Bad results of CARS

[sven9972]

Thanks for your impressive work. During experiments, I get bad results like 92% or 93% by the cars.yml. And the param is like this: flops: 0.22893817600000002 , params:1898.3880000000001 I think this param is too small, how can l get another model of large params? In Model Zoo, the Size of CARS-A is 7.72M, but the Size of CARS-A is 2.4M in paper of CARS. This confuses me. Can you please give an explanation, thank you very much!!!

[xinghaochen]

Thanks for your interest.

The cars.yml provided in this repository is a toy example for understanding how CARS algorithm works. For benchmarking CARS, you should change the hyper-parameters accordingly (e.g. the num_individual for the search algorithm etc.) , which are all well described in the paper.

The "Size of CARS-A" in Model Zoo means the actual storing size in MB, while the paper reports the number of parameters (2.4M). Besides, the pre-trained models in the Model Zoo are slightly different from that reported in the paper, due to re-benchmarking from vega.

[sven9972]

Thank you for your reply! But I changed the num_individual to 64 or 128, the result of all models is still only 94%-95%. I will continue to make some attempts to benchmark CARS.

And if you can publish the hyper-parameters of benchmarking CARS in cars.yml, I will be very grateful! Thank you again for this impressive work.