1.首先對給的數據進行劃分,類型爲每個類單獨放在一個文件夾中
import json
import shutil
import os
from glob import glob
from tqdm import tqdm
# 此文件的作用是創建每個類的文件夾,以及根據給出來的Json中已經做好的分類,對數據進行對號入座劃分。
# 加載json文件得出一個字典,然後根據Key值來提取每個文件到相應的文件夾中,(注意去除了不合理數據)
try:
for i in range(0,59):
os.mkdir("./data/train/" + str(i))
except:
pass
file_train = json.load(open("./data/temp/labels/AgriculturalDisease_train_annotations.json","r",encoding="utf-8"))
file_val = json.load(open("./data/temp/labels/AgriculturalDisease_validation_annotations.json","r",encoding="utf-8"))
file_list = file_train + file_val
for file in tqdm(file_list):
filename = file["image_id"]
origin_path = "./data/temp/images/" + filename
ids = file["disease_class"]
if ids == 44:
continue
if ids == 45:
continue
if ids > 45:
ids = ids -2
save_path = "./data/train/" + str(ids) + "/"
shutil.copy(origin_path,save_path)
2.獲取增強數據集類的定義
# 數據增強的多種方式,使用自定義的方法。調用只需在dataloader.py文件中的get_item函數中調用類自身參數
# transforms,transforms中集合了compose,compose中列出詳細所使用的增強方式。
from __future__ import division
import cv2
import numpy as np
from numpy import random
import math
from sklearn.utils import shuffle
# 常用的增強方式幾乎都在這裏,只需在compose中列出類名即可
__all__ = ['Compose','RandomHflip', 'RandomUpperCrop', 'Resize', 'UpperCrop', 'RandomBottomCrop',
"RandomErasing",'BottomCrop', 'Normalize', 'RandomSwapChannels', 'RandomRotate',
'RandomHShift',"CenterCrop","RandomVflip",'ExpandBorder', 'RandomResizedCrop',
'RandomDownCrop', 'DownCrop', 'ResizedCrop',"FixRandomRotate"]
# 組合
# “隨機翻轉”,“隨機頂部切割”,“調整大小”,“上切割”,“隨機底部切割”、
# “隨機擦除”,“底部切割”,“正則化”,“隨機交換頻道”,“隨機旋轉”,
# “隨機HShift”,“中央切割”,“隨機Vflip”,“擴展邊界”,“隨機調整切割”,
# “隨機下降”,“下降切割”, “調整切割”,“固定隨機化”。
# 每個增強方式類需要調用普通方法描述如下:
def rotate_nobound(image, angle, center=None, scale=1.):
(h, w) = image.shape[:2]
# if the center is None, initialize it as the center of
# the image
if center is None:
center = (w // 2, h // 2)
# perform the rotation
M = cv2.getRotationMatrix2D(center, angle, scale)
rotated = cv2.warpAffine(image, M, (w, h))
return rotated
def scale_down(src_size, size):
w, h = size
sw, sh = src_size
if sh < h:
w, h = float(w * sh) / h, sh
if sw < w:
w, h = sw, float(h * sw) / w
return int(w), int(h)
def fixed_crop(src, x0, y0, w, h, size=None):
out = src[y0:y0 + h, x0:x0 + w]
if size is not None and (w, h) != size:
out = cv2.resize(out, (size[0], size[1]), interpolation=cv2.INTER_CUBIC)
return out
# 固定隨機旋轉
class FixRandomRotate(object):
def __init__(self, angles=[0,90,180,270], bound=False):
self.angles = angles
self.bound = bound
def __call__(self,img):
do_rotate = random.randint(0, 4)
angle=self.angles[do_rotate]
if self.bound:
img = rotate_bound(img, angle)
else:
img = rotate_nobound(img, angle)
return img
def center_crop(src, size):
h, w = src.shape[0:2]
new_w, new_h = scale_down((w, h), size)
x0 = int((w - new_w) / 2)
y0 = int((h - new_h) / 2)
out = fixed_crop(src, x0, y0, new_w, new_h, size)
return out
def bottom_crop(src, size):
h, w = src.shape[0:2]
new_w, new_h = scale_down((w, h), size)
x0 = int((w - new_w) / 2)
y0 = int((h - new_h) * 0.75)
out = fixed_crop(src, x0, y0, new_w, new_h, size)
return out
def rotate_bound(image, angle):
# grab the dimensions of the image and then determine the
# center
h, w = image.shape[:2]
(cX, cY) = (w // 2, h // 2)
M = cv2.getRotationMatrix2D((cX, cY), angle, 1.0)
cos = np.abs(M[0, 0])
sin = np.abs(M[0, 1])
# compute the new bounding dimensions of the image
nW = int((h * sin) + (w * cos))
nH = int((h * cos) + (w * sin))
# adjust the rotation matrix to take into account translation
M[0, 2] += (nW / 2) - cX
M[1, 2] += (nH / 2) - cY
rotated = cv2.warpAffine(image, M, (nW, nH))
return rotated
# 常用增強方式,以類的方式體現:
# 將多個transform組合起來使用
crop切割 filp旋轉
class Compose(object):
def __init__(self, transforms):
self.transforms = transforms
def __call__(self, img):
for t in self.transforms:
img = t(img)
return img
class RandomRotate(object):
def __init__(self, angles, bound=False):
self.angles = angles
self.bound = bound
def __call__(self,img):
do_rotate = random.randint(0, 2)
if do_rotate:
angle = np.random.uniform(self.angles[0], self.angles[1])
if self.bound:
img = rotate_bound(img, angle)
else:
img = rotate_nobound(img, angle)
return img
class RandomBrightness(object):
def __init__(self, delta=10):
assert delta >= 0
assert delta <= 255
self.delta = delta
def __call__(self, image):
if random.randint(2):
delta = random.uniform(-self.delta, self.delta)
image = (image + delta).clip(0.0, 255.0)
# print('RandomBrightness,delta ',delta)
return image
class RandomContrast(object):
def __init__(self, lower=0.9, upper=1.05):
self.lower = lower
self.upper = upper
assert self.upper >= self.lower, "contrast upper must be >= lower."
