(doc) needs help on how to deal with batch data

[qlibp]

In run_tensor.py, how to implement the forward function to handle batch data? As the final result of Network.forward should return Tensor data, so the method in Module 1 using for and list is not a choice. For now, I merely figure out one possible solution: add one more dimension to the dataset. For dataset of shape (50, 2), I will change it to (50, 1, 2) and then using the broadcasting mechanism to pass it through network.

Is there a better way to implement it?

[srush]

You are correct that this is the expected method to implement it => View, then Broadcast, and then reduce. I will add some hints to make that more clear.

[srush]

I will add some more docs here, in the mean time, this diagram from the next section explains the main approach https://minitorch.github.io/matrixmult.html

[qlibp]

You are correct that this is the expected method to implement it => View, then Broadcast, and then reduce. I will add some hints to make that more clear.

Thanks for the confirmation and then I will try to figure out what goes wrong with my code.

[qlibp]

I finally find out that the reason why my training procedure get stuck(derivative vanish at around epoch=60). I miss the relu activation in the Network forward function. I comment here in case someone come up with the same problem which I waste for about a week to find it out.

[srush]

Oh no! I didn't mean for that to be a hard part. I'll add more docs in the network class to describe the exact shape of the network.

Did you eventually get it working? I see why that would be frustrating.

[qlibp]

Oh yeah, I finally make it work(maybe). It's still a little bit different from the result of Module 1.

[srush]

Looks great!