epochs=epochs,
validation_data=(test_dataset_x[0],
feature_test_data),
verbose=verbose)
print('{} train:'.format(student.name),
student_history.history['mean_squared_error'])
print('{} test:'.format(student.name),
student_history.history['val_mean_squared_error'])
#dataset_update(student_history, student,parameters)
student_predictions = student(test_dataset_x[0][:evaluate_prediction_size])
return student_predictions
# student1_predictions = train_student(student1(), model, )
# student2_predictions = train_student(student2(), model, )
student3_predictions = train_student(
student3(),
model,
)
# student_cnn_predictions = train_student(studentcnn(), model, )
with open(
prediction_data_path +
'predictions_2_{}_{}_{}'.format(layer_name, feature, run_index),
'wb') as file_pi:
pickle.dump(
(feature_test_data[:evaluate_prediction_size], student3_predictions),
file_pi)
pickle.dump(all_weights, file_pi)
#LOAD WITH - pickle.load - and load manualy to model.get_layer.set_weights()
#save weights with keras
keras_weights_path = child_folder + '{}_keras_weights/'.format(
feature_list)
if not os.path.exists(keras_weights_path):
os.mkdir(keras_weights_path)
net.save_weights(keras_weights_path +
'keras_weights_{}_{}'.format(feature_list, traject))
#LOADING WITH - load_status = sequential_model.load_weights("ckpt")
student = student3(sample=sample,
res=res,
activation='relu',
dropout=dropout,
rnn_dropout=rnn_dropout,
num_feature=num_feature)
student.evaluate(test_dataset_x[0], feature_test_data, verbose=2)
student_history = student.fit(
train_dataset_x[0],
feature_train_data,
batch_size=32,
epochs=epochs,
validation_data=(test_dataset_x[0], feature_test_data),
verbose=verbose,
callbacks=[model_checkpoint_callback, lr_reducer, early_stopper])
print('{} train:'.format(student.name),
student_history.history['mean_squared_error'])
print('{} test:'.format(student.name),
t_f = True
feature_list = 'all'
temp_path = path + 'saved_models/{}_feature/'.format(feature_list)
home_folder = temp_path + '{}_{}_saved_models/'.format(feature_list, trajectory_index)
checkpoint = t_f
if checkpoint:
child_folder = home_folder + 'checkpoint/'
else:
child_folder = home_folder + 'end_of_run_model/'
#loading weights as numpy array
numpy_weights_path = child_folder + '{}_numpy_weights/'.format(feature_list)
with open(numpy_weights_path + 'numpy_weights_{}_{}'.format(feature_list,trajectory_index), 'rb') as file_pi:
np_weights = pickle.load(file_pi)
numpy_student = student3(activation = 'relu', dropout = 0.2, rnn_dropout = 0, num_feature = 64)
layer_index = 0
for i in range(3):
layer_name = 'convLSTM{}'.format(i+1)
saved_weights = [np_weights[layer_index], np_weights[layer_index+ 1], np_weights[layer_index+ 2]]
numpy_student.get_layer(layer_name).set_weights(saved_weights)
layer_index += 3
res = 8
sample = 10
def full_student(student, decoder):
input = keras.layers.Input(shape=(sample, res,res,3))\
student_features = student(input)
decoder_prediction = decoder(student_features)
model = keras.models.Model(inputs=input,outputs=decoder_prediction)
student_history.history['val_mean_squared_error'])
dataset_update(student_history, student, parameters, name='sample_5')
print(student.get_weights()[3][:4, :4])
save_model(student, save_model_path, parameters, checkpoint=False)
student.load_weights(checkpoint_filepath)
save_model(student, save_model_path, parameters, checkpoint=True)
student.evaluate(test_dataset_x[0], feature_test_data, verbose=2)
student_predictions = student(test_dataset_x[0])
train_predictions = student(train_dataset_x[0])
return student_predictions, train_predictions
#jul 15 on 13:00 I chainged the activation to relu befor it was tanh (defoult) run 184770 /184790
#student1_predictions = train_student(student1(activation = 'relu'), model, )
#student2_predictions = train_student(student2(activation = 'relu'), model, )
student3_predictions, train3_predictions = train_student(
student3(activation='relu', dropout=dropout, rnn_dropout=rnn_dropout),
model,
)
#student3_predictions = train_student(student3(activation = 'relu', dropout = dropout, rnn_dropout = rnn_dropout), model, )
#student_cnn_predictions = train_student(studentcnn(), model, )
with open(
prediction_data_path + 'predictions_traject_{}_{}_{}_{}'.format(
layer_name,
feature,
trajectory_index,
run_index,
), 'wb') as file_pi:
pickle.dump((feature_test_data, student3_predictions), file_pi)
validation_data=(test_dataset_x[0], feature_test_data),
verbose = verbose,
callbacks=[model_checkpoint_callback])
print('{} train:'.format(student.name), student_history.history['mean_squared_error'])
print('{} test:'.format(student.name), student_history.history['val_mean_squared_error'])
dataset_update(student_history, student,parameters, name = 'compare_traject')
print(student.get_weights()[3][:4,:4])
save_model(student, save_model_path, parameters, checkpoint = False)
student.load_weights(checkpoint_filepath)
save_model(student, save_model_path, parameters, checkpoint = True)
student.evaluate(test_dataset_x[0],
feature_test_data, verbose = 2)
student_predictions = student(test_dataset_x[0])
train_predictions = student(train_dataset_x[0])
return student_predictions, train_predictions
#jul 15 on 13:00 I chainged the activation to relu befor it was tanh (defoult) run 184770 /184790
#student1_predictions = train_student(student1(activation = 'relu'), model, )
#student2_predictions = train_student(student2(activation = 'relu'), model, )
student3_predictions,train3_predictions = train_student(student3(activation = 'relu', dropout = dropout, rnn_dropout = rnn_dropout), model, )
#student3_predictions = train_student(student3(activation = 'relu', dropout = dropout, rnn_dropout = rnn_dropout), model, )
#student_cnn_predictions = train_student(studentcnn(), model, )
with open(prediction_data_path + 'predictions_traject_{}_{}_{}'.format(layer_name, feature, run_index), 'wb') as file_pi:
pickle.dump((feature_test_data, student3_predictions, train3_predictions), file_pi)