app_dialogue.py

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, set_seed
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:<fake_token_around_image><image:https://i1.pickpik.com/photos/515/906/355/kittens-cats-pet-cute-preview.jpg><fake_token_around_image>Describe this image.<end_of_utterance>""",
    """\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.<end_of_utterance>""",
    (
        "\nUser:How about this"
        " image?<fake_token_around_image><image:https://cdn.pixabay.com/photo/2017/09/25/13/12/puppy-2785074_1280.jpg><fake_token_around_image>Can"
        " you describe it too?<end_of_utterance>"
    ),
    """\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.<end_of_utterance>""",
    "\nUser: What kind of breed is it?<end_of_utterance>",
    """\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.<end_of_utterance>""",
    "\nUser: What can you tell me about this breed of dogs?<end_of_utterance>",
    """\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.<end_of_utterance>""",
    "\nUser: ghjkhjabnufs<end_of_utterance>",
    """\nAssistant: That doesn’t seem to be a word. Could you ask me another way?<end_of_utterance>""",
    "\nUser: Do you like Cavalier King Charles Spaniel?<end_of_utterance>\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.<end_of_utterance>""",
    (
        "\nUser: How many dogs do you see in this"
        " image?<fake_token_around_image><image:https://i.dailymail.co.uk/i/pix/2011/07/01/article-2010308-0CD22A8300000578-496_634x414.jpg><fake_token_around_image><end_of_utterance>"
    ),
    """\nAssistant: There is no dogs in this image. The picture shows a tennis player jumping to volley the ball.<end_of_utterance>""",
]

BAN_TOKENS = "<image>;<fake_token_around_image>"
EOS_TOKENS = "</s>;<end_of_utterance>;User"


TOKENIZER_FAST = True
MAX_SEQ_LEN = 2048

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger()

SEED = 38
set_seed(38)

import requests 
from io import BytesIO
import uuid

from PIL import Image

from PIL import Image
import tempfile


def convert_to_rgb_pil(image):
    print(f"***** convert_to_rgb_pil ******")
    print(f"params: image is - {image}")
    #if image.mode == "RGB":
    #    return image
    # Save the converted image to a temporary file
    #temp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".jpg")
    #temp_file_path = temp_file.name
    filename = f"{uuid.uuid4()}.jpg"
    local_path = f"{filename}"
    
    if image.mode != "RGB":
        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")
        alpha_composite.save(local_path)
    else:
        image.save(local_path)
        
    #temp_file.close()
    print(f"# Return the path to the saved image as - {local_path}")
    return temp_file_path  # Return the path to the saved image
    

def convert_to_rgb(filepath_or_pilimg):
    # `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
    print(f"***** convert_to_rgb ******")
    print(f"params: image is - {filepath_or_pilimg}")
    
    if isinstance(filepath_or_pilimg, PIL.Image.Image):
        return convert_to_rgb_pil(filepath_or_pilimg)
        
    with Image.open(filepath_or_pilimg) as image:
        # Check if the image is already in the RGB format
        if image.mode == "RGB":
            return filepath_or_pilimg  # If already in RGB, return the original path

        # Convert image to RGBA
        image_rgba = image.convert("RGBA")
        
        # Create a white background image of the same size
        background = Image.new("RGBA", image_rgba.size, (255, 255, 255))
        
        # Composite the original image over the white background
        alpha_composite = Image.alpha_composite(background, image_rgba)
        
        # Convert the composited image to RGB format
        alpha_composite = alpha_composite.convert("RGB")

        # Save the converted image to a temporary file
        filename = f"{uuid.uuid4()}.jpg"
        local_path = f"{filename}"
        #temp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".jpg")
        #temp_file_path = temp_file.name
        alpha_composite.save(local_path)
        #temp_file.close()

        print(f"# Return the path to the saved image as - {local_path}")
        return temp_file_path  # Return the path to the saved image

