scripts/run_training.py

"""
End-to-End Thumbnail VLM Training Pipeline

This script:
1. Downloads and prepares training data (PosterCraft/Poster100K)
2. Fine-tunes Janus-Pro-7B for thumbnail generation
3. Evaluates the model with sample generations
4. Pushes to HuggingFace Hub

Run with: python run_training.py
"""

import os
import sys
import json
import random
import logging
import io
import base64
from pathlib import Path

import numpy as np
import torch
import torch.nn.functional as F
from PIL import Image
from tqdm import tqdm

import trackio

logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
logger = logging.getLogger(__name__)

# ─── Configuration ──────────────────────────────────────────────────────────
MODEL_PATH = "deepseek-ai/Janus-Pro-7B"
HUB_MODEL_ID = "asats/thumbnail-vlm-janus-pro"
OUTPUT_DIR = "/app/results/janus_thumbnail"
DATA_DIR = "/app/thumbnail_data"
IMAGE_DIR = os.path.join(DATA_DIR, "images")
JSONL_PATH = os.path.join(DATA_DIR, "train.jsonl")

# Training hyperparameters (Janus-4o recipe)
EPOCHS = 3
BATCH_SIZE = 1  # per device (single GPU)
GRADIENT_ACCUMULATION = 16  # effective batch = 16
LR = 5e-6
IMAGE_SIZE = 384
IMAGE_TOKEN_NUM = 576
VQ_CODEBOOK_SIZE = 16384
CFG_MASK_PROB = 0.10
SAVE_EVERY = 200
LOG_EVERY = 5
SEED = 42

os.makedirs(IMAGE_DIR, exist_ok=True)
os.makedirs(OUTPUT_DIR, exist_ok=True)


# ─── Step 1: Prepare Training Data ──────────────────────────────────────────

def prepare_training_data(max_poster_samples=8000, max_synthetic=2000):
    """Download PosterCraft images and create JSONL training file."""
    logger.info("=" * 60)
    logger.info("Step 1: Preparing Training Data")
    logger.info("=" * 60)
    
    if os.path.exists(JSONL_PATH):
        with open(JSONL_PATH) as f:
            count = sum(1 for _ in f)
        logger.info(f"Training data already exists: {count} samples")
        return count
    
    from datasets import load_dataset
    
    entries = []
    
    # 1a. Download PosterCraft/Poster100K images
    logger.info(f"Downloading PosterCraft/Poster100K (up to {max_poster_samples} samples)...")
    
    T2I_TEMPLATES = [
        "Generate a professional thumbnail: {caption}",
        "Create an eye-catching visual thumbnail: {caption}",
        "Design a compelling thumbnail image: {caption}",
        "Generate a high-quality thumbnail: {caption}",
        "Create a visually striking thumbnail: {caption}",
    ]
    
    try:
        ds = load_dataset("PosterCraft/Poster100K", split="train", streaming=True)
        count = 0
        for sample in tqdm(ds, desc="Downloading posters", total=max_poster_samples):
            if count >= max_poster_samples:
                break
            
            caption = sample.get("caption", "")
            if not caption or len(caption) < 20:
                continue
            
            image_data = sample.get("image")
            if image_data is None:
                continue
            
            fname = f"poster_{count:06d}.jpg"
            fpath = os.path.join(IMAGE_DIR, fname)
            
            try:
                if isinstance(image_data, bytes):
                    img = Image.open(io.BytesIO(image_data))
                elif isinstance(image_data, str):
                    img_bytes = base64.b64decode(image_data)
                    img = Image.open(io.BytesIO(img_bytes))
                elif isinstance(image_data, Image.Image):
                    img = image_data
                else:
                    continue
                
                # Resize and save
                img = img.convert("RGB")
                w, h = img.size
                min_dim = min(w, h)
                left = (w - min_dim) // 2
                top = (h - min_dim) // 2
                img = img.crop((left, top, left + min_dim, top + min_dim))
                img = img.resize((IMAGE_SIZE, IMAGE_SIZE), Image.LANCZOS)
                img.save(fpath, "JPEG", quality=95)
                
                # Truncate long captions
                if len(caption) > 500:
                    caption = caption[:500] + "..."
                
                template = random.choice(T2I_TEMPLATES)
                instruction = template.format(caption=caption)
                entries.append({
                    "instruction": instruction,
                    "output_image": fname,
                })
                count += 1
                
                if count % 500 == 0:
                    logger.info(f"  Downloaded {count}/{max_poster_samples} posters")
                    
            except Exception as e:
                continue
                
    except Exception as e:
        logger.error(f"Error loading PosterCraft: {e}")
    
    logger.info(f"  Downloaded {len(entries)} poster images")
    
