Spaces:
Sleeping
Sleeping
File size: 11,820 Bytes
3303c2f 00ddcd8 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 |
import torch
import types
import timm
import requests
import random
import yaml
import gradio as gr
from PIL import Image
from timm import create_model
from torchvision import transforms
from timm.data import resolve_data_config
from modelguidedattacks.guides.unguided import Unguided
from timm.data.transforms_factory import create_transform
from modelguidedattacks.cls_models.registry import TimmPretrainModelWrapper
# Download human-readable labels for ImageNet.
IMAGENET_LABELS_URL = "https://storage.googleapis.com/bit_models/ilsvrc2012_wordnet_lemmas.txt"
LABELS = requests.get(IMAGENET_LABELS_URL).text.strip().split("\n")
SORTED_LABELS = sorted(LABELS.copy(), key=lambda s: s.lower())
def get_timm_model(name):
"""Retrieves model from timm library by name with weights loaded.
"""
model = create_model(name,pretrained="true")
transform = create_transform(**resolve_data_config({}, model=model))
model = model.eval()
return model, transform
def create_attacker(model, transform, iterations):
""" Instantiates an QuadAttack Model.
"""
# config_dict = {"cvx_proj_margin" : 0.2,
# "opt_warmup_its": 5}
with open("base_config.yaml") as f:
config_dict = yaml.safe_load(f)
config = types.SimpleNamespace(**config_dict)
attacker = Unguided(TimmPretrainModelWrapper(model, transform,"", "", ""), config, iterations=iterations,
lr=0.002, topk_loss_coef_upper=10)
return attacker
def predict_topk_accuracies(img, k, iters, model_name, desired_labels, button=None, progress=gr.Progress(track_tqdm=True)):
""" Predict the top K results using base model and attacker model.
"""
label_inds = list(range(0,1000)) #label indices
# convert user desired labels to desired inds
desired_inds = [LABELS.index(name) for name in desired_labels]
# remove selected before randomly sampling the rest
for ind in desired_inds:
label_inds.remove(ind)
# fill up user selections to top k results
desired_inds = desired_inds + random.sample(label_inds,k-len(desired_inds))
tensorized_desired_inds = torch.tensor(desired_inds).unsqueeze(0) #[B,K]
model, transform = get_timm_model(model_name)
# Define a transformation to convert PIL image to a tensor
normalization = transforms.Compose([
transform.transforms[-1] # Converts to a PyTorch tensor
])
preprocess = transforms.Compose(
transform.transforms[:-1] # Converts to a PyTorch tensor
)
attacker = create_attacker(model, normalization, iters)
img = img.convert('RGB')
orig_img = img.copy()
orig_img = preprocess(orig_img)
orig_img = orig_img.unsqueeze(0)
img = transform(img).unsqueeze(0)
with torch.no_grad():
outputs = model(img)
attack_outputs, attack_img = attacker(orig_img, tensorized_desired_inds, None)
probabilities = torch.nn.functional.softmax(outputs[0], dim=0)
attacker_probs = torch.nn.functional.softmax(attack_outputs[0], dim=0)
values, indices = torch.topk(probabilities, k)
attack_vals, attack_inds = torch.topk(attacker_probs, k)
attack_img_out = orig_img + attack_img #B C H W
# Convert the PyTorch tensor to a NumPy array
attack_img_out = attack_img_out.squeeze(0) # C H W
attack_img_out = attack_img_out.permute(1, 2, 0).numpy() # H W C
orig_img = orig_img.squeeze(0)
orig_img = orig_img.permute(1, 2, 0).numpy()
attack_img = attack_img.squeeze(0)
attack_img = attack_img.permute(1, 2, 0).numpy()
# Convert the NumPy array to a PIL image
attack_img_out = Image.fromarray((attack_img_out * 255).astype('uint8'))
orig_img = Image.fromarray((orig_img * 255).astype('uint8'))
attack_img = Image.fromarray((attack_img * 255).astype('uint8'))
return (orig_img, attack_img_out, attack_img,{LABELS[i]: v.item() for i, v in zip(indices, values)}, {LABELS[i]: v.item() for i, v in zip(attack_inds, attack_vals)})
def random_fill_classes(desired_labels, k):
label_inds = list(range(0,1000)) #label indices
# convert user desired labels to desired inds
if len(desired_labels) > k:
desired_labels = desired_labels[:k]
desired_inds = [LABELS.index(name) for name in desired_labels]
# remove selected before randomly sampling the rest
for ind in desired_inds:
label_inds.remove(ind)
# fill up user selections to top k results
desired_inds = desired_inds + random.sample(label_inds,k-len(desired_inds))
return [LABELS[ind] for ind in desired_inds]
input_img = gr.Image(type='pil')
top_k_slider = gr.Slider(2, 20, value=10, step=1, label="Top K predictions", info="Choose between 2 and 20")
iteration_slider = gr.Slider(30, 1000, value=60, step=1, label="QuadAttack Iterations", info="Choose how many iterations to optimize using QuadAttack! (Usually <= 60 is enough)")
model_choice_list = gr.Dropdown(
timm.list_models(), value="vit_base_patch16_224", label="timm model name", info="Currently only supporting timm models! See code for models used in paper."
