Data augmentation

Overview

This tutorial demonstrates data augmentation: a technique to increase the diversity of your training set by applying random (but realistic) transformations such as image rotation. You will learn how to apply data augmentation in two ways. First, you will use Keras Preprocessing Layers. Next, you will use tf.image.

Setup

import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf
import tensorflow_datasets as tfds

from tensorflow.keras import layers

2021-07-31 01:20:29.398577: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudart.so.11.0

Download a dataset

This tutorial uses the tf_flowers dataset. For convenience, download the dataset using TensorFlow Datasets. If you would like to learn about others ways of importing data, see the load images tutorial.

(train_ds, val_ds, test_ds), metadata = tfds.load(
    'tf_flowers',
    split=['train[:80%]', 'train[80%:90%]', 'train[90%:]'],
    with_info=True,
    as_supervised=True,
)

2021-07-31 01:20:32.658409: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcuda.so.1
2021-07-31 01:20:33.245494: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-07-31 01:20:33.246357: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1733] Found device 0 with properties: 
pciBusID: 0000:00:05.0 name: Tesla V100-SXM2-16GB computeCapability: 7.0
coreClock: 1.53GHz coreCount: 80 deviceMemorySize: 15.78GiB deviceMemoryBandwidth: 836.37GiB/s
2021-07-31 01:20:33.246390: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudart.so.11.0
2021-07-31 01:20:33.249453: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcublas.so.11
2021-07-31 01:20:33.249541: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcublasLt.so.11
2021-07-31 01:20:33.250610: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcufft.so.10
2021-07-31 01:20:33.250929: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcurand.so.10
2021-07-31 01:20:33.251983: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcusolver.so.11
2021-07-31 01:20:33.252852: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcusparse.so.11
2021-07-31 01:20:33.253031: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudnn.so.8
2021-07-31 01:20:33.253121: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-07-31 01:20:33.253980: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-07-31 01:20:33.254770: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1871] Adding visible gpu devices: 0
2021-07-31 01:20:33.255502: I tensorflow/core/platform/cpu_feature_guard.cc:142] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 AVX512F FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2021-07-31 01:20:33.256126: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-07-31 01:20:33.256945: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1733] Found device 0 with properties: 
pciBusID: 0000:00:05.0 name: Tesla V100-SXM2-16GB computeCapability: 7.0
coreClock: 1.53GHz coreCount: 80 deviceMemorySize: 15.78GiB deviceMemoryBandwidth: 836.37GiB/s
2021-07-31 01:20:33.257035: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-07-31 01:20:33.257857: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-07-31 01:20:33.258642: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1871] Adding visible gpu devices: 0
2021-07-31 01:20:33.258683: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudart.so.11.0
2021-07-31 01:20:33.827497: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1258] Device interconnect StreamExecutor with strength 1 edge matrix:
2021-07-31 01:20:33.827531: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1264]      0 
2021-07-31 01:20:33.827538: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1277] 0:   N 
2021-07-31 01:20:33.827734: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-07-31 01:20:33.828649: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-07-31 01:20:33.829480: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-07-31 01:20:33.830357: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1418] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 14646 MB memory) -> physical GPU (device: 0, name: Tesla V100-SXM2-16GB, pci bus id: 0000:00:05.0, compute capability: 7.0)

The flowers dataset has five classes.

num_classes = metadata.features['label'].num_classes
print(num_classes)

5

Let's retrieve an image from the dataset and use it to demonstrate data augmentation.

get_label_name = metadata.features['label'].int2str

image, label = next(iter(train_ds))
_ = plt.imshow(image)
_ = plt.title(get_label_name(label))

2021-07-31 01:20:33.935374: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:176] None of the MLIR Optimization Passes are enabled (registered 2)
2021-07-31 01:20:33.935908: I tensorflow/core/platform/profile_utils/cpu_utils.cc:114] CPU Frequency: 2000160000 Hz
2021-07-31 01:20:34.470092: W tensorflow/core/kernels/data/cache_dataset_ops.cc:768] The calling iterator did not fully read the dataset being cached. In order to avoid unexpected truncation of the dataset, the partially cached contents of the dataset  will be discarded. This can happen if you have an input pipeline similar to `dataset.cache().take(k).repeat()`. You should use `dataset.take(k).cache().repeat()` instead.

