Naga Sai Abhinay Devarinti

[Community Pipeline] Checkpoint Merger based on Automatic1111

3

The checkpoint merger pipeline based on the discussion at #877. Tested the checkpoint merger for the case with two checkpoints.

[Community Pipeline] Unclip Image Interpolation

2

Added the code sample in the README.md

BPE Trainer doesn't respect the `vocab_size` parameter when dataset size is increased

1

I'm training a new tokenizer on an Indic language, Tamil. I tried two different runs: Test run with part of the data used for training ~0.3Gb ``` from datasets import...

Regression Test Pipeline

3

- [x] Add English text normlaization. - [x] Add WER calculations. - [x] Compare and check norm outputs from Python implementation - [x] Add WER to the regression tests once...

Draft for reading files in chunks

1

TODO: From Zack: > The tricky part is keeping track of timestamps as we iterate through, there can be drift > If we just do every 30s we’re going to...

[Bug Fix] HfUploader - use the right local filename path.

1

## Description of changes: Fixes the `HfUploader.upload_file` method to use the right path to the local file for uploading. ## Issue #, if available: #875 ## Merge Checklist: _Put an...

## To reproduce Steps to reproduce the behavior: 1. Setup mosaicml-streaming from the main branch `pip install git+https://github.com/mosaicml/streaming.git` 2. Run sample code: ```python from streaming import MDSWriter #replace with the...

[WIP] Differentiable HPSv3

1

It is differentiable. Tested it with DRAFT training. But reward is slightly higher than current implementation in the repo. Eg: 7.610 vs 7.5410. Needs comparison with the original codebase as...

HPSv3 gradient computation broken ?

4

It looks like the gradient computation is broken after flowing through HPSv3. ``` latents = decode_latent(vae=vae, vae_scale_factor=vae_scale_factor, height=data["height"], width=data["width"], latents=latents) latents = image_processor.denormalize(latents).clamp(0,1) ## check latents are in range 0-1...