Auto-format by https://ultralytics.com/actions

adv_patch_gen/conv_visdrone_2_yolo/conv_visdrone_2_yolo_fmt.py

@@ -1,7 +1,8 @@
 """
-Convert VisDrone annotation format to YOLO labels format
-Works for training of YOLOv5 and YOLOv7.
-YOLOv7 requires an additional txt file (Same name as the first parent directory) with paths to the images for the train, val & test splits
+Convert VisDrone annotation format to YOLO labels format Works for training of YOLOv5 and YOLOv7.
+
+YOLOv7 requires an additional txt file (Same name as the first parent directory) with paths to the images for the train,
+val & test splits
 """
 import os
 import os.path as osp

adv_patch_gen/utils/common.py

@@ -1,6 +1,4 @@
-"""
-Common utils
-"""
+"""Common utils."""
 import socket
 from typing import Tuple, Union
 
@@ -30,17 +28,13 @@ class BColors:
 
 
 def is_port_in_use(port: int) -> bool:
- """
- Checks if a port is free for use
- """
+ """Checks if a port is free for use."""
  with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as stream:
  return stream.connect_ex(("localhost", int(port))) == 0
 
 
 def pad_to_square(img: Image, pad_rgb: Tuple[int, int, int] = (127, 127, 127)) -> Image:
- """
- Pads a PIL image to a square with pad_rgb values to the longest side
- """
+ """Pads a PIL image to a square with pad_rgb values to the longest side."""
  w, h = img.size
  if w == h:
  padded_img = img
@@ -57,9 +51,7 @@ def pad_to_square(img: Image, pad_rgb: Tuple[int, int, int] = (127, 127, 127)) -
 
 
 def calc_mean_and_std_err(arr: Union[list, np.ndarray]) -> Tuple[float, float]:
- """ "
- Calculate mean and standard error
- """
+ """" Calculate mean and standard error."""
  mean = np.mean(arr)
  std_err = np.std(arr, ddof=1) / np.sqrt(len(arr))
  return mean, std_err

adv_patch_gen/utils/config_parser.py

@@ -1,25 +1,19 @@
-"""
-Create argparse options for config files
-"""
+"""Create argparse options for config files."""
 import json
 import argparse
 from easydict import EasyDict as edict
 
 
 def load_config_object(cfg_path: str) -> edict:
- """
- Loads a config json and returns a edict object
- """
+ """Loads a config json and returns a edict object."""
  with open(cfg_path, "r", encoding="utf-8") as json_file:
  cfg_dict = json.load(json_file)
 
  return edict(cfg_dict)
 
 
 def get_argparser(desc="Config file load for training adv patches") -> argparse.ArgumentParser:
- """
- Get parser with the default config argument
- """
+ """Get parser with the default config argument."""
  parser = argparse.ArgumentParser(description=desc)
  parser.add_argument(
  "--cfg",

adv_patch_gen/utils/dataset.py

@@ -1,6 +1,5 @@
-"""
-Dataset Class for loading YOLO format datasets where the source data dir has the image and labels subdirs
-where each image must have a corresponding label file with the same name
+"""Dataset Class for loading YOLO format datasets where the source data dir has the image and labels subdirs where each
+image must have a corresponding label file with the same name.
 """
 import glob
 import os.path as osp
@@ -18,7 +17,8 @@
 
 
 class YOLODataset(Dataset):
- """Create a dataset for adversarial-yolt.
+ """
+ Create a dataset for adversarial-yolt.
 
