Closed codemonkey9000 closed 6 years ago
We need to remove this:
and could also be read as
Luminance(84, 84, 4)
for 4 layers of an 84x84 image.
We should add proper support for multichannel convolution. We haven't. The dimensions of that image are a cheap hack (appending the images along one axis so all four images form one weirdly shaped image) and it's really bad. You can point out why the image is such an odd shape, but we can't say 'could also be read as' because there are fundamental differences. This should be fixed, and if anyone knows what they're doing deep learning wise it will look ridiculous and deeply concerning. Sadly, this has never been a product priority and it isn't on the roadmap to address.
Done! Look okay now though? Anything I should add?
Added description to explain how Luminance works. Looking for input on how to improve this.