Merge branch 'master' of github.com:SamSamhuns/yolov5_adversarial

adv_patch_gen/README.md

@@ -33,4 +33,4 @@ pip install Cython
 sudo apt-get install python3.8-dev # or any desired python version
 pip install pycocotools # https://github.com/ppwwyyxx/cocoapi
 # pip install git+https://github.com/philferriere/cocoapi.git#subdirectory=PythonAPI # depreciated
-```
+```

adv_patch_gen/configs/README.md

@@ -1,46 +1,47 @@
 # Adversarial Patch Training Config Reference
 
-
-  "image_dir": "data/train/images",
-  "label_dir": "data/train/labels",
-  "val_image_dir": "data/val/images", # epoch freq for running validation run. 1 means validate after every epoch. 0 or null means no val
-  "use_even_odd_images": "all", # (str), ('all', 'even', 'odd'): use images with even/odd numbers in the last char of their filenames
-  "log_dir": "runs/train_adversarial",
-  "tensorboard_port": 8994,
-  "tensorboard_batch_log_interval": 15,
-  "weights_file": "runs/weights/best.pt",
-  "triplet_printfile": "triplets.csv",
-  "device": "cuda:0", # (str): 'cpu' or 'cuda' or 'cuda:0,1,2,3'
-  "use_amp": true,
-  "patch_name": "base",
-  "val_epoch_freq": 100,
-  "patch_save_epoch_freq": 1, # int freq for saving patches. 1 means save after every epoch
-  "model_in_sz": [640, 640], # (int, int): model input height, width
-  "patch_src": "gray", # str: gray random, or path_to_init_patch
-  "patch_img_mode": "RGB", # str: patch channel image mode. Currently RGB * L supported
-  "patch_size": [64, 64], # (int, int): must be (height, width)
-  "objective_class_id": null, # int: class id to target for adv attack. Use null for general attack for all classes
-  "min_pixel_area": null, # int: min pixel area to use for training. Pixel area chosen after resizing to model in size
-  "target_size_frac": 0.3, # float: patch proportion size compared to bbox size. Range also accepted i.e. [0.25, 0.4]
-  "use_mul_add_gau": true, # bool: use mul & add gaussian noise or not to patches
-  "mul_gau_mean": 0.5, # float: mul gaussian noise mean (reduces contrast) mean. Range also accepted i.e. [0.25, 0.4]
-  "mul_gau_std": 0.1, # float: mul gaussian noise std (Adds rand noise)
-  "random_patch_loc": true, # bool: place/translate patches randomly on bbox
-  "x_off_loc": [-0.25, 0.25], # [float, float]: left, right x-axis disp from bbox center
-  "y_off_loc": [-0.25, 0.25], # [float, float]: top, bottom y-axis disp from bbox center
-  "rotate_patches": true, # bool: rotate patches or not
-  "transform_patches": true, # bool: add bightness, contrast and noise transforms to patches or not
-  "patch_pixel_range": [0, 255], # [int, int]: patch pixel range, range is [0, 255], numbers div by 255 in patches
-  "patch_alpha": 1, # float: patch opacity, recommended to set to 1
-  "class_list": ["class1", "class2"],
-  "n_classes": 2,
-  "n_epochs": 300,
-  "max_labels": 48,
-  "start_lr": 0.03,
-  "min_tv_loss": 0.1,
-  "sal_mult": 1.0,
-  "tv_mult": 2.5,
-  "nps_mult": 0.01, # float: Use 0.01 when not using sal. With sal use 0.001
-  "batch_size": 8,
-  "debug_mode": false, # bool: if yes, images with adv drawn saved during each batch
-  "loss_target": "obj * cls" # str: 'obj', 'cls', 'obj * cls'
+```
+ "image_dir": "data/train/images",
+ "label_dir": "data/train/labels",
+ "val_image_dir": "data/val/images", # epoch freq for running validation run. 1 means validate after every epoch. 0 or null means no val
+ "use_even_odd_images": "all", # (str), ('all', 'even', 'odd'): use images with even/odd numbers in the last char of their filenames
+ "log_dir": "runs/train_adversarial",
+ "tensorboard_port": 8994,
+ "tensorboard_batch_log_interval": 15,
+ "weights_file": "runs/weights/best.pt",
+ "triplet_printfile": "triplets.csv",
+ "device": "cuda:0", # (str): 'cpu' or 'cuda' or 'cuda:0,1,2,3'
+ "use_amp": true,
+ "patch_name": "base",
+ "val_epoch_freq": 100,
+ "patch_save_epoch_freq": 1, # int freq for saving patches. 1 means save after every epoch
+ "model_in_sz": [640, 640], # (int, int): model input height, width
+ "patch_src": "gray", # str: gray random, or path_to_init_patch
+ "patch_img_mode": "RGB", # str: patch channel image mode. Currently RGB * L supported
+ "patch_size": [64, 64], # (int, int): must be (height, width)
+ "objective_class_id": null, # int: class id to target for adv attack. Use null for general attack for all classes
+ "min_pixel_area": null, # int: min pixel area to use for training. Pixel area chosen after resizing to model in size
+ "target_size_frac": 0.3, # float: patch proportion size compared to bbox size. Range also accepted i.e. [0.25, 0.4]
+ "use_mul_add_gau": true, # bool: use mul & add gaussian noise or not to patches
+ "mul_gau_mean": 0.5, # float: mul gaussian noise mean (reduces contrast) mean. Range also accepted i.e. [0.25, 0.4]
+ "mul_gau_std": 0.1, # float: mul gaussian noise std (Adds rand noise)
+ "random_patch_loc": true, # bool: place/translate patches randomly on bbox
+ "x_off_loc": [-0.25, 0.25], # [float, float]: left, right x-axis disp from bbox center
+ "y_off_loc": [-0.25, 0.25], # [float, float]: top, bottom y-axis disp from bbox center
+ "rotate_patches": true, # bool: rotate patches or not
+ "transform_patches": true, # bool: add bightness, contrast and noise transforms to patches or not
+ "patch_pixel_range": [0, 255], # [int, int]: patch pixel range, range is [0, 255], numbers div by 255 in patches
+ "patch_alpha": 1, # float: patch opacity, recommended to set to 1
+ "class_list": ["class1", "class2"],
+ "n_classes": 2,
+ "n_epochs": 300,
+ "max_labels": 48,
+ "start_lr": 0.03,
+ "min_tv_loss": 0.1,
+ "sal_mult": 1.0,
+ "tv_mult": 2.5,
+ "nps_mult": 0.01, # float: Use 0.01 when not using sal. With sal use 0.001
+ "batch_size": 8,
+ "debug_mode": false, # bool: if yes, images with adv drawn saved during each batch
+ "loss_target": "obj * cls" # str: 'obj', 'cls', 'obj * cls'
+```

