app.py

import os
import subprocess
from typing import Union
from huggingface_hub import whoami, HfApi
from fastapi import FastAPI
from starlette.middleware.sessions import SessionMiddleware
import sys

# ai-toolkit이 없으면 설치
if not os.path.exists("ai-toolkit"):
    subprocess.run("git clone https://github.com/ostris/ai-toolkit.git", shell=True)
    subprocess.run("cd ai-toolkit && git submodule update --init --recursive", shell=True)

# ai-toolkit 경로 추가
toolkit_path = os.path.join(os.getcwd(), "ai-toolkit")
sys.path.append(toolkit_path)

# 필요한 패키지 설치
subprocess.run("pip install -r ai-toolkit/requirements.txt", shell=True)


is_spaces = True if os.environ.get("SPACE_ID") else False
    
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
import sys

from dotenv import load_dotenv

load_dotenv()

# Add the current working directory to the Python path
sys.path.insert(0, os.getcwd())

import gradio as gr
from PIL import Image
import torch
import uuid
import shutil
import json
import yaml
from slugify import slugify
from transformers import AutoProcessor, AutoModelForCausalLM

# Gradio app 설정
app = FastAPI()
app.add_middleware(SessionMiddleware, secret_key="your-secret-key")

if not is_spaces:
    sys.path.insert(0, "ai-toolkit")
    from toolkit.job import get_job
    gr.OAuthProfile = None
    gr.OAuthToken = None

    
MAX_IMAGES = 150


# Hugging Face 토큰 설정
HF_TOKEN = os.getenv("HF_TOKEN")
if not HF_TOKEN:
    raise ValueError("HF_TOKEN environment variable is not set")


if is_spaces:
    subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
    import spaces
    
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
os.environ["HUGGING_FACE_HUB_TOKEN"] = HF_TOKEN

# HF API 초기화
api = HfApi(token=HF_TOKEN)

def load_captioning(uploaded_files, concept_sentence):
    uploaded_images = [file for file in uploaded_files if not file.endswith('.txt')]
    txt_files = [file for file in uploaded_files if file.endswith('.txt')]
    txt_files_dict = {os.path.splitext(os.path.basename(txt_file))[0]: txt_file for txt_file in txt_files}
    updates = []
    if len(uploaded_images) <= 1:
        raise gr.Error(
            "Please upload at least 2 images to train your model (the ideal number with default settings is between 4-30)"
        )
    elif len(uploaded_images) > MAX_IMAGES:
        raise gr.Error(f"For now, only {MAX_IMAGES} or less images are allowed for training")
    # Update for the captioning_area
    # for _ in range(3):
    updates.append(gr.update(visible=True))
    # Update visibility and image for each captioning row and image
    for i in range(1, MAX_IMAGES + 1):
        # Determine if the current row and image should be visible
        visible = i <= len(uploaded_images)
        
        # Update visibility of the captioning row
        updates.append(gr.update(visible=visible))

        # Update for image component - display image if available, otherwise hide
        image_value = uploaded_images[i - 1] if visible else None
        updates.append(gr.update(value=image_value, visible=visible))
        
        corresponding_caption = False
        if(image_value):
            base_name = os.path.splitext(os.path.basename(image_value))[0]
            print(base_name)
            print(image_value)
            if base_name in txt_files_dict:
                print("entrou")
                with open(txt_files_dict[base_name], 'r') as file:
                    corresponding_caption = file.read()
                    
        # Update value of captioning area
        text_value = corresponding_caption if visible and corresponding_caption else "[trigger]" if visible and concept_sentence else None
        updates.append(gr.update(value=text_value, visible=visible))

