balgot / ib031-cifar10 Goto Github PK

Create new model which will merge some categories to one (let's say <dog, cats> <ships, planes> ...) based on some metrics, achieve high accuracy on those categories and then will run sub-model to differentiate between merged categories

Remove unnecessary opening of file, especially if there results range is small -> in this case 10

Detect edges via color gradients

Write general information about the project to README, so once it's public, it is clear what we are doing

I think we should notice:
"If you're going to use this dataset, please cite the tech report at the bottom of this page."

We shall measure efficiencies and avoid overfitting via learning curves. Also add these plots to final notebook. Might be interesting to create ROC curve

We might achieve greater results by merging common categories (planes, ships), (dogs, cats), (cars, trucks) and then training new model to decide them

Take argument for passing just one label, so we don't have to create whole new array af the same labels

Load all files directly from source rather than loading them by batch

It may be more readable to print average image and then a bunch of outliers than having to scroll up to find the corresponding average image

It is expected that after calling init(), the cache/ directory is in the same state always - i.e. empty

https://github.com/MichalBarnisin/ib031-cifar10/blob/48830655a7ee09d310bdaeaf20cd12bfbd55dca6/cache.py#L64

As said on the first-review, we shall work with hsv representations of pictures