architecture derived problem about densenas HOT 6 CLOSED

Hi, thanks for your interest in our work. Could you please provide the complete searching log file? Thus, I can get detailed information about your running. In addition, how many GPUs do you use in your experiment?

from densenas.

log.txt
I’m so glad you can help me out of this LOL.
I use only one gpu since I got a bug when I use multi-gpu. I guess maybe it's my pytorch version did not match(I use pytorch 1.7.0/cuda10.2).
Btw, I miss another bug when I run the search script before I got the log file.
That is, in OPTIMIZER.arch_step, I change the parameter 'search_stage' from 2 to 1 in function 'sample_branch', since it return a bug in function 'comp_rescale_value'. And I dont know if this change makes some difference of the derived model.
Thanks a lot!

from densenas.

Hi! Could please provide all the config details of your experiment, which should have been written into the log file but I didn't find it? In addition, running with a single GPU may cause the batch size too small to support the optimization. And what is the bug you mentioned on comp_rescale_value? Manually changing the sample_branch parameter may lead to some problems.

from densenas.

###############
12/11 15:59:08 epoch 0 arch_lr 3.000000e-04
Traceback (most recent call last):
File "/hdd1/wyn/DenseNAS/run_apis/search.py", line 158, in
super_model, epoch, 'Arch', search_stage)
File "/hdd1/wyn/DenseNAS/run_apis/trainer.py", line 153, in train
logits, loss, sub_obj = self.search_optim.arch_step(input, target, model, search_stage)
File "/hdd1/wyn/DenseNAS/run_apis/optimizer.py", line 54, in arch_step
model)
File "/hdd1/wyn/DenseNAS/run_apis/optimizer.py", line 113, in rescale_arch_params
rescale_params(alpha_head_weights_drop, model.module.alpha_head_weights, alpha_head_index)
File "/hdd1/wyn/DenseNAS/run_apis/optimizer.py", line 109, in rescale_params
indices_branch)
File "/hdd1/wyn/DenseNAS/run_apis/optimizer.py", line 100, in comp_rescale_value
rescale_mat = torch.zeros_like(new_weights).scatter_(0, index, rescale_value)
RuntimeError: Expected index [2] to be smaller than self [6] apart from dimension 0 and to be smaller size than src []

Process finished with exit code 1

###############
config.txt

Hi! Im not sure if these are enough. Please tell me if I miss some important information.

from densenas.

The higher version of PyTorch seems to carry incompatibility issues, as also mentioned in issues-17. We have tested the code on PyTorch 1.0 - 1.2. Could you have a try with a lower PyTorch version with the defaut parameters, e.g. search_stage, not changed, if needed?

from densenas.

The higher version of PyTorch seems to carry incompatibility issues, as also mentioned in issues-17. We have tested the code on PyTorch 1.0 - 1.2. Could you have a try with a lower PyTorch version with the defaut parameters, e.g. search_stage, not changed, if needed?

Thx a lot for your kind advice. I'll have a try then!

from densenas.