cvlab-yonsei/ewgs

An official implementation of "Network Quantization with Element-wise Gradient Scaling" (CVPR 2021) in PyTorch.

ewgs's Issues

About OOM issue related to 'create_graph=True'

Dear Author,

Above all, thank you for sharing nice codes.

BTW, about quant training on CIFAR10,
Have you ever faced with OOM issues by loss.backward(create_graph=True) in update_grad_scales?
When I tried it by below args, I was faced with the "RuntimeError: CUDA out of memory" issue.

python train_quant.py --gpu_id '0'
--weight_levels 8
--act_levels 8
--baseline False
--use_hessian True
--load_pretrain True
--pretrain_path '../results/ResNet20_CIFAR10/fp/checkpoint/last_checkpoint.pth'
--log_dir '../results/ResNet20_CIFAR10/ours(hess)/W8A8/

Do you have some idea to avoid this issue?

Thank you in advance.

Imagenet with Mobilenetv2

Can you please provide the script to reproduce the results in Table 5 for STE and EWGS on Imagenet trained using Mobilenetv2?

scaling_factor seems no affect on grad

pretrained models not available

Hello,
Thanks for the great work.
I'm trying to run CIFAR code as shown below from run.sh.
However, I couldn't find a reference to the pretrain_path models.
Is there any way/hint to get them?

Thanks

python train_quant.py --gpu_id '0' \
                       --weight_levels 2 \
                       --act_levels 2 \
                       --baseline False \
                       --use_hessian True \
                       --load_pretrain True \
                       --pretrain_path '../results/ResNet20_CIFAR10/fp/checkpoint/last_checkpoint.pth' \
                       --log_dir '../results/ResNet20_CIFAR10/ours(hess)/W1A1/'

In detection task，bkwd factor usually appear to nan

When I train yolox，in EWGS with use_hessian，the final delta can reach at nan：

def backward(ctx, g):
diff = ctx.saved_tensors[0]
delta = ctx._scaling_factor
scale = 1 + delta * torch.sign(g)*diff
return g * scale, None, None

where the delta is calculated by hessian factor

Can this code do distributed traininig?

using model = torch.nn.DataParallel(model), buff_weight and buff_act cannot return correctly in update_grad_scales. @junghyup-lee

training time on ImageNet

Thanks for your great paper, can you share your total training time and the number of GPU used, or to say, GPU hours? because in my experiment of EWGS, it is very slow.

Training problems on ImageNet

Hello, author.Thank you so much for open-source code,I have some problems with this code.
First,I run the code based on the Cifar10 dataset, I get the same accuracy as the original paper.
Then,I use ImageNet dataset(ILSVRC2012) to do experiments,I do a 4bit quantization based on Resnet34,I train my model using pretrained model,and I didn't modify the network structure and Super parameter configuration,unfortunately,after 100epochs,I only get 71.4% (Top1) Classification accuracy,It is 73.9% in the original paper.
So I hope you can help me sincerely.

Shortcut quantization (+, *)

I'm leaving an issue to see if there's something I haven't caught in the code.

Did you guys quantize the results of the shortcut??

because at QBasicBlock, it does not have a quantization part after addition (EWGS/CIFAR10/custom_models.py Line 83~85).

Thanks.