feat: add data augmentation
Browse files- detector/data.py +117 -14
- train.py +2 -0
detector/data.py
CHANGED
@@ -5,20 +5,102 @@ from . import config
|
|
5 |
|
6 |
import math
|
7 |
import os
|
|
|
8 |
import pickle
|
9 |
import torch
|
10 |
import torchvision.transforms as transforms
|
|
|
11 |
from typing import List, Dict, Tuple
|
12 |
from torch.utils.data import Dataset, DataLoader
|
13 |
from pytorch_lightning import LightningDataModule
|
14 |
from PIL import Image
|
15 |
|
16 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
class FontDataset(Dataset):
|
18 |
-
def __init__(
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
self.path = path
|
20 |
self.fonts = load_font_with_exclusion(config_path)
|
21 |
self.regression_use_tanh = regression_use_tanh
|
|
|
22 |
|
23 |
self.images = [
|
24 |
os.path.join(path, f) for f in os.listdir(path) if f.endswith(".jpg")
|
@@ -51,9 +133,6 @@ class FontDataset(Dataset):
|
|
51 |
out[7:10] = out[2:5]
|
52 |
out[10] = label.line_spacing / label.image_width
|
53 |
out[11] = label.angle / 180.0 + 0.5
|
54 |
-
|
55 |
-
if self.regression_use_tanh:
|
56 |
-
out[2:12] = out[2:12] * 2 - 1
|
57 |
|
58 |
return out
|
59 |
|
@@ -62,6 +141,25 @@ class FontDataset(Dataset):
|
|
62 |
image_path = self.images[index]
|
63 |
image = Image.open(image_path).convert("RGB")
|
64 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
65 |
transform = transforms.Compose(
|
66 |
[
|
67 |
transforms.Resize((config.INPUT_SIZE, config.INPUT_SIZE)),
|
@@ -70,13 +168,9 @@ class FontDataset(Dataset):
|
|
70 |
)
|
71 |
image = transform(image)
|
72 |
|
73 |
-
#
|
74 |
-
|
75 |
-
|
76 |
-
label: FontLabel = pickle.load(f)
|
77 |
-
|
78 |
-
# encode label
|
79 |
-
label = self.fontlabel2tensor(label, label_path)
|
80 |
|
81 |
return image, label
|
82 |
|
@@ -91,6 +185,9 @@ class FontDataModule(LightningDataModule):
|
|
91 |
train_shuffle: bool = True,
|
92 |
val_shuffle: bool = False,
|
93 |
test_shuffle: bool = False,
|
|
|
|
|
|
|
94 |
regression_use_tanh: bool = False,
|
95 |
**kwargs,
|
96 |
):
|
@@ -99,9 +196,15 @@ class FontDataModule(LightningDataModule):
|
|
99 |
self.train_shuffle = train_shuffle
|
100 |
self.val_shuffle = val_shuffle
|
101 |
self.test_shuffle = test_shuffle
|
102 |
-
self.train_dataset = FontDataset(
|
103 |
-
|
104 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
105 |
|
106 |
def get_train_num_iter(self, num_device: int) -> int:
|
107 |
return math.ceil(
|
|
|
5 |
|
6 |
import math
|
7 |
import os
|
8 |
+
import random
|
9 |
import pickle
|
10 |
import torch
|
11 |
import torchvision.transforms as transforms
|
12 |
+
import torchvision.transforms.functional as TF
|
13 |
from typing import List, Dict, Tuple
|
14 |
from torch.utils.data import Dataset, DataLoader
|
15 |
from pytorch_lightning import LightningDataModule
|
16 |
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)
|
36 |
+
|
37 |
+
brightness = random.uniform(1 - self.brightness, 1 + self.brightness)
|
38 |
+
image = TF.adjust_brightness(image, brightness)
|
39 |
+
text_color = TF.adjust_brightness(text_color, brightness)
|
40 |
+
stroke_color = TF.adjust_brightness(stroke_color, brightness)
|
41 |
+
|
42 |
+
contrast = random.uniform(1 - self.contrast, 1 + self.contrast)
|
43 |
+
image = TF.adjust_contrast(image, contrast)
|
44 |
+
text_color = TF.adjust_contrast(text_color, contrast)
|
45 |
+
stroke_color = TF.adjust_contrast(stroke_color, contrast)
|
46 |
+
|
47 |
+
saturation = random.uniform(1 - self.saturation, 1 + self.saturation)
|
48 |
+
image = TF.adjust_saturation(image, saturation)
|
49 |
+
text_color = TF.adjust_saturation(text_color, saturation)
|
50 |
+
stroke_color = TF.adjust_saturation(stroke_color, saturation)
|
51 |
+
|
52 |
+
hue = random.uniform(-self.hue, self.hue)
|
53 |
+
image = TF.adjust_hue(image, hue)
|
54 |
+
text_color = TF.adjust_hue(text_color, hue)
|
55 |
+
stroke_color = TF.adjust_hue(stroke_color, hue)
|
56 |
+
|
57 |
+
label[2:5] = text_color.view(3)
|
58 |
+
label[7:10] = stroke_color.view(3)
|
59 |
+
return image, label
|
60 |
+
|
61 |
+
|
62 |
+
class RandomCrop(object):
|
63 |
+
def __init__(self, crop_factor: float = 0.1, preserve: float = 0.2):
