def tokenize_cleaned_tweets(tweets: Dict[str, List[str]],
create_tokenizer=False):
if create_tokenizer is False:
max_tweet_word_count = load_pickle('max_tweet_word_count.pickle')
tok = load_tokenizer()
ret = {}
for hashtag, tweets in tweets.items():
padded_x, _ = encode_embed_docs(tweets,
tok,
max_tweets=max_tweet_word_count)
ret[hashtag] = padded_x
dest = 'hashtag_' + hashtag
numpy.save(dest, padded_x)
return ret
else:
all_tweets = tweets.values()
flat_list = []
for sublist in all_tweets:
for item in sublist:
flat_list.append(item)
tok = tokenize(flat_list, verbose=False)
temp_padded_x, _ = encode_embed_docs(flat_list, tok)
max_tweet_word_count = len(temp_padded_x[0])
save_pickle(max_tweet_word_count, 'max_tweet_word_count.pickle')
save_tokenizer(tok)
return tokenize_cleaned_tweets(tweets)
def load_training_sentiment_data_small():
t = load_tokenizer('learn')
y = load_pickle('tokenized/learn/small_y.pickle')
padded_x = load_pickle('tokenized/learn/small_padded_x.pickle')
unpadded_x = load_pickle('tokenized/learn/small_unpadded_x.pickle')
max_tweet_word_count = load_pickle(
'tokenized/learn/max_tweet_word_count.pickle')
vocab_size = load_pickle('tokenized/learn/vocab_size.pickle')
return t, y, padded_x, unpadded_x, max_tweet_word_count, vocab_size
def load_training_sentiment_data():
t = load_tokenizer()
y = load_csv('tokenized/learn/lables.csv')
padded_x = load_pickle('tokenized/learn/padded_x.pickle')
unpadded_x = load_pickle('tokenized/learn/unpadded_x.pickle')
max_tweet_word_count = load_pickle(
'tokenized/learn/max_tweet_word_count.pickle')
vocab_size = t.num_words
return t, y, padded_x, unpadded_x, max_tweet_word_count, vocab_size
def learn_embedding_word(x, y):
padded_x = x
tokenizer = load_tokenizer()
max_tweet_word_count = load_pickle('max_tweet_word_count.pickle')
print('learning word...')
e = EmbeddingWord(tokenizer, padded_x, None, max_tweet_word_count,
tokenizer.num_words, y)
e.create_embedding()
weights = e.get_weights()
save_model_mat(weights, 'embedding_word')
return weights
def learn_embedding_skip_gram(x, y, texts):
padded_x = x
tokenizer = load_tokenizer()
max_tweet_word_count = load_pickle('max_tweet_word_count.pickle')
print('learning skip_gram...')
e = EmbeddingGensimSkipGram(tokenizer, padded_x, None,
max_tweet_word_count, tokenizer.num_words, y,
texts)
e.create_embedding()
weights = e.get_weights()
save_model_mat(weights, 'embedding_skip_gram')
return weights
from sklearn.model_selection import train_test_split
from bix.twitter.base.utils import load_csv, encode_embed_docs, save_pickle, load_pickle, save_csv
from bix.twitter.learn.tokenizer.tokenizer_utils import load_tokenizer
if __name__ == '__main__':
print('loading saved state')
tokenizer = load_tokenizer('learn')
x = load_csv('learn/tweets.csv')
y = load_csv('learn/lables.csv')
max_tweet_word_count = load_pickle(
'tokenized/learn/max_tweet_word_count.pickle')
print('reducing learning data')
x_learn, _, y_learn, _ = train_test_split(
x, y, test_size=0.995, random_state=4) # 16k are more than enough
print('encoding data')
padded_x, unpadded_x = encode_embed_docs(x_learn, tokenizer,
max_tweet_word_count)
print('saving')
save_pickle(padded_x, 'tokenized/learn/small_padded_x.pickle')
save_pickle(unpadded_x, 'tokenized/learn/small_unpadded_x.pickle')
save_pickle(y_learn, 'tokenized/learn/small_y.pickle')
save_csv('learn/tweets_learn.csv', x_learn)
def train_model_convolutional(x, y, embedding_mats):
# print(device_lib.list_local_devices())
# exit(0)
padded_x = x
max_tweet_word_count = load_pickle('max_tweet_word_count.pickle')
tok = load_tokenizer()
vocab_size = tok.num_words
#if 'test_all_data' in args:
# padded_x = load_pickle('tokenized/learn/padded_x.pickle')
# y = load_csv('learn/lables.csv')
enc_y = np_utils.to_categorical(y)
x_train, x_test, y_train, y_test = train_test_split(padded_x,
enc_y,
test_size=0.2,
random_state=5)
#model_mat_word = load_model_mat('embedding_word')
#print(f'model_mat_word.shape: {model_mat_word[0].shape}')
#model_mat_glove = load_model_mat('embedding_glove')
#print(f'model_mat_glove.shape: {model_mat_glove[0].shape}')
#model_mat_skip_gram = load_model_mat('embedding_skip_gram')
#print(f'model_mat_skip_gram.shape: {model_mat_skip_gram[0].shape}')
# model_mat_skip_gram = load_model_mat('embedding_skip_gram')
# vocab_size = len(tok.word_index) + 1
# model = Sequential()
# model.add(Embedding(vocab_size, model_mat_word[0].shape[1], input_length=max_tweet_word_count,
# weights=model_mat_word))
# model.add(Flatten())
# model.add(Dense(1, activation='sigmoid'))
# word
input = Input(shape=(max_tweet_word_count, ))
xs = []
for mat in embedding_mats:
xs.append(
Embedding(vocab_size,
mat[0].shape[1],
input_length=max_tweet_word_count,
weights=mat,
trainable=False)(input))
combined = concatenate(xs)
z = Conv1D(100, 5, activation='relu')(combined)
z = Conv1D(100, 5, activation='relu')(z)
z = MaxPooling1D()(z)
z = Conv1D(160, 5, activation='relu')(z)
z = Conv1D(160, 5, activation='relu')(z)
z = GlobalMaxPooling1D()(z)
z = Dropout(0.5)(z)
#f = Flatten()(z)
# conv
#polling
#flatten
#dense
# (evntl. residual conv)
#z = Dense(10, activation="relu")(z)
z = Dense(len(y_test[0]), activation="softmax")(z)
# ca 20kk total params
model = Model(inputs=[input], outputs=z)
# run_opts = tensorflow.RunOptions(report_tensor_allocations_upon_oom=True)
# compile the model
model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['acc']) # , options=run_opts)
# experiment optimizer (adam vs rmsprop)
# expeniment activation function (liki_relu, elu)
# summarize the model
print(model.summary())
# fit the model
es = EarlyStopping(monitor='val_loss')
model.fit([x_train],
y_train,
epochs=50,
verbose=1,
batch_size=8000,
validation_split=0.1,
callbacks=[es])
# todo: use return value
# evaluate the model
loss, accuracy = model.evaluate([x_test],
y_test,
verbose=1,
batch_size=8000)
print('Accuracy: %f' % (accuracy * 100))
# small: Accuracy: 93.041664
# all data - Accuracy: 79.458750
#3 embedding Layers:
model.save('sentiment_conv_ep100.h5')
print('finished')
