[Bug]autoslim, use splitted checkpoint to test subnet, accuracy is wrong about mmrazor HOT 4 CLOSED

Thanks for your issue. Is the order of the channel-cfgs in step 4 the same as that of algorithm.channel_cfg in step 3?

from mmrazor.

Thanks for your issue. Is the order of the channel-cfgs in step 4 the same as that of algorithm.channel_cfg in step 3?

Yes. the channel-cfgs in step 4 is the same as that of algorithm.channel_cfg in step 3, the same file.

python -m torch.distributed.launch --nproc_per_node=8 ./tools/mmcls/train_mmcls.py
configs/pruning/autoslim/autoslim_mbv2_subnet_8xb256_in1k.py
--cfg-options algorithm.channel_cfg=/root/work/mmrazor/work_dirs/autoslim_mbv2_search_8xb1024_in1k/subnet_199761362.yaml
--launcher pytorch

By the way, for step3 if 'algorithm.channel_cfg' is 3 files which is the same as 'channel-cfgs' in step 4, that's all right.
But if 'algorithm.channel_cfg' is only 1 file which is the same as 'channel-cfgs' in step 4, the test result in step 5 is wrong.

from mmrazor.

I'm a little confused. Maybe you do not need to split checkpoint when you only retrain one subnet('algorithm.channel_cfg' is only 1 file). Only when you retrain multiple subnets together (just like autoslim), you need to split the checkpoint before you test subnets with different flops.

By the way, experiments show that replacing LabelSmoothLoss with CrossEntropyLoss may lead to better results. And we will fix it later.

from mmrazor.

When i only retrain one subnet， i dont need to split checkpoint, I knew it when I tried to retrain 3 subnets later.
But when the first time i used it, i did not know. you should fix the split step or include a detailed explanation.

I'll try the CrossEntropyLoss. How good?
Is autoslim suitable for mmdet.FasterRCNN or yolo? @HIT-cwh

Thank you for your great work.

from mmrazor.