Spaces:
Running
Running
| from typing import List | |
| import gradio as gr | |
| import PIL | |
| from gradio import ChatMessage | |
| from smolagents.gradio_ui import stream_to_gradio | |
| from agents.all_agents import get_master_agent | |
| from llm import get_default_model | |
| gr.set_static_paths(paths=["images/"]) | |
| master_agent = get_master_agent(get_default_model()) | |
| print(master_agent) | |
| def resize_image(image): | |
| width, height = image.size | |
| if width > 1200 or height > 800: | |
| ratio = min(1200 / width, 800 / height) | |
| new_width = int(width * ratio) | |
| new_height = int(height * ratio) | |
| resized_image = image.resize((new_width, new_height), PIL.Image.Resampling.LANCZOS) | |
| return resized_image | |
| return image | |
| def chat_interface_fn(input_request, history: List[ChatMessage], gallery): | |
| if gallery is None: | |
| gallery = [] | |
| else: | |
| gallery = [value[0] for value in gallery] | |
| message = input_request["text"] | |
| image_paths = input_request["files"] | |
| prompt = f""" | |
| You are given the following message from the user: | |
| {message} | |
| """ | |
| if len(image_paths) > 0: | |
| prompt += """ | |
| The user also provided the additional images that you can find in "images" variable | |
| """ | |
| if len(history) > 0: | |
| prompt += "This request follows a previous request, you can use the previous request to help you answer the current request." | |
| prompt += """ | |
| Before your final answer, if you have any images to show, store them in the "final_images" variable. | |
| Always return a text of what you did. | |
| """ | |
| images = [PIL.Image.open(image_path) for image_path in image_paths] | |
| if len(gallery) > 0: | |
| images.extend(gallery) | |
| resized_images = [resize_image(image) for image in images] | |
| for message in stream_to_gradio( | |
| master_agent, | |
| task=prompt, | |
| task_images=resized_images, | |
| additional_args={"images": images}, | |
| reset_agent_memory=False, | |
| ): | |
| history.append(message) | |
| yield history, None | |
| final_images = master_agent.python_executor.state.get("final_images", []) | |
| gallery.extend(final_images) | |
| yield history, gallery | |
| def example_selected(example): | |
| textbox.value = example[0] | |
| image_box.value = example[1] | |
| example = { | |
| "text": example[0], | |
| "files": [ | |
| { | |
| "url": example[1], | |
| "path": example[1], | |
| "name": example[1], | |
| } | |
| ], | |
| } | |
| return example | |
| with gr.Blocks() as demo: | |
| gr.Markdown( | |
| """ | |
| # ScouterAI | |
| { width="800" height="600" style="display: block; margin: 0 auto" } | |
| Welcome to ScouterAI, the Agent that is capable of detecting over 9000 entities in images using the best models of the HuggingFace Hub. | |
| """) | |
| output_gallery = gr.Gallery(label="Output Gallery", type="pil", format="png") | |
| textbox = gr.MultimodalTextbox() | |
| gr.ChatInterface( | |
| chat_interface_fn, | |
| type="messages", | |
| multimodal=True, | |
| textbox=textbox, | |
| additional_inputs=[output_gallery], | |
| additional_outputs=[output_gallery], | |
| ) | |
| text_box = gr.Textbox(label="Text", visible=False) | |
| image_box = gr.Image(label="Image", visible=False) | |
| dataset = gr.Dataset( | |
| samples=[ | |
| [ | |
| "I would like to detect all the cars in the image", | |
| "https://upload.wikimedia.org/wikipedia/commons/5/51/Crossing_the_Hudson_River_on_the_George_Washington_Bridge_from_Fort_Lee%2C_New_Jersey_to_Manhattan%2C_New_York_%287237796950%29.jpg", | |
| ], | |
| [ | |
| "Find vegetables in the image and annotate the image with their masks", | |
| "https://media.istockphoto.com/id/1203599923/fr/photo/fond-de-nourriture-avec-lassortiment-des-l%C3%A9gumes-organiques-frais.jpg?s=612x612&w=0&k=20&c=Yu8nfOYI9YZ0UTpb7iFqX8OHp9wfvd9keMQ0BZIzhWs=", | |
| ], | |
| ], | |
| components=[text_box, image_box], | |
| label="Examples", | |
| ) | |
| dataset.select(example_selected, [dataset], [textbox]) | |
| demo.launch() | |