Roshan S Sharma

OOM Handling during training

3

OOMs can be handled during model training in ESPNet2 by skipping the problematic batch or reducing the batch size dynamically as done in Fairseq here: https://github.com/pytorch/fairseq/blob/50a671f78d0c8de0392f924180db72ac9b41b801/fairseq/trainer.py#L188 This feature would be...

ESP-summ PR

11

## What? - Building the ESPNet-SUMM toolkit to support more summarization recipes, model architectures and evaluation metrics ## TODO: Code Additions - [x] Truncating feature inputs - [x] FNet Encoder...

Init asr2/sum2 for How2

6

## What? Added recipe for asr2, sum2 on How2-2000h Extras: - [x] Added segmentation support for KMeans based Discretization over long input recordings - this is needed for SSL feature...

Is there a way to obtain utterance level and corpus level metrics simultaneously by default? I want utterance-level metrics to do fine-grained analysis afterwards. Would appreciate any pointers - thanks...

Attn Mask for Non-causal Models

2

We are examining non-NLP applications of the cosformer self-attention, and would need to use attention masking for the padded tokens in the batch. Is there a way to incorporate this...

### Author Pages https://aclanthology.org/people/r/roshan-s-sharma/ https://aclanthology.org/people/r/roshan-sharma/ ### Type of Author Metadata Correction - [ ] The author page wrongly conflates different people with the same name. - [X] This author has...