    # 1b. Generate synthetic thumbnail prompts
    logger.info(f"Generating {max_synthetic} synthetic thumbnail prompts...")
    
    categories = {
        "tech": [
            "Professional tech review thumbnail: sleek {product} product shot with dramatic studio lighting, bold comparison graphics, modern gradient background, text overlay '{title}'",
            "Technology unboxing thumbnail: hands opening {product} box, excited reaction, bright colorful background, text '{title}'",
        ],
        "cooking": [
            "Appetizing cooking thumbnail: close-up of {dish} with steam rising, warm golden lighting, rustic kitchen background, text '{title}'",
            "Food recipe thumbnail: beautifully plated {dish}, overhead flat-lay shot, marble countertop, fresh ingredients scattered, text '{title}'",
        ],
        "gaming": [
            "Epic gaming thumbnail: dramatic gameplay scene from {game}, character in action pose, particle effects, neon glow, bold text '{title}'",
            "Gaming reaction thumbnail: intense player reaction face, {game} gameplay in background, fire effects, text '{title}'",
        ],
        "fitness": [
            "Fitness motivation thumbnail: athletic person doing {exercise}, dynamic studio lighting, energetic colors, bold text '{title}'",
            "Health transformation thumbnail: before/after split screen, clean modern design, motivational text '{title}'",
        ],
        "education": [
            "Educational explainer thumbnail: clean whiteboard graphic about {topic}, colorful diagrams, professional font, text '{title}'",
            "Science video thumbnail: fascinating visualization of {topic}, bright curiosity-inducing colors, text '{title}'",
        ],
        "travel": [
            "Travel vlog thumbnail: stunning panoramic view of {place}, golden hour lighting, wanderlust aesthetic, text '{title}'",
            "Adventure thumbnail: breathtaking landscape of {place}, person silhouette, dramatic sky, text '{title}'",
        ],
    }
    
    fill_values = {
        "product": ["iPhone 16 Pro", "MacBook Air M4", "PS5 Pro", "Tesla Model Y", "Quest 4", "Galaxy S25"],
        "dish": ["pasta carbonara", "ramen bowl", "sushi platter", "chocolate lava cake", "grilled salmon"],
        "game": ["Zelda Echoes", "GTA VI", "Minecraft", "Fortnite", "Elden Ring DLC", "Cyberpunk 2078"],
        "exercise": ["deadlifts", "yoga flow", "HIIT workout", "marathon training", "calisthenics"],
        "topic": ["quantum computing", "AI revolution", "black holes", "climate change", "human brain"],
        "place": ["Tokyo streets", "Santorini sunset", "Iceland northern lights", "Bali rice terraces", "Swiss Alps"],
        "title": [
            "You NEED To See This!", "GAME CHANGER!", "The Ultimate Guide", "Top 10 SECRETS",
            "I Tried This For 30 Days", "Watch Before You Buy!", "Is It Worth It?",
            "HONEST Review", "This Changed EVERYTHING", "Don't Make This Mistake!",
        ],
    }
    
    for i in range(max_synthetic):
        cat_name = random.choice(list(categories.keys()))
        template = random.choice(categories[cat_name])
        
        prompt = template
        for key, values in fill_values.items():
            placeholder = "{" + key + "}"
            if placeholder in prompt:
                prompt = prompt.replace(placeholder, random.choice(values))
        
        # For synthetic entries, we'll use a placeholder image
        # These will be used only for prompt diversity — we need poster images as targets
        # Pair with a random poster image
        if entries:
            random_poster = random.choice(entries)
            entries.append({
                "instruction": f"Generate a YouTube thumbnail: {prompt}",
                "output_image": random_poster["output_image"],  # reuse poster as target
            })
    
    # Shuffle
    random.seed(SEED)
    random.shuffle(entries)
    
    # Write JSONL
    with open(JSONL_PATH, "w") as f:
        for entry in entries:
            f.write(json.dumps(entry) + "\n")
    
    logger.info(f"Total training samples: {len(entries)}")
    logger.info(f"JSONL saved to: {JSONL_PATH}")
    return len(entries)


