train.py

#!/usr/bin/env python3

from collections import defaultdict
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path
import glob
import os
import random
import sys

from huggingface_hub import snapshot_download
import pandas as pd
from PIL import Image
from safetensors.torch import load_file, save_file
import schedulefree
import torch
import wandb
from torch import nn
from torch.utils.data import Dataset
from tqdm.auto import tqdm
import torchvision.transforms.v2 as v2
from model import DINOv3ViTH, TaggerAestheticModel, _split_and_clean_state_dict

PATCH_SIZE = 16
N_REGISTERS = 4
MAX_SIZE = 512
IMAGENET_MEAN = [0.485, 0.456, 0.406]
IMAGENET_STD = [0.229, 0.224, 0.225]
IMAGE_PREPROCESS_WORKERS = 16

random.seed(3407)
torch.set_grad_enabled(True)
torch.set_float32_matmul_precision("high")


def load_votes_and_split():
    repo_id = "taigasan/e6-visual-ratings"
    local_repo_path = snapshot_download(repo_id=repo_id, repo_type="dataset", revision="daaa857ffab11075c2fc6912e7f23879d324dcc9")
    print("Downloaded repo to:", local_repo_path)

    rating_log_dir = os.path.join(local_repo_path, "ratings_log")
    parquet_files = sorted(glob.glob(os.path.join(rating_log_dir, "*.parquet")))
    assert len(parquet_files) > 0
    df_list = [pd.read_parquet(path) for path in parquet_files]
    combined_df = pd.concat(df_list, ignore_index=True)
    print("Total votes before pool filter:", len(combined_df))

    pool_path = os.path.join(local_repo_path, "pool.parquet")
    pool_df = pd.read_parquet(pool_path)
    assert "md5" in pool_df.columns
    pool_md5_list = sorted(pool_df["md5"].astype(str).tolist())
    valid_md5 = set(pool_md5_list)
    combined_df = combined_df[
        combined_df["md5a"].isin(valid_md5) & combined_df["md5b"].isin(valid_md5)
    ].reset_index(drop=True)
    print("Total votes after pool filter:", len(combined_df))
    print("Pool rows:", len(pool_df))

    df = combined_df.sample(frac=1, random_state=42).reset_index(drop=True)
    df_first = df.iloc[:2000].reset_index(drop=True)
    df_second = df.iloc[2000:].reset_index(drop=True)
    print("Val rows:", len(df_first))
    print("Train rows:", len(df_second))
    return df_first, df_second, pool_md5_list


def _snap(x: int, m: int) -> int:
    return max(m, (x // m) * m)


def preprocess_pil(img: Image.Image, max_size: int = MAX_SIZE) -> torch.Tensor:
    img = img.convert("RGB")
    w, h = img.size
    long_edge = max(w, h)
    target_long = _snap(min(long_edge, max_size), PATCH_SIZE)
    scale = target_long / long_edge
    new_w = _snap(max(PATCH_SIZE, round(w * scale)), PATCH_SIZE)
    new_h = _snap(max(PATCH_SIZE, round(h * scale)), PATCH_SIZE)
    return v2.Compose(
        [
            v2.Resize((new_h, new_w), interpolation=v2.InterpolationMode.LANCZOS),
            v2.ToImage(),
            v2.ToDtype(torch.float32, scale=True),
            v2.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD),
        ]
    )(img).unsqueeze(0)


def find_local_image_path(md5: str, image_dir) -> Path:
    image_dir = Path(image_dir)
    for ext in ("webp", "jpg", "jpeg", "png"):
        p = image_dir / f"{md5}.{ext}"
        if p.exists():
            return p
    raise AssertionError(f"Missing local image for md5={md5}; checked dir={image_dir}")


def load_and_preprocess_image(item):
    md5, image_dir = item
    p = find_local_image_path(md5, image_dir)
    with Image.open(p) as img:
        pixel_values = preprocess_pil(img)[0]
    h = int(pixel_values.shape[1])
    w = int(pixel_values.shape[2])
    return md5, (h, w), pixel_values


def pool_features(outputs) -> torch.Tensor:
    hidden = outputs if isinstance(outputs, torch.Tensor) else outputs.last_hidden_state
    assert hidden.ndim == 3 and hidden.shape[1] >= 1 + N_REGISTERS
    cls = hidden[:, 0, :]
    regs = hidden[:, 1 : 1 + N_REGISTERS, :].flatten(1)  # [b, reg, d] -> [b, reg*d]
    return torch.cat([cls, regs], dim=-1).to(torch.float32)


