A problem about the loss and accuracy.

[Hanzy1996]

When I train the VAE, the loss decreases first, but then the loss increases and will never decrease any more. For example, during training: epoch 0 | iter 0 | loss 8543. epoch 0 | iter 50 | loss 5377. epoch 0 | iter 100 | loss 4846. epoch 1 | iter 0 | loss 4675. .... epoch 19 | iter 0 | loss 3503. epoch 19 | iter 50 | loss 3372. epoch 19 | iter 100 | loss 3357. epoch 20 | iter 0 | loss 3681. ... epoch 35 | iter 0 | loss 6382. epoch 35 | iter 50 | loss 6518. epoch 35 | iter 100 | loss 6475. ... epoch 99 | iter 0 | loss 13399 epoch 99 | iter 50 | loss 12820 epoch 99 | iter 100 | loss 13844

Then, during test, the results are not good, like: [0.0] novel=0.0020, seen=0.0004, h=0.0007 , loss=5.0318 [1.0] novel=0.0024, seen=0.0014, h=0.0018 , loss=4.7751 [2.0] novel=0.0014, seen=0.0074, h=0.0023 , loss=4.5471 ... [20.0] novel=0.0020, seen=0.0351, h=0.0038 , loss=2.5270 [21.0] novel=0.0020, seen=0.0345, h=0.0038 , loss=2.4551 [22.0] novel=0.0020, seen=0.0369, h=0.0038 , loss=2.4068

How could I fix this problem?

[Hanzy1996]

SEE https://github.com/edgarschnfld/CADA-VAE-PyTorch/issues/3#issuecomment-500818861

[Hanzy1996]

I have got the similar results with the paper. May I ask why the choice of pytorch have so huge effect on the performance? Since the effect can't be ignored, which version of pytorch is the best choice until now?

[edgarschnfld]

I only worked with pytorch 0.4.1, so that seems the best choice. Apart from that, I do not know why different pytorch versions behave differently.

[Hanzy1996]

Much appreciation! One last confusion, do you use python2 or python3?

[edgarschnfld]

You're welcome :) I used python 3.5.6, to be very precise.

[Hanzy1996]

Thanks so much. I think the main problem is the version of python. The code works well with python3 no matter which pytorch I use. But the results are bad when I use python2. Maybe you could remind others that python3 is a better choice in your README. By the way, your work is so excellent and I am so inspired.

[edgarschnfld]

I updated the readme. Thank you for the kind words, and good luck with your work :)

[divyanshusharma1709]

Hi! I know this issue is really old, but I am having the same error as @Hanzy1996. I have setup the environment exactly as mentioned by the author, but still get these results:

