image_classification_project
kaggle图像分类比赛项目，提供了项目bsaeline.
Usage
- git clone https://github.com/HarleysZhang/image_classification_project.git
- Configure the environment and change data drectory into my data directory form
Python the main.py
data direcrory
├── dataset
│ ├── sample_submission.csv
│ ├── test
│ ├── train
│ └── train.csv
├── __init__.py
├── logs
│ ├── models
│ └── training_log.csv
├── main.py
├── model.py
└── submitiMet Collection 2019 - FGVC6
Data_Generator
Here is code:
# Artwork dataset generator
class ArtworkDataset(Sequence):
"""input data generator sequence
# Argument:
@label_path: The path of train_label.csv, file
@batch_size: The size of input_data's batch, int
@target_size: The image size, tuple
@mode: default is to 'train', str
@aug: default is to True, bool
# Returns:
A batch of input_data sequence
"""
def __init__(self, label_path, batch_size, target_size, num_classes, mode='train', aug=True, one_hot=False):
# 初始化类实例参数
self.label_path = label_path
# self.df_train = df_train
self.batch_size = batch_size
self.target_size = target_size
self.num_classes = num_classes
self.mode = mode
self.aug = aug
self.one_hot = one_hot
if isfile(self.label_path):
self.df_train = pd.read_csv(self.label_path)
self.train_dataset_info = []
columns = list(self.df_train.columns)
for name, labels in zip(self.df_train[columns[0]], self.df_train[columns[1]].str.split(' ')):
self.train_dataset_info.append({
'id': name,
'labels': np.array([int(label) for label in labels])})
self.train_dataset_info = np.array(self.train_dataset_info)
else:
print('The train_labels.csv is not exist!')
# split data into train, valid
self.indexes = np.arange(self.train_dataset_info.shape[0])
self.train_indexes, self.valid_indexes = train_test_split(self.indexes, test_size=0.1, random_state=8)
# self.data = []
if self.mode == "train":
self.data = self.train_dataset_info[self.train_indexes]
if self.mode == "val":
self.data = self.train_dataset_info[self.valid_indexes]
def __len__(self):
"""Number of batch in the Sequence.
"""
return len(self.data) // self.batch_size
def __getitem__(self, index):
"""Gets batch at position `index`.
"""
assert self.target_size[2] == 3
start = self.batch_size * index
size = min(self.data.shape[0] - start, self.batch_size)
# dataset_info = shuffle(dataset_info)
batch_images = []
X_train_batch = self.data[start:start + size]
# print(X_train_batch.shape[0])
batch_labels = np.zeros((len(X_train_batch), self.num_classes))
assert size == X_train_batch.shape[0]
for i in range(size):
image = self.load_image(X_train_batch[i]['id'])
if self.aug:
image = self.augment_img(image)
batch_images.append(preprocess_input(image))
batch_labels[i][X_train_batch[i]['labels']] = 1
assert len(batch_images) == X_train_batch.shape[0]
return np.array(batch_images), batch_labels
def augment_img(self, image):
"""
Return the array of augment image
:param image: image id
:return: image array
"""
seq = iaa.Sequential([
iaa.OneOf([
iaa.Affine(rotate=90),
iaa.Affine(rotate=180),
iaa.Affine(rotate=270),
iaa.Fliplr(0.5),
iaa.Flipud(0.5),
iaa.Multiply((0.8, 1.2), per_channel=0.5),
iaa.Affine(shear=(-5, 5)), # shear by -5 to +5 degrees
iaa.Affine(rotate=(-5, 5)), # rotate by -5 to +5 degrees
])], random_order=True)
image_aug = seq.augment_image(image)
return image_aug
def load_image(self, x):
# 将图像尺寸固定为target_size
try:
x_file = self.expand_path(x)
img = cv2.imread(x_file) # cv2读取图片速度比pillow快
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # cv2读取通道顺序为BGR，所以要转换成RGB
