feat: update v2 data augmentation
Browse files- detector/data.py +86 -20
- train.py +9 -2
detector/data.py
CHANGED
@@ -17,19 +17,13 @@ from PIL import Image
|
|
17 |
|
18 |
|
19 |
class RandomColorJitter(object):
|
20 |
-
def __init__(
|
21 |
-
self, brightness=0.5, contrast=0.5, saturation=0.5, hue=0.05, preserve=0.2
|
22 |
-
):
|
23 |
self.brightness = brightness
|
24 |
self.contrast = contrast
|
25 |
self.saturation = saturation
|
26 |
self.hue = hue
|
27 |
-
self.preserve = preserve
|
28 |
|
29 |
def __call__(self, batch):
|
30 |
-
if random.random() < self.preserve:
|
31 |
-
return batch
|
32 |
-
|
33 |
image, label = batch
|
34 |
text_color = label[2:5].clone().view(3, 1, 1)
|
35 |
stroke_color = label[7:10].clone().view(3, 1, 1)
|
@@ -60,14 +54,10 @@ class RandomColorJitter(object):
|
|
60 |
|
61 |
|
62 |
class RandomCrop(object):
|
63 |
-
def __init__(self, crop_factor: float = 0.1
|
64 |
self.crop_factor = crop_factor
|
65 |
-
self.preserve = preserve
|
66 |
|
67 |
def __call__(self, batch):
|
68 |
-
if random.random() < self.preserve:
|
69 |
-
return batch
|
70 |
-
|
71 |
image, label = batch
|
72 |
width, height = image.size
|
73 |
|
@@ -89,15 +79,37 @@ class RandomCrop(object):
|
|
89 |
return image, label
|
90 |
|
91 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
92 |
class FontDataset(Dataset):
|
93 |
def __init__(
|
94 |
self,
|
95 |
path: str,
|
96 |
config_path: str = "configs/font.yml",
|
97 |
regression_use_tanh: bool = False,
|
98 |
-
transforms:
|
99 |
crop_roi_bbox: bool = False,
|
100 |
):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
101 |
self.path = path
|
102 |
self.fonts = load_font_with_exclusion(config_path)
|
103 |
self.regression_use_tanh = regression_use_tanh
|
@@ -109,6 +121,9 @@ class FontDataset(Dataset):
|
|
109 |
]
|
110 |
self.images.sort()
|
111 |
|
|
|
|
|
|
|
112 |
def __len__(self):
|
113 |
return len(self.images)
|
114 |
|
@@ -148,25 +163,71 @@ class FontDataset(Dataset):
|
|
148 |
with open(label_path, "rb") as f:
|
149 |
label: FontLabel = pickle.load(f)
|
150 |
|
151 |
-
if self.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
left, top, width, height = label.bbox
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
153 |
image = TF.crop(image, top, left, height, width)
|
154 |
label.image_width = width
|
155 |
label.image_height = height
|
156 |
|
157 |
-
|
158 |
-
|
159 |
|
160 |
-
# data augmentation
|
161 |
-
if self.transforms:
|
162 |
transform = transforms.Compose(
|
163 |
[
|
164 |
-
RandomColorJitter(),
|
165 |
-
RandomCrop(),
|
|
|
166 |
]
|
167 |
)
|
168 |
image, label = transform((image, label))
|
169 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
170 |
# resize and to tensor
|
171 |
transform = transforms.Compose(
|
172 |
[
|
@@ -176,6 +237,11 @@ class FontDataset(Dataset):
|
|
176 |
)
|
177 |
image = transform(image)
|
178 |
|
|
|
|
|
|
|
|
|
|
|
179 |
# normalize label
|
180 |
if self.regression_use_tanh:
|
181 |
label[2:12] = label[2:12] * 2 - 1
|
|
|
17 |
|
18 |
|
19 |
class RandomColorJitter(object):
|
20 |
+
def __init__(self, brightness=0.5, contrast=0.5, saturation=0.5, hue=0.05):
|
|
|
|
|
21 |
self.brightness = brightness
|
22 |
self.contrast = contrast
|
23 |
self.saturation = saturation
|
24 |
self.hue = hue
|
|
|
25 |
|
26 |
def __call__(self, batch):
|
|
|
|
|
|
|
27 |
image, label = batch
|
28 |
text_color = label[2:5].clone().view(3, 1, 1)
|
29 |
stroke_color = label[7:10].clone().view(3, 1, 1)
|
|
|
54 |
|
55 |
|
56 |
class RandomCrop(object):
|
57 |
+
def __init__(self, crop_factor: float = 0.1):
|
58 |
self.crop_factor = crop_factor
|
|
|
59 |
|
60 |
def __call__(self, batch):
|
|
|
|
|
|
|
61 |
image, label = batch
|
62 |
width, height = image.size
|
63 |
|
|
|
79 |
return image, label
|
80 |
|
81 |
|
82 |
+
class RandomRotate(object):
|
83 |
+
def __init__(self, max_angle: int = 15):
|
84 |
+
self.max_angle = max_angle
|
85 |
+
|
86 |
+
def __call__(self, batch):
|
87 |
+
image, label = batch
|
88 |
+
|
89 |
+
angle = random.uniform(-self.max_angle, self.max_angle)
|
90 |
+
image = TF.rotate(image, angle)
|
91 |
+
label[11] = label[11] + angle / 180
|
92 |
+
return image, label
|
93 |
+
|
94 |
+
|
95 |
class FontDataset(Dataset):
|
96 |
def __init__(
|
97 |
self,
|
98 |
path: str,
|
99 |
config_path: str = "configs/font.yml",
|
100 |
regression_use_tanh: bool = False,
|
101 |
+
transforms: str = None,
|
102 |
crop_roi_bbox: bool = False,
|
103 |
):
|
104 |
+
"""Font dataset
|
105 |
+
|
106 |
+
Args:
|
107 |
+
path (str): path to the dataset
|
108 |
+
config_path (str, optional): path to font config file. Defaults to "configs/font.yml".
|
109 |
+
regression_use_tanh (bool, optional): whether use tanh as regression normalization. Defaults to False.
|
110 |
+
transforms (str, optional): choose from None, 'v1', 'v2'. Defaults to None.
|
111 |
+
crop_roi_bbox (bool, optional): whether to crop text roi bbox, must be true when transform='v2'. Defaults to False.
|
112 |
+
"""
|
113 |
self.path = path
|
114 |
self.fonts = load_font_with_exclusion(config_path)
|
115 |
self.regression_use_tanh = regression_use_tanh
|
|
|
121 |
]
|
122 |
self.images.sort()
|
123 |
|
124 |
+
if transforms == "v2":
|
125 |
+
assert crop_roi_bbox, "crop_roi_bbox must be true when transform='v2'"
|
126 |
+
|
127 |
def __len__(self):
|
128 |
return len(self.images)
|
129 |
|
|
|
163 |
with open(label_path, "rb") as f:
|
164 |
label: FontLabel = pickle.load(f)
|
165 |
|
166 |
+
if (self.transforms == "v1") or (self.transforms is None):
|
167 |
+
if self.crop_roi_bbox:
|
168 |
+
left, top, width, height = label.bbox
|
169 |
+
image = TF.crop(image, top, left, height, width)
|
170 |
+
label.image_width = width
|
171 |
+
label.image_height = height
|
172 |
+
|
173 |
+
# encode label
|
174 |
+
label = self.fontlabel2tensor(label, label_path)
|
175 |
+
|
176 |
+
# data augmentation
|
177 |
+
if self.transforms is not None:
|
178 |
+
transform = transforms.Compose(
|
179 |
+
[
|
180 |
+
transforms.RandomApply(RandomColorJitter(), p=0.8),
|
181 |
+
transforms.RandomApply(RandomCrop(), p=0.8),
|
182 |
+
]
|
183 |
+
)
|
184 |
+
image, label = transform((image, label))
|
185 |
+
elif self.transforms == "v2":
|
186 |
+
# crop from 30% to 130% of bbox
|
187 |
left, top, width, height = label.bbox
