Closed
Description
When I tried the mario_rl_tutorial.py using CUDA, I received an out of memory error:
RuntimeError: CUDA error: out of memory
Monitoring the GPU memory I saw a constant incremental of the use of its memory, specially when the Agent starts the training. As a workaround I changed the cache method of the Agent, as it seems that all the experiencies are sent to the GPU:
if self.use_cuda:
state = torch.tensor(state).cuda()
next_state = torch.tensor(next_state).cuda()
action = torch.tensor([action]).cuda()
reward = torch.tensor([reward]).cuda()
done = torch.tensor([done]).cuda()
(...)
self.memory.append((state, next_state, action, reward, done,))Instead of that, I modified this piece for using CPU and changed the recall method, to send at this moment the batch of experiencies to GPU:
def recall(self):
"""
Retrieve a batch of experiences from memory
"""
batch = random.sample(self.memory, self.batch_size)
state, next_state, action, reward, done = map(torch.stack, zip(*batch))
return state.cuda(), next_state.cuda(), action.squeeze().cuda(), reward.squeeze().cuda(), done.squeeze().cuda()Not sure if this is the right approach, but it works. However, it might impact on the learning performance.
Could anyone help here?