Keras：图像的不同网格块的不同生成器答案

【问题标题】：Keras: Different generators for different grid patches of an imageKeras：图像的不同网格块的不同生成器
【发布时间】：2019-04-01 12:17:44
【问题描述】：

我有一个庞大的图像数据集，我想将其拆分为一个补丁网格，例如 10 乘以 10 个补丁的网格。对于这些不同的补丁中的每一个，我都想训练一个自动编码器，所以在 10x10 网格的情况下，我将有 100 个自动编码器。

到目前为止，我的解决方案是为每个补丁创建一个新的 ImageDataGenerator。但我认为这太低效了，因为那时所有图像都需要完全加载 100 次（每个自动编码器一次），尽管只需要一个补丁。理论上一次就足够了。有没有更好的方法我看不到？提前致谢！

def crop_to_patch_function(patch_x: int, patch_y: int, grid_size: int):
    def crop_to_patch(img):
        x, y = patch_x*grid_size, patch_y*grid_size
        return img[y:(y+grid_size), x:(x+grid_size), :]
    return crop_to_patch


def patch_generator(patch_x, patch_y, grid_size):
    datagen = ImageDataGenerator(rescale=1/255)
    train_batches_tmp = datagen.flow_from_directory(
        directory=train_data_dir,
        target_size=(img_height, img_width),
        batch_size=batch_size,
        color_mode='rgb',
        class_mode='input',
    )
    while True:
        batch_x, batch_y = next(train_batches_tmp)
        batch_patches = np.zeros((batch_x.shape[0], grid_size, grid_size, 3))
        for i in range(batch_x.shape[0]):
            batch_patches[i] = crop_to_patch_function(patch_x, patch_y, grid_size)(batch_x[i])
        yield (batch_patches, batch_patches)


# batches of patch at pos (2, 4)
patch_x, patch_y = 2, 4
train_patch_batches = patch_generator(patch_x, patch_y, grid_size)

【问题讨论】：

标签： python tensorflow keras

【解决方案1】：

通过预先创建补丁来预处理图像行不通？将它们保存到不同的目录，并将每个 ImageDataGenerator 分配到 100 个目录之一，为每个模型加载数据。

类似：

def images_to_patches(images_list):
    for idx,image in enumerator(images_list):
        for patch_x in range(10):
            for patch_y in range(10):
                //returns the patch image
                patch_img = crop_patch(image, patch_x, patch_y, grid_size)
                img_dir = str(patch_x)+str(patch_y)
                patch_img.save(os.path.join(img_dir,idx))

【讨论】：