CODEWITHSUNDEEP
pythontensorflowtensorflow-datasets
Sreejith.S
Sreejith.S

Reputation: 31

TF Dataset API for Image augmentation

I am using tf Dataset API to read images and its labels. I like to do multiple image augmentations on images and increase my training data size. What i have done now is like below.

def flip(self, img, lbl):
  image = tf.image.flip_left_right(img)
  return image, lbl

def transpose(self, img, lbl):
  image = tf.image.transpose_image(img)
  return image, lbl

# just read and resize the image.
process_fn = lambda img, lbl: self.read_convert_image(img, lbl, self.args)
flip_fn = lambda img, lbl: self.flip(img,lbl)
transpose_fn = lambda img, lbl: self.transpose(img,lbl)

train_set = self.train_set.repeat()
train_set = train_set.shuffle(args.batch_size)
train_set = train_set.map(process_fn)

fliped_data = train_set.map(flip_fn)
transpose_data = train_set.map(transpose_fn)

train_set = train_set.concatenate(fliped_data)
train_set = train_set.concatenate(transpose_data)

train_set = train_set.batch(args.batch_size)
iterator = train_set.make_one_shot_iterator()

images, labels = iterator.get_next()

Is there a better way to do multiple augmentations. The problem with above approach is if i add more augmentation function , that many map and concatenate is required.

Thank You

Upvotes: 1

Views: 2921

Answers (2)

MPękalski
MPękalski

Reputation: 7103

If you want to do augmentations yourself, without relying on Keras's ImageDataGenerator you can create a function like img_aug and then use it in your model or in the Dataset API pipeline. The code below is just a pseudocode, but it shows the idea. You define all your transformations, then you have some generic threshold above which you apply a transformation and try to apply them up to X times (in the code below it is 4)

def img_aug(image):
  image = distorted_image

  def h_flip():
    return tf.image.flip_left_right(distorted_image)                
  def v_flip():
    return tf.image.flip_up_down(distorted_image)

  threshold = tf.constant(0.9, dtype=tf.float32)      

  def body(i, distorted_image):
    p_order = tf.random_uniform(shape=[2], minval=0., maxval=1., dtype=tf.float32)
    distorted_image = tf.case({                                      
                               tf.greater(p_order[0], threshold): h_flip,  
                               tf.greater(p_order[1], threshold): v_flip, 
                              }
                              ,default=identity, exclusive=False)
    return (i+1, distorted_image)

  def cond(i, *args):
    return i < 4 # max number of transformations

  parallel_iterations = 1
  tf.while_loop(cond, body, [0,distorted_image], 
                parallel_iterations=parallel_iterations)
  return distorted_image

Upvotes: 3

Uzzal Podder
Uzzal Podder

Reputation: 3205

A simple alternative for Image Augmentation is using Tensorflow implemented Keras which contains easy to use api

It's looks like this 

ImageDataGenerator(rescale=1./255, 
    rotation_range=40,
    width_shift_range=0.2,
    height_shift_range=0.2,
    shear_range=0.2,
    zoom_range = 0.2, 
    horizontal_flip = True)

And you are ready to use the augmented image as much as you want.

Here is an working github code example Conv_net_with_augmentation

Upvotes: 0

Related Questions