Update app.py

app.py

@@ -1,224 +1,228 @@
+import gradio as gr
 import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from PIL import Image 
-import numpy as np 
-import os 
-import tempfile
-import spaces
+from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer, StoppingCriteria
 import gradio as gr
-
+import spaces
+import torch
+import numpy as np
+import torch
+import torchvision.transforms as T
+from PIL import Image
+from torchvision.transforms.functional import InterpolationMode
+from transformers import AutoModel, AutoTokenizer
+
+from threading import Thread
+import re
+import time 
+from PIL import Image
+import torch
+import spaces
 import subprocess
-import sys
-
-def install_flash_attn_wheel():
-    flash_attn_wheel_url = "https://github.com/Dao-AILab/flash-attention/releases/download/v2.6.3/flash_attn-2.6.3+cu123torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl"
-    try:
-        # Call pip to install the wheel file
-        subprocess.check_call([sys.executable, "-m", "pip", "install", flash_attn_wheel_url])
-        print("Wheel installed successfully!")
-    except subprocess.CalledProcessError as e:
-        print(f"Failed to install the flash attnetion wheel. Error: {e}")
-
-install_flash_attn_wheel()
-
-import cv2
-try:
-    from mmengine.visualization import Visualizer
-except ImportError:
-    Visualizer = None
-    print("Warning: mmengine is not installed, visualization is disabled.")
-    
-# Load the model and tokenizer 
-model_path = "ByteDance/Sa2VA-4B"
- 
-model = AutoModelForCausalLM.from_pretrained(
-    model_path,
-    torch_dtype="auto",
-    device_map="cuda:0",
+import os
+
+subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
+
+torch.set_default_device('cuda')
+
+
+IMAGENET_MEAN = (0.485, 0.456, 0.406)
+IMAGENET_STD = (0.229, 0.224, 0.225)
+
+def build_transform(input_size):
+    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
+    transform = T.Compose([
+        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
+        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
+        T.ToTensor(),
+        T.Normalize(mean=MEAN, std=STD)
+    ])
+    return transform
+
+def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
+    best_ratio_diff = float('inf')
+    best_ratio = (1, 1)
+    area = width * height
+    for ratio in target_ratios:
+        target_aspect_ratio = ratio[0] / ratio[1]
+        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
+        if ratio_diff < best_ratio_diff:
+            best_ratio_diff = ratio_diff
+            best_ratio = ratio
+        elif ratio_diff == best_ratio_diff:
+            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
+                best_ratio = ratio
+    return best_ratio
+
+def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
+    orig_width, orig_height = image.size
+    aspect_ratio = orig_width / orig_height
+
+    # calculate the existing image aspect ratio
+    target_ratios = set(
+        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
+        i * j <= max_num and i * j >= min_num)
+    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+    # find the closest aspect ratio to the target
+    target_aspect_ratio = find_closest_aspect_ratio(
+        aspect_ratio, target_ratios, orig_width, orig_height, image_size)
+
+    # calculate the target width and height
+    target_width = image_size * target_aspect_ratio[0]
+    target_height = image_size * target_aspect_ratio[1]
+    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
+
+    # resize the image
+    resized_img = image.resize((target_width, target_height))
+    processed_images = []
+    for i in range(blocks):
+        box = (
+            (i % (target_width // image_size)) * image_size,
+            (i // (target_width // image_size)) * image_size,
+            ((i % (target_width // image_size)) + 1) * image_size,
+            ((i // (target_width // image_size)) + 1) * image_size
+        )
+        # split the image
+        split_img = resized_img.crop(box)
+        processed_images.append(split_img)
+    assert len(processed_images) == blocks
+    if use_thumbnail and len(processed_images) != 1:
+        thumbnail_img = image.resize((image_size, image_size))
+        processed_images.append(thumbnail_img)
+    return processed_images
+
+def load_image(image_file, input_size=448, max_num=12):
+    image = Image.open(image_file).convert('RGB')
+    transform = build_transform(input_size=input_size)
+    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
+    pixel_values = [transform(image) for image in images]
+    pixel_values = torch.stack(pixel_values)
+    return pixel_values
+
+model = AutoModel.from_pretrained(
+    "5CD-AI/Vintern-3B-beta",
+    torch_dtype=torch.bfloat16,
+    low_cpu_mem_usage=True,
     trust_remote_code=True,
 ).eval().cuda()
+tokenizer = AutoTokenizer.from_pretrained("5CD-AI/Vintern-3B-beta", trust_remote_code=True, use_fast=False)
 
-tokenizer = AutoTokenizer.from_pretrained(
-    model_path,
-    trust_remote_code = True,
-)
-
-from third_parts import VideoReader
-def read_video(video_path, video_interval):
-    vid_frames = VideoReader(video_path)[::video_interval]
-    
-    temp_dir = tempfile.mkdtemp()
-    os.makedirs(temp_dir, exist_ok=True)
-    image_paths = []  # List to store paths of saved images
-    
-    for frame_idx in range(len(vid_frames)):
-        frame_image = vid_frames[frame_idx]
-        frame_image = frame_image[..., ::-1]  # BGR (opencv system) to RGB (numpy system)
-        frame_image = Image.fromarray(frame_image)
-        vid_frames[frame_idx] = frame_image
-
-        # Save the frame as a .jpg file in the temporary folder
-        image_path = os.path.join(temp_dir, f"frame_{frame_idx:04d}.jpg")
-        frame_image.save(image_path, format="JPEG")
-
-        # Append the image path to the list
-        image_paths.append(image_path)
-    return vid_frames, image_paths
-
-def visualize(pred_mask, image_path, work_dir):
-    visualizer = Visualizer()
-    img = cv2.imread(image_path)
-    visualizer.set_image(img)
-    visualizer.draw_binary_masks(pred_mask, colors='g', alphas=0.4)
-    visual_result = visualizer.get_image()
-
-    output_path = os.path.join(work_dir, os.path.basename(image_path))
-    cv2.imwrite(output_path, visual_result)
-    return output_path
 
 @spaces.GPU
-def image_vision(image_input_path, prompt):
-    image_path = image_input_path
-    text_prompts = f"<image>{prompt}"
-    image = Image.open(image_path).convert('RGB')
-    input_dict = {
-        'image': image,
-        'text': text_prompts,
-        'past_text': '',
-        'mask_prompts': None,
-        'tokenizer': tokenizer,
-    }
-    return_dict = model.predict_forward(**input_dict)
-    print(return_dict)
-    answer = return_dict["prediction"] # the text format answer
-    
-    seg_image = return_dict["prediction_masks"]
+def chat(message, history):
+    print("history",history)
+    print("message",message)
+
+    if len(history) != 0 and len(message["files"]) != 0:
+        return """Chúng tôi hiện chỉ hổ trợ 1 ảnh ở đầu ngữ cảnh! Vui lòng tạo mới cuộc trò chuyện.
+We currently only support one image at the start of the context! Please start a new conversation."""
     
