Reduce time for reproducibility and DGI save/load tests

[huonw]

The time for the 3.x pytest testing step seems to be bimodal at the moment: usually around 18 minutes, but sometimes hitting the 30 minute timeout. This may be due to the agent running out of burst credits, and thus running slower.

This causes problems both with the time it takes for the build to run and with the build failing due to the timeout. Cutting the test times will make development easier.

In https://buildkite.com/stellar/stellargraph-public/builds/5012, the 3.7 and 3.8 steps take about 30 minutes, but get in just under the timeout. Based on the JUnit XML files there, the slowest tests are:

3.7:

class	name	time
layer.test_deep_graph_infomax	test_dgi_save_load[HinSAGE]	189.596
reproducibility.test_graphsage	test_link_prediction[False]	104.020
reproducibility.test_deep_graph_infomax	test_dgi[True-APPNP]	48.177
layer.test_hinsage	test_hinsage_save_load	45.856
layer.test_deep_graph_infomax	test_dgi[False-HinSAGE]	39.722
reproducibility.test_graphsage	test_link_prediction[True]	39.340
reproducibility.test_graphsage	test_nai[True]	38.290
layer.test_graphsage	test_graphsage_save_load	34.116
test_ensemble	test_ensemble_init_parameters	32.852
test_ensemble	test_evaluate_link_prediction	32.474
reproducibility.test_graphsage	test_unsupervised[True]	31.503
test_ensemble	test_predict	28.937
layer.test_knowledge_graph	test_save_load[RotH]	27.562
test_ensemble	test_evaluate	27.181
reproducibility.test_deep_graph_infomax	test_dgi[False-APPNP]	26.203
reproducibility.test_deep_graph_infomax	test_dgi[True-GAT]	25.441
reproducibility.test_deep_graph_infomax	test_dgi[False-GAT]	24.703
test_ensemble	test_predict_link_prediction	23.889

3.8:

class	name	time
layer.test_deep_graph_infomax	test_dgi_save_load[HinSAGE]	196.829
reproducibility.test_graphsage	test_link_prediction[False]	77.198
reproducibility.test_graphsage	test_nai[True]	53.483
layer.test_hinsage	test_hinsage_save_load	45.713
reproducibility.test_deep_graph_infomax	test_dgi[True-APPNP]	43.655
layer.test_deep_graph_infomax	test_dgi[False-HinSAGE]	39.480
layer.test_graphsage	test_graphsage_save_load	38.373
reproducibility.test_graphsage	test_link_prediction[True]	36.478
test_ensemble	test_evaluate_link_prediction	32.547
test_ensemble	test_evaluate	32.126
test_ensemble	test_ensemble_init_parameters	31.851
test_ensemble	test_predict	27.299
layer.test_knowledge_graph	test_save_load[RotH]	27.001
reproducibility.test_deep_graph_infomax	test_dgi[True-GAT]	26.443
test_ensemble	test_predict_link_prediction	23.212
test_ensemble	test_BaggingEnsemble_fit	22.931
reproducibility.test_deep_graph_infomax	test_dgi[False-GAT]	22.925
reproducibility.test_graphsage	test_nai[False]	22.844

Clearly saving/loading a HinSAGE Deep Graph Infomax model is extremely slow, as are all of the reproducibility tests. This PR removes the HinSAGE DGI test save-load test, and reduces the repeats for the reproducibility ones to just two: the reference model and then one repeat. (The reproducibility tests are quite flaky at the moment, so it's kinda a waste to spend a lot of CPU time on them.)

[codeclimate]

Code Climate has analyzed commit 6634af0 and detected 0 issues on this pull request.

View more on Code Climate.

[timpitman]

This looks 4 to 5 minutes faster 🚀