classifier_optimizer = torch.optim.Adam(cnn.parameters(), lr=0.001)
running_reward = 0
batch = []
labels = []
total_steps = 0
if args.mode == "train" or args.mode == "all":
for i_episode in count(1000):
observation = env.reset()
print("episode: ", i_episode)
for t in range(1000):
action = select_action(observation,env.action_space_n(),args.epsilon)
observation, reward, done, info = env.step(action)
model.rewards.append(reward)
if env.is_touching():
print("touching!")
#print("batch size", len(batch))
if len(batch) > args.batch_size:
#TODO GPU support
#batch = torch.from_numpy(np.asarray(batch))
batch = torch.LongTensor(torch.from_numpy(np.asarray(batch)))
labels = torch.from_numpy(np.asarray(labels))
#labels = torch.LongTensor(torch.from_numpy(np.asarray(labels)))
if args.gpu and torch.cuda.is_available():
batch = batch.cuda()
labels = labels.cuda()
batch = Variable(batch)
labels = Variable(labels)
classifier_optimizer.zero_grad()
outputs = cnn(batch)
done = False
while not done:
action = RL.choose_action(observation)
observation_, reward, done, info = env.step(action)
# something to consider - should we modify the reward if it's the terminal state and
# we haven't touched yet? Massive penalty for finishing the round with no touch
RL.store_transition(observation, action, reward, observation_)
ep_r[i_episode] += reward
if total_steps > 1000:
cost = RL.learn()
if env.is_touching():
print('\ntouching at step', env.steps, 'total reward is ',
ep_r[i_episode])
games_where_touched += 1
cnn_features_TD[TD_cnt] = observation_
cnn_labels_TD[TD_cnt] = env.class_label
TD_cnt += 1
ep_touch[i_episode] += 1
if (env.steps % 500 == 0):
print('\nepisode: ', i_episode + 1, 'step: ', env.steps,
'episode reward ', ep_r[i_episode])
observation = observation_
total_steps += 1