img = cv2.resize(img, self.target_size[:2])
# print(img.shape)
# Normalize to zero mean and unit variance
# img -= np.mean(img, keepdims=True)
# img /= np.std(img, keepdims=True) + K.epsilon()
img = img.astype(np.uint8)
except Exception as e:
print(e)
return img
def expand_path(self, id):
# 根据图像id,获取文件完整路径
if isfile(os.path.join(TRAIN, id + '.png')):
return os.path.join(TRAIN, id + '.png')
if isfile(os.path.join(TEST, id + '.png')):
return os.path.join(TEST, id + '.png')
return idHistopathologic Cancer Detection
Data_Generator
Here are data generator code:
# Cancer dataset generator
class CancerDataset(Sequence):
"""input data generator, a sequence object
# Argument:
@label_path: The path of train_label.csv, file
@batch_size: The size of input_data's batch, int
@target_size: The image size, tuple
@mode: default is to 'train', str
@aug: default is to True, bool
# Returns:
A batch of input_data sequence
"""
def __init__(self, label_path, batch_size, target_size, mode='train', aug=True, one_hot=False):
# 初始化类实例参数
self.label_path = label_path
# self.df_train = df_train
self.batch_size = batch_size
self.target_size = target_size
# self.num_class = num_class
self.mode = mode
self.aug = aug
self.one_hot = one_hot
if isfile(self.label_path):
self.df_train = pd.read_csv(self.label_path)
self.id_label_map = {k: v for k, v in zip(self.df_train.id.values, self.df_train.label.values)}
else:
print('The train_labels.csv is not exist!')
self.train_id, self.val_id = train_test_split(self.df_train['id'], test_size=0.15,
random_state=8, shuffle=True)
self.data = []
if self.mode == "train":
self.data = self.train_id
self.data = [x for x in self.train_id]
if self.mode == "val":
self.data = self.val_id
self.data = [x for x in self.val_id]
def __len__(self):
"""Number of batch in the Sequence.
"""
return len(self.data) // self.batch_size
def __getitem__(self, index):
"""Gets batch at position `index`.
"""
start = self.batch_size * index
size = min(len(self.data) - start, self.batch_size)
X = []
Y = []
for i in range(size):
img = self.resize_image(self.data[start + i])
img = self.augment_img(img)
img = img.astype(np.uint8)
label = self.id_label_map[self.data[start + i]]
X.append(img)
Y.append(label)
# Y = [self.id_label_map[x] for x in self.data[start: start + size]]
X = [preprocess_input(x) for x in X]
return np.array(X), np.array(Y)
def augment_img(self, image):
"""
Return the array of augment image
:param image: image id
:return: image array
"""
seq = iaa.Sequential([
iaa.OneOf([
iaa.Affine(rotate=90),
iaa.Affine(rotate=180),
iaa.Affine(rotate=270),
iaa.Fliplr(0.5),
iaa.Flipud(0.5),
iaa.Multiply((0.8, 1.2), per_channel=0.5),
iaa.Affine(shear=(-5, 5)), # shear by -5 to +5 degrees
iaa.Affine(rotate=(-5, 5)), # rotate by -5 to +5 degrees
])], random_order=True)
image_aug = seq.augment_image(image)
return image_aug
def resize_image(self, x):
# 将图像尺寸固定为target_size
try:
x_file = self.expand_path(x)
img = cv2.imread(x_file) # cv2读取图片速度比pillow快
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # cv2读取通道顺序为BGR，所以要转换成RGB
img = cv2.resize(img, self.target_size[:2])
# Normalize to zero mean and unit variance
# img -= np.mean(img, keepdims=True)
# img /= np.std(img, keepdims=True) + K.epsilon()
img = img.astype(np.uint8)
except Exception as e:
print(e)
return img
def expand_path(self, id):
# 根据图像id,获取文件完整路径
if isfile(os.path.join(TRAIN, id + '.tif')):
return os.path.join(TRAIN, id + '.tif')
if isfile(os.path.join(TEST, id + '.tif')):
return os.path.join(TEST, id + '.tif')
return id