adversarial_examples / task_template.py
IvoHoese's picture
Create task_template.py
ae7f001 verified
import torch
import requests
import sys
import os
import numpy as np
# --------------------------------
# DATASET
# --------------------------------
"""
Dataset contents:
-"image_ids": Tensor containing the IDs of the 100 natural images, has shape (100)
-"images": Tensor containing the 100 natural images, has shape (100, 3, 28, 28)
-"labels": Tensor of true labels for the images, has shape (100)
"""
# Load the dataset
dataset = torch.load("natural_images.pt", weights_only=False)
print("Dataset keys:", dataset.keys())
print("Image IDs shape:", dataset["image_ids"].shape)
print("Images shape:", dataset["images"].shape)
print("Labels shape:", dataset["labels"].shape)
print("First 10 image IDs:", dataset["image_ids"][:10])
print("First 10 labels:", dataset["labels"][:10])
print("First image tensor:", dataset["images"][:1])
# --------------------------------
# SUBMISSION FORMAT
# --------------------------------
"""
The submission must be a .npz file of the following format:
-"image_ids": Tensor containing the IDs of the images corresponding to your adversarial examples, has shape (100)
-"images": Tensor containing the generated adversarial examples in the same order as the corresponding
natural images, has shape (100, 3, 28, 28)
"""
# Example Submission:
adversarial_examples = torch.randint(0, 256, size=(100, 3, 28, 28), dtype=torch.uint8)
adversarial_examples = adversarial_examples.float() / 255.0 # normalize to [0, 1] range
images_np = adversarial_examples.detach().cpu().numpy()
image_ids = np.arange(len(images_np))
np.savez_compressed("example_submission.npz", image_ids=image_ids, images=images_np)
# --------------------------------
# SUBMISSION PROCESS
# --------------------------------
"""
Adversarial Examples Task β€” Participant Submission Guide
========================================================
You will upload a single **.npz** file that contains ONLY an array named **'images'**.
The evaluator will load your file, run shape/dtype checks against the natural images,
and then score it by running a fixed classifier and measuring perturbations.
Follow these rules carefully to avoid automatic rejection.
1) File format
--------------
- **Extension:** `.npz` (NumPy compressed archive)
- **Content:** must contain exactly one required key: `'images'`
- **Max file size:** 200 MB (hard limit). Larger files are rejected.
2) Array requirements
---------------------
Let `G` be the ground-truth tensor loaded:
- **Shape:** `images.shape` must match `G["images"].shape` **exactly**.
- If `G["images"]` is `(N, 3, H, W)`, your array must also be `(N, 3, H, W)`.
- No extra samples; no fewer; no different dimensions.
- **Dtype:** `images.dtype` must match `G["images"].dtype` **exactly**.
- If the GT uses `float32`, you must submit `float32`.
- Safe cast example: `images = np.asarray(images, dtype=np.float32)`
- **Finite values only:** No NaN or Inf anywhere.
- The evaluator checks: `torch.isfinite(images).all()`.
- **Contiguity:** The server will convert to a contiguous Torch tensor; standard NumPy arrays are fine.
3) Typical failure messages & what they mean
--------------------------------------------
- "File must be .npz and contain an 'images' array."
β†’ Wrong extension or missing `'images'` key.
- "File too large: X bytes (limit 209715200)."
β†’ Your file exceeds 200 MB.
- "Failed to read .npz: ..."
β†’ The file is corrupted or not a valid `.npz` created with `allow_pickle=False`.
- "Failed to convert 'images' to torch tensor: ..."
β†’ Your `'images'` array has an unsupported dtype or structure (e.g., object array).
- "Submitted images must have shape (N, C, H, W), but got (...)."
β†’ Shape mismatch with the ground-truth images.
- "Submitted images must be of type torch.float32, but got torch.float64."
β†’ Dtype mismatch with the ground-truth images.
- "Images must not contain NaN or Inf values."
β†’ Clean your array: `np.isfinite(images).all()` must be True.
"""
BASE_URL = "http://34.122.51.94:80"
API_KEY = "YOUR_API_KEY_HERE"
TASK_ID = "10-adversarial-examples"
# Path to the .npz file containing the images you want to get logits for
QUERY_PATH = "PATH/TO/YOUR/QUERY_FILE.npz"
# Path to the .npz file you want to send
FILE_PATH = "PATH/TO/YOUR/SUBMISSION.npz"
GET_LOGITS = False # set True to get logits from the API
SUBMIT = False # set True to submit your solution
def die(msg):
print(f"{msg}", file=sys.stderr)
sys.exit(1)
if GET_LOGITS:
with open(QUERY_PATH, "rb") as f:
files = {"npz": (QUERY_PATH, f, "application/octet-stream")}
response = requests.post(
f"{BASE_URL}/{TASK_ID}/logits",
files=files,
headers={"X-API-Key": API_KEY},
)
if response.status_code == 200:
data = response.json()
print("Request successful")
print(data)
else:
print("Request failed")
print("Status code:", response.status_code)
print("Detail:", response.text)
if SUBMIT:
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, "csv"),
}
resp = requests.post(
f"{BASE_URL}/submit/{TASK_ID}",
headers={"X-API-Key": API_KEY},
files=files,
timeout=(10, 120),
)
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)