Shap: Support for tf.keras?

[artemmavrin]

I am trying to run this example from the README:

https://github.com/slundberg/shap/blob/8f736c0cf9050183f3885db87f8fc1b4eb3833e9/README.md#L134-L149

I get the following error:

AttributeError: 'NoneType' object has no attribute 'backend'

which comes from this line:

https://github.com/slundberg/shap/blob/8f736c0cf9050183f3885db87f8fc1b4eb3833e9/shap/explainers/deep/deep_tf.py#L116

My SHAP and TensorFlow versions are

• shap: 0.32.0
• tensorflow: 2.0.0

I do not have standalone Keras installed (see below).

A full working example to reproduce the error is

import tensorflow as tf
import shap
import numpy as np

# Load data
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()
x_train = tf.image.convert_image_dtype(x_train, dtype=tf.dtypes.float32)
x_test = tf.image.convert_image_dtype(x_test, dtype=tf.dtypes.float32)
x_train = x_train.numpy()
x_test = x_test.numpy()

# Dummy model
model = tf.keras.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28)),
    tf.keras.layers.Dense(units=10),
])

background = x_train[np.random.choice(x_train.shape[0], 100, replace=False)]
e = shap.DeepExplainer(model, background)

Explanation

The culprit is the following lines:

https://github.com/slundberg/shap/blob/8f736c0cf9050183f3885db87f8fc1b4eb3833e9/shap/explainers/deep/deep_tf.py#L69-L76

I do not have standalone Keras installed (because I use tf.keras), so the import statement silently fails, and keras remains None. Later on, the following line is executed, without checking if keras is None:

https://github.com/slundberg/shap/blob/8f736c0cf9050183f3885db87f8fc1b4eb3833e9/shap/explainers/deep/deep_tf.py#L116

Question

Are there plans to add support for tf.keras? (NB: François Chollet recommends switching Keras code to tf.keras.)

Addendum

If I do install standalone Keras (version 2.3.1), then my example still fails, this time with the following error:

TypeError: Tensor is unhashable if Tensor equality is enabled. Instead, use tensor.experimental_ref() as the key.

coming from this line:

https://github.com/slundberg/shap/blob/8f736c0cf9050183f3885db87f8fc1b4eb3833e9/shap/explainers/deep/deep_tf.py#L296

Like the error says, in TensorFlow 2, tensors are not hashable, and hence cannot be used as dictionary keys (tf.Tensor.__hash__ is not available; this has to do with the new behavior of tf.Tensor.__eq__ performing element-wise comparison). Related: #850 and #880.

[shan224]

Yep running into this as well

[marksal87]

I ran into this today, too.

[nboley]

This is an issue for me as well.

[slundberg]

The fundamental issue here is that DeepExplainer does not yet support TF 2.0. GradientExplainer does, so I recommend using that until we build the new version of DeepExplainer for TF 2.0 (we are working on it).