def pil_to_markdown_im(image):
    """
    Convert a PIL image into markdown filled with the base64 string representation.
    """
    print(f"***** pil_to_markdown_im ******")
    print(f"params: image is - {image}")
    #if isinstance(image, PIL.Image.Image):
    #img_b64_str = pil_to_base64(image)
    #img_str = f'<img src="data:image/png;base64,{img_b64_str}" />'
    #if path_or_url.startswith(("http://", "https://")):
    #response = requests.get(image)
    #image = Image.open(BytesIO(response.content))
    # Generate a unique filename using UUID
    filename = f"{uuid.uuid4()}.jpg"
    local_path = f"{filename}" 
    image.save(local_path)
    img_str =  f"![](/file={local_path})"
    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'<img src="data:image/png;base64,([^"]+)" />'
    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'<img src="data:image/png;base64,([^"]+)" />'
    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'<img src="data:image/png;base64,{part.strip()}" />'
            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 <fake_token_around_image>) should be their own elements.
    From:
    ```
    [
        "bonjour<fake_token_around_image><image:IMG_URL><fake_token_around_image>hello",
        PIL.Image.Image,
        "Aurevoir",
    ]
    ```
    to:
    ```
    [
        "bonjour",
        IMG_URL,
        "hello",
        PIL.Image.Image,
        "Aurevoir",
    ]
    ```
    """
    print(f"******* isolate_images_urls *******")
    print(f"params: prompt_list is - {prompt_list}")
    
    linearized_list = []
    for prompt in prompt_list:
        print(f"inside FOR loop: prompt in prompt_list is - {prompt}")
        # Prompt can be either a string, or a PIL image
        if isinstance(prompt, PIL.Image.Image):
            print(f"inside first IF in FOR loop: prompt is of type PIL.Image.Image")
            linearized_list.append(prompt)
            print(f"linearized_list after append is - {linearized_list}")
        elif isinstance(prompt, str) and "/tmp/gradio/" in prompt: #isinstance(prompt, PIL.Image.Image):
            print(f"inside IF in FOR loop: prompt is a string and is a path for temporary file")
            linearized_list.append(prompt)
            print(f"linearized_list after append is - {linearized_list}")
        elif isinstance(prompt, str) and "/tmp/gradio/" not in prompt:
            print(f"inside ELIF in FOR loop: prompt is a string and is NOT a path for temporary file")
            if "<fake_token_around_image>" not in prompt:
                print(f"inside IF inside ELIF in FOR loop: '<fake_token_around_image>' is NOT in prompt")
                linearized_list.append(prompt)
                print(f"linearized_list after append is - {linearized_list}")
            else:
                print(f"inside ELSE inside ELIF in FOR loop: '<fake_token_around_image>' IS IN prompt")
                prompt_splitted = prompt.split("<fake_token_around_image>")
                print(f"prompt_splitted is - {prompt_splitted}")
                for ps in prompt_splitted:
                    print(f"Inside FOR loop inside FOR loop: ps in prompt_split is {ps}")
                    if ps == "":
                        continue
                    if ps.startswith("<image:"):
                        linearized_list.append(ps[7:-1])
                    else:
                        linearized_list.append(ps)
        else:
            raise TypeError(
                f"Unrecognized type for `prompt`. Got {type(type(prompt))}. Was expecting something in [`str`,"
                " `PIL.Image.Image`]"
            )
            print(f"linearized_list to be returned is - {linearized_list}")
    return linearized_list


# Chatbot handling utils
def handle_manual_images_in_user_prompt(user_prompt: str) -> List[Union[str, PIL.Image.Image]]:
    """
    Handle the case of textually manually inputted images (i.e. the `<fake_token_around_image><image:IMG_URL><fake_token_around_image>`) in the user prompt
    by fetching them and replacing the whole sub-sequence by a PIL image.
    """
    if "<fake_token_around_image>" 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.
    """
    print("********** user_prompt_list_to_markdown *********")
    print(f" param : user_prompt_list is - {user_prompt_list}")
    resulting_string = ""
    for elem in user_prompt_list:
        print(f"inside user_prompt_list_to_markdown, for loop on user_prompt_list")
        print(f"elem is - {elem} ")
        if isinstance(elem, str):
            if "/tmp/gradio/" not in elem:
                resulting_string += elem
                print(f"inside IF - when elem is string and is not temp image filepath. resulting_string is - {resulting_string}")
            elif "/tmp/gradio/" in elem:            
                resulting_string += f"![](/file={convert_to_rgb(elem)})"
                print(f"inside IF - when elem is string and is a temp image filepath. resulting_string is - {resulting_string}")       
        #elif isinstance(elem, str) and "/tmp/gradio/" in elem:
        #    resulting_string += f"![](/file={convert_to_rgb(elem)})"  #f"![](/file={image})" 
        #    print(f"inside first ELIF - when elem is string and is the temp image filepath. resulting_string is - {resulting_string}")
        elif isinstance(elem, PIL.Image.Image): #or "/tmp/gradio/" in elem:  #and "/tmp/gradio/" in elem:
            resulting_string += f"![](/file={convert_to_rgb(elem)})" #pil_to_markdown_im(convert_to_rgb(elem)) <---------------
            print(f"inside the ELIF - when elem is an instance of  PIL.Image.Image. The resulting_string after convert_to_rgb() operation is - {resulting_string}")
        else:
            raise ValueError(
                "Unknown type for `user_prompt_list`. Expected an element of type `str` or `PIL.Image.Image` and got"
                f" `{type(elem)}`"
            )
    print(f" final resulting_string that will be returned is - {resulting_string}")
    return resulting_string


