Anas BELFADIL

I think the problem is in `RecurrentActorProb` class. From this part of the code it seems that it expects an input of shape [bsz, len*dim]: ```python self.nn = nn.LSTM( input_size=int(np.prod(state_shape)),...

@Trinkle23897 Thank you for your answer. I understand that, and that's not the problem. To explain better: I see a part of the code where the input expected is of...

You mean I should have `dim` instead of `len*dim`? Even when I'm working with `stack_num!=1`? But anyway self.nn is getting `obs` of shape `[bsz, len, dim]` when it is expecting...

If I have an observation of shape `[bsz, len, dim]` what is the `state_shape` argument that I should pass to `RecurrentActorProb`?

Okay, thanks for the support. It is a little bit confusing since `state_shape` for the normal `ActorProb` is equal to `obs.shape`, maybe you can consider making them (`ActorProb` and `RecurrentActorProb`)...

For making CQL work with RNN, I changed `tmp_obs` and `tmp_obs_next` in cql.py>CQLPolicy>learn as follows: ```python tmp_obs = obs.unsqueeze(1) \ .repeat(1, self.num_repeat_actions, 1, 1) \ .view(batch_size * self.num_repeat_actions, obs.shape[-2], obs.shape[-1])...

Which task would you recommend for testing this solution? Ideally it should be a task in d4rl datasets and where SAC has been tried with RNNs and worked correctly, since...

Anyone got any RNN based algo to learn correctly?

@adriangonz Thank you for the swift response. I'm trying to implement it, but I run to a problem that I'm not able to solve. I want to pass a dataframe...

Hi @adriangonz Thanks for your comment. Yes I use datetime for content_type, but it's the index that I need to pass along with the columns, because I want to do...