so hard to train

[buptlj]

Hi, I use the preprocessed CelebA dataset that you provided to train, but the attention masks become 1 after several iters. I didn't change the parameters in the code.
lambda_cls=160, lambda_rec=10, lambda_gp=10, lambda_sat=0.1, lambda_smooth=1e-4.
When I change lambda_sat to 1, the attention masks also saturate to 1.
But when I change lambda_sat to 1.5, the attention masks become 0.
Your pretrained model works well，did you use the same parameters？
Thanks！

[vipermu]

Hey @buptlj glad that at least the pre-trained model worked :)

When you set lambda_sat to 1.5 it becomes 0? This is really weird....

Are you using a batch size of 16 when you try the default hyperparameter values? Could you also try them with a batch size of 25?

[buptlj]

I didn't change the params and the batch size is 16.
I also tried them with a batch size of 25, but still not work.
I also tried lambda_cls=4000, but the results are almost the same with lambda_cls=160.
The attention mask is either too white or too dark.
I got the attention mask:

[vipermu]

Okay this has happened to me before. In my case the problem was because I had not computed the Action Units of each face correctly.

You mentioned that you were using the dataset that I've provided so that shouldn't be a problem (I trained my models with those images and Action Units).

Do you think that somehow you can be doing something different that makes that the pair AU-image do not correspond anymore?

I also assume that you are using all the images for training right?

[buptlj]

Finally, I get the correct result.
You mentioned that you got the above problem because you had not computed the Action Units of each face correctly, so I use openface to compute the AUs and everything goes right.
I find that the AUs I get are different from yours.
For example, your face img and AUs:

0.56 0.03 0.32 0.84 1.23 0.70 0.31 0.64 1.89 0.75 0.00 0.00 0.45 0.00 1.53 0.04 0.04
My result:

0.00 0.00 0.00 0.00 0.50 0.80 0.00 0.15 1.08 1.46 0.02 0.62 1.14 0.00 0.07 0.14 0.00
The two faces are almost the same，but the AUs are different.

[vipermu]

I see, it's weird because I uploaded the file I used to train my final models. I'll check it out though, thanks!

[duxingren14]

Finally, I get the correct result.
You mentioned that you got the above problem because you had not computed the Action Units of each face correctly, so I use openface to compute the AUs and everything goes right.
I find that the AUs I get are different from yours.
For example, your face img and AUs:

0.56 0.03 0.32 0.84 1.23 0.70 0.31 0.64 1.89 0.75 0.00 0.00 0.45 0.00 1.53 0.04 0.04
My result:

0.00 0.00 0.00 0.00 0.50 0.80 0.00 0.15 1.08 1.46 0.02 0.62 1.14 0.00 0.07 0.14 0.00
The two faces are almost the same，but the AUs are different.

Could you please also share the AUs you generated. I would like to try to reproduce the results. Thanks.

[pytyeah]

Could you show me the correct attention mask and color mask that you've obtained? I think my result is a little weird. Thanks for your help!

[Changezi001]

Finally, I get the correct result.
You mentioned that you got the above problem because you had not computed the Action Units of each face correctly, so I use openface to compute the AUs and everything goes right.
I find that the AUs I get are different from yours.
For example, your face img and AUs:

0.56 0.03 0.32 0.84 1.23 0.70 0.31 0.64 1.89 0.75 0.00 0.00 0.45 0.00 1.53 0.04 0.04
My result:

0.00 0.00 0.00 0.00 0.50 0.80 0.00 0.15 1.08 1.46 0.02 0.62 1.14 0.00 0.07 0.14 0.00
The two faces are almost the same，but the AUs are different.

I am also having the same problem of getting the attention maps saturated to 1 by setting all the parameters to default values and having the batch size of 16. Could you please share the AUs that you have generated?

[Sycamoretree]

Finally, I get the correct result.
You mentioned that you got the above problem because you had not computed the Action Units of each face correctly, so I use openface to compute the AUs and everything goes right.
I find that the AUs I get are different from yours.
For example, your face img and AUs:

0.56 0.03 0.32 0.84 1.23 0.70 0.31 0.64 1.89 0.75 0.00 0.00 0.45 0.00 1.53 0.04 0.04
My result:

0.00 0.00 0.00 0.00 0.50 0.80 0.00 0.15 1.08 1.46 0.02 0.62 1.14 0.00 0.07 0.14 0.00
The two faces are almost the same，but the AUs are different.

I am also having the same problem of getting the attention maps saturated to 1 by setting all the parameters to default values and having the batch size of 16. Could you please share the AUs that you have generated?

Could you please share the AUs that you have generated? Because I have found my AUs are wrong.

[ForeverPs]

Great, I use OpenFace to re-generate the AUs for all images, and the results become normal. Great work!