在Google Colab中tf.case无法正确二次运行的原因咨询

I'm currently learning Stanford's CS20 TensorFlow course on Google Colab and encountered an unusual issue with tf.case:

1. My initial code ran correctly with this line:

   out_1b = tf.case({tf.less(x1, y1): lambda: tf.add(x1, y1), tf.greater(x1, y1): lambda: tf.subtract(x1, y1)}, default=lambda: tf.constant(0.0), exclusive=True)
   

2. When I changed tf.constant(0.0) to tf.constant(0), I got the expected type mismatch error (since the other branches return float tensors, while this returns an integer tensor).
3. But after switching back to tf.constant(0.0), I started getting this error:

   ValueError: Operation 'case_2/cond/Merge' has been marked as not fetchable.
   

4. The only way to resolve this was to restart my Colab session entirely. I'm hoping to understand why this happens.

Let me break down what's going on here—this is a common quirk tied to how TensorFlow manages computation graphs in Colab's persistent execution environment:

Why this error occurs after fixing the type issue #

When you first modified the default branch to return tf.constant(0), TensorFlow attempted to build a computation graph where the output types of the tf.case branches didn't align. This invalid graph construction caused TensorFlow to internally mark the Merge operation (a core part of how tf.case handles branching logic) as un-fetchable, since it couldn't resolve the type conflict.

Even after you fixed the type mismatch by reverting to 0.0, TensorFlow doesn't automatically clear the corrupted state of that existing graph node. Since Colab keeps your TensorFlow execution context alive between code cell runs, re-running the code tries to reuse the broken graph structure—hence the error about the operation being marked as not fetchable.

Restarting your Colab session wipes out all existing graph state completely, allowing you to build a fresh, valid graph from scratch, which is why that resolves the problem.

How to avoid this in the future #

• Instead of editing and re-running the same code cell, try writing your corrected code in a new cell. This minimizes the chance of reusing broken graph nodes from previous failed runs.
• If you need to re-run modified tensor operations, you can explicitly reset the default graph before executing your code again:

  tf.compat.v1.reset_default_graph()
  

  (Note: This is most relevant if you're working in graph execution mode. For eager execution (TensorFlow 2.x's default), simply redefining your tensors and operations from scratch should also work.)

内容的提问来源于stack exchange，提问作者Dongqing