Use Keras preprocessing layers

Note: The Keras Preprocesing Layers introduced in this section are currently experimental.

Resizing and rescaling

You can use preprocessing layers to resize your images to a consistent shape, and to rescale pixel values.

IMG_SIZE = 180

resize_and_rescale = tf.keras.Sequential([
  layers.experimental.preprocessing.Resizing(IMG_SIZE, IMG_SIZE),
  layers.experimental.preprocessing.Rescaling(1./255)
])

Note: the rescaling layer above standardizes pixel values to [0,1]. If instead you wanted [-1,1], you would write Rescaling(1./127.5, offset=-1).

You can see the result of applying these layers to an image.

result = resize_and_rescale(image)
_ = plt.imshow(result)

You can verify the pixels are in [0-1].

print("Min and max pixel values:", result.numpy().min(), result.numpy().max())

Min and max pixel values: 0.0 1.0

Data augmentation

You can use preprocessing layers for data augmentation as well.

Let's create a few preprocessing layers and apply them repeatedly to the same image.

data_augmentation = tf.keras.Sequential([
  layers.experimental.preprocessing.RandomFlip("horizontal_and_vertical"),
  layers.experimental.preprocessing.RandomRotation(0.2),
])

# Add the image to a batch
image = tf.expand_dims(image, 0)

plt.figure(figsize=(10, 10))
for i in range(9):
  augmented_image = data_augmentation(image)
  ax = plt.subplot(3, 3, i + 1)
  plt.imshow(augmented_image[0])
  plt.axis("off")

There are a variety of preprocessing layers you can use for data augmentation including layers.RandomContrast, layers.RandomCrop, layers.RandomZoom, and others.

Two options to use the preprocessing layers

There are two ways you can use these preprocessing layers, with important tradeoffs.

Option 1: Make the preprocessing layers part of your model

model = tf.keras.Sequential([
  resize_and_rescale,
  data_augmentation,
  layers.Conv2D(16, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  # Rest of your model
])

There are two important points to be aware of in this case:

• Data augmentation will run on-device, synchronously with the rest of your layers, and benefit from GPU acceleration.

• When you export your model using model.save, the preprocessing layers will be saved along with the rest of your model. If you later deploy this model, it will automatically standardize images (according to the configuration of your layers). This can save you from the effort of having to reimplement that logic server-side.

Note: Data augmentation is inactive at test time so input images will only be augmented during calls to model.fit (not model.evaluate or model.predict).

Option 2: Apply the preprocessing layers to your dataset

aug_ds = train_ds.map(
  lambda x, y: (resize_and_rescale(x, training=True), y))

With this approach, you use Dataset.map to create a dataset that yields batches of augmented images. In this case:

• Data augmentation will happen asynchronously on the CPU, and is non-blocking. You can overlap the training of your model on the GPU with data preprocessing, using Dataset.prefetch, shown below.
• In this case the prepreprocessing layers will not be exported with the model when you call model.save. You will need to attach them to your model before saving it or reimplement them server-side. After training, you can attach the preprocessing layers before export.

You can find an example of the first option in the image classification tutorial. Let's demonstrate the second option here.

Apply the preprocessing layers to the datasets

Configure the train, validation, and test datasets with the preprocessing layers you created above. You will also configure the datasets for performance, using parallel reads and buffered prefetching to yield batches from disk without I/O become blocking. You can learn more dataset performance in the Better performance with the tf.data API guide.

