def pre_training(df, cfg):
# Train autoencoder as pre training
encoder_train = df[:-int(len(df) * cfg['test_size'])]
print(encoder_train)
train_x = []
for name, values in df.iteritems():
x, _ = split_sequence(values.values.reshape(-1, 1), cfg)
train_x.append(x)
train_x = np.asarray(train_x)
train_x = train_x.reshape([
train_x.shape[0] * train_x.shape[1], train_x.shape[2], train_x.shape[3]
])
print(train_x.shape)
encoder, decoder, cfg = build_autoencoder(
train_x, cfg, weights='weights//pretrained_encoder.hdf5')
# Test autoencoder on holdout data
#encoder_test = data.x_test
#predictions = decoder.predict(encoder_test)
#mse = 0
#for i in range(len(predictions)):
# mse += mean_squared_error(encoder_test[i], predictions[i])
#print('Test mean mse:', mse/len(predictions))
return encoder, cfg
def pre_training(data, cfg):
# Train autoencoder as pre training
train, test = data.get_x()
encoder, decoder, cfg = build_autoencoder(train, cfg, weights='weights//pretrained_encoder.hdf5')
# Test autoencoder on holdout data
predictions = decoder.predict(test)
mse = 0
for i in range(len(predictions)):
mse += mean_squared_error(test[i], predictions[i])
print('Test mean mse:', mse/len(predictions))
return encoder, cfg
def train_autoencoder(data, cfg):
# Train autoencoder as pre training
encoder_train = np.concatenate([data.train_conv, data.train_org], axis=0)
encoder, decoder, cfg = build_autoencoder(encoder_train, cfg, weights='weights//pretrained_encoder.hdf5')
return encoder, cfg