急切执行 function 的输入不能是 Keras 符号张量，但可以找到

Question

I'm trying to obtain the jacobian of my model wrt an input (sample_x which is a binary vector in numpy).

print("Initiating gradient checker")
    sample_x_tensor = sample_x.toarray()
    sample_x_tensor = tf.convert_to_tensor(sample_x.toarray())
    sample_x_tensor = tf.cast(sample_x_tensor, tf.float32)

    with tf.GradientTape() as tape:
        tape.watch(sample_x_tensor)
        y_pred = model(sample_x_tensor)
        print(y_pred[0])

    jacobian = tape.jacobian(y_pred, sample_x_tensor)

Model is a straightforward Keras binary classification model, Keras 2.15 and Tensorflow 2. Getting the following exception:

tensorflow.python.eager.core._SymbolicException: Inputs to eager execution function cannot be Keras symbolic tensors, but found [<tf.Tensor 'Reshape:0' 
shape=(1,) dtype=float32>]

To my understanding, TF2 has eager execution by default. Any idea how I can rectify this?

Answer 1

Using the Code Snippet you provided and adding basic Keras Binary Classification Model.

Code used for reproduction:

%tensorflow_version 2.x  # Using Google Colab

import tensorflow as tf  # Tensorflow 2.2.0-rc3
from tensorflow.keras.layers import Dense
from tensorflow.keras.models import Sequential

model = Sequential()
model.add(Dense(32, input_shape = (2,)))
model.add(Dense(1, activation = 'sigmoid'))

print("Initiating gradient checker")
sample_x_tensor = tf.random.normal((100,2))
# sample_x_tensor = sample_x.toarray()
# sample_x_tensor = tf.convert_to_tensor(sample_x.toarray())
sample_x_tensor = tf.cast(sample_x_tensor, tf.float32)

with tf.GradientTape() as tape:
    tape.watch(sample_x_tensor)
    y_pred = model(sample_x_tensor)
    print(y_pred[0])

jacobian = tape.jacobian(y_pred, sample_x_tensor)

This returns:

Initiating gradient checker
tf.Tensor([0.38266268], shape=(1,), dtype=float32)

It executed successfully, but I would suggest a few things for you:

1. Check your Model Layers and see if it is processing the right data types .
2. Check your Input Data Shapes and Type , make sure that it is compatible with the layers used. (Tensor Datatype as much as possible)

Hope this helps you.