def _train_ff_network(
hyperparameter_dict: dict,
data: SignalData) -> Tuple[FeedForward, List, List, List, List]:
"""Trains a feed-forward network using the specified hyperparameters.
"""
# Ensure reproducibility by giving PyTorch the same seed every time we train.
torch.manual_seed(1)
# Print hyperparameters.
print(f'Hyperparameters: {hyperparameter_dict}')
# Get hyperparameters.
learning_rate = hyperparameter_dict['learning_rate']
batch_size = hyperparameter_dict['batch_size']
optimizer_str = hyperparameter_dict['optimizer']
# There are 6 labels, and Pytorch expects them to go from 0 to 5.
full_train_labels = data.train_labels - 1
# Get generators.
signal_dataset = SignalDataset(data.train_signals, full_train_labels)
(training_generator,
validation_generator) = utils_nn.get_trainval_generators(
signal_dataset, batch_size, num_workers=0, training_fraction=0.8)
# Crete feed forward network.
input_size = data.num_timesteps * data.num_components
feed_forward = FeedForward(input_size, input_size,
data.num_activity_labels)
print(feed_forward)
# Parameters should be moved to GPU before constructing the optimizer.
device = torch.device('cuda:0' if USE_CUDA else 'cpu')
feed_forward = feed_forward.to(device)
# Get optimizer.
optimizer = None
if optimizer_str == 'adam':
optimizer = torch.optim.Adam(feed_forward.parameters(),
lr=learning_rate)
elif optimizer_str == 'sgd':
optimizer = torch.optim.SGD(feed_forward.parameters(),
lr=learning_rate)
else:
raise Exception(f'Specified optimizer not valid: {optimizer_str}')
training_accuracy_list = []
training_loss_list = []
validation_accuracy_list = []
validation_loss_list = []
max_epochs = 10
for epoch in range(max_epochs):
print(f'Epoch {epoch}')
# Training data.
(training_accuracy,
training_loss) = utils_nn.fit(feed_forward, training_generator,
optimizer, USE_CUDA)
training_accuracy_list.append(training_accuracy)
training_loss_list.append(training_loss)
# Validation data.
(validation_accuracy,
validation_loss) = utils_nn.evaluate(feed_forward,
validation_generator,
'Validation', USE_CUDA)
validation_accuracy_list.append(validation_accuracy)
validation_loss_list.append(validation_loss)
return (feed_forward, training_accuracy_list, training_loss_list,
validation_accuracy_list, validation_loss_list)
