Tune in to the first Women in ML Symposium this Tuesday, October 19 at 9am PST Register now

Load and preprocess images

View on TensorFlow.org Run in Google Colab View source on GitHub Download notebook

This tutorial shows how to load and preprocess an image dataset in three ways:

Setup

import numpy as np
import os
import PIL
import PIL.Image
import tensorflow as tf
import tensorflow_datasets as tfds
print(tf.__version__)
2.6.0

Download the flowers dataset

This tutorial uses a dataset of several thousand photos of flowers. The flowers dataset contains five sub-directories, one per class:

flowers_photos/
  daisy/
  dandelion/
  roses/
  sunflowers/
  tulips/
import pathlib
dataset_url = "https://storage.googleapis.com/download.tensorflow.org/example_images/flower_photos.tgz"
data_dir = tf.keras.utils.get_file(origin=dataset_url,
                                   fname='flower_photos',
                                   untar=True)
data_dir = pathlib.Path(data_dir)

After downloading (218MB), you should now have a copy of the flower photos available. There are 3,670 total images:

image_count = len(list(data_dir.glob('*/*.jpg')))
print(image_count)
3670

Each directory contains images of that type of flower. Here are some roses:

roses = list(data_dir.glob('roses/*'))
PIL.Image.open(str(roses[0]))

png

roses = list(data_dir.glob('roses/*'))
PIL.Image.open(str(roses[1]))

png

Load data using a Keras utility

Let's load these images off disk using the helpful tf.keras.utils.image_dataset_from_directory utility.

Create a dataset

Define some parameters for the loader:

batch_size = 32
img_height = 180
img_width = 180

It's good practice to use a validation split when developing your model. You will use 80% of the images for training and 20% for validation.

train_ds = tf.keras.utils.image_dataset_from_directory(
  data_dir,
  validation_split=0.2,
  subset="training",
  seed=123,
  image_size=(img_height, img_width),
  batch_size=batch_size)
Found 3670 files belonging to 5 classes.
Using 2936 files for training.
val_ds = tf.keras.utils.image_dataset_from_directory(
  data_dir,
  validation_split=0.2,
  subset="validation",
  seed=123,
  image_size=(img_height, img_width),
  batch_size=batch_size)
Found 3670 files belonging to 5 classes.
Using 734 files for validation.

You can find the class names in the class_names attribute on these datasets.

class_names = train_ds.class_names
print(class_names)
['daisy', 'dandelion', 'roses', 'sunflowers', 'tulips']

Visualize the data

Here are the first nine images from the training dataset.

import matplotlib.pyplot as plt

plt.figure(figsize=(10, 10))
for images, labels in train_ds.take(1):
  for i in range(9):
    ax = plt.subplot(3, 3, i + 1)
    plt.imshow(images[i].numpy().astype("uint8"))
    plt.title(class_names[labels[i]])
    plt.axis("off")

png

You can train a model using these datasets by passing them to model.fit (shown later in this tutorial). If you like, you can also manually iterate over the dataset and retrieve batches of images:

for image_batch, labels_batch in train_ds:
  print(image_batch.shape)
  print(labels_batch.shape)
  break
(32, 180, 180, 3)
(32,)

The image_batch is a tensor of the shape (32, 180, 180, 3). This is a batch of 32 images of shape 180x180x3 (the last dimension refers to color channels RGB). The label_batch is a tensor of the shape (32,), these are corresponding labels to the 32 images.

You can call .numpy() on either of these tensors to convert them to a numpy.ndarray.

Standardize the data

The RGB channel values are in the [0, 255] range. This is not ideal for a neural network; in general you should seek to make your input values small.

Here, you will standardize values to be in the [0, 1] range by using tf.keras.layers.Rescaling:

normalization_layer = tf.keras.layers.Rescaling(1./255)

There are two ways to use this layer. You can apply it to the dataset by calling Dataset.map:

normalized_ds = train_ds.map(lambda x, y: (normalization_layer(x), y))
image_batch, labels_batch = next(iter(normalized_ds))
first_image = image_batch[0]
# Notice the pixel values are now in `[0,1]`.
print(np.min(first_image), np.max(first_image))
0.0 0.96902645

Or, you can include the layer inside your model definition to simplify deployment. You will use the second approach here.

Configure the dataset for performance

Let's make sure to use buffered prefetching so you can yield data from disk without having I/O become blocking. These are two important methods you should use when loading data:

  • Dataset.cache keeps the images in memory after they're loaded off disk during the first epoch. This will ensure the dataset does not become a bottleneck while training your model. If your dataset is too large to fit into memory, you can also use this method to create a performant on-disk cache.
  • Dataset.prefetch overlaps data preprocessing and model execution while training.

Interested readers can learn more about both methods, as well as how to cache data to disk in the Prefetching section of the Better performance with the tf.data API guide.

AUTOTUNE = tf.data.AUTOTUNE

train_ds = train_ds.cache().prefetch(buffer_size=AUTOTUNE)
val_ds = val_ds.cache().prefetch(buffer_size=AUTOTUNE)

Train a model

For completeness, you will show how to train a simple model using the datasets you have just prepared.

