ImageNet search question about pc-darts HOT 13 CLOSED

@xxsgcjwddsg ,the subset still has 1000 classes. Actually, we sampled 10% and 2.5% from each class for training and validation respectively. Yes, hyper-parameters are architecture weights.

from pc-darts.

@xxsgcjwddsg ,the subset still has 1000 classes. Actually, we sampled 10% and 2.5% from each class for training and validation respectively. Yes, hyper-parameters are architecture weights.

Thanks.
How do you search with multiple GPUs? I add model = nn.DataParallel(model) before model = model.cuda() and model = model.module after it, but still search on one GPU.

from pc-darts.

@xxsgcjwddsg ,you need to comment this line torch.cuda.set_device(args.gpu)

from pc-darts.

@xxsgcjwddsg ,you need to comment this line torch.cuda.set_device(args.gpu)

Thanks !
I has comment torch.cuda.set_device(args.gpu), but don't work.

from pc-darts.

Can you show the errors? I think maybe the error still comes from the model.module? E.g. in the SGD optimizer, it should be model.parameters()not model.module.parameters(). Besides, do you change the model.module in the train function and validation function?

from pc-darts.

Thanks a lot.
I shouldn't add "model = model.module" after "model = nn.DataParallel(model).cuda()"

from pc-darts.

Hi, when run architecture.step(), it encountered an error of OOM.
Thanks for your reply!

from pc-darts.

@xxsgcjwddsg ，just for sure that you use 8 v100gpus. Besides，I notice that your parameter size is nearly two times of mine, how many layers are stacked in the search period? We use 8 in our experiments. And the initial channels are 16.

from pc-darts.

Thanks for your reply.
I use 8 v100 gpus, but the first one takes up most. As the picture shows，the batch size is 256.
I think the reason is using self.model.module._loss(input, target). However, it will run only on one GPU.

from pc-darts.

@xxsgcjwddsg ，hi，maybe you need to remove the .to(..... .device) e.g. to(xtemp.device) code in the model_search_imagenet.py. I run this code in the company， may be there are differences within devices. If it worked, please tell me and I will update the code.

from pc-darts.

Thanks for your reply. It cann't works.
I have another question. The paper randomly sample two subsets from the 1.3M training set of ImageNet, with 10% and 2.5% images, respectively. The batch_size of valid_queue is also 1024? If it is 1024, then the frequency of architect.step() is 1/4 of optimizer.step()?

from pc-darts.

@xxsgcjwddsg ,hi, I have no idea now, I can run the code with 8 V100(16G each). Maybe you can add my wechat, we can talk about more details. Yes, and I also change the validation batch-size to balance the frequency and found no difference. Which version of pytorch you use by the way？

from pc-darts.

@yuhuixu1993, hi. I notice that the validation batch-size is same as the train batch-size.According to my understanding, it means that the valid dataset will be used four times?
Or another way， architect steps one time while optimizer steps four times?

from pc-darts.