def remove_spaces_around_token(text):
    pattern = r'\s*(<fake_token_around_image>)\s*'
    replacement = r'\1'
    result = re.sub(pattern, replacement, text)
    return result


# 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.
    """
    print(f"*********format_user_prompt_with_im_history_and_system_conditioning*********")
    print(f"format_user_prompt_with_im_history_and_system_conditioning -- param current_user_prompt_str is - {current_user_prompt_str} ")
    print(f"format_user_prompt_with_im_history_and_system_conditioning -- param current_image is - {current_image} ")
    print(f"format_user_prompt_with_im_history_and_system_conditioning -- param history is - {history} ")
    
    resulting_list = copy.deepcopy(SYSTEM_PROMPT)

    # Format history
    for turn in history:
        print(f"inside for loop, turn is - {turn}")
        user_utterance, assistant_utterance = turn
        print("calling split_str_on_im_markdown from inside for loop inside format_user_prompt_with_im_history_and_system_conditioning")
        splitted_user_utterance = split_str_on_im_markdown(user_utterance)
        print(f"splitted_user_utterance from split_str_on_im_markdown is - {splitted_user_utterance} ")
        splitted_user_utterance = [
            im_markdown_to_pil(s) if s.startswith('<img src="data:image/png;base64,') else s
            for s in splitted_user_utterance
            if s != ""
        ]
        print(f"splitted_user_utterance after im_markdown_to_pil() is - {splitted_user_utterance} ")
        
        if isinstance(splitted_user_utterance[0], str):
            resulting_list.append("\nUser: ")
        else:
            resulting_list.append("\nUser:")
        print(f"resulting_list after if..else block is - {resulting_list}")
        resulting_list.extend(splitted_user_utterance)
        print(f"resulting_list after extend is - {resulting_list}")
        resulting_list.append(f"<end_of_utterance>\nAssistant: {assistant_utterance}")
        print(f"resulting_list after append is - {resulting_list}")
        

    # Format current input
    current_user_prompt_str = remove_spaces_around_token(current_user_prompt_str)
    print(f"current_user_prompt_str is - {current_user_prompt_str}")
    
    if current_image is None:
        print("inside IF : current_image is NONE")
        if "<img src=data:image/png;base64" in current_user_prompt_str:
            raise ValueError("The UI does not support inputing via the text box an image in base64.")
        current_user_prompt_list = handle_manual_images_in_user_prompt(current_user_prompt_str)
        print(f"current_user_prompt_list (or [user_prompt]/resulting_user_prompt((most likely this one)) from handle_manual_images_in_user_prompt ) is - {current_user_prompt_list}")
        resulting_list.append("\nUser: ")
        print(f"resulting_list with append user - {resulting_list}")
        resulting_list.extend(current_user_prompt_list)
        print(f"resulting_list after extend with current_user_prompt_list is - {resulting_list}")
        resulting_list.append("<end_of_utterance>\nAssistant:")
        print(f"resulting_list after append with end_of_utteranceAssistant is - {resulting_list}")
        return resulting_list, current_user_prompt_list
    else:
        print("inside ELSE : current_image is not NONE")
        # Choosing to put the image first when the image is inputted through the UI, but this is an arbiratrary choice.
        resulting_list.extend(["\nUser:", Image.open(current_image), f"{current_user_prompt_str}<end_of_utterance>\nAssistant:"]) #current_image
        print(f"final resulting_list passed on to calling function is - {resulting_list}")
        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.convert_tokens_to_ids(eos_token)
                if tokenized_eos_token == 0: # <unk> with our llama tokenizer
                    raise ValueError(f"Unknown tokens specified for exit condition.")
                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 = {
        "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",
        "Top P Sampling",
        "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["temperature"] = temperature
        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["temperature"] = temperature
        generation_args["do_sample"] = True
        generation_args["top_k"] = top_k
    elif decoding_strategy == "Top P Sampling":
        generation_args["temperature"] = temperature
        generation_args["do_sample"] = True
        generation_args["top_p"] = top_p
    elif decoding_strategy == "contrastive_sampling":
        generation_args["temperature"] = temperature
        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


