Keras: Progress bar doesn't show enough digits

[entron]

The error of my problem is around or below 0.01. As I am using mean_squared_error, the progress bar is on longer showing enough digits after the 2nd epoch as example below:

Epoch 1/20
819007/819007 [==============================] - 123s - loss: 0.0002 - val_loss: 0.0001
Epoch 2/20
819007/819007 [==============================] - 127s - loss: 0.0001 - val_loss: 0.0001
Epoch 3/20
819007/819007 [==============================] - 126s - loss: 0.0001 - val_loss: 0.0001
Epoch 4/20
819007/819007 [==============================] - 127s - loss: 0.0001 - val_loss: 0.0001

[smauq]

You can for example create a callback for on_epoch_end and print the loss yourself with whatever precision you want.

[entron]

Yes indeed. Still it would be good that the progress bar can automatically use scientific notation or other way to show the extra digits in this case.

[fchollet]

Another trivial solution is to increase your loss by an arbitrary factor
(e.g. x1000), by using a custom loss:

loss = lambda y_true, y_pred: 1000 * mse(y_true, y_pred)

On 7 December 2015 at 11:56, Cheng Guo [email protected] wrote:

Yes indeed. Still it would be good that the progress bar can automatically
use scientific notation or other way to show the extra digits in this case.

—
Reply to this email directly or view it on GitHub
https://github.com/fchollet/keras/issues/1203#issuecomment-162641440.

[rpinsler]

I agree with @entron. What's the benefit of showing this useless output by default? Sure, it's easy to find a workaround for it, but I think that misses the point. I'd also go with scientific notation. Don't make common use cases harder than they really are.

[entron]

Thanks a lot for the answers and the quick fix! I am really amazed by the Keras community!