  Attributes:
  image_dir: Directory containing the images of the YOLO format dataset.
@@ -106,10 +106,8 @@ def __getitem__(self, idx):
  return image, label
 
  def pad_and_scale(self, img, lab):
- """
- Pad image and adjust label
- img is a PIL image
- lab is of fmt class x_center y_center width height with normalized coords
+ """Pad image and adjust label img is a PIL image lab is of fmt class x_center y_center width height with
+ normalized coords.
  """
  img_w, img_h = img.size
  if img_w == img_h:
@@ -132,9 +130,7 @@ def pad_and_scale(self, img, lab):
  return padded_img, lab
 
  def pad_label(self, label: torch.Tensor) -> torch.Tensor:
- """
- Pad labels with zeros if fewer labels than max_n_labels present
- """
+ """Pad labels with zeros if fewer labels than max_n_labels present."""
  pad_size = self.max_n_labels - label.shape[0]
  if pad_size > 0:
  padded_lab = F.pad(label, (0, 0, 0, pad_size), value=0)

adv_patch_gen/utils/loss.py

@@ -1,6 +1,4 @@
-"""
-Loss functions used in patch generation
-"""
+"""Loss functions used in patch generation."""
 from typing import Tuple
 
 import torch
@@ -19,9 +17,7 @@ def __init__(self, config):
  self.config = config
 
  def forward(self, output: torch.Tensor):
- """
- output must be of the shape [batch, -1, 5 + num_cls]
- """
+ """Output must be of the shape [batch, -1, 5 + num_cls]"""
  # get values necessary for transformation
  assert output.size(-1) == (5 + self.config.n_classes)
 
@@ -43,7 +39,9 @@ def forward(self, output: torch.Tensor):
 
 
 class SaliencyLoss(nn.Module):
- """Implementation of the colorfulness metric as the saliency loss.
+ """
+ Implementation of the colorfulness metric as the saliency loss.
+
  The smaller the value, the less colorful the image.
  Reference: https://infoscience.epfl.ch/record/33994/files/HaslerS03.pdf
  """

adv_patch_gen/utils/median_pool.py

@@ -4,7 +4,8 @@
 
 
 class MedianPool2d(nn.Module):
- """Median pool (usable as median filter when stride=1) module.
+ """
+ Median pool (usable as median filter when stride=1) module.
 
  Args:
  kernel_size: size of pooling kernel, int or 2-tuple

adv_patch_gen/utils/patch.py

@@ -1,4 +1,4 @@
-""" Modules for creating adversarial object patch """
+"""Modules for creating adversarial object patch."""
 import math
 from typing import Union, Tuple
 

adv_patch_gen/utils/scheduler.py

@@ -17,8 +17,9 @@
 
 
 class GradualWarmupScheduler(_LRScheduler):
- """Sets the learning rate of parameter group to gradually increase for num_epochs from start_lr
- to the original lr set for the optimizer
+ """
+ Sets the learning rate of parameter group to gradually increase for num_epochs from start_lr to the original lr set
+ for the optimizer.
 
  Args:
  optimizer (Optimizer): Wrapped optimizer.

adv_patch_gen/utils/video.py

@@ -10,9 +10,7 @@ def ffmpeg_create_video_from_image_dir(
  title_text: Optional[str] = None,
  framerate: int = 25,
 ) -> None:
- """
- Create a video from images in source_image_dir
- """
+ """Create a video from images in source_image_dir."""
  source = osp.join(source_image_dir, f"*{etxn}")
  cmd = ["ffmpeg", "-y", "-framerate", str(framerate), "-pattern_type", "glob", "-i", source]
  if title_text is not None:
@@ -23,17 +21,13 @@ def ffmpeg_create_video_from_image_dir(
 
 
 def ffmpeg_combine_two_vids(vid1: str, vid2: str, target_video_path: str) -> None:
- """
- Attaches two videos side by side horizontally and saves to target_video_path
- """
+ """Attaches two videos side by side horizontally and saves to target_video_path."""
  cmd = ["ffmpeg", "-y", "-i", vid1, "-i", vid2, "-filter_complex", "hstack", target_video_path]
  output, error = Popen(cmd, universal_newlines=True, stdin=PIPE, stdout=PIPE, stderr=PIPE).communicate()
 
