def __init__(self, training_keylog="log.json", model_name="naive_model", clean_imgs=False):
self.screenshot_dir = os.path.join(helper.get_home_folder(), '.dolphin-emu', 'ScreenShots')
self.keylog_filename = training_keylog
self.output_dir = helper.get_output_folder()
self.image_dir = os.path.join(helper.get_output_folder(), "images")
self.models_dir = os.path.join(helper.get_models_folder())
self.model_name = model_name
self.clean_imgs = clean_imgs
self.state_counts = dict()
self.state_decision_map = dict()
self.defaults = {k.name: 0 for k in keylog.Keyboard}
def test_nn_single_imge():
model = torch.load(os.path.join(helper.get_models_folder(), "mkrnn.pkl"))
image_dir = os.path.join(helper.get_dataset_folder(), 'mario_kart', "images")
image_file_name = "140.png"
img_path = os.path.join(image_dir, image_file_name)
image_data = cv2.imread(img_path)
# Generate tensor from image
x = helper.get_tensor(image_data)
pred = model(x)
key_state = helper.get_key_state_from_vector(pred)
print(pred, key_state)
if torch.cuda.is_available():
valid_y_label = valid_y_label.cuda()
valid_nn_label = Variable(valid_y_label)
valid_forward_pass = mkcrnn(valid_x)
valid_loss_eval = loss_func(
valid_forward_pass,
valid_nn_label) # compute validation loss
valid_loss += valid_loss_eval.data[0]
print('Epoch: ', epoch, 'Step: ', step,
'| validation loss: %.4f' % valid_loss)
validation_losses.append(valid_loss)
# Save model
torch.save(
mkcrnn,
os.path.join(helper.get_models_folder(),
"mkcrnn_{}_gru.pkl".format(history)))
# Save validation curve data
fig_data = [
validation_losses, history, num_epochs, batch_size, learning_rate
]
helper.pickle_object(fig_data,
"mkcrnn_training_data_{}_gru".format(history))
# Plot validation curve
f = plt.figure()
plt.plot(validation_losses)
plt.ylabel('Validation Error')
plt.xlabel('Number of Iterations')
plt.title(
def test_nn(nn_name, history=1):
""" Check that neural network Mario Kart AI can process a Dolphin screenshot and choose an action. """
agent = mk_nn.MarioKartNN(os.path.join(helper.get_models_folder(), nn_name), history_length=history)
while True:
agent.process_frame()
def test_process_frame():
""" Check that the basic Mario Kart AI can process a Dolphin screenshot and choose an action. """
agent = mk_naive_agent.MarioKartAgent(os.path.join(helper.get_models_folder(), "naive_model.pickle"))
while True:
agent.process_frame()
valid_y_label = valid_y_label.cuda()
valid_nn_label = Variable(valid_y_label)
valid_forward_pass = mkcrnn(valid_x)
valid_loss_eval = loss_func(
valid_forward_pass,
valid_nn_label) # compute validation loss
valid_loss += valid_loss_eval.data[0]
print('Epoch: ', epoch, 'Step: ', step,
'| validation loss: %.4f' % valid_loss)
validation_losses.append(valid_loss)
# Save model
torch.save(
mkcrnn,
os.path.join(
helper.get_models_folder(),
"mkcrnn_{}_frames_{}_lstm.pkl".format(num_input_frames, history)))
# Save validation curve data
fig_data = [
validation_losses, num_input_frames, history, num_epochs, batch_size,
learning_rate
]
helper.pickle_object(
fig_data, "mkcrnn_training_data_{}_frames_{}_lstm".format(
num_input_frames, history))
# Plot validation curve
f = plt.figure()
plt.plot(validation_losses)
plt.ylabel('Validation Error')
# Log training data
if step % 50 == 0:
print('Epoch: ', epoch, 'Step: ', step, '| training loss: %.4f' % loss.data[0])
valid_loss = 0
for (valid_x, valid_y) in valid_loader:
valid_y_label = valid_y.view(-1, output_vec)
if torch.cuda.is_available():
valid_y_label = valid_y_label.cuda()
valid_nn_label = Variable(valid_y_label)
valid_forward_pass = mknn(valid_x)
valid_loss_eval = loss_func(valid_forward_pass, valid_nn_label) # compute validation loss
valid_loss += valid_loss_eval.data[0]
print('Epoch: ', epoch, 'Step: ', step, '| validation loss: %.4f' % valid_loss)
validation_losses.append(valid_loss)
# Save model
torch.save(mknn, os.path.join(helper.get_models_folder(), "mknn.pkl"))
# Save validation curve data
fig_data = [validation_losses, num_epochs, batch_size, learning_rate]
helper.pickle_object(fig_data, "mknn_training_data_lr{}_epoch{}".format(learning_rate, num_epochs))
# Plot validation curve
f = plt.figure()
plt.plot(validation_losses)
plt.ylabel('Validation Error')
plt.xlabel('Number of Iterations')
plt.title('NN Cross Validation Error, Learning Rate = %s, Batch Size = %i, Number of Epochs = %i' % (
learning_rate, batch_size, num_epochs))
plt.show(block=True)