gtx 1060 with 6G ram, out of memory about faceswap-gan

[shaoanlu]

You can

1. Reduce batch size.
2. Reduce the dense layer dimension from 1024 to 512.
3. Remove residual blocks in the decoder.

[chikiuso]

Hi Shaoanlu,
About (1), is it batchSize=8?
About (2), is it x = Dense(1024)(Flatten()(x)) and x = Dense(4 * 4 * 1024)(x)?

Thanks!

[shaoanlu]

Yes. Also, you can decrease the number of filters in Conv2D() by factor of 2. i.e., 512 to 256, 256 to 128 and so on.

[chikiuso]

Thanks for your help Shaoanlu, but I am not lucky enough as it consume 5.9G ram out of 6G once I started the training.

[chikiuso]

May I ask how many Gb of graphic card ram do you have? thanks.

[shaoanlu]

I used Tesla K80 for training, which provides 12GB vRAM.

[chilian1996]

You can

1. Reduce batch size.
2. Reduce the dense layer dimension from 1024 to 512.
3. Remove residual blocks in the decoder.

Sorry, I'm just a new user of faceswap. How do we adjust the dense layer dimension? Can we do it in the GUI Interface? Thanks!