Evaluation about mcunet-320KB(Imagenet)

[andrewwang0612]

Thanks for the great work. I run this line to evluate the performance if this model python eval_torch.py --net_id mcunet-320kB --dataset {imagenet/} --data-dir PATH/TO/DATA/val But the accuracy just gets about 11%,

And I use this github to preprate the Imagenet dataset https://gist.github.com/antoinebrl/7d00d5cb6c95ef194c737392ef7e476a The validation just like this setting,it split to 1000 folders and each folder have about 50 images

Could you tell me the possible reason? Or I use the wrong way to split the Imagenet on validation?

[tonylins]

Hi, could you change the argument to --dataset imagenet (actually I think it will raise an error with your command)?

[andrewwang0612]

Oh, I haven't type it clearly. I used the argument to --dataset imagenet

[tonylins]

This is weird. Can you try evaluating the VWW dataset, which I have provided as a zip copy?

[andrewwang0612]

I run this command and provided by yours VWW dataset

python eval_torch.py --net_id mcunet-320kB --dataset vww --data-dir ./vww-s256/val --batch-size 4 Still have some problem,I haven't modify eval_torch.py .

[tonylins]

Hi, for vww dataset, you need to test on vww models like mcunet-320kB-vww, instead of ImageNet models mcunet-320kB. Please use the following script instead:

python eval_torch.py --net_id mcunet-320kB-vww --dataset vww --data-dir ./vww-s256/val --batch-size 4

[andrewwang0612]

Thank you!!For vww dataset, I get the right accyracy. But for imagenet dataset,still have the same problem! Or I have to split 10,000 samples from the training set of ImageNet?

[tonylins]

Hi, thanks for confirming. I just pulled the repo and verified that I could reproduce the number. It should be related to the dataset processing. The number of iterations for testing also does not match. Can you try testing other torchvision models to see if you can get the correct results? Please see https://github.com/pytorch/examples/tree/main/imagenet and use --evaluate.

[andrewwang0612]

Thank you so much!! For imagenet dataset that provided by you. I get the correct results. It's related to the data processing.