Simple Model for CIFAR-10 using Flux not converging

Specific Domains Machine Learning
1

I’m trying to train a simple model on the CIFAR-10 dataset.
I’m using the model proposed by the flux documentation on cifar-10 found here: Deep Learning with Julia & Flux: A 60 Minute Blitz · Flux

This is my code.

When training it I got 50% accuracy after 10 iterations but it doesn’t improve much after next iterations. For example after 35 iterations I got around 0.56% accuracy.

How can I improve the model to converge on this simple dataset? I’m doing something wrong?

2

Your training code currently differentiates through the augmentation and I’m not sure it should be. Have you tried removing the augmentation from the loss function?