assert self.lower >= 0, "contrast lower must be non-negative."
# expects float image
def __call__(self, image):
if random.randint(2):
alpha = random.uniform(self.lower, self.upper)
# print('contrast:', alpha)
image = (image * alpha).clip(0.0,255.0)
return image
class RandomSaturation(object):
def __init__(self, lower=0.8, upper=1.2):
self.lower = lower
self.upper = upper
assert self.upper >= self.lower, "contrast upper must be >= lower."
assert self.lower >= 0, "contrast lower must be non-negative."
def __call__(self, image):
if random.randint(2):
alpha = random.uniform(self.lower, self.upper)
image[:, :, 1] *= alpha
# print('RandomSaturation,alpha',alpha)
return image
class RandomHue(object):
def __init__(self, delta=18.0):
assert delta >= 0.0 and delta <= 360.0
self.delta = delta
def __call__(self, image):
if random.randint(2):
alpha = random.uniform(-self.delta, self.delta)
image[:, :, 0] += alpha
image[:, :, 0][image[:, :, 0] > 360.0] -= 360.0
image[:, :, 0][image[:, :, 0] < 0.0] += 360.0
# print('RandomHue,alpha:', alpha)
return image
class ConvertColor(object):
def __init__(self, current='BGR', transform='HSV'):
self.transform = transform
self.current = current
def __call__(self, image):
if self.current == 'BGR' and self.transform == 'HSV':
image = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)
elif self.current == 'HSV' and self.transform == 'BGR':
image = cv2.cvtColor(image, cv2.COLOR_HSV2BGR)
else:
raise NotImplementedError
return image
class RandomSwapChannels(object):
def __call__(self, img):
if np.random.randint(2):
order = np.random.permutation(3)
return img[:,:,order]
return img
class RandomCrop(object):
def __init__(self, size):
self.size = size
def __call__(self, image):
h, w, _ = image.shape
new_w, new_h = scale_down((w, h), self.size)
if w == new_w:
x0 = 0
else:
x0 = random.randint(0, w - new_w)
if h == new_h:
y0 = 0
else:
y0 = random.randint(0, h - new_h)
out = fixed_crop(image, x0, y0, new_w, new_h, self.size)
return out
class RandomResizedCrop(object):
def __init__(self, size,scale=(0.49, 1.0), ratio=(1., 1.)):
self.size = size
self.scale = scale
self.ratio = ratio
def __call__(self,img):
if random.random() < 0.2:
return cv2.resize(img,self.size)
h, w, _ = img.shape
area = h * w
d=1
for attempt in range(10):
target_area = random.uniform(self.scale[0], self.scale[1]) * area
aspect_ratio = random.uniform(self.ratio[0], self.ratio[1])
new_w = int(round(math.sqrt(target_area * aspect_ratio)))
new_h = int(round(math.sqrt(target_area / aspect_ratio)))
if random.random() < 0.5:
new_h, new_w = new_w, new_h
if new_w < w and new_h < h:
x0 = random.randint(0, w - new_w)
y0 = (random.randint(0, h - new_h))//d
out = fixed_crop(img, x0, y0, new_w, new_h, self.size)
return out
# Fallback
return center_crop(img, self.size)
class DownCrop():
def __init__(self, size, select, scale=(0.36,0.81)):
self.size = size
self.scale = scale
self.select = select
def __call__(self,img, attr_idx):
if attr_idx not in self.select:
return img, attr_idx
if attr_idx == 0:
self.scale=(0.64,1.0)
h, w, _ = img.shape
area = h * w
s = (self.scale[0]+self.scale[1])/2.0
target_area = s * area
new_w = int(round(math.sqrt(target_area)))
new_h = int(round(math.sqrt(target_area)))
if new_w < w and new_h < h:
dw = w-new_w
x0 = int(0.5*dw)
y0 = h-new_h
out = fixed_crop(img, x0, y0, new_w, new_h, self.size)
return out, attr_idx
# Fallback
return center_crop(img, self.size), attr_idx
class ResizedCrop(object):
def __init__(self, size, select,scale=(0.64, 1.0), ratio=(3. / 4., 4. / 3.)):
self.size = size
self.scale = scale
self.ratio = ratio
self.select = select
def __call__(self,img, attr_idx):
if attr_idx not in self.select:
return img, attr_idx
h, w, _ = img.shape
area = h * w
d=1
if attr_idx == 2:
self.scale=(0.36,0.81)
d=2
if attr_idx == 0:
self.scale=(0.81,1.0)
target_area = (self.scale[0]+self.scale[1])/2.0 * area