)
desired_labels = gr.Dropdown(
SORTED_LABELS, max_choices=20,filterable=True, multiselect=True, label="Desired Labels for QuadAttack", info="Select classes you wish to output from an attack. \
Classes will be ranked in order listed and randomly filled up to \
K if < K options are selected."
)
button = gr.Button("Randomly fill Top-K attack classes.")
desc = r'<div align="center">Authors: Thomas Paniagua, Ryan Grainger, Tianfu Wu <p><a href="https://arxiv.org/abs/2312.11510">Paper</a><br><a href="https://github.com/thomaspaniagua/quadattack">Code</a></p> </div>'
with gr.Interface(predict_topk_accuracies,
inputs=[input_img,
top_k_slider,
iteration_slider,
model_choice_list,
desired_labels,
button],
outputs=[
gr.Image(type='pil', label="Input Image"),
gr.Image(type='pil', label="Perturbed Image"),
gr.Image(type='pil', label="Added Noise"),
gr.Label(label="Original Top K"),
gr.Label(label="QuadAttack Top K"),
# gr.Image(type='pil', label="Perturbed Image")
],
title='QuadAttack!',
description= desc,
cache_examples=False,
allow_flagging="never",
thumbnail= "quadattack_pipeline.pdf",
examples = [["image_examples/RV.jpeg", 5, 30, "vit_base_patch16_224", None, None
# ["lemon", "plastic_bag", "hay", "tripod", "bell_cote, bell_cot"]
],
# ["image_examples/biker.jpeg", 10, 60, "swinv2_cr_base_224", None, None
# ["hog, pig, grunter, squealer, Sus_scrofa",
# "lesser_panda, red_panda, panda, bear_cat, cat_bear, Ailurus_fulgens",
# "caldron, cauldron", "dowitcher", "water_tower", "quill, quill_pen",
# "balance_beam, beam", "unicycle, monocycle", "pencil_sharpener",
# "puffer, pufferfish, blowfish, globefish"
# ]
# ],
["image_examples/mower.jpeg", 15, 100,"wide_resnet101_2", None , None
# ["washbasin, handbasin, washbowl, lavabo, wash-hand_basin",
# "cucumber, cuke", "bolete", "oboe, hautboy, hautboi", "crane",
# "wolf_spider, hunting_spider", "Norfolk_terrier", "nail", "sidewinder, horned_rattlesnake, Crotalus_cerastes",
# "cannon", "beaker", "Shetland_sheepdog, Shetland_sheep_dog, Shetland",
# "monitor", "restaurant, eating_house, eating_place, eatery", "electric_fan, blower"
# ]
],
# ["image_examples/dog.jpeg", 20, 150, "xcit_small_12_p8_224", None, None
# ["church, church_building", "axolotl, mud_puppy, Ambystoma_mexicanum",
# "Scotch_terrier, Scottish_terrier, Scottie", "black-footed_ferret, ferret, Mustela_nigripes",
# "lab_coat, laboratory_coat", "gyromitra", "grasshopper, hopper", "snail", "tabby, tabby_cat",
# "bell_cote, bell_cot", "Indian_cobra, Naja_naja", "robin, American_robin, Turdus_migratorius",
# "tiger_cat", "book_jacket, dust_cover, dust_jacket, dust_wrapper", "loudspeaker, speaker, speaker_unit, loudspeaker_system, speaker_system",
# "washbasin, handbasin, washbowl, lavabo, wash-hand_basin", "electric_guitar", "armadillo", "ski_mask",
# "convertible"
# ]
# ],
["image_examples/fish.jpeg", 10, 100, "pvt_v2_b0", None, None
# ["ground_beetle, carabid_beetle", "sunscreen, sunblock, sun_blocker", "brass, memorial_tablet, plaque", "Irish_terrier", "head_cabbage", "bathtub, bathing_tub, bath, tub",
# "centipede", "squirrel_monkey, Saimiri_sciureus", "Chihuahua", "hourglass"
# ]
]
]
).queue() as app:
#turn off clear button as it erases globals
for block in app.blocks:
if isinstance(app.blocks[block],gr.Button):
if app.blocks[block].value == "Clear":
app.blocks[block].visible=False
button.click(random_fill_classes, inputs=[desired_labels,top_k_slider], outputs=desired_labels)
if __name__ == "__main__":
app.launch() |