I think there is a small mistake in the model of the i3d. https://github.com/tomrunia/PyTorchConv3D/blob/e6cc440e5613d6c826288402c66681a5e6ea5fa0/models/i3d.py#L303 In the paper QuVadis Action Recognition this is actuall 2,7,7 not 8,7,7.