# aspect_ratio = random.uniform(self.ratio[0], self.ratio[1])
new_w = int(round(math.sqrt(target_area)))
new_h = int(round(math.sqrt(target_area)))
# if random.random() < 0.5:
# new_h, new_w = new_w, new_h
if new_w < w and new_h < h:
x0 = (w - new_w)//2
y0 = (h - new_h)//d//2
out = fixed_crop(img, x0, y0, new_w, new_h, self.size)
# cv2.imshow('{}_img'.format(idx2attr_map[attr_idx]), img)
# cv2.imshow('{}_crop'.format(idx2attr_map[attr_idx]), out)
#
# cv2.waitKey(0)
return out, attr_idx
# Fallback
return center_crop(img, self.size), attr_idx
class RandomHflip(object):
def __call__(self, image):
if random.randint(2):
return cv2.flip(image, 1)
else:
return image
class RandomVflip(object):
def __call__(self, image):
if random.randint(2):
return cv2.flip(image, 0)
else:
return image
class Hflip(object):
def __init__(self,doHflip):
self.doHflip = doHflip
def __call__(self, image):
if self.doHflip:
return cv2.flip(image, 1)
else:
return image
class CenterCrop(object):
def __init__(self, size):
self.size = size
def __call__(self, image):
return center_crop(image, self.size)
class UpperCrop():
def __init__(self, size, scale=(0.09, 0.64)):
self.size = size
self.scale = scale
def __call__(self,img):
h, w, _ = img.shape
area = h * w
s = (self.scale[0]+self.scale[1])/2.0
target_area = s * area
new_w = int(round(math.sqrt(target_area)))
new_h = int(round(math.sqrt(target_area)))
if new_w < w and new_h < h:
dw = w-new_w
x0 = int(0.5*dw)
y0 = 0
out = fixed_crop(img, x0, y0, new_w, new_h, self.size)
return out
# Fallback
return center_crop(img, self.size)
class RandomUpperCrop(object):
def __init__(self, size, select, scale=(0.09, 0.64), ratio=(3. / 4., 4. / 3.)):
self.size = size
self.scale = scale
self.ratio = ratio
self.select = select
def __call__(self,img, attr_idx):
if random.random() < 0.2:
return img, attr_idx
if attr_idx not in self.select:
return img, attr_idx
h, w, _ = img.shape
area = h * w
for attempt in range(10):
s = random.uniform(self.scale[0], self.scale[1])
d = 0.1 + (0.3 - 0.1) / (self.scale[1] - self.scale[0]) * (s - self.scale[0])
target_area = s * area
aspect_ratio = random.uniform(self.ratio[0], self.ratio[1])
new_w = int(round(math.sqrt(target_area * aspect_ratio)))
new_h = int(round(math.sqrt(target_area / aspect_ratio)))
# new_w = int(round(math.sqrt(target_area)))
# new_h = int(round(math.sqrt(target_area)))
if new_w < w and new_h < h:
dw = w-new_w
x0 = random.randint(int((0.5-d)*dw), int((0.5+d)*dw)+1)
y0 = (random.randint(0, h - new_h))//10
out = fixed_crop(img, x0, y0, new_w, new_h, self.size)
return out, attr_idx
# Fallback
return center_crop(img, self.size), attr_idx
class RandomDownCrop(object):
def __init__(self, size, select, scale=(0.36, 0.81), ratio=(3. / 4., 4. / 3.)):
self.size = size
self.scale = scale
self.ratio = ratio
self.select = select
def __call__(self,img, attr_idx):
if random.random() < 0.2:
return img, attr_idx
if attr_idx not in self.select:
return img, attr_idx
if attr_idx == 0:
self.scale=(0.64,1.0)
h, w, _ = img.shape
area = h * w
for attempt in range(10):
s = random.uniform(self.scale[0], self.scale[1])
d = 0.1 + (0.3 - 0.1) / (self.scale[1] - self.scale[0]) * (s - self.scale[0])
target_area = s * area
aspect_ratio = random.uniform(self.ratio[0], self.ratio[1])
new_w = int(round(math.sqrt(target_area * aspect_ratio)))
new_h = int(round(math.sqrt(target_area / aspect_ratio)))
#
# new_w = int(round(math.sqrt(target_area)))
# new_h = int(round(math.sqrt(target_area)))
if new_w < w and new_h < h:
dw = w-new_w
x0 = random.randint(int((0.5-d)*dw), int((0.5+d)*dw)+1)
y0 = (random.randint((h - new_h)*9//10, h - new_h))
out = fixed_crop(img, x0, y0, new_w, new_h, self.size)
# cv2.imshow('{}_img'.format(idx2attr_map[attr_idx]), img)
# cv2.imshow('{}_crop'.format(idx2attr_map[attr_idx]), out)
#
# cv2.waitKey(0)
return out, attr_idx
# Fallback
return center_crop(img, self.size), attr_idx
class RandomHShift(object):
def __init__(self, select, scale=(0.0, 0.2)):