Note: data augmentation should only be applied to the training set.

batch_size = 32
AUTOTUNE = tf.data.AUTOTUNE

def prepare(ds, shuffle=False, augment=False):
  # Resize and rescale all datasets
  ds = ds.map(lambda x, y: (resize_and_rescale(x), y), 
              num_parallel_calls=AUTOTUNE)

  if shuffle:
    ds = ds.shuffle(1000)

  # Batch all datasets
  ds = ds.batch(batch_size)

  # Use data augmentation only on the training set
  if augment:
    ds = ds.map(lambda x, y: (data_augmentation(x, training=True), y), 
                num_parallel_calls=AUTOTUNE)

  # Use buffered prefecting on all datasets
  return ds.prefetch(buffer_size=AUTOTUNE)

train_ds = prepare(train_ds, shuffle=True, augment=True)
val_ds = prepare(val_ds)
test_ds = prepare(test_ds)

Train a model

For completeness, you will now train a model using these datasets. This model has not been tuned for accuracy (the goal is to show you the mechanics).

model = tf.keras.Sequential([
  layers.Conv2D(16, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Conv2D(32, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Conv2D(64, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Flatten(),
  layers.Dense(128, activation='relu'),
  layers.Dense(num_classes)
])

model.compile(optimizer='adam',
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

epochs=5
history = model.fit(
  train_ds,
  validation_data=val_ds,
  epochs=epochs
)

Epoch 1/5
2021-07-31 01:20:39.448244: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudnn.so.8
2021-07-31 01:20:41.475212: I tensorflow/stream_executor/cuda/cuda_dnn.cc:359] Loaded cuDNN version 8100
2021-07-31 01:20:46.496035: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcublas.so.11
2021-07-31 01:20:46.860481: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcublasLt.so.11
92/92 [==============================] - 17s 66ms/step - loss: 1.3434 - accuracy: 0.4271 - val_loss: 1.1534 - val_accuracy: 0.5368
Epoch 2/5
92/92 [==============================] - 3s 27ms/step - loss: 1.0960 - accuracy: 0.5565 - val_loss: 1.0718 - val_accuracy: 0.5695
Epoch 3/5
92/92 [==============================] - 3s 26ms/step - loss: 1.0115 - accuracy: 0.5988 - val_loss: 1.0322 - val_accuracy: 0.6022
Epoch 4/5
92/92 [==============================] - 3s 27ms/step - loss: 0.9503 - accuracy: 0.6202 - val_loss: 0.8811 - val_accuracy: 0.6730
Epoch 5/5
92/92 [==============================] - 3s 27ms/step - loss: 0.8758 - accuracy: 0.6570 - val_loss: 0.8760 - val_accuracy: 0.6485

loss, acc = model.evaluate(test_ds)
print("Accuracy", acc)

12/12 [==============================] - 1s 13ms/step - loss: 0.8319 - accuracy: 0.6812
Accuracy 0.6811988949775696

Custom data augmentation

You can also create custom data augmenation layers. This tutorial shows two ways of doing so. First, you will create a layers.Lambda layer. This is a good way to write concise code. Next, you will write a new layer via subclassing, which gives you more control. Both layers will randomly invert the colors in an image, according to some probability.

def random_invert_img(x, p=0.5):
  if  tf.random.uniform([]) < p:
    x = (255-x)
  else:
    x
  return x

def random_invert(factor=0.5):
  return layers.Lambda(lambda x: random_invert_img(x, factor))

random_invert = random_invert()

plt.figure(figsize=(10, 10))
for i in range(9):
  augmented_image = random_invert(image)
  ax = plt.subplot(3, 3, i + 1)
  plt.imshow(augmented_image[0].numpy().astype("uint8"))
  plt.axis("off")

Next, implement a custom layer by subclassing.

class RandomInvert(layers.Layer):
  def __init__(self, factor=0.5, **kwargs):
    super().__init__(**kwargs)
    self.factor = factor

  def call(self, x):
    return random_invert_img(x)

_ = plt.imshow(RandomInvert()(image)[0])

Both of these layers can be used as described in options 1 and 2 above.

Using tf.image

The above layers.preprocessing utilities are convenient. For finer control, you can write your own data augmentation pipelines or layers using tf.data and tf.image. You may also want to check out TensorFlow Addons Image: Operations and TensorFlow I/O: Color Space Conversions

Since the flowers dataset was previously configured with data augmentation, let's reimport it to start fresh.

(train_ds, val_ds, test_ds), metadata = tfds.load(
    'tf_flowers',
    split=['train[:80%]', 'train[80%:90%]', 'train[90%:]'],
    with_info=True,
    as_supervised=True,
)

Retrieve an image to work with.

image, label = next(iter(train_ds))
_ = plt.imshow(image)
_ = plt.title(get_label_name(label))

2021-07-31 01:21:08.829596: W tensorflow/core/kernels/data/cache_dataset_ops.cc:768] The calling iterator did not fully read the dataset being cached. In order to avoid unexpected truncation of the dataset, the partially cached contents of the dataset  will be discarded. This can happen if you have an input pipeline similar to `dataset.cache().take(k).repeat()`. You should use `dataset.take(k).cache().repeat()` instead.

Let's use the following function to visualize and compare the original and augmented images side-by-side.

def visualize(original, augmented):
  fig = plt.figure()
  plt.subplot(1,2,1)
  plt.title('Original image')
  plt.imshow(original)

  plt.subplot(1,2,2)
  plt.title('Augmented image')
  plt.imshow(augmented)

Data augmentation

Flipping the image

Flip the image either vertically or horizontally.

flipped = tf.image.flip_left_right(image)
visualize(image, flipped)

Grayscale the image

Grayscale an image.

grayscaled = tf.image.rgb_to_grayscale(image)
visualize(image, tf.squeeze(grayscaled))
_ = plt.colorbar()

Saturate the image

Saturate an image by providing a saturation factor.

saturated = tf.image.adjust_saturation(image, 3)
visualize(image, saturated)

Change image brightness

Change the brightness of image by providing a brightness factor.

bright = tf.image.adjust_brightness(image, 0.4)
visualize(image, bright)

Center crop the image

Crop the image from center up to the image part you desire.

cropped = tf.image.central_crop(image, central_fraction=0.5)
visualize(image,cropped)

Rotate the image

Rotate an image by 90 degrees.

rotated = tf.image.rot90(image)
visualize(image, rotated)

Random transformations

Warning: There are two sets of random image operations: tf.image.random* and tf.image.stateless_random*. Using tf.image.random* operations is strongly discouraged as they use the old RNGs from TF 1.x. Instead, please use the random image operations introduced in this tutorial. For more information, please refer to Random number generation.

Applying random transformations to the images can further help generalize and expand the dataset. Current tf.image API provides 8 such random image operations (ops):

• tf.image.stateless_random_brightness
• tf.image.stateless_random_contrast
• tf.image.stateless_random_crop
• tf.image.stateless_random_flip_left_right
• tf.image.stateless_random_flip_up_down
• tf.image.stateless_random_hue
• tf.image.stateless_random_jpeg_quality
• tf.image.stateless_random_saturation

These random image ops are purely functional: the ouput only depends on the input. This makes them simple to use in high performance, deterministic input pipelines. They require a seed value be input each step. Given the same seed, they return the same results independent of how many times they are called.

Note: seed is a Tensor of shape (2,) whose values are any integers.

In the following sections, you will:

1. Go over examples of using random image operations to transform an image; and
2. Demonstrate how to apply random transformations to a training dataset.

Randomly change image brightness

Randomly change the brightness of image by providing a brightness factor and seed. The brightness factor is chosen randomly in the range [-max_delta, max_delta) and is associated with the given seed.

for i in range(3):
  seed = (i, 0)  # tuple of size (2,)
  stateless_random_brightness = tf.image.stateless_random_brightness(
      image, max_delta=0.95, seed=seed)
  visualize(image, stateless_random_brightness)

Randomly change image contrast

Randomly change the contrast of image by providing a contrast range and seed. The contrast range is chosen randomly in the interval [lower, upper] and is associated with the given seed.

for i in range(3):
  seed = (i, 0)  # tuple of size (2,)
  stateless_random_contrast = tf.image.stateless_random_contrast(
      image, lower=0.1, upper=0.9, seed=seed)
  visualize(image, stateless_random_contrast)

Randomly crop an image

Randomly crop image by providing target size and seed. The portion that gets cropped out of image is at a randomly chosen offset and is associated with the given seed.

for i in range(3):
  seed = (i, 0)  # tuple of size (2,)
  stateless_random_crop = tf.image.stateless_random_crop(
      image, size=[210, 300, 3], seed=seed)
  visualize(image, stateless_random_crop)

Apply augmentation to a dataset

Let's first download the image dataset again in case they are modified in the previous sections.