# ─── Step 2: Train Model ────────────────────────────────────────────────────

def train_model():
    """Fine-tune Janus-Pro-7B for thumbnail generation."""
    logger.info("=" * 60)
    logger.info("Step 2: Training Janus-Pro-7B")
    logger.info("=" * 60)
    
    from transformers import AutoModelForCausalLM, get_cosine_schedule_with_warmup
    from janus.models import MultiModalityCausalLM, VLChatProcessor
    
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    dtype = torch.bfloat16
    
    # Initialize tracking
    trackio.init(project="thumbnail-vlm", name="janus-pro-thumbnail")
    
    # Load model
    logger.info(f"Loading {MODEL_PATH}...")
    processor = VLChatProcessor.from_pretrained(MODEL_PATH)
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_PATH, trust_remote_code=True, torch_dtype=dtype
    )
    model = model.to(device)
    
    # Freeze encoders (SigLIP + VQ) — only train LLM backbone + gen_head + aligners
    for name, param in model.named_parameters():
        if 'vision_model' in name or 'gen_vision_model' in name:
            param.requires_grad = False
    
    trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
    total = sum(p.numel() for p in model.parameters())
    logger.info(f"Parameters: {total/1e6:.1f}M total, {trainable/1e6:.1f}M trainable")
    
    model.train()
    
    # Load data
    entries = []
    with open(JSONL_PATH) as f:
        for line in f:
            entries.append(json.loads(line.strip()))
    logger.info(f"Training on {len(entries)} samples")
    
    # Optimizer
    optimizer = torch.optim.AdamW(
        filter(lambda p: p.requires_grad, model.parameters()),
        lr=LR, betas=(0.9, 0.95), weight_decay=0.0,
    )
    
    num_steps = (len(entries) * EPOCHS) // (BATCH_SIZE * GRADIENT_ACCUMULATION)
    warmup_steps = int(num_steps * 0.03)
    scheduler = get_cosine_schedule_with_warmup(optimizer, warmup_steps, num_steps)
    
    logger.info(f"Training config: epochs={EPOCHS}, batch={BATCH_SIZE}x{GRADIENT_ACCUMULATION}={BATCH_SIZE*GRADIENT_ACCUMULATION}")
    logger.info(f"Total steps: {num_steps}, warmup: {warmup_steps}")
    
    # Training loop
    global_step = 0
    best_loss = float("inf")
    running_loss = 0.0
    
    for epoch in range(EPOCHS):
        random.shuffle(entries)
        epoch_loss = 0.0
        epoch_steps = 0
        
        for batch_idx in range(0, len(entries), BATCH_SIZE):
            entry = entries[batch_idx]
            
            # Load target image
            img_path = os.path.join(IMAGE_DIR, entry["output_image"])
            if not os.path.exists(img_path):
                continue
            
            try:
                target_img = Image.open(img_path).convert("RGB")
                # Ensure 384x384
                target_img = target_img.resize((IMAGE_SIZE, IMAGE_SIZE), Image.LANCZOS)
                
                # Convert to tensor [-1, 1]
                arr = np.array(target_img).astype(np.float32) / 255.0 * 2.0 - 1.0
                target_tensor = torch.from_numpy(arr).permute(2, 0, 1).unsqueeze(0).to(device, dtype=dtype)
                
                # Encode target → VQ tokens
                with torch.no_grad():
                    quant, emb_loss, info = model.gen_vision_model.encode(target_tensor)
                    target_tokens = info[2].detach().reshape(1, -1)  # [1, 576]
                
                # Build prompt
                instruction = entry["instruction"]
                if random.random() < CFG_MASK_PROB:
                    instruction = ""  # CFG training: mask 10% of prompts
                
                conversation = [
                    {"role": "<|User|>", "content": instruction},
                    {"role": "<|Assistant|>", "content": ""},
                ]
                sft_format = processor.apply_sft_template_for_multi_turn_prompts(
                    conversations=conversation,
                    sft_format=processor.sft_format,
                    system_prompt="",
                )
                prompt = sft_format + processor.image_start_tag
                
                # Tokenize
                input_ids = processor.tokenizer.encode(prompt)
                input_ids = torch.LongTensor(input_ids).unsqueeze(0).to(device)
                
                # Get embeddings
                text_embeds = model.language_model.get_input_embeddings()(input_ids)
                img_embeds = model.prepare_gen_img_embeds(target_tokens.reshape(-1))
                img_embeds = img_embeds.reshape(1, IMAGE_TOKEN_NUM, -1)
                
