L1 Pruning on resnet18 from mmpretrain

[mathieu-charbonnel]

Good morning :)
I am new to pruning and to model compression in general and I have been playing with mmrazor and took as an example the following config from pretrain.
_base_ = ['mmpretrain::resnet/resnet18_8xb32_in1k.py']
I added to this config the pruning related elements

data_preprocessor` = {'type': 'mmpretrain.ClsDataPreprocessor'}
architecture = _base_.model
architecture.update({
    'init_cfg': {
        'type':
        'Pretrained',
        'checkpoint':
        'https://download.openmmlab.com/mmclassification/v0/resnet/resnet18_8xb32_in1k_20210831-fbbb1da6.pth'  # noqa
    }
})
un_prune = 1.0
stage_ratio_1 = 0.5
stage_ratio_2 = 0.4
stage_ratio_3 = 0.6
stage_ratio_4 = un_prune

target_pruning_ratio = {
    # stage 1
    'backbone.conv1_(0, 64)_64': un_prune,  # short cut layers
    'backbone.layer1.0.conv1_(0, 64)_64': stage_ratio_1,
    'backbone.layer1.1.conv1_(0, 64)_64': stage_ratio_1,
    'backbone.layer1.2.conv1_(0, 64)_64': un_prune,
    # stage 2
    'backbone.layer2.0.conv1_(0, 128)_128': un_prune,
    'backbone.layer2.0.conv2_(0, 128)_128': un_prune,  # short cut layers
    'backbone.layer2.1.conv1_(0, 128)_128': stage_ratio_2,
    'backbone.layer2.2.conv1_(0, 128)_128': stage_ratio_2,
    'backbone.layer2.3.conv1_(0, 128)_128': un_prune,
    # stage 3
    'backbone.layer3.0.conv1_(0, 256)_256': un_prune,
    'backbone.layer3.0.conv2_(0, 256)_256': un_prune,  # short cut layers
    'backbone.layer3.1.conv1_(0, 256)_256': stage_ratio_3,
    'backbone.layer3.2.conv1_(0, 256)_256': stage_ratio_3,
    'backbone.layer3.3.conv1_(0, 256)_256': stage_ratio_3,
    'backbone.layer3.4.conv1_(0, 256)_256': stage_ratio_3,
    'backbone.layer3.5.conv1_(0, 256)_256': un_prune,
    # stage 4
    'backbone.layer4.0.conv1_(0, 512)_512': stage_ratio_4,
    'backbone.layer4.0.conv2_(0, 512)_512': un_prune,  # short cut layers
    'backbone.layer4.1.conv1_(0, 512)_512': stage_ratio_4,
    'backbone.layer4.2.conv1_(0, 512)_512': stage_ratio_4
}
model = dict(
    _delete_=True,
    _scope_='mmrazor',
    type='ItePruneAlgorithm',
    architecture=architecture,
    mutator_cfg=dict(
        type='ChannelMutator',
        channel_unit_cfg=dict(
            type='L1MutableChannelUnit',
            default_args=dict(choice_mode='ratio'))),
    target_pruning_ratio=target_pruning_ratio,
    step_freq=1,
    prune_times=1,
)

Then ran a short training on one image using
! timeout 2m python mmrazor/tools/train.py $pruning_cfg_path --work-dir 'demo/pruning'
Then run the following

! python mmrazor/tools/pruning/get_static_model_from_algorithm.py \
  $pruning_cfg_path \
  demo/pruning/epoch_14.pth \
  -o "/mnt/deeplearning/users/mcharbon/demo/pruning"

and finally

!python /workspace/mmlab/MMDeploy/tools/deploy.py \
    /workspace/mmlab/MMDeploy/configs/mmpretrain/classification_onnxruntime_dynamic.py \
    demo/pruning/pruned_resnet_cfg.py \
    demo/pruning/epoch_14_pruned.pth \
    /workspace/mmlab/MMDeploy/tests/data/tiger.jpeg \
    --work-dir demo/pruning \
    --device cpu \
    --show \
    --dump-info

Here is the onnx visualized on Netron. It has those gatherND operations everywhere, plus it seems the masks (from Mutable Channel Units ?) are still there. Is this really the deployed model in the final pruning stage or is there something I am missing ? I would imagine a model similar to the base Resnet18 but with lower number of channels for considered convolution layers.

Thank you very much in advance :),

Mathieu