About loss...

[AMzhanghan]

Thanks for your great work! When I train, I report an error： ValueError: When passing a list as loss, it should have one entry per model outputs. The model has 1 outputs, but you passed loss=['mse', <bound method WCT2.gram_loss of <model.WCT2 object at 0x7fa85e2974d0>>]

After modification：(self.wct.compile(optimizer=Adam(self.lr), loss=["mse", self.gram_loss]) ) to（self.wct.compile(optimizer=Adam(self.lr), loss=[self.gram_loss])）

It works;

But the loss is very bigger, What's the problem？ thx

Epoch 1/10

1/1250 [..............................] - ETA: 1:54:33 - loss: 3585171783680.0000 2/1250 [..............................] - ETA: 1:01:06 - loss: 2059623989248.0000 3/1250 [..............................] - ETA: 43:16 - loss: 1529374932992.0000
4/1250 [..............................] - ETA: 34:19 - loss: 1247480791040.0000 5/1250 [..............................] - ETA: 28:56 - loss: 1053487792128.0000 6/1250 [..............................] - ETA: 25:21 - loss: 901079168341.3334 7/1250 [..............................] - ETA: 22:47 - loss: 797250371584.0000 8/1250 [..............................] - ETA: 20:52 - loss: 704853544960.0000 9/1250 [..............................] - ETA: 19:22 - loss: 639602908273.7778 10/1250 [..............................] - ETA: 18:11 - loss: 579075877683.2000 11/1250 [..............................] - ETA: 17:12 - loss: 532678425506.9091 12/1250 [..............................] - ETA: 16:23 - loss: 492782063616.0000 13/1250 [..............................] - ETA: 15:42 - loss: 458466281944.6154

[ptran1203]

Could you tell me your Tf version?

[ptran1203]

But the loss is very bigger, What's the problem？ thx

It's Ok, Maybe I will add some visualization after a period of epochs to check if the result is good

[AMzhanghan]

Could you tell me your Tf version?

1.15.0

[ptran1203]

I train this model on Tf2.

[ptran1203]

I recommend you upgrade to TF2 or fix like this

https://stackoverflow.com/questions/51705464/keras-tensorflow-combined-loss-function-for-single-output

You need two loss functions: Mean_squared_error and gram_loss, You may want to combine them to single loss

[AMzhanghan]

How many epochs have you trained ？

[AMzhanghan]

I trained 10 epochs， and the results vis is very bad

[ptran1203]

Your modification self.wct.compile(optimizer=Adam(self.lr), loss=[self.gram_loss]) is missing mean_squared_error loss function

[ptran1203]

python3 train.py --train-tfrec /content/tfrecords/train.tfrec\
                --val-tfrec /content/tfrecords/val.tfrec\
                --epochs 100\
                --resume\
                --batch-size 8\
                --lr 2e-4\

I trained with this script

[AMzhanghan]

Your modification self.wct.compile(optimizer=Adam(self.lr), loss=[self.gram_loss]) is missing mean_squared_error loss function

in your code I not find the 'mean_squared_error'

[AMzhanghan]

Your modification self.wct.compile(optimizer=Adam(self.lr), loss=[self.gram_loss]) is missing mean_squared_error loss function

in your code I not find the 'mean_squared_error'

I fixed the bug; update tf 2.2 and I find the begin loss of [loss: 5860.0898]; is bigger

[ptran1203]

Does the result look better?

[AMzhanghan]

Still training

[AMzhanghan]

are you Vietnamese?

[ptran1203]

are you Vietnamese?

Yes

[AMzhanghan]

Nice to meet you. Is your research GAN？

[ptran1203]

yes, a little bit during my thesis

[AMzhanghan]

What social software do you use? We can communicate

[ptran1203]

Do you use Skype?

[AMzhanghan]

no, Wechat do you use ?