[Day 76] Finishing the Retrieval-based LM talk, and learning about distillation, quantization and pruning

 Hello :)
Today is Day 76!


  • Finished section 6 and 7 about multilingual retrieval-based LMs and retrieval-based LMs' challenges and opportunities (ACL 2023)
  • Covered lecture 11 of CMU 11-711 Advanced NLP - Distillation, Quantization, and Pruning

Section 6:





Section 7:





Lecture 11: Distillation, Quantization, and Pruning

Problem: The best models for NLP tasks are massive. So how can we cheaply, efficiently and equitably deploy NLP systems at the expense of performance?

Answer: Model compression

Quantization - keep the model the same but reduce the number of bits
Pruning - remove parts of the model while retaining performance
Distillation - train a smaller model to imitate the larger model

Quantization - no parameters are changed, up to k bits of precision

Amongst other methods, we can use post-training quantization

We can binarize the parameters and activations 

Pruning - a number of parameters are set to zero, the rest are unchanged

There is Magnitute pruning (Han et al., 2015; See et al., 2016)

set to 0, a % of parameters with the least magnitute


With machine translation, researchers found that we can remove almost half the params with almost none effect on downstream tasks. This is a type of unstructured pruning, because we remove params throughout the model, anywhere we see good.

Problems with unstructured pruning:


we can make our model(vectors) and make it sparse, but if our hardware cannot take advantage of this sparsity, we don't gain anything - and this is currently an active research area.

A more currently useful idea is structured pruning (Xia et al, 2022).

Instead of picking params from the model, we remove entire components/layers from the model, example


We can remove almost half the heads of a transformers, and we get a negligent impact on performance.


Distillation - train one model (the 'student') to replicate the behaviour of another large(r) model (the 'teacher'); ~all params are changed

Train the teacher

Generate soft targets - instead of using the actual(hard) labels for training, the teacher outputs soft targets which are probability distributions over the classes

Train the student using the soft targets, and it learns to mimic the behaviour of the teacher by minimizing the difference between its predictions and the soft targets

Example is DistilBERT (Sanh et al., 2019)





That is all for today!
See you tomorrow :)

Popular posts from this blog

[Day 198] Transactions Data Streaming Pipeline Porject [v1 completed]

Image
 Hello :) Today is Day 198! A quick summary of today: data streaming pipeline project [v1 done] Here is a link to the project's repo. Well ... I did not know I can do it in a day (~14 hours) after yesterday's issues but here we are. Turns out in order to insert the full (~70 variables with nested/list structure), I need the proper pyspark schema. And yesterday I did not have that and that is why when I was reading data in the kafka producer I was getting NULL in the columns - my schema was wrong. Well today I not only fixed the schema for the 4 variables I had yesterday, but included *all* the variables that come from the Stripe API ~ 70 (for completeness).  When I run docker-compose, the data streams and is input into the postgres db (and is still running). Unfortunately, the free Stripe API for creating realistic transactions has a limit of 25, so every 3 seconds, 25 new transactions are sent to the db. It has been running half the day (since I got that set up) and as I am wr
Read more

[Day 107] Transforming natural language to charts

Image
 Hello :) Today is Day 107! A quick summary of today: created text2chart - transforming natural language to charts [ github ] [ webapp ] I am super sick, but in the around 3 hours of my fever not being high, I managed to study a bit. I was again thinking about creating some nlp project but was stuck on the idea of how to host a model. After looking around I found this article where they create a similar app called chat2vis  and they use codellama. I know codellama is an open source model that I can get from huggingface and I wondered, how did they host it. Turns out - they ask the user to input a huggingface api key, which then allows for querying and generating content. (AWESOME! this is applicable to some of the other PDF chat apps that I have, but I will look to make an app of them later).  My fever eased a bit, and I sat on my chair to do some work ^^ Here is a breakdown of the app. I used streamlit. Also, 2 default datasets during dev: Financial Statements of Major Companies(2009-
Read more

[Day 54] I became a backprop ninja! (woohoo)

Image
 Hello :) Today is Day 54! A quick summary of today: Read through cs231n's neural network lectures [1] [2] [3] Spent 8 hours going over Andrej Karpathy's backprop ninja tutorial After yesterday's confidence boost in doing backprop, I came in today with the intention of becoming of, as Andrej Karpathy calls it, a backprop ninja.  The cs231n NN lectures were the usual about how NN became to be, param initialization, activation functions, loss functions, regularization and optimizers. Nothing out of the ordinary, but still was good to reassure myself of that knowledge. But that was definitely not the main even of today. As I mentioned, I wanted to become (or at least attempt to become) a backprop ninja today. I have watched that particular tutorial twice, the 1st time I had no idea what is going on, the 2nd time I could follow, and now it was the 3rd time - and now, I wanted to participate and actually see the why and the how of the things we have to do, when we do manual back
Read more