Bienqq
/
car-detection-bayes
2
1
Fork 0
Code Issues Pull Requests Releases Wiki Activity

updates

This commit is contained in:
Glenn Jocher 2019-07-20 14:54:37 +02:00
parent cb30d60f4e
commit 4816969933
2 changed files with 32 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
train.py
View File

@ -33,11 +33,17 @@ hyp = {'giou': 1.153, # giou loss gain
'cls_pw': 3.05, # cls BCELoss positive_weight 'cls_pw': 3.05, # cls BCELoss positive_weight
'obj': 20.93, # obj loss gain 'obj': 20.93, # obj loss gain
'obj_pw': 2.842, # obj BCELoss positive_weight 'obj_pw': 2.842, # obj BCELoss positive_weight
'iou_t': 0.2759, # iou target-anchor training threshold 'iou_t': 0.2759, # iou training threshold
'lr0': 0.001357, # initial learning rate 'lr0': 0.001357, # initial learning rate
'lrf': -4., # final learning rate = lr0 * (10 ** lrf) 'lrf': -4., # final LambdaLR learning rate = lr0 * (10 ** lrf)
'momentum': 0.916, # SGD momentum 'momentum': 0.916, # SGD momentum
'weight_decay': 0.000572} # optimizer weight decay 'weight_decay': 0.000572, # optimizer weight decay
'hsv_s': 0.5, # image HSV-Saturation augmentation (fraction)
'hsv_v': 0.5, # image HSV-Value augmentation (fraction)
'degrees': 10, # image rotation (+/- deg)
'translate': 0.1, # image translation (+/- fraction)
'scale': 0.1, # image scale (+/- gain)
'shear': 2} # image shear (+/- deg)
# # Training hyperparameters e # # Training hyperparameters e
@ -50,7 +56,7 @@ hyp = {'giou': 1.153, # giou loss gain
# 'obj_pw': 2.634, # obj BCELoss positive_weight # 'obj_pw': 2.634, # obj BCELoss positive_weight
# 'iou_t': 0.273, # iou target-anchor training threshold # 'iou_t': 0.273, # iou target-anchor training threshold
# 'lr0': 0.001542, # initial learning rate # 'lr0': 0.001542, # initial learning rate
# 'lrf': -4., # final learning rate = lr0 * (10 ** lrf) # 'lrf': -4., # final LambdaLR learning rate = lr0 * (10 ** lrf)
# 'momentum': 0.8364, # SGD momentum # 'momentum': 0.8364, # SGD momentum
# 'weight_decay': 0.0008393} # optimizer weight decay # 'weight_decay': 0.0008393} # optimizer weight decay
@ -149,6 +155,7 @@ def train(cfg,
img_size, img_size,
batch_size, batch_size,
augment=True, augment=True,
hyp=hyp, # augmentation hyperparameters
rect=opt.rect) # rectangular training rect=opt.rect) # rectangular training
# Initialize distributed training # Initialize distributed training
@ -375,14 +382,14 @@ if __name__ == '__main__':
# Mutate # Mutate
init_seeds(seed=int(time.time())) init_seeds(seed=int(time.time()))
s = [.15, .15, .15, .15, .15, .15, .15, .15, .15, .00, .05, .10] # fractional sigmas s = [.15, .15, .15, .15, .15, .15, .15, .15, .15, .00, .05, .10, .15, .15, .15, .15, .15, .15] # sigmas
for i, k in enumerate(hyp.keys()): for i, k in enumerate(hyp.keys()):
x = (np.random.randn(1) * s[i] + 1) ** 2.0 # plt.hist(x.ravel(), 300) x = (np.random.randn(1) * s[i] + 1) ** 2.0 # plt.hist(x.ravel(), 300)
hyp[k] *= float(x) # vary by 20% 1sigma hyp[k] *= float(x) # vary by 20% 1sigma
# Clip to limits # Clip to limits
keys = ['lr0', 'iou_t', 'momentum', 'weight_decay'] keys = ['lr0', 'iou_t', 'momentum', 'weight_decay', 'hsv_s', 'hsv_v', 'translate', 'scale']
limits = [(1e-4, 1e-2), (0.00, 0.70), (0.60, 0.95), (0, 0.01)] limits = [(1e-4, 1e-2), (0.00, 0.70), (0.60, 0.95), (0, 0.001), (0, .8), (0, .8), (0, .8), (0, .8)]
for k, v in zip(keys, limits): for k, v in zip(keys, limits):
hyp[k] = np.clip(hyp[k], v[0], v[1]) hyp[k] = np.clip(hyp[k], v[0], v[1])

