Status

[trevorlapay]

How is everyone doing?

Tasking for me: I plan on working out of the ImageDataGenerator rabbit hole and trying a few premade kernels to see how far I can get.

https://www.kaggle.com/rejpalcz/cnn-128x128x4-keras-from-scratch-lb-0-328/code

https://www.kaggle.com/iafoss/pretrained-resnet34-with-rgby-0-460-public-lb/code

https://www.kaggle.com/wordroid/inceptionresnetv2-resize256-f1loss-lb0-419

I'll be working it on and off this weekend.

What are you guys up to? Anything I can help with?

T

[hankyusa]

I was trying to put together the ensemble of 28 networks that each predict one label. Eventually I realized that this way of going about things is extremely inefficient. Today I was trying to decide if I should plow through it to show that it's a bad idea or switch to premade kernels like is being suggested on the discussion boards. What do you think?

[trevorlapay]

I think we should chase down whatever is the most interesting for us. I think we can get to 30% using 2D CNNs (I'll verify that) but I think it will be a monumental task to get to >50%. I'm happy with either of your approaches. We can document it in the report regardless.

Since we did so well on the projects (and I assume on the exams) I'm not sweating too hard about getting accuracy points. I'd love to do well, though.

On Nov 30, 2018 5:00 PM, "Luke Hanks" notifications@github.com wrote:

I was trying to put together the ensemble of 28 networks that each predict one label. Eventually I realized that this way of going about things is extremely inefficient. Today I was trying to decide if I should plow through it to show that it's a bad idea or switch to premade kernels like is being suggested on the discussion boards. What do you think?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/trevorlapay/UNMFall18_ML_Project4/issues/3#issuecomment-443375990, or mute the thread https://github.com/notifications/unsubscribe-auth/ALDXesN5MzHziu0asghYrK8PIe5uumxuks5u0cafgaJpZM4Y8R8P .

[trevorlapay]

[trevorlapay]

Not great, but a decent start?

[trevorlapay]

I should mention that I used this kernel:

https://www.kaggle.com/rejpalcz/cnn-128x128x4-keras-from-scratch-lb-0-328

[trevorlapay]

[hankyusa]

Wow. That's pretty good for this project.

Regarding the importance of this project on our grades. My grade is greater than 115% thanks to the excessively generous undergraduate grade boost. I could probably skip this project and still get an A. So the only reason I'm working on it at all is because I don't think it's ethical to leave you hanging when you don't have the grade boost. Basically, I'll put into this project exactly as much as you want me too. Otherwise, I need to focus on the classes that I'm not guaranteed to ace.

I'll just get this ensemble of binary classifiers working. It will probably do terribly, but I'll learn something about simply putting it together. I'll play with neural networks for real after I graduate.

[trevorlapay]

Luke, I'm really in the same boat. I can't afford to quit the project, but I would need to get a 50% for it to ruin my A+. I'm doing this mostly because I'm curious how well we can do. So, if you have more pressing stuff, don't sweat this. However, if you have any tips or ideas, I'd love to hear them.

I'm running my updated kernel using Nadam, a set of image augs, and a number of other changes to see what happens. I am getting decent loss reduction on 100 epochs over a size 32 batch (I can't fit much more than that in memory, especially if I do augmentation). I dont expect it to work a miracle, but I'm curious to see if we get anything out of the augmentations (Flip, contrast, gaussian blur, others).

If this doesn't buy us anything, I'm going to try and use some out of the box regularizers and tinker with the model. I think the model in this kernel is just a random set of layers and has a lot of room for improvement. After that, I'll go for out best model at that point and do the report.