KeyError: 'acc' on end of epoch about slideflow

[skochanny]

Have early_stop_value='loss' in HP

[jamesdolezal]

What is your Tensorflow version? And what branch are you on?

[skochanny]

tensorflow 2.0.0-beta1, on nightly branch. i switched to master but that doesn't have latest updates we've made like adding the newest hyperparameters

[jamesdolezal]

This is due to the fact that Tensorflow changed how the accuracy metric is stored in the "logs" dictionary, from 'acc' to 'accuracy'. The latest dev branch reflects this update but has not yet been merged into nightly. To circumvent the issue until I merge into nightly, search for all instances of 'acc' in slideflow/model.py and replace with 'accuracy'. There should be 9 instances.

I'll hopefully be able to merge my dev changes into nightly soon.

Edit: I'd also recommend upgrading Tensorflow to v2.1, as there was a security flaw found in 2.0 that has been subsequently patched. Requirements for the Slideflow package is now TF v > 2.1

[skochanny]

i replaced all of the 'acc' with 'accuracy' but that did not solve the problem. I deleted the acc: logs['accuracy'] part from the string in line 705 and that did solve the problem. i think that the logs object must not contain 'accuracy' (I wasn't sure what exactly logs was). The logs['loss'] works, so it's something with 'accuracy' specifically. So i'm not sure how to fix that, but it's working right now.
I upgraded to TF 2.1, which required update on the machine to CUDA 10.1.

[jamesdolezal]

What outcomes are you training to, and are the outcomes linear or categorical?

[skochanny]

right now, linear. haven't tested with categorical

[jamesdolezal]

Figured it out. It was a problem in the logging of the exponential moving average, which assumed you were using a categorical model. I've now fixed it so that it will work with linear models. Doing some final bug checking before pushing the update, will let you know when the update hits the dev or nightly branch.

[jamesdolezal]

Fixed in c856426