Aaron Briel comments

Results 16 comments of


                                            Aaron Briel

Bump numpy from 1.19.5 to 1.21.6

Hello @melindaloubser1 , I'm wondering if this is going to be addressed soon, as I believe that it may resolve issues with running on the Mac M1 (see [here](https://stackoverflow.com/questions/65336789/numpy-build-fail-in-m1-big-sur-11-1) :...

Results with custom dataset

I'll assume that the low scores overall, as similar to what was stated in the paper, could be related to the complexity of the question/answer pairs combined with the large...

Pretraining not generating added tokens file

It might be best to submit single issues at a time and label them appropriately as bug/enhancement/help wanted, etc.

Using multiple training instances in AWS Sagemaker.

@nectario You would set this in the initialization of BertDataBunch.

Accessing BERT encoding layer output.

You can access them in the model within fast-bert.

ValueError: Target size (torch.Size([9, 4])) must be the same as input size (torch.Size([12, 4]))

There's not enough information here to pursue this.

New issue with bert tokenizer

This is not enough information. Kindly provide steps to reproduce the issue.

ERROR: 'NoneType' object is not iterable error during loading of training data

Is there something unique about example 10000? I wonder if there are any null examples or ones with non-parsable characters?

i write a model for single class, but get y_pred like this [0.4,0.5]

I think this is because y_pred contains values for pos and neg. I think you would thus need to customize the fbeta by adding logic to convert y_pred to 'pos'...

Does the multi-label classifier optimize the entire model or it freezes BERT?

I noticed the freeze/requires_grad logic is commented in learner_cls. @kaushaltrivedi how does one freeze the bert layers and only train the added custom layer?