self.scale = scale
self.select = select
def __call__(self,img, attr_idx):
if attr_idx not in self.select:
return img, attr_idx
do_shift_crop = random.randint(0, 2)
if do_shift_crop:
h, w, _ = img.shape
min_shift = int(w*self.scale[0])
max_shift = int(w*self.scale[1])
shift_idx = random.randint(min_shift, max_shift)
direction = random.randint(0,2)
if direction:
right_part = img[:, -shift_idx:, :]
left_part = img[:, :-shift_idx, :]
else:
left_part = img[:, :shift_idx, :]
right_part = img[:, shift_idx:, :]
img = np.concatenate((right_part, left_part), axis=1)
# Fallback
return img, attr_idx
class RandomBottomCrop(object):
def __init__(self, size, select, scale=(0.4, 0.8)):
self.size = size
self.scale = scale
self.select = select
def __call__(self,img, attr_idx):
if attr_idx not in self.select:
return img, attr_idx
h, w, _ = img.shape
area = h * w
for attempt in range(10):
s = random.uniform(self.scale[0], self.scale[1])
d = 0.25 + (0.45 - 0.25) / (self.scale[1] - self.scale[0]) * (s - self.scale[0])
target_area = s * area
new_w = int(round(math.sqrt(target_area)))
new_h = int(round(math.sqrt(target_area)))
if new_w < w and new_h < h:
dw = w-new_w
dh = h - new_h
x0 = random.randint(int((0.5-d)*dw), min(int((0.5+d)*dw)+1,dw))
y0 = (random.randint(max(0,int(0.8*dh)-1), dh))
out = fixed_crop(img, x0, y0, new_w, new_h, self.size)
return out, attr_idx
# Fallback
return bottom_crop(img, self.size), attr_idx
class BottomCrop():
def __init__(self, size, select, scale=(0.4, 0.8)):
self.size = size
self.scale = scale
self.select = select
def __call__(self,img, attr_idx):
if attr_idx not in self.select:
return img, attr_idx
h, w, _ = img.shape
area = h * w
s = (self.scale[0]+self.scale[1])/3.*2.
target_area = s * area
new_w = int(round(math.sqrt(target_area)))
new_h = int(round(math.sqrt(target_area)))
if new_w < w and new_h < h:
dw = w-new_w
dh = h-new_h
x0 = int(0.5*dw)
y0 = int(0.9*dh)
out = fixed_crop(img, x0, y0, new_w, new_h, self.size)
return out, attr_idx
# Fallback
return bottom_crop(img, self.size), attr_idx
class Resize(object):
def __init__(self, size, inter=cv2.INTER_CUBIC):
self.size = size
self.inter = inter
def __call__(self, image):
return cv2.resize(image, (self.size[0], self.size[0]), interpolation=self.inter)
class ExpandBorder(object):
def __init__(self, mode='constant', value=255, size=(336,336), resize=False):
self.mode = mode
self.value = value
self.resize = resize
self.size = size
def __call__(self, image):
h, w, _ = image.shape
if h > w:
pad1 = (h-w)//2
pad2 = h - w - pad1
if self.mode == 'constant':
image = np.pad(image, ((0, 0), (pad1, pad2), (0, 0)),
self.mode, constant_values=self.value)
else:
image = np.pad(image,((0,0), (pad1, pad2),(0,0)), self.mode)
elif h < w:
pad1 = (w-h)//2
pad2 = w-h - pad1
if self.mode == 'constant':
image = np.pad(image, ((pad1, pad2),(0, 0), (0, 0)),
self.mode,constant_values=self.value)
else:
image = np.pad(image, ((pad1, pad2), (0, 0), (0, 0)),self.mode)
if self.resize:
image = cv2.resize(image, (self.size[0], self.size[0]),interpolation=cv2.INTER_LINEAR)
return image
class AstypeToInt():
def __call__(self, image, attr_idx):
return image.clip(0,255.0).astype(np.uint8), attr_idx
class AstypeToFloat():
def __call__(self, image, attr_idx):
return image.astype(np.float32), attr_idx
import matplotlib.pyplot as plt
class Normalize(object):
def __init__(self,mean, std):
'''
:param mean: RGB order
:param std: RGB order
'''
self.mean = np.array(mean).reshape(3,1,1)
self.std = np.array(std).reshape(3,1,1)
def __call__(self, image):
'''
:param image: (H,W,3) RGB
:return:
'''
# plt.figure(1)
# plt.imshow(image)
# plt.show()
return (image.transpose((2, 0, 1)) / 255. - self.mean) / self.std
class RandomErasing(object):
def __init__(self, select,EPSILON=0.5,sl=0.02, sh=0.09, r1=0.3, mean=[0.485, 0.456, 0.406]):
self.EPSILON = EPSILON
self.mean = mean
self.sl = sl
self.sh = sh
self.r1 = r1
self.select = select
