size mismatch when using FullyObsWrapper about minigrid HOT 5 CLOSED

This is normal. The default model is made to take as input a 7x7x3 image. It applies convolutions hence assumes the output has size 64 (https://github.com/lcswillems/pytorch-a2c-ppo/blob/9e87fdad55c1139526a8228dfe15d0b667fbedae/model.py#L25). So you should modify this number.

I am modifying my code to do it automatically. I will tell you when it is done.

from minigrid.

This is done. Now, the model adapts to the observation space (https://github.com/lcswillems/pytorch-a2c-ppo/blob/e37dd17f9115c08b4234cc85b38f7429c9623560/model.py#L36). However, keep in mind that the model created automatically for the fully-observable space is not necessarily the best.

from minigrid.

Yes, the problem is that with a fully observable grid, the size of your observation is the size of the grid. I unfortunately don't know of a way to automatically adjust the convolutional layer sizes in PyTorch, so you have to manually adjust these so that the size of the tensors being passed through operations match:

Convolutional layers:
https://github.com/lcswillems/pytorch-a2c-ppo/blob/master/model.py#L26

You could also pad the size of the embedding after the convolution, or make it go through a fully-connected layer to get an embedding of the size you want:
https://github.com/lcswillems/pytorch-a2c-ppo/blob/master/model.py#L80

from minigrid.

@lcswillems Awesome if you can do it automatically! 👍

from minigrid.

Of course. Browsed through that code to see if the input size was hardcoded, but didn't realize a fixed CNN output size implies a fixed input size too. Thanks for the quick replies!

from minigrid.