Missing callbacks

[khatchad]

Missing Strategy.run()

Called here: https://github.com/mead-ml/mead-baseline/blob/5d7632bb151c2d09501ebf49f36ba8c4204df4c8/mead/api_examples/pretrain_discrim_tf.py#L414.

The callback function _replicated_train_step() is defined here: https://github.com/mead-ml/mead-baseline/blob/5d7632bb151c2d09501ebf49f36ba8c4204df4c8/mead/api_examples/pretrain_discrim_tf.py#L394-L405.

But, we don't see it in the call graph. The method reference should be: < PythonLoader, Lscript pretrain_discrim_tf.py/train/_replicated_train_step, do()LRoot; >. The call graph nodes: https://gist.github.com/khatchad/ab56e3be454103829275c7507999b7d1

Regression

There are some callbacks defined in tensorflow.xml, as well as others. I see the return values are specified, but I wonder if adding the callback summary will also consider the given function argument as invoked:

ML/com.ibm.wala.cast.python.ml/data/tensorflow.xml

Lines 758 to 766 in 1b1ffac

	<package name="tensorflow/estimator">
	<class name="Estimator" allocatable="true">
	<method name="do" descriptor="()LRoot;" numArgs="2" paramNames="self model">
	<new def="x" class="Ltensorflow/estimator/train/train" />
	<putfield class="LRoot" field="train" fieldType="LRoot" ref="self" value="x" />
	<putfield class="LRoot" field="$callback" fieldType="LRoot" ref="x" value="model" />
	<return value="arg0" />
	</method>
	</class>

ML/com.ibm.wala.cast.python.ml/data/tensorflow.xml

Lines 778 to 787 in 1b1ffac

	<package name="tensorflow/estimator/train">
	<class name="train" allocatable="true">
	<method name="do" descriptor="()LRoot;" numArgs="3">
	<getfield class="LRoot" field="$callback" fieldType="LRoot" ref="arg0" def="xx" />
	<getfield class="LRoot" field="data" fieldType="LRoot" ref="arg1" def="data" />
	<call class="LRoot" name="do" descriptor="()LRoot;" type="virtual" arg0="xx" arg1="data" arg2="3" numArgs="3" def="v" />
	<return value="v" />
	</method>
	</class>
	</package>

The ones there are for estimators (looks like there's two definitions). I wonder if we just need to add more for, e.g., TF2 APIs. I found these docs for the estimator.

[khatchad]

Actually, above, we can see that xx represents the passed function (arg0; but represented as a field?):

ML/com.ibm.wala.cast.python.ml/data/tensorflow.xml

Line 781 in 1b1ffac

<getfield class="LRoot" field="$callback" fieldType="LRoot" ref="arg0" def="xx" />

Then, a call is added to this function:

ML/com.ibm.wala.cast.python.ml/data/tensorflow.xml

Line 783 in 1b1ffac

<call class="LRoot" name="do" descriptor="()LRoot;" type="virtual" arg0="xx" arg1="data" arg2="3" numArgs="3" def="v" />

The first argument of the original function is passed:

ML/com.ibm.wala.cast.python.ml/data/tensorflow.xml

Line 782 in 1b1ffac

<getfield class="LRoot" field="data" fieldType="LRoot" ref="arg1" def="data" />

[khatchad]

For experimental_distribute_datasets_from_function, that function:

Generates tf.distribute.DistributedValues from value_fn.

Tracking DistributedValues isn't currently supported.