[0.0] novel=0.0000, seen=0.0182, h=0.0000 , loss=4.0405 [1.0] novel=0.0000, seen=0.0180, h=0.0000 , loss=3.9586 [2.0] novel=0.0000, seen=0.0177, h=0.0000 , loss=3.8913 [3.0] novel=0.0000, seen=0.0168, h=0.0000 , loss=3.8519 [4.0] novel=0.0000, seen=0.0165, h=0.0000 , loss=3.8334 [5.0] novel=0.0000, seen=0.0162, h=0.0000 , loss=3.7746 [6.0] novel=0.0000, seen=0.0166, h=0.0000 , loss=3.7464 [7.0] novel=0.0000, seen=0.0176, h=0.0000 , loss=3.7225 [8.0] novel=0.0000, seen=0.0166, h=0.0000 , loss=3.7080 [9.0] novel=0.0000, seen=0.0165, h=0.0000 , loss=3.6731 [10.0] novel=0.0000, seen=0.0175, h=0.0000 , loss=3.6680 [11.0] novel=0.0000, seen=0.0171, h=0.0000 , loss=3.6450 [12.0] novel=0.0011, seen=0.0162, h=0.0020 , loss=3.6349 [13.0] novel=0.0022, seen=0.0171, h=0.0038 , loss=3.6105 [14.0] novel=0.0022, seen=0.0173, h=0.0038 , loss=3.6129 [15.0] novel=0.0022, seen=0.0176, h=0.0038 , loss=3.6025 [16.0] novel=0.0032, seen=0.0175, h=0.0054 , loss=3.5328 [17.0] novel=0.0032, seen=0.0174, h=0.0054 , loss=3.5407 [18.0] novel=0.0022, seen=0.0175, h=0.0038 , loss=3.5225 [19.0] novel=0.0022, seen=0.0182, h=0.0038 , loss=3.5302 [20.0] novel=0.0022, seen=0.0184, h=0.0039 , loss=3.5368 [21.0] novel=0.0022, seen=0.0167, h=0.0038 , loss=3.5115 [22.0] novel=0.0032, seen=0.0170, h=0.0054 , loss=3.4782 [23.0] novel=0.0056, seen=0.0173, h=0.0085 , loss=3.4906 [24.0] novel=0.0043, seen=0.0175, h=0.0069 , loss=3.4371 [25.0] novel=0.0045, seen=0.0166, h=0.0071 , loss=3.4738 [26.0] novel=0.0067, seen=0.0175, h=0.0097 , loss=3.4174 [27.0] novel=0.0056, seen=0.0184, h=0.0086 , loss=3.4653 [28.0] novel=0.0067, seen=0.0177, h=0.0097 , loss=3.4711 [29.0] novel=0.0077, seen=0.0175, h=0.0107 , loss=3.4394 [30.0] novel=0.0077, seen=0.0178, h=0.0108 , loss=3.4429 [31.0] novel=0.0099, seen=0.0174, h=0.0126 , loss=3.4128 [32.0] novel=0.0109, seen=0.0177, h=0.0135 , loss=3.4260 [33.0] novel=0.0120, seen=0.0178, h=0.0143 , loss=3.4201 [34.0] novel=0.0120, seen=0.0180, h=0.0144 , loss=3.3901 [35.0] novel=0.0109, seen=0.0177, h=0.0135 , loss=3.3965 [36.0] novel=0.0142, seen=0.0173, h=0.0156 , loss=3.4226 [37.0] novel=0.0142, seen=0.0174, h=0.0156 , loss=3.4043 [38.0] novel=0.0142, seen=0.0188, h=0.0162 , loss=3.4088 [39.0] novel=0.0152, seen=0.0188, h=0.0168 , loss=3.4133 [40.0] novel=0.0152, seen=0.0175, h=0.0163 , loss=3.4013 [41.0] novel=0.0174, seen=0.0179, h=0.0176 , loss=3.4220 [42.0] novel=0.0184, seen=0.0175, h=0.0180 , loss=3.4031 [43.0] novel=0.0163, seen=0.0185, h=0.0173 , loss=3.3607 [44.0] novel=0.0152, seen=0.0179, h=0.0165 , loss=3.3655 [45.0] novel=0.0152, seen=0.0180, h=0.0165 , loss=3.3808 [46.0] novel=0.0163, seen=0.0179, h=0.0170 , loss=3.3966 [47.0] novel=0.0195, seen=0.0187, h=0.0191 , loss=3.3460 [48.0] novel=0.0195, seen=0.0179, h=0.0187 , loss=3.3633 [49.0] novel=0.0206, seen=0.0172, h=0.0187 , loss=3.3768

[edgarschnfld]

Hi, I added cadavae.yaml, so with conda env create --file=cadavae.yml you will have a replica of my currently installed environment for this repository. Let me know if it is still not working after that.

[divyanshusharma1709]

Hi @edgarschnfld ! Thanks for the swift reply. It seems that I made a mistake in loading the data. Confirmed working for PyTorch 1.4 and Python 3.6

[edgarschnfld]

Great, thanks for confirming that it works for other PyTorch and Python versions :)