search

mit-han-lab / amc

[ECCV 2018] AMC: AutoML for Model Compression and Acceleration on Mobile Devices
https://arxiv.org/abs/1802.03494
MIT License
430 stars 114 forks source link

Imagenet dataset ? #11

Open luunguyen97 opened 5 years ago

luunguyen97 commented 5 years ago

Hi, thanks for your great job. I wonder to know how can I get the right imagenet dataset you used for strategy search on mobilenet?

tonylins commented 5 years ago

We used the regular ImageNet dataset. You can find more information here: http://image-net.org/download

luunguyen97 commented 5 years ago

@tonylins Did you mean the whole ILSVRC2012 dataset of ImageNet. I see that it very large dataset (about 150Gb for both training set and validation set). However, It is said in your AMC paper that you used 3000 images from the training set to evaluate the reward function. So I just need to download 3000 image of the dataset ? Thanks.

chaopengz commented 5 years ago

@tonylins Did you mean the whole ILSVRC2012 dataset of ImageNet. I see that it very large dataset (about 150Gb for both training set and validation set). However, It is said in your AMC paper that you used 3000 images from the training set to evaluate the reward function. So I just need to download 3000 image of the dataset ? Thanks.

Although the author used only 3000 images for reward evaluation, about 15000(300*50) images from training set is needed to do weights regression during pruning kernel. What's more, you need the whole training set to finetune the pruned model. Thus it is necessary to download the whole ImageNet2012 dataset I think.