20.03.2021 Views

Deep-Learning-with-PyTorch

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Summary

441

14.9 Exercises

14.10 Summary

1 Implement a test set for classification, or reuse the test set from chapter 13’s

exercises. Use the validation set to pick the best epochs while training, but use

the test set to evaluate the end-to-end project. How well does performance on

the validation set line up with performance on the test set?

2 Can you train a single model that is able to do three-way classification, distinguishing

among non-nodules, benign modules, and malignant nodules in one

pass?

a What class-balancing split works best for training?

b How does this single-pass model perform, compared to the two-pass

approach we are using in the book?

3 We trained our classifier on annotations, but expect it to perform on the output

of our segmentation. Use the segmentation model to build a list of non-nodules

to use during training instead of the non-nodules provided.

a Does the classification model performance improve when trained on this

new set?

b Can you characterize what kinds of nodule candidates see the biggest

changes with the newly trained model?

4 The padded convolutions we use result in less than full context near the edges

of the image. Compute the loss for segmented pixels near the edges of the CT

scan slice, versus those in the interior. Is there a measurable difference between

the two?

5 Try running the classifier on the entire CT by using overlapping 32 × 48 × 48

patches. How does this compare to the segmentation approach?

• An unambiguous split between training and validation (and test) sets is crucial.

Here, splitting by patient is much less prone to getting things wrong. This is

even more true when you have several models in your pipeline.

• Getting from pixel-wise marks to nodules can be achieved using very traditional

image processing. We don’t want to look down on the classics, but value these

tools and use them where appropriate.

• Our diagnosis script performs both segmentation and classification. This allows

us to diagnose a CT that we have not seen before, though our current Dataset

implementation is not configured to accept series_uids from sources other

than LUNA.

• Fine-tuning is a great way to fit a model while using a minimum of training data.

Make sure the pretrained model has features relevant to your task, and make

sure that you retrain a portion of the network with enough capacity.

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!