initial commit

LLavaTagger/LLavaTagger.py

@@ -0,0 +1,142 @@
+import warnings
+warnings.simplefilter(action='ignore')
+from transformers import AutoProcessor, LlavaForConditionalGeneration, BitsAndBytesConfig, logging
+import argparse
+import cv2
+import torch
+import os
+import numpy
+from typing import Iterator
+from torch.multiprocessing import Process, Queue
+import json
+from tqdm import tqdm
+
+
+image_ext_ocv = [".bmp", ".jpeg", ".jpg", ".png"]
+
+
+def find_image_files(path: str) -> list[str]:
+	paths = list()
+	for root, dirs, files in os.walk(path):
+		for filename in files:
+			name, extension = os.path.splitext(filename)
+			if extension.lower() in image_ext_ocv:
+				paths.append(os.path.join(root, filename))
+	return paths
+
+
+def image_loader(paths: list[str]) -> Iterator[numpy.ndarray]:
+	for path in paths:
+		name, extension = os.path.splitext(path)
+		extension = extension.lower()
+		imagebgr = cv2.imread(path)
+		image = cv2.cvtColor(imagebgr, cv2.COLOR_BGR2RGB)
+		if image is None:
+			print(f"Warning: could not load {path}")
+		else:
+			yield image, path
+
+
+def pipeline(queue: Queue, image_paths: list[str], prompt: str, device: torch.device, model_name_or_path: str, batch_size: int):
+	model = LlavaForConditionalGeneration.from_pretrained(model_name_or_path, torch_dtype=torch.float16, low_cpu_mem_usage=None,
+		quantization_config=BitsAndBytesConfig(
+			load_in_4bit=True,
+			bnb_4bit_compute_dtype=torch.float16,
+			bnb_4bit_use_double_quant=False,
+			bnb_4bit_quant_type='nf4',
+			), device_map=device, attn_implementation="flash_attention_2")
+	processor = AutoProcessor.from_pretrained(model_name_or_path)
+	image_generator = image_loader(image_paths)
+
+	stop = False
+	finished_count = 0
+	while not stop:
+		prompts = list()
+		images = list()
+		filenames = list()
+		for i in range(0, batch_size):
+			image, filename = next(image_generator, (None, None))
+			if image is None:
+				stop = True
+				break
+
+			filenames.append(filename)
+			images.append(image)
+			prompts.append(prompt)
+
+		if len(images) == 0:
+			break
+
+		inputs = processor(text=prompts, images=images, return_tensors="pt").to(model.device)
+		generate_ids = model.generate(**inputs, max_new_tokens=100, min_new_tokens=3, length_penalty=1.0, do_sample=False, temperature=1.0, top_k=50, top_p=1.0)
+		decodes = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+		finished_count += len(images)
+		for i, decoded in enumerate(decodes):
+			trim = len(prompt) - len("<image>")
+			queue.put({"file_name": filenames[i], "text": decoded[trim:].strip()})
+
+
+def split_list(input_list, count):
+	target_length = int(len(input_list) / count)
+	for i in range(0, count - 1):
+		yield input_list[i * target_length: (i + 1) * target_length]
+	yield input_list[(count - 1) * target_length: len(input_list)]
+
+
+def save_meta(meta_file, meta, reldir, common_description):
+	meta["file_name"] = os.path.relpath(meta["file_name"], reldir)
+	if common_description is not None:
+		meta["text"] = common_description + meta["text"]
+	meta_file.write(json.dumps(meta) + '\n')
+
+
+if __name__ == "__main__":
+	parser = argparse.ArgumentParser("A script to tag images via llava")
+	parser.add_argument('--model', '-m', default="llava-hf/llava-1.5-13b-hf", help="model to use")
+	parser.add_argument('--quantize', '-q', action='store_true', help="load quantized")
+	parser.add_argument('--prompt', '-p', default="Please describe this image in 10 to 20 words.", help="Prompt to use on eatch image")
+	parser.add_argument('--batch', '-b', default=4, type=int, help="Batch size to use for inference")
+	parser.add_argument('--common_description', '-c', help="An optional description that will be preended to the ai generated one")
+	parser.add_argument('--image_dir', '-i', required=True, help="A directory containg the images to tag")
+	args = parser.parse_args()
+
+	prompt = "USER: <image>\n" + args.prompt + "\nASSISTANT: "
+	os.environ["BITSANDBYTES_NOWELCOME"] = "1"
+
+	image_paths = find_image_files(args.image_dir)
+	image_path_chunks = list(split_list(image_paths, torch.cuda.device_count()))
+
+	print(f"Will use {torch.cuda.device_count()} processies to create tags")
+
+	logging.set_verbosity_error()
+	warnings.filterwarnings("ignore")
+	torch.multiprocessing.set_start_method('spawn')
+
+	queue = Queue()
+	processies = list()
+	for i in range(0, torch.cuda.device_count()):
+		processies.append(Process(target=pipeline, args=(queue, image_path_chunks[i], prompt, torch.device(i), args.model, args.batch)))
+		processies[-1].start()
+
+	progress = tqdm(desc="Generateing tags", total=len(image_paths))
+	exit = False
+	with open(os.path.join(args.image_dir, "metadata.jsonl"), mode='w') as output_file:
+		while not exit:
+			if not queue.empty():
+				meta = queue.get()
+				save_meta(output_file, meta, args.image_dir, args.common_description)
+				progress.update()
+			exit = True
+			for process in processies:
+				if process.is_alive():
+					exit = False
+					break
+
+		while not queue.empty():
+			meta = queue.get()
+			save_meta(output_file, meta, args.image_dir, args.common_description)
+			progress.update()
+
+	for process in processies:
+		process.join()
+

LLavaTagger/README.md

@@ -0,0 +1,21 @@
+# LLavaTagger
+
+LLavaTagger is a python script that tags images based on a given prompt using the [LLaVA](https://llava-vl.github.io/) multi modal llm. LLavaTagger supports using any number of gpus in ddp parralel for this task.
+
+## How to use
+
+first create a python venv and install the required packages into it:
+
+	$ python -m venv venv
+	$ source venv/bin/activate
+	$ pip install -r requirements.txt
+
+Then run LLavaTagger for instance like so:
+
+	$ python LLavaTagger.py --common_description "a image of a cat, " --prompt "describe the cat in 10 to 20 words" --batch 8 --quantize --image_dir ~/cat_images
+
+By default LLavaTagger will run in parallel on all available gpus, if this is undesriable please use the ROCR_VISIBLE_DEVICES= or CUDA_VISIBLE_DEVICES= environment variable to hide unwanted gpus
+
+LLavaTagger will then create a meta.jsonl in the image directory sutable to be used by the scripts of [diffusers](https://github.com/huggingface/diffusers) to train stable diffusion (xl) if other formats are desired ../utils contains scripts to transform the metadata into other formats for instace for the use with [kohya](https://github.com/bmaltais/kohya_ss)
+
+If editing the created tags is desired, [QImageTagger](https://uvos.xyz/git/uvos/QImageTagger) can be used for this purpose

LLavaTagger/requirements.txt

@@ -0,0 +1,11 @@
+accelerate==0.29.0
+bitsandbytes
+huggingface-hub==0.22.2
+ninja==1.11.1.1
+safetensors==0.4.2
+tokenizers==0.15.2
+transformers
+torch
+opencv-python
+numpy
+tqdm