Implement the gradient for depthwiseConv2d. about tfjs HOT 5 CLOSED

Let's take conv2d since it's very similar to depthwiseConv2d. Here are the lines that specify forward mode and gradients for conv2d. That op then uses 2 kernel/backend ops specified here.

DepthwiseConv2d is analogous but it's missing the gradients. You'd have to add depthwiseConv2dDerInput and depthwiseConv2DDerFilter to the backend. The math should be simpler than the conv2dDerInput and conv2dDerFilter. Hope this helps!

from tfjs.

Status on this? I'd really love to be able to calculate gradients for doing saliency maps.

from tfjs.

Is someone working on this? I'd like to get the gradients for MobileNet to perform finetuning in the browser and visualize features on the original image like Deepdream does.

After some digging through, I confirm the need for this ticket as part of my research and when I printed out filteredTape in Engine.prototype.gradients(), I also see some undefined layers interlaced between consecutive pairs of reshape ops (see pic). @nsthorat I see in #108 (comment) that the team is working on this, so I was wondering if I could be pointed somewhere.

Also, this comment is a followup to #153. Nikhil you said

We don't have gradients defined for depthwise convs, so MobileNet gradients should have worked in the old way.

I'm not sure what you mean by the old way. If you're referring to getting the gradients, that's precisely what we want. How do you go about it?

from tfjs.

You need the gradients for depthwise conv in order to analyze MobileNet's gradients. We will add this early next week. Stay tuned. In the meantime if anyone else is interested in taking a stab at this, I'll be happy to review.

from tfjs.

Gotcha. If I wanted to take a quick jab at it, what are some files I should specifically look out for?
My project deadline is next week so I'd like to have it added by early next week. I'd love to help given some pointers!

Thanks,
Jason

from tfjs.