使用 AugLy 打造端到端多模態資料擴增與對抗穩健性基準：涵蓋影像、文字、音訊與 PyTorch

In this tutorial, we build a comprehensive multimodal augmentation and robustness workflow with AugLy for images, text, and audio.We start by addressing modern dependency compatibility issues and generating deterministic synthetic datasets so the experiments remain self-contained and reproducible.

We then explore AugLy’s functional and class-based APIs, metadata, and intensity tracking, probabilistic composition, bounding-box-aware transformations, and custom transforms.

We extend the workflow into practical robustness experiments by benchmarking perceptual-hash copy detection under image distortions and evaluating text classifiers against adversarial perturbations, Unicode obfuscation, sanitization, and adversarial training.

We also integrate audio augmentation, build a queryable metadata warehouse, and connect AugLy transformations directly to PyTorch datasets and DataLoaders, giving us an end-to-end view of augmentation as both a data-generation mechanism and a measurable robustness tool.

Copy CodeCopiedUse a different Browserimport subprocess, sys, importlib def sh(cmd): print(f"$ {cmd}") subprocess.run(cmd, shell=True, check=False, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL) def need(mod): try: importlib.

importmodule(mod) return False except ImportError: return True if need("augly"): sh("apt-get -qq install -y libmagic1 > /dev/null 2>&1") sh(f'"{sys.executable}" -m pip install -q --no-deps augly') sh(f'"{sys.executable}" -m pip install -q "iopath>=0.1.8" "python-magic>=0.4.22" ' f'"regex>=2021.4.

4" "nlpaug==1.1.3"') import numpy as np from PIL import Image, ImageDraw, ImageFont, ImageFilter for name, builtin in (("float", float), ("int", int), ("bool", bool)): if not hasattr(np, name): setattr(np, name, builtin) def size(font, text): left, top, right, bottom = font.

getbbox(text) return (right, bottom) if not hasattr(ImageFont.FreeTypeFont, "getsize"): ImageFont.FreeTypeFont.getsize = lambda self, t, *a, **k: size(self, t) if not hasattr(ImageFont.

FreeTypeFont, "getsizemultiline"): def getsizemultiline(self, text, direction=None, spacing=4, features=None, language=None, strokewidth=0): lines = text.

split("\n") w = max((size(self, ln)[0] for ln in lines), default=0) h = sum(size(self, ln)[1] for ln in lines) + spacing (len(lines) - 1) return (w, h) ImageFont.FreeTypeFont.

getsizemultiline = getsizemultiline import os, io, json, math, random, string, textwrap, unicodedata, warnings from dataclasses import dataclass from typing import Any, Dict, List, Optional, Tuple import matplotlib.pyplot as plt import pandas as pd import augly.image as imaugs import augly.

text as textaugs import augly.utils as augutils from augly.image.transforms import BaseTransform as ImageBaseTransform warnings.filterwarnings("ignore") pd.setoption("display.width", 160) SEED = 1234 random.seed(SEED) np.random.seed(SEED) print("\n" + "=" 78) print("AugLy ready.

assets at:", augutils.ASSETSBASEDIR) print("image augs :", len([f for f in dir(imaugs) if f[0].islower()])) print("text augs :", len([f for f in dir(textaugs) if f[0].islower()])) print("=" 78 + "\n") def makeimage(idx: int, w: int = 320, h: int = 240) -> Tuple[Image.

Image, Tuple[int, int, int, int]]: """Procedurally generated 'photo' + a ground-truth bbox in pascalvoc format.""" rng = random.Random(SEED + idx) img = Image.new("RGB", (w, h), tuple(rng.randint(20, 90) for in range(3))) d = ImageDraw.Draw(img) for in range(70): x0, y0 = rng.randint(0, w), rng.

randint(0, h) d.line([x0, y0, x0 + rng.randint(-60, 60), y0 + rng.randint(-60, 60)], fill=tuple(rng.randint(60, 160) for in range(3)), width=rng.randint(1, 3)) ow, oh = rng.randint(70, 130), rng.randint(60, 110) ox, oy = rng.randint(10, w - ow - 10), rng.

randint(10, h - oh - 10) box = (ox, oy, ox + ow, oy + oh) colour = tuple(rng.randint(150, 255) for in range(3)) if idx % 3 == 0: d.ellipse(box, fill=colour, outline=(255, 255, 255), width=3) elif idx % 3 == 1: d.rectangle(box, fill=colour, outline=(255, 255, 255), width=3) else: d.

polygon([(ox + ow // 2, oy), (ox + ow, oy + oh), (ox, oy + oh)], fill=colour, outline=(255, 255, 255)) return img, box NIMAGES = 24 IMAGES, BOXES = zip([makeimage(i) for i in range(NIMAGES)]) IMAGES, BOXES = list(IMAGES), list(BOXES) DEMOIMG, DEMOBOX = IMAGES[0], BOXES[0] def maketextdataset(nperclass: int = 260): """Tiny sentiment corpus built from templates -> learnable but not trivial.

