Keras: UnboundLocalError: local variable 'epoch_logs' referenced before assignment

[valexandersaulys]

Traceback (most recent call last):
  File "pretrained_mlp.py", line 102, in <module>
    batch_size=1);
  File ".../keras/keras/models.py", line 465, in fit
    shuffle=shuffle, metrics=metrics)
  File ".../keras/keras/models.py", line 228, in _fit
    callbacks.on_epoch_end(epoch, epoch_logs)
UnboundLocalError: local variable 'epoch_logs' referenced before assignment

Has anybody ever had this issue? It infrequently comes up for me when training with autoencoders. I'm using python 2.7.8 on Ubuntu if that means anything.

[w0nk0]

I've actually had the same issue yesterday, when reconstructing a model from a saved yaml. I didn't want to post anything here before I had more of an idea what could have caused it, and in which circumstances it did/didn't appear.

I was training a RNN without autoencoders, on 64-bit Python 3.4 on Win7.

From what I've gleaned from the source, epoch_logs is defined in the final batch of batch processing, and the error is thrown (as shown in your stack trace) when the end-of-epoch callbacks are activated.

[fchollet]

It can happen when the model is trained on empty data (which won't work anyway). I can turn that into a clearer error message.

[valexandersaulys]

Ah-hah! That was indeed my error, silly mistake. Thanks for the update!

[kevinmartinmauro]

This can also be due to having a negative batch size.

[luis-i-reyes-castro]

Hello! I wanted to report that I got this error too, and I've found that in my case it was (at least partially) caused by having steps_per_epoch = 0. In this case line 1852 of "training.py" fails to enter the while loop which defines 'epochs_log'.

[jeremykohn]

For me, the solution was a smaller batch size. Here's why:

history = model.fit_generator(train_generator,
                              steps_per_epoch=nb_train_samples//batch_size,
                              ...
                              )

Since I mistakenly used a batch_size larger than the number of training samples, nb_train_samples//batch_size was rounded down to 0. And as @luis-i-reyes-castro explains, steps_per_epoch must be above 0.