[OTX] Replace forked mm libraries with public ones #1497

cih9088 · 2023-01-09T02:28:45Z

I highly encourage you to look into this PR thoroughly as this PR impacts on all tasks in OTX.

Changes

replaced forked mm libraries with public ones
employed mmdeploy to export model to OpenVINO IR
enabled multi-node distributed training by extending [OTX] Enable multi-GPU training #1392
used existing hooks (AdaptiveTrainSchedulingHook, EvalHook) to evaluate model before and after training instead of createing new one [OTX] Evaluate a model before training starts #1472
refactored (align code base across cls/det/seg to future refactoring)
fixed bugs

Tests

all pre-commit tests and pytests are passed

Related PRs

mmdet (OTX)
mmseg (OTX, MPA)
mmcls (OTX, MPA)
mmdeploy (OTX, MPA)

Signed-off-by: Songki Choi <songki.choi@intel.com>

Enable model training and NNCF in mmdet (openvinotoolkit#1355) * Enable detection training on latest mmcv/det - ATSS / SSD / YOLOX - NNCF support for ATSS * fix: import errors * feat: add monkey patch to mmdet modules - most of patches would be just wrapping for not tracing in nncf context * feat: add trainable yolox - add trainable yolox - recursively search dataset cfg for nested dataset classes * fix: change device to cpu when nncf tracing * feat: add trainable ssd * refactor: rearange nncf adapter * feat: add trainable mask rcnn models * refactor: move out common utils * fix: ssd head bug * feat: add lr scheduler for accuracy aware runner * refactor: nncf module and monkey patch * fix: proper clustering anchors for ssd * fix: unable to trace the first module in NNCFNetwork * fix: bring back ssd head structure * feat: add train_step method to NNCFNetwork * fix: mismatches * fix: update pipeline for wrapper * fix: add missing file * Fix merge error Signed-off-by: Songki Choi <songki.choi@intel.com> Co-authored-by: Inhyuk Cho <andy.inhyuk.jo@intel.com>

* refactor: remove redundant * feat: enable mmseg training * feat: add nncf related stuff * fix: change lr config * fix: align nncf target metric * refactor: use mpa for training and inference * test: enable tests * fix: minor bug * refactor: patcher * fix: build consistent nncf graph * fix: minor bug * fix: remove unused backup * fix: dealt with datacontainer

* fix: use patcher * feat: update mmcls version * feat: enable NNCF for mmcls * refactor: add build NNCF model functions * fix: minor bug * fix: typo * fix: make sure importing nncf when enabled only * fix: inherit from base super class of otx

…t#1466) * feat: export using mmdeploy * fix: adapt mmdeploy exported model * test: enable openvino export * fix: patch depending on fn type * feat: mmdeploy for classification model * test: enable export and openvino performance test * fix: change temporary requirements * refactor: use builder * fix: do not propagate logger * fix: remove image channel format conversion * fix: handle unlabeled data * fix: run eval before optimizing nncf network * feat: change confidence threshold after nncf optimization * fix: remove redundant attribute * fix: official released openvino version * fix: remove redundants

…t as 0. (openvinotoolkit#1475)" This reverts commit c076902.

…t#1472)" This reverts commit f728295.

cih9088 · 2023-01-13T14:46:35Z

@harimkang @goodsong81 @wonjuleee @sungmanc
As we agreed to merge this PR without review and let contributors fix bugs while they are developing on this PR, can we do it without CI result?
CI seems to have a problem with this PR. It gets stuck with no log and reaches to the time limit.

For a reference, here are test results in local with a clean python 3.8.5 environment installed with otx/algorithms/init_venv.sh.
(pre-commit, tests/units, tests/integration/cli, in order)

NNCF optimisation for YOLOX failed from time to time.

AssertionError: trained_performance[k]=0.9512195121951218, evaluated_performance[k]=0.9367088607594937

JihwanEom · 2023-01-13T15:29:19Z

I agree, CI seems like not working properly due to multi-gpu training issues. (Fails after 30 minutes for each TC)

sungmanc · 2023-01-16T01:16:06Z

@cih9088 , What I have experienced during the CI tests, FAIL cases could be shown in CI although all test cases were passed on the local test. So, checking with CI test is needed before merging the branch unless we are in a hurry.

So, I suggest waiting for the fixing of the CI.

harimkang

There is an issue on the HPO side, but we decided to fix it in the next PR, not this PR.

goodsong81

Cheers!

sungmanc · 2023-01-16T08:51:07Z

@cih9088 , maybe we need to change the URL of repo to more public one. What do you think?

goodsong81 · 2023-01-16T09:09:53Z

@cih9088 , maybe we need to change the URL of repo to more public one. What do you think?

This PR #1393 might be helpful for reference.

cih9088 · 2023-01-16T09:20:35Z

@sungmanc @goodsong81
Agreed. Let me change the repo to the one in #1393 and see how things work out.

otx/cli/tools/train.py

sovrasov · 2023-01-16T14:43:38Z

@sungmanc @goodsong81 Agreed. Let me change the repo to the one in #1393 and see how things work out.

You may need #1486 to use the updated NNCF commit.

goodsong81 and others added 17 commits January 9, 2023 10:13

Update MPA submodule to origin/otx

8a418cc

Signed-off-by: Songki Choi <songki.choi@intel.com>

feat: public mm series libraries

d4b8242

feat: otx refactoring and bug fix

1d385d9

Revert "[OTX] Bugfix: multi GPU raise error when num_workers isn't se…

3a64319

…t as 0. (openvinotoolkit#1475)" This reverts commit c076902.

feat: enable multi-nodes distributed training

bf35549

fix: redundant parts

a2a45e0

Revert "[OTX] Evaluate a model before training starts (openvinotoolki…

123cada

…t#1472)" This reverts commit f728295.

feat: enable evaluation before and after training

1bbdee1

style: fix failed cases

6f1ebfb

fix: disable sam optimizer for nncf task

c6cbcb9

fix: add frezelayer hook for segmentation

c1d32c8

fix: deepcopy instead of shallowcopy

9ac33fa

fix: enable temporary disabled features

52e3b1d

cih9088 requested a review from a team as a code owner January 9, 2023 02:28

cih9088 requested review from wonjuleee and goodsong81 January 9, 2023 02:29

cih9088 force-pushed the feature/otx-public-mmcv branch from 52ea65a to 52e3b1d Compare January 9, 2023 03:03

github-actions bot added CLI Any changes in OTE CLI DEPENDENCY Any changes in any dependencies (new dep or its version) should be produced via Change Request on PM ALGO Any changes in OTX Algo Tasks implementation TEST Any changes in tests labels Jan 9, 2023

cih9088 added 3 commits January 9, 2023 22:10

fix: handle nncf state simply

98041ab

fix: remove submodule

adc22be

Merge branch 'feature/otx' into feature/otx-public-mmcv

abdb636

cih9088 force-pushed the feature/otx-public-mmcv branch from 79e1baa to 3031515 Compare January 10, 2023 04:04

cih9088 marked this pull request as draft January 10, 2023 04:09

cih9088 force-pushed the feature/otx-public-mmcv branch 2 times, most recently from 1a75de3 to 2c21b40 Compare January 12, 2023 13:35

fix: unclutter things

a6a3056

cih9088 force-pushed the feature/otx-public-mmcv branch 6 times, most recently from d781dcd to 33fe7ec Compare January 13, 2023 01:10

fix: ignore annoying leftover data.yaml

e9a3994

cih9088 force-pushed the feature/otx-public-mmcv branch 2 times, most recently from d51cf4a to 734db24 Compare January 13, 2023 04:58

goodsong81 mentioned this pull request Jan 13, 2023

[DOCUMENTATION][feature/otx] update explanation templates. Introduce explanation section for classification task. #1467

Merged

5 tasks

Merge branch 'feature/otx' into feature/otx-public-mmcv

c821e5b

cih9088 force-pushed the feature/otx-public-mmcv branch from 734db24 to c821e5b Compare January 13, 2023 09:54

JihwanEom approved these changes Jan 13, 2023

View reviewed changes

harimkang approved these changes Jan 16, 2023

View reviewed changes

goodsong81 approved these changes Jan 16, 2023

View reviewed changes

goodsong81 merged commit 91fff1d into openvinotoolkit:feature/otx Jan 16, 2023

cih9088 mentioned this pull request Jan 16, 2023

[OTX] Drop legacy-resolver #1448

Merged

harimkang reviewed Jan 16, 2023

View reviewed changes

otx/cli/tools/train.py Show resolved Hide resolved

harimkang mentioned this pull request Jan 16, 2023

[OTX][HOT-FIX] Revert otx cli tools with data.yaml check #1530

Merged

JihwanEom mentioned this pull request Jan 25, 2023

[OTX][Hot-fix] Add init_weights call in the mmcls builder.py #1562

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[OTX] Replace forked mm libraries with public ones #1497

[OTX] Replace forked mm libraries with public ones #1497

cih9088 commented Jan 9, 2023 •

edited

Loading

cih9088 commented Jan 13, 2023 •

edited

Loading

JihwanEom commented Jan 13, 2023 •

edited

Loading

sungmanc commented Jan 16, 2023

harimkang left a comment

goodsong81 left a comment

sungmanc commented Jan 16, 2023

goodsong81 commented Jan 16, 2023

cih9088 commented Jan 16, 2023

sovrasov commented Jan 16, 2023

[OTX] Replace forked mm libraries with public ones #1497

[OTX] Replace forked mm libraries with public ones #1497

Conversation

cih9088 commented Jan 9, 2023 • edited Loading

cih9088 commented Jan 13, 2023 • edited Loading

JihwanEom commented Jan 13, 2023 • edited Loading

sungmanc commented Jan 16, 2023

harimkang left a comment

Choose a reason for hiding this comment

goodsong81 left a comment

Choose a reason for hiding this comment

sungmanc commented Jan 16, 2023

goodsong81 commented Jan 16, 2023

cih9088 commented Jan 16, 2023

sovrasov commented Jan 16, 2023

cih9088 commented Jan 9, 2023 •

edited

Loading

cih9088 commented Jan 13, 2023 •

edited

Loading

JihwanEom commented Jan 13, 2023 •

edited

Loading