""" rng = random.

Random(SEED) posadj = ["excellent", "delightful", "superb", "charming", "brilliant", "flawless", "wonderful", "outstanding", "impressive", "lovely"] negadj = ["terrible", "awful", "dreadful", "disappointing", "clumsy", "broken", "miserable", "useless", "painful", "sloppy"] subj = ["the movie", "this restaurant", "the hotel room", "their support team", "the new phone", "the sequel", "this laptop", "the delivery service"] tailp = ["and I would recommend it to anyone", "worth every rupee", "I left completely satisfied", "easily the best of the year", "it exceeded all my expectations"] tailn = ["and I want a refund", "a total waste of money", "I left extremely frustrated", "easily the worst of the year", "it failed every expectation"] rows = [] for in range(nperclass): rows.

append((f"{rng.choice(subj)} was {rng.choice(posadj)} {rng.choice(tailp)}", 1)) rows.append((f"{rng.choice(subj)} was {rng.choice(negadj)} {rng.choice(tailn)}", 0)) rng.

shuffle(rows) return [r[0] for r in rows], [r[1] for r in rows] TEXTS, LABELS = maketextdataset() DEMOTEXT = "The quick brown fox jumps over the lazy dog near the river bank" def makeaudio(seconds: float = 2.0, sr: int = 16000) -> Tuple[np.

ndarray, int]: """A chirp + harmonics + a little noise = something you can actually hear change.""" t = np.linspace(0, seconds, int(sr seconds), endpoint=False) f = np.linspace(220, 880, t.size) sig = 0.5 np.sin(2 np.pi f t) + 0.2 np.sin(2 np.pi 2 f t) sig += 0.02 np.random.RandomState(SEED).

randn(t.size) env = np.minimum(1.0, np.minimum(t 8, (seconds - t) 8)) return (sig env).astype(np.float32), sr AUDIO, SR = makeaudio() def showgrid(pairs, cols=4, title="", figsizescale=2.9): """pairs: list of (caption, PIL.Image).""" rows = math.ceil(len(pairs) / cols) fig, axes = plt.

subplots(rows, cols, figsize=(cols figsizescale, rows figsizescale)) axes = np.atleast1d(axes).ravel() for ax, (cap, im) in zip(axes, pairs): ax.imshow(im) ax.settitle(cap, fontsize=8) ax.axis("off") for ax in axes[len(pairs):]: ax.axis("off") if title: fig.suptitle(title, fontsize=13, y=1.0) plt.

tightlayout() plt.show() def asstr(out) -> str: """AugLy text augs return str for str input in some transforms, list in others.""" return out[0] if isinstance(out, list) else out print("\n### §2 IMAGE AUGMENTATION + METADATA " + "#" * 38) functionalresult = imaugs.pixelization(DEMOIMG, ratio=0.

25) classresult = imaugs.Pixelization(ratio=0.25, p=1.0)(DEMOIMG) print("functional == class:", np.arrayequal(np.array(functionalresult), np.array(classresult))) IMAGEZOO = { "blur": lambda im, m: imaugs.blur(im, radius=3.0, metadata=m), "brightness": lambda im, m: imaugs.brightness(im, factor=1.

7, metadata=m), "colorjitter": lambda im, m: imaugs.colorjitter(im, brightnessfactor=1.3, contrastfactor=1.4, saturationfactor=1.6, metadata=m), "crop": lambda im, m: imaugs.crop(im, x1=.15, y1=.15, x2=.85, y2=.85, metadata=m), "encodingquality": lambda im, m: imaugs.

encodingquality(im, quality=8, metadata=m), "grayscale": lambda im, m: imaugs.grayscale(im, metadata=m), "hflip": lambda im, m: imaugs.hflip(im, metadata=m), "memeformat": lambda im, m: imaugs.memeformat(im, text="TOP TEXT", captionheight=90, metadata=m), "opacity": lambda im, m: imaugs.

opacity(im, level=0.45, metadata=m), "overlayemoji": lambda im, m: imaugs.overlayemoji(im, opacity=0.9, emojisize=0.35, metadata=m), "overlayscreenshot": lambda im, m: imaugs.overlayontoscreenshot(im, metadata=m), "overlaystripes": lambda im, m: