Enhance image processing workflow in wdv3.py

This commit introduces a new function, `process_image`, to streamline the image processing logic by ensuring the global processor is initialized before processing each image. The batch processing setup is updated to sort and deduplicate image paths, improving efficiency. Additionally, comments are added for clarity, and the number of workers remains set to one to minimize CUDA contention. These changes enhance the maintainability and robustness of the WDV3Processor class.

caption/wdv3.py

@@ -1,4 +1,7 @@
 #!/usr/bin/env python
+# -*- coding: utf-8 -*-
+# 
+# caption/wdv3.py
 
 import sys
 import os
@@ -232,12 +235,19 @@ class WDV3Processor(BatchProcessor[Path, None]):
         if torch.cuda.is_available():
             torch.cuda.empty_cache()
 
+# Function to process a single image, required to be at the top level for pickling
+def process_image(img_path: Path):
+    global global_processor
+    if global_processor is None:
+        raise ValueError("Global processor not initialized. Ensure initializer is set correctly.")
+    global_processor.process_item(img_path)
+
 def main(opts: ScriptOptions):
     target_path = Path(opts.image_file).resolve()
     
     batch_opts = BatchOptions(
         batch_size=16,
-        num_workers=1,  # Single worker
+        num_workers=1,  # Single worker to minimize CUDA contention
         device="cpu" if opts.cpu else "cuda",
         debug=False,
         skip_existing=True,
@@ -253,8 +263,12 @@ def main(opts: ScriptOptions):
             image_paths.extend(target_path.rglob(f'*{ext}'))
             image_paths.extend(target_path.rglob(f'*{ext.upper()}'))
         
+        # Sort and deduplicate image paths
+        sorted_unique_image_paths = sorted(set(image_paths))
+        
+        # Initialize the multiprocessing Pool with the initializer
         with mp.Pool(processes=batch_opts.num_workers, initializer=initializer, initargs=(batch_opts, opts.model)) as pool:
-            pool.map(lambda img: global_processor.process_item(img), sorted(set(image_paths)))
+            pool.map(process_image, sorted_unique_image_paths)
             pool.close()
             pool.join()
     else: