use autoslim for yolof about mmrazor

[HIT-cwh]

Sorry for taking so long to get back to you.
We will solve the related issues you raised as soon as possible.

[HIT-cwh]

AutoSlim can be used in Yolof in pr[#126 ]. Pls make sure the following codes works fine first:

# model settings
model_cfg = dict(
    type='mmdet.YOLOF',
    backbone=dict(
        type='ResNet',
        depth=50,
        num_stages=4,
        out_indices=(3, ),
        frozen_stages=1,
        norm_cfg=dict(type='BN', requires_grad=False),
        norm_eval=True,
        style='caffe',
        init_cfg=dict(
            type='Pretrained',
            checkpoint='open-mmlab://detectron/resnet50_caffe')),
    neck=dict(
        type='DilatedEncoder',
        in_channels=2048,
        out_channels=512,
        block_mid_channels=128,
        num_residual_blocks=4),
    bbox_head=dict(
        type='YOLOFHead',
        num_classes=80,
        in_channels=512,
        reg_decoded_bbox=True,
        anchor_generator=dict(
            type='AnchorGenerator',
            ratios=[1.0],
            scales=[1, 2, 4, 8, 16],
            strides=[32]),
        bbox_coder=dict(
            type='DeltaXYWHBBoxCoder',
            target_means=[.0, .0, .0, .0],
            target_stds=[1., 1., 1., 1.],
            add_ctr_clamp=True,
            ctr_clamp=32),
        loss_cls=dict(
            type='FocalLoss',
            use_sigmoid=True,
            gamma=2.0,
            alpha=0.25,
            loss_weight=1.0),
        loss_bbox=dict(type='GIoULoss', loss_weight=1.0)),
    # training and testing settings
    train_cfg=dict(
        assigner=dict(
            type='UniformAssigner', pos_ignore_thr=0.15, neg_ignore_thr=0.7),
        allowed_border=-1,
        pos_weight=-1,
        debug=False),
    test_cfg=dict(
        nms_pre=1000,
        min_bbox_size=0,
        score_thr=0.05,
        nms=dict(type='nms', iou_threshold=0.6),
        max_per_img=100))
algorithm_cfg = ConfigDict(
    type='AutoSlim',
    architecture=dict(type='MMDetArchitecture', model=model_cfg),
    pruner=dict(
        type='RatioPruner',
        ratios=(2 / 12, 3 / 12, 4 / 12, 5 / 12, 6 / 12, 7 / 12, 8 / 12, 9 / 12,
                10 / 12, 11 / 12, 1.0)),
    retraining=False,
    bn_training_mode=True,
    input_shape=None)

algorithm = build_algorithm(algorithm_cfg)
subnet_dict = algorithm.pruner.sample_subnet()
assert isinstance(subnet_dict, dict)
algorithm.pruner.set_subnet(subnet_dict)
subnet_dict = algorithm.pruner.export_subnet()
assert isinstance(subnet_dict, dict)
algorithm.pruner.deploy_subnet(algorithm.architecture, subnet_dict)
imgs = torch.randn(16, 3, 224, 224)
losses = algorithm.architecture.forward_dummy(imgs)

By the way, AutoSlim can also be used to prune models with shared modules such as detection head in RetinaNet in this pr.

[NickYangMin]

ok, I will try it, thx!

[NickYangMin]

I try it with branch 'dev_v0.3.0'，but it still doesn't work

[HIT-cwh]

I try it with branch 'dev_v0.3.0'，but it still doesn't work

Pls try it with #126 , this pr is based on branch 'dev_v0.3.0' and hasn't been merged yet.

[NickYangMin]

ok

[NickYangMin]

it works, thx! @HIT-cwh