    # Update for the sample caption area
    updates.append(gr.update(visible=True))
    # Update prompt samples
    updates.append(gr.update(placeholder=f'A portrait of person in a bustling cafe {concept_sentence}', value=f'A person in a bustling cafe {concept_sentence}'))
    updates.append(gr.update(placeholder=f"A mountainous landscape in the style of {concept_sentence}"))
    updates.append(gr.update(placeholder=f"A {concept_sentence} in a mall"))
    return updates

def hide_captioning():
    return gr.update(visible=False), gr.update(visible=False), gr.update(visible=False), gr.update(visible=False) 

def create_dataset(*inputs):
    print("Creating dataset")
    images = inputs[0]
    destination_folder = str(f"datasets/{uuid.uuid4()}")
    if not os.path.exists(destination_folder):
        os.makedirs(destination_folder)

    jsonl_file_path = os.path.join(destination_folder, "metadata.jsonl")
    with open(jsonl_file_path, "a") as jsonl_file:
        for index, image in enumerate(images):
            new_image_path = shutil.copy(image, destination_folder)

            original_caption = inputs[index + 1]
            file_name = os.path.basename(new_image_path)

            data = {"file_name": file_name, "prompt": original_caption}

            jsonl_file.write(json.dumps(data) + "\n")

    return destination_folder


def run_captioning(images, concept_sentence, *captions):
    device = "cuda" if torch.cuda.is_available() else "cpu"
    torch_dtype = torch.float16
    model = AutoModelForCausalLM.from_pretrained(
        "microsoft/Florence-2-large", torch_dtype=torch_dtype, trust_remote_code=True
    ).to(device)
    processor = AutoProcessor.from_pretrained("microsoft/Florence-2-large", trust_remote_code=True)

    captions = list(captions)
    for i, image_path in enumerate(images):
        print(captions[i])
        if isinstance(image_path, str):  # If image is a file path
            image = Image.open(image_path).convert("RGB")

        prompt = "<DETAILED_CAPTION>"
        inputs = processor(text=prompt, images=image, return_tensors="pt").to(device, torch_dtype)

        generated_ids = model.generate(
            input_ids=inputs["input_ids"], pixel_values=inputs["pixel_values"], max_new_tokens=1024, num_beams=3
        )

        generated_text = processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
        parsed_answer = processor.post_process_generation(
            generated_text, task=prompt, image_size=(image.width, image.height)
        )
        caption_text = parsed_answer["<DETAILED_CAPTION>"].replace("The image shows ", "")
        if concept_sentence:
            caption_text = f"{caption_text} [trigger]"
        captions[i] = caption_text

        yield captions
    model.to("cpu")
    del model
    del processor

if is_spaces:
    run_captioning = spaces.GPU()(run_captioning)

def recursive_update(d, u):
    for k, v in u.items():
        if isinstance(v, dict) and v:
            d[k] = recursive_update(d.get(k, {}), v)
        else:
            d[k] = v
    return d

def start_training(
    lora_name,
    concept_sentence,
    which_model,
    steps,
    lr,
    rank,
    dataset_folder,
    sample_1,
    sample_2,
    sample_3,
    use_more_advanced_options,
    more_advanced_options,
):
    if not lora_name:
        raise gr.Error("You forgot to insert your LoRA name! This name has to be unique.")
    
    try:
        username = whoami()["name"]
    except:
        raise gr.Error("Failed to get username. Please check your HF_TOKEN.")
        
    print("Started training")
    slugged_lora_name = slugify(lora_name)

    # Load the default config
    with open("train_lora_flux_24gb.yaml", "r") as f:
        config = yaml.safe_load(f)

    # dev 모델 설정
    config["config"]["name"] = slugged_lora_name
    config["config"]["process"][0]["model"]["name_or_path"] = "black-forest-labs/FLUX.1-dev"
    config["config"]["process"][0]["model"]["assistant_lora_path"] = None  # adapter 없이 설정
    config["config"]["process"][0]["model"]["low_vram"] = False
    config["config"]["process"][0]["train"]["skip_first_sample"] = True
    config["config"]["process"][0]["train"]["steps"] = int(steps)
    config["config"]["process"][0]["train"]["lr"] = float(lr)
    config["config"]["process"][0]["network"]["linear"] = int(rank)
    config["config"]["process"][0]["network"]["linear_alpha"] = int(rank)
    config["config"]["process"][0]["datasets"][0]["folder_path"] = dataset_folder
    config["config"]["process"][0]["save"]["push_to_hub"] = True
    config["config"]["process"][0]["save"]["hf_repo_id"] = f"{username}/{slugged_lora_name}"
    config["config"]["process"][0]["save"]["hf_private"] = True
    config["config"]["process"][0]["save"]["hf_token"] = HF_TOKEN
    config["config"]["process"][0]["sample"]["sample_steps"] = 28