class PreferenceDataset(Dataset):
    def __init__(self, pair_df, embed_cache):
        self.embed_cache = embed_cache

        pairs = []
        has_a_won = False
        has_b_won = False
        for row in pair_df.itertuples(index=False):
            md5a = row.md5a
            md5b = row.md5b
            winner_md5 = row.winner_md5

            if not winner_md5:
                continue
            if winner_md5 == md5a:
                outcome = 1
                has_a_won = True
            else:
                outcome = 0
                has_b_won = True

            assert md5a in self.embed_cache
            assert md5b in self.embed_cache
            pairs.append(
                {
                    "md5a": md5a,
                    "md5b": md5b,
                    "outcome": outcome,
                }
            )

        assert has_a_won
        assert has_b_won
        self.pairs = pairs

    def __len__(self):
        return len(self.pairs)

    def __getitem__(self, idx):
        sample = self.pairs[idx]
        md5a = sample["md5a"]
        md5b = sample["md5b"]
        outcome = sample["outcome"]
        embed_a = self.embed_cache[md5a]
        embed_b = self.embed_cache[md5b]
        outcome = torch.tensor([outcome], dtype=torch.float32)
        return {
            "embed_a": embed_a,
            "embed_b": embed_b,
            "outcome": outcome,
        }


def build_feature_table(md5_list, image_dir, backbone, batch_size=64):
    out = {}
    total = len(md5_list)
    assert total > 0
    progress_step = max(1, total // 100)
    next_progress = progress_step
    feature_dim = None

    backbone.eval()
    with torch.no_grad():
        for start in tqdm(range(0, total, batch_size), total=(total + batch_size - 1) // batch_size, leave=True):
            batch_md5 = md5_list[start : start + batch_size]
            size_to_md5 = defaultdict(list)
            size_to_tensors = defaultdict(list)

            worker_count = min(IMAGE_PREPROCESS_WORKERS, len(batch_md5))
            assert worker_count > 0
            with ThreadPoolExecutor(max_workers=worker_count) as executor:
                preprocessed = list(executor.map(load_and_preprocess_image, [(m, image_dir) for m in batch_md5]))

            for md5, size, pixel_values in preprocessed:
                size_to_md5[size].append(md5)
                size_to_tensors[size].append(pixel_values)

            for size in size_to_md5:
                md5_group = size_to_md5[size]
                tensor_group = size_to_tensors[size]
                for gstart in range(0, len(md5_group), batch_size):
                    gend = min(gstart + batch_size, len(md5_group))
                    pixel_values = torch.stack(tensor_group[gstart:gend], dim=0).to("cuda")
                    outputs = backbone(pixel_values=pixel_values)
                    features = pool_features(outputs).cpu()
                    assert features.ndim == 2 and features.shape[0] == (gend - gstart)
                    assert torch.isfinite(features).all(), f"Non-finite features in size group {size}"
                    if feature_dim is None:
                        feature_dim = int(features.shape[1])
                    assert int(features.shape[1]) == feature_dim
                    for md5, feat in zip(md5_group[gstart:gend], features):
                        out[md5] = feat

            done = len(out)
            if done >= next_progress or done == total:
                print(f"Feature progress: {done}/{total}")
                while next_progress <= done:
                    next_progress += progress_step

    assert len(out) == total
    return out


def load_or_build_embed_cache(md5_list, image_dir, backbone, cache_path, batch_size=64):
    assert len(md5_list) > 0

    cache = {}
    cache_path = Path(cache_path)
    if cache_path.exists():
        payload = torch.load(cache_path, map_location="cpu")
        cached_md5 = payload["md5"]
        cached_features = payload["features"]
        assert len(cached_md5) == len(cached_features)
        assert cached_features.ndim == 2 and torch.isfinite(cached_features).all()
        cache = {m: cached_features[i] for i, m in enumerate(cached_md5)}
        print(f"Loaded embed cache: {len(cache)} items from {cache_path}")

    missing = [m for m in md5_list if m not in cache]
    if missing:
        print(f"Building missing embeddings: {len(missing)}")
        built = build_feature_table(missing, image_dir=image_dir, backbone=backbone, batch_size=batch_size)
        cache.update(built)

        ordered = sorted(cache.keys())
        features = torch.stack([cache[m] for m in ordered], dim=0)
        assert features.ndim == 2 and torch.isfinite(features).all()
        cache_path.parent.mkdir(parents=True, exist_ok=True)
        torch.save({"md5": ordered, "features": features}, cache_path)
        print(f"Saved embed cache: {cache_path}")

    return {m: cache[m] for m in md5_list}


def load_frozen_backbone(backbone_path):
    backbone_path = Path(backbone_path).resolve()
    assert backbone_path.exists(), f"Missing backbone checkpoint: {backbone_path}"
    sd = load_file(str(backbone_path), device="cpu")
    backbone_sd, _ = _split_and_clean_state_dict(sd)
    backbone = DINOv3ViTH()
    backbone.load_state_dict(backbone_sd, strict=True)
    backbone = backbone.to("cuda").eval().to(torch.float32)
    for p in backbone.parameters():
        p.requires_grad_(False)
    return backbone


def test_code(df_second, pool_md5_list):
    batch_size = 4
    backbone_path = Path("./tagger_proto.safetensors")
    backbone = load_frozen_backbone(backbone_path)
    embed_cache = load_or_build_embed_cache(
        md5_list=pool_md5_list,
        image_dir=Path("./samples").resolve(),
        backbone=backbone,
        cache_path="tagger_scorer/frozen_embed_cache_512.pt",
        batch_size=64,
    )

    train_set = PreferenceDataset(
        pair_df=df_second,
        embed_cache=embed_cache,
    )
    train_loader = torch.utils.data.DataLoader(
        train_set,
        batch_size=batch_size,
        shuffle=True,
        num_workers=0,
    )
    print(len(train_set))
    for sample in train_loader:
        print(sample["embed_a"].shape, sample["embed_b"].shape, sample["outcome"].shape)
        break


def count_parameters(model):
    return sum(p.numel() for p in model.parameters() if p.requires_grad)


class AestheticPredictor(nn.Module):
    def __init__(self, feature_dim: int):
        super().__init__()
        self.scoring_head = TaggerAestheticModel(feature_dim).scoring_head

    def forward(self, features):
        return self.scoring_head(features)


class dotdict(dict):
    __getattr__ = dict.get
    __setattr__ = dict.__setitem__
    __delattr__ = dict.__delitem__


def eval_model(model, optimizer, val_loader, wandb_run, step):
    model.eval()
    optimizer.eval()
    acc_loss = 0.0
    num_steps = 0
    with torch.no_grad():
        for data in tqdm(val_loader, total=len(val_loader), leave=True):
            for k in data:
                data[k] = data[k].to("cuda")
            output_one = model(data["embed_a"])
            output_two = model(data["embed_b"])
            pred_outcome = output_one - output_two
            criterion = nn.BCEWithLogitsLoss()
            loss = criterion(pred_outcome, data["outcome"])
            acc_loss += loss.item()
            num_steps += 1

    acc_loss /= num_steps
    wandb_run.log({"val_loss": acc_loss}, step=step)

    model.train()
    optimizer.train()


def save_model(model, step, model_name):
    out_dir = Path("./checkpoints")
    out_dir.mkdir(parents=True, exist_ok=True)
    out_path = out_dir / f"{model_name}_{step}.safetensors"
    save_file({k: v.detach().to(torch.float32).contiguous() for k, v in model.state_dict().items()}, str(out_path))


def train_main():
    print(
        "WARNING: Using pinned dataset revision daaa857ffab11075c2fc6912e7f23879d324dcc9. "
        "This is an old snapshot — embed_cache.pt covers this pool so you do not need to "
        "if you want to use the latest version have sample images locally in /samples."
    )
    df_first, df_second, pool_md5_list = load_votes_and_split()

    settings = {
        "model_name": "classifer_head",
        "remarks": "Training head only with cached DINOv3 embeddings at 512.",
        "optimizer": "AdamW schedulefree",
        "batch_size": 4,
        "val_batch_size": 16,
        "image_dir": "./samples",
        "do_eval": True,
        "backbone_path": "./tagger_proto.safetensors",
        "embed_cache_path": "./embed_cache.pt",
        "lr": 1e-3,
        "warmup_steps": 100,
        "weight_decay": 1e-2,
        "betas": (0.9, 0.999),
        "save_every": 5000,
        "eval_every": 2000,
        "num_accumulation_steps": 8,
        "train_target_steps": 6000 * 8,
        "embed_cache_batch_size": 64,
    }
    s = dotdict(settings)

