Keras 3, Kaggle, CLI, Streaming #295

nglehuy · 2025-03-13T17:26:15Z

Migrate code base from keras v2 -> v3
Add cli tensorflow_asr when installing repo, move scripts to inside to run as modules
Add kaggle callbacks to backup and restore on kaggle's models
Add streaming conformer as paper http://arxiv.org/abs/2010.11395

…ence

nglehuy added 30 commits June 16, 2024 23:11

feat(streaming): training, inference, gradient accumulation

232a80c

fix: remove ds2 dropout on conv module

9ea9b87

fix: add sync batch norm, remove wrong bn in ds2

7c21c1d

fix: only wrap tf.function in jit compile

4e0e8f5

fix: use autograph do_not_convert for batchnorm sync to work

de38407

chore: config

43d6054

fix: update train/test step

d55fd40

fix: nan to num

2f502bb

fix: update compute mask ds2

7441c95

fix: nan to num

f2241cf

fix: ctc loss

ebb6930

fix: update train step

305ddab

chore: config

8268afe

chore: config

c9e4d38

fix: ctc

9b46cbe

fix: add custom batch norm to avoid tf.cond

b5bfe92

fix: env utils

dc08e6a

fix: log batch that cause invalid loss

cf37206

chore: buffer size

4596609

fix: handle unknown dataset size with no metadata provided

ba9d6b2

chore: add option use loss scale

fe594ad

fix: support log debug

7aed458

fix: update gradient accumulation

4330dec

fix: ga

522f080

feat: tf2.16 with keras 3

90cabc2

fix: ga

8285f6d

Merge branch 'tf2.16' into feat-streaming

fd446d6

feat: fix layers, models to tf2.16 with keras 3

5037896

feat: update models to compatible with keras 3

a853eba

fix: loss compute using add_loss, loss tracking

f8a7b91

nglehuy added 13 commits January 26, 2025 00:38

chore: list devices

5ff3163

fix: attention mask

3dbab33

fix: general layers to show outputshape, invalid loss show outputs

1545543

fix: models configs

5f784b7

fix: config streaming

70ac41e

fix: configs

eebc361

fix: configs

6b0bec4

Merge branch 'main' into feat-streaming

2ac8e7f

fix: streaming masking mha

7dcd145

fix: streaming masking mha

e209040

fix: update mha attention mask

5649fdd

feat: add support for layer norm in conformer conv module

26c4a5f

chore: update configs

4f77a52

nglehuy force-pushed the feat-streaming branch from d40c2a0 to db6de4e Compare March 17, 2025 14:07

fix: feature extraction layer dtype tf.float32 to ensure loss converg…

c3ab865

…ence

nglehuy force-pushed the feat-streaming branch from db6de4e to c3ab865 Compare March 17, 2025 14:07

fix: ctc loss tpu - case logits to float32

778c1a2

nglehuy force-pushed the feat-streaming branch from 4e40fca to 778c1a2 Compare March 17, 2025 15:37

nglehuy added 9 commits March 20, 2025 01:54

fix: use auto mask

e68ceee

fix: pad logits length to label length

f1e2a88

fix: ctc loss tpu

f1a0ed6

chore: config

6f7f246

fix: disable bias/activity regularizer as not needed

d538e69

chore: config

7611ff8

chore: setup mxp

0c8e7c1

chore: setup mxp

56d2afa

fix: small kaggle

454163c

nglehuy force-pushed the feat-streaming branch from 49238de to 454163c Compare March 25, 2025 14:07

nglehuy added 2 commits March 25, 2025 21:20

chore: transformer-ctc streaming

a333bfc

chore: config

cf435a3

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Keras 3, Kaggle, CLI, Streaming #295

Keras 3, Kaggle, CLI, Streaming #295

nglehuy commented Mar 13, 2025

Keras 3, Kaggle, CLI, Streaming #295

Are you sure you want to change the base?

Keras 3, Kaggle, CLI, Streaming #295

Conversation

nglehuy commented Mar 13, 2025