 
 def ffmpeg_combine_three_vids(vid1: str, vid2: str, vid3: str, target_video_path: str) -> None:
- """
- Attaches three videos side by side horizontally and saves to target_video_path
- """
+ """Attaches three videos side by side horizontally and saves to target_video_path."""
  filter_pat = "[1:v][0:v]scale2ref=oh*mdar:ih[1v][0v];[2:v][0v]scale2ref=oh*mdar:ih[2v][0v];[0v][1v][2v]hstack=3,scale='2*trunc(iw/2)':'2*trunc(ih/2)'"
  cmd = ["ffmpeg", "-y", "-i", vid1, "-i", vid2, "-i", vid3, "-filter_complex", filter_pat, target_video_path]
  output, error = Popen(cmd, universal_newlines=True, stdin=PIPE, stdout=PIPE, stderr=PIPE).communicate()

detect_oak.py

@@ -118,7 +118,8 @@ def inference(
  debug: bool = False,
  **kwrags,
 ) -> None:
- """Run Object Detection Application
+ """
+ Run Object Detection Application.
 
  Args:
  weights: str = path to yolov5 model
@@ -234,7 +235,8 @@ def inference_threaded(
  debug: bool = False,
  **kwrags,
 ) -> None:
- """Run Object Detection Application
+ """
+ Run Object Detection Application.
 
  Args:
  weights: str = path to yolov5 model
@@ -319,7 +321,8 @@ def inference_threaded_with_defense(
  cam_fps: int = 20,
  debug: bool = False,
 ) -> None:
- """Run Object Detection Application
+ """
+ Run Object Detection Application.
 
  Args:
  weights: str = path to yolov5 model

explainer.py

@@ -164,9 +164,8 @@ def __init__(self, model):
  self.model = model
 
  def forward(self, x):
- """
- first one is a 3 dim array which contains predictions
- second one is a list of heads with their corresponding predictions
+ """First one is a 3 dim array which contains predictions second one is a list of heads with their corresponding
+ predictions.
  """
  total_prediction, _ = self.model(x)
  return total_prediction

test_patch.py

@@ -1,6 +1,4 @@
-"""
-Testing code for evaluating Adversarial patches against object detection
-"""
+"""Testing code for evaluating Adversarial patches against object detection."""
 import io
 import os
 import os.path as osp
@@ -47,9 +45,7 @@
 
 
 def eval_coco_metrics(anno_json: str, pred_json: str, txt_save_path: str, w_mode: str = "a") -> np.ndarray:
- """
- Compare and eval pred json producing coco metrics
- """
+ """Compare and eval pred json producing coco metrics."""
  anno = COCO(anno_json) # init annotations api
  pred = anno.loadRes(pred_json) # init predictions api
  evaluator = COCOeval(anno, pred, "bbox")
@@ -69,9 +65,7 @@ def eval_coco_metrics(anno_json: str, pred_json: str, txt_save_path: str, w_mode
 
 
 class PatchTester:
- """
- Module for testing patches on dataset against object detection models
- """
+ """Module for testing patches on dataset against object detection models."""
 
  def __init__(self, cfg: edict) -> None:
  self.cfg = cfg
@@ -96,8 +90,9 @@ def calc_asr(
  recompute_asr_all: bool = False,
  ) -> Tuple[float, float, float, float]:
  """
- Calculate attack success rate (How many bounding boxes were hidden from the detector)
- for all predictions and for different bbox areas.
+ Calculate attack success rate (How many bounding boxes were hidden from the detector) for all predictions and
+ for different bbox areas.
+
  Note cls_id is None, misclassifications are ignored and only missing detections are considered attack success.
  Args:
  boxes: torch.Tensor, first pass boxes (gt unpatched boxes) [class, x1, y1, x2, y2]
@@ -172,9 +167,7 @@ def calc_asr(
 
