autoslim后的模型测试

[Hiwyl]

• 其中origin 是使用mmcls训练的mobilenetv2_1.5
• autoslim是下载的权重（计算量和参数量是拿的README），参见：
  

在推理速度上没有体现？还是操作有误，分别推理了10000次，预热100次取平均

[mm-assistant]

We recommend using English or English & Chinese for issues so that we could have broader discussion.

[Hiwyl]

测试过程

1. 国产芯片测试

1. 使用mmcls 训练mobilenetv2_1.5模型并转为onnx
2. 下载autoslim，使用mmdeploy自带的torch2onnx转为onnx，其中相关mcfg如下：

# deploy_cfg
onnx_config = dict(
    type="onnx",
    export_params=True,
    keep_initializers_as_inputs=False,
    opset_version=10,
    save_file="mobilenetv2_mmdeploy",
    input_names=["input"],
    output_names=["output"],
    input_shape=[224, 224],
)

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


# model_cfg 为mmrazor中pruning中mobilenet
_base_ = [
    "./autoslim_mbv2_supernet_8xb256_in1k.py",
]

model = dict(head=dict(loss=dict(type="LabelSmoothLoss", mode="original", label_smooth_val=0.1, loss_weight=1.0)))

# FIXME: you may replace this with the channel_cfg searched by yourself
channel_cfg = [
    "https://download.openmmlab.com/mmrazor/v0.1/pruning/autoslim/autoslim_mbv2_subnet_8xb256_in1k/autoslim_mbv2_subnet_8xb256_in1k_flops-0.22M_acc-71.39_20211222-43117c7b_channel_cfg.yaml",  # noqa: E501
]

algorithm = dict(
    architecture=dict(type="MMClsArchitecture", model=model),
    distiller=None,
    retraining=True,
    bn_training_mode=False,
    channel_cfg=channel_cfg,
)

runner = dict(type="EpochBasedRunner", max_epochs=300)

find_unused_parameters = True

torch 速度测试

均使用mmcls的image_demo测试，只是加载的权重不同，使用相同的cfg

# model settings
model = dict(
    type="ImageClassifier",
    backbone=dict(type="MobileNetV2", widen_factor=1.5),
    neck=dict(type="GlobalAveragePooling"),
    head=dict(
        type="LinearClsHead",
        num_classes=1000,
        in_channels=1920,
        loss=dict(type="LabelSmoothLoss", mode="original", label_smooth_val=0.1, loss_weight=1.0),
        topk=(1, 5),
    ),
)
# dataset settings
dataset_type = "CDataset"
img_norm_cfg = dict(mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
train_pipeline = [
    dict(type="LoadImageFromFile"),
    dict(type="RandomResizedCrop", size=224, backend="pillow"),
    dict(type="RandomFlip", flip_prob=0.5, direction="horizontal"),
    dict(type="Normalize", **img_norm_cfg),
    dict(type="ImageToTensor", keys=["img"]),
    dict(type="ToTensor", keys=["gt_label"]),
    dict(type="Collect", keys=["img", "gt_label"]),
]
test_pipeline = [
    dict(type="LoadImageFromFile"),
    dict(type="Resize", size=(256, -1), backend="pillow"),
    dict(type="CenterCrop", crop_size=224),
    dict(type="Normalize", **img_norm_cfg),
    dict(type="ImageToTensor", keys=["img"]),
    dict(type="Collect", keys=["img"]),
]
data = dict(
    samples_per_gpu=2,
    workers_per_gpu=1,
    train=dict(
        type=dataset_type,
        data_prefix="/projects/release/data/train",
        ann_file="/projects/release/data/meta/train.txt",
        pipeline=train_pipeline,
    ),
    val=dict(
        type=dataset_type,
        data_prefix="/projects/release/data/val",
        ann_file="/projects/release/data/meta/val.txt",
        pipeline=test_pipeline,
    ),
    test=dict(
        # replace `data/val` with `data/test` for standard test
        type=dataset_type,
        data_prefix="/projects/release/data/val",
        ann_file="/projects/release/data/meta/val.txt",
        pipeline=test_pipeline,
    ),
)
evaluation = dict(interval=1, metric="accuracy")
# optimizer
optimizer = dict(type="SGD", lr=0.045, momentum=0.9, weight_decay=0.00004)
optimizer_config = dict(grad_clip=None)
# learning policy
lr_config = dict(policy="step", gamma=0.98, step=1)
runner = dict(type="EpochBasedRunner", max_epochs=300)
# checkpoint saving
checkpoint_config = dict(interval=1)
# yapf:disable
log_config = dict(
    interval=100,
    hooks=[
        dict(type="TextLoggerHook"),
        # dict(type='TensorboardLoggerHook')
    ],
)
# yapf:enable

dist_params = dict(backend="nccl")
log_level = "INFO"
load_from = None
resume_from = None
workflow = [("train", 1)]

[pppppM]

We found that the file size of the three different checkpoint sizes is the same, but the parameter size in each checkpoint is different.
This seems to be some problem with splitting the checkpoint, we will fix this issue soon

[HIT-cwh]

Hi! Sorry for the inconvenience to you. We split the checkpoint based on algorithm.pruner.deploy_subnet (refer to here). When overwriting the original parameter with the sliced parameter of a nn.Module, it is necessary to use the copy of the sliced parameter like module.weight = nn.Parameter(temp_weight.data.clone()). After that, the file size of the three different checkpoint sizes will be completely different.
Could you make a PR to fix this bug?

[Hiwyl]

Thank you for the guidance, I will look at the source code recently and try to fix it

[Hiwyl]

@HIT-cwh #193

• origin mobilenet_v2_1.5 39.25MB
• search config

_base_ = [
    "./supernet.py",
]

algorithm = dict(distiller=None, input_shape=(3, 224, 224))

searcher = dict(
    type="GreedySearcher",
    target_flops=[300000000, 200000000],
    max_channel_bins=12,
    metrics="accuracy",
    metric_options={"topk": (1,)},
)

data = dict(samples_per_gpu=4, workers_per_gpu=1)

• split_checkpoints->checkpoint_1：4.92MB；checkpoint_2：6.21MB

[Hiwyl]

oh no, export onnx weight size are equal

[happygds]

@Hiwyl 你好，请问现在测试正常了么