def __call__(self, img,attr_idx):
if attr_idx not in self.select:
return img,attr_idx
if random.uniform(0, 1) > self.EPSILON:
return img,attr_idx
for attempt in range(100):
area = img.shape[1] * img.shape[2]
target_area = random.uniform(self.sl, self.sh) * area
aspect_ratio = random.uniform(self.r1, 1 / self.r1)
h = int(round(math.sqrt(target_area * aspect_ratio)))
w = int(round(math.sqrt(target_area / aspect_ratio)))
if w <= img.shape[2] and h <= img.shape[1]:
x1 = random.randint(0, img.shape[1] - h)
y1 = random.randint(0, img.shape[2] - w)
if img.shape[0] == 3:
# img[0, x1:x1+h, y1:y1+w] = random.uniform(0, 1)
# img[1, x1:x1+h, y1:y1+w] = random.uniform(0, 1)
# img[2, x1:x1+h, y1:y1+w] = random.uniform(0, 1)
img[0, x1:x1 + h, y1:y1 + w] = self.mean[0]
img[1, x1:x1 + h, y1:y1 + w] = self.mean[1]
img[2, x1:x1 + h, y1:y1 + w] = self.mean[2]
# img[:, x1:x1+h, y1:y1+w] = torch.from_numpy(np.random.rand(3, h, w))
else:
img[0, x1:x1 + h, y1:y1 + w] = self.mean[1]
# img[0, x1:x1+h, y1:y1+w] = torch.from_numpy(np.random.rand(1, h, w))
return img,attr_idx
return img,attr_idx
if __name__ == '__main__':
import matplotlib.pyplot as plt
class FSAug(object):
def __init__(self):
self.augment = Compose([
AstypeToFloat(),
# RandomHShift(scale=(0.,0.2),select=range(8)),
# RandomRotate(angles=(-20., 20.), bound=True),
ExpandBorder(select=range(8), mode='symmetric'),# symmetric
# Resize(size=(336, 336), select=[ 2, 7]),
AstypeToInt()
])
def __call__(self, spct,attr_idx):
return self.augment(spct,attr_idx)
trans = FSAug()
img_path = '/media/gserver/data/FashionAI/round2/train/Images/coat_length_labels/0b6b4a2146fc8616a19fcf2026d61d50.jpg'
img = cv2.cvtColor(cv2.imread(img_path),cv2.COLOR_BGR2RGB)
img_trans,_ = trans(img,5)
# img_trans2,_ = trans(img,6)
plt.figure()
plt.subplot(221)
plt.imshow(img)
plt.subplot(222)
plt.imshow(img_trans)
# plt.subplot(223)
# plt.imshow(img_trans2)
# plt.imshow(img_trans2)
plt.show()
方式二: 用於線下增強數據,採用的方法是
- 高斯噪聲
- 亮度變化
- 左右翻轉
- 上下翻轉
- 色彩抖動
- 對化
- 銳度變化
from PIL import Image,ImageEnhance,ImageFilter,ImageOps import os import shutil import numpy as np import cv2 import random from skimage.util import random_noise from skimage import exposure image_number = 0 raw_path = "./data/train/" new_path = "./aug/train/" # 加高斯噪聲 def addNoise(img): ''' 注意:輸出的像素是[0,1]之間,所以乘以5得到[0,255]之間 ''' return random_noise(img, mode='gaussian', seed=13, clip=True)*255 def changeLight(img): rate = random.uniform(0.5, 1.5) # print(rate) img = exposure.adjust_gamma(img, rate) #大於1爲調暗,小於1爲調亮;1.05 return img try: for i in range(59): os.makedirs(new_path + os.sep + str(i)) except: pass for raw_dir_name in range(59): raw_dir_name = str(raw_dir_name) saved_image_path = new_path + raw_dir_name+"/" raw_image_path = raw_path + raw_dir_name+"/" if not os.path.exists(saved_image_path): os.mkdir(saved_image_path) raw_image_file_name = os.listdir(raw_image_path) raw_image_file_path = [] for i in raw_image_file_name: raw_image_file_path.append(raw_image_path+i) for x in raw_image_file_path: img = Image.open(x) cv_image = cv2.imread(x) # 高斯噪聲 gau_image = addNoise(cv_image) # 隨機改變 light = changeLight(cv_image) light_and_gau = addNoise(light) cv2.imwrite(saved_image_path + "gau_" + os.path.basename(x),gau_image) cv2.imwrite(saved_image_path + "light_" + os.path.basename(x),light) cv2.imwrite(saved_image_path + "gau_light" + os.path.basename(x),light_and_gau) #img = img.resize((800,600)) #1.翻轉 img_flip_left_right = img.transpose(Image.FLIP_LEFT_RIGHT) img_flip_top_bottom = img.transpose(Image.FLIP_TOP_BOTTOM) #2.旋轉 #img_rotate_90 = img.transpose(Image.ROTATE_90) #img_rotate_180 = img.transpose(Image.ROTATE_180) #img_rotate_270 = img.transpose(Image.ROTATE_270) #img_rotate_90_left = img_flip_left_right.transpose(Image.ROTATE_90) #img_rotate_270_left = img_flip_left_right.transpose(Image.ROTATE_270) #3.亮度 #enh_bri = ImageEnhance.Brightness(img) #brightness = 1.5 #image_brightened = enh_bri.enhance(brightness) #4.