dope_callback = gr.CSVLogger()
dope_hf_callback = gr.HuggingFaceDatasetSaver(
    hf_token=os.getenv("HF_AUTH_TOKEN"),
    dataset_name="HuggingFaceM4/gradio_dope_data_points",
    private=True,
)
problematic_callback = gr.CSVLogger()

textbox = gr.Textbox(
    show_label=False,
    value="Describe the battle against the fierce dragons.",
    visible=True,
    container=False,
    label="Text input",
)
with gr.Blocks(title="IDEFICS-Chat", theme=gr.themes.Base()) as demo:
    gr.Markdown(
        """
        # IDEFICS
        This is a demo for [IDEFICS](https://huggingface.co/HuggingFaceM4/idefics-80b), a open-access large visual lanugage model built built solely on publicly available data and models.
        <br>Like GPT-4, the multimodal model accepts arbitrary sequences of image and text inputs and produces text outputs.
        <br>IDEFICS (which stans for **I**mage-aware **D**ecoder **E**nhanced à la **F**lamingo with **I**nterleaved **C**ross-attention**S**) is an open-access reproduction of [Flamingo](https://huggingface.co/papers/2204.14198), a closed-source visual language model developed by Deepmind.

        The [model cards](https://huggingface.co/HuggingFaceM4/idefics-80b) and [dataset card](https://huggingface.co/datasets/HuggingFaceM4/OBELISC) provide plenty of information about the model and training data.
        <br>We provide an [interactive visualization](https://atlas.nomic.ai/map/f2fba2aa-3647-4f49-a0f3-9347daeee499/ee4a84bd-f125-4bcc-a683-1b4e231cb10f) (TODO: change to official link when have it) that allows exploring the content of the training data.
        <br>You can also [read more about](https://github.com/huggingface/m4-logs/blob/master/memos/README.md) some of the technical challenges encountered during training IDEFICS.
        """
    )

    with gr.Row():
        with gr.Column(scale=3):
            with gr.Row(elem_id="model_selector_row"):
                model_selector = gr.Dropdown(
                    choices=MODELS,
                    value="HuggingFaceM4/idefics-9b-instruct",
                    interactive=True,
                    show_label=False,
                    container=False,
                    label="Model"
                )
            processor, tokenizer, model = load_processor_tokenizer_model(model_selector.value)

            imagebox = gr.Image(type="filepath", label="Image input")

            with gr.Accordion("Advanced 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",
                        "Top P Sampling",
                    ],
                    value="Top P Sampling",
                    label="Decoding strategy",
                    interactive=True,
                )
                temperature = gr.Slider(
                    minimum=0.0,
                    maximum=5.0,
                    value=1.2,
                    step=0.1,
                    interactive=True,
                    label="Sampling temperature",
                )
                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,
                )
                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=0.8,
                    step=0.01,
                    interactive=True,
                    visible=True,
                    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 ["Top P Sampling"])),
                    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="ChatIDEFICS",
                visible=True,
                height=550,
                value=[
                    [
                        (
                            prefetch_images_in_history(
                                """Write a funny story including the following characters.
There should be drama and most importantly romance.
Character 1:<fake_token_around_image><image:https://m.media-amazon.com/images/M/MV5BODg1OTczMWEtNTU3MS00OTUzLThjODEtNDg1MWQwZmExYmFlXkEyXkFqcGdeQWFybm8@._V1_QL75_UX500_CR0,0,500,281_.jpg><fake_token_around_image>
Character 2:<fake_token_around_image><image:https://static.wikia.nocookie.net/dreamworks/images/0/09/Shrek_dreamworks.png/revision/latest?cb=20170921225439><fake_token_around_image>
Character 3:<fake_token_around_image><image:https://upload.wikimedia.org/wikipedia/en/e/ef/Marsupilami_1.jpg><fake_token_around_image>
The plot should take place in the world pictured here:<fake_token_around_image><image:https://www.astronomy.com/wp-content/uploads/sites/2/2021/09/ON1.jpg><fake_token_around_image>"""
                            )
                        ),
                        (
                            """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.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])
                with gr.Column(scale=1, min_width=15):
                    dope_bttn = gr.Button("Dope🔥")
                with gr.Column(scale=1, min_width=15):
                    problematic_bttn = gr.Button("Problematic😬")