|
188 |
+
|
189 |
+
right = left + width
|
190 |
+
bottom = top + height
|
191 |
+
|
192 |
+
width_delta = width * 0.07
|
193 |
+
height_delta = height * 0.07
|
194 |
+
|
195 |
+
left = max(0, int(left - width_delta))
|
196 |
+
top = max(0, int(top - height_delta))
|
197 |
+
|
198 |
+
right = min(image.width, int(right + width_delta))
|
199 |
+
bottom = min(image.height, int(bottom + height_delta))
|
200 |
+
|
201 |
+
width = right - left
|
202 |
+
height = bottom - top
|
203 |
+
|
204 |
image = TF.crop(image, top, left, height, width)
|
205 |
label.image_width = width
|
206 |
label.image_height = height
|
207 |
|
208 |
+
# encode label
|
209 |
+
label = self.fontlabel2tensor(label, label_path)
|
210 |
|
|
|
|
|
211 |
transform = transforms.Compose(
|
212 |
[
|
213 |
+
transforms.RandomApply(RandomColorJitter(), p=0.8),
|
214 |
+
RandomCrop(crop_factor=0.54),
|
215 |
+
transforms.RandomApply(RandomRotate(), p=0.8),
|
216 |
]
|
217 |
)
|
218 |
image, label = transform((image, label))
|
219 |
|
220 |
+
transform = transforms.Compose(
|
221 |
+
[
|
222 |
+
transforms.RandomApply(
|
223 |
+
transforms.GaussianBlur(random.randint(2, 5), sigma=(0.1, 5.0)),
|
224 |
+
p=0.8,
|
225 |
+
),
|
226 |
+
]
|
227 |
+
)
|
228 |
+
|
229 |
+
image = transform(image)
|
230 |
+
|
231 |
# resize and to tensor
|
232 |
transform = transforms.Compose(
|
233 |
[
|
|
|
237 |
)
|
238 |
image = transform(image)
|
239 |
|
240 |
+
if self.transforms == "v2":
|
241 |
+
# noise
|
242 |
+
if random.random() < 0.9:
|
243 |
+
image = image + torch.randn_like(image) * random.random() * 0.05
|
244 |
+
|
245 |
# normalize label
|
246 |
if self.regression_use_tanh:
|
247 |
label[2:12] = label[2:12] * 2 - 1
|
train.py
CHANGED
@@ -54,6 +54,14 @@ parser.add_argument(
|
|
54 |
action="store_true",
|
55 |
help="Crop ROI bounding box (default: False)",
|
56 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
|
58 |
args = parser.parse_args()
|
59 |
|
@@ -73,7 +81,6 @@ lambda_direction = 0.5
|
|
73 |
lambda_regression = 1.0
|
74 |
|
75 |
regression_use_tanh = False
|
76 |
-
augmentation = True
|
77 |
|
78 |
num_warmup_epochs = 5
|
79 |
num_epochs = 100
|
@@ -90,7 +97,7 @@ data_module = FontDataModule(
|
|
90 |
val_shuffle=False,
|
91 |
test_shuffle=False,
|
92 |
regression_use_tanh=regression_use_tanh,
|
93 |
-
train_transforms=augmentation,
|
94 |
crop_roi_bbox=args.crop_roi_bbox,
|
95 |
)
|
96 |
|
|
|
54 |
action="store_true",
|
55 |
help="Crop ROI bounding box (default: False)",
|
56 |
)
|
57 |
+
parser.add_argument(
|
58 |
+
"-a",
|
59 |
+
"--augmentation",
|
60 |
+
type=str,
|
61 |
+
default=None,
|
62 |
+
choices=["v1", "v2"],
|
63 |
+
help="Augmentation strategy to use (default: None)",
|
64 |
+
)
|
65 |
|
66 |
args = parser.parse_args()
|
67 |
|
|
|
81 |
lambda_regression = 1.0
|
82 |
|
83 |
regression_use_tanh = False
|
|
|
84 |
|
85 |
num_warmup_epochs = 5
|
86 |
num_epochs = 100
|
|
|
97 |
val_shuffle=False,
|
98 |
test_shuffle=False,
|
99 |
regression_use_tanh=regression_use_tanh,
|
100 |
+
train_transforms=args.augmentation,
|
101 |
crop_roi_bbox=args.crop_roi_bbox,
|
102 |
)
|
103 |
|