import ast import copy import glob import hashlib import logging import os import re from pathlib import Path from typing import List, Optional, Tuple from urllib.parse import urlparse import gradio as gr import PIL from gradio import processing_utils from gradio_client.client import DEFAULT_TEMP_DIR from text_generation import Client from transformers import AutoProcessor MODELS = [ # "HuggingFaceM4/idefics-9b-instruct", "HuggingFaceM4/idefics-80b-instruct", ] API_PATHS = { "HuggingFaceM4/idefics-9b-instruct": ( "https://api-inference.huggingface.co/models/HuggingFaceM4/idefics-9b-instruct" ), "HuggingFaceM4/idefics-80b-instruct": ( "https://api-inference.huggingface.co/models/HuggingFaceM4/idefics-80b-instruct" ), } SYSTEM_PROMPT = [ """The following is a conversation between a highly knowledgeable and intelligent visual AI assistant, called Assistant, and a human user, called User. In the following interactions, User and Assistant will converse in natural language, and Assistant will answer in a sassy way. Assistant's main purpose is to create funny meme texts from the images User provides. Assistant should be funny, sassy, and impertinent, and sometimes Assistant roasts people. Assistant should not be mean. It should not say toxic, homophobic, sexist, racist, things or any demeaning things that can make people uncomfortable. Assistant was created by Hugging Face. Here's a conversation example:""", """\nUser:""", "https://ichef.bbci.co.uk/news/976/cpsprodpb/7727/production/_103330503_musk3.jpg", "Write a meme for that image.", """\nAssistant: When you're trying to quit smoking but the cravings are too strong.""", "\nUser:How about this image?", "https://www.boredpanda.com/blog/wp-content/uploads/2017/01/image-copy-copy-587d0e7918b57-png__700.jpg", "Write something funny about this image.", """\nAssistant: Eggcellent service!""", "\nUser: Roast this person", "https://i.pinimg.com/564x/98/34/4b/98344b2483bd7c8b71a5c0fed6fe20b6.jpg", "", """\nAssistant: Damn your handwritting is pretty awful. But I suppose it must be pretty hard to hold a pen, considering you are a hammerhead shark.""", ] BAN_TOKENS = ( # For documentation puporse. We are not using this list, it is hardcoded inside `idefics_causal_lm.py` inside TGI. ";" ) EOS_STRINGS = ["", "\nUser:"] STOP_SUSPECT_LIST = [] GRADIO_LINK = "https://victorsanh-meme-it.hf.space" API_TOKEN = os.getenv("HF_AUTH_TOKEN") IDEFICS_LOGO = "https://huggingface.co/spaces/HuggingFaceM4/idefics_playground/resolve/main/IDEFICS_logo.png" PROCESSOR = AutoProcessor.from_pretrained( "HuggingFaceM4/idefics-9b-instruct", token=API_TOKEN, ) BOT_AVATAR = "IDEFICS_logo.png" logging.basicConfig(level=logging.INFO) logger = logging.getLogger() # Monkey patch adapted from gradio.components.image.Image - mostly to make the `save` step optional in `pil_to_temp_file` def hash_bytes(bytes: bytes): sha1 = hashlib.sha1() sha1.update(bytes) return sha1.hexdigest() def pil_to_temp_file(img: PIL.Image.Image, dir: str = DEFAULT_TEMP_DIR, format: str = "png") -> str: """Save a PIL image into a temp file""" bytes_data = processing_utils.encode_pil_to_bytes(img, format) temp_dir = Path(dir) / hash_bytes(bytes_data) temp_dir.mkdir(exist_ok=True, parents=True) filename = str(temp_dir / f"image.{format}") if not os.path.exists(filename): img.save(filename, pnginfo=processing_utils.get_pil_metadata(img)) return filename def add_file(file): return file.name, gr.update(label='๐Ÿ–ผ๏ธ Uploaded!') # This is a hack to make pre-computing the default examples work. # During normal inference, we pass images as url to a local file using the method `gradio_link` # which allows the tgi server to fetch the local image from the frontend server. # however, we are building the space (and pre-computing is part of building the space), the frontend is not available # and won't answer. So tgi server will try to fetch an image that is not available yet, which will result in a timeout error # because tgi will never be able to return the generation. # To bypass that, we pass instead the