How to preprocess images using Tensorflow.data.Dataset?-CodePudding

My goal is to prepare a tf.data.Dateset from a list of image paths and some metadata. I need to preprocess images (flip) based on the metadata, so not every image is preprocessed in the same way.

So far I've tried the following approach:

Build the initial dataset from a list.

    train_ds = Dataset.from_tensor_slices(train_samples)

where train_samples is n x [img_path, label, 'RIGHT' / 'LEFT'] list.

Transform it to the final form using the map function.

def _process_sample(self, sample):
   img_path = join(self.main_folder, sample[0])
   img = load_img(img_path)

   if sample[2] == 'R':
       img = img.transpose(Image.FLIP_LEFT_RIGHT)

   img = img.resize(input_shape)

   input_array = img_to_array(img)
   input_array /= 255.0

   return input_array, sample[1]

train_ds = train_ds.map(self._process_sample)

I know that sample is a Tensor and I cannot just access the value in the graph mode. I also know that I could run the code in the eager mode, but I don't want to lose the performance.

Is there any way to solve this?

CodePudding user response：

Thanks for the hints. I've finally solved it by using tf.Dataset.from_generator(). Here is the working code:

def gen():
    for sample in train_samples:
        img_path = join(self.main_folder, sample[0])
        img = load_img(img_path)

        if sample[2] == 'R':
            img = img.transpose(Image.FLIP_LEFT_RIGHT)

        img = img.resize(input_shape)

        input_array = img_to_array(img)
        input_array /= 255.0

        yield input_array, sample[1]

train_ds = tf.data.Dataset.from_generator(gen, output_types=(tf.float32, tf.int8))