Hi, @KumoLiu
I am very interested in this issue, but I have the following questions for you:

1. What does the problem you mentioned with using logging.conf to document the arguments recorded by _log_input_summary mean? Could you please describe it in more detail?
2. Regarding the issue that output becomes messy and unmanageable, I have a method. I want to use buffer to record the GPU ID. Does this meet your needs?

Thanks.

Hi @hyacinth97223, thanks for your interest here.

1. What does the problem you mentioned with using logging.conf to document the arguments recorded by _log_input_summary mean? Could you please describe it in more detail?

For this one, I mean users can directly log the arguments to a local file using logging.conf.

2. Regarding the issue that output becomes messy and unmanageable, I have a method. I want to use a log to record the GPU ID. Does this meet your needs?

For multi-gpu, there will be a lot of duplicate code that is not very well understood. Such as:

2024-05-27 12:31:45,604 - INFO - --- input summary of monai.bundle.scripts.run ---
2024-05-27 12:31:45,604 - INFO - --- input summary of monai.bundle.scripts.run ---
2024-05-27 12:31:45,604 - INFO - > config_file: ['configs/train.json', 'configs/multi_gpu_train.json']
2024-05-27 12:31:45,604 - INFO - ---
2024-05-27 12:31:45,604 - INFO - > config_file: ['configs/train.json', 'configs/multi_gpu_train.json']
2024-05-27 12:31:45,604 - INFO - ---
monai.bundle.workflows ConfigWorkflow.__init__:workflow_type: Current default value of argument `workflow_type=None` has been deprecated since version 1.2. It will be changed to `workflow_type=train` in version 1.4.
monai.bundle.workflows ConfigWorkflow.__init__:workflow_type: Current default value of argument `workflow_type=None` has been deprecated since version 1.2. It will be changed to `workflow_type=train` in version 1.4.
2024-05-27 12:31:45,604 - INFO - Setting logging properties based on config: configs/logging.conf.
2024-05-27 12:31:45,604 - INFO - Setting logging properties based on config: configs/logging.conf.
Loading dataset: 100%|██████████████████████████████████████████| 32/32 [00:41<00:00,  1.30s/it]
Loading dataset: 100%|██████████████████████████████████████████| 32/32 [00:41<00:00,  1.31s/it]
Loading dataset: 100%|███████████████| 9/9 [00:17<00:00,  1.93s/it]
Loading dataset: 100%|███████████████| 9/9 [00:17<00:00,  1.94s/it]

Hi, @KumoLiu
I have an idea and implemented it as follows. Would this meet your requirements and expectations?

I modified the _log_input_summary() function to include the following enhancements:

Single GPU (rank 0 only): If there is only one GPU, the function will exclusively output all arguments of rank 0 and create a log file to record these arguments.
Multiple GPUs: In a multi-GPU environment, the function will output the content of each thread separately, creating a distinct log file for each thread to record its respective arguments.

I wrote a multi-threaded program to test the functionality of this function, and the output is as shown in the image below.

The generated log files are as shown in the image below.

I would greatly appreciate your feedback on whether this implementation meets your expectations and requirements. Please let me know if you have any further questions or modifications.

Hi @hyacinth97223, thanks for your interest here again!
The expected behavior looks good to me.
For the first one, did you try to modify the logic which make users can directly set the logging path in the "logging.conf" ?
https://docs.python.org/3/library/logging.config.html#logging-config-fileformat
For the second one, I guess we can utilize the RankFilter.

BTW, I didn't see where you push the code?