the predicted result is nan about keras-fcn

[JihongJu]

Hi, I'm not sure if I understand your problem correctly. By "parameters in some layers are ## nan", I assume you mean the weights in some layers are nan. Do you get nan losses in that case?

Usually nan loss (or inf loss) is caused by improper configuration of hyperparameters, for example, the learning rate is to0 large.

[kaijie-qin]

Hi, I find that the loss is nan and the parameters after conv5 are nan.

1. where can I set the learning rate?
   2.can you upload you trained model?

[JihongJu]

Hi,

1. You can change the arguments of the optimizer while fitting a data generator. See Keras Optimizers
2. Unfortunately, I don't have a trained model. At the moment, I am using a pre-trained model converted from the Caffe model zoo. (See #7)