Final Update

Added argument for prompt 
--prompt="your prompt here"

qwen-batch-single-pass.py

@@ -2,45 +2,50 @@
 import re
 import torch
 import time
+import argparse
 from tqdm import tqdm
 from transformers import AutoModelForCausalLM, AutoTokenizer
-from transformers.generation import GenerationConfig
-import argparse
 
+# Argument Parsing
+parser = argparse.ArgumentParser(description='Image Captioning Script')
+parser.add_argument('--imgdir', type=str, default='path/to/img/dir', help='Directory containing images')
+parser.add_argument('--exist', type=str, default='replace', choices=['skip', 'add', 'replace'], help='Handling of existing captions')
+parser.add_argument('--prompt', type=str, default='describe this image in detail, in less than 35 words', help='Prompt to use for image captioning')
+args = parser.parse_args()
+
+# Function to check for unwanted elements in the caption
 def has_unwanted_elements(caption):
-    patterns = [r'<ref>.*?</ref>', r'<box>.*?</box>']
+    patterns = [r'<ref>.*?</ref>', r'<box>.*?</box>', r'\[\d+\]', r'\(\[\d+\]\)']
     return any(re.search(pattern, caption) for pattern in patterns)
 
+# Function to clean up the caption
 def clean_caption(caption):
     caption = re.sub(r'<ref>(.*?)</ref>', r'\1', caption)
     caption = re.sub(r'<box>.*?</box>', '', caption)
     return caption.strip()
 
-# Argument parsing
-parser = argparse.ArgumentParser(description='Image Captioning Script')
-parser.add_argument('--imgdir', type=str, default='img/dir/here', help='Path to image directory')
-parser.add_argument('--exist', type=str, choices=['skip', 'add', 'replace'], default='replace', help='Handling of existing txt files')
-args = parser.parse_args()
-
-image_directory = args.imgdir
+# Supported image types
 image_types = ['.png', '.jpg', '.jpeg', '.bmp', '.gif']
 
+# Initialize the model and tokenizer
 tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen-VL-Chat-Int4", trust_remote_code=True)
-model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-VL-Chat-Int4", device_map="cuda", trust_remote_code=True, use_flash_attn=True).eval()
+model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-VL-Chat-Int4", device_map="cuda", trust_remote_code=True).eval()
 
-files = [f for f in os.listdir(image_directory) if os.path.splitext(f)[1].lower() in image_types]
+# Get the list of image files in the specified directory
+files = [f for f in os.listdir(args.imgdir) if os.path.splitext(f)[1].lower() in image_types]
 
+# Initialize the progress bar
 pbar = tqdm(total=len(files), desc="Captioning", dynamic_ncols=True, position=0, leave=True)
 start_time = time.time()
 
 print("Captioning phase:")
 for i in range(len(files)):
     filename = files[i]
-    image_path = os.path.join(image_directory, filename)
+    image_path = os.path.join(args.imgdir, filename)
 
-    # Check for existing txt file and handle based on the argument
+    # Handle based on the argument 'exist'
     txt_filename = os.path.splitext(filename)[0] + '.txt'
-    txt_path = os.path.join(image_directory, txt_filename)
+    txt_path = os.path.join(args.imgdir, txt_filename)
 
     if args.exist == 'skip' and os.path.exists(txt_path):
         pbar.update(1)
@@ -49,25 +54,28 @@ def clean_caption(caption):
         with open(txt_path, 'r', encoding='utf-8') as f:
             existing_content = f.read()
 
+    # Generate the caption using the model
     query = tokenizer.from_list_format([
         {'image': image_path},
-        {'text': 'describe this image in detail, as if you are an art critic in less than 35 words'},
+        {'text': args.prompt},
     ])
     response, _ = model.chat(tokenizer, query=query, history=None)
-
+
+    # Clean up the caption if necessary
     if has_unwanted_elements(response):
         response = clean_caption(response)
 
+    # Write the caption to the corresponding .txt file
     with open(txt_path, 'w', encoding='utf-8') as f:
         if args.exist == 'add' and os.path.exists(txt_path):
             f.write(existing_content + "\n" + response)
         else:
             f.write(response)
 
+    # Update progress bar with some additional information about the process
     elapsed_time = time.time() - start_time
     images_per_sec = (i + 1) / elapsed_time
     estimated_time_remaining = (len(files) - i - 1) / images_per_sec
-
     pbar.set_postfix({"Time Elapsed": f"{elapsed_time:.2f}s", "ETA": f"{estimated_time_remaining:.2f}s", "Speed": f"{images_per_sec:.2f} img/s"})
     pbar.update(1)