[WIP] Create examples

[monatis]

We can also create some basic documentation based on these examples. It has some bugs now --I'll check it out.

[monatis]

ArcFace loss has a learnable parameter, but it needs to know labels to calculate the loss. So I implemented it as a subclass of GroupLoss, and it should be used with EmptyHead as in train_cifar100.py.

[generall]

Could you please add a link to the source of this numbers?

[generall]

That is a good example for demonstrating ArcFaceLoss, but definitely not the first example a new user should see.

First of all, if I understood right, it is not a fine-tuning, but training of the model from scratch, right?
Secondly, it does not show anything related to model management - saving checkpoints, monitoring metrics, e.t.c.

I guess it is Ok to keep it as is for ArcFace example - in my mind each example should demonstrate only a single feature on top of minimal functionality. But we would need to have another example with very basic dataset and loss, but with proper model management.

[monatis]

definitely not the first example a new user should see.

Agree. An example with Contrastive Loss or Triplet Loss (not implemented yet) would work better as a starter.

but training of the model from scratch, right?

It's fine-tuning of a pretrained model from torchvision, but ArcFace can also be used for cold starts (i.e. pretraining from scratch).

But we would need to have another example with very basic dataset and loss, but with proper model management.

Yes I'll look into a proper dataset for it. And this example was also for me, well, to better understand the inner working of the library :)

I'll also write a short and step-by-step guide for readme like the following, with each item proceeded by a short code snippet:

1. Subclass quaterion_models.encoder.Encoder, and implement this and that.
2. Subclass quaterion.TrainableModel, and implement this hook and the other.
3. ...