(train_datasets, val_ds, test_ds), metadata = tfds.load(
    'tf_flowers',
    split=['train[:80%]', 'train[80%:90%]', 'train[90%:]'],
    with_info=True,
    as_supervised=True,
)

Let's define a utility function for resizing and rescaling the images. This function will be used in unifying the size and scale of images in the dataset:

def resize_and_rescale(image, label):
  image = tf.cast(image, tf.float32)
  image = tf.image.resize(image, [IMG_SIZE, IMG_SIZE])
  image = (image / 255.0)
  return image, label

Let's also define augment function that can apply the random transformations to the images. This function will be used on the dataset in the next step.

def augment(image_label, seed):
  image, label = image_label
  image, label = resize_and_rescale(image, label)
  image = tf.image.resize_with_crop_or_pad(image, IMG_SIZE + 6, IMG_SIZE + 6)
  # Make a new seed
  new_seed = tf.random.experimental.stateless_split(seed, num=1)[0, :]
  # Random crop back to the original size
  image = tf.image.stateless_random_crop(
      image, size=[IMG_SIZE, IMG_SIZE, 3], seed=seed)
  # Random brightness
  image = tf.image.stateless_random_brightness(
      image, max_delta=0.5, seed=new_seed)
  image = tf.clip_by_value(image, 0, 1)
  return image, label

Option 1: Using tf.data.experimental.Counter()

Create a tf.data.experimental.Counter() object (let's call it counter) and zip the dataset with (counter, counter). This will ensure that each image in the dataset gets associated with a unique value (of shape (2,)) based on counter which later can get passed into the augment function as the seed value for random transformations.

# Create counter and zip together with train dataset
counter = tf.data.experimental.Counter()
train_ds = tf.data.Dataset.zip((train_datasets, (counter, counter)))

Map the augment function to the training dataset.

train_ds = (
    train_ds
    .shuffle(1000)
    .map(augment, num_parallel_calls=AUTOTUNE)
    .batch(batch_size)
    .prefetch(AUTOTUNE)
)

val_ds = (
    val_ds
    .map(resize_and_rescale, num_parallel_calls=AUTOTUNE)
    .batch(batch_size)
    .prefetch(AUTOTUNE)
)

test_ds = (
    test_ds
    .map(resize_and_rescale, num_parallel_calls=AUTOTUNE)
    .batch(batch_size)
    .prefetch(AUTOTUNE)
)

Option 2: Using tf.random.Generator

Create a tf.random.Generator object with an intial seed value. Calling make_seeds function on the same generator object returns a new, unique seed value always. Define a wrapper function that 1) calls make_seeds function and that 2) passes the newly generated seed value into the augment function for random transformations.

Note: tf.random.Generator objects store RNG state in a tf.Variable, which means it can be saved as a checkpoint or in a SavedModel. For more details, please refer to Random number generation.

# Create a generator
rng = tf.random.Generator.from_seed(123, alg='philox')

# A wrapper function for updating seeds
def f(x, y):
  seed = rng.make_seeds(2)[0]
  image, label = augment((x, y), seed)
  return image, label

Map the wrapper function f to the training dataset.

train_ds = (
    train_datasets
    .shuffle(1000)
    .map(f, num_parallel_calls=AUTOTUNE)
    .batch(batch_size)
    .prefetch(AUTOTUNE)
)

val_ds = (
    val_ds
    .map(resize_and_rescale, num_parallel_calls=AUTOTUNE)
    .batch(batch_size)
    .prefetch(AUTOTUNE)
)

test_ds = (
    test_ds
    .map(resize_and_rescale, num_parallel_calls=AUTOTUNE)
    .batch(batch_size)
    .prefetch(AUTOTUNE)
)

These datasets can now be used to train a model as shown previously.

Next steps

This tutorial demonstrated data augmentation using Keras Preprocessing Layers and tf.image. To learn how to include preprocessing layers inside your model, see the Image classification tutorial. You may also be interested in learning how preprocessing layers can help you classify text, as shown in the Basic text classification tutorial. You can learn more about tf.data in this guide, and you can learn how to configure your input pipelines for performance here.