images URLs from the spaces repo. all_images = glob.glob(f"{os.path.dirname(__file__)}/example_images/*") DEFAULT_IMAGES_TMP_PATH_TO_URL = {} for im_path in all_images: H = gr.Image(im_path, visible=False, type="filepath") tmp_filename = H.preprocess(H.value) DEFAULT_IMAGES_TMP_PATH_TO_URL[tmp_filename] = f"https://huggingface.co/spaces/HuggingFaceM4/idefics_playground/resolve/main/example_images/{os.path.basename(im_path)}" # Utils to handle the image markdown display logic def split_str_on_im_markdown(string: str) -> List[str]: """ Extract from a string (typically the user prompt string) the potential images from markdown Examples: - `User:![](https://favurl.com/chicken_on_money.png)Describe this image.` would become `["User:", "https://favurl.com/chicken_on_money.png", "Describe this image."]` - `User:![](/file=/my_temp/chicken_on_money.png)Describe this image.` would become `["User:", "/my_temp/chicken_on_money.png", "Describe this image."]` """ IMAGES_PATTERN = re.compile(r"!\[[^\]]*\]\((.*?)\s*(\"(?:.*[^\"])\")?\s*\)") parts = [] cursor = 0 for pattern in IMAGES_PATTERN.finditer(string): start = pattern.start() if start != cursor: parts.append(string[cursor:start]) image_url = pattern.group(1) if image_url.startswith("/file="): image_url = image_url[6:] # Remove the 'file=' prefix parts.append(image_url) cursor = pattern.end() if cursor != len(string): parts.append(string[cursor:]) return parts def is_image(string: str) -> bool: """ There are two ways for images: local image path or url. """ return is_url(string) or string.startswith(DEFAULT_TEMP_DIR) def is_url(string: str) -> bool: """ Checks if the passed string contains a valid url and nothing else. e.g. if space is included it's immediately invalidated the url """ if " " in string: return False result = urlparse(string) return all([result.scheme, result.netloc]) def isolate_images_urls(prompt_list: List) -> List: """ Convert a full string prompt to the list format expected by the processor. In particular, image urls (as delimited by ) should be their own elements. From: ``` [ "bonjourhello", PIL.Image.Image, "Aurevoir", ] ``` to: ``` [ "bonjour", IMG_URL, "hello", PIL.Image.Image, "Aurevoir", ] ``` """ linearized_list = [] for prompt in prompt_list: # Prompt can be either a string, or a PIL image if isinstance(prompt, PIL.Image.Image): linearized_list.append(prompt) elif isinstance(prompt, str): if "" not in prompt: linearized_list.append(prompt) else: prompt_splitted = prompt.split("") for ps in prompt_splitted: if ps == "": continue if ps.startswith(" PIL.Image.Image: """Fetching images""" return PROCESSOR.image_processor.fetch_images(url_list) def handle_manual_images_in_user_prompt(user_prompt: str) -> List[str]: """ Handle the case of textually manually inputted images (i.e. the ``) in the user prompt by fetching them, saving them locally and replacing the whole sub-sequence the image local path. """ if "" in user_prompt: splitted_user_prompt = isolate_images_urls([user_prompt]) resulting_user_prompt = [] for u_p in splitted_user_prompt: if is_url(u_p): img = fetch_images([u_p])[0] tmp_file = pil_to_temp_file(img) resulting_user_prompt.append(tmp_file) else: resulting_user_prompt.append(u_p) return resulting_user_prompt else: return [user_prompt] def gradio_link(img_path: str) -> str: url = f"{GRADIO_LINK}/file={img_path}" return url def prompt_list_to_markdown(prompt_list: List[str]) -> str: """ Convert a user prompt in the list format (i.e. elements are either a PIL image or a string) into the markdown format that is used for the chatbot history and rendering. """ resulting_string = "" for elem in prompt_list: if is_image(elem): if is_url(elem): resulting_string += f"![]({elem})" else: resulting_string += f"![](/file={elem})" else: resulting_string += elem return resulting_string def prompt_list_to_tgi_input(prompt_list: List[str]) -> str: """ TGI expects a string that contains both text and images in the image markdown format (i.e. the `![]()` ). The images links are parsed on TGI side """ result_string_input = "" for elem in prompt_list: if is_image(elem): if is_url(elem): result_string_input += f"![]({elem})" else: result_string_input += f"![]({gradio_link(img_path=elem)})" else: result_string_input += elem return result_string_input def remove_spaces_around_token(text: str) -> str: pattern = r"\s*()\s*" replacement = r"\1" result = re.sub(pattern, replacement, text) return result # Chatbot utils def format_user_prompt_with_im_history_and_system_conditioning( system_prompt: List[str], current_user_prompt_str: str, current_image: Optional[str], history: List[Tuple[str, str]] ) -> Tuple[List[str], List[str]]: """ Produces the resulting list that needs to go inside the processor. It handles the potential image box input, the history and the system conditionning. """ # resulting_list = copy.deepcopy(SYSTEM_PROMPT) resulting_list = system_prompt # Format history for turn in history: user_utterance, assistant_utterance = turn splitted_user_utterance = split_str_on_im_markdown(user_utterance) optional_space = "" if not is_image(splitted_user_utterance[0]): optional_space = " " resulting_list.append(f"\nUser:{optional_space}") resulting_list.extend(splitted_user_utterance) resulting_list.append(f"\nAssistant: {assistant_utterance}") # Format current input current_user_prompt_str = remove_spaces_around_token(current_user_prompt_str) if current_image is None: if "![](" in current_user_prompt_str: current_user_prompt_list = split_str_on_im_markdown(current_user_prompt_str) else: current_user_prompt_list = handle_manual_images_in_user_prompt(current_user_prompt_str) optional_space = "" if not is_image(current_user_prompt_list[0]): # Check if the first element is an image (and more precisely a path to an image) optional_space = " " resulting_list.append(f"\nUser:{optional_space}") resulting_list.extend(current_user_prompt_list) resulting_list.append("\nAssistant:") else: # Choosing to put the image first when the image is inputted through the UI, but this is an arbiratrary choice. resulting_list.extend(["\nUser:", current_image, f"{current_user_prompt_str}\nAssistant:"]) current_user_prompt_list = [current_user_prompt_str] return resulting_list, current_user_prompt_list # dope_callback = gr.CSVLogger() # problematic_callback = gr.CSVLogger() textbox = gr.Textbox( placeholder="Upload an image and start conversing by sending a message! You can add an image at each turn, but don't have to.", show_label=False, # value="Write something funny about that image.", visible=True, container=False, label="Text input", scale=6, max_lines=5, ) with gr.Blocks(title="AI Dad Jokes", theme=gr.themes.Base()) as demo: gr.HTML("""

AI Dad Jokes

""") with gr.Row(variant="panel"): with gr.Column(scale=1): gr.Image(IDEFICS_LOGO, elem_id="banner-image", show_label=False, show_download_button=False) with gr.Column(scale=5): gr.HTML("""

AI Dad Jokes is an AI system that writes humorous content inspired by images. Whether that's crafting memes, sharing light-hearted yet amiable jests, or playfully witty remarks, AI Dad Jokes assists you in creating delightful jokes!

AI Dad Jokes is powered by IDEFICS, an open-access large visual language model developped by Hugging Face. Like GPT-4, the multimodal model accepts arbitrary sequences of image and text inputs and produces text outputs. IDEFICS can answer questions about images, describe visual content, create stories grounded in multiple images, etc.

โ›”๏ธ Intended uses and limitations: This demo is provided as research artifact to the community showcasing IDEFIC's capabilities. We detail misuses and out-of-scope uses here. In particular, the system should not be used to engage in harassment, abuse and bullying. The model can produce factually incorrect texts, hallucinate facts (with or without an image) and will struggle with small details in images. While the system will tend to refuse answering questionable user requests, it can produce problematic outputs (including racist, stereotypical, and disrespectful texts), in particular when prompted to do so.