色彩 #enh_col = ImageEnhance.Color(img) #color = 1.5 #image_colored = enh_col.enhance(color) #5.對比度 enh_con = ImageEnhance.Contrast(img) contrast = 1.5 image_contrasted = enh_con.enhance(contrast) #6.銳度 #enh_sha = ImageEnhance.Sharpness(img) #sharpness = 3.0 #image_sharped = enh_sha.enhance(sharpness) #保存 img.save(saved_image_path + os.path.basename(x)) img_flip_left_right.save(saved_image_path + "left_right_" + os.path.basename(x)) img_flip_top_bottom.save(saved_image_path + "top_bottom_" + os.path.basename(x)) #img_rotate_90.save(saved_image_path + "rotate_90_" + os.path.basename(x)) #img_rotate_180.save(saved_image_path + "rotate_180_" + os.path.basename(x)) #img_rotate_270.save(saved_image_path + "rotate_270_" + os.path.basename(x)) #img_rotate_90_left.save(saved_image_path + "rotate_90_left_" + os.path.basename(x)) #img_rotate_270_left.save(saved_image_path + "rotate_270_left_" + os.path.basename(x)) #image_brightened.save(saved_image_path + "brighted_" + os.path.basename(x)) #image_colored.save(saved_image_path + "colored_" + os.path.basename(x)) image_contrasted.save(saved_image_path + "contrasted_" + os.path.basename(x)) #image_sharped.save(saved_image_path + "sharped_" + os.path.basename(x)) image_number += 1 print("convert pictur" "es :%s size:%s mode:%s" % (image_number, img.size, img.mode))
加載數據的類(自定義繼承)
- 與pytorch中的加載數據類差不多,只是多了自己的某些功能。
-
from torch.utils.data import Dataset from torchvision import transforms as T from config import config from PIL import Image from itertools import chain from glob import glob from tqdm import tqdm import random import numpy as np import pandas as pd import os import cv2 import torch #1.set random seed random.seed(config.seed) np.random.seed(config.seed) torch.manual_seed(config.seed) torch.cuda.manual_seed_all(config.seed) #2.define dataset class ZiyiDataset(Dataset): def __init__(self,label_list,transforms=None,train=True,test=False): self.test = test self.train = train imgs = [] if self.test: for index,row in label_list.iterrows(): imgs.append((row["filename"])) self.imgs = imgs else: for index,row in label_list.iterrows(): imgs.append((row["filename"],row["label"])) self.imgs = imgs if transforms is None: if self.test or not train: self.transforms = T.Compose([ T.Resize((config.img_weight,config.img_height)), T.ToTensor(), T.Normalize(mean = [0.485,0.456,0.406], std = [0.229,0.224,0.225])]) else: self.transforms = T.Compose([ T.Resize((config.img_weight,config.img_height)), T.RandomRotation(30), T.RandomHorizontalFlip(), T.RandomVerticalFlip(), T.RandomAffine(45), T.ToTensor(), T.Normalize(mean = [0.485,0.456,0.406], std = [0.229,0.224,0.225])]) else: self.transforms = transforms def __getitem__(self,index): if self.test: filename = self.imgs[index] img = Image.open(filename) img = self.transforms(img) return img,filename else: filename,label = self.imgs[index] img = Image.open(filename) img = self.transforms(img) return img,label def __len__(self): return len(self.imgs) def collate_fn(batch): imgs = [] label = [] for sample in batch: imgs.append(sample[0]) label.append(sample[1]) return torch.stack(imgs, 0), \ label def get_files(root,mode): #for test if mode == "test": files = [] for img in os.listdir(root): files.append(root + img) files = pd.DataFrame({"filename":files}) return files elif mode != "test": #for train and val all_data_path,labels = [],[] image_folders = list(map(lambda x:root+x,os.listdir(root))) jpg_image_1 = list(map(lambda x:glob(x+"/*.jpg"),image_folders)) jpg_image_2 = list(map(lambda x:glob(x+"/*.JPG"),image_folders)) all_images = list(chain.from_iterable(jpg_image_1 + jpg_image_2)) print("loading train dataset") for file in tqdm(all_images): all_data_path.append(file) labels.append(int(file.split("/")[-2])) all_files = pd.DataFrame({"filename":all_data_path,"label":labels}) return all_files else: print("check the mode please!")