32
utils/datasets.py
View File

@ -152,7 +152,7 @@ class LoadWebcam: # for inference
class LoadImagesAndLabels(Dataset): # for training/testing class LoadImagesAndLabels(Dataset): # for training/testing
def __init__(self, path, img_size=416, batch_size=16, augment=False, rect=True, image_weights=False): def __init__(self, path, img_size=416, batch_size=16, augment=False, hyp=None, rect=True, image_weights=False):
with open(path, 'r') as f: with open(path, 'r') as f:
img_files = f.read().splitlines() img_files = f.read().splitlines()
self.img_files = [x for x in img_files if os.path.splitext(x)[-1].lower() in img_formats] self.img_files = [x for x in img_files if os.path.splitext(x)[-1].lower() in img_formats]
@ -166,6 +166,7 @@ class LoadImagesAndLabels(Dataset): # for training/testing
self.batch = bi # batch index of image self.batch = bi # batch index of image
self.img_size = img_size self.img_size = img_size
self.augment = augment self.augment = augment
self.hyp = hyp
self.image_weights = image_weights self.image_weights = image_weights
self.rect = False if image_weights else rect self.rect = False if image_weights else rect
@ -271,6 +272,7 @@ class LoadImagesAndLabels(Dataset): # for training/testing
img_path = self.img_files[index] img_path = self.img_files[index]
label_path = self.label_files[index] label_path = self.label_files[index]
hyp = self.hyp
# Load image # Load image
img = self.imgs[index] img = self.imgs[index]
@ -289,13 +291,12 @@ class LoadImagesAndLabels(Dataset): # for training/testing
augment_hsv = True augment_hsv = True
if self.augment and augment_hsv: if self.augment and augment_hsv:
# SV augmentation by 50% # SV augmentation by 50%
fraction = 0.50 # must be < 1.0
img_hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV) # hue, sat, val img_hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV) # hue, sat, val
S = img_hsv[:, :, 1].astype(np.float32) # saturation S = img_hsv[:, :, 1].astype(np.float32) # saturation
V = img_hsv[:, :, 2].astype(np.float32) # value V = img_hsv[:, :, 2].astype(np.float32) # value
a = random.uniform(-1, 1) * fraction + 1 a = random.uniform(-1, 1) * hyp['hsv_s'] + 1
b = random.uniform(-1, 1) * fraction + 1 b = random.uniform(-1, 1) * hyp['hsv_v'] + 1
S *= a S *= a
V *= b V *= b
@ -331,7 +332,11 @@ class LoadImagesAndLabels(Dataset): # for training/testing
# Augment image and labels # Augment image and labels
if self.augment: if self.augment:
img, labels = random_affine(img, labels, degrees=(-3, 3), translate=(0.05, 0.05), scale=(0.90, 1.10)) img, labels = random_affine(img, labels,
degrees=hyp['degrees'],
translate=hyp['translate'],
scale=hyp['scale'],
shear=hyp['shear'])
nL = len(labels) # number of labels nL = len(labels) # number of labels
if nL: if nL:
@ -410,8 +415,7 @@ def letterbox(img, new_shape=416, color=(128, 128, 128), mode='auto'):
return img, ratiow, ratioh, dw, dh return img, ratiow, ratioh, dw, dh
def random_affine(img, targets=(), degrees=(-10, 10), translate=(.1, .1), scale=(.9, 1.1), shear=(-2, 2), def random_affine(img, targets=(), degrees=10, translate=.1, scale=.1, shear=10):
borderValue=(128, 128, 128)):
# torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(.1, .1), scale=(.9, 1.1), shear=(-10, 10)) # torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(.1, .1), scale=(.9, 1.1), shear=(-10, 10))
# https://medium.com/uruvideo/dataset-augmentation-with-random-homographies-a8f4b44830d4 # https://medium.com/uruvideo/dataset-augmentation-with-random-homographies-a8f4b44830d4
@ -423,24 +427,24 @@ def random_affine(img, targets=(), degrees=(-10, 10), translate=(.1, .1), scale=
# Rotation and Scale # Rotation and Scale
R = np.eye(3) R = np.eye(3)
a = random.uniform(degrees[0], degrees[1]) a = random.uniform(-degrees, degrees)
# a += random.choice([-180, -90, 0, 90]) # add 90deg rotations to small rotations # a += random.choice([-180, -90, 0, 90]) # add 90deg rotations to small rotations
s = random.uniform(scale[0], scale[1]) s = random.uniform(1 - scale, 1 + scale)
R[:2] = cv2.getRotationMatrix2D(angle=a, center=(img.shape[1] / 2, img.shape[0] / 2), scale=s) R[:2] = cv2.getRotationMatrix2D(angle=a, center=(img.shape[1] / 2, img.shape[0] / 2), scale=s)
# Translation # Translation
T = np.eye(3) T = np.eye(3)
T[0, 2] = random.uniform(-1, 1) * translate[0] * img.shape[0] + border # x translation (pixels) T[0, 2] = random.uniform(-translate, translate) * img.shape[0] + border # x translation (pixels)
T[1, 2] = random.uniform(-1, 1) * translate[1] * img.shape[1] + border # y translation (pixels) T[1, 2] = random.uniform(-translate, translate) * img.shape[1] + border # y translation (pixels)
# Shear # Shear
S = np.eye(3) S = np.eye(3)
S[0, 1] = math.tan(random.uniform(shear[0], shear[1]) * math.pi / 180) # x shear (deg) S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # x shear (deg)
S[1, 0] = math.tan(random.uniform(shear[0], shear[1]) * math.pi / 180) # y shear (deg) S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # y shear (deg)
M = S @ T @ R # Combined rotation matrix. ORDER IS IMPORTANT HERE!! M = S @ T @ R # Combined rotation matrix. ORDER IS IMPORTANT HERE!!
imw = cv2.warpAffine(img, M[:2], dsize=(width, height), flags=cv2.INTER_AREA, imw = cv2.warpAffine(img, M[:2], dsize=(width, height), flags=cv2.INTER_AREA,
borderValue=borderValue) # BGR order borderValue borderValue=(128, 128, 128)) # BGR order borderValue
# Return warped points also # Return warped points also
if len(targets) > 0: if len(targets) > 0: