我正在尝试实现的体系结构在这里:Patient-data adapted model architecture: ResNet-50。我的图像按如下标签分为文件夹:
root/
├── train/
│ ├── class1/
│ ├── class2/
│ ...
│
└── validation/
├── class1/
├── class2/
...
我还有一个CSV文件,其中包含图像名称,图像标签(一个图像可以具有多个类标签)和其他信息:
+--------+---------------+-------+------+
| File | Labels | Info1 | Info2 |
+-------+---------------+-------+-------+
| 1.png | class1 | 0.512 | 1 |
| 2.png | class2 | 0.4 | 0 |
| 3.png | class1|class2 | 0.64 | 1 |
+-------+---------------+-------+-------+
我的网络模型有两个输入,一个输入将用于处理图像,另一个输入将连接到密集层之前的最后一层:
input_shape = (img_height, img_width, 1)
img_input= Input(input_shape)
vec_input = Input((2,))
res = ZeroPadding2D((3, 3))(img_input)
# Processing ...
res = Flatten()(res)
res = Concatenate()([res, vec_input])
res = Dense(classes, activation='softmax', name='fc' + str(classes))(res)
要获取图像,我正在将ImageDataGenerator与flow_from_directory配合使用,这对于仅获取图像数据而言效果很好:
validation_datagen = ImageDataGenerator(rescale=1. / 255)
validation_generator = validation_datagen.flow_from_directory(
validation_dir,
target_size=(target_size, target_size),
batch_size=batch_size,
class_mode=class_mode,
color_mode=color_mode)
# Similarly for the train data generator ...
# Train the model using above defined data generators
history = model.fit_generator(
train_generator,
epochs=epochs,
validation_data=validation_generator)
现在,我需要将每个图像的附加信息用作模型中的vec_input。我已经看过使用flow_from_dataframe并创建自定义生成器,但是不确定如何执行此操作。我可以通过将图像放置在相同的文件夹中(如果需要)来重组图像,尽管那以后我想我不能使用flow_from_directory。关于如何实现此目标的任何想法?
编辑:
[如果有人需要解决方案,这就是我能想到的:
class CustomSequenceGenerator(Sequence):
def __init__(self, image_dir, csv_file_path, label_path, dim=448, batch_size=8,
n_classes=15, n_channels=1, vec_size=3, shuffle=True):
# Keras generator
self.image_dir = image_dir
self.image_file_list = os.listdir(image_dir)
self.batch_size = batch_size
self.csv_file = pd.read_csv(csv_file_path)
self.n_classes = n_classes
self.dim = dim
self.n_channels = n_channels
self.shuffle = shuffle
self.vec_size = vec_size
self.labels = get_class_labels(label_path)
self.labels_dict = dict(zip(self.labels, range(0, len(self.labels))))
self.csv_file.set_index('File', inplace=True, drop=True)
def __len__(self):
"""It is mandatory to implement it on Keras Sequence"""
return int(np.ceil(len(self.image_file_list) / float(self.batch_size)))
def __getitem__(self, index):
# Generate indexes of the batch
samples = self.image_file_list[index * self.batch_size:(index + 1) * self.batch_size]
x, y = self.__data_generation(samples, index)
return x, y
def __data_generation(self, samples, start_index):
x_batch_image = np.empty((self.batch_size, self.dim, self.dim, self.n_channels))
x_batch_vector = np.empty((self.batch_size, self.vec_size))
y_batch = np.empty((self.batch_size, self.n_classes))
self.csv_file.reindex()
for i, sample in enumerate(samples):
image_file_path = self.image_dir + "/" + sample
image = self.preprocess_image(Image.open(image_file_path), 448)
features, labels = self.preprocess_csv(self.csv_file, sample, self.labels_dict, self.n_classes)
x_batch_image[i] = image
x_batch_vector[i] = features
y_batch[i] = labels
return [x_batch_image, x_batch_vector], y_batch
我认为实现此目标的最佳方法是实现自定义Sequence
object,并可能继承Sequence
的方法。也许您不需要的是ImageDataGenerator
的所有复杂性(即随机变换,图像保存,插值),在这种情况下,您不需要继承它。