hjxdl 0.2.15__py3-none-any.whl → 0.2.17__py3-none-any.whl

hdl/_version.py

@@ -12,5 +12,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.2.15'
-__version_tuple__ = version_tuple = (0, 2, 15)
+__version__ = version = '0.2.17'
+__version_tuple__ = version_tuple = (0, 2, 17)

hdl/utils/llm/ocrrag.py

@@ -0,0 +1,224 @@
+import argparse
+from PIL import Image
+import hashlib
+import torch
+import fitz
+import gradio as gr
+import os
+import numpy as np
+import json
+from transformers import AutoModel, AutoTokenizer
+
+from .chat import OpenAI_M
+from .vis import pilimg_to_base64
+
+def get_image_md5(img: Image.Image):
+    img_byte_array = img.tobytes()
+    hash_md5 = hashlib.md5()
+    hash_md5.update(img_byte_array)
+    hex_digest = hash_md5.hexdigest()
+    return hex_digest
+
+def calculate_md5_from_binary(binary_data):
+    hash_md5 = hashlib.md5()
+    hash_md5.update(binary_data)
+    return hash_md5.hexdigest()
+
+def add_pdf_gradio(pdf_file_binary, progress=gr.Progress(), cache_dir=None, model=None, tokenizer=None):
+    model.eval()
+
+    knowledge_base_name = calculate_md5_from_binary(pdf_file_binary)
+
+    this_cache_dir = os.path.join(cache_dir, knowledge_base_name)
+    os.makedirs(this_cache_dir, exist_ok=True)
+
+    with open(os.path.join(this_cache_dir, f"src.pdf"), 'wb') as file:
+        file.write(pdf_file_binary)
+
+    dpi = 200
+    doc = fitz.open("pdf", pdf_file_binary)
+
+    reps_list = []
+    images = []
+    image_md5s = []
+
+    for page in progress.tqdm(doc):
+        pix = page.get_pixmap(dpi=dpi)
+        image = Image.frombytes("RGB", [pix.width, pix.height], pix.samples)
+        image_md5 = get_image_md5(image)
+        image_md5s.append(image_md5)
+        with torch.no_grad():
+            reps = model(text=[''], image=[image], tokenizer=tokenizer).reps
+        reps_list.append(reps.squeeze(0).cpu().numpy())
+        images.append(image)
+
+    for idx in range(len(images)):
+        image = images[idx]
+        image_md5 = image_md5s[idx]
+        cache_image_path = os.path.join(this_cache_dir, f"{image_md5}.png")
+        image.save(cache_image_path)
+
+    np.save(os.path.join(this_cache_dir, f"reps.npy"), reps_list)
+
+    with open(os.path.join(this_cache_dir, f"md5s.txt"), 'w') as f:
+        for item in image_md5s:
+            f.write(item+'\n')
+
+    return knowledge_base_name
+
+def retrieve_gradio(knowledge_base, query, topk, cache_dir=None, model=None, tokenizer=None):
+    model.eval()
+
+    target_cache_dir = os.path.join(cache_dir, knowledge_base)
+
+    if not os.path.exists(target_cache_dir):
+        return None
+
+    md5s = []
+    with open(os.path.join(target_cache_dir, f"md5s.txt"), 'r') as f:
+        for line in f:
+            md5s.append(line.rstrip('\n'))
+
+    doc_reps = np.load(os.path.join(target_cache_dir, f"reps.npy"))
+
+    query_with_instruction = "Represent this query for retrieving relevant document: " + query
+    with torch.no_grad():
+        query_rep = model(text=[query_with_instruction], image=[None], tokenizer=tokenizer).reps.squeeze(0).cpu()
+
+    query_md5 = hashlib.md5(query.encode()).hexdigest()
+
+    doc_reps_cat = torch.stack([torch.Tensor(i) for i in doc_reps], dim=0)
+
+    similarities = torch.matmul(query_rep, doc_reps_cat.T)
+
+    topk_values, topk_doc_ids = torch.topk(similarities, k=topk)
+
+    images_topk = [Image.open(os.path.join(target_cache_dir, f"{md5s[idx]}.png")) for idx in topk_doc_ids.cpu().numpy()]
+
+    with open(os.path.join(target_cache_dir, f"q-{query_md5}.json"), 'w') as f:
+        f.write(json.dumps(
+            {
+                "knowledge_base": knowledge_base,
+                "query": query,
+                "retrieved_docs": [os.path.join(target_cache_dir, f"{md5s[idx]}.png") for idx in topk_doc_ids.cpu().numpy()]
+            }, indent=4, ensure_ascii=False
+        ))
+
+    return images_topk
+
+def answer_question_stream(images, question, gen_model):
+    # Load images from the image paths in images[0]
+    pil_images = [Image.open(image[0]).convert('RGB') for image in images]
+
+    # Calculate the total size of the new image (for vertical concatenation)
+    widths, heights = zip(*(img.size for img in pil_images))
+
+    # Assuming vertical concatenation, so width is the max width, height is the sum of heights
+    total_width = max(widths)
+    total_height = sum(heights)
+
+    # Create a new blank image with the total width and height
+    new_image = Image.new('RGB', (total_width, total_height))
+
+    # Paste each image into the new image
+    y_offset = 0
+    for img in pil_images:
+        new_image.paste(img, (0, y_offset))
+        y_offset += img.height  # Move the offset down by the height of the image
+
+    # Convert the concatenated image to base64
+    new_image_base64 = pilimg_to_base64(new_image)
+
+    # Call the model with the base64-encoded concatenated image and stream=True
+    for partial_answer in gen_model.chat(
+        prompt=question,
+        images=[new_image_base64],  # Use the concatenated image
+        stream=True  # Enable streaming
+    ):
+        # Yield the partial answer as it comes in
+        yield partial_answer  # Stream the output to Gradio
+
+def upvote(knowledge_base, query, cache_dir):
+    target_cache_dir = os.path.join(cache_dir, knowledge_base)
+    query_md5 = hashlib.md5(query.encode()).hexdigest()
+
+    with open(os.path.join(target_cache_dir, f"q-{query_md5}.json"), 'r') as f:
+        data = json.loads(f.read())
+
+    data["user_preference"] = "upvote"
+
+    with open(os.path.join(target_cache_dir, f"q-{query_md5}-withpref.json"), 'w') as f:
+        f.write(json.dumps(data, indent=4, ensure_ascii=False))
+
+def downvote(knowledge_base, query, cache_dir):
+    target_cache_dir = os.path.join(cache_dir, knowledge_base)
+    query_md5 = hashlib.md5(query.encode()).hexdigest()
+
+    with open(os.path.join(target_cache_dir, f"q-{query_md5}.json"), 'r') as f:
+        data = json.loads(f.read())
+
+    data["user_preference"] = "downvote"
+
+    with open(os.path.join(target_cache_dir, f"q-{query_md5}-withpref.json"), 'w') as f:
+        f.write(json.dumps(data, indent=4, ensure_ascii=False))
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description="MiniCPMV-RAG-PDFQA Script")
+    parser.add_argument('--cache-dir', dest='cache_dir', type=str, required=True, help='Cache directory path')
+    parser.add_argument('--device', dest='device', type=str, default='cuda:0', help='Device for model inference')
+    parser.add_argument('--model-path', dest='model_path', type=str, required=True, help='Path to the embedding model')
+    parser.add_argument('--llm-host', dest='llm_host', type=str, default='127.0.0.1', help='LLM server IP address')
+    parser.add_argument('--llm-port', dest='llm_port', type=int, default=22299, help='LLM server port')
+    parser.add_argument('--server-name', dest='server_name', type=str, default='0.0.0.0', help='Gradio server name')
+    parser.add_argument('--server-port', dest='server_port', type=int, default=10077, help='Gradio server port')
+
+    args = parser.parse_args()
+
+    print("Loading embedding model...")
+    tokenizer = AutoTokenizer.from_pretrained(args.model_path, trust_remote_code=True)
+    model = AutoModel.from_pretrained(args.model_path, trust_remote_code=True)
+    model.to(args.device)
+    model.eval()
+    print("Embedding model loaded!")
+
+    gen_model = OpenAI_M(
+        server_ip=args.llm_host,
+        server_port=args.llm_port
+    )
+
+    with gr.Blocks() as app:
+        gr.Markdown("# MiniCPMV-RAG-PDFQA: Two Vision Language Models Enable End-to-End RAG")
+
+        file_input = gr.File(type="binary", label="Step 1: Upload PDF")
+        file_result = gr.Text(label="Knowledge Base ID")
+        process_button = gr.Button("Process PDF")
+
+        process_button.click(lambda pdf: add_pdf_gradio(pdf, cache_dir=args.cache_dir, model=model, tokenizer=tokenizer),
+                             inputs=file_input, outputs=file_result)
+
+        kb_id_input = gr.Text(label="Knowledge Base ID")
+        query_input = gr.Text(label="Your Question")
+        topk_input = gr.Number(value=5, minimum=1, maximum=10, step=1, label="Number of pages to retrieve")
+        retrieve_button = gr.Button("Retrieve Pages")
+        images_output = gr.Gallery(label="Retrieved Pages")
+
+        retrieve_button.click(lambda kb, query, topk: retrieve_gradio(kb, query, topk, cache_dir=args.cache_dir, model=model, tokenizer=tokenizer),
+                              inputs=[kb_id_input, query_input, topk_input], outputs=images_output)
+
+        button = gr.Button("Answer Question")
+        gen_model_response = gr.Textbox(label="MiniCPM-V-2.6's Answer", lines=10)
+
+        # Use answer_question_stream for streaming response and pass gen_model
+        button.click(lambda images, query: answer_question_stream(images, query, gen_model),
+                     inputs=[images_output, query_input],
+                     outputs=gen_model_response)
+
+        upvote_button = gr.Button("🤗 Upvote")
+        downvote_button = gr.Button("🤣 Downvote")
+
+        upvote_button.click(lambda kb, query: upvote(kb, query, cache_dir=args.cache_dir),
+                            inputs=[kb_id_input, query_input], outputs=None)
+        downvote_button.click(lambda kb, query: downvote(kb, query, cache_dir=args.cache_dir),
+                              inputs=[kb_id_input, query_input], outputs=None)
+
+    app.launch(server_name=args.server_name, server_port=args.server_port)

{hjxdl-0.2.15.dist-info → hjxdl-0.2.17.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: hjxdl
-Version: 0.2.15
+Version: 0.2.17
 Summary: A collection of functions for Jupyter notebooks
 Home-page: https://github.com/huluxiaohuowa/hdl
 Author: Jianxing Hu

{hjxdl-0.2.15.dist-info → hjxdl-0.2.17.dist-info}/RECORD

@@ -1,5 +1,5 @@
 hdl/__init__.py,sha256=GffnD0jLJdhkd-vo989v40N90sQbofkayRBwxc6TVhQ,72
-hdl/_version.py,sha256=jLEv5wW_x8zLfp1bwdtLZaGYh6V6kSSSoRN6ETWwP6c,413
+hdl/_version.py,sha256=VhyMkib4e74TxEaAAAUCZ36wnl5wS9k5JtYxBa3f_mw,413
 hdl/args/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 hdl/args/loss_args.py,sha256=s7YzSdd7IjD24rZvvOrxLLFqMZQb9YylxKeyelSdrTk,70
 hdl/controllers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -132,13 +132,14 @@ hdl/utils/llm/chatgr.py,sha256=GO2G7g6YybduA5VCUuGjvEsJfC_6L7rycSnPeHMcxyM,2820
 hdl/utils/llm/embs.py,sha256=Tf0FOYrOFZp7qQpEPiSCXzlgyHH0X9HVTUtsup74a9E,7174
 hdl/utils/llm/extract.py,sha256=2sK_WJzmYIc8iuWaM9DA6Nw3_6q1O4lJ5pKpcZo-bBA,6512
 hdl/utils/llm/llama_chat.py,sha256=watcHGOaz-bv3x-yDucYlGk5f8FiqfFhwWogrl334fk,4387
+hdl/utils/llm/ocrrag.py,sha256=Quv3wUhkCodGeJRwTGrH9cYbkAvV0u7JY3pwcespyss,9049
 hdl/utils/llm/vis.py,sha256=-6QvxSVzKqxLh_l0aYg2wN2G5HOiQvCpfp-jn9twXw0,16210
 hdl/utils/llm/visrag.py,sha256=vNj4cHsvfC_Vc0eDPKZc-yflLUMGApZGpggjAqAlwS8,9215
 hdl/utils/schedulers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 hdl/utils/schedulers/norm_lr.py,sha256=bDwCmdEK-WkgxQMFBiMuchv8Mm7C0-GZJ6usm-PQk14,4461
 hdl/utils/weather/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 hdl/utils/weather/weather.py,sha256=k11o6wM15kF8b9NMlEfrg68ak-SfSYLN3nOOflFUv-I,4381
-hjxdl-0.2.15.dist-info/METADATA,sha256=8TpagjCOPvQJbvgS4FPGB6gqHRyvScegPcxhL7ZIwlM,836
-hjxdl-0.2.15.dist-info/WHEEL,sha256=OVMc5UfuAQiSplgO0_WdW7vXVGAt9Hdd6qtN4HotdyA,91
-hjxdl-0.2.15.dist-info/top_level.txt,sha256=-kxwTM5JPhylp06z3zAVO3w6_h7wtBfBo2zgM6YZoTk,4
-hjxdl-0.2.15.dist-info/RECORD,,
+hjxdl-0.2.17.dist-info/METADATA,sha256=ZuE3sjFH07ZsR1rrIPjyLUW4ZbabwMe5eCTIxag2rdY,836
+hjxdl-0.2.17.dist-info/WHEEL,sha256=OVMc5UfuAQiSplgO0_WdW7vXVGAt9Hdd6qtN4HotdyA,91
+hjxdl-0.2.17.dist-info/top_level.txt,sha256=-kxwTM5JPhylp06z3zAVO3w6_h7wtBfBo2zgM6YZoTk,4
+hjxdl-0.2.17.dist-info/RECORD,,