Reproduce ICML '21 experiments about programl

[ChrisCummins]

Hi @akash-isu, the input/output datasets for the experiments are available online. Unfortunately the code to repro the experiments is old and hasn't been ported to the latest version of ProGraML yet. Please see this comment for further details.

Cheers,
Chris

[akash-isu]

Hi Chris,

Thank you for your prompt response. I appreciate you pointing me to the right place.
However, I was looking more for the Device Mapping experiments. I know where to get the dataset. I was looking more for the model you implemented for the paper and would like to replicate the exact model. If you had any pointers to where I could find that, it'd be really helpful.

Thank you for your response and help.

Akash

[ChrisCummins]

Hi Akash, ah I see, I didn't realize that the older commit doesn't have the code for graph classification checked in! That's a shame. Hmm.

I was looking more for the model you implemented for the paper and would like to replicate the exact model. If you had any pointers to where I could find that, it'd be really helpful.

We used the same GGNN implementation for both node classification and graph classification. You could start by adapting the existing node classification code to graph classification by changing the way you construct the GGNN. For device mapping, graph_y_dimensionality=2 and node_y_dimensionality=0, for algorithmic classification graph_y_dimensionality=104 and node_y_dimensionality=0:

ProGraML/models/ggnn/ggnn.py

Lines 195 to 207 in a8fff63

	class Ggnn(Model):
	"""A gated graph neural network."""
	def __init__(
	self,
	vocabulary: Dict[str, int],
	node_y_dimensionality: int,
	graph_y_dimensionality: int,
	graph_x_dimensionality: int,
	use_selector_embeddings: bool,
	test_only: bool = False,
	name: str = "ggnn",
	):

Hope that helps.

Cheers,
Chris