the reimplement result of distgan+ss

[czzerone]

hello, I use the parameters provided in your code and paper and then want to reproduce your result of Dist-gan+SS, but the FID scores of 100K-500K are shown as below:
step: 10000 - FID (10K, 5K): 52.14751911204845
step: 20000 - FID (10K, 5K): 42.20208117605698
step: 30000 - FID (10K, 5K): 45.02294306743511
step: 40000 - FID (10K, 5K): 46.25833011921277
step: 50000 - FID (10K, 5K): 48.59530929102222

I think the FID scores seem to be a little large, how do you think that?

[tntrung]

Hi, it looks like you just run till 50K iterations, dist-gan needs some more iterations to converge. If there is still any problem up to 300K iterations, let me know I will check again the code? Thanks.

[czzerone]

I trained the model till 300K iterations with the batchsize=128, but I still got the FID scores a little large. I kept the same with your code except for the batch size.

step: 10000 - FID (10K, 5K): 56.55471167378875
step: 20000 - FID (10K, 5K): 48.948138169161425
step: 30000 - FID (10K, 5K): 49.36886535856645
step: 40000 - FID (10K, 5K): 44.443539741565274
step: 50000 - FID (10K, 5K): 45.39865798468646
step: 60000 - FID (10K, 5K): 44.162925041817815
step: 70000 - FID (10K, 5K): 41.791836263334524
step: 80000 - FID (10K, 5K): 42.625322239113
step: 90000 - FID (10K, 5K): 40.927866242410744
step: 100000 - FID (10K, 5K): 39.63991590152894
step: 110000 - FID (10K, 5K): 38.17071822170243
step: 120000 - FID (10K, 5K): 40.474300249710694
step: 130000 - FID (10K, 5K): 39.67866145187291
step: 140000 - FID (10K, 5K): 40.25284372060353
step: 150000 - FID (10K, 5K): 37.59208368572543
step: 160000 - FID (10K, 5K): 37.588404485573236
step: 170000 - FID (10K, 5K): 38.00086816808115
step: 180000 - FID (10K, 5K): 36.15899764367717
step: 190000 - FID (10K, 5K): 36.2833215392126
step: 200000 - FID (10K, 5K): 36.41073655210392
step: 210000 - FID (10K, 5K): 35.81622445631332
step: 220000 - FID (10K, 5K): 33.25562462943372
step: 230000 - FID (10K, 5K): 32.82274447747653
step: 240000 - FID (10K, 5K): 30.74282333899842
step: 250000 - FID (10K, 5K): 30.502626768164514
step: 260000 - FID (10K, 5K): 28.76506592514643
step: 270000 - FID (10K, 5K): 28.577641944714465
step: 280000 - FID (10K, 5K): 28.131259062217474
step: 290000 - FID (10K, 5K): 26.535512047756512
step: 300000 - FID (10K, 5K): 26.01967472361651

[tntrung]

Hi, not sure with this, we never tried with 128 batch size. What architecture are you using, e.g., dcgan or resnet? You may try to train more when having more samples in the batch, it looks the FID score is decreasing. Thanks.

[czzerone]

the architecture is resnet

…

---Original--- From: "Ngoc-Trung Tran"<[email protected]&gt; Date: Sat, Oct 24, 2020 14:15 PM To: "tntrung/msgan"<[email protected]&gt;; Cc: "Author"<[email protected]&gt;;"czzerone"<[email protected]&gt;; Subject: Re: [tntrung/msgan] the reimplement result of distgan+ss (#3) Hi, not sure with this, we never tried with 128 batch size. What architecture are you using, e.g., dcgan or resnet? You may try to train more when having more samples in the batch, it looks the FID score is decreasing. Thanks. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.