Update: The bug is caused by Serial option (light, medium and heavy), non-serial option resolved the bug. Further testing suggested that the bug might be related to the "remember" attribute not be able to turn on from a serial model.
@nicholas-leonard This problem has bothered me for 2 weeks now. The loss from the training indicate the learning is happening, but whenever I load the pre-trained model using another script, the sampling from the model (seq2seq using SeqLSTM) return gibberish.
So I did a test on an overfitting example, and the sampling still fails when loading from another script.
It turned out if I insert this piece of code right after the training script, the sampling work again. So it must be the error from the model saving function (or maybe the saving code I wrote is wrong? I have tried both Serial and the non-Serial option for saving.)
Update: The bug is caused by Serial option (light, medium and heavy), non-serial option resolved the bug. Further testing suggested that the bug might be related to the "remember" attribute not be able to turn on from a serial model.
@nicholas-leonard This problem has bothered me for 2 weeks now. The loss from the training indicate the learning is happening, but whenever I load the pre-trained model using another script, the sampling from the model (seq2seq using SeqLSTM) return gibberish.
So I did a test on an overfitting example, and the sampling still fails when loading from another script.
It turned out if I insert this piece of code right after the training script, the sampling work again. So it must be the error from the model saving function (or maybe the saving code I wrote is wrong? I have tried both Serial and the non-Serial option for saving.)
save file code:
model definition: