Log multiple losses

[Landanjs]

Resolves CO-842. Adds logic to log multiple losses. Also, adds loss dictionary to DeepLabv3+ as an example. A couple of scenarios:

1. Scalar tensor loss

This will be logged as loss/train/total. Bert example:

2. Tuple of tensor losses

Individual losses will be logged as 'loss/train/loss{i}' where i is the index of the individual loss. There will also be 'loss/train/total' which is the sum of the individual losses.

3. Dictionary of losses without total key

Individual losses will be logged as 'loss/train/{loss_name}'. There will also be 'loss/train/total' which is the sum of the individual losses.

4. Dictionary of losses with total key

Individual losses will be logged as 'loss/train/{loss_name}'. Loss given at 'loss/train/total' will be used as the loss for the backpropagation, individual losses are not summed when 'total' is present.

Questions

• Assumes returned loss is a torch.Tensor, is this fine?
• Should these different scenarios be documented somewhere?

[abhi-mosaic]

Is there any way we can get rid of the trailing / when there is only 1 loss value, like loss/train instead of loss/train/ ? I'm just wondering if there could be some parsing issues later on.

[EDIT] or wait maybe it has to be that way for WandB to put plots together in the same place? Hmmm
I wonder if we can add a default like loss/train/total that always gets logged... would match the notation a bit better too.

Also, I wonder if we can drop the _loss suffix for things like loss/train/dice rather than loss/train/dice_loss, as it seems a bit repetitive.

[Landanjs]

Thanks for the suggestion Abhi, good points! Yeah, that would be cleaner, let me try to set it up 🙂

Hmmm that might actually make the closure stuff easier as well!

[eracah]

''

[mvpatel2000]

I don't understand pyright same :')

[abhi-mosaic]

This looks great to me!!

@Landanjs I'm assuming you have some convergence tests on DeepLabV3 that work? And hopefully we have some trainer tests that do loss as a Tensor, tuple, dict?

Approving now for velocity and trusting you on the tests :)

[Landanjs]

I added some convergence tests for tuple losses and dict losses with 'total'. TBH I do not know how to make logging, but I'm running DeepLabv3 and ResNet experiments (wandb here group by group). The results appear to be align with dev.

[Landanjs]

I mentioned this to Evan, but just for the record, it looks like i need someone from @mosaicml/composer-team-eng to review as well