ClearML support for configuration

[pyb0924]

❓ Questions/Help/Support

Hello! I'd like to know how to transfer the ClearML configuration logs from pure PyTorch to PyTorch-Ignite, that is:

task = Task.init("project-name", task_name="task-name")
task.connect_configuration(config) # or task.connect({k: v for k, v in config.items()})

In official docs, I only found ClearMLLogger class which implemented Task.init() inside, but with no support for Task.connect(). Is there any good solution to this problem in current verison of release?

Thanks!

[vfdev-5]

@pyb0924 thanks for asking! At the moment we do not expose task publicly from ClearMLLogger. There is a private argument ClearMLLogger._task but it is not advised to use it... In our examples, we just import Task from clearml package directly:

ignite/examples/contrib/cifar10/main.py

Lines 49 to 68 in ad37a17

	if config["with_clearml"]:
	try:
	from clearml import Task
	except ImportError:
	# Backwards-compatibility for legacy Trains SDK
	from trains import Task
	task = Task.init("CIFAR10-Training", task_name=output_path.stem)
	task.connect_configuration(config)
	# Log hyper parameters
	hyper_params = [
	"model",
	"batch_size",
	"momentum",
	"weight_decay",
	"num_epochs",
	"learning_rate",
	"num_warmup_epochs",
	]
	task.connect({k: config[k] for k in hyper_params})

I agree that it could be helpful to expose the task from ClearMLLogger. Actually, we could do 2 things:

1. Like TensorboardLogger

   ignite/ignite/contrib/handlers/tensorboard_logger.py

   Lines 163 to 164 in ad37a17

   	def __getattr__(self, attr: Any) -> Any:
   	return getattr(self.writer, attr)

   such that we can pass through directly to ClearMLLogger.clearml_logger
2. Provide new ClearMLLogger.get_task() method to return ClearMLLogger._task.

I'll mark this topic as "help wanted" and "good first issue" if someone would like to help with it.

[pyb0924]

@vfdev-5 Thanks for your clear answer!
But there is still a slight question: I can directly import Task from clearml package directly, but suppose I want to simultaneously use ClearMLLogger during my training , I have to do like this:

task = Task.init(project_name=project_name, task_name=task_name)
task.connect_configuration(config)
clearml_logger = ClearMLLogger(project_name=project_name, task_name=task_name) 
clearml_logger.attach_output_handler(
    trainer,
    event_name=Events.ITERATION_COMPLETED(every=log_every),
    tag="training",
    output_transform=lambda loss: {"batchloss": loss},
)
...

So, is there any potential conflict produced by this kind of duplicated code?

Thanks!

[pyb0924]

By the way, I've tried in one of my own project but raise no conflict. It's just a supposed question to check the answer in principle. Thanks!

[vfdev-5]

So, is there any potential conflict produced by this kind of duplicated code?

yeah, if task is already created in the user code, then we fetch current task inside ClearMLLogger and use it if it exists. So, basically, there is no conflict.

[pyb0924]

Really appreciate to your answering!