pytorch 0.4 compatibility

[quark0]

TODO

[robintibor]

What would be necessary for this? I would like to run it with pytorch 0.4 and willing to try to adapt necessary code :)

[quark0]

You may follow the official migration guide and/or refer to several forks of this repository where people have adapted the code for pytorch 4.0. It's primarily about getting rid of the depreciated volatile flag.

[robintibor]

Thanks, do you know one fork already where it was adapted?

[dragen1860]

@robintibor @cclauss I have adopted part of the source code from 0.3 to 1.0. you can try it by python train_search.py with pytorch 1.0.

[adammenges]

@dragen1860 when you say part of it what doesn't work?

Also for others, it's off your repo: https://github.com/dragen1860/DARTS-PyTorch

Submit a pull request?

[dragen1860]

@adammenges . Currently i just adopt the part of train_search on cifar10 related experiment. So when i say part of it I mean i will adopt the other exps, such as RNN part, in the future. It's not because not working.

[adammenges]

Sweet, thanks!

---

From: Jackie Loong [email protected] Sent: Wednesday, January 23, 2019 3:49 PM To: quark0/darts Cc: Adam Menges; Mention Subject: Re: [quark0/darts] pytorch 0.4 compatibility (#35)

@adammengeshttps://github.com/adammenges . Currently i just adopt the part of train_search on cifar10 related experiment. So when i say part of it I mean i will adopt the other exps, such as RNN part, in the future. It's not because not working.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHubhttps://github.com/quark0/darts/issues/35#issuecomment-457011669, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AAx5Hk3tnQAhAg0VSaV2tctBThdN-n58ks5vGPTwgaJpZM4WBuzf.

[dkumazaw]

@dragen1860 Thanks for the awesome work! If you do not mind sharing, aside from getting rid of volatile flags and taking care of syntax changes, did you have to do anything else to make it work? I am doing migration myself as well and encountering some memory issues...

[dragen1860]

@dkumazaw I did not meet the memory issues yet. I felt confused why did not I meet these issues~~. So I can not give you too much insight on how to overcome the bugs. But I test the train_search.py already and it works. Maybe you can compare my repo. to find it why works：

Epoch: 40 lr: 3.291796e-03
01/24 02:01:26 PM Genotype: Genotype(normal=[('max_pool_3x3', 0), ('sep_conv_5x5', 1), ('max_pool_3x3', 0), ('sep_conv_5x5', 1), ('sep_conv_3x3', 1), ('max_pool_3x3', 0), ('sep_conv_3x3', 0), ('sep_conv_5x5', 4)], normal_concat=range(2, 6), reduce=[('max_pool_3x3', 0), ('sep_conv_3x3', 1), ('max_pool_3x3', 0), ('dil_conv_5x5', 2), ('max_pool_3x3', 0), ('sep_conv_5x5', 3), ('max_pool_3x3', 0), ('skip_connect', 2)], reduce_concat=range(2, 6))
01/24 02:01:34 PM Step:000 loss:0.023659 acc1:100.000000 acc5:100.000000
01/24 02:08:00 PM Step:050 loss:0.017888 acc1:99.521760 acc5:100.000000
^[01/24 02:14:26 PM Step:100 loss:0.017079 acc1:99.529099 acc5:100.000000
01/24 02:20:52 PM Step:150 loss:0.017631 acc1:99.515425 acc5:100.000000

and rmb to tell me why.~~

[dkumazaw]

@dragen1860 Thanks! Let me move to your repo to discuss more

[alphadl]

I improved the code to make it compatible with PyTorch 1.1 while allowing multi-GPU training on both RNN and CNN experiments.~ you can refer: https://github.com/alphadl/darts.pytorch1.1