    if concept_sentence:
        config["config"]["process"][0]["trigger_word"] = concept_sentence
    
    if sample_1 or sample_2 or sample_3:
        config["config"]["process"][0]["train"]["disable_sampling"] = False
        config["config"]["process"][0]["sample"]["sample_every"] = steps
        config["config"]["process"][0]["sample"]["prompts"] = []
        if sample_1:
            config["config"]["process"][0]["sample"]["prompts"].append(sample_1)
        if sample_2:
            config["config"]["process"][0]["sample"]["prompts"].append(sample_2)
        if sample_3:
            config["config"]["process"][0]["sample"]["prompts"].append(sample_3)
    else:
        config["config"]["process"][0]["train"]["disable_sampling"] = True

    if(use_more_advanced_options):
        more_advanced_options_dict = yaml.safe_load(more_advanced_options)
        config["config"]["process"][0] = recursive_update(config["config"]["process"][0], more_advanced_options_dict)
        print(config)
    
    try:
        # Save the updated config
        random_config_name = str(uuid.uuid4())
        os.makedirs("tmp", exist_ok=True)
        config_path = f"tmp/{random_config_name}-{slugged_lora_name}.yaml"
        with open(config_path, "w") as f:
            yaml.dump(config, f)

        # 직접 로컬 GPU에서 학습 실행
        from toolkit.job import get_job
        job = get_job(config_path)
        job.run()
        job.cleanup()
    except Exception as e:
        raise gr.Error(f"Training failed: {str(e)}")

    return f"""# Training completed successfully! 
    ## Your model is available at: <a href='https://huggingface.co/{username}/{slugged_lora_name}'>{username}/{slugged_lora_name}</a>"""

    
def update_pricing(steps):
    try:
        seconds_per_iteration = 7.54
        total_seconds = (steps * seconds_per_iteration) + 240
        cost_per_second = 0.80/60/60
        cost = round(cost_per_second * total_seconds, 2)
        cost_preview = f'''To train this LoRA, a paid L4 GPU will be hooked under the hood during training and then removed once finished.
        ### Estimated to cost <b>< US$ {str(cost)}</b> for {round(int(total_seconds)/60, 2)} minutes with your current train settings <small>({int(steps)} iterations at {seconds_per_iteration}s/it)</small>'''
        return gr.update(visible=True), cost_preview, gr.update(visible=False), gr.update(visible=True)
    except:
        return gr.update(visible=False), "", gr.update(visible=False), gr.update(visible=True)
    


def swap_base_model(model):
    return gr.update(visible=True) if model == "[dev] (high quality model, non-commercial license)" else gr.update(visible=False)    

config_yaml = '''
device: cuda:0
model:
  is_flux: true
  quantize: true
network:
  linear: 16 #it will overcome the 'rank' parameter
  linear_alpha: 16 #you can have an alpha different than the ranking if you'd like
  type: lora
sample:
  guidance_scale: 3.5
  height: 1024
  neg: '' #doesn't work for FLUX
  sample_every: 1000
  sample_steps: 28
  sampler: flowmatch
  seed: 42
  walk_seed: true
  width: 1024
save:
  dtype: float16
  hf_private: true
  max_step_saves_to_keep: 4
  push_to_hub: true
  save_every: 10000
train:
  batch_size: 1
  dtype: bf16
  ema_config:
    ema_decay: 0.99
    use_ema: true
  gradient_accumulation_steps: 1
  gradient_checkpointing: true
  noise_scheduler: flowmatch 
  optimizer: adamw8bit #options: prodigy, dadaptation, adamw, adamw8bit, lion, lion8bit
  train_text_encoder: false #probably doesn't work for flux
  train_unet: true
'''



custom_theme = gr.themes.Base(
    primary_hue="indigo",
    secondary_hue="slate",
    neutral_hue="slate",
).set(
    # 기본 배경 및 보더
    background_fill_primary="#1a1a1a",
    background_fill_secondary="#2d2d2d",
    border_color_primary="#404040",
    