    backbone = load_frozen_backbone(s.backbone_path)

    image_dir = Path(s.image_dir).resolve()

    embed_cache = load_or_build_embed_cache(
        md5_list=pool_md5_list,
        image_dir=image_dir,
        backbone=backbone,
        cache_path=s.embed_cache_path,
        batch_size=s.embed_cache_batch_size,
    )
    print("Cached embeds:", len(embed_cache))

    feature_dim = int(next(iter(embed_cache.values())).shape[0])
    model = AestheticPredictor(feature_dim).to(torch.float32).to("cuda")
    print(count_parameters(model))

    train_set = PreferenceDataset(
        pair_df=df_second,
        embed_cache=embed_cache,
    )
    print("Train set size:", len(train_set))
    train_loader = torch.utils.data.DataLoader(
        train_set,
        batch_size=s.batch_size,
        shuffle=True,
        num_workers=12,
        drop_last=True,
    )

    if s.do_eval:
        val_set = PreferenceDataset(
            pair_df=df_first,
            embed_cache=embed_cache,
        )
        val_loader = torch.utils.data.DataLoader(
            val_set,
            batch_size=s.val_batch_size,
            shuffle=True,
            num_workers=12,
            drop_last=True,
        )

    wandb_api_key = os.getenv("WANDB_API_KEY")
    assert wandb_api_key is not None and wandb_api_key != "", "WANDB_API_KEY env var is required"
    wandb.login(key=wandb_api_key)
    wandb_run = wandb.init(
        project=os.getenv("WANDB_PROJECT", "aesthetic_bradley_terry"),
        name=settings["model_name"],
        config=settings,
    )
    assert wandb_run is not None

    decay_head = []
    no_decay_head = []
    for name, param in model.scoring_head.named_parameters():
        param.requires_grad = True
        if "bias" in name or "embed" in name or "norm" in name:
            no_decay_head.append(param)
        else:
            decay_head.append(param)

    optimizer = schedulefree.AdamWScheduleFree(
        [
            {"params": decay_head, "lr": s.lr},
            {"params": no_decay_head, "weight_decay": 0.0, "lr": s.lr},
        ],
        lr=s.lr,
        warmup_steps=s.warmup_steps,
        weight_decay=s.weight_decay,
        betas=s.betas,
    )
    model.train()
    optimizer.train()
    step_counter = 0

    optimizer.zero_grad()
    try:
        for _ in range(200000):
            acc_loss = 0.0
            for i, data in tqdm(enumerate(train_loader, 0), total=len(train_loader)):
                for k in data:
                    data[k] = data[k].to("cuda")
                output_one = model(data["embed_a"])
                output_two = model(data["embed_b"])
                pred_outcome = output_one - output_two
                criterion = nn.BCEWithLogitsLoss()
                loss = criterion(pred_outcome, data["outcome"])
                pre_acc_loss = loss.item()
                wandb_run.log({"pre_acc_train_loss": pre_acc_loss}, step=step_counter)

                acc_loss += pre_acc_loss / s.num_accumulation_steps
                loss = loss / s.num_accumulation_steps
                loss.backward()
                if ((i + 1) % s.num_accumulation_steps == 0) or (i + 1 == len(train_loader)):
                    grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
                    optimizer.step()
                    optimizer.zero_grad()
                    wandb_run.log(
                        {
                            "train_loss": acc_loss,
                            "lr": s.lr,
                            "grad_norm": float(grad_norm),
                        },
                        step=step_counter,
                    )
                    acc_loss = 0.0
                    if step_counter + 1 >= s.train_target_steps:
                        save_model(model, step_counter, s.model_name)
                        if s.do_eval:
                            eval_model(model, optimizer, val_loader, wandb_run, step_counter)
                        return

                if step_counter % s.save_every == 0:
                    save_model(model, step_counter, s.model_name)
                if s.do_eval and step_counter % s.eval_every == 0:
                    eval_model(model, optimizer, val_loader, wandb_run, step_counter)
                step_counter += 1
    except KeyboardInterrupt:
        pass
    finally:
        save_model(model, step_counter, s.model_name)
        if s.do_eval:
            eval_model(model, optimizer, val_loader, wandb_run, step_counter)
        wandb.finish()

    print("Finished Training")


if __name__ == "__main__":
    train_main()