Below is my implementation of a2c using PyTorch. Upon learning about backpropagation in PyTorch, I have known to zero_grad() the optimizer after each update iteration. However, there is still a RunTime error on second-time backpropagation.
def torchworker(number, model):
worker_env = gym.make("Taxi-v3").env
max_steps_per_episode = 2000
worker_opt = optim.Adam(lr=5e-4, params=model.parameters())
p_history = []
val_history = []
r_history = []
running_reward = 0
episode_count = 0
under = 0
start = time.time()
for i in range(2):
state = worker_env.reset()
episode_reward = 0
penalties = 0
drop = 0
print("Episode {} begins ({})".format(episode_count, number))
worker_env.render()
criterion = nn.SmoothL1Loss()
time_solve = 0
for _ in range(1, max_steps_per_episode):
#worker_env.render()
state = torch.tensor(state, dtype=torch.long)
action_probs = model.forward(state)[0]
critic_value = model.forward(state)[1]
val_history.append((state, critic_value[0]))
# Choose action
action = np.random.choice(6, p=action_probs.detach().numpy())
p_history.append(torch.log(action_probs[action]))
# Apply chosen action
state, reward, done, _ = worker_env.step(action)
r_history.append(reward)
episode_reward += reward
time_solve += 1
if reward == -10:
penalties += 1
elif reward == 20:
drop += 1
if done:
break
# Update running reward to check condition for solving
running_reward = (running_reward * (episode_count) + episode_reward) / (episode_count + 1)
# Calculate discounted returns
returns = deque(maxlen=3500)
discounted_sum = 0
for r in r_history[::-1]:
discounted_sum = r + gamma * discounted_sum
returns.appendleft(discounted_sum)
# Calculate actor losses and critic losses
loss_actor_value = 0
loss_critic_value = 0
history = zip(p_history, val_history, returns)
for log_prob, value, ret in history:
diff = ret - value[1]
loss_actor_value += -log_prob * diff
ret_tensor = torch.tensor(ret, dtype=torch.float32)
loss_critic_value += criterion(value[1], ret_tensor)
loss = loss_actor_value + 0.1 * loss_critic_value
print(loss)
# Update params
loss.backward()
worker_opt.step()
worker_opt.zero_grad()
# Log details
end = time.time()
episode_count += 1
if episode_count % 1 == 0:
worker_env.render()
if running_reward > -50: # Condition to consider the task solved
under += 1
if under > 5:
print("Solved at episode {} !".format(episode_count))
break
I believe there may be something to do with the architecture of my AC model, so I also include it here for reference.
class ActorCriticNetwork(nn.Module):
def __init__(self, num_inputs, num_hidden, num_actions):
super(ActorCriticNetwork, self).__init__()
self.embed = nn.Embedding(500, 10)
self.fc1 = nn.Linear(10, num_hidden * 2)
self.fc2 = nn.Linear(num_hidden * 2, num_hidden)
self.c = nn.Linear(num_hidden, 1)
self.fc3 = nn.Linear(num_hidden, num_hidden)
self.a = nn.Linear(num_hidden, num_actions)
def forward(self, x):
out = F.relu(self.embed(x))
out = F.relu(self.fc1(out))
out = F.relu(self.fc2(out))
critic = self.c(out)
out = F.relu(self.fc3(out.detach()))
actor = F.softmax(self.a(out), dim=-1)
return actor, critic
Would you please tell me what the mistake here is? Thank you in advance.
