Add KuzushijiMNIST example

[goodhamgupta]

Hi everyone,

Thanks for this excellent library! This PR aims to add an example for training the KuzushijiMNIST dataset using Axon(#47). The file is almost an exact replica of mnist.exs, with the only difference being the Scidata version used and how we transform the images and labels.

Training the model locally on my laptop, I get the following metrics:

--------------------------------------------------
                      Model
==================================================
 Layer                    Shape        Parameters
==================================================
 input_8 ( input )        {nil, 784}   0
 dense_11 ( dense )       {nil, 128}   100480
 relu_12 ( relu )         {nil, 128}   0
 dropout_13 ( dropout )   {nil, 128}   0
 dense_16 ( dense )       {nil, 10}    1290
 softmax_17 ( softmax )   {nil, 10}    0
--------------------------------------------------


Training Model

Epoch: 4, Batch: 1750, Loss: 0.52946 Accuracy: 0.97277

 Testing Model

Epoch: 0, Batch: 100,  Accuracy: 0.98372

Thanks!

[goodhamgupta]

I accidentally closed this PR. Sorry for the notification noise! 😓

[seanmor5]

Thank you @goodhamgupta for the PR! I want this example to get merged, my concern is that it is too similar to the original MNIST example and does not demonstrate or highlight a unique feature of Axon. I think we can fix that though :)

I think we can adapt one of the examples in: https://keras.io/examples/ using this dataset and Axon. At a quick glance, https://keras.io/examples/vision/metric_learning/ and https://keras.io/examples/vision/near_dup_search/ and anything in https://keras.io/examples/generative/ seem very do-able for now. Or if you find another example you'd like to do, but you aren't sure how to make it work in Axon, I'd be glad to help walk you through putting the example together in this PR! Just let me know what looks interesting and we can get something up and running

Thanks again for the contribution!

[goodhamgupta]

Thank you for the kind review @seanmor5! I fully agree that this PR is a copy of the MNIST example for now and doesn't provide any new insight into the features of Axon.

I want to implement the metric learning example for similarity search(https://keras.io/examples/vision/metric_learning/). Would you prefer I close this PR for now and open a new PR once it's ready, or should I continue on this PR?

[seanmor5]

You can continue working in this PR! Let me know if you run into any problems and I'll be glad to help!

[goodhamgupta]

Hi @seanmor5,

Sorry for pinging you directly, but if you could provide any hints on how to proceed from my previous comment it would be super helpful! 😅

[seanmor5]

Hi @goodhamgupta I am closing this for now as stale, if you want to re-open and continue work feel free and I will do my best to guide you in the right direction :)