#5 PyTorch Book Thread: Sunday, 26th Sept 8AM PT

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

:tv: YouTube Link: PyTorch Book Reading - 5. CT Scans & PyTorch - YouTube

Hi all!
This thread is for discussing, Q&A, and everything else for the #5 meetup of the book reading group.

This week we’re officially wrapping up Part 1 of the book and will start diving into an end-to-end project using Part 2! We’ll start understanding what CT Scans are and how to work with them using torch.

:round_pushpin:Link to sign up


<<<< Previous Session Thread


What is a CNN in <15 words?


A CNN is a Deep Learning algorithm that can take in an input image, assign importance (learnable weights and biases) to various aspects/objects in the image, and be able to differentiate one from the other.


Convolutions or traversing of a kernel matrix over an image in order to highlight key features.


convolutional NN is a neural network architecture mostly used to capture the spatial information in data, mostly used for images


how to design a cnn based on my own requirements?

Is it possible to change the convolutional kernel to other shapes? For example, instead of this cross shape of 1s, can we choose a different shape? see kernels types Kernel (image processing) - Wikipedia

excel-conv from fastai courses/conv-example.xlsx at master · fastai/courses · GitHub


okay, i will learn,thx

values in kernel matrix is arbitrary values, as we will have number of kernels in a filter?

I always start with a Resnet too. But this is an interesting question, I think it’s good to explore and build intuitions on different architectures, ex. when to go for mobilenet, resnet etc… Are there any resources on this? I’ll try to explore too.

Yup, we can do that but it’s going to be a bit tricky and will require little bit of an extra effort.
You can check this answer from ptrblck on Pytorch Forum

Why do many models stick with the 3x3 kernel? I was looking into action recognition (video data), and even there they cite LeCun’s approach

You can get more details here




Nice. This is very helpful.

1 Like

Higher dimension kernels mean more parameters and the numbver of parameters increase exponentially. Also seeing the receptive field compunds as we go down deeper in the network, 3 x 3 kernels do the job well so we may not necessarily need to use larger size kernels.

However, one can always experiment and find new things emerge :slight_smile: !


i think that Resnet 50 is okay, i often apply it when i am working on object detection, i want to explore whether i can revise them or not with my own project or my own custom dataset !


Goto paperswithcode OR


1 Like