Aston Zhang

icon indicating a website link https://astonzhang.com/

icon company Meta icon location Menlo Park icon location Building large language models on the Llama Team at Meta Generative AI

Results 70 comments of Aston Zhang

Inconsistent learning curves of two NiN trainings

We force_reinit=True since it may be invoked multiple times in a notebook and each invocation should train the model from scratch (re-initialize params). The re-initialization takes place *before* the training...

solutions manual

For exercises on the book, probably no. For assignment in the Berkeley course: https://courses.d2l.ai/berkeley-stat-157/homework.html Instructors may request all the homework solutions by emailing rlhu at amazon dot com.

CI: Add ability to build specific chapters or sections

This feature will be helpful for changes that may affect a subset of notebooks!

GRU and LSTM are running very slow in TF

I think it's always. Can you look into it?

[RFC] Code structure refactor

Thanks. The idea is very good and the workaround of breaking a class def into multiple cells may address our earlier concerns. One thing we need to think carefully is...

Remember a user's framework selection?

Thanks and this sounds a nice feature to have! FYI, the entire website is built using sphinx: https://github.com/d2l-ai/d2l-book/blob/master/d2lbook/build.py#L307

Figure caption is lost when upgrading from 0.3.5 to 0.4.2

Is this feature (show figure caption in latex and html) removed in 0.4.2? https://github.com/spatialaudio/nbsphinx/issues/157

Figure caption is lost when upgrading from 0.3.5 to 0.4.2

Thanks. However, showing figure captions is quite essential. For instance, our website and pdf need figure captions with figure numberings: source code in markdown: https://github.com/d2l-ai/d2l-en html: http://www.d2l.ai/ pdf: http://www.d2l.ai/d2l-en.pdf Do...

[Feat] Imrove error verbosity during d2lbook build pdf

Indeed, a nice feature to have -- so no need to debug latex locally.

Increase number of iterations for MLE code

Thanks. It's not necessary to make the estimate *exactly* equal to the ground-truth. It's more ideal to have a close estimate while reducing training time in our demo :)