Converting a model into PyTorch

[p-christ]

Hi,

I'm trying to load a pre-trained model and convert it into a PyTorch model but can't get it to work and was wondering if someone could help me.

I'm able to load the pre-trained model using stable baselines and copy the weights over to a PyTorch model. But then when I play the game the pytorch agent is not able to get the same score as the baselines agent and i am not sure why. It could potentially be because the baselines agent does some extra pre-processing behind the scenes besides just normalising the state to the 0-1 range? Is anyone able to help me?

I've made a colab to demonstrate the problem here: https://colab.research.google.com/drive/1-IIjA1oKUjg5eoctajpl06OoHzU-5-_9

[araffin]

Hello,

It could potentially be because the baselines agent does some extra pre-processing behind the scenes besides just normalising the state to the 0-1 range?

There is no such up to my knowledge. However, I would recommend you first trying with a simpler network (e.g. on CartPole-v1), because it gets tricky with convolution (different conventions for pytorch and tensorflow).
I would also suggest you to check all shapes and names before assigning the weights, using .named_parameters() for pytorch.

EDIT: I made it work with CartPole, will share the notebook soon

[araffin]

Update: I made it work with CartPole, you can find the notebook here: https://colab.research.google.com/drive/1R-wHO2gLQScx46EIjqj7Sj6MjK-i5Hey

Will try to make it work with the cnn if I have some time this weekend.

[araffin]

Update: I'm working on the CNN now, it seems that the problem comes with the first fully connected layer (the conv layer outputs the right thing).

[araffin]

The problem comes from the reshape (from conv to fc)

[araffin]

@p-christ , I solved the issue doing that before flattening:

# shape before flattening
# tf: (?, 7, 7, 64)
# pytorch: [1, 64, 7, 7]
x = x.permute(0, 2, 3, 1).contiguous()
x = x.view(x.size(0), -1)

[araffin]

@p-christ I made a working colab notebook: https://colab.research.google.com/drive/1XwCWeZPnogjz7SLW2kLFXEJGmynQPI-4

The problem came from tensorflow/pytorch differences, not SB.

Closing the issue.

[p-christ]

thanks a lot

[Chainesh]

Can learnings from this be documented, also if it's possible to make a function to do same in sb3?