[Feature] Add mmrazor support

[wutongshenqiu]

Motivation

Add mmrazor support.

Modification

To support mmrazor, only need to add an attribute from_mmrazor = True in config file and modify init_pytorch_model method in corresponding task.

Use cases

Here we take mmclassification onnx for example.

A minimum independent config is like

onnx_config = dict(
    type='onnx',
    export_params=True,
    keep_initializers_as_inputs=False,
    opset_version=11,
    save_file='end2end.onnx',
    input_names=['input'],
    output_names=['output'],
    input_shape=None)

backend_config = dict(type='onnxruntime')
codebase_config = dict(type='mmcls', task='Classification', from_mmrazor=True)

The only defference is an extra from_mmrazor = True attribute

- codebase_config = dict(type='mmcls', task='Classification')
+ codebase_config = dict(type='mmcls', task='Classification', from_mmrazor=True)

Besides, model_cfg cound be downloaded here and checkpoint here, to use mmdeploy, model config file should manully add mutable_cfg attribute like follows.

algorithm = dict(
    ...
    mutable_cfg=${MUTABLE_CFG_PATH}
    ...
)

The mutable config file could be downloaded here

Checklist

• Pre-commit or other linting tools are used to fix the potential lint issues.

[codecov]

Codecov Report

Merging #220 (9567a3b) into dev-v0.5.0 (e96077f) will increase coverage by 0.06%.
The diff coverage is 100.00%.

@@              Coverage Diff               @@
##           dev-v0.5.0     #220      +/-   ##
==============================================
+ Coverage       62.93%   62.99%   +0.06%     
==============================================
  Files             212      212              
  Lines            7168     7181      +13     
  Branches         1107     1111       +4     
==============================================
+ Hits             4511     4524      +13     
  Misses           2334     2334              
  Partials          323      323              

Flag	Coverage Δ
unittests	62.99% <100.00%> (+0.06%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
mmdeploy/codebase/base/task.py	81.33% <100.00%> (+1.92%)	⬆️
mmdeploy/codebase/mmcls/deploy/classification.py	67.30% <100.00%> (+0.64%)	⬆️
mmdeploy/codebase/mmdet/deploy/object_detection.py	73.77% <100.00%> (+0.43%)	⬆️
mmdeploy/codebase/mmseg/deploy/segmentation.py	75.26% <100.00%> (+0.54%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update e96077f...9567a3b. Read the comment docs.

[grimoire]

The result is different from mobilenetv2 in mmcls model zoo.
image: https://github.com/grimoire/mmclassification/blob/master/tests/data/color.jpg
mmcls: tabby, tabby cat 0.30
mmrazor: egyptian cat 0.19
Is it expected?

[wutongshenqiu]

The result is different from mobilenetv2 in mmcls model zoo. image: https://github.com/grimoire/mmclassification/blob/master/tests/data/color.jpg mmcls: tabby, tabby cat 0.30 mmrazor: egyptian cat 0.19 Is it expected?

Yep, we obtain the architecture of model by SPOS, which might be different from original implementation with mmcls. Besides, the result of the checkpoint could be found here

[AllentDan]

Test data could be simplified if possible

[wutongshenqiu]

Simplified model configs in test data.

[grimoire]

LGTM

[AllentDan]

LGTM after fixing unit tests

[wutongshenqiu]

🙌 Unit tests should be possible to fix after pr open-mmlab/mmrazor#128 is merged.

And sorry 😭 that I might misuse the git rebase, which makes the reviews been dismissed accidentally.

[grimoire]

Don't worry, approval will be dismissed when any change happened. That is expected.

[wutongshenqiu]

Thanks!
The reason for the failure of ci this time seems to be that git is not pre-installed in the cuda container(as showed in the following picture), but it may not be until mmrazor officially releases version 0.3.0 that its support for mmdeploy can be installed through pip.
Is there any work around through this?

[grimoire]

Edit this

mmdeploy/.github/workflows/build.yml

Line 81 in 73cf3b5

apt-get update && apt-get install -y ffmpeg libsm6 libxext6 git ninja-build libglib2.0-0 libxrender-dev python${{matrix.python-version}}-dev

might be a workaround.
It did surprise me that the environment does not include git...

[wutongshenqiu]

That's a little wired. Line 81 already include apt-get install -y git.

mmdeploy/.github/workflows/build.yml

Line 81 in 73cf3b5

apt-get update && apt-get install -y ffmpeg libsm6 libxext6 git ninja-build libglib2.0-0 libxrender-dev python${{matrix.python-version}}-dev

And after dig into the error, I find that step Install system dependencies in this action seems not excute correctly(as showed below).

However, other actions seem no such problem(A example is as below)

Do github or nvidia also encounter network problems? 😂

[grimoire]

Will it be fixed in next mmrazor release(we can change the requirement then)?

[wutongshenqiu]

Yes! After mmrazor 0.3.0 is released, it can be installed via pip instead of git

[wutongshenqiu]

Sorry for waiting so long.
MMRazor no longer needs to be installed via git, so the problem is supposed to be solved.

[grimoire]

LGTM

[AllentDan]

LGTM