Cleanup examples

[xnuohz]

• Combine 15 examples related to cora into 1 file
• Update related docs, github workflow, readme

[puririshi98]

code changes look good at first review, but for such an extensive change i would ask that you run every one of the deleted examples, and then the new version with its corresponding args for this planetoid example. I would ask that you share those logs and also check yourself to make sure that the new version is as good or better in terms of accuracy, speed, and readability. and if not for any reason try to understand why. This is a very valuable PR but we want to be careful with something like this.

If compute is an issue, i would ask that you write a bash script for me that will automatically run the above experiments and have the logs saved to some shared cloud drive or a github repo so you can do the above mentioned anlaysis after i run the script.

regardless once you have said analysis, i will look through myself as well and make sure our analyses align. then i think this is safe to merge

[puririshi98]

also please update the changelog

[akihironitta]

Looks great so far! ✨

Tbh, I'd vote for having many separate 90~100 lines of code examples to allow users to immediately understand what the script does and easily modify the scirpt that they're interested in (at the cost of maintaining duplicate code in the example directory), but I guess this is just my preference.

Could we also add type annotations?

[akihironitta]

Does this script need the package in all cases?

[xnuohz]

updated and only judged when gnn_choice = splinegnn

[akihironitta]

Suggested change

# define args

[akihironitta]

I prefer to remove comments that don't add any information.

Suggested change

# detect device

[akihironitta]

Let's accumulate the loss as a tensor and not as a float so that we don't need to cuda sync every iteration:

Suggested change

	total_loss = 0
	for pos_rw, neg_rw in loader:
	optimizer.zero_grad()
	loss = model.loss(pos_rw.to(device), neg_rw.to(device))
	loss.backward()
	optimizer.step()
	total_loss += loss.item()
	return total_loss / len(loader)
	total_loss = torch.zeros(1, device=device).squeeze()
	for pos_rw, neg_rw in loader:
	optimizer.zero_grad()
	loss = model.loss(pos_rw.to(device), neg_rw.to(device))
	loss.backward()
	optimizer.step()
	total_loss += loss
	return total_loss.item() / len(loader)

[akihironitta]

nit for consistency:

Suggested change

	start = time.time()
	start = time.perf_counter()

[xnuohz]

@puririshi98 @akihironitta ready for review, thank you

1. total 14 models, 4 improved, 9 keep the same, 1 worse
   overall.csv
2. the new file aligns config and params of the old file, when ready to merge, should delete the redundant parts, which may affect the metrics
3. the old file does not init model params explicitly, while i updated in the new file. commented out for aligning config and params of the old file
4. script for the log of new file, refer to planetoid_train.sh,
   new_logs.zip
5. script for the log of old file, refer to run_cora.sh
   old_logs.zip

[puririshi98]

Tbh, I'd vote for having many separate 90~100 lines of code examples to allow users to immediately understand what the script does and easily modify the scirpt that they're interested in (at the cost of maintaining duplicate code in the example directory), but I guess this is just my preference.

I think that is a valid point, instead of just deleting them all, @xnuohz could you leave them in and update the readme to say that

"planetoid_train.py combines (insert list of python examples it covers) into a single unified framework that can all be run run with planetoid_train.sh. (insert description of the planetoid datasets it covers in general and explain why they can all be covered with a unified codebase)".

@akihironitta thoughts?

[puririshi98]

also @xnuohz please figure out why that 1 example is worse and add some special condition to the framework that fixes the issue. this should be considered a strict requirement for merging

[xnuohz]

after added model params init, their results are consistent

[puririshi98]

okay lets wait till @akihironitta replies to my previous comment about restructuring this PR to satisfy his concerns.