Skip to content

[QUANTIZE] Refactor quantization codebase and fix model accuracy - #3543

Merged
ZihengJiang merged 8 commits into
apache:masterfrom
ZihengJiang:quantize-lr
Aug 15, 2019
Merged

[QUANTIZE] Refactor quantization codebase and fix model accuracy#3543
ZihengJiang merged 8 commits into
apache:masterfrom
ZihengJiang:quantize-lr

Conversation

@ZihengJiang

@ZihengJiangZihengJiang commented Jul 14, 2019•
edited
Loading

Copy link
Copy Markdown
Contributor
  • Separate quantization code base into different files: partition.cc, annotate.cc, realize.cc
  • Change rewrite_for_vta to extra partition pass and enable it by default
  • Change annotation.force_cast(x) to annotation.cast_hint(x, dtype)
  • Remove qconfig.store_lowbit_output and enable it by default
  • Fixed accuracy of models like mobilenet：
    • resnet18_v1(8-16bit): 69.29%
    • resnet18_v1(8-32bit): 69.29%
    • resnet34_v1: 73.33%
    • resnet50_v1: 74.78%
    • resnet101_v1: 75.66%
    • mobilenetv2_1.0: 66.64%

cc @tqchen@eqy@vinx13@tmoreau89

@ZihengJiangZihengJiang changed the title [WIP] Add nightly quantization regression tests[QUANTIZE] Refactor codebase, fix accuracy, add nightly regression testsJul 19, 2019
@ZihengJiangZihengJiang changed the title [QUANTIZE] Refactor codebase, fix accuracy, add nightly regression tests[QUANTIZE] Refactor codebase and fix accuracyJul 19, 2019
@ZihengJiang

ZihengJiang commented Jul 19, 2019

Copy link
Copy Markdown
ContributorAuthor

After discussing offline with Tianqi, we decide to build the nightly regression tests in another repo.

@ZihengJiangZihengJiang changed the title [QUANTIZE] Refactor codebase and fix accuracy[QUANTIZE] Refactor quantization codebase and fix model accuracyJul 19, 2019
@ZihengJiang

ZihengJiang commented Jul 19, 2019

Copy link
Copy Markdown
ContributorAuthor

@tqchen@eqy@vinx13@tmoreau89 Could you please help to review this change?

vinx13
vinx13 requested changes Jul 19, 2019
View reviewed changes
Comment threadpython/tvm/relay/quantize/quantize.py Outdated
Comment threadpython/tvm/relay/quantize/quantize.py Outdated
Comment threadpython/tvm/relay/quantize/_annotate.py Outdated
Comment threadpython/tvm/relay/quantize/_partition.py
Comment threadpython/tvm/relay/quantize/_partition.py
@eqy

eqy commented Jul 19, 2019

Copy link
Copy Markdown
Contributor

@ZihengJiang Do you think we we could try to get the calibration PR first? I have to port it over to the new pass infra and I think this is likely more easy to replay on top of calibration than vice-versa.

@ZihengJiang

ZihengJiang commented Jul 19, 2019•
edited
Loading

Copy link
Copy Markdown
ContributorAuthor

@eqy Do you mean this one? #3294
Sure if you have time recently

@vinx13vinx13 mentioned this pull request Jul 22, 2019
Merged
tmoreau89
tmoreau89 approved these changes Jul 23, 2019
View reviewed changes

@tmoreau89tmoreau89 left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

Comment threadsrc/relay/pass/quantize/quantize.h Outdated
@tqchen

tqchen commented Aug 2, 2019

Copy link
Copy Markdown
Member

@ZihengJiang please followup now that #3538 is merged

@tmoreau89

tmoreau89 commented Aug 7, 2019

Copy link
Copy Markdown
Contributor

What is the status on this PR? @tqchen@ZihengJiang

@tmoreau89tmoreau89 mentioned this pull request Aug 9, 2019
Merged
ZihengJiang
ZihengJiang commented Aug 12, 2019
View reviewed changes
Comment threadpython/tvm/relay/quantize/_partition.py
def add_partition_function(ref_call, new_args, ctx):
"""Rewrite function for ewise add for partition"""
if 'cuda' in _target.current_target().keys:
#TODO(wuwei/ziheng) cuda specific rules

@ZihengJiangZihengJiangAug 12, 2019•
edited
Loading

Copy link
Copy Markdown
ContributorAuthor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@vinx13 Since general devices and VTA are okay/required to insert stop_fusion in both side, let's use different rewrite rules for specific target here,

@ZihengJiangZihengJiang mentioned this pull request Aug 13, 2019
Closed
vinx13
vinx13 approved these changes Aug 15, 2019
View reviewed changes
Comment threadtests/python/nightly/quantization/test_quantization_accuracy.py Outdated
@ZihengJiang

ZihengJiang commented Aug 15, 2019•
edited
Loading

Copy link
Copy Markdown
ContributorAuthor

@mingwayzhang those links should be helpful:

@ZihengJiang
ZihengJiang merged commit 7eb1f35 into apache:masterAug 15, 2019
@ZihengJiang
ZihengJiang deleted the quantize-lr branch August 15, 2019 09:31
@vinx13vinx13 mentioned this pull request Aug 16, 2019
Merged
wweic pushed a commit to neo-ai/tvm that referenced this pull request Aug 16, 2019
@ZihengJiangWei Chen
[QUANTIZE] Refactor quantization codebase and fix model accuracy (apa…
d537774
…che#3543)
* Refactor.
* update
* update
* update
* update
* update
* update
anijain2305 pushed a commit to anijain2305/tvm that referenced this pull request Aug 22, 2019
@ZihengJiang@anijain2305
[QUANTIZE] Refactor quantization codebase and fix model accuracy (apa…
addf4b4
…che#3543)
* Refactor.
* update
* update
* update
* update
* update
* update
wweic pushed a commit to neo-ai/tvm that referenced this pull request Sep 6, 2019
@ZihengJiangWei Chen
[QUANTIZE] Refactor quantization codebase and fix model accuracy (apa…
c4b9fbc
…che#3543)
* Refactor.
* update
* update
* update
* update
* update
* update
@tqchentqchen mentioned this pull request Nov 8, 2019
Closed
Sign up for freeto join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@tmoreau89tmoreau89tmoreau89 approved these changes
@vinx13vinx13vinx13 approved these changes

Assignees

No one assigned

Labels

status: need review

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

6 participants

@ZihengJiang@eqy@tqchen@tmoreau89@vinx13@mingwayzhang