|
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 |
+
|
74 |
+
# use random value to decide scaling factor on x and y axis
|
75 |
+
random_height = random.random() * self.crop_factor
|
76 |
+
random_width = random.random() * self.crop_factor
|
77 |
+
# use random value again to decide scaling factor for 4 borders
|
78 |
+
random_top = random.random() * random_height
|
79 |
+
random_left = random.random() * random_width
|
80 |
+
# calculate new width and height and position
|
81 |
+
top = int(random_top * height)
|
82 |
+
left = int(random_left * width)
|
83 |
+
height = int(height - random_height * height)
|
84 |
+
width = int(width - random_width * width)
|
85 |
+
# crop image
|
86 |
+
image = TF.crop(image, top, left, height, width)
|
87 |
+
|
88 |
+
label[[5, 6, 10]] = label[[5, 6, 10]] * (1 - random_height)
|
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: bool = False,
|
99 |
+
):
|
100 |
self.path = path
|
101 |
self.fonts = load_font_with_exclusion(config_path)
|
102 |
self.regression_use_tanh = regression_use_tanh
|
103 |
+
self.transforms = transforms
|
104 |
|
105 |
self.images = [
|
106 |
os.path.join(path, f) for f in os.listdir(path) if f.endswith(".jpg")
|
|
|
133 |
out[7:10] = out[2:5]
|
134 |
out[10] = label.line_spacing / label.image_width
|
135 |
out[11] = label.angle / 180.0 + 0.5
|
|
|
|
|
|
|
136 |
|
137 |
return out
|
138 |
|
|
|
141 |
image_path = self.images[index]
|
142 |
image = Image.open(image_path).convert("RGB")
|
143 |
|
144 |
+
# Load label
|
145 |
+
label_path = image_path.replace(".jpg", ".bin")
|
146 |
+
with open(label_path, "rb") as f:
|
147 |
+
label: FontLabel = pickle.load(f)
|
148 |
+
|
149 |
+
# encode label
|
150 |
+
label = self.fontlabel2tensor(label, label_path)
|
151 |
+
|
152 |
+
# data augmentation
|
153 |
+
if self.transforms:
|
154 |
+
transform = transforms.Compose(
|
155 |
+
[
|
156 |
+
RandomColorJitter(),
|
157 |
+
RandomCrop(),
|
158 |
+
]
|
159 |
+
)
|
160 |
+
image, label = transform((image, label))
|
161 |
+
|
162 |
+
# resize and to tensor
|
163 |
transform = transforms.Compose(
|
164 |
[
|
165 |
transforms.Resize((config.INPUT_SIZE, config.INPUT_SIZE)),
|
|
|
168 |
)
|
169 |
image = transform(image)
|
170 |
|
171 |
+
# normalize label
|
172 |
+
if self.regression_use_tanh:
|
173 |
+
label[2:12] = label[2:12] * 2 - 1
|
|
|
|
|
|
|
|
|
174 |
|
175 |
return image, label
|
176 |
|
|
|
185 |
train_shuffle: bool = True,
|
186 |
val_shuffle: bool = False,
|
187 |
test_shuffle: bool = False,
|
188 |
+
train_transforms: bool = False,
|
189 |
+
val_transforms: bool = False,
|
190 |
+
test_transforms: bool = False,
|
191 |
regression_use_tanh: bool = False,
|
192 |
**kwargs,
|
193 |
):
|
|
|
196 |
self.train_shuffle = train_shuffle
|
197 |
self.val_shuffle = val_shuffle
|
198 |
self.test_shuffle = test_shuffle
|
199 |
+
self.train_dataset = FontDataset(
|
200 |
+
train_path, config_path, regression_use_tanh, train_transforms
|
201 |
+
)
|
202 |
+
self.val_dataset = FontDataset(
|
203 |
+
val_path, config_path, regression_use_tanh, val_transforms
|
204 |
+
)
|
205 |
+
self.test_dataset = FontDataset(
|
206 |
+
test_path, config_path, regression_use_tanh, test_transforms
|
207 |
+
)
|
208 |
|
209 |
def get_train_num_iter(self, num_device: int) -> int:
|
210 |
return math.ceil(
|
train.py
CHANGED
@@ -31,6 +31,7 @@ lambda_direction = 0.5
|
|
31 |
lambda_regression = 1.0
|
32 |
|
33 |
regression_use_tanh = True
|
|
|
34 |
|
35 |
num_warmup_epochs = 1
|
36 |
num_epochs = 100
|
@@ -47,6 +48,7 @@ data_module = FontDataModule(
|
|
47 |
val_shuffle=False,
|
48 |
test_shuffle=False,
|
49 |
regression_use_tanh=regression_use_tanh,
|
|
|
50 |
)
|
51 |
|
52 |
num_iters = data_module.get_train_num_iter(num_device) * num_epochs
|
|
|
31 |
lambda_regression = 1.0
|
32 |
|
33 |
regression_use_tanh = True
|
34 |
+
augmentation = True
|
35 |
|
36 |
num_warmup_epochs = 1
|
37 |
num_epochs = 100
|
|
|
48 |
val_shuffle=False,
|
49 |
test_shuffle=False,
|
50 |
regression_use_tanh=regression_use_tanh,
|
51 |
+
train_transforms=augmentation,
|
52 |
)
|
53 |
|
54 |
num_iters = data_module.get_train_num_iter(num_device) * num_epochs
|