Update app.py

app.py

@@ -5,6 +5,7 @@ import json
 import time
 import asyncio
 from threading import Thread
+import tempfile
 
 import gradio as gr
 import spaces
@@ -47,16 +48,6 @@ model_x = Qwen2_5_VLForConditionalGeneration.from_pretrained(
     torch_dtype=torch.float16
 ).to(device).eval()
 
-# Load typhoon-ocr-3b
-MODEL_ID_T = "scb10x/typhoon-ocr-3b"
-processor_t = AutoProcessor.from_pretrained(MODEL_ID_T, trust_remote_code=True)
-model_t = Qwen2_5_VLForConditionalGeneration.from_pretrained(
-    MODEL_ID_T,
-    trust_remote_code=True,
-    torch_dtype=torch.float16
-).to(device).eval()
-
-
 # Load Gemma3n-E4B-it
 MODEL_ID_G = "google/gemma-3n-E4B-it"
 processor_g = AutoProcessor.from_pretrained(MODEL_ID_G, trust_remote_code=True)
@@ -66,29 +57,40 @@ model_g = AutoModelForImageTextToText.from_pretrained(
     torch_dtype=torch.float16
 ).to(device).eval()
 
+# Load Gemma3n-E2B-it
+MODEL_ID_N = "google/gemma-3n-E2B-it"
+processor_n = AutoProcessor.from_pretrained(MODEL_ID_N, trust_remote_code=True)
+model_n = AutoModelForImageTextToText.from_pretrained(
+    MODEL_ID_N,
+    trust_remote_code=True,
+    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.
+    Downsamples the video to evenly spaced frames and saves them to temporary files.
+    Returns a list of (frame_path, timestamp) and the temp directory.
     """
     vidcap = cv2.VideoCapture(video_path)
     total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
     fps = vidcap.get(cv2.CAP_PROP_FPS)
-    frames = []
     frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)
+    temp_dir = tempfile.mkdtemp()
+    frames = []
     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)
+            frame_path = os.path.join(temp_dir, f"frame_{i}.jpg")
+            Image.fromarray(image).save(frame_path)
             timestamp = round(i / fps, 2)
-            frames.append((pil_image, timestamp))
+            frames.append((frame_path, timestamp))
     vidcap.release()
-    return frames
+    return frames, temp_dir
 
 @spaces.GPU
-def generate_image(model_name: str, text: str, image: Image.Image,
+def generate_image(model_name: str, text: str, image_path: str,
                    max_new_tokens: int = 1024,
                    temperature: float = 0.6,
                    top_p: float = 0.9,
@@ -103,30 +105,43 @@ def generate_image(model_name: str, text: str, image: Image.Image,
     elif model_name == "DREX-062225-7B-exp":
         processor = processor_x
         model = model_x
-    elif model_name == "Typhoon-OCR-3B":
-        processor = processor_t
-        model = model_t
     elif model_name == "Gemma3n-E4B-it":
         processor = processor_g
         model = model_g
+    elif model_name == "Gemma3n-E2B-it":
+        processor = processor_n
+        model = model_n
     else:
         yield "Invalid model selected.", "Invalid model selected."
         return
 
-    if image is None:
+    if image_path is None:
         yield "Please upload an image.", "Please upload an image."
         return
 
-    messages = [{"role": "user", "content": [{"type": "image", "image": image}, {"type": "text", "text": text}]}]
-    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)
+    messages = [{"role": "user", "content": [{"type": "text", "text": text}, {"type": "image", "image": image_path}]}]
+
+    if model_name in ["Gemma3n-E4B-it", "Gemma3n-E2B-it"]:
+        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)
+    else:
+        prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+        inputs = processor(
+            text=[prompt_full],
+            images=[image_path],
+            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,
@@ -162,12 +177,12 @@ def generate_video(model_name: str, text: str, video_path: str,
     elif model_name == "DREX-062225-7B-exp":
         processor = processor_x
         model = model_x
-    elif model_name == "Typhoon-OCR-3B":
-        processor = processor_t
-        model = model_t
     elif model_name == "Gemma3n-E4B-it":
         processor = processor_g
         model = model_g
+    elif model_name == "Gemma3n-E2B-it":
+        processor = processor_n
+        model = model_n
     else:
         yield "Invalid model selected.", "Invalid model selected."
         return
@@ -176,26 +191,35 @@ def generate_video(model_name: str, text: str, video_path: str,
         yield "Please upload a video.", "Please upload a video."
         return
 
-    frames = downsample_video(video_path)
+    frames, temp_dir = downsample_video(video_path)
     content = [{"type": "text", "text": text}]
-    if model_name == "Gemma3n-E4B-it":
-        for frame, _ in frames:
-            content.append({"type": "image", "image": frame})
-    else:
-        for frame in frames:
-            image, timestamp = frame
-            content.append({"type": "text", "text": f"Frame {timestamp}:"})
-            content.append({"type": "image", "image": image})
+    for frame_path, timestamp in frames:
+        content.append({"type": "text", "text": f"Frame {timestamp}:"})
+        content.append({"type": "image", "image": frame_path})
     messages = [{"role": "user", "content": content}]
-    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)
+
+    if model_name in ["Gemma3n-E4B-it", "Gemma3n-E2B-it"]:
+        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)
+    else:
+        prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+        images = [frame_path for frame_path, _ in frames]
+        inputs = processor(
+            text=[prompt_full],
+            images=images,
+            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,
@@ -253,7 +277,7 @@ with gr.Blocks(css=css, theme=gr.themes.Citrus()) as demo:
             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_upload = gr.Image(type="filepath", label="Image")
                     image_submit = gr.Button("Submit", elem_classes="submit-btn")
                     gr.Examples(
                         examples=image_examples,
@@ -281,18 +305,11 @@ with gr.Blocks(css=css, theme=gr.themes.Citrus()) as demo:
                 markdown_output = gr.Markdown(label="Formatted Result (Result.Md)")
 
             model_choice = gr.Radio(
-                choices=["DREX-062225-7B-exp", "VIREX-062225-7B-exp", "Typhoon-OCR-3B", "Gemma3n-E4B-it"],
+                choices=["DREX-062225-7B-exp", "VIREX-062225-7B-exp", "Gemma3n-E4B-it", "Gemma3n-E2B-it"],
                 label="Select Model",
                 value="DREX-062225-7B-exp"
             )
 
-            gr.Markdown("**Model Info 💻** | [Report Bug](https://huggingface.co/spaces/prithivMLmods/Doc-VLMs/discussions)")          
-            gr.Markdown("> [DREX-062225-7B-exp](https://huggingface.co/prithivMLmods/DREX-062225-exp): the drex-062225-exp (document retrieval and extraction expert) model is a specialized fine-tuned version of docscopeocr-7b-050425-exp, optimized for document retrieval, content extraction, and analysis recognition. built on top of the qwen2.5-vl architecture.")
-            gr.Markdown("> [VIREX-062225-7B-exp](https://huggingface.co/prithivMLmods/VIREX-062225-exp): the virex-062225-exp (video information retrieval and extraction expert - experimental) model is a fine-tuned version of qwen2.5-vl-7b-instruct, specifically optimized for advanced video understanding, image comprehension, sense of reasoning, and natural language decision-making through cot reasoning.")
-            gr.Markdown("> [Typhoon-OCR-3B](https://huggingface.co/scb10x/typhoon-ocr-3b): a bilingual document parsing model built specifically for real-world documents in thai and english, inspired by models like olmocr, based on qwen2.5-vl-instruction. this model is intended to be used with a specific prompt only.")
-            gr.Markdown("> [Gemma3n-E4B-it](https://huggingface.co/google/gemma-3n-E4B-it): A multimodal model capable of processing images and videos for various tasks.")
-            gr.Markdown(">⚠️note: all the models in space are not guaranteed to perform well in video inference use cases.")  
-    
     image_submit.click(
         fn=generate_image,
         inputs=[model_choice, image_query, image_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],