PyTorch Course (2022), Part 4: Image Classification (MNIST)

In this video we use the network constructed in the previous video to train a neural network on the MNIST data set. The goal of this network is to take in images of hand written digits, and predict what digit they correspond to.
Code:
github.com/luk...
MNIST Data
Discord:
/ discord
PyTorch website:
pytorch.org/ge...

[@lomash_irl]

What a hidden gem! This video isn't just about image classification, it also explains how to use datasets, dataloaders, and create training functions, step by step. I highly recommend watching the first video in this series on tensors.

[@Iwillseeyouagainin25years]

Great series! In case anyone in the future has a problem as me: If you run into the error message cannot import name 'PILLOW_VERSION' from 'PIL' when importing torchvision, update torchvision to a version >=0.5.0. Alternatively, downgrade pillow < version 7. Pillow changed 'PILLOW_VERSION'' to '__version__'.

[@cianjones2728]

I find this series amazing. You're a very good teacher! Are you planning on continuing this series in the future?

[@rankinstudio]

Hey man. Just wanted to say thanks. After watching this about 100x I was able to successfully build my own image dataset and get this working. Great tutorial, and excellent explanations. Looking forward to another installment and hearing more about your research. Cheers

[@panoskotoulas759]

That's a great series. Seeing both theory and practice at the same time is awesome and makes it easier to understand the process. Really hope you'll continue the series. I know it's too much work for little views but I people really benefit from watching these videos, they're top quality.

[@DEEPAKSV99]

Your teaching is so good!! You really deserve way more views & subscribers. I was able to find this tutorial series since I specifically searched for a Pytorch tutorial with the latest version. Wonder why doesn't the youtube algorithm put your video at the top. Also would you consider making a Part-5 of this series on building CNN with Pytorch?

[@963shots]

Please continue whenever you get time out of your busy schedule sire!

[@jakstrike1]

Top vids! Got hooked on the memes but this is good pytorch intro content too coming from tabular data ML background.

[@kaba1996]

Wow I was thinking why we're using vectors in programming Thanks great video

[@christinecannon6811]

77

[@propjoe8296]

My go-to channel when I need the basics of NN💯🎉. Please can make a playlist on CNN and RNN.

[@oludelehalleluyah6723]

You've stopped this series. I am enjoying it.

[@cvicracer]

Your explanation is very easy to follow, thank you very much.

[@yamanarslanca8325]

Thanks for the video. Around @4:00 you used one-hot encoding and I understand the concept. But I wonder a case where you had too many categorical data (for example 100 different numbers in your target) so your num_classes would be higher than 10 (a lot higher). Then if you apply one-hot encoding in this case you would increase your feature space dramatically and I assume this would lead to "curse of dimensionality" or something similar that which makes your model's performance worse. Do you know how to deal with such situation ? Also, I know this is a noob question but I just want to be sure, when you initialize your model (model=ModelClass()) initial parameters of the model (initial weights and bias) is automatically and randomly assigned by the -torch.nn.Module- or by optimization function -SGD(f.parameters())- ? Thanks.

[@KapardhiMarla]

whatever be the categories the dimension of the tensor or in this the vector remains 2

[@rishidixit7939]

At 24:00 when reshaping isn’t the row major order going to affect how rows are made from the original data ?

[@RonivaldoPassosSampaio]

Well done! Very clear explanation !!

[@qualimania]

Hey, great tutorial! I have a question regarding the dataset. For some reason, the link you provided is not working. I wonder if someone else has the same problem. Any help will be appreciated!

[@cur1ousss2047]

Great tutorial thanks

[@aleksnader8853]

thanks for your teaching

[@mrmkl9839]

If I increase my batch size the predictions (and the loss) gets worse and worse. Why can that be the case?

[@aliabasnezhad7872]

Can you explain, when defining CTDataset, why you didn't use super? Thanks!

[@sivuyilesifuba]

Goated explanation

[@HimanshuGauttam]

The given code segment L(f(xs), ys) (line number 63) produces the error message. RuntimeError: 1D target tensor expected, multi-target not supported

[@joshmarion8640]

glad im not the only one. Did you ever find a solution?

[@ИванСергиенко-б9з]

thank you very much for your tutorial. it's very useful. Could you pls explain, how can i test this model with random img? problem is that img.shape is ([28, 28, 3]) with color channel. how can i remove 3d dimension?

[@sould3mon271]

hopefully it's to late and you figured it out already but you can either convert img.convert("L" ) if loaded with PIL.image before converting to array or leave as is and use img[:,:,x] as slice method where x is one of the channels RGB and R is 0 sadly i have had no luck with it getting it to recognise own made test samples at all, only the testdata works for me and can't figure out why tried all methods of image conversions on multiple digits. realy wished he would have showed us an own made sample

[@sinchanmaity8825]

Thank you brother

[@danl8994]

Thank You so much

[@ekashama]

Is it real to do this with softmax?

[@nicolasmagee1780]

Please make more :)

[@himanshuchaudhary5796]

How can i contact you

[@trafo222]

discord

[@bloozism]

Billy?

[@himanshuchaudhary5796]

I had been trying to contact with you, I am working on few project in Cosmology. I want to to regarding about those topic

[@officiallyaninja]

he has a discord server

[@MrPSolver]

Messaged you on discord

[@masoudazadkhah4153]

you are the best !