utils

Small I/O and bookkeeping helpers: atomic JSON writes, boolean-string parsing, and the task-list / run-status tracking that makes long benchmark sweeps resumable.

medvision_bm.utils.utils.atomic_write_json(json_path, data, indent=4)

Write data as JSON to json_path atomically.

Writes to a temp file in the same directory, fsyncs it, then os.replace-es it over the target. If the write fails (e.g. ENOSPC on a full disk), the original file is left untouched instead of being truncated to 0 bytes, which is what opening the target directly with mode "w" would do.

Parameters:

• json_path (str) – Destination path for the JSON file.

• data – Any JSON-serializable object to write.

• indent (int) – Indentation passed to json.dump. Defaults to 4.

Raises:

BaseException – Re-raises any error raised while writing or replacing the file, after removing the partial temp file.

medvision_bm.utils.utils.str2bool(v)

Coerce a string (or bool) to a boolean, for use as an argparse type.

Parameters:

v – A boolean, or a string such as "yes", "true", "1", "no", "false", or "0" (case-insensitive).

Returns:

True for truthy tokens and False for falsy tokens. A value that is already a bool is returned unchanged.

Return type:

bool

Raises:

argparse.ArgumentTypeError – If v is not a recognized boolean token.

medvision_bm.utils.utils.set_cuda_num_processes()

Determine the number of GPU processes to launch from the CUDA environment.

Reads CUDA_VISIBLE_DEVICES. When it is unset, all GPUs reported by torch.cuda.device_count() are used; otherwise the number of device ids listed in the variable is used (at least 1).

Returns:

The number of processes to run, one per visible GPU.

Return type:

int

medvision_bm.utils.utils.update_task_status(json_path, model_name, task_name)

Mark a (model, task) pair as completed in a JSON tracking file.

Loads the existing status file (creating the parent directory and treating a missing file as empty), sets data[model_name][task_name] to True, and writes the file back atomically.

Parameters:

• json_path (str) – Path to the JSON tracking file.

• model_name (str) – Model whose status is updated.

• task_name (str) – Task to mark as completed.

Returns:

Always False; the return value is unused by callers.

Return type:

bool

medvision_bm.utils.utils.load_tasks(json_file_path)

Load task names from a JSON file mapping task names to their definitions.

Parameters:

json_file_path (str) – Path to a JSON file whose top-level keys are task names.

Returns:

The task names (the top-level keys), in file order.

Return type:

list

medvision_bm.utils.utils.load_tasks_status(tasks_status_file, model_name)

Return the completion-status mapping for a single model.

Parameters:

• tasks_status_file (str) – Path to the JSON status file. A missing file is treated as empty.

• model_name (str) – Model whose status entry is returned.

Returns:

Mapping of task name to completion flag for model_name, or an empty dict if the file or the model entry is absent.

Return type:

dict

Raises:

ValueError – If the status file exists but cannot be read or parsed.