    # 버튼 스타일
    button_primary_background_fill="#4F46E5",
    button_primary_background_fill_dark="#4338CA",
    button_primary_background_fill_hover="#6366F1",
    button_primary_border_color="#4F46E5",
    button_primary_border_color_dark="#4338CA",
    button_primary_text_color="white",
    button_primary_text_color_dark="white",
    
    button_secondary_background_fill="#374151",
    button_secondary_background_fill_dark="#1F2937",
    button_secondary_background_fill_hover="#4B5563",
    button_secondary_text_color="white",
    button_secondary_text_color_dark="white",
    
    # 블록 스타일
    block_background_fill="#2d2d2d",
    block_background_fill_dark="#1F2937",
    block_label_background_fill="#4F46E5",
    block_label_background_fill_dark="#4338CA",
    block_label_text_color="white",
    block_label_text_color_dark="white",
    block_title_text_color="white",
    block_title_text_color_dark="white",
    
    # 입력 필드 스타일
    input_background_fill="#374151",
    input_background_fill_dark="#1F2937",
    input_border_color="#4B5563",
    input_border_color_dark="#374151",
    input_placeholder_color="#9CA3AF",
    input_placeholder_color_dark="#6B7280",
    
    # 그림자 효과
    shadow_spread="8px",
    shadow_inset="0px 2px 4px 0px rgba(0,0,0,0.1)",
    
    # 컨테이너 스타일
    panel_background_fill="#2d2d2d",
    panel_background_fill_dark="#1F2937",
    
    # 보더 스타일
    border_color_accent="#4F46E5",
    border_color_accent_dark="#4338CA"
)

css='''
/* 기본 스타일 */
h1 {
    font-size: 3em;
    text-align: center;
    margin-bottom: 0.5em;
    color: white !important;
}

h3 {
    margin-top: 0;
    font-size: 1.2em;
    color: white !important;
}

/* Markdown 텍스트 스타일 */
.markdown {
    color: white !important;
}

.markdown h1, 
.markdown h2, 
.markdown h3, 
.markdown h4, 
.markdown h5, 
.markdown h6,
.markdown p {
    color: white !important;
}

/* 컴포넌트 스타일 */
.container {
    max-width: 1200px;
    margin: 0 auto;
    padding: 20px;
}

/* 입력 필드 스타일 */
.input-group {
    background: var(--block-background-fill);
    padding: 15px;
    border-radius: 12px;
    margin-bottom: 20px;
    box-shadow: 0 2px 4px rgba(0,0,0,0.1);
}

/* 모든 입력 필드 텍스트 색상 */
input, textarea, .gradio-textbox input, .gradio-textbox textarea, .gradio-number input {
    color: white !important;
}

/* 라벨 텍스트 스타일 */
label, .label-text {
    color: white !important;
}

/* 라디오 버튼 텍스트 */
.gradio-radio label span {
    color: white !important;
}

/* 체크박스 텍스트 */
.gradio-checkbox label span {
    color: white !important;
}

/* 버튼 스타일 */
.button {
    height: 40px;
    border-radius: 8px;
    transition: all 0.3s ease;
    color: white !important;
}

.button:hover {
    transform: translateY(-2px);
    box-shadow: 0 4px 6px rgba(0,0,0,0.1);
}

/* 이미지 업로드 영역 */
.image-upload-area {
    border: 2px dashed var(--input-border-color);
    border-radius: 12px;
    padding: 20px;
    text-align: center;
    margin-bottom: 20px;
    color: white !important;
}

/* 캡션 영역 */
.caption-area {
    background: var(--block-background-fill);
    padding: 15px;
    border-radius: 12px;
    margin-top: 20px;
    color: white !important;
}

