#8 PyTorch Book Thread: Sunday, 17th Oct 8am PT

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

:tv: YouTube Link: PyTorch Book Reading - 8. U-net, Image Segmentation and Image Augmentations in PyTorch - YouTube

Hey again, everybody!

Last week we got a vanilla introduction to image augmentations, this week we’ll take an in depth look, and understand what is image segmentation and learn about the U-net Architecture!

:link: Link to sign up

:books: Resources:

5 Likes

Ravi’s Kernel: PyTorch vs. Cancer | Kaggle

Mrigank’s blog: Understanding ConvMixer (with a simple PyTorch implementation) | by Mri Nath | Oct, 2021 | Medium

6 Likes

When would you rather use Segmentation over Classification?

1 Like

Maybe Self-Driving Cars

1 Like

Multi-class classfication

1 Like

I guess we use classification when we care about the presence/absence of an object in an image. And we use segmentation when we care about the exact position of an object in the image.

2 Likes

When we want to also know the location of objects (thanks for the feedback :grinning_face_with_smiling_eyes:) in the images.

1 Like

Medical imaging/ classify the pixels

2 Likes

Basically, classify different objects in one image

1 Like
1 Like

Image Segmentation is primarily used when we want to understand the entire context of various contents in the image. The context here means: accurately identifying the position of different objects in an image.
Use Case:

  1. Image Masking(the one we use in zoom to add virtual background)
  2. Self-driving cars.
  3. Medical Imaging
2 Likes

When would you Segment a tumour? After classification or before?

After classification, because we want to see finer level details where the tumour is present.

3 Likes

My guess is that we should segment BEFORE classification so that we can only send the part of the image that may contain a tumor to our classification model to classify it as benign/malignant.

3 Likes

to train the model we need to segment it before

1 Like

I am thinking, before classification. Maybe we can segment( not sure how? but…) and then easily crop and classify.

1 Like

Based on the pipeline detailed in the book, segmentation is performed before classification to isolate nodules.

2 Likes

UNet has been out for a few years. Do you know of any significant improvements to UNet architecture for segmentation?

Up sampling is basically resizing . Right?

1 Like

Homework convert this to Torch

6 Likes