Fastai v2 chat

fastai fastai dev
1032

You might be interested in other approach for Siamese. One samples pairs of the same class in the dataset, pack in batches and negatives are mined in the batch. No specific class needed.

I have implemented it for patch descriptor learning:

3 Likes
1033

Btw, is there a way to skip computing validation loss?

1034

You can raiske a CancelValidation error (canā€™t remember the exact name but you should find it easily) that will skip the validation phase.
If the questions is skipping loss but still computing other metrics, the answer is no however.

1035

I always use

pip install git+https://github.com/fastai/fastcore.git --upgrade
pip install git+https://github.com/fastai/fastai2.git --upgrade

which is the better approach?

1 Like
1036

I donā€™t know, I just added the branch at the end.

1 Like
1037

Skipping loss, but do val. Ok, Iā€™ll do my callback then

1038

So youā€™ve written imshow_torch to work with that, thatā€™s pretty cool !

I wonder what would happen if you call _pre_show_batch on your dataloaders :thinking:

1039

I was running lr_find and I see this printed

@log_args had an issue on LabelSmoothingCrossEntropy.__init__ -> missing a required argument: 'self'

Itā€™s not throwing any error though, anything to be concerned about?

1 Like
1040

I think that you need to install it again, they updated something and need last fastcore.

You should try to install it again:

pip install git+https://github.com/fastai/fastcore.git@master
pip install git+https://github.com/fastai/fastai2.git@master
1041

I did, this happened after I updated it. Before installing the latest one I was getting an error.

1042

Itā€™s a warning you can ignore, we are working on fixing those.

1 Like
1044

Can you share the notebook?
Even if it does not happen anymore Iā€™d like to check it works as intended and you may have an edge case.

1045

Here is a minimal example: https://colab.research.google.com/drive/1S1HyJs7a0ehwkltsBjGvRLDPSeOcXm0u

1 Like
1046

Quick qeustion, is there a Resize (pre size) images method on V2? To resize all images before running the trianing?

1047

No, there is no utility for that.

1048

Thanks, I found the issue and am sending a PR. This should not affect your program. This is mainly used by WandbCallback.

1 Like
1049

Does everything in the CONTRIBUTING.md for fastai v1 still apply to v2, or are there important differences?

1050

A post was split to a new topic: Questions on nbdev

1051

I only have labels for my validation data, how should I build my dsets? I tried doing the following:

tls_raw = TfmdLists(fns, [lambda o: o.read()], splits=splits)
tls_train = TfmdLists(tls_raw.train, [tkzer, nmzer], split_idx=0)
dset_valid = Datasets(tls_raw.valid, [[tkzer, nmzer], [parent_label]], split_idx=1)

tknzer and nmzer are just instances of Tokenizer and Numericalize

The problem is that dset_valid is still considering itā€™s items to be part of the training set

At the end I would like to have a single Datasets object, where if I do dset.train[0] it returns only an item and dset.valid[0] returns (item,label)

Itā€™s not necessary to follow this TfmdLists approach, this is just my best idea to solve this so farā€¦ Iā€™m open to new easier ways :smile:

1052

If you have such different structures for training and validation, you canā€™t have them in a single Datasets objects. When you are finished and independently created the DataLoaders, you can combine them in a DataLoaders.

1 Like