import base64 import copy import logging import os import re from io import BytesIO from typing import List, Optional, Tuple, Union from urllib.parse import urlparse import gradio as gr import PIL from accelerate.utils import get_max_memory from PIL import Image from transformers import AutoConfig, AutoProcessor, IdeficsForVisionText2Text MODELS = [ "HuggingFaceM4/idefics-9b-instruct", # "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 do its best to answer User’s questions. Assistant has the ability to perceive images and reason about the content of visual inputs. Assistant was built to be respectful, polite and inclusive. It knows a lot, and always tells the truth. When prompted with an image, it does not make up facts. The conversation begins:""", """\nUser:Describe this image.""", """\nAssistant: Five kittens are standing together in the center of the photograph. From the left to right, there is one orange kitten, two kittens white and black stripes, and two brown kittens with an orange head. They are in bright green grass and it looks like they are moving forward.""", ( "\nUser:How about this" " image?Can" " you describe it too?" ), """\nAssistant: A dog is lying on the floor, looking at the camera. It is looking directly at you. The dog has a white body and brown patches on its face and ears. Its eyes are dark. Its nose is black, and it has long, floppy ears, white paws, long fur, big eyes.""", "\nUser: What kind of breed is it?", """\nAssistant: It looks like a Cavalier King Charles Spaniel. These breeds are known for their gentle, affectionate nature. Their long ears and big, brown eyes give them a sweet, expressive face. Cavaliers are small dogs weighing between 13 and 20 pounds. They have a silky, medium-length coat that comes in various colors, including black and tan, tricolor, red and white, or black.""", "\nUser: What can you tell me about this breed of dogs?", """\nAssistant: One specific characteristic of the Cavalier King Charles Spaniel is its friendly and affectionate nature. This breed is known for being extremely sociable and forming strong bonds with their owners. They are often described as "velcro dogs" because they love to be close to their human companions, whether it's sitting on their laps, cuddling on the couch, or simply following them around the house.""", "\nUser: ghjkhjabnufs", """\nAssistant: That doesn’t seem to be a word. Could you ask me another way?""", "\nUser: Do you like Cavalier King Charles Spaniel?\n", """\nAssistant: I do not have personal opinions as I’m just a computer program. However, cavaliers are known for being excellent family pets due to their gentle and patient demeanor, making them great with children and other pets. Their affectionate nature also makes them ideal therapy dogs, as they can provide comfort and emotional support to those in need.""", ( "\nUser: How many dogs do you see in this" " image?" ), """\nAssistant: There is no dogs in this image. The picture shows a tennis player jumping to volley the ball.""", ] BAN_TOKENS = ";" EOS_TOKENS = ";;User" TOKENIZER_FAST = True MAX_SEQ_LEN = 2048 logging.basicConfig(level=logging.INFO) logger = logging.getLogger() def convert_to_rgb(image): # `image.convert("RGB")` would only work for .jpg images, as it creates a wrong background # for transparent images. The call to `alpha_composite` handles this case if image.mode == "RGB": return image image_rgba = image.convert("RGBA") background = Image.new("RGBA", image_rgba.size, (255, 255, 255)) alpha_composite = Image.alpha_composite(background, image_rgba) alpha_composite = alpha_composite.convert("RGB") return alpha_composite # Conversion between PIL Image <-> base64 <-> Markdown utils def pil_to_base64(pil_image): """ Convert an PIL image into base64 string representation """ buffered = BytesIO() pil_image.save(buffered, format="JPEG") # You can change the format as per your image type encoded_image = base64.b64encode(buffered.getvalue()).decode("utf-8") return encoded_image def pil_to_markdown_im(image): """ Convert a PIL image into markdown filled with the base64 string representation. """ img_b64_str = pil_to_base64(image) img_str = f'' return img_str def base64_to_pil(encoded_image): decoded_image = base64.b64decode(encoded_image) pil_image = Image.open(BytesIO(decoded_image)) return pil_image def im_markdown_to_pil(im_markdown_str): pattern = r'' match = re.search(pattern, im_markdown_str) img_b64_str = match.group(1) return base64_to_pil(img_b64_str) def split_str_on_im_markdown(string_with_potential_im_markdown): """ Extract from a string (typically the user prompt string) the potentional images saved as a base64 representation inside a markdown. """ pattern = r'' parts = re.split(pattern, string_with_potential_im_markdown) result = [] for i, part in enumerate(parts): if i % 2 == 0: result.append(part) else: img_tag = f'' result.append(img_tag) return result # Fetching utils def is_url(string): """ 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): """ 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(" List[Union[str, PIL.Image.Image]]: """ Handle the case of textually manually inputted images (i.e. the ``) in the user prompt by fetching them and replacing the whole sub-sequence by a PIL image. """ if "" in user_prompt: splitted_user_prompt = isolate_images_urls([user_prompt]) resulting_user_prompt = [] for up in splitted_user_prompt: if is_url(up): img = processor.image_processor.fetch_images([up])[0] resulting_user_prompt.append(img) else: resulting_user_prompt.append(up) return resulting_user_prompt else: return [user_prompt] def user_prompt_list_to_markdown(user_prompt_list: List[Union[str, PIL.Image.Image]]): """ 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 user_prompt_list: if isinstance(elem, str): resulting_string += elem elif isinstance(elem, PIL.Image.Image): resulting_string += pil_to_markdown_im(convert_to_rgb(elem)) else: raise ValueError( "Unknown type for `user_prompt_list`. Expected an element of type `str` or `PIL.Image.Image` and got" f" `{type(elem)}`" ) return resulting_string # Model and generation utils def load_processor_tokenizer_model(model_name): processor = AutoProcessor.from_pretrained( model_name, token=os.getenv("HF_AUTH_TOKEN", True), truncation_side="left", ) tokenizer = processor.tokenizer config = AutoConfig.from_pretrained(model_name, use_auth_token=os.getenv("HF_AUTH_TOKEN", True)) max_memory_map = get_max_memory() for key in max_memory_map.keys(): if key != "cpu": # Get this in GB max_memory_map[key] = max_memory_map[key] // (1024 * 1024 * 1024) # Decrease 2 for Pytorch overhead and 2 for the forward to be safe max_memory_map[key] = f"{max_memory_map[key] - 4} GiB" model = IdeficsForVisionText2Text.from_pretrained( model_name, token=os.getenv("HF_AUTH_TOKEN", True), device_map="auto", offload_folder="./offload", torch_dtype=config.torch_dtype, max_memory=max_memory_map, ) model.eval() print("Current device map:", model.hf_device_map) print("Model default generation config:", model.generation_config) # TODO: the device_map looks very inefficien right now. that could be improved return processor, tokenizer, model def format_user_prompt_with_im_history_and_system_conditioning( current_user_prompt_str: str, current_image: Optional[PIL.Image.Image], history: List[Tuple[str, str]] ) -> List[Union[str, PIL.Image.Image]]: """ 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) # Format history for turn in history: user_utterance, assistant_utterance = turn splitted_user_utterance = split_str_on_im_markdown(user_utterance) splitted_user_utterance = [ im_markdown_to_pil(s) if s.startswith('\nAssistant: {assistant_utterance}") # Format current input if current_image is None: if "\nAssistant:") return resulting_list, current_user_prompt_list 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:"]) return resulting_list, [current_user_prompt_str] def model_generation( prompt_list, processor, tokenizer, model, temperature, no_repeat_ngram_size, max_new_tokens, min_length, ban_tokens, eos_tokens, force_words, repetition_penalty, hide_special_tokens, decoding_strategy, num_beams, length_penalty, top_k, top_p, penalty_alpha, ): input_args = processor( isolate_images_urls(prompt_list), truncation=True, max_length=MAX_SEQ_LEN - max_new_tokens, padding=True, # add_end_of_utterance_token=False, # Already taken care of inside the prompts, so bypassing the processor's handling of this token ) for k, v in input_args.items(): input_args[k] = v.to(0) # Excluding some words from the generation bad_words_ids = None ban_tokens = ban_tokens.replace("\\n", "\n") bad_words = ban_tokens.split(";") if len(bad_words) > 0: bad_words_ids = tokenizer(bad_words, add_special_tokens=False).input_ids # Forcing some words in the generation force_words_ids = None if force_words != "": force_words = force_words.replace("\\n", "\n") force_words = force_words.split(";") if len(force_words) > 0: force_words_ids = tokenizer(force_words, add_special_tokens=False).input_ids eos_token_ids = None if eos_tokens != "": eos_tokens = eos_tokens.replace("\\n", "\n") eos_tokens = eos_tokens.split(";") if len(eos_tokens) > 0: eos_token_ids = [] for eos_token in eos_tokens: tokenized_eos_token = tokenizer(eos_token, add_special_tokens=False).input_ids if len(tokenized_eos_token) > 1: raise ValueError( f"eos_tokens should be one token, here {eos_token} is {len(tokenized_eos_token)} tokens:" f" {tokenized_eos_token}" ) eos_token_ids += tokenized_eos_token # Common parameters to all decoding strategies # This documentation is useful to read: https://huggingface.co/docs/transformers/main/en/generation_strategies generation_args = { "temperature": temperature, "no_repeat_ngram_size": no_repeat_ngram_size, "max_new_tokens": max_new_tokens, "min_length": min_length, "bad_words_ids": bad_words_ids, "force_words_ids": force_words_ids, "repetition_penalty": repetition_penalty, "eos_token_id": eos_token_ids, } assert decoding_strategy in [ "greedy", "beam_search", "beam_sampling", "sampling_top_k", "sampling_top_p", "contrastive_sampling", ] if decoding_strategy == "greedy": pass elif decoding_strategy == "beam_search": generation_args["num_beams"] = num_beams generation_args["length_penalty"] = length_penalty assert generation_args["num_beams"] > 1 elif decoding_strategy == "beam_sampling": generation_args["num_beams"] = num_beams generation_args["length_penalty"] = length_penalty generation_args["do_sample"] = True assert generation_args["num_beams"] > 1 elif decoding_strategy == "sampling_top_k": generation_args["do_sample"] = True generation_args["top_k"] = top_k elif decoding_strategy == "sampling_top_p": generation_args["do_sample"] = True generation_args["top_p"] = top_p elif decoding_strategy == "contrastive_sampling": generation_args["do_sample"] = True generation_args["penalty_alpha"] = penalty_alpha generation_args["top_k"] = top_k generated_tokens = model.generate( **input_args, **generation_args, ) tokens = tokenizer.convert_ids_to_tokens(generated_tokens[0]) decoded_skip_special_tokens = repr( tokenizer.batch_decode(generated_tokens, skip_special_tokens=hide_special_tokens)[0] ) actual_generated_tokens = generated_tokens[:, input_args["input_ids"].shape[-1] :] first_end_token = len(actual_generated_tokens[0]) actual_generated_tokens = actual_generated_tokens[:, :first_end_token] generated_text = tokenizer.batch_decode(actual_generated_tokens, skip_special_tokens=hide_special_tokens)[0] logger.info( "Result: \n" f"----Prompt: `{prompt_list}`\n" f"----Tokens ids - prompt + generation: `{generated_tokens[0].tolist()}`\n" f"----Tokens converted - prompt + generation: `{tokens}`\n" f"----String decoded with skipped special tokens - prompt + generation: `{decoded_skip_special_tokens}`\n" f"----Total length - prompt + generation `{len(generated_tokens[0].tolist())}`\n" f"----Token ids - generation: `{actual_generated_tokens[0].tolist()}`\n" f"----Tokens converted - generation: `{tokenizer.convert_ids_to_tokens(actual_generated_tokens[0])}`\n" f"----String decoded with skipped special tokens - generation: `{generated_text}`\n" f"----Total length - generation: `{len(actual_generated_tokens[0].tolist())}`\n" f"----Generation mode: `{decoding_strategy}`\n" f"----Generation parameters: `{generation_args}`\n" ) return generated_text textbox = gr.Textbox( show_label=False, value="What color are the cat's eyes?", placeholder=( "To input images, use the following syntax:" " `textexttext`" ), visible=True, container=False, ) with gr.Blocks(title="IDEFICS", theme=gr.themes.Base()) as demo: # state = gr.State() with gr.Row(): with gr.Column(scale=3): with gr.Row(elem_id="model_selector_row"): model_selector = gr.Dropdown( choices=MODELS, value=MODELS[0] if len(MODELS) > 0 else "", interactive=True, show_label=False, container=False, ) processor, tokenizer, model = load_processor_tokenizer_model(model_selector.value) imagebox = gr.Image(type="pil", label="Image input") with gr.Accordion("Generation parameters", open=False, visible=True) as parameter_row: max_new_tokens = gr.Slider( minimum=0, maximum=2048, value=512, step=1, interactive=True, label="Maximum number of new tokens to generate", ) min_length = gr.Slider( minimum=0, maximum=50, value=0, step=1, interactive=True, label="Minimum number of new tokens to generate", ) repetition_penalty = gr.Slider( minimum=0.0, maximum=5.0, value=1.0, step=0.1, interactive=True, label="Repetition penalty", info="1.0 means no penalty", ) no_repeat_ngram_size = gr.Slider( minimum=0, maximum=10, value=0, step=1, interactive=True, label="N-gram repetition threshold", info="If set to int > 0, all ngrams of that size can only occur once.", ) decoding_strategy = gr.Radio( [ "greedy", # "beam_search", # "beam_sampling", # "sampling_top_k", "sampling_top_p", ], value="greedy", label="Decoding strategy", interactive=True, ) temperature = gr.Slider( minimum=0.0, maximum=5.0, value=1.0, step=0.1, interactive=True, label="Sampling temperature", ) decoding_strategy.change( fn=lambda selection: gr.Slider.update( visible=( selection in ["contrastive_sampling", "beam_sampling", "sampling_top_p", "sampling_top_k"] ) ), inputs=decoding_strategy, outputs=temperature, ) num_beams = gr.Slider( minimum=0, maximum=20, value=3.0, step=1.0, interactive=True, visible=False, label="Number of beams", info="Only used if `decoding_strategy` is `beam_search` or `beam_sampling`.", ) decoding_strategy.change( fn=lambda selection: gr.Slider.update(visible=(selection in ["beam_search", "beam_sampling"])), inputs=decoding_strategy, outputs=num_beams, ) top_p = gr.Slider( minimum=0.0, maximum=1.0, value=0.95, step=0.01, interactive=True, visible=False, label="Top P", info=( "If set to float < 1, only the smallest set of most probable tokens with probabilities that" " add up to top_p or higher are kept for generation." ), ) decoding_strategy.change( fn=lambda selection: gr.Slider.update(visible=(selection in ["sampling_top_p"])), inputs=decoding_strategy, outputs=top_p, ) top_k = gr.Slider( minimum=0, maximum=500, value=50, step=1, interactive=True, visible=False, label="Top K", info="The number of highest probability vocabulary tokens to keep for top-k-filtering.", ) decoding_strategy.change( fn=lambda selection: gr.Slider.update(visible=(selection in ["sampling_top_k"])), inputs=decoding_strategy, outputs=top_k, ) length_penalty = gr.Slider( minimum=-1000.0, maximum=1000.0, value=1.0, step=0.1, interactive=True, visible=False, label="Length penalty", info=( "length_penalty > 0.0 promotes longer sequences, while length_penalty < 0.0 encourages shorter" " sequences. Only used if `decoding_strategy` is `beam_search` or `beam_sampling`." ), ) decoding_strategy.change( fn=lambda selection: gr.Slider.update(visible=(selection in ["beam_search", "beam_sampling"])), inputs=decoding_strategy, outputs=length_penalty, ) penalty_alpha = gr.Slider( minimum=0.0, maximum=5.0, value=0.95, step=0.05, interactive=True, visible=False, label="Penalty alpha", info="Only used if `decoding_strategy` is `contrastive_sampling`.", ) decoding_strategy.change( fn=lambda selection: gr.Slider.update(visible=(selection in ["contrastive_sampling"])), inputs=decoding_strategy, outputs=penalty_alpha, ) with gr.Column(scale=6): def prefetch_images_in_history(user_prompt_str): """ Pre-fetch the images that are passed in the chatbot default history. """ return user_prompt_list_to_markdown(handle_manual_images_in_user_prompt(user_prompt_str)) chatbot = gr.Chatbot( elem_id="chatbot", label="Idefics Chatbot", visible=True, height=550, value=[ [ ( prefetch_images_in_history( "What's unusual about this" " image?" ) ), ( "The unusual aspect of this image is that there is a cat lying on a bed with an orange on" " its stomach. The cat is white and fluffy, and it appears to be sleeping or resting" " comfortably with the orange on its belly. This is an uncommon sight, as cats typically" " do not interact with fruits in such a manner." ), ], [ ( prefetch_images_in_history( "What" " about this cat? Why do people find it funny?" ) ), ( "The cat in the image is a gray and white long-haired cat with a surprised expression on" " its face. The cat's eyes are wide open, and it appears to be looking directly at the" " camera. The cat's mouth is also open, which adds to its surprised and funny appearance." " The combination of the cat's surprised expression and its wet fur, possibly from being" " in the shower or bath, makes it an amusing and entertaining sight for many people." ), ], ], ) with gr.Row(): with gr.Column(scale=8): textbox.render() with gr.Column(scale=1, min_width=60): submit_btn = gr.Button(value="Submit", visible=True) with gr.Column(scale=1, min_width=20): clear_btn = gr.ClearButton([textbox, chatbot]) cur_dir = os.path.dirname(os.path.abspath(__file__)) def model_inference( user_prompt_str, chat_history, image, decoding_strategy, num_beams, temperature, no_repeat_ngram_size, max_new_tokens, min_length, repetition_penalty, length_penalty, top_k, top_p, penalty_alpha, ): # global processor, model, tokenizer force_words = "" hide_special_tokens = False 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, ) generated_text = model_generation( prompt_list=formated_prompt_list, processor=processor, tokenizer=tokenizer, model=model, temperature=temperature, no_repeat_ngram_size=no_repeat_ngram_size, max_new_tokens=max_new_tokens, min_length=min_length, ban_tokens=BAN_TOKENS, eos_tokens=EOS_TOKENS, force_words=force_words, repetition_penalty=repetition_penalty, hide_special_tokens=hide_special_tokens, decoding_strategy=decoding_strategy, num_beams=num_beams, length_penalty=length_penalty, top_k=top_k, top_p=top_p, penalty_alpha=penalty_alpha, ) if image is None: # Case where there is no image OR the image is passed as `` chat_history.append( (user_prompt_list_to_markdown(user_prompt_list), generated_text.strip("")) ) 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"{user_prompt_list_to_markdown([image] + user_prompt_list)}", generated_text.strip(""), ) ) return "", None, chat_history def process_example(message, image): clear_msg, image_value, chat = model_inference( user_prompt_str=message, chat_history=[], image=image, decoding_strategy="greedy", num_beams=3, temperature=1.0, no_repeat_ngram_size=0, max_new_tokens=512, min_length=16, repetition_penalty=1.0, length_penalty=1.0, top_k=50, top_p=0.95, penalty_alpha=0.95, ) return clear_msg, image_value, chat textbox.submit( fn=model_inference, inputs=[ textbox, chatbot, imagebox, decoding_strategy, num_beams, temperature, no_repeat_ngram_size, max_new_tokens, min_length, repetition_penalty, length_penalty, top_k, top_p, penalty_alpha, ], outputs=[textbox, imagebox, chatbot], ) submit_btn.click( fn=model_inference, inputs=[ textbox, chatbot, imagebox, decoding_strategy, num_beams, temperature, no_repeat_ngram_size, max_new_tokens, min_length, repetition_penalty, length_penalty, top_k, top_p, penalty_alpha, ], outputs=[ textbox, imagebox, chatbot, ], ) examples_path = os.path.dirname(__file__) gr.Examples( examples=[ ["What are the armed baguettes guarding?", f"{examples_path}/example_images/baguettes_guarding_paris.png"], ["Can you describe the image?", f"{examples_path}/example_images/bear_costume.png"], [ "What is this object and do you think it is horrifying?", f"{examples_path}/example_images/can_horror.png", ], [ "Can you tell me a very short story based on this image?", f"{examples_path}/example_images/chicken_on_money.png", ], ["Can you describe this image in details please?", f"{examples_path}/example_images/dragons_playing.png"], ["Can you tell me what is unique about this image?", f"{examples_path}/example_images/ironman_cap.png"], [ "Can you write an advertisement for Coca-Cola based on this image?", f"{examples_path}/example_images/polar_bear_coke.png", ], [ "What is the rabbit doing in this image? Do you think this image is real?", f"{examples_path}/example_images/rabbit_force.png", ], ["What is happening in this image and why is it unusual?", f"{examples_path}/example_images/ramen.png"], [ "What I should look most forward to when I visit this place?", f"{examples_path}/example_images/tree_fortress.jpg", ], [ ( "What" " do you think the dog is doing and is it unusual?" ), None, ], ], inputs=[textbox, imagebox], outputs=[textbox, imagebox, chatbot], fn=process_example, cache_examples=True, examples_per_page=5, label="Click on any example below to get started", ) demo.queue() demo.launch()