def my_initializer(shape=None,
dtype=tf.float32,
partition_info=None,
glove_activation=False):
assert dtype is tf.float32
embedding_matrix = preprocessing.load_glove_embeddings(
word_index, model)
return embedding_matrix
def init_weights(self):
glove_embeddings = preprocessing.load_glove_embeddings()
embedding_weights = torch.FloatTensor(self.vocab_size, self.input_size)
init.uniform(embedding_weights, a=-0.25, b=0.25)
for k, v in glove_embeddings.items():
embedding_weights[k] = torch.FloatTensor(v)
embedding_weights[0] = torch.FloatTensor([0] * self.input_size)
del self.embedding.weight
self.embedding.weight = nn.Parameter(embedding_weights)
def train_embedding_to_sentiment():
_, lexicon = load_sentiment_lexicon()
if not os.path.exists('data/lexicon_words_indices.npy'):
lexicon_pandas = pd.read_csv('lexicons/ts.lex.txt',
sep=" ",
usecols=[0],
header=None)
word2index, embedding_matrix = load_glove_embeddings(
'data/glove.twitter.27B.100d.txt',
embedding_dim=100,
vocab_size=1193514)
lexicon_pandas.columns = ['word']
X = np.array([0] + [
word2index[word] if word in word2index else 0
for word in lexicon_pandas.word.values.tolist()
])
np.save('data/lexicon_words_indices', X)
else:
X = np.load('data/lexicon_words_indices.npy')
embedding_matrix = np.load('data/embeddings_matrix2.npy')
mask = [True] + (X[1:] > 0).tolist()
y = lexicon[mask]
X = X[mask]
X = X.reshape(-1, 1)
model = embedding_to_sentiment_model(X[0].shape, embedding_matrix, 1,
y.shape[1])
opt = k.optimizers.Adam(lr=0.001, amsgrad=True)
model.compile(optimizer=opt,
loss='mean_squared_error',
metrics=['mse', 'mae'])
# checkpoint
checkpoint = k.callbacks.ModelCheckpoint(
'models/embedding_sentiment_model-{epoch:02d}-{val_loss:.2f}.h5',
monitor='val_loss',
verbose=1,
save_best_only=True,
save_weights_only=True,
mode='min')
csv_logger = k.callbacks.CSVLogger('logs/embedding_sentiment_model.log')
model.fit(X,
y,
epochs=200,
batch_size=5000,
shuffle=True,
callbacks=[checkpoint, csv_logger],
validation_split=0.2)