def k_fold_cross_validation(mlp,
dataset,
k=5,
learning_rate=0.01,
momentum=0.7,
epochs=100,
threshold=None):
MSE_train_mean = 0.0
MSE_test_mean = 0.0
parts = split_dataset(dataset, k)
target_test = []
output_test = []
for k_i in np.arange(k):
mlp.init_weights()
training_parts = set(np.arange(k))
training_parts.remove(k_i)
dataset_train = np.concatenate(
[parts[i] for i in list(training_parts)])
dataset_test = parts[k_i]
input_data = dataset_train[:, 0:mlp.n_inputs]
output_data = dataset_train[:, mlp.n_inputs:(mlp.n_inputs +
mlp.n_outputs)]
input_data_test = dataset_test[:, 0:mlp.n_inputs]
output_data_test = dataset_test[:, mlp.n_inputs:(mlp.n_inputs +
mlp.n_outputs)]
mlp.fit((input_data, output_data),
learning_rate=learning_rate,
momentum=momentum,
epochs=epochs)
MSE_train, _ = mlp.compute_MSE((input_data, output_data))
MSE_train_mean += MSE_train
MSE_test, temp_out = mlp.compute_MSE(
(input_data_test, output_data_test))
MSE_test_mean += MSE_test
output_test.append(temp_out)
target_test.append(output_data_test)
target_test = np.concatenate(target_test, axis=0)
output_test = np.concatenate(output_test, axis=0)
if threshold is None:
return (MSE_train_mean / k, MSE_test_mean / k)
else:
return (MSE_train_mean / k, MSE_test_mean / k,
compute_confusion_matrix(target_test, output_test,
threshold), target_test, output_test)
def k_fold_cross_validation_per_epoch(mlp,
dataset,
k=5,
learning_rate=0.01,
momentum=0.7,
epochs=100):
MSE_train = np.zeros((k, epochs))
MSE_test = np.zeros((k, epochs))
parts = split_dataset(dataset, k)
for k_i in np.arange(k):
mlp.init_weights()
training_parts = set(np.arange(k))
training_parts.remove(k_i)
dataset_train = np.concatenate(
[parts[i] for i in list(training_parts)])
dataset_test = parts[k_i]
input_data = dataset_train[:, 0:mlp.n_inputs]
output_data = dataset_train[:, mlp.n_inputs:(mlp.n_inputs +
mlp.n_outputs)]
input_data_test = dataset_test[:, 0:mlp.n_inputs]
output_data_test = dataset_test[:, mlp.n_inputs:(mlp.n_inputs +
mlp.n_outputs)]
MSE_train[k_i, :], MSE_test[k_i, :] = mlp.fit(
(input_data, output_data), (input_data_test, output_data_test),
learning_rate=learning_rate,
momentum=momentum,
epochs=epochs)
return (np.mean(MSE_train, axis=0), np.mean(MSE_test, axis=0))
def k_fold_cross_validation(mlp, dataset, k=5, learning_rate=0.01, momentum=0.7, epochs=100, threshold=None):
MSE_train_mean = 0.0
MSE_test_mean = 0.0
parts = split_dataset(dataset, k)
target_test = []
output_test = []
for k_i in np.arange(k):
mlp.init_weights()
training_parts = set(np.arange(k))
training_parts.remove(k_i)
dataset_train = np.concatenate([parts[i] for i in list(training_parts)])
dataset_test = parts[k_i]
input_data = dataset_train[:,0:mlp.n_inputs]
output_data = dataset_train[:,mlp.n_inputs:(mlp.n_inputs+mlp.n_outputs)]
input_data_test = dataset_test[:,0:mlp.n_inputs]
output_data_test = dataset_test[:,mlp.n_inputs:(mlp.n_inputs+mlp.n_outputs)]
mlp.fit((input_data, output_data),
learning_rate=learning_rate, momentum=momentum, epochs=epochs)
MSE_train, _ = mlp.compute_MSE((input_data, output_data))
MSE_train_mean += MSE_train
MSE_test, temp_out = mlp.compute_MSE((input_data_test, output_data_test))
MSE_test_mean += MSE_test
output_test.append(temp_out)
target_test.append(output_data_test)
target_test = np.concatenate(target_test, axis=0)
output_test = np.concatenate(output_test, axis=0)
if threshold is None:
return (MSE_train_mean / k, MSE_test_mean / k)
else:
return (MSE_train_mean / k, MSE_test_mean / k, compute_confusion_matrix(target_test, output_test, threshold))
def k_fold_cross_validation_per_epoch(mlp, dataset, k=5, learning_rate=0.01, momentum=0.7, epochs=100):
MSE_train = np.zeros((k, epochs))
MSE_test = np.zeros((k, epochs))
parts = split_dataset(dataset, k)
for k_i in np.arange(k):
mlp.init_weights()
training_parts = set(np.arange(k))
training_parts.remove(k_i)
dataset_train = np.concatenate([parts[i] for i in list(training_parts)])
dataset_test = parts[k_i]
input_data = dataset_train[:,0:mlp.n_inputs]
output_data = dataset_train[:,mlp.n_inputs:(mlp.n_inputs+mlp.n_outputs)]
input_data_test = dataset_test[:,0:mlp.n_inputs]
output_data_test = dataset_test[:,mlp.n_inputs:(mlp.n_inputs+mlp.n_outputs)]
MSE_train[k_i,:], MSE_test[k_i,:] = mlp.fit((input_data, output_data), (input_data_test, output_data_test),
learning_rate=learning_rate, momentum=momentum, epochs=epochs)
return (np.mean(MSE_train, axis=0), np.mean(MSE_test, axis=0))