adv_patch_gen/conv_visdrone_2_yolo/README.md

@@ -20,14 +20,16 @@ Annotations for the detections are follows:
 
 `<bbox_left>, <bbox_top>, <bbox_width>, <bbox_height>, <score>, <object_category>, <truncation>, <occlusion>`
 
-  - <bbox_left> The x coordinate of the top-left corner of the predicted bounding box
-  - <bbox_top> The y coordinate of the top-left corner of the predicted object bounding box
-  - <bbox_width> The width in pixels of the predicted object bounding box
-  - <bbox_height> The height in pixels of the predicted object bounding box
-  - <score> The score in the DETECTION file indicates the confidence of the predicted bounding box enclosing an object instance. The score in GROUNDTRUTH file is set to 1 or 0. 1 indicates the bounding box is considered in evaluation, while 0 indicates the bounding box will be ignored.
-  - <object_category> The object category indicates the type of annotated object, (i.e., ignored regions(0), pedestrian(1), people(2), bicycle(3), car(4), van(5), truck(6), tricycle(7), awning-tricycle(8), bus(9), motor(10), others(11))
-  - <truncation> The score in the DETECTION result file should be set to the constant -1.The score in the GROUNDTRUTH file indicates the degree of object parts appears outside a frame (i.e., no truncation = 0 (truncation ratio 0%), and partial truncation = 1 (truncation ratio 1% ~ 50%)).
-  - <occlusion> The score in the DETECTION file should be set to the constant -1. The score in the GROUNDTRUTH file indicates the fraction of objects being occluded (i.e., no occlusion = 0 (occlusion ratio 0%), partial occlusion = 1 (occlusion ratio 1% ~ 50%), and heavy occlusion = 2 (occlusion ratio 50% ~ 100%)).
+```
+ - <bbox_left> The x coordinate of the top-left corner of the predicted bounding box
+ - <bbox_top> The y coordinate of the top-left corner of the predicted object bounding box
+ - <bbox_width> The width in pixels of the predicted object bounding box
+ - <bbox_height> The height in pixels of the predicted object bounding box
+ - <score> The score in the DETECTION file indicates the confidence of the predicted bounding box enclosing an object instance. The score in GROUNDTRUTH file is set to 1 or 0. 1 indicates the bounding box is considered in evaluation, while 0 indicates the bounding box will be ignored.
+ - <object_category> The object category indicates the type of annotated object, (i.e., ignored regions(0), pedestrian(1), people(2), bicycle(3), car(4), van(5), truck(6), tricycle(7), awning-tricycle(8), bus(9), motor(10), others(11))
+ - <truncation> The score in the DETECTION result file should be set to the constant -1.The score in the GROUNDTRUTH file indicates the degree of object parts appears outside a frame (i.e., no truncation = 0 (truncation ratio 0%), and partial truncation = 1 (truncation ratio 1% ~ 50%)).
+ - <occlusion> The score in the DETECTION file should be set to the constant -1. The score in the GROUNDTRUTH file indicates the fraction of objects being occluded (i.e., no occlusion = 0 (occlusion ratio 0%), partial occlusion = 1 (occlusion ratio 1% ~ 50%), and heavy occlusion = 2 (occlusion ratio 50% ~ 100%)).
+```
 
 ## Download VisDrone Dataset
 
@@ -81,4 +83,4 @@ Note: To convert to COCO annotation format refer to <https://github.com/SamSamhu
 
 ## References
 
--  [VisDrone](https://github.com/VisDrone/VisDrone-Dataset)
+- [VisDrone](https://github.com/VisDrone/VisDrone-Dataset)