How does GradientTape "record" operation

TensorFlow General Discussion
,
1

Hello.

Would like to know how exactly does the GradientTape context, record operations.

Browsing through github i followed the code to the C definition of the GradientTape PyObject, and functions to manipulate the tape and operations.

But what exactly triggers the recording.

The comments in the relevant files mention context / trace with regards to recording operations.

Any directions would be helpful.

Best regards.

3

Hi @Nihal_Kenkre, recording of gradients in GradientTape starts when tf.GradientTape.watch is called. Thank You.

4

Some more information on the question.

This is the GradientTape doc - tf.GradientTape  |  TensorFlow v2.16.1

It records the operations happening inside the GradientTape context.

Any pointers on how this is possible would be helpful.

There is python code for gradient tape which contains a tape class which then calls C extensions for GradientTape and Tape.

e.g. Tape Class

But i am not able to figure out the “Record Operations Trigger” which happens within the GradientTape context.

Is there a way to use the trace and traceback modules in python to do something like this? but how would they know where to start and stop.

Cheers