def evaluate(model, data_input, gold_output):
predictions = model.predict(
data_input,
batch_size=keras_models.model_params['batch_size'],
verbose=1)
if len(predictions.shape) == 3:
predictions_classes = np.argmax(predictions, axis=2)
train_batch_f1 = metrics.accuracy_per_sentence(predictions_classes,
gold_output)
print("Results (per sentence): ", train_batch_f1)
train_y_properties_stream = gold_output.reshape(gold_output.shape[0] *
gold_output.shape[1])
predictions_classes = predictions_classes.reshape(
predictions_classes.shape[0] * predictions_classes.shape[1])
class_mask = train_y_properties_stream != 0
train_y_properties_stream = train_y_properties_stream[class_mask]
predictions_classes = predictions_classes[class_mask]
else:
predictions_classes = np.argmax(predictions, axis=1)
train_y_properties_stream = gold_output
accuracy = metrics.accuracy(predictions_classes, train_y_properties_stream)
micro_scores = metrics.compute_micro_PRF(predictions_classes,
train_y_properties_stream,
empty_label=keras_models.p0_index)
print("Results: Accuracy: ", accuracy)
print("Results: Micro-Average F1: ", micro_scores)
return predictions_classes, predictions
def f_train(params):
model = getattr(keras_models, model_name)(params, embedding_matrix, max_sent_len, n_out)
callback_history = model.fit(train_as_indices[:-1],
[train_y_properties_one_hot],
epochs=20, batch_size=keras_models.model_params['batch_size'], verbose=1,
validation_data=(
val_as_indices[:-1], val_y_properties_one_hot),
callbacks=[callbacks.EarlyStopping(monitor="val_loss", patience=1, verbose=1)])
predictions = model.predict(val_as_indices[:-1], batch_size=16, verbose=1)
predictions_classes = np.argmax(predictions, axis=1)
_, _, acc = metrics.compute_micro_PRF(predictions_classes, val_as_indices[-1])
return {'loss': -acc, 'status': hy.STATUS_OK}