                # Teacher forcing: [text | img[:-1]] → predict img
                full_embeds = torch.cat([text_embeds, img_embeds[:, :-1, :]], dim=1)
                
                with torch.amp.autocast('cuda', dtype=dtype):
                    outputs = model.language_model.model(inputs_embeds=full_embeds)
                    hidden = outputs.last_hidden_state
                    
                    # Logits for image token positions
                    text_len = text_embeds.shape[1]
                    image_hidden = hidden[:, text_len - 1:, :]  # [1, 576, 4096]
                    logits = model.gen_head(image_hidden)  # [1, 576, 16384]
                    
                    loss = F.cross_entropy(
                        logits.reshape(-1, VQ_CODEBOOK_SIZE),
                        target_tokens.reshape(-1),
                    )
                    loss = loss / GRADIENT_ACCUMULATION
                
                loss.backward()
                running_loss += loss.item() * GRADIENT_ACCUMULATION
                epoch_loss += loss.item() * GRADIENT_ACCUMULATION
                epoch_steps += 1
                
                # Optimizer step
                if (batch_idx // BATCH_SIZE + 1) % GRADIENT_ACCUMULATION == 0:
                    torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
                    optimizer.step()
                    scheduler.step()
                    optimizer.zero_grad()
                    global_step += 1
                    
                    # Logging
                    if global_step % LOG_EVERY == 0:
                        avg_loss = running_loss / LOG_EVERY
                        current_lr = scheduler.get_last_lr()[0]
                        print(f"step={global_step}/{num_steps}, epoch={epoch+1}/{EPOCHS}, "
                              f"loss={avg_loss:.4f}, lr={current_lr:.2e}")
                        trackio.log({
                            "train/loss": avg_loss,
                            "train/lr": current_lr,
                            "train/epoch": epoch + 1,
                            "train/step": global_step,
                        })
                        running_loss = 0.0
                    
                    # Save checkpoint
                    if global_step % SAVE_EVERY == 0:
                        ckpt_dir = os.path.join(OUTPUT_DIR, f"checkpoint-{global_step}")
                        os.makedirs(ckpt_dir, exist_ok=True)
                        model.save_pretrained(ckpt_dir)
                        processor.save_pretrained(ckpt_dir)
                        logger.info(f"Checkpoint saved: {ckpt_dir}")
                        
                        if epoch_loss / max(epoch_steps, 1) < best_loss:
                            best_loss = epoch_loss / epoch_steps
                            best_dir = os.path.join(OUTPUT_DIR, "best")
                            os.makedirs(best_dir, exist_ok=True)
                            model.save_pretrained(best_dir)
                            processor.save_pretrained(best_dir)
                            logger.info(f"New best model! loss={best_loss:.4f}")
                
            except Exception as e:
                logger.warning(f"Error at batch {batch_idx}: {e}")
                optimizer.zero_grad()
                continue
        
        avg_epoch_loss = epoch_loss / max(epoch_steps, 1)
        logger.info(f"Epoch {epoch+1}/{EPOCHS} complete. Avg loss: {avg_epoch_loss:.4f}")
    
    # Final save
    final_dir = os.path.join(OUTPUT_DIR, "final")
    os.makedirs(final_dir, exist_ok=True)
    model.save_pretrained(final_dir)
    processor.save_pretrained(final_dir)
    
    # Push to Hub
    logger.info(f"Pushing model to {HUB_MODEL_ID}...")
    try:
        model.push_to_hub(HUB_MODEL_ID, token=os.environ.get("HF_TOKEN"))
        processor.push_to_hub(HUB_MODEL_ID, token=os.environ.get("HF_TOKEN"))
        logger.info(f"Model pushed to: https://huggingface.co/{HUB_MODEL_ID}")
    except Exception as e:
        logger.error(f"Failed to push to hub: {e}")
    
    trackio.finish()
    return model, processor


# ─── Step 3: Evaluate ────────────────────────────────────────────────────────

def evaluate_model(model=None, processor=None):
    """Generate sample thumbnails to verify the model works."""
    logger.info("=" * 60)
    logger.info("Step 3: Evaluating Model")
    logger.info("=" * 60)
    
    if model is None:
        from transformers import AutoModelForCausalLM
        from janus.models import MultiModalityCausalLM, VLChatProcessor
        
        model_path = os.path.join(OUTPUT_DIR, "final")
        if not os.path.exists(model_path):
            model_path = MODEL_PATH
        
