New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[TVMC] Keep quantized weights when importing PyTorch model #9417

Merged

manupak merged 1 commit into apache:main from lhutton1:keep-quantized-weight

Nov 2, 2021

Contributor

lhutton1 commented Nov 1, 2021

BYOC requires keep_quantized_weight be set to true when converting PyTorch models using from_torch. Setting this to be True when importing a model using TVMC.

cc @leandron @masahi @ekalda


          [TVMC] Keep quantized weights when importing PyTorch model

b7772f9

BYOC requires `keep_quantized_weight` be set to true when converting
PyTorch models using `from_torch`. Setting this to be True when using
TVMC.

Change-Id: I8c183f9f802ea54d24679a4017e56481d84e5655

lhutton1 requested review from jwfromm and leandron as code owners

November 1, 2021 18:19

masahi approved these changes

View reviewed changes

Member

masahi left a comment

I'm glad you find this feature useful!

manupak approved these changes

View reviewed changes

Contributor

manupak left a comment

LGTM

manupak merged commit d1aebcb into apache:main

Contributor

manupak commented Nov 2, 2021

Thanks @masahi @lhutton1 . This is merged now.

Contributor

u99127 commented Nov 2, 2021

If this was merged in before the release branch was cut - can we back port this fix for 0.8 ? It seems pretty inappropriate to let something like this not get fixed in the release.

lhutton1 deleted the keep-quantized-weight branch

November 2, 2021 09:16

Contributor Author

lhutton1 commented Nov 2, 2021

cc @junrushao1994 would it be possible to include this fix in the v8.0 draft?

Contributor

u99127 commented Nov 2, 2021

Further should there be a test for this ?

Contributor

manupak commented Nov 2, 2021

I think "from_pytorch API" is tested seperately. We can add a test to see if PyTorchFrontend.load(...) issues the correct a call to "from_pytorch" API. @lhutton1 WDYT ?

Contributor Author

lhutton1 commented Nov 2, 2021 •

edited

Loading

We can check that the output of converting a quantized model has either int8 or uint8 weights, rather than float32 - which I think would be slightly better. I did look at this, but it looks like there's issues running PyTorch TVMC tests in CI - #7455, so it will need further investigation. Edit: It looks related to #9362, so I'll look at re-enabling the test once CI is updated and adding a similar quantized version.

lhutton1 added a commit to lhutton1/tvm that referenced this pull request


          [TVMC] Add test for quantized pytorch model

843e738

As a follow up to apache#9417 and now that apache#9362 is resolved, this PR adds a
test to check quantized pytorch mobilenetv2 is converted correctly.

Change-Id: Iaf2d38ce71c008e0141a4a2536bd54c2c9f3fe3d

lhutton1 mentioned this pull request

[TVMC] Add test for quantized pytorch model #9467

Merged

leandron pushed a commit that referenced this pull request


          [TVMC] Add test for quantized pytorch model (#9467)

As a follow up to #9417 and now that #9362 is resolved, this PR adds a
test to check quantized pytorch mobilenetv2 is converted correctly.

Change-Id: Iaf2d38ce71c008e0141a4a2536bd54c2c9f3fe3d

mehrdadh pushed a commit to mehrdadh/tvm that referenced this pull request


          [TVMC] Add test for quantized pytorch model (apache#9467)

0949e15

As a follow up to apache#9417 and now that apache#9362 is resolved, this PR adds a
test to check quantized pytorch mobilenetv2 is converted correctly.

Change-Id: Iaf2d38ce71c008e0141a4a2536bd54c2c9f3fe3d

mehrdadh pushed a commit to mehrdadh/tvm that referenced this pull request


          [TVMC] Keep quantized weights when importing PyTorch model (apache#9417)

0adf6e6

BYOC requires `keep_quantized_weight` be set to true when converting
PyTorch models using `from_torch`. Setting this to be True when using
TVMC.

Change-Id: I8c183f9f802ea54d24679a4017e56481d84e5655

mehrdadh pushed a commit to mehrdadh/tvm that referenced this pull request


          [TVMC] Add test for quantized pytorch model (apache#9467)

b28874d

As a follow up to apache#9417 and now that apache#9362 is resolved, this PR adds a
test to check quantized pytorch mobilenetv2 is converted correctly.

Change-Id: Iaf2d38ce71c008e0141a4a2536bd54c2c9f3fe3d

ylc pushed a commit to ylc/tvm that referenced this pull request


          [TVMC] Keep quantized weights when importing PyTorch model (apache#9417)

0646dc4

BYOC requires `keep_quantized_weight` be set to true when converting
PyTorch models using `from_torch`. Setting this to be True when using
TVMC.

Change-Id: I8c183f9f802ea54d24679a4017e56481d84e5655

ylc pushed a commit to ylc/tvm that referenced this pull request


          [TVMC] Add test for quantized pytorch model (apache#9467)

59c5d3f

As a follow up to apache#9417 and now that apache#9362 is resolved, this PR adds a
test to check quantized pytorch mobilenetv2 is converted correctly.

Change-Id: Iaf2d38ce71c008e0141a4a2536bd54c2c9f3fe3d

yangulei pushed a commit to yangulei/tvm that referenced this pull request


          [TVMC] Add test for quantized pytorch model (apache#9467)

2e9d2f8

As a follow up to apache#9417 and now that apache#9362 is resolved, this PR adds a
test to check quantized pytorch mobilenetv2 is converted correctly.

Change-Id: Iaf2d38ce71c008e0141a4a2536bd54c2c9f3fe3d

ylc pushed a commit to ylc/tvm that referenced this pull request


          [TVMC] Keep quantized weights when importing PyTorch model (apache#9417)

f196171

BYOC requires `keep_quantized_weight` be set to true when converting
PyTorch models using `from_torch`. Setting this to be True when using
TVMC.

Change-Id: I8c183f9f802ea54d24679a4017e56481d84e5655

ylc pushed a commit to ylc/tvm that referenced this pull request


          [TVMC] Add test for quantized pytorch model (apache#9467)

a40a909

As a follow up to apache#9417 and now that apache#9362 is resolved, this PR adds a
test to check quantized pytorch mobilenetv2 is converted correctly.

Change-Id: Iaf2d38ce71c008e0141a4a2536bd54c2c9f3fe3d

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet