removed video inference settings

app.py

@@ -3,7 +3,6 @@ import random
 import uuid
 import json
 import time
-import asyncio
 from threading import Thread
 from typing import Iterable
 
@@ -12,7 +11,6 @@ import spaces
 import torch
 import numpy as np
 from PIL import Image
-import cv2
 
 from transformers import (
     Qwen2VLForConditionalGeneration,
@@ -158,7 +156,6 @@ div.no-padding { padding: 0 !important; }
 """
 
 # Constants for text generation
-MAX_MAX_NEW_TOKENS = 2048
 DEFAULT_MAX_NEW_TOKENS = 1024
 # Increased max_length to accommodate more complex inputs, especially with multiple images
 MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "8192"))
@@ -208,7 +205,7 @@ model_a = AutoModelForImageTextToText.from_pretrained(
 MODEL_ID_W = "allenai/olmOCR-7B-0725"
 processor_w = AutoProcessor.from_pretrained(MODEL_ID_W, trust_remote_code=True)
 model_w = Qwen2_5_VLForConditionalGeneration.from_pretrained(
-    MODEL_ID_W, 
+    MODEL_ID_W,
     trust_remote_code=True,
     torch_dtype=torch.float16
 ).to(device).eval()
@@ -222,35 +219,9 @@ model_m = Qwen2_5_VLForConditionalGeneration.from_pretrained(
     torch_dtype=torch.float16
 ).to(device).eval()
 
-def downsample_video(video_path):
-    """
-    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 = []
-    # Use a maximum of 10 frames to avoid excessive memory usage
-    frame_indices = np.linspace(0, total_frames - 1, min(total_frames, 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)
-            pil_image = Image.fromarray(image)
-            timestamp = round(i / fps, 2)
-            frames.append((pil_image, timestamp))
-    vidcap.release()
-    return frames
 
 @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):
+def generate_image(model_name: str, text: str, image: Image.Image):
     """
     Generates responses using the selected model for image input.
     Yields raw text and Markdown-formatted text.
@@ -286,8 +257,7 @@ def generate_image(model_name: str, text: str, image: Image.Image,
         ]
     }]
     prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
-    
-    # FIX: Set truncation to False to avoid the ValueError
+
     inputs = processor(
         text=[prompt_full],
         images=[image],
@@ -296,7 +266,7 @@ def generate_image(model_name: str, text: str, image: Image.Image,
     ).to(device)
 
     streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
-    generation_kwargs = {**inputs, "streamer": streamer, "max_new_tokens": max_new_tokens}
+    generation_kwargs = {**inputs, "streamer": streamer, "max_new_tokens": DEFAULT_MAX_NEW_TOKENS}
     thread = Thread(target=model.generate, kwargs=generation_kwargs)
     thread.start()
     buffer = ""
@@ -306,138 +276,43 @@ def generate_image(model_name: str, text: str, image: Image.Image,
         time.sleep(0.01)
         yield buffer, buffer
 
-@spaces.GPU
-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 responses using the selected model for video input.
-    Yields raw text and Markdown-formatted text.
-    """
-    if model_name == "RolmOCR-7B":
-        processor = processor_m
-        model = model_m
-    elif model_name == "Qwen2-VL-OCR-2B":
-        processor = processor_x
-        model = model_x
-    elif model_name == "Nanonets-OCR2-3B":
-        processor = processor_v
-        model = model_v
-    elif model_name == "Aya-Vision-8B":
-        processor = processor_a
-        model = model_a
-    elif model_name == "olmOCR-7B-0725":
-        processor = processor_w
-        model = model_w
-    else:
-        yield "Invalid model selected.", "Invalid model selected."
-        return
-
-    if video_path is None:
-        yield "Please upload a video.", "Please upload a video."
-        return
-
-    frames_with_ts = downsample_video(video_path)
-    images_for_processor = [frame for frame, ts in frames_with_ts]
 
-    messages = [{"role": "user", "content": [{"type": "text", "text": text}]}]
-    for frame in images_for_processor:
-        messages[0]["content"].insert(0, {"type": "image"})
-        
-    prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
-    
-    inputs = processor(
-        text=[prompt_full],
-        images=images_for_processor,
-        return_tensors="pt",
-        padding=True
-    ).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, buffer
-
-# Define examples for image and video inference
+# Define examples for image inference
 image_examples = [
-    ["Extract the full page.", "images/ocr.png"],     
-    ["Extract the content.", "images/4.png"],          
+    ["Extract the full page.", "images/ocr.png"],
+    ["Extract the content.", "images/4.png"],
     ["Convert this page to doc [table] precisely for markdown.", "images/0.png"]
 ]
 
-video_examples = [
-    ["Explain the Ad in Detail.", "videos/1.mp4"],
-]
 
 # Create the Gradio Interface
 with gr.Blocks(css=css, theme=thistle_theme) as demo:
     gr.Markdown("# **Multimodal OCR**", elem_id="main-title")
     with gr.Row():
         with gr.Column(scale=2):
-            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="Upload Image", height=290)
-                    image_submit = gr.Button("Submit", variant="primary")
-                    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="Upload Video", height=290)
-                    video_submit = gr.Button("Submit", variant="primary")
-                    gr.Examples(
-                        examples=video_examples,
-                        inputs=[video_query, video_upload]
-                    )
-                    gr.Markdown("> Only the olmOCR and RolmOCR models currently support video inference (max video length: 30 secs).")
-            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)
-                
+            image_query = gr.Textbox(label="Query Input", placeholder="Enter your query here...")
+            image_upload = gr.Image(type="pil", label="Upload Image", height=290)
+            image_submit = gr.Button("Submit", variant="primary")
+            gr.Examples(
+                examples=image_examples,
+                inputs=[image_query, image_upload]
+            )
         with gr.Column(scale=3):
                 gr.Markdown("## Output", elem_id="output-title")
                 output = gr.Textbox(label="Raw Output Stream", interactive=False, lines=11, show_copy_button=True)
                 with gr.Accordion("(Result.md)", open=False):
                     markdown_output = gr.Markdown(label="(Result.Md)")
-                                   
+
                 model_choice = gr.Radio(
                     choices=["Nanonets-OCR2-3B", "olmOCR-7B-0725", "RolmOCR-7B",
                          "Aya-Vision-8B", "Qwen2-VL-OCR-2B"],
                     label="Select Model",
                     value="Nanonets-OCR2-3B"
                 )
-     
+
     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, markdown_output]
-    )
-    video_submit.click(
-        fn=generate_video,
-        inputs=[model_choice, video_query, video_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],
+        inputs=[model_choice, image_query, image_upload],
         outputs=[output, markdown_output]
     )