.caption-row {
    display: flex;
    align-items: center;
    margin-bottom: 10px;
    gap: 10px;
}

/* 고급 옵션 영역 */
.advanced-options {
    background: var(--block-background-fill);
    padding: 15px;
    border-radius: 12px;
    margin-top: 20px;
    color: white !important;
}

/* 진행 상태 표시 */
.progress-area {
    background: var(--block-background-fill);
    padding: 15px;
    border-radius: 12px;
    margin-top: 20px;
    text-align: center;
    color: white !important;
}

/* 플레이스홀더 텍스트 */
::placeholder {
    color: rgba(255, 255, 255, 0.5) !important;
}

/* 코드 에디터 텍스트 */
.gradio-code {
    color: white !important;
}

/* 아코디언 텍스트 */
.gradio-accordion .label-wrap {
    color: white !important;
}

/* 반응형 디자인 */
@media (max-width: 768px) {
    .caption-row {
        flex-direction: column;
    }
}

/* 스크롤바 스타일 */
::-webkit-scrollbar {
    width: 8px;
}

::-webkit-scrollbar-track {
    background: var(--background-fill-primary);
    border-radius: 4px;
}

::-webkit-scrollbar-thumb {
    background: var(--primary-500);
    border-radius: 4px;
}

::-webkit-scrollbar-thumb:hover {
    background: var(--primary-600);
}

/* 모든 텍스트 입력 요소 */
.gradio-container input[type="text"],
.gradio-container textarea,
.gradio-container .input-text,
.gradio-container .input-textarea {
    color: white !important;
}

/* 드롭다운 텍스트 */
select, option {
    color: white !important;
}

/* 버튼 텍스트 */
button {
    color: white !important;
}
'''

# Gradio 앱 수정
with gr.Blocks(theme=custom_theme, css=css) as demo:

    gr.Markdown(
    """# 🆔 Gini LoRA 학습
    ### 1)LoRA 이름 영어로 '입력' 2)트리거 단어 영어로 '입력'  3)기본 모델 '클릭' 4)이미지(최소 2장~최대 150장 미만) '업로드' 5)비전 인식 LLM 라벨링 '클릭' 6)START 클릭""",
    elem_classes=["markdown"]
    )    
    
    with gr.Tab("Train"):
        with gr.Column(elem_classes="container"):
            # LoRA 설정 그룹
            with gr.Group(elem_classes="input-group"):
                with gr.Row():
                    lora_name = gr.Textbox(
                        label="LoRA 이름",
                        info="고유한 이름이어야 합니다",
                        placeholder="예: Persian Miniature Painting style, Cat Toy"
                    )
                    concept_sentence = gr.Textbox(
                        label="트리거 단어/문장",
                        info="사용할 트리거 단어나 문장",
                        placeholder="p3rs0n이나 trtcrd같은 특이한 단어, 또는 'in the style of CNSTLL'같은 문장"
                    )
                
                model_warning = gr.Markdown(visible=False)
                which_model = gr.Radio(
                    ["고퀄리티 맞춤 학습 모델"],
                    label="기본 모델",
                    value="[dev] (high quality model)"
                )

            # 이미지 업로드 영역
            with gr.Group(visible=True, elem_classes="image-upload-area") as image_upload:

                with gr.Row():
                    images = gr.File(
                        file_types=["image", ".txt"],
                        label="Upload your images",
                        file_count="multiple",
                        interactive=True,
                        visible=True,
                        scale=1,
                    )
                    with gr.Column(scale=3, visible=False) as captioning_area:
                        with gr.Column():
                            gr.Markdown(
                                """# 이미지 라벨링
    <p style="margin-top:0"> 비전인식 LLM이 이미지를 인식하여 자동으로 라벨링(이미지 인식을 위한 필수 설명). [trigger] '트리거 워드'는 학습한 모델을 실행하는 고유 키값 /trigger word.</p>
    """, elem_classes="group_padding")
                            do_captioning = gr.Button("비전 인식 LLM 자동 라벨링")
                            output_components = [captioning_area]
                            caption_list = []
                            for i in range(1, MAX_IMAGES + 1):
                                locals()[f"captioning_row_{i}"] = gr.Row(visible=False)
                                with locals()[f"captioning_row_{i}"]:
                                    locals()[f"image_{i}"] = gr.Image(
                                        type="filepath",
                                        width=111,
                                        height=111,
                                        min_width=111,
                                        interactive=False,
                                        scale=2,
                                        show_label=False,
                                        show_share_button=False,
                                        show_download_button=False,
                                    )
                                    locals()[f"caption_{i}"] = gr.Textbox(
                                        label=f"Caption {i}", scale=15, interactive=True
                                    )