        processor = VLChatProcessor.from_pretrained(model_path)
        model = AutoModelForCausalLM.from_pretrained(
            model_path, trust_remote_code=True, torch_dtype=torch.bfloat16
        ).cuda().eval()
    else:
        model.eval()
    
    device = next(model.parameters()).device
    eval_dir = os.path.join(OUTPUT_DIR, "eval_samples")
    os.makedirs(eval_dir, exist_ok=True)
    
    test_prompts = [
        "A professional tech review YouTube thumbnail showing a sleek smartphone with dramatic lighting and bold text 'BEST PHONE 2025'",
        "Cooking video thumbnail with a delicious pasta dish, steam rising, warm kitchen lighting, text 'EASY 15-MIN RECIPE'",
        "Gaming thumbnail with an epic battle scene, neon glow effects, excited gamer face, text 'INSANE GAMEPLAY'",
        "Travel vlog thumbnail showing a stunning sunset over Santorini, Greece, golden hour, text 'DREAM VACATION'",
        "Fitness transformation thumbnail with before/after split, motivational energy, text '30 DAY CHALLENGE'",
    ]
    
    for i, prompt in enumerate(test_prompts):
        logger.info(f"Generating thumbnail {i+1}/{len(test_prompts)}: {prompt[:50]}...")
        
        conversation = [
            {"role": "<|User|>", "content": prompt},
            {"role": "<|Assistant|>", "content": ""},
        ]
        sft_format = processor.apply_sft_template_for_multi_turn_prompts(
            conversations=conversation,
            sft_format=processor.sft_format,
            system_prompt="",
        )
        prompt_text = sft_format + processor.image_start_tag
        
        with torch.inference_mode():
            input_ids = processor.tokenizer.encode(prompt_text)
            input_ids = torch.LongTensor(input_ids)
            
            tokens = torch.zeros((2, len(input_ids)), dtype=torch.int).to(device)
            tokens[0, :] = input_ids  # conditional
            tokens[1, :] = input_ids  # unconditional
            tokens[1, 1:-1] = processor.pad_id
            
            inputs_embeds = model.language_model.get_input_embeddings()(tokens)
            generated = torch.zeros((1, IMAGE_TOKEN_NUM), dtype=torch.int).to(device)
            
            past_kv = None
            for t in range(IMAGE_TOKEN_NUM):
                outputs = model.language_model.model(
                    inputs_embeds=inputs_embeds, use_cache=True, past_key_values=past_kv
                )
                past_kv = outputs.past_key_values
                logits = model.gen_head(outputs.last_hidden_state[:, -1, :])
                
                logit_cond = logits[0:1]
                logit_uncond = logits[1:2]
                logits_guided = logit_uncond + 5.0 * (logit_cond - logit_uncond)
                probs = torch.softmax(logits_guided, dim=-1)
                
                next_tok = torch.multinomial(probs, num_samples=1)
                generated[:, t] = next_tok.squeeze(-1)
                
                next_tok_exp = torch.cat([next_tok, next_tok], dim=0)
                img_emb = model.prepare_gen_img_embeds(next_tok_exp.squeeze(-1))
                inputs_embeds = img_emb.unsqueeze(1)
            
            dec = model.gen_vision_model.decode_code(
                generated, shape=[1, 8, IMAGE_SIZE // 16, IMAGE_SIZE // 16]
            )
            dec = dec.to(torch.float32).cpu().numpy().transpose(0, 2, 3, 1)
            dec = np.clip((dec + 1) / 2 * 255, 0, 255).astype(np.uint8)
            
            img = Image.fromarray(dec[0])
            save_path = os.path.join(eval_dir, f"thumbnail_{i}.png")
            img.save(save_path)
            logger.info(f"  Saved: {save_path}")
    
    logger.info(f"Evaluation complete! Samples saved to {eval_dir}/")


# ─── Main ────────────────────────────────────────────────────────────────────

if __name__ == "__main__":
    random.seed(SEED)
    np.random.seed(SEED)
    torch.manual_seed(SEED)
    
    # Step 1: Prepare data
    num_samples = prepare_training_data(max_poster_samples=8000, max_synthetic=2000)
    
    # Step 2: Train
    model, processor = train_model()
    
    # Step 3: Evaluate
    evaluate_model(model, processor)
    
    logger.info("=" * 60)
    logger.info("PIPELINE COMPLETE!")
    logger.info(f"  Model: https://huggingface.co/{HUB_MODEL_ID}")
    logger.info(f"  Eval samples: {OUTPUT_DIR}/eval_samples/")
    logger.info("=" * 60)