""") with gr.Row(elem_id="model_selector_row"): model_selector = gr.Dropdown( choices=MODELS, value="HuggingFaceM4/idefics-80b-instruct", interactive=True, show_label=False, container=False, label="Model", visible=False, ) with gr.Row(): with gr.Column(): imagebox = gr.Image(type="filepath", label="Image input", visible=True) with gr.Column(): chatbot = gr.Chatbot( elem_id="chatbot", label="IDEFICS", visible=True, height=750, avatar_images=[None, BOT_AVATAR] # value=[ # [ # ( # prefetch_images_in_history( # """Write a funny story including the following characters. # There should be drama and most importantly romance. # Character 1: # Character 2: # Character 3: # The plot should take place in the world pictured here:""" # ) # ), # """Once upon a time, in a world where the sky was a dome and the ground was covered in grass, there lived a beautiful princess named Mulan. She was kind, brave, and always ready to help others. One day, while walking through the forest, she stumbled upon a strange creature. It was a green ogre named Shrek, who was lost and looking for his way back home. # Mulan, being the kind-hearted person she was, decided to help Shrek find his way back. As they traveled together, they began to develop a strong bond. Shrek was impressed by Mulan's bravery and determination, while Mulan admired Shrek's loyalty and sense of humor. # However, their journey was not without obstacles. They encountered a mischievous cat named Hoboua, who was always causing trouble and getting into fights. Despite his troublemaking ways, Hoboua had a good heart and eventually became a valuable ally to Mulan and Shrek. # As they traveled, they faced many challenges, including battling fierce dragons and navigating treacherous mountains. Through it all, Mulan and Shrek grew closer, and their feelings for each other deepened. # Finally, they reached Shrek's home, and he was reunited with his family and friends. Mulan, however, was sad to leave him behind. But Shrek had a surprise for her. He had fallen in love with her and wanted to be with her forever. # Mulan was overjoyed, and they shared a passionate kiss. From that day on, they lived happily ever after, exploring the world together and facing any challenges that came their way. # And so, the story of Mulan and Shrek's romance came to an end, leaving a lasting impression on all who heard it.""", # ], # ], ) with gr.Group(): with gr.Row(): textbox.render() submit_btn = gr.Button(value="โ–ถ๏ธ Submit", visible=True) clear_btn = gr.ClearButton([textbox, imagebox, chatbot], value="๐Ÿงน Clear") regenerate_btn = gr.Button(value="๐Ÿ”„ Regenerate", visible=True) upload_btn = gr.UploadButton("๐Ÿ“ Upload image", file_types=["image"]) with gr.Row(): with gr.Accordion("Advanced settings", open=False, visible=True) as parameter_row: system_prompt = gr.Textbox( value=SYSTEM_PROMPT, visible=False, lines=20, max_lines=50, interactive=True, ) max_new_tokens = gr.Slider( minimum=8, maximum=256, value=64, step=1, interactive=True, label="Maximum number of new tokens to generate", ) repetition_penalty = gr.Slider( minimum=0.0, maximum=5.0, value=1.2, step=0.01, interactive=True, label="Repetition penalty", info="1.0 is equivalent to no penalty", ) decoding_strategy = gr.Radio( [ "Greedy", "Top P Sampling", ], value="Top P Sampling", label="Decoding strategy", interactive=True, info="Higher values is equivalent to sampling more low-probability tokens.", ) temperature = gr.Slider( minimum=0.0, maximum=5.0, value=0.6, step=0.1, interactive=True, visible=True, label="Sampling temperature", info="Higher values will produce more diverse outputs.", ) decoding_strategy.change( fn=lambda selection: gr.Slider.update( visible=( selection in ["contrastive_sampling", "beam_sampling", "Top P Sampling", "sampling_top_k"] ) ), inputs=decoding_strategy, outputs=temperature, ) top_p = gr.Slider( minimum=0.01, maximum=0.99, value=0.8, step=0.01, interactive=True, visible=True, label="Top P", info="Higher values is equivalent to sampling more low-probability tokens.", ) decoding_strategy.change( fn=lambda selection: gr.Slider.update(visible=(selection in ["Top P Sampling"])), inputs=decoding_strategy, outputs=top_p, ) def model_inference( model_selector, system_prompt, user_prompt_str, chat_history, image, decoding_strategy, temperature, max_new_tokens, repetition_penalty, top_p, ): if user_prompt_str.strip() == "" and image is None: return "", None, chat_history system_prompt = ast.literal_eval(system_prompt) formated_prompt_list, user_prompt_list = format_user_prompt_with_im_history_and_system_conditioning( system_prompt=system_prompt, current_user_prompt_str=user_prompt_str.strip(), current_image=image, history=chat_history, ) client_endpoint = API_PATHS[model_selector] client = Client( base_url=client_endpoint, headers={"x-use-cache": "0", "Authorization": f"Bearer {API_TOKEN}"}, ) # Common parameters to all decoding strategies # This documentation is useful to read: https://huggingface.co/docs/transformers/main/en/generation_strategies generation_args = { "max_new_tokens": max_new_tokens, "repetition_penalty": repetition_penalty, "stop_sequences": EOS_STRINGS, } assert decoding_strategy in [ "Greedy", "Top P Sampling", ] if decoding_strategy == "Greedy": generation_args["do_sample"] = False elif decoding_strategy == "Top P Sampling": generation_args["temperature"] = temperature generation_args["do_sample"] = True generation_args["top_p"] = top_p if image is None: # Case where there is no image OR the image is passed as `` chat_history.append([prompt_list_to_markdown(user_prompt_list), '']) else: # Case where the image is passed through the Image Box. # Convert the image into base64 for both passing it through the chat history and # displaying the image inside the same bubble as the text. chat_history.append( [ f"{prompt_list_to_markdown([image] + user_prompt_list)}", '', ] ) query = prompt_list_to_tgi_input(formated_prompt_list) stream = client.generate_stream(prompt=query, **generation_args) acc_text = "" for idx, response in enumerate(stream): text_token = response.token.text if response.details: # That's the exit condition return if text_token in STOP_SUSPECT_LIST: acc_text += text_token continue if idx == 0 and text_token.startswith(" "): text_token = text_token.lstrip() acc_text += text_token last_turn = chat_history.pop(-1) last_turn[-1] += acc_text if last_turn[-1].endswith("\nUser"): # Safeguard: sometimes (rarely), the model won't generate the token `` and will go directly to generating `\nUser:` # It will thus stop the generation on `\nUser:`. But when it exits, it will have already generated `\nUser` # This post-processing ensures that we don't have an additional `\nUser` wandering around. last_turn[-1] = last_turn[-1][:-5] chat_history.append(last_turn) yield "", None, chat_history acc_text = "" def process_example(message, image): """ Same as `model_inference` but in greedy mode and with the 80b-instruct. Specifically for pre-computing the default examples. """ model_selector="HuggingFaceM4/idefics-80b-instruct" user_prompt_str=message chat_history=[] max_new_tokens=512 formated_prompt_list, user_prompt_list = format_user_prompt_with_im_history_and_system_conditioning( current_user_prompt_str=user_prompt_str.strip(), current_image=image, history=chat_history, ) client_endpoint = API_PATHS[model_selector] client = Client( base_url=client_endpoint, headers={"x-use-cache": "0", "Authorization": f"Bearer {API_TOKEN}"}, timeout=240, # Generous time out just in case because we are in greedy. All examples should be computed in less than 30secs with the 80b-instruct. ) # Common parameters to all decoding strategies # This documentation is useful to read: https://huggingface.co/docs/transformers/main/en/generation_strategies