""" Sample submission script for the MGI task. End-to-end pipeline: 1. Download the 900 reference images from the SprintML/MGI dataset repo on Hugging Face (data/img_000.png .. data/img_899.png). 2. Download the RAR-XL generator weights (yucornetto/RAR) and the MaskGIT-VQ tokenizer weights. 3. Build a valid 1800-slot submission.npz, following the same six-block transition convention in the task description: 0000-0299 M->N 0300-0599 M->G 0600-0899 N->M 0900-1199 N->G 1200-1499 G->M 1500-1799 G->N 4. Submit it to the evaluation API. build_submission() below is a placeholder that just returns random noise for every slot. This is for familiarzing you with submission shape/format. Please replace your own attack in its place. """ import os import sys import numpy as np import requests from pathlib import Path from PIL import Image from huggingface_hub import hf_hub_download, snapshot_download BASE_DIR = Path(__file__).resolve().parent # --- submission / API config ------------------------------------------------- BASE_URL = "http://35.192.205.84" API_KEY = "YOUR_API_KEY_HERE" TASK_ID = "29-mgi" OUTPUT_PATH = "submission.npz" # --- submission format --------------------------------------------------- # 1800 slots, six 300-image blocks for the six required misclassifications: # 0000-0299 M->N 0300-0599 M->G 0600-0899 N->M # 0900-1199 N->G 1200-1499 G->M 1500-1799 G->N BASE_IMAGES = 900 # underlying reference dataset (img_000.png .. img_899.png) IMAGE_SIZE = 256 TOTAL_IMAGES = 1800 # submission slots EXPECTED_NAMES = tuple(f"{index:04d}" for index in range(TOTAL_IMAGES)) # --- Hugging Face sources ----------------------------------------------------- HF_DATASET_REPO = "SprintML/MGI" HF_DATA_SUBFOLDER = "data" HF_RAR_REPO = "yucornetto/RAR" # RAR generator checkpoints (rar_xl.bin, ...) HF_MASKGIT_REPO = "fun-research/TiTok" # MaskGIT-VQ tokenizer weight used by RAR RAR_MODEL_SIZE = "rar_xl" MODEL_DIR = BASE_DIR / "model" # RAR-XL architecture config: the hyperparameters RAR/demo_util.py needs to # build the model class match those in the official RAR repo's # configs/training/generator/rar.yaml for the XL size (see rar/README_RAR.md). # Written locally only if not already present. RAR_XL_CONFIG = """\ experiment: generator_checkpoint: "" model: vq_model: codebook_size: 1024 token_size: 256 num_latent_tokens: 256 finetune_decoder: False pretrained_tokenizer_weight: "" generator: hidden_size: 1280 num_hidden_layers: 32 num_attention_heads: 16 intermediate_size: 5120 dropout: 0.1 attn_drop: 0.1 class_label_dropout: 0.1 image_seq_len: 256 condition_num_classes: 1000 use_checkpoint: False """ def ensure_dataset() -> Path: """Download the 900 reference images from the HF dataset repo, if missing.""" local_dir = snapshot_download( repo_id=HF_DATASET_REPO, repo_type="dataset", allow_patterns=[f"{HF_DATA_SUBFOLDER}/*.png"], ) data_dir = Path(local_dir) / HF_DATA_SUBFOLDER print(f"Reference dataset ready: {data_dir}") return data_dir def ensure_model_weights() -> tuple[Path, Path, Path]: """Download RAR-XL + MaskGIT-VQ weights and write a matching config, if missing.""" MODEL_DIR.mkdir(parents=True, exist_ok=True) generator_ckpt = Path( hf_hub_download(repo_id=HF_RAR_REPO, filename=f"{RAR_MODEL_SIZE}.bin") ) tokenizer_ckpt = Path( hf_hub_download( repo_id=HF_MASKGIT_REPO, filename="maskgit-vqgan-imagenet-f16-256.bin" ) ) config_path = MODEL_DIR / "rar.yaml" if not config_path.exists(): config_path.write_text(RAR_XL_CONFIG) print(f"Model weights ready: generator={generator_ckpt}, tokenizer={tokenizer_ckpt}") return config_path, generator_ckpt, tokenizer_ckpt def load_reference_images(data_dir: Path) -> np.ndarray: """Load the 900 reference dataset images as uint8 (BASE_IMAGES, 256, 256, 3).""" images = np.empty((BASE_IMAGES, IMAGE_SIZE, IMAGE_SIZE, 3), dtype=np.uint8) for i in range(BASE_IMAGES): with Image.open(data_dir / f"img_{i:03d}.png") as img: images[i] = np.asarray( img.convert("RGB").resize((IMAGE_SIZE, IMAGE_SIZE), Image.BILINEAR), dtype=np.uint8, ) return images def build_submission(original: np.ndarray, seed: int = 0) -> np.ndarray: """ Placeholder -- fill this in with your own attack. Returns random noise for every one of the 1800 slots, just to show the submission shape/format you need to produce. Scores 0 as-is. """ rng = np.random.default_rng(seed) return rng.integers( 0, 256, size=(TOTAL_IMAGES, IMAGE_SIZE, IMAGE_SIZE, 3), dtype=np.uint8 ) def make_submission_file(images: np.ndarray, output_path: str) -> None: assert images.shape == (TOTAL_IMAGES, IMAGE_SIZE, IMAGE_SIZE, 3), images.shape assert images.dtype == np.uint8, images.dtype names = np.array(EXPECTED_NAMES) np.savez_compressed(output_path, images=images, names=names) print(f"Saved submission -> {output_path}") def die(msg: str) -> None: print(msg, file=sys.stderr) sys.exit(1) def submit(file_path: str) -> None: if not os.path.isfile(file_path): die(f"File not found: {file_path}") try: with open(file_path, "rb") as f: files = { "file": (os.path.basename(file_path), f, "application/octet-stream"), } resp = requests.post( f"{BASE_URL}/submit/{TASK_ID}", headers={"X-API-Key": API_KEY}, files=files, ) try: body = resp.json() except Exception: body = {"raw_text": resp.text} if resp.status_code == 413: die("Upload rejected: file too large (HTTP 413). Reduce size and try again.") resp.raise_for_status() submission_id = body.get("submission_id") print("Successfully submitted.") print("Server response:", body) if submission_id: print(f"Submission ID: {submission_id}") except requests.exceptions.RequestException as e: detail = getattr(e, "response", None) print(f"Submission error: {e}") if detail is not None: try: print("Server response:", detail.json()) except Exception: print("Server response (text):", detail.text) sys.exit(1) if __name__ == "__main__": data_dir = ensure_dataset() ensure_model_weights() # downloads RAR-XL + MaskGIT-VQ weights for your own attack original = load_reference_images(data_dir) submitted = build_submission(original) make_submission_file(submitted, OUTPUT_PATH) submit(OUTPUT_PATH)