def make_coco_transforms_v2(image_set):
### keep only one scale as 960*608
normalize = T.Compose(
[T.ToTensor(),
T.Normalize([0.408, 0.459, 0.482], [1., 1., 1.])])
print('NVdata Norm rgb: [0.408, 0.459, 0.482], [1., 1., 1.]')
# scales = [608]
re_size = (608, 960)
if image_set == 'train':
print(re_size)
return T.Compose([
T.RandomHorizontalFlip(),
T.RandomSelect(
T.RandomResize(re_size),
T.Compose([
T.RandomResize([400, 500, 600]),
T.RandomSizeCrop(384, 600),
T.RandomResize(re_size),
])),
normalize,
])
if image_set == 'test':
# print("608 960")
# print("800 1333")
return T.Compose([
# T.RandomResize([608], max_size=980), #800 1333, 604 960, not604, should be 608*966 604*960
T.RandomResize((604, 960)),
# T.RandomResize([800], max_size=1333),
normalize,
])
raise ValueError(f'unknown {image_set}')
def make_coco_transforms(image_set):
# MEANS = (104, 117, 123) as bgr
normalize = T.Compose([
T.ToTensor(),
# T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
# mean=(104, 117, 123) ==> (0.40784313725490196 0.4588235294117647 0.4823529411764706) transpose(2,1,0) = [0.482, 0.459, 0.408]
# T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) # coco mean and std
T.Normalize([0.482, 0.459, 0.408], [1., 1., 1.]) # mean=(104, 117, 123) for not for rgb, bgr instead, after transpose [0.482, 0.459, 0.408]
])
print('NVdata Norm: [0.482, 0.459, 0.408], [1., 1., 1.]')
scales = [480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800]
if image_set == 'train':
return T.Compose([
T.RandomHorizontalFlip(),
T.RandomSelect(
T.RandomResize(scales, max_size=1333),
T.Compose([
T.RandomResize([400, 500, 600]),
T.RandomSizeCrop(384, 600),
T.RandomResize(scales, max_size=1333),
])
),
normalize,
])
if image_set == 'test':
return T.Compose([
T.RandomResize([800], max_size=1333),
normalize,
])
raise ValueError(f'unknown {image_set}')
def make_coco_transforms(image_set):
# MEANS = (104, 117, 123) as bgr
normalize = T.Compose([
T.ToTensor(),
# T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
# mean=(104, 117, 123) ==> (0.40784313725490196 0.4588235294117647 0.4823529411764706) transpose(2,1,0) = [0.482, 0.459, 0.408]
# T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) # coco mean and std this shuold be in bgr
## before for 150 epochs
# T.Normalize([0.482, 0.459, 0.408], [1., 1., 1.]) # mean=(104, 117, 123) for not for rgb, bgr instead, after transpose [0.482, 0.459, 0.408]
# New implement from 12.11
# This is rgb already # mean=(104, 117, 123) ==> (0.40784313725490196 0.4588235294117647 0.4823529411764706)
T.Normalize([0.408, 0.459, 0.482], [1., 1., 1.])
])
print('NVdata Norm rgb: [0.408, 0.459, 0.482], [1., 1., 1.]')
scales = [480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800]
# scales = [480, 512, 544, 576, 604, 640, 672, 704, 736, 768, 800]
if image_set == 'train':
return T.Compose([
T.RandomHorizontalFlip(),
T.RandomSelect(
T.RandomResize(scales, max_size=1333),
T.Compose([
T.RandomResize([400, 500, 600]),
T.RandomSizeCrop(384, 600),
T.RandomResize(scales, max_size=1333),
])),
normalize,
])
if image_set == 'test':
# print("604 960")
print("608 966")
# print("800 1333")
return T.Compose([
T.RandomResize(
[608], max_size=980
), #800 1333, 604 960, not604, should be 608*966 604*960
# T.RandomResize([604], max_size=960),
# T.RandomResize([800], max_size=1333),
normalize,
])
raise ValueError(f'unknown {image_set}')