-    if '[SEG]' in answer and Visualizer is not None:
-        pred_masks = seg_image[0]
-        temp_dir = tempfile.mkdtemp()
-        pred_mask = pred_masks
-        os.makedirs(temp_dir, exist_ok=True)
-        seg_result = visualize(pred_mask, image_input_path, temp_dir)
-        return answer, seg_result
+    if len(history) == 0 and len(message["files"]) != 0:
+        test_image = message["files"][0]["path"]
+        pixel_values = load_image(test_image, max_num=6).to(torch.bfloat16).cuda()
+    elif len(history) == 0 and len(message["files"]) == 0:
+        pixel_values = None     
+    elif history[0][0][0] is not None and os.path.isfile(history[0][0][0]):
+        test_image = history[0][0][0]
+        pixel_values = load_image(test_image, max_num=6).to(torch.bfloat16).cuda()
     else:
-        return answer, None
-
-@spaces.GPU(duration=80)
-def video_vision(video_input_path, prompt, video_interval):
-    # Open the original video
-    cap = cv2.VideoCapture(video_input_path)
-
-    # Get original video properties
-    original_fps = cap.get(cv2.CAP_PROP_FPS)
-
-    frame_skip_factor = video_interval
-
-    # Calculate new FPS
-    new_fps = original_fps / frame_skip_factor
-
-    vid_frames, image_paths = read_video(video_input_path, video_interval)
-    # create a question (<image> is a placeholder for the video frames)
-    question = f"<image>{prompt}"
-    result = model.predict_forward(
-        video=vid_frames,
-        text=question,
-        tokenizer=tokenizer,
-    )
-    prediction = result['prediction']
-    print(prediction)
-
-    if '[SEG]' in prediction and Visualizer is not None:
-        _seg_idx = 0
-        pred_masks = result['prediction_masks'][_seg_idx]
-        seg_frames = []
-        for frame_idx in range(len(vid_frames)):
-            pred_mask = pred_masks[frame_idx]
-            temp_dir = tempfile.mkdtemp()
-            os.makedirs(temp_dir, exist_ok=True)
-            seg_frame = visualize(pred_mask, image_paths[frame_idx], temp_dir)
-            seg_frames.append(seg_frame)
-
-        output_video = "output_video.mp4"
-
-        # Read the first image to get the size (resolution)
-        frame = cv2.imread(seg_frames[0])
-        height, width, layers = frame.shape
-
-        # Define the video codec and create VideoWriter object
-        fourcc = cv2.VideoWriter_fourcc(*'mp4v')  # Codec for MP4
-        video = cv2.VideoWriter(output_video, fourcc, new_fps, (width, height))
-
-        # Iterate over the image paths and write to the video
-        for img_path in seg_frames:
-            frame = cv2.imread(img_path)
-            video.write(frame)
-
-        # Release the video writer
-        video.release()
-
-        print(f"Video created successfully at {output_video}")
-
-        return result['prediction'], output_video
-            
-    else:
-        return result['prediction'], None
+        pixel_values = None 
+        
     
-
-
-# Gradio UI
-
-with gr.Blocks(analytics_enabled=False) as demo:
-    with gr.Column():
-        gr.Markdown("# Sa2VA: Marrying SAM2 with LLaVA for Dense Grounded Understanding of Images and Videos")
-        gr.HTML("""
-        <div style="display:flex;column-gap:4px;">
-            <a href="https://github.com/magic-research/Sa2VA">
-                <img src='https://img.shields.io/badge/GitHub-Repo-blue'>
-            </a> 
-            <a href="https://arxiv.org/abs/2501.04001">
-                <img src='https://img.shields.io/badge/ArXiv-Paper-red'>
-            </a>
-            <a href="https://huggingface.co/spaces/fffiloni/Sa2VA-simple-demo?duplicate=true">
-                <img src="https://huggingface.co/datasets/huggingface/badges/resolve/main/duplicate-this-space-sm.svg" alt="Duplicate this Space">
-            </a>
-            <a href="https://huggingface.co/fffiloni">
-                <img src="https://huggingface.co/datasets/huggingface/badges/resolve/main/follow-me-on-HF-sm-dark.svg" alt="Follow me on HF">
-            </a>
-        </div>
-        """)
-        with gr.Tab("Single Image"):
-            with gr.Row():
-                with gr.Column():
-                    image_input = gr.Image(label="Image IN", type="filepath")
-                    with gr.Row():
-                        instruction = gr.Textbox(label="Instruction", scale=4)
-                        submit_image_btn = gr.Button("Submit", scale=1)
-                with gr.Column():
-                    output_res = gr.Textbox(label="Response")
-                    output_image = gr.Image(label="Segmentation", type="numpy")
+    generation_config = dict(max_new_tokens= 512, do_sample=False, num_beams = 3, repetition_penalty=2.0)
     
-            submit_image_btn.click(
-                fn = image_vision,
-                inputs = [image_input, instruction],
-                outputs = [output_res, output_image]
-            )
-        with gr.Tab("Video"):
-            with gr.Row():
-                with gr.Column():
-                    video_input = gr.Video(label="Video IN")
-                    frame_interval = gr.Slider(label="Frame interval", step=1, minimum=1, maximum=12, value=6)
-                    with gr.Row():
-                        vid_instruction = gr.Textbox(label="Instruction", scale=4)
-                        submit_video_btn = gr.Button("Submit", scale=1)
-                with gr.Column():
-                    vid_output_res = gr.Textbox(label="Response")
-                    output_video = gr.Video(label="Segmentation")
-            
-            submit_video_btn.click(
-                fn = video_vision,
-                inputs = [video_input, vid_instruction, frame_interval],
-                outputs = [vid_output_res, output_video]
-            )
+    if len(history) == 0:
+        if pixel_values is not None:
+            question = '<image>\n'+message["text"]
+        else:
+            question = message["text"]
+        response, conv_history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
+    else:
+        conv_history = []
+        if history[0][0][0] is not None and os.path.isfile(history[0][0][0]):
+            start_index = 1
+        else:
+            start_index = 0
+        
+        for i, chat_pair in enumerate(history[start_index:]):
+            if i == 0 and start_index == 1:
+                 conv_history.append(tuple(['<image>\n'+chat_pair[0],chat_pair[1]]))
+            else:
+                conv_history.append(tuple(chat_pair))
 
-demo.queue().launch(show_api=False, show_error=True)
+            
+        print("conv_history",conv_history)
+        question = message["text"]
+        response, conv_history = model.chat(tokenizer, pixel_values, question, generation_config, history=conv_history, return_history=True)
+        
+    print(f'User: {question}\nAssistant: {response}')
+
+    return response
+    # buffer = ""
+    # for new_text in response:
+    #   buffer += new_text
+    #   generated_text_without_prompt = buffer[:]
+    #   time.sleep(0.005)
+    #   yield generated_text_without_prompt
+
+CSS ="""
+# @media only screen and (max-width: 600px){
+#     #component-3 {
+#       height: 90dvh !important;
+#       transform-origin: top; /* Đảm bảo rằng phần tử mở rộng từ trên xuống */
+#       border-style: solid;
+#       overflow: hidden;
+#       flex-grow: 1;
+#       min-width: min(160px, 100%);
+#       border-width: var(--block-border-width);
+#     }
+# }
+#component-3 {
+  height: 50dvh !important;
+  transform-origin: top; /* Đảm bảo rằng phần tử mở rộng từ trên xuống */
+  border-style: solid;
+  overflow: hidden;
+  flex-grow: 1;
+  min-width: min(160px, 100%);
+  border-width: var(--block-border-width);
+}
+/* Đảm bảo ảnh bên trong nút hiển thị đúng cách cho các nút có aria-label chỉ định */
+button.svelte-1lcyrx4[aria-label="user's message: a file of type image/jpeg, "] img.svelte-1pijsyv {
+  width: 100%;
+  object-fit: contain;
+  height: 100%;
+  border-radius: 13px; /* Thêm bo góc cho ảnh */
+  max-width: 50vw;     /* Giới hạn chiều rộng ảnh */
+}
+/* Đặt chiều cao cho nút và cho phép chọn văn bản chỉ cho các nút có aria-label chỉ định */
+button.svelte-1lcyrx4[aria-label="user's message: a file of type image/jpeg, "] {
+  user-select: text;
+  text-align: left;
+  height: 300px;
+}
+/* Thêm bo góc và giới hạn chiều rộng cho ảnh không thuộc avatar container */
+.message-wrap.svelte-1lcyrx4 > div.svelte-1lcyrx4 .svelte-1lcyrx4:not(.avatar-container) img {
+  border-radius: 13px;
+  max-width: 50vw;
+}
+.message-wrap.svelte-1lcyrx4 .message.svelte-1lcyrx4 img {
+    margin: var(--size-2);
+    max-height: 500px;
+}
+"""
+
+
+demo = gr.ChatInterface(
+    fn=chat,
+    description="""Try [Vintern-3B-beta](https://huggingface.co/5CD-AI/Vintern-3B-beta) in this demo. Vintern-3B-beta consists of [InternViT-300M-448px](https://huggingface.co/OpenGVLab/InternViT-300M-448px), an MLP projector, and [Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct).
+Bias, Risks, and Limitations
+The model might have biases because it learned from data that could be biased.
+Users should be cautious of these possible biases when using the model.""",
+    examples=[{"text": "Mô tả hình ảnh.", "files":["./demo_3.jpg"]},
+              {"text": "Trích xuất các thông tin từ ảnh.", "files":["./demo_1.jpg"]}, 
+              {"text": "Mô tả hình ảnh một cách chi tiết.", "files":["./demo_2.jpg"]}],
+    title="❄️ Vintern-3B-beta Test ❄️",
+    multimodal=True,
+    css=CSS
+)
+demo.queue().launch()