# Fully connected layer model
model = ModelUtils.getFcModel(num_ftrs, NUM_CLASSES,
parameters['hidden_layers'],
parameters['fc_neurons'],
parameters['activation'],
parameters['dropout_bp'])
model.to(device) # Move model to device
# summary(model,input_size=(TRAIN_BATCH_SIZE,num_ftrs))
# Predictive Coding model
pc_model_architecture = ModelUtils.getPcModelArchitecture(
num_ftrs, NUM_CLASSES, parameters['hidden_layers'],
parameters['fc_neurons'])
pc_model = PcTorch(pc_model_architecture)
pc_model.set_training_parameters(TRAIN_BATCH_SIZE,
INFERENCE_STEPS,
parameters['activation'],
parameters['optimizer'],
parameters['lr_pc'],
parameters['momentum_pc'],
normalize_input=True)
# Loss and optmizer
criterion = nn.CrossEntropyLoss()
optimizer = 'None'
if parameters['optimizer'] == 'sgd':
optimizer = optim.SGD(model.parameters(),
lr=parameters['lr_bp'],
for i in range(len(x_train)):
x_train_list.append(x_train[i].flatten().astype(np.float))
for i in range(len(x_test)):
x_test_list.append(x_test[i].flatten().astype(np.float))
# normalize dataset
x_train_list, mi, ma = util.normalize_dataset(x_train_list)
x_test_list, mi, ma = util.normalize_dataset(x_test_list)
# Get time before training
t_start = datetime.datetime.now()
print("Starting timer")
# Initialize network and train
model_torch = PcTorch(NETWORK_ARCHITECTURE)
model_torch.train(x_train_list,
y_train_list,
x_test_list,
y_test_list,
batch_size=BATCH_SIZE,
epochs=EPOCHS,
max_it=INFERENCE_STEPS,
optmizer=OPTIMIZER,
activation=ACTIVATION,
dataset_perc=DATA_PERC,
learning_rate=LR)
# Get time after training
t_end = datetime.datetime.now()
elapsedTime = (t_end - t_start)
