Thomas Rochefort-Beaudoin

@araffin the error happens at the start of the training. I am using `VecNormalize` because of the general recommendation of the documentation. I guess I could only normalize the rewards...

@araffin What is curious is the drastic difference in memory usage between using only the Flatten() layer and the two Conv2D layer.

OS: Linux-5.14.18-100.fc33.x86_64-x86_64-with-fedora-33-Thirty_Three #1 SMP Fri Nov 12 17:38:44 UTC 2021 Python: 3.7.10 Stable-Baselines3: 1.5.0 PyTorch: 1.11.0+cu113 GPU Enabled: True Numpy: 1.20.3 Gym: 0.21.0

@araffin I guess your right! Although its strange that the difference between the FeatureExtractor with the nn.Flatten() layer only vs. the one with the two Conv2D should only be a...

replacing the two nn.Conv2D layers by two nn.Linear layers of 1024 units works like a charm and has no significant impact on the memory usage. Those 2 dense layers should...

@Liuzy0908 , Here is how I understand it: The environment is wrapped when you instantiate GAIL so that the reward used in the .learn method corresponds to the GAIL generator...

Also, the ResNet embedding is done on single image patches instead of the whole complete image as it is hinted on the figure "Full Episode Sequence" (See Figure 15 in...

@OrigamiDream The way I understand the paper, all of the inputs are tokenized which implies that the main transformer block sees a constant shape at its input. The way I...