[1] [Handwritten Chinese Font Generation with Collaborative Stroke Refinement]之數據增長實現

[1] [Handwritten Chinese Font Generation with Collaborative Stroke Refinement]之數據增長實現

1.原理

《Handwritten Chinese Font Generation with Collaborative Stroke Refinement》arxiv.org/abs/1904.13268

文中網絡架構圖如下：

圖中的online zoom-augmentation是實現網絡訓練時，對輸入數據進行水平和垂直方向上的transform

2.編碼實現

def hv_transform(self, image):
        '''
        '''
        orig_shp = image.shape

        image = tf.cast(image, tf.float32)

        rd = tf.random.uniform([])
        less_than_025 = tf.less(rd, tf.constant(0.25))
        less_than_050 = tf.less(rd, tf.constant(0.50))
        less_than_075 = tf.less(rd, tf.constant(0.75))

        def func_less_than_025():
            x = tf.image.resize_images(image, [orig_shp[0]//2, orig_shp[1]])

            h_off = tf.random.uniform([1], minval=0, maxval=orig_shp[0] - orig_shp[0]//2, dtype=tf.int32)
            x     = self.pad_to_bounding_box(
                                    x,
                                    offset_height  = h_off, 
                                    offset_width   = tf.constant([0], dtype=tf.int32), 
                                    target_height  = orig_shp[0], 
                                    target_width   = orig_shp[1])
            return x

        def func_less_than_050():
            x = tf.image.resize_images(image, [orig_shp[0], orig_shp[1]//2])
            
            w_off = tf.random.uniform([1], minval=0, maxval=orig_shp[1] - orig_shp[1]//2, dtype=tf.int32)
            x     = self.pad_to_bounding_box(
                                    x,
                                    offset_height  = tf.constant([0], dtype=tf.int32),
                                    offset_width   = w_off,
                                    target_height  = orig_shp[0], 
                                    target_width   = orig_shp[1])
            return x

        def func_less_than_075():
            x = tf.image.resize_images(image, [orig_shp[0]//2, orig_shp[1]//2])
            
            h_off = tf.random.uniform([1], minval=0, maxval=orig_shp[0] - orig_shp[0]//2, dtype=tf.int32)
            w_off = tf.random.uniform([1], minval=0, maxval=orig_shp[1] - orig_shp[1]//2, dtype=tf.int32)
            x     = self.pad_to_bounding_box(
                                    x,
                                    offset_height  = h_off,
                                    offset_width   = w_off,
                                    target_height  = orig_shp[0], 
                                    target_width   = orig_shp[1])
            return x

        def func_less_than_100():
            return image

        x = tf.case({less_than_025: func_less_than_025, less_than_050: func_less_than_050, less_than_075: func_less_than_075},
                default=func_less_than_100, exclusive=False)
        return x

• 上面的代碼實現是基於TensorFlow 1.13.1
• TensorFlow編碼時，通過random實現隨機參數值的數據增長，注意通過tf.reshape()告訴TF當前數據的shape，否則會出現一些很難校正的錯誤