Hyper-parameter

[mmnn-00]

Thanks for your codes. However, I have a question about setting hyper-parameters.
When using CIFAR10, the DQ paper has the following description:

According to the above description, I set the hyperparameters as below, although it is different from the sample code written inthe README. Are the following correct? (Changed '-se' from 0 to 10, and removed '--pretrained')
# Dataset bin generation (By default we use a bin number of 10) CUDA_VISIBLE_DEVICES=0 python -u quantize_sample.py \ --fraction 0.1 --dataset CIFAR10 --data_path ~/data_cifar \ --num_exp 10 --workers 10 -se 10 --selection Submodular --model ResNet18 \ -sp ../results/bin_cifar_010 \ --batch 128 --submodular GraphCut --submodular_greedy NaiveGreedy

[vimar-gu]

Thanks for pointing out the inconsistency between the paper and the implementation. In the paper we kept it consistent with the original Deepcore setting to conduct 10-epoch pre-training. And directly adopting pre-trained model is also practical. Please try running these two scripts and compare the performance.

[mmnn-00]

I got it. I will try. Thank you very much!

[haiduo]

I got it. I will try. Thank you very much!

Hi @mmnn-00 , what are your reproducible results, please?