#7 PyTorch Book Thread: Sunday, 10th Oct 8AM PT

Note: This is a wiki, please edit it & add resources!

:tv: YouTube Link: PyTorch Book Reading - 7. Training a Tumor classifier - YouTube

Hi y’all!

What to expect for #7 of the PyTorch Book Reading Group: We’ve just learned how to grab the LUNA 16 dataset and make it work with torch datasets. In this meetup, we’ll train a classifier on top of this.

:link: Link to sign up

:books: Resources:

3 Likes

Q: Would you be interested in being a part of fastai ML author group?

12 Likes
2 Likes

apologies for asking a obvious question. What is fastai ML author group?

Fastai has a machine learning course that last ran in 2018 (and may be a little outdated). Some of us are planning to reboot it and make it work in 2021 :slight_smile:

4 Likes

one output (or channel) for each filter (or kernel) applied

3 Likes

can you share this excel sheet?

1 Like
3 Likes

I think accuracy isn’t a good metric for highly imbalanced datasets.

2 Likes

In an imbalanced dataset, we have many many more data points of one category compared to another?

3 Likes

because 99.999% of sample belong to one class (benign nodules)

2 Likes

when %of one class is quite more than compared to other. Say 99.99% data belongs to 1 class and just 0.01% to other.
If u build a model that just predicts the majority class, in that case our accuracy will be already 99.99% but we know that this model is not good.

2 Likes

Because of Class imbalance

1 Like

Is it a good idea to use fewer samples from the majority class, and heavy augmentation with the minority class in an attempt to balance a dataset?

1 Like

It do work sometime, but it will depend on your data. But in general it’s not a good idea to drop the labelled data(unless it’s the only option left)

2 Likes

Would making a custom loss function also fix this, like making wrong classification of a bad tumor more expensive in the loss?

2 Likes

I guess, the CrossEntropy Loss should help here . My understanding is it will penalize wrong prediction with high confidence quite heavily . Thoughts?

2 Likes

Could you post the link to the sources or that paper specifically?

I think cross-entropy will penalize wrong predictions equally - if a scan has a malignant tumor but the model says everything’s fine, and if a scan has a benign tumor but the model says it’s malignant.

What @meerio meant (I think) is that if a scan has a malignant tumor and the model says it is benign, then it should be penalized more compared to if a scan has a benign tumor but the model says it’s malignant.

3 Likes
3 Likes