YH Michael Wang

My key takeaways from Stanford XCS236, from likelihood-based modeling to diffusion and model tradeoffs in practice.

I did some experimental work in April, where I build bindings for Nvidia's raytracing engine PyOptiX to support writing device kernel in python. Original post: https://developer.nv...

Point in polygon test is one of the most fundamental building block in ray tracing. Given a point on a surface and a vertices array representing a polygon, is the point in the poly...

This is a series of notes taken from the . Types are essential components of a programming language. From practicality, types separates the definition of different categories of da...

This is a series of notes taken from the . From our of higher order functions, functions can be evaluated, stored and passed as arguments just as other variables. Storing a functio...

This is a series of notes taken from the by professor Dan Grossman. In this note and next one, I will discuss higher-order functions and their applications. One key feature that di...

This article introduces the Pointer Generator Network and Coverage Mechanism. After the seminal paper from Sutzkever in 2014 and Bahdanau in 2015, before Transformer shook the grou...

Fairseq distributed training is largely built on top of the distributed training feature provided by . A couple important notes from their tutorial that will be useful: 1. The exam...

This is a 2 part tutorial for the Fairseq model BART. In the first part I have walked through the details how a Transformer model is built. Please refer to . A BART class is, in es...

BART is a novel denoising autoencoder that achieved excellent result on Summarization. It is proposed by FAIR and a great implementation is included in its production grade seq2seq...

So I wrapped my Tiger type checker assignment. It's 973 lines of code as of first submission. And it's not even styled in the most readable way. But anyway, I hope only minor modif...

Tensorflow Dataset defines a unified wrapper for all kinds of dataset handling. Traditionally, to process a dataset, one needs to download a dataset, process them into the type tha...