen=max(map(len, leftoversX)))
# print (len(supporting_sentences))
# for sentence in supporting_sentences:
# print(sentence)
supporting_sentences = pad(supporting_sentences, maxlen=max(map(len, supporting_sentences)))
#X = reverse(X, word_idx)
#Xq = reverse(Xq, word_idx)
# import collections
# y=collections.Counter([tuple(list(x[0]) + list(x[1])) for x in zip(X,Xq)])
# for yi,v in y.items():
# if(v > 1):
# print yi, v
print('X.shape = {}'.format(X.shape))
print('Xq.shape = {}'.format(Xq.shape))
print('Y.shape = {}'.format(Y.shape))
print('leftover.shape = {}'.format(leftoversX.shape))
print('supporting_sentences.shape = {}'.format(supporting_sentences.shape))
print(bcolors.ENDC)
return X, Xq, Y, supporting_sentences, leftoversX
list_sent_train = X_train["comment_text"] list_sent_test = X_test["comment_text"] sent_text = t_data['comment_text'] l_toxic = t_label['toxic'] l_sever_t = t_label['severe_toxic'] l_obscene = t_label['obscene'] l_idh = t_label['identity_hate'] l_in = t_label['insult'] l_th = t_label['threat'] max_features = 20000 tokenizer = Tokenizer(num_words=max_features, char_level=True) tokenizer.fit_on_texts(list(list_sent_train)) list_token_train = tokenizer.texts_to_sequences(list_sent_train) list_sent_test = tokenizer.texts_to_sequences(list_sent_test) maxlength = 500 X_t = pad(list_token_train, maxlen=maxlength) X_te = pad(list_sent_test, maxlen=maxlength) inp = Input(shape=(maxlength, )) embedding_size = 240 x = Embedding(len(tokenizer.word_index) + 1, embedding_size)(inp) x = Conv1D(filters=100, kernel_size=4, padding='same', activation='relu')(x) x = MaxPooling1D(pool_size=4)(x) #x = Bidirectional(GRU(60, return_sequences=True,name='lstm_layer',dropout=0.2,recurrent_dropout=0.2))(x) x = GlobalMaxPool1D()(x) x = Dense(50, activation="relu")(x)
def coco_generator(
mappings,
captions, #used for one hotting the target should be vocabularies size
dict_size,
max_len,
image_batch_szie=1,
path_to_pkl_files=".",
pkl_file_extension='.pkl',
epochs=1):
for _ in range(epochs):
acc_features = np.array([[[0 for i in range(4096)]]])
acc_caption = np.array([[[0 for i in range(max_len)]]])
acc_target = np.array([[[0 for i in range(dict_size + 1)]]])
counter = 0
for pkl_file, image_subset in mappings.items():
with open(path_to_pkl_files + '/' + pkl_file + pkl_file_extension,
'rb') as file:
feature_dict = pickle.load(file)
for image_name in image_subset:
image_name = image_name.split('/')[-1]
temp = []
caption = captions[image_name]
[[temp.append(line[:i]) for i in range(1, len(line))]
for line in caption]
caption = np.array(
pad(temp, maxlen=max_len, padding='post', value=0))
temp = []
[
np.array([temp.append(i) for i in line[1:]])
for line in captions[image_name]
]
temp = np.asarray(temp)
target = np.array(
[np.array(one_hot(i, dict_size)) for i in temp])
features = feature_dict[image_name]
features = (features.repeat(len(caption),
axis=0)).reshape(-1, 1, 4096)
# caption = caption.reshape( -1, 1 , max_len)
target = target.reshape(-1, 1, dict_size + 1)
counter += 1
if image_batch_szie > 1:
acc_features = np.append(acc_features, features, axis=0)
acc_caption = np.append(acc_caption, caption, axis=0)
acc_target = np.append(acc_target, target, axis=0)
if counter == image_batch_szie:
acc_features = acc_features.reshape(-1, 1, 4096)
# acc_caption = acc_caption.reshape(-1,1, max_len)
acc_target = acc_target.reshape(-1, 1, dict_size + 1)
yield [acc_features[1:],
acc_caption[1:]], acc_target[1:]
acc_features = np.array([[[0 for i in range(4096)]]])
acc_caption = np.array([[[0 for i in range(max_len)]]])
acc_target = np.array(
[[[0 for i in range(dict_size + 1)]]])
counter = 0
continue
else:
yield [[features, caption], target]
if len(acc_caption) > 1:
acc_features = acc_features.reshape(-1, 1, 4096)
# acc_caption = acc_caption.reshape(-1,1, max_len)
acc_target = acc_target.reshape(-1, 1, dict_size + 1)
yield [acc_features[1:], acc_caption[1:]], acc_target[1:]
