TensorFlow functions with Keras

Reading time ~2 minutes


Recently, I was trying to use Cohen’s Kappa as a metric with Keras. I decided I would use the TensorFlow contrib function that already existed. While trying to get TensorFlow working with Keras, I discovered there were no easily-findable documents describing how to do this. The example from Keras’ blog is a few years old, and wasn’t working anymore. So after figuring out how to get TensorFlow working with Keras, I decided to document it (for the children).

Why use TensorFlow with Keras? TF, particularly the contrib portion, has many functions that are not available within Keras’ backend. Ideally you’d want to use Keras’ backend for things like TF functions, but for creating custom loss functions, metrics, or other custom code, it can be nice to use TF’s codebase.

The usual route

Usually you want to use Keras’ backend to implement custom functions. For example:

import keras.backend as K

def customLoss(yTrue,yPred):
    return K.sum(K.log(yTrue) - K.log(yPred))

(borrowed from here)

However, the backend will be missing many functions from TensorFlow, like much of the contrib section. Functions are also named differently.

As a solution, you can use TensorFlow with Keras in the following manner:

import keras
import tensorflow as tf
from keras import backend as K

# build model...(not shown)

# custom metric with TF
def cohens_kappa(y_true, y_pred):
    y_true_classes = tf.argmax(y_true, 1)
    y_pred_classes = tf.argmax(y_pred, 1)
    return tf.contrib.metrics.cohen_kappa(y_true_classes, y_pred_classes, 10)[1]

# compile model...
              metrics=['accuracy', cohens_kappa, tf_accuracy])

# IMPORTANT: init TF variables (only needed for some functions/implementations)

# fit model
model.fit(x_train, y_train,
          validation_data=(x_test, y_test))

All the examples I could find on the web (which weren’t many) used the tf.global_variables_initializer() function. It turns out this no longer works (it must’ve worked in 2016 when Keras made their blog post on the subject). Now it seems it should be tf.local_variables_initializer(). This can be run with Keras’ backend, or with TF directly. With Keras’ backend, it looks like it does in the code sample above (K.get_session() gets the TF session). Initializing the variables with TF directly would look more like:

sess = tf.Session()

There are a few examples in a Github repo I created here. This uses the Cohen’s Kappa metric from tf.contrib.

A few other small things

One other important thing for using TF metrics with Keras is you usually want the op_update object which is returned. Most (maybe all) TF metrics will return a tuple. The first part of the tuple is supposed to be the metric, and the second part an op_update. However, I found with Keras, the TF metrics’ op_update is the one that contains the actual metric value. So as you can see in the codeblock above, I’m getting the second part of the tuple for my metric: tf.contrib.metrics.cohen_kappa(y_true_classes, y_pred_classes, 10)[1].

The other small thing I learned through this is that to get the inverse of a one-hot encoded vector, all you have to use is argmax.

In the end, running TF with Keras turned out to be pretty simple. All that needs to be done is to initialiaze the variables with tf.local_variables_initializer() after creating any code with TF functions, and before running model.fit(). Again, this is only for certain TF functions that are creating/using TF variables within their source code. Some other TF functions don’t create any variables, so don’t require the initialization.

Good luck with your neural nets, and praise be to our future machine overlords.

Version details

Some small details: the TensorFlow version used was 1.9.0 (GPU), and Keras was 2.2.4. This was on an Ubuntu 18.0.4 LTS OS with Anaconda Python 3.6.8.

Sentiment analysis

A review of textblob's sentiment analysis, and a better way to do it with neural networks. Continue reading

Cannabis recommender and science

Published on November 20, 2016

Doing data science => Grupo Bimbo deliveries

Published on September 15, 2016