  @staticmethod
  def draw_bbox_on_pil_image(bbox: np.ndarray, padded_img_pil: Image, class_list: List[str]) -> Image:
- """
- Draw bounding box on a PIL image and return said image after drawing
- """
+ """Draw bounding box on a PIL image and return said image after drawing."""
  padded_img_np = np.ascontiguousarray(padded_img_pil)
  label_2_class = dict(enumerate(class_list))
 

train.py

@@ -262,7 +262,8 @@ def train(hyp, opt, device, callbacks): # hyp is path/to/hyp.yaml or hyp dictio
  prefix=colorstr("train: "),
  shuffle=True,
  seed=opt.seed,
- patch_dir=opt.patch_dir)
+ patch_dir=opt.patch_dir,
+ )
  labels = np.concatenate(dataset.labels, 0)
  mlc = int(labels[:, 0].max()) # max label class
  assert mlc < nc, f"Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}"
@@ -544,7 +545,9 @@ def parse_opt(known=False):
  parser.add_argument("--freeze", nargs="+", type=int, default=[0], help="Freeze layers: backbone=10, first3=0 1 2")
  parser.add_argument("--save-period", type=int, default=-1, help="Save checkpoint every x epochs (disabled if < 1)")
  parser.add_argument("--seed", type=int, default=0, help="Global training seed")
- parser.add_argument('--patch_dir', type=str, default='', help='Directory with images to apply as patch augmentation in train imgs')
+ parser.add_argument(
+ "--patch_dir", type=str, default="", help="Directory with images to apply as patch augmentation in train imgs"
+ )
  parser.add_argument("--local_rank", type=int, default=-1, help="Automatic DDP Multi-GPU argument, do not modify")
 
  # Logger arguments

train_patch.py

@@ -1,5 +1,5 @@
 """
-Training code for Adversarial patch training
+Training code for Adversarial patch training.
 
 python train_patch.py --cfg config_json_file
 """
@@ -48,9 +48,7 @@
 
 
 class PatchTrainer:
- """
- Module for training on dataset to generate adv patches
- """
+ """Module for training on dataset to generate adv patches."""
 
  def __init__(self, cfg: edict):
  self.cfg = cfg
@@ -111,9 +109,7 @@ def __init__(self, cfg: edict):
  self.epoch_length = len(self.train_loader)
 
  def init_tensorboard(self, log_dir: str = None, port: int = 6006, run_tb=True):
- """
- Initialize tensorboard with optional name
- """
+ """Initialize tensorboard with optional name."""
  if run_tb:
  while is_port_in_use(port) and port < 65535:
  port += 1
@@ -146,7 +142,7 @@ def generate_patch(self, patch_type: str, pil_img_mode: str = "RGB") -> torch.Te
 
  def read_image(self, path, pil_img_mode: str = "RGB") -> torch.Tensor:
  """
- Read an input image to be used as a patch
+ Read an input image to be used as a patch.
 
  Arguments:
  path: Path to the image to be read.
@@ -157,9 +153,7 @@ def read_image(self, path, pil_img_mode: str = "RGB") -> torch.Tensor:
  return adv_patch_cpu
 
  def train(self) -> None:
- """
- Optimize a patch to generate an adversarial example.
- """
+ """Optimize a patch to generate an adversarial example."""
  # make output dirs
  patch_dir = osp.join(self.cfg.log_dir, "patches")
  os.makedirs(patch_dir, exist_ok=True)
@@ -284,9 +278,7 @@ def train(self) -> None:
  print(f"Total training time {time.time() - start_time:.2f}s")
 
  def val(self, epoch: int, patchfile: str, conf_thresh: float = 0.4, nms_thresh: float = 0.4) -> None:
- """
- Calculates the attack success rate according for the patch with respect to different bounding box areas
- """
+ """Calculates the attack success rate according for the patch with respect to different bounding box areas."""
  # load patch from file
  patch_img = Image.open(patchfile).convert(self.cfg.patch_img_mode)
  patch_img = T.Resize(self.cfg.patch_size)(patch_img)