Update app.py

app.py

@@ -12,14 +12,14 @@ import torch
 import numpy as np
 from PIL import Image
 import cv2
-import edge_tts
 
 from transformers import (
-    AutoModelForCausalLM,
-    AutoTokenizer,
-    TextIteratorStreamer,
     Qwen2VLForConditionalGeneration,
+    Qwen2_5_VLForConditionalGeneration,
+    AutoModelForVision2Seq,
+    AutoModelForImageTextToText,
     AutoProcessor,
+    TextIteratorStreamer,
 )
 from transformers.image_utils import load_image
 
@@ -30,271 +30,251 @@ MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))
 
 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
 
-# Load text-only model and tokenizer
-model_id = "prithivMLmods/Galactic-Qwen-14B-Exp2"
-tokenizer = AutoTokenizer.from_pretrained(model_id)
-model = AutoModelForCausalLM.from_pretrained(
-    model_id,
-    device_map="auto",
-    torch_dtype=torch.bfloat16,
-)
-model.eval()
+# Load Nanonets-OCR-s
+MODEL_ID_M = "nanonets/Nanonets-OCR-s"
+processor_m = AutoProcessor.from_pretrained(MODEL_ID_M, trust_remote_code=True)
+model_m = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+    MODEL_ID_M,
+    trust_remote_code=True,
+    torch_dtype=torch.float16
+).to(device).eval()
 
-# Load multimodal processor and model
-MODEL_ID = "prithivMLmods/Imgscope-OCR-2B-0527"
-processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
-model_m = Qwen2VLForConditionalGeneration.from_pretrained(
-    MODEL_ID,
+# Load SmolDocling-256M-preview
+MODEL_ID_X = "ds4sd/SmolDocling-256M-preview"
+processor_x = AutoProcessor.from_pretrained(MODEL_ID_X, trust_remote_code=True)
+model_x = AutoModelForVision2Seq.from_pretrained(
+    MODEL_ID_X,
     trust_remote_code=True,
     torch_dtype=torch.float16
-).to("cuda").eval()
+).to(device).eval()
 
-# Edge TTS voices mapping for new tags.
-TTS_VOICE_MAP = {
-    "@jennyneural": "en-US-JennyNeural",
-    "@guyneural": "en-US-GuyNeural",
-    "@palomaneural": "es-US-PalomaNeural",
-    "@alonsoneural": "es-US-AlonsoNeural",
-    "@madhurneural": "hi-IN-MadhurNeural"
-}
+#--------------------------------------------------------------------------------------#
+#Load MonkeyOCR
+MODEL_ID_G = "echo840/MonkeyOCR"
+SUBFOLDER = "Recognition"
 
-async def text_to_speech(text: str, voice: str, output_file="output.mp3"):
-    """
-    Convert text to speech using Edge TTS and save as MP3.
-    """
-    communicate = edge_tts.Communicate(text, voice)
-    await communicate.save(output_file)
-    return output_file
+processor_g = AutoProcessor.from_pretrained(
+    MODEL_ID_G,
+    trust_remote_code=True,
+    subfolder=SUBFOLDER
+)
 
-def clean_chat_history(chat_history):
-    """
-    Filter out any chat entries whose "content" is not a string.
-    This helps prevent errors when concatenating previous messages.
-    """
-    cleaned = []
-    for msg in chat_history:
-        if isinstance(msg, dict) and isinstance(msg.get("content"), str):
-            cleaned.append(msg)
-    return cleaned
+model_g = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+    MODEL_ID_G,
+    trust_remote_code=True,
+    subfolder=SUBFOLDER,
+    torch_dtype=torch.float16
+).to(device).eval()
+#--------------------------------------------------------------------------------------#
 
 def downsample_video(video_path):
     """
-    Downsamples the video to 10 evenly spaced frames.
+    Downsamples the video to evenly spaced frames.
     Each frame is returned as a PIL image along with its timestamp.
     """
     vidcap = cv2.VideoCapture(video_path)
     total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
     fps = vidcap.get(cv2.CAP_PROP_FPS)
     frames = []
-    # Sample 10 evenly spaced frames.
     frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)
     for i in frame_indices:
         vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
         success, image = vidcap.read()
         if success:
-            image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)  # Convert BGR to RGB
+            image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
             pil_image = Image.fromarray(image)
             timestamp = round(i / fps, 2)
             frames.append((pil_image, timestamp))
     vidcap.release()
     return frames
 
-def progress_bar_html(label: str) -> str:
+@spaces.GPU
+def generate_image(model_name: str, text: str, image: Image.Image,
+                   max_new_tokens: int = 1024,
+                   temperature: float = 0.6,
+                   top_p: float = 0.9,
+                   top_k: int = 50,
+                   repetition_penalty: float = 1.2):
     """
-    Returns an HTML snippet for a thin progress bar with a label.
-    The progress bar is styled as a light cyan animated bar.
+    Generates responses using the selected model for image input.
     """
-    return f'''
-<div style="display: flex; align-items: center;">
-    <span style="margin-right: 10px; font-size: 14px;">{label}</span>
-    <div style="width: 110px; height: 5px; background-color: #B0E0E6; border-radius: 2px; overflow: hidden;">
-        <div style="width: 100%; height: 100%; background-color: #00FFFF; animation: loading 1.5s linear infinite;"></div>
-    </div>
-</div>
-<style>
-@keyframes loading {{
-    0% {{ transform: translateX(-100%); }}
-    100% {{ transform: translateX(100%); }}
-}}
-</style>
-    '''
+    if model_name == "Nanonets-OCR-s":
+        processor = processor_m
+        model = model_m
+    elif model_name == "SmolDocling-256M-preview":
+        processor = processor_x
+        model = model_x
+    elif model_name == "MonkeyOCR-Recognition":
+        processor = processor_g
+        model = model_g
+    else:
+        yield "Invalid model selected."
+        return
+
+    if image is None:
+        yield "Please upload an image."
+        return
+
+    messages = [{
+        "role": "user",
+        "content": [
+            {"type": "image", "image": image},
+            {"type": "text", "text": text},
+        ]
+    }]
+    prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    inputs = processor(
+        text=[prompt_full],
+        images=[image],
+        return_tensors="pt",
+        padding=True,
+        truncation=False,
+        max_length=MAX_INPUT_TOKEN_LENGTH
+    ).to(device)
+    streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
+    generation_kwargs = {**inputs, "streamer": streamer, "max_new_tokens": max_new_tokens}
+    thread = Thread(target=model.generate, kwargs=generation_kwargs)
+    thread.start()
+    buffer = ""
+    for new_text in streamer:
+        buffer += new_text
+        buffer = buffer.replace("<|im_end|>", "")
+        time.sleep(0.01)
+        yield buffer
 
 @spaces.GPU
-def generate(input_dict: dict, chat_history: list[dict],
-             max_new_tokens: int = 1024,
-             temperature: float = 0.6,
-             top_p: float = 0.9,
-             top_k: int = 50,
-             repetition_penalty: float = 1.2):
+def generate_video(model_name: str, text: str, video_path: str,
+                   max_new_tokens: int = 1024,
+                   temperature: float = 0.6,
+                   top_p: float = 0.9,
+                   top_k: int = 50,
+                   repetition_penalty: float = 1.2):
     """
-    Generates chatbot responses with support for multimodal input, video processing,
-    and Edge TTS when using the new tags @JennyNeural or @GuyNeural.
-    Special command:
-      - "@video-infer": triggers video processing using Imgscope-OCR
+    Generates responses using the selected model for video input.
     """
-    text = input_dict["text"]
-    files = input_dict.get("files", [])
-    lower_text = text.strip().lower()
-
-    # Check for TTS tag in the prompt.
-    tts_voice = None
-    for tag, voice in TTS_VOICE_MAP.items():
-        if lower_text.startswith(tag):
-            tts_voice = voice
-            text = text[len(tag):].strip()  # Remove the tag from the prompt.
-            break
+    if model_name == "Nanonets-OCR-s":
+        processor = processor_m
+        model = model_m
+    elif model_name == "SmolDocling-256M-preview":
+        processor = processor_x
+        model = model_x
+    elif model_name == "MonkeyOCR-Recognition":
+        processor = processor_g
+        model = model_g
+    else:
+        yield "Invalid model selected."
+        return
 
-    # Branch for video processing with Callisto OCR3.
-    if lower_text.startswith("@video-infer"):
-        prompt = text[len("@video-infer"):].strip() if not tts_voice else text
-        if files:
-            # Assume the first file is a video.
-            video_path = files[0]
-            frames = downsample_video(video_path)
-            messages = [
-                {"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]},
-                {"role": "user", "content": [{"type": "text", "text": prompt}]}
-            ]
-            # Append each frame with its timestamp.
-            for frame in frames:
-                image, timestamp = frame
-                image_path = f"video_frame_{uuid.uuid4().hex}.png"
-                image.save(image_path)
-                messages[1]["content"].append({"type": "text", "text": f"Frame {timestamp}:"})
-                messages[1]["content"].append({"type": "image", "url": image_path})
-        else:
-            messages = [
-                {"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]},
-                {"role": "user", "content": [{"type": "text", "text": prompt}]}
-            ]
-        # Enable truncation to avoid token/feature mismatch.
-        inputs = processor.apply_chat_template(
-            messages,
-            tokenize=True,
-            add_generation_prompt=True,
-            return_dict=True,
-            return_tensors="pt",
-            truncation=True,
-            max_length=MAX_INPUT_TOKEN_LENGTH
-        ).to("cuda")
-        streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
-        generation_kwargs = {
-            **inputs,
-            "streamer": streamer,
-            "max_new_tokens": max_new_tokens,
-            "do_sample": True,
-            "temperature": temperature,
-            "top_p": top_p,
-            "top_k": top_k,
-            "repetition_penalty": repetition_penalty,
-        }
-        thread = Thread(target=model_m.generate, kwargs=generation_kwargs)
-        thread.start()
-        buffer = ""
-        yield progress_bar_html("Processing video with Imgscope-OCR")
-        for new_text in streamer:
-            buffer += new_text
-            buffer = buffer.replace("<|im_end|>", "")
-            time.sleep(0.01)
-            yield buffer
+    if video_path is None:
+        yield "Please upload a video."
         return
 
-    # Multimodal processing when files are provided.
-    if files:
-        if len(files) > 1:
-            images = [load_image(image) for image in files]
-        elif len(files) == 1:
-            images = [load_image(files[0])]
-        else:
-            images = []
-        messages = [{
-            "role": "user",
-            "content": [
-                *[{"type": "image", "image": image} for image in images],
-                {"type": "text", "text": text},
-            ]
-        }]
-        prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
-        # Enable truncation explicitly here as well.
-        inputs = processor(
-            text=[prompt_full],
-            images=images,
-            return_tensors="pt",
-            padding=True,
-            truncation=True,
-            max_length=MAX_INPUT_TOKEN_LENGTH
-        ).to("cuda")
-        streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
-        generation_kwargs = {**inputs, "streamer": streamer, "max_new_tokens": max_new_tokens}
-        thread = Thread(target=model_m.generate, kwargs=generation_kwargs)
-        thread.start()
-        buffer = ""
-        yield progress_bar_html("Processing image with Imgscope-OCR")
-        for new_text in streamer:
-            buffer += new_text
-            buffer = buffer.replace("<|im_end|>", "")
-            time.sleep(0.01)
-            yield buffer
-    else:
-        # Normal text conversation processing with Pocket Llama.
-        conversation = clean_chat_history(chat_history)
-        conversation.append({"role": "user", "content": text})
-        input_ids = tokenizer.apply_chat_template(conversation, add_generation_prompt=True, return_tensors="pt")
-        if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
-            input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
-            gr.Warning(f"Trimmed input from conversation as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.")
-        input_ids = input_ids.to(model.device)
-        streamer = TextIteratorStreamer(tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True)
-        generation_kwargs = {
-            "input_ids": input_ids,
-            "streamer": streamer,
-            "max_new_tokens": max_new_tokens,
-            "do_sample": True,
-            "top_p": top_p,
-            "top_k": top_k,
-            "temperature": temperature,
-            "num_beams": 1,
-            "repetition_penalty": repetition_penalty,
-        }
-        t = Thread(target=model.generate, kwargs=generation_kwargs)
-        t.start()
-        outputs = []
-        yield progress_bar_html("Processing With Galactic Qwen")
-        for new_text in streamer:
-            outputs.append(new_text)
-            yield "".join(outputs)
-        final_response = "".join(outputs)
-        yield final_response
+    frames = downsample_video(video_path)
+    messages = [
+        {"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]},
+        {"role": "user", "content": [{"type": "text", "text": text}]}
+    ]
+    for frame in frames:
+        image, timestamp = frame
+        messages[1]["content"].append({"type": "text", "text": f"Frame {timestamp}:"})
+        messages[1]["content"].append({"type": "image", "image": image})
+    inputs = processor.apply_chat_template(
+        messages,
+        tokenize=True,
+        add_generation_prompt=True,
+        return_dict=True,
+        return_tensors="pt",
+        truncation=False,
+        max_length=MAX_INPUT_TOKEN_LENGTH
+    ).to(device)
+    streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
+    generation_kwargs = {
+        **inputs,
+        "streamer": streamer,
+        "max_new_tokens": max_new_tokens,
+        "do_sample": True,
+        "temperature": temperature,
+        "top_p": top_p,
+        "top_k": top_k,
+        "repetition_penalty": repetition_penalty,
+    }
+    thread = Thread(target=model.generate, kwargs=generation_kwargs)
+    thread.start()
+    buffer = ""
+    for new_text in streamer:
+        buffer += new_text
+        buffer = buffer.replace("<|im_end|>", "")
+        time.sleep(0.01)
+        yield buffer
 
-        # If a TTS voice was specified, convert the final response to speech.
-        if tts_voice:
-            output_file = asyncio.run(text_to_speech(final_response, tts_voice))
-            yield gr.Audio(output_file, autoplay=True)
+# Define examples for image and video inference
+image_examples = [
+    ["fill the correct numbers", "example/image3.png"],
+    ["ocr the image", "example/image1.png"],
+    ["explain the scene", "example/image2.jpg"],
+]
 
-# Create the Gradio ChatInterface with the custom CSS applied
-demo = gr.ChatInterface(
-    fn=generate,
-    additional_inputs=[
-        gr.Slider(label="Max new tokens", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS),
-        gr.Slider(label="Temperature", minimum=0.1, maximum=4.0, step=0.1, value=0.6),
-        gr.Slider(label="Top-p (nucleus sampling)", minimum=0.05, maximum=1.0, step=0.05, value=0.9),
-        gr.Slider(label="Top-k", minimum=1, maximum=1000, step=1, value=50),
-        gr.Slider(label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05, value=1.2),
-    ],
-    examples=[
-        ["Write the code that converts temperatures between Celsius and Fahrenheit in short"],
-        [{"text": "Create a short story based on the image.", "files": ["examples/1.jpg"]}],
-        ["@JennyNeural Who was Nikola Tesla and what were his contributions?"],
-        [{"text": "@video-infer Describe the video", "files": ["examples/Missing.mp4"]}]
-    ],
-    cache_examples=False,
-    description="# **Imgscope-OCR-Mini**",
-    type="messages",
-    fill_height=True,
-    textbox=gr.MultimodalTextbox(label="Query Input", file_types=["image", "video"], file_count="multiple"),
-    stop_btn="Stop Generation",
-    multimodal=True,
-)
+video_examples = [
+    ["Explain the ad in detail", "example/1.mp4"],
+    ["Identify the main actions in the coca cola ad...", "example/2.mp4"]
+]
+
+css = """
+.submit-btn {
+    background-color: #2980b9 !important;
+    color: white !important;
+}
+.submit-btn:hover {
+    background-color: #3498db !important;
+}
+"""
+
+# Create the Gradio Interface
+with gr.Blocks(css=css, theme="bethecloud/storj_theme") as demo:
+    gr.Markdown("# **[core OCR](https://huggingface.co/collections/prithivMLmods/core-and-docscope-ocr-models-6816d7f1bde3f911c6c852bc)**")
+    with gr.Row():
+        with gr.Column():
+            with gr.Tabs():
+                with gr.TabItem("Image Inference"):
+                    image_query = gr.Textbox(label="Query Input", placeholder="Enter your query here...")
+                    image_upload = gr.Image(type="pil", label="Image")
+                    image_submit = gr.Button("Submit", elem_classes="submit-btn")
+                    gr.Examples(
+                        examples=image_examples,
+                        inputs=[image_query, image_upload]
+                    )
+                with gr.TabItem("Video Inference"):
+                    video_query = gr.Textbox(label="Query Input", placeholder="Enter your query here...")
+                    video_upload = gr.Video(label="Video")
+                    video_submit = gr.Button("Submit", elem_classes="submit-btn")
+                    gr.Examples(
+                        examples=video_examples,
+                        inputs=[video_query, video_upload]
+                    )
+            with gr.Accordion("Advanced options", open=False):
+                max_new_tokens = gr.Slider(label="Max new tokens", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)
+                temperature = gr.Slider(label="Temperature", minimum=0.1, maximum=4.0, step=0.1, value=0.6)
+                top_p = gr.Slider(label="Top-p (nucleus sampling)", minimum=0.05, maximum=1.0, step=0.05, value=0.9)
+                top_k = gr.Slider(label="Top-k", minimum=1, maximum=1000, step=1, value=50)
+                repetition_penalty = gr.Slider(label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05, value=1.2)
+        with gr.Column():
+            output = gr.Textbox(label="Output", interactive=False, lines=3, scale=2)
+            model_choice = gr.Radio(
+                choices=["Nanonets-OCR-s", "MonkeyOCR-Recognition", "SmolDocling-256M-preview"],
+                label="Select Model",
+                value="Nanonets-OCR-s"
+            )
+            
+    image_submit.click(
+        fn=generate_image,
+        inputs=[model_choice, image_query, image_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],
+        outputs=output
+    )
+    video_submit.click(
+        fn=generate_video,
+        inputs=[model_choice, video_query, video_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],
+        outputs=output
+    )
 
 if __name__ == "__main__":
-    demo.queue(max_size=20).launch(share=True)
+    demo.queue(max_size=30).launch(share=True, mcp_server=True, ssr_mode=False, show_error=True)