def main():
path_to_vec = '../glove/glove.6B.100d.txt'
dir_name = '../'
reader = SSTDataReader(dir_name, nclasses=2)
embedding_params = reader.get_word_embedding(path_to_vec,
orthonormalized=False)
lookup_table = get_lookup_table(embedding_params)
max_sequence_length = 60
sequence_input = Input(shape=(max_sequence_length, ), dtype='int32')
phase_embedding = phase_embedding_layer(max_sequence_length,
lookup_table.shape[0])
amplitude_embedding = amplitude_embedding_layer(np.transpose(lookup_table),
max_sequence_length)
# [embed_seq_real, embed_seq_imag] = ComplexMultiply()([phase_embedding, amplitude_embedding])
output = phase_embedding(sequence_input)
model = Model(sequence_input, output)
model.compile(loss='binary_crossentropy',
optimizer='rmsprop',
metrics=['accuracy'])
model.summary()
train_test_val = reader.create_batch(embedding_params=embedding_params,
batch_size=-1)
training_data = train_test_val['train']
test_data = train_test_val['test']
validation_data = train_test_val['dev']
# for x, y in batch_gen(training_data, max_sequence_length):
# model.train_on_batch(x,y)
train_x, train_y = data_gen(training_data, max_sequence_length)
test_x, test_y = data_gen(test_data, max_sequence_length)
val_x, val_y = data_gen(validation_data, max_sequence_length)
# sequence_input = Input(shape=(max_sequence_length,), dtype='int32')
# path_to_vec = '../glove/glove.6B.100d.txt'
# embedded_sequences = amplitude_embedding_layer(path_to_vec, 10)
# output = embedded_sequences(sequence_input)
# model = Model(sequence_input, output)
# model.compile(loss='categorical_crossentropy',
# optimizer='rmsprop',
# metrics=['acc'])
# model.summary()
x = train_x
y = model.predict(x)
print(y)
print(y.shape)
df1.ix[l[params.network_type],params.dataset_name] = max(grid_result.best_score_)
df1.to_excel(experiment_results_path)
if __name__ == '__main__':
params = Params()
params.parse_config('config/waby.ini')
reader = data_reader_initialize(params.dataset_name,params.datasets_dir)
if(params.wordvec_initialization == 'orthogonalize'):
embedding_params = reader.get_word_embedding(params.wordvec_path,orthonormalized=True)
elif( (params.wordvec_initialization == 'random') | (params.wordvec_initialization == 'word2vec')):
embedding_params = reader.get_word_embedding(params.wordvec_path,orthonormalized=False)
else:
raise ValueError('The input word initialization approach is invalid!')
# print(embedding_params['word2id'])
lookup_table = get_lookup_table(embedding_params)
gridsearch(params)
def complex_embedding(params):
# datasets_dir, dataset_name, wordvec_initialization ='random', wordvec_path = None, loss = 'binary_crossentropy', optimizer = 'rmsprop', batch_size = 16, epochs= 4
reader = data_reader_initialize(params.dataset_name, params.datasets_dir)
if (params.wordvec_initialization == 'orthogonalize'):
embedding_params = reader.get_word_embedding(params.wordvec_path,
orthonormalized=True)
elif ((params.wordvec_initialization == 'random') |
(params.wordvec_initialization == 'word2vec')):
embedding_params = reader.get_word_embedding(params.wordvec_path,
orthonormalized=False)
else:
raise ValueError('The input word initialization approach is invalid!')
# print(embedding_params['word2id'])
lookup_table = get_lookup_table(embedding_params)
max_sequence_length = reader.max_sentence_length
random_init = True
if not (params.wordvec_initialization == 'random'):
random_init = False
if params.network_type == 'complex_superposition':
model = run_complex_embedding_network_superposition(
lookup_table,
max_sequence_length,
reader.nb_classes,
random_init=random_init)
elif params.network_type == 'complex_mixture':
model = run_complex_embedding_network_mixture(lookup_table,
max_sequence_length,
reader.nb_classes,
random_init=random_init)
else:
model = run_real_embedding_network(lookup_table,
max_sequence_length,
reader.nb_classes,
random_init=random_init)
model.compile(loss=params.loss,
optimizer=params.optimizer,
metrics=['accuracy'])
model.summary()
weights = model.get_weights()
train_test_val = reader.create_batch(embedding_params=embedding_params,
batch_size=-1)
training_data = train_test_val['train']
test_data = train_test_val['test']
validation_data = train_test_val['dev']
# for x, y in batch_gen(training_data, max_sequence_length):
# model.train_on_batch(x,y)
train_x, train_y = data_gen(training_data, max_sequence_length)
test_x, test_y = data_gen(test_data, max_sequence_length)
val_x, val_y = data_gen(validation_data, max_sequence_length)
print(len(train_x))
print(len(test_x))
print(len(val_x))
# assert len(train_x) == 67349
# assert len(test_x) == 1821
# assert len(val_x) == 872
train_y = to_categorical(train_y)
test_y = to_categorical(test_y)
val_y = to_categorical(val_y)
history = model.fit(x=train_x,
y=train_y,
batch_size=params.batch_size,
epochs=params.epochs,
validation_data=(test_x, test_y))
val_acc = history.history['val_acc']
train_acc = history.history['acc']
if not (os.path.exists(params.eval_dir)):
os.mkdir(params.eval_dir)
learning_curve_path = os.path.join(params.eval_dir, 'learning_curve')
epoch_indexes = [x + 1 for x in range(len(val_acc))]
line_1, = plt.plot(epoch_indexes, val_acc)
line_2, = plt.plot(epoch_indexes, train_acc)
# plt.axis([0, 6, 0, 20])
plt.legend([line_1, line_2], ['test_acc', 'train_acc'])
fig = plt.gcf()
fig.savefig(learning_curve_path, dpi=fig.dpi)
evaluation = model.evaluate(x=test_x, y=test_y)
eval_file_path = os.path.join(params.eval_dir, 'eval.txt')
with open(eval_file_path, 'w') as eval_file:
eval_file.write('acc: {}, loss: {}'.format(evaluation[1],
evaluation[0]))
embedding_dir = os.path.join(params.eval_dir, 'embedding')
if not (os.path.exists(embedding_dir)):
os.mkdir(embedding_dir)
np.save(os.path.join(embedding_dir, 'phase_embedding'),
model.get_weights()[0])
np.save(os.path.join(embedding_dir, 'amplitude_embedding'),
model.get_weights()[1])
np.save(os.path.join(embedding_dir, 'word2id'),
embedding_params['word2id'])
save_model(model, os.path.join(params.eval_dir, 'model'))
experiment_results_path = 'eval/experiment_result.xlsx'
xls_file = pd.ExcelFile(experiment_results_path)
df1 = xls_file.parse('Sheet1')
l = {'complex_mixture': 0, 'complex_superposition': 1, 'real': 2}
df1.ix[l[params.network_type], params.dataset_name] = max(val_acc)
df1.to_excel(experiment_results_path)