Aditya Oke

``` train_transforms = T.Compose([T.ToTensor(), T.Normalize((0.5,), (0.5,))]) valid_transforms = T.Compose([T.ToTensor(), T.Normalize((0.5,), (0.5,))]) train_set = datasets.CIFAR10("./data", download=True, train=True, transform=train_transforms) valid_set = datasets.CIFAR10("./data", download=True, train=False, transform=valid_transforms) train_loader = DataLoader(train_set, 32, shuffle=True, num_workers=2) valid_loader...

We can't force end users to has `self.classes` attribute. Rather now I propose a parameter called `metrics`. Which users can pass.

We need discussion on metrics parameter and how it should work, should It work for all models ? If yes then how.

It's high time now we do this. If CI time reduces by 3 mins per build we save lot of time in a matrix build like ours.

Yes ! I was about to tell about these. We include `weights` and `reduction`. Reduction can be `none`, `sum` or `mean`. P.S. We also keep consistent with torchvision and if...

@zhiqwang you are welcome to contribute 😄

Join on Slack [here](https://join.slack.com/t/quickai/shared_invite/zt-iz7tqk3r-IQa4SoxJGIK5WS8VdZhzeQ) All the development talks go here ! You can freely communicate your ideas, RFCs and thoughts.

Each loss in separate file is better This keeps abstraction minimal. If We could create a folder called `nn` but that would interfere with `torch` which we can avoid. Also,...

If we call our folder as `nn` people will have to do `from quickvision.nn import loss_fn` While `nn` is something torch uses whose code is bonded to C++ API. So...

@zhiqwang @hassiahk I think we can start with this for release 2.0, Let;s propsose a template API for this.