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
Fix numeric overflow issue that causes loss to Inf/NaN

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 2 commits April 5, 2025 16:52

fix: soft device placement

90e6695

fix: option TF_CUDNN

f4e459d

nglehuy force-pushed the feat-streaming branch from 1bc4ebf to f4e459d Compare April 5, 2025 10:02

nglehuy added 2 commits April 5, 2025 17:04

chore: logging

7473458

fix: configs

57aef49

nglehuy force-pushed the feat-streaming branch from 496b65c to 57aef49 Compare April 5, 2025 14:07

nglehuy added 6 commits April 5, 2025 22:09

fix: feature extraction layer dtype

a2eaf15

fix: numeric stability with dtype compatible

5a1f054

chore: summary

b70bdd6

fix: softmax numberic overflow with mask

edf5eb2

chore: remove commented code

81a336c

chore: config

6be5428

nglehuy force-pushed the feat-streaming branch 2 times, most recently from 0d6d3bf to 6be5428 Compare April 7, 2025 05:10

nglehuy added 16 commits April 7, 2025 12:11

chore: remove commented code

0d75686

fix: use keras-nightly

f5886a5

fix: requirements

e524af6

fix: requirements

20d30c6

fix: requirements with docker

7f145f7

fix: kaggle backup & restore callback

2d557cc

fix: deps

6dc3410

fix: deps

36f3cc0

fix: deps

cea79ef

fix: deps

5e31671

fix: deps

6ca9c61

fix: file util

63f6280

fix: env util

4390d1b

fix: add setup.sh script, update make function

64aeb2f

fix: update kaggel callback

65fdf46

fix: update gradient accumulation

c8d3614

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 •

edited

Loading

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 • edited Loading

nglehuy commented Mar 13, 2025 •

edited

Loading