generation_args = { "max_new_tokens": max_new_tokens, "repetition_penalty": None, "stop_sequences": EOS_STRINGS, "do_sample": False, } if image is None: # Case where there is no image OR the image is passed as `` chat_history.append([prompt_list_to_markdown(user_prompt_list), '']) else: # Case where the image is passed through the Image Box. # Convert the image into base64 for both passing it through the chat history and # displaying the image inside the same bubble as the text. chat_history.append( [ f"{prompt_list_to_markdown([image] + user_prompt_list)}", '', ] ) # Hack - see explanation in `DEFAULT_IMAGES_TMP_PATH_TO_URL` for idx, i in enumerate(formated_prompt_list): if i.startswith(DEFAULT_TEMP_DIR): for k, v in DEFAULT_IMAGES_TMP_PATH_TO_URL.items(): if k == i: formated_prompt_list[idx] = v break query = prompt_list_to_tgi_input(formated_prompt_list) generated_text = client.generate(prompt=query, **generation_args).generated_text if generated_text.endswith("\nUser"): generated_text = generated_text[:-5] last_turn = chat_history.pop(-1) last_turn[-1] += generated_text chat_history.append(last_turn) return "", None, chat_history textbox.submit( fn=model_inference, inputs=[ model_selector, system_prompt, textbox, chatbot, imagebox, decoding_strategy, temperature, max_new_tokens, repetition_penalty, top_p, ], outputs=[textbox, imagebox, chatbot], ) submit_btn.click( fn=model_inference, inputs=[ model_selector, system_prompt, textbox, chatbot, imagebox, decoding_strategy, temperature, max_new_tokens, repetition_penalty, top_p, ], outputs=[ textbox, imagebox, chatbot, ], ) def remove_last_turn(chat_history): if len(chat_history) == 0: return gr.Update(), gr.Update() last_interaction = chat_history[-1] chat_history = chat_history[:-1] chat_update = gr.update(value=chat_history) text_update = gr.update(value=last_interaction[0]) return chat_update, text_update regenerate_btn.click(fn=remove_last_turn, inputs=chatbot, outputs=[chatbot, textbox]).then( fn=model_inference, inputs=[ model_selector, system_prompt, textbox, chatbot, imagebox, decoding_strategy, temperature, max_new_tokens, repetition_penalty, top_p, ], outputs=[ textbox, imagebox, chatbot, ], ) upload_btn.upload(add_file, [upload_btn], [imagebox, upload_btn], queue=False) submit_btn.click(lambda : gr.update(label='๐Ÿ“ Upload image', interactive=True), [], upload_btn) textbox.submit(lambda : gr.update(label='๐Ÿ“ Upload image', interactive=True), [], upload_btn) clear_btn.click(lambda : gr.update(label='๐Ÿ“ Upload image', interactive=True), [], upload_btn) examples_path = os.path.dirname(__file__) gr.Examples( examples=[ [ "Write a meme text for that image.", f"{examples_path}/example_images/citibike.webp", ], [ "Write something funny about that. A meme for instance.", f"{examples_path}/example_images/zuck.jpeg", ], [ "Craft a humorous caption for this image!", f"{examples_path}/example_images/echasse.jpg", ], [ "How about adding a dash of humor to this image with your words?", f"{examples_path}/example_images/jesus.jpg", ], [ "Give this image a comedic twist.", f"{examples_path}/example_images/owl.jpg", ], [ "Tell me a joke about that image.", f"{examples_path}/example_images/pigeon.jpg", ], [ "Let your sense of humor shine with that image!", f"{examples_path}/example_images/plotorange.jpg", ], [ "Make me laugh by commenting that image.", f"{examples_path}/example_images/rats.jpg", ], [ "Craft a meme text for that image.", f"{examples_path}/example_images/sugardaddy.jpg", ], [ "Ready to make this image even better? Write something funny to go with it!", f"{examples_path}/example_images/wtf.jpg", ], ], inputs=[textbox, imagebox], outputs=[textbox, imagebox, chatbot], fn=process_example, cache_examples=False, examples_per_page=6, label=( "Click on any example below to get started." ), ) demo.queue(concurrency_count=40, max_size=40) demo.launch()