    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
        print("***********Model_inference*************")
        print(f"Inside Model_inference, user_prompt_str is - {user_prompt_str} ")
        print(f"Inside Model_inference, chat_history is - {chat_history} ")
        print(f"Inside Model_inference, image type is - {type(image)} ")
        print(f"Inside Model_inference, image is - {image} ")

        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,
        )

        print(f"formated_prompt_list (or resulting_list)  is {formated_prompt_list}")
        print(f"user_prompt_list (or [current_user_prompt_str]) is {user_prompt_list}")
        
        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 `<fake_token_around_image><image:IMAGE_URL><fake_token_around_image>`
            chat_history.append(
                (user_prompt_list_to_markdown(user_prompt_list), generated_text.strip("<end_of_utterance>"))
            )
            print(f"chat_history (IF image is None or is with fake token) is -{chat_history}")
        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("<end_of_utterance>"),
                )
            )
            int(f"chat_history (ELSE IF image is available) is -{chat_history}")
        return "", None, chat_history

    def process_example(message, image):
        print("********* process_example **********")
        clear_msg, image_value, chat = model_inference(
            user_prompt_str=message,
            chat_history=[],
            image=image,
            decoding_strategy="Greedy",
            num_beams=None,
            temperature=None,
            no_repeat_ngram_size=None,
            max_new_tokens=512,
            min_length=16,
            repetition_penalty=None,
            length_penalty=None,
            top_k=None,
            top_p=0.95,
            penalty_alpha=None,
        )
        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,
        ],
    )

    # Using Flagging for saving dope and problematic examples
    # Dope examples flagging
    dope_hf_callback.setup(
        [
            model_selector,
            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,
        ],
        "gradio_dope_data_points"
    )
    dope_bttn.click(
        lambda *args: dope_hf_callback.flag(args),
        [
            model_selector,
            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,
        ],
        None,
        preprocess=False
    )
    # Problematic examples flagging
    problematic_callback.setup(
        [
            model_selector,
            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,
        ],
        "gradio_problematic_data_points"
    )
    problematic_bttn.click(
        lambda *args: problematic_callback.flag(args),
        [
            model_selector,
            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,
        ],
        None,
        preprocess=False
    )


    gr.Markdown(
        """## How to use?

        There are two ways to provide image inputs:
        - Using the image box on the left panel
        - Using the inline syntax: `text<fake_token_around_image><image:URL_IMAGE><fake_token_around_image>text`

        The second syntax allows inputting an arbitrary number of images."""
    )

    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 tell me a very short story based on this image?",
                 f"{examples_path}/example_images/chicken_on_money.png",
             ],
            # ["Can you describe the image?", f"{examples_path}/example_images/bear_costume.png"],
            # ["What is this animal and why is it unusual?", f"{examples_path}/example_images/blue_dog.png"],
            # [
            #     "What is this object and do you think it is horrifying?",
            #     f"{examples_path}/example_images/can_horror.png",
            # ],
            # ["What is this sketch for? How would you make an argument to prove this sketch was made by Picasso himself?", f"{examples_path}/example_images/cat_sketch.png"],
            # ["Which celebrity does this claymation figure look like?", f"{examples_path}/example_images/kanye.jpg"],
            # [
            #     "Which famous person does the person in the image look like? Could you craft an engaging narrative featuring this character from the image as the main protagonist?",
            #     f"{examples_path}/example_images/obama-harry-potter.jpg",
            # ],
            # [
            #     "Is there a celebrity look-alike in this image? What is happening to the person?",
            #     f"{examples_path}/example_images/ryan-reynolds-borg.jpg",
            # ],
            # ["Can you describe this image in details please?", f"{examples_path}/example_images/dragons_playing.png"],
            # ["What can you tell me about the cap in 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",
            # ],
            # ["Who is the person in the image and what is he doing?", f"{examples_path}/example_images/tom-cruise-astronaut-pegasus.jpg"],
            # [
            #     "What is happening in this image? Which famous personality does this person in center looks like?",
            #     f"{examples_path}/example_images/gandhi_selfie.jpg",
            # ],
            # [
            #     (
            #         "<fake_token_around_image><image:https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/stable-diffusion-xl-coreml/a_high_quality_photo_of_a_surfing_dog.7667.final_float16_original.jpg><fake_token_around_image>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(share=False)