                                output_components.append(locals()[f"captioning_row_{i}"])
                                output_components.append(locals()[f"image_{i}"])
                                output_components.append(locals()[f"caption_{i}"])
                                caption_list.append(locals()[f"caption_{i}"])

            with gr.Accordion("Advanced options", open=False):
                steps = gr.Number(label="Steps", value=1000, minimum=1, maximum=10000, step=1)
                lr = gr.Number(label="Learning Rate", value=4e-4, minimum=1e-6, maximum=1e-3, step=1e-6)
                rank = gr.Number(label="LoRA Rank", value=16, minimum=4, maximum=128, step=4)
                with gr.Accordion("Even more advanced options", open=False):
                    if(is_spaces):
                        gr.Markdown("Attention: changing this parameters may make your training fail or go out-of-memory if training on Spaces. Only change settings here it if you know what you are doing. Beware that training is done in an L4 GPU with 24GB of RAM")
                    use_more_advanced_options = gr.Checkbox(label="Use more advanced options", value=False)
                    more_advanced_options = gr.Code(config_yaml, language="yaml")

            with gr.Accordion("Sample prompts (optional)", visible=False) as sample:
                gr.Markdown(
                    "Include sample prompts to test out your trained model. Don't forget to include your trigger word/sentence (optional)"
                )
                sample_1 = gr.Textbox(label="Test prompt 1")
                sample_2 = gr.Textbox(label="Test prompt 2")
                sample_3 = gr.Textbox(label="Test prompt 3")
            with gr.Group(visible=False) as cost_preview:
                cost_preview_info = gr.Markdown(elem_id="cost_preview_info", elem_classes="group_padding")
                payment_update = gr.Button("I have set up a payment method", visible=False)
            output_components.append(sample)
            output_components.append(sample_1)
            output_components.append(sample_2)
            output_components.append(sample_3)
            start = gr.Button("START 클릭('약 25~30분 후 학습이 종료되고 완료 메시지가 출력됩니다.)'", visible=False)
        progress_area = gr.Markdown("")



    dataset_folder = gr.State()

    images.upload(
        load_captioning,
        inputs=[images, concept_sentence],
        outputs=output_components
    ).then(
        update_pricing,
        inputs=[steps],
        outputs=[cost_preview, cost_preview_info, payment_update, start]
    )
    
    images.clear(
        hide_captioning,
        outputs=[captioning_area, cost_preview, sample, start]
    )
    
    images.delete(
        load_captioning,
        inputs=[images, concept_sentence],
        outputs=output_components
    ).then(
        update_pricing,
        inputs=[steps],
        outputs=[cost_preview, cost_preview_info, payment_update, start]
    )
    
    gr.on(
        triggers=[steps.change],
        fn=update_pricing,
        inputs=[steps],
        outputs=[cost_preview, cost_preview_info, payment_update, start]
    )

    start.click(fn=create_dataset, inputs=[images] + caption_list, outputs=dataset_folder).then(
        fn=start_training,
        inputs=[
            lora_name,
            concept_sentence,
            which_model,
            steps,
            lr,
            rank,
            dataset_folder,
            sample_1,
            sample_2,
            sample_3,
            use_more_advanced_options,
            more_advanced_options
        ],
        outputs=progress_area,
    )

    do_captioning.click(fn=run_captioning, inputs=[images, concept_sentence] + caption_list, outputs=caption_list)


if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860, auth=("gini", "pick"), show_error=True)