Problem with provided weights and Source Code

[elkleis]

Hi, I recently downloded the weights provided here, and found some inconsistencies.

• First, the best weights seem to be for a slightly modified ResNet18 model, and not the ResNet50 as stated.
• The channels of the image are 2 but on the Source Code are left as 3.
• This Error comes up : ValueError: Dimension 1 in both shapes must be equal, but are 1 and 4. Shapes are [8192,1] and [8192,4]. for 'Assign_106' (op: 'Assign') with input shapes: [8192,1], [8192,4].
  As I understand, this is refferred to the outputs of the model. But why is it 4 if we're only expecting steering prediction?