3.獲取模型
- 獲取模型較爲簡單,單一模型採取pytorch中的預訓練模型,添加所需要的層,進行微調然後遷移學習新數據。
import torchvision import torch.nn.functional as F from torch import nn from config import config def generate_model(): class DenseModel(nn.Module): def __init__(self, pretrained_model): super(DenseModel, self).__init__() self.classifier = nn.Linear(pretrained_model.classifier.in_features, config.num_classes) for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_normal(m.weight) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() elif isinstance(m, nn.Linear): m.bias.data.zero_() self.features = pretrained_model.features self.layer1 = pretrained_model.features._modules['denseblock1'] self.layer2 = pretrained_model.features._modules['denseblock2'] self.layer3 = pretrained_model.features._modules['denseblock3'] self.layer4 = pretrained_model.features._modules['denseblock4'] def forward(self, x): features = self.features(x) out = F.relu(features, inplace=True) out = F.avg_pool2d(out, kernel_size=8).view(features.size(0), -1) out = F.sigmoid(self.classifier(out)) return out return DenseModel(torchvision.models.densenet169(pretrained=True)) def get_net(): #return MyModel(torchvision.models.resnet101(pretrained = True)) model = torchvision.models.resnet50(pretrained = True) #for param in model.parameters(): # param.requires_grad = False # pytorch添加層的方式直接在Model.層名=層具體形式 model.avgpool = nn.AdaptiveAvgPool2d(1) model.fc = nn.Linear(2048,config.num_classes) #添加全連接層以作分類任務,num_classes爲分類個數 return model
4.開始訓練
-
import os import random import time import json import torch import torchvision import numpy as np import pandas as pd import warnings from datetime import datetime from torch import nn,optim from config import config from collections import OrderedDict from torch.autograd import Variable from torch.utils.data import DataLoader from dataset.dataloader import * from sklearn.model_selection import train_test_split,StratifiedKFold from timeit import default_timer as timer from models.model import * from utils import * #1. 設置隨機種子 and cudnn performance random.seed(config.seed) np.random.seed(config.seed) torch.manual_seed(config.seed) torch.cuda.manual_seed_all(config.seed) os.environ["CUDA_VISIBLE_DEVICES"] = config.gpus torch.backends.cudnn.benchmark = True warnings.filterwarnings('ignore') #2. 評估函數,通過Losses,topk的不斷更新來評估模型 def evaluate(val_loader,model,criterion): #2.1 AverageMeter類是Computes and stores the average and current value # 創建三個其對象,以用於評估 losses = AverageMeter() top1 = AverageMeter() top2 = AverageMeter() #2.2 開啓評估模式 and confirm model has been transfered to cuda model.cuda() model.eval() with torch.no_grad(): for i,(input,target) in enumerate(val_loader): input = Variable(input).cuda() target = Variable(torch.from_numpy(np.array(target)).long()).cuda() #target = Variable(target).cuda() #2.2.1 compute output output = model(input) loss = criterion(output,target) #2.2.2 measure accuracy and record loss precision1,precision2 = accuracy(output,target,topk=(1,2)) losses.update(loss.item(),input.size(0)) top1.update(precision1[0],input.size(0)) top2.update(precision2[0],input.size(0)) return [losses.avg,top1.avg,top2.avg] #3. test model on public dataset and save the probability matrix def test(test_loader,model,folds): #3.1 confirm the model converted to cuda # 得出的結果是概率,再用softmax得出最終分類結果 csv_map = OrderedDict({"filename":[],"probability":[]}) model.cuda() model.eval() with open("./submit/baseline.json","w",encoding="utf-8") as f : submit_results = [] for i,(input,filepath) in enumerate(tqdm(test_loader)): # filepath?????? # 通過模型得到輸出概率結果,再用softmax得出預測結果,寫入文件。 #3.2 change everything to cuda and get only basename filepath = [os.path.basename(x) for x in filepath] with torch.no_grad(): image_var = Variable(input).cuda() #3.3.output #print(filepath) #print(input,input.shape) y_pred = model(image_var) #print(y_pred.shape) smax = nn.Softmax(1) smax_out = smax(y_pred) #3.4 save probability to csv files csv_map["filename"].extend(filepath) for output in smax_out: prob = ";".join([str(i) for i in output.data.tolist()]) csv_map["probability"].append(prob) result = pd.DataFrame(csv_map) result["probability"] = result["probability"].map(lambda x : [float(i) for i in x.split(";")]) for index, row in result.iterrows(): # 因爲44,45類刪除,所以預測結果加2 pred_label = np.argmax(row['probability']) if pred_label > 43: pred_label = pred_label + 2 submit_results.append({"image_id":row['filename'],"disease_class":pred_label}) json.dump(submit_results,f,ensure_ascii=False,cls = MyEncoder) #4. more details to build main function def main(): fold = 0 #4.1 mkdirs if not os.path.exists(config.submit): os.mkdir(config.submit) if not os.path.exists(config.weights): os.mkdir(config.weights) if not os.path.exists(config.best_models): os.mkdir(config.best_models) if not os.path.exists(config.logs): os.mkdir(config.logs) if not os.path.exists(config.weights + config.model_name + os.sep +str(fold) + os.sep): os.makedirs(config.weights + config.model_name + os.sep +str(fold) + os.sep) if not os.path.exists(config.best_models + config.model_name + os.sep +str(fold) + os.sep): os.makedirs(config.best_models + config.model_name + os.sep +str(fold) + os.sep) #4.2 get model and optimizer model = get_net() #model = torch.nn.DataParallel(model) model.cuda() #optimizer = optim.SGD(model.parameters(),lr = config.lr,momentum=0.9,weight_decay=config.weight_decay) optimizer = optim.Adam(model.parameters(),lr = config.lr,amsgrad=True,weight_decay=config.weight_decay) criterion = nn.CrossEntropyLoss().cuda() #criterion = FocalLoss().cuda() log = Logger() log.open(config.logs + "log_train.txt",mode="a") log.write("\n----------------------------------------------- [START %s] %s\n\n" % (datetime.now().strftime('%Y-%m-%d %H:%M:%S'), '-' * 51)) #4.3 some parameters for K-fold and restart model start_epoch = 0 best_precision1 = 0 best_precision_save = 0 resume = False #4.4 restart the training process if resume: checkpoint = torch.load(config.best_models + str(fold) + "/model_best.pth.tar") start_epoch = checkpoint["epoch"] fold = checkpoint["fold"] best_precision1 = checkpoint["best_precision1"] model.load_state_dict(checkpoint["state_dict"]) optimizer.load_state_dict(checkpoint["optimizer"]) #4.5 get files and split for K-fold dataset #4.5.1 read files train_ = get_files(config.train_data,"train") #val_data_list = get_files(config.val_data,"val") test_files = get_files(config.test_data,"test") """ #4.5.2 split split_fold = StratifiedKFold(n_splits=3) folds_indexes = split_fold.split(X=origin_files["filename"],y=origin_files["label"]) folds_indexes = np.array(list(folds_indexes)) fold_index = folds_indexes[fold] #4.5.3 using fold index to split for train data and val data train_data_list = pd.concat([origin_files["filename"][fold_index[0]],origin_files["label"][fold_index[0]]],axis=1) val_data_list = pd.concat([origin_files["filename"][fold_index[1]],origin_files["label"][fold_index[1]]],axis=1) """ train_data_list,val_data_list = train_test_split(train_,test_size = 0.15,stratify=train_["label"]) #4.5.4 load dataset train_dataloader = DataLoader(ZiyiDataset(train_data_list),batch_size=config.batch_size,shuffle=True,collate_fn=collate_fn,pin_memory=True) val_dataloader = DataLoader(ZiyiDataset(val_data_list,train=False),batch_size=config.batch_size,shuffle=True,collate_fn=collate_fn,pin_memory=False) test_dataloader = DataLoader(ZiyiDataset(test_files,test=True),batch_size=1,shuffle=False,pin_memory=False) #scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,"max",verbose=1,patience=3) scheduler = optim.lr_scheduler.StepLR(optimizer,step_size = 10,gamma=0.1) # optim.lr_scheduler 提供了基於多種epoch數目調整學習率的方法 # step_size(整數類型): 調整學習率的步長,每過step_size次,更新一次學習率 # gamma(float 類型):學習率下降的乘數因子 #4.5.5.1 define metrics train_losses = AverageMeter() train_top1 = AverageMeter() train_top2 = AverageMeter() valid_loss = [np.inf,0,0] model.train() #logs log.write('** start training here! **\n') log.write(' |------------ VALID -------------|----------- TRAIN -------------|------Accuracy------|------------|\n') log.write('lr iter epoch | loss top-1 top-2 | loss top-1 top-2 | Current Best | time |\n') log.write('-------------------------------------------------------------------------------------------------------------------------------\n') #4.5.5 train start = timer() for epoch in range(start_epoch,config.epochs): # 一個epoch爲所有數據迭代一次進入模型擬合的過程,其中又分爲batch_size來分批次進行 scheduler.step(epoch) # train #global iter for iter,(input,target) in enumerate(train_dataloader): #4.5.5 switch to continue train process model.train() input = Variable(input).cuda() target = Variable(torch.from_numpy(np.array(target)).long()).cuda() #target = Variable(target).cuda() output = model(input) loss = criterion(output,target) precision1_train,precision2_train = accuracy(output,target,topk=(1,2)) train_losses.update(loss.item(),input.size(0)) train_top1.update(precision1_train[0],input.size(0)) train_top2.update(precision2_train[0],input.size(0)) #backward optimizer.zero_grad() loss.backward() optimizer.step() lr = get_learning_rate(optimizer) print('\r',end='',flush=True) print('%0.4f %5.1f %6.1f | %0.3f %0.3f %0.3f | %0.3f %0.3f %0.3f | %s | %s' % (\ lr, iter/len(train_dataloader) + epoch, epoch, valid_loss[0], valid_loss[1], valid_loss[2], train_losses.avg, train_top1.avg, train_top2.avg,str(best_precision_save), time_to_str((timer() - start),'min')) , end='',flush=True) #evaluate lr = get_learning_rate(optimizer) #evaluate every half epoch valid_loss = evaluate(val_dataloader,model,criterion) is_best = valid_loss[1] > best_precision1 best_precision1 = max(valid_loss[1],best_precision1) try: best_precision_save = best_precision1.cpu().data.numpy() except: pass save_checkpoint({ "epoch":epoch + 1, "model_name":config.model_name, "state_dict":model.state_dict(), "best_precision1":best_precision1, "optimizer":optimizer.state_dict(), "fold":fold, "valid_loss":valid_loss, },is_best,fold) #adjust learning rate #scheduler.step(valid_loss[1]) print("\r",end="",flush=True) log.write('%0.4f %5.1f %6.1f | %0.3f %0.3f %0.3f | %0.3f %0.3f %0.3f | %s | %s' % (\ lr, 0 + epoch, epoch, valid_loss[0], valid_loss[1], valid_loss[2], train_losses.avg, train_top1.avg, train_top2.avg, str(best_precision_save), time_to_str((timer() - start),'min')) ) log.write('\n') time.sleep(0.01) best_model = torch.load(config.best_models + os.sep+config.model_name+os.sep+ str(fold) +os.sep+ 'model_best.pth.tar') model.load_state_dict(best_model["state_dict"]) test(test_dataloader,model,fold) if __name__ =="__main__": main()