Debugging in TensorFlow

Unlike in Torch, the variables in TensorFlow are symbolic by nature.

Good slides about debugging in TF can be found here

The associated code can be found here

Basic Approaches

The most simple method is to convert tensors to numpy array and print.

For example x is a TF tensor,

x_np = sess.run(x)
print(x_np)

Here we convert x into a numpy object x_np.

Or, if x needs some feed_dict, we can

real_input = 199999
x_np = sess.run(x, feed_dict=real_input)
print(x_np)

Or we can use tf.InteractiveSession() to experiment in shell or Jupyter Notebook.

TensorBoard is too heavy.

Use tf.Assert() as often as possible.