Closed Debrove closed 4 years ago
Thanks for your interest at first. Would you mind giving me more details? Since this repo contains multiple implementations of different algorithms, I'm a little bit confused about which part are you referring to. Which script in this repo are you used for searching? Which script in this repo are you used for re-training? By "result", do you mean the error on CIFAR-10?
Sorry for carelessness.
For 1, I use ./scripts-search/DARTS1V-search-NASNet-space.sh
for searching and bash ./scripts/nas-infer-train.sh cifar10 DARTS_V1 96 -1
for re-training.
For 2, I use bash ./scripts/nas-infer-train.sh cifar10 DARTS_V1 96 -1
to re-train the genotype searched by quark0/darts.
For 3, I use bash ./scripts/nas-infer-train.sh cifar10 DARTS_V1 96 -1
to re-train the genotype searched by the script bash ./scripts-search/NASNet-space-search-by-GDAS.sh cifar10 1 -1
.
The reason I did these is to see whether the same genotypes will have similar results on different codes.(yours and quark0s)
Thank you for your quick reply.
I see. The re-training codes are almost the same as that in quark0/darts, but for the specific hyper-parameters, I can not remember all the details but I think I check it several months ago.
For 2, you said that my codes obtain "2.79+-0.11" but quark0/darts gets "2.99+-0.04"? Would it be a good thing, as you can obtain a better result using my codes :)
One question is that the decay in configs/opts/NAS-CIFAR-V2.config and configs/search-opts/DARTS-NASNet-CIFAR.config are 0.0005, while in the paper the decay is 0.0003. Is there any suggestion?
Well, it is interesting and thank you very much.
Nice catch! it should be my mistake
Since the weight decay value has been corrected in the recent commit, I'm closing this issue.
Hi, there are some confusions
1.I try to reproduce the DARTS-V1 results by your codes, but final result is
3.28+-0.04
. 2.I replaced the DARTS genotype searched by quark0/darts with DARTS_V1 and train it by your codes. Finally the result is2.79+-0.11
. It costs 35 hours. 3.I replaced the GDAS genotype with DARTS_V1 and train it. The final result is 2.72. It costs 38 hours to train.For 1, I guess the searched architecture is not good enough though I have searched four times and pick up the best. For 2, it is lower than the result train by quark0/darts, which is
2.99+-0.04
and costs 41 hours. For 3, the reproduced GDAS result is2.89+-0.05
(costs 54 hours to train) , which is higher than the result in 3.So, for training phase, is any differences from quark0/darts?(I can not find any differences if I did not miss something.)
Thank you very much.