Pranay Agrawal

[Pranay Agrawal @codepragrr]

Im switching my progress updates to an alt account as an experiment.

goodbye.

[Pranay Agrawal @codepragrr]

#100DaysOfCode Day 30:Checked whether the model relied on a curriculum for learning for higher number of hops and the answer is no as I reran the experiment removing the hop =1..4 data and it still generalized.Set off a 20000 overnight epoch training run to see if grokking occurs

[Pranay Agrawal @codepragrr]

#100DaysOfCode Day 29: The model successfully extrapolated till hop=5 with no end in sight, ensured a fixed cycle length and proper train test split involving unseen arrays to genuinely test multihop reasoning.

[Pranay Agrawal @codepragrr]

#100DaysOfCode Day 28: The model was able to extrapolate the pointer chasing algorithm till hop=3 but was struggling beyond that. Will diagnose the issue and solve it.

[Pranay Agrawal @codepragrr]

@DoGYS_tutorials Crazy work 🔥, always inspired by your work ethic.

[Pranay Agrawal @codepragrr]

@DoGYS_tutorials @IGN lmfaoo

[Pranay Agrawal @codepragrr]

#100DaysOfCode Was supposed to finish the datagen work today but got distracted, will continue tomorrow.

[Pranay Agrawal @codepragrr]

#100DaysOfCode Day 27: Finished synthetic data generation for my multihop reasoning project.

[Pranay Agrawal @codepragrr]

#100DaysOfCode Day 26: Continued with the design of the experiment to test whether neural nets can do multihop reasoning by training on the pointer chasing algorithm. The design is 80% done need to iron out a few small permutation details.

[Pranay Agrawal @codepragrr]

#100DaysOfCode Day 25: Started @karpathy's GPT video to implement a transformer from scratch. Implemented till Bigram model.

[Pranay Agrawal @codepragrr]

#100DaysOfCode Day 24: Started on the experimental design of my next project to see if neural nets can do multihop reasoning.

[Pranay Agrawal @codepragrr]

#100DaysOfCode Day 23: Finished the autoencoder experiment.

[Pranay Agrawal @praggr]

1/4🧵

How much can you compress an image before it falls apart?
I trained an autoencoder on CIFAR-10 and pushed the compression until things broke.

[Pranay Agrawal @codepragrr]

#100DaysOfCode Day 22: Worked on and finished the UI Demo for the autoencoder.

[Pranay Agrawal @codepragrr]

#100DaysOfCode Day 21: Implemented learned downsampling instead of basic maxpooling to further improve the results and ensured that the prelatent size. For 6x compression, getting a PSNR of 30 now which is near indistinguishable from OG unless u look really closely.

[Pranay Agrawal @codepragrr]

#100DaysOfCode Was engaged in the past week in college work, will restart tomorrow or day after.

[Pranay Agrawal @codepragrr]

@TheJokerHut The UI looks really clean. Great work!!

[Pranay Agrawal @codepragrr]

#100DaysOfCode Day 20: Was able to improve the psnr results by reducing the downsampling that I had done before the latent. At 12x compression I'm getting a psnr of 26 now.

[Pranay Agrawal @codepragrr]

#100DaysOfCode Day 19: Finished implementing the autoencoder and was able to get a PSNR of 23dB with 12x compression. But increasing latent size didnt improve PSNR, will check.

[Pranay Agrawal @codepragrr]

#100DaysOfCode Day 18: Started implementation of a basic autoencoder to reproduce CIFAR dataset.

[Pranay Agrawal @codepragrr]

#100DaysOfCode Day 17: Started learning about autoencoders