Week 1 roundup

Hi folks,

I thought it might be worth doing a quick round up of what we have all been up to and what issues we have overcome or been stopped by. I know this topic is covered at length in the main forum but I thought there might be value in having a smaller thread run within our group.

I’ll kick it off.

I’ve been working on my first ever Kaggle submission by following Radek Osmulskis work over at https://github.com/radekosmulski/quickdraw

While I’ve been plagued by technical issues I’m now working on my second submission with my third and fourth already planned out.

For the next week, in addition to whatever the lecture throws at us, I need to learn how to use the Paperspace Gradient Job builder. The ability to leave a big/long learning cycle to run without having to worry about starting/stopping the notebooks would be invaluable going forward.


Nice one @timbo72 .

I tried training resnet-based models to classify my own hand-curated images of different types of galaxies, but even with Jeremy’s help the best I can do is 75% accuracy. Let me know if you’ve any bright ideas on how to improve further!

There’s a much more involved Kaggle competition that I’d like to attempt in the future, but need to understand multi-label classification before I do. I believe that’s coming up later in the course, so looking forward to that.

Ben I’ve had a look through your notebook (which is great by the way) and I have a weird idea.

remember those slides of the first few layers of resnet how they were just dots, tiny lines, gradients, etc… (I’m sorry I can’t lay my hands on the slide right now but you know the one i mean right?)

Looking at your galaxy images they are all exactly the same on a granular level, i.e. all stars. perhaps you could run them through some sort of filter and see what you can get. I’m thinking photoshops fast masks, unsharp masks, gradients, etc… I’m sorry I don’t know enough of the image modification libraries in Python to suggest one.

Nice one @timbo72 and @followben .I have been also been playing around with the starter code of quick draw by Radek. Looks a really interesting problem with so many possibilities on how to tackle it.

1 Like

@followben I think 75% is pretty good. The challenge here I think is the dataset. The images are taken at different focal lengths. It is going to be difficult to learn that a barred galaxy at focal length X is the same as a barred galaxy at focal length X+Y where Y is large. This will be particularly problematic when other galaxies at focal length X look like barred galaxies because there is very little information (pixels) at short focal lengths that help the network discriminate between the images. If you amended the dataset so the focal lengths are consistent (by removing images with short focal lengths) my guess is the accuracy will improve.

I have been trying this kaggle competition (https://www.kaggle.com/c/pubg-finish-placement-prediction) using a tabular model and random forest. At the moment, Mean Absolute Error is around 0.06, which I need to improve :sweat_smile: Hoping the lesson on structured data would help with this. I’ve only used the basic steps to train the model (https://docs.fast.ai/tabular.html) so any advice would be appreciated.

1 Like

Hi all, hope you are having a good weekend. I wanted to start to read some deep learning papers (say from AlexNet) and have an attempt at blogging to see whether I could understand and explain some aspects from each paper. I would truly appreciate your feedback and advice.

Edit: I tried to train ZFNet from scratch but failed to do so (https://github.com/YJAJ/Deep_learning_studies/blob/master/ZFNet-babies.ipynb). The dataset has only ~5000 images I guess, but the loss does not go down smoothly like what happened in Week 5 lecture notebook. How do you think I can improve this? Thanks in advance.

1 Like

Looks very interesting, @YJP! I haven’t had a chance to read it yet, but I’ll read it later tonight and share any thoughts back to you that occur to me.