Andrey Klochkov

Here's a colab demoing the problem. It's a modified [mnist tutorial](https://keras.io/examples/vision/mnist_convnet/) that shows that when a callback is used to run model validation more often than once per epoch, training...

The comment from `google-ml-butler` says to reopen the issue if it needs further work, but I don't seem to have permissions to do that. @[sushreebarsa](https://github.com/sushreebarsa), can you please reopen this?