The Sequential model consists of three convolution blocks (tf.keras.layers.Conv2D) with a max pooling layer (tf.keras.layers.MaxPooling2D) in each of them. There's a fully-connected layer (tf.keras.layers.Dense) with 128 units on top of it that is activated by a ReLU activation function ('relu'). This model has not been tuned in any way—the goal is to show you the mechanics using the datasets you just created. To learn more about image classification, visit the Image classification tutorial.

num_classes = 5

model = tf.keras.Sequential([
  tf.keras.layers.Rescaling(1./255),
  tf.keras.layers.Conv2D(32, 3, activation='relu'),
  tf.keras.layers.MaxPooling2D(),
  tf.keras.layers.Conv2D(32, 3, activation='relu'),
  tf.keras.layers.MaxPooling2D(),
  tf.keras.layers.Conv2D(32, 3, activation='relu'),
  tf.keras.layers.MaxPooling2D(),
  tf.keras.layers.Flatten(),
  tf.keras.layers.Dense(128, activation='relu'),
  tf.keras.layers.Dense(num_classes)
])

Choose the tf.keras.optimizers.Adam optimizer and tf.keras.losses.SparseCategoricalCrossentropy loss function. To view training and validation accuracy for each training epoch, pass the metrics argument to Model.compile.

model.compile(
  optimizer='adam',
  loss=tf.losses.SparseCategoricalCrossentropy(from_logits=True),
  metrics=['accuracy'])
model.fit(
  train_ds,
  validation_data=val_ds,
  epochs=3
)
Epoch 1/3
92/92 [==============================] - 4s 22ms/step - loss: 1.3233 - accuracy: 0.4326 - val_loss: 1.0868 - val_accuracy: 0.5450
Epoch 2/3
92/92 [==============================] - 1s 11ms/step - loss: 0.9960 - accuracy: 0.6056 - val_loss: 0.9698 - val_accuracy: 0.6117
Epoch 3/3
92/92 [==============================] - 1s 11ms/step - loss: 0.8776 - accuracy: 0.6625 - val_loss: 0.9740 - val_accuracy: 0.6417
<keras.callbacks.History at 0x7fcb6076f4d0>

You may notice the validation accuracy is low compared to the training accuracy, indicating your model is overfitting. You can learn more about overfitting and how to reduce it in this tutorial.

Using tf.data for finer control

The above Keras preprocessing utility—tf.keras.utils.image_dataset_from_directory—is a convenient way to create a tf.data.Dataset from a directory of images.

For finer grain control, you can write your own input pipeline using tf.data. This section shows how to do just that, beginning with the file paths from the TGZ file you downloaded earlier.

list_ds = tf.data.Dataset.list_files(str(data_dir/'*/*'), shuffle=False)
list_ds = list_ds.shuffle(image_count, reshuffle_each_iteration=False)
for f in list_ds.take(5):
  print(f.numpy())
b'/home/kbuilder/.keras/datasets/flower_photos/dandelion/8663932737_0a603ab718_n.jpg'
b'/home/kbuilder/.keras/datasets/flower_photos/dandelion/4634716478_1cbcbee7ca.jpg'
b'/home/kbuilder/.keras/datasets/flower_photos/daisy/21402054779_759366efb0_n.jpg'
b'/home/kbuilder/.keras/datasets/flower_photos/sunflowers/5970301989_fe3a68aac8_m.jpg'
b'/home/kbuilder/.keras/datasets/flower_photos/sunflowers/45045003_30bbd0a142_m.jpg'

The tree structure of the files can be used to compile a class_names list.

class_names = np.array(sorted([item.name for item in data_dir.glob('*') if item.name != "LICENSE.txt"]))
print(class_names)
['daisy' 'dandelion' 'roses' 'sunflowers' 'tulips']

Split the dataset into training and validation sets:

val_size = int(image_count * 0.2)
train_ds = list_ds.skip(val_size)
val_ds = list_ds.take(val_size)

You can print the length of each dataset as follows:

print(tf.data.experimental.cardinality(train_ds).numpy())
print(tf.data.experimental.cardinality(val_ds).numpy())
2936
734

Write a short function that converts a file path to an (img, label) pair:

def get_label(file_path):
  # Convert the path to a list of path components
  parts = tf.strings.split(file_path, os.path.sep)
  # The second to last is the class-directory
  one_hot = parts[-2] == class_names
  # Integer encode the label
  return tf.argmax(one_hot)
def decode_img(img):
  # Convert the compressed string to a 3D uint8 tensor
  img = tf.io.decode_jpeg(img, channels=3)
  # Resize the image to the desired size
  return tf.image.resize(img, [img_height, img_width])
def process_path(file_path):
  label = get_label(file_path)
  # Load the raw data from the file as a string
  img = tf.io.read_file(file_path)
  img = decode_img(img)
  return img, label

Use Dataset.map to create a dataset of image, label pairs:

# Set `num_parallel_calls` so multiple images are loaded/processed in parallel.
train_ds = train_ds.map(process_path, num_parallel_calls=AUTOTUNE)
val_ds = val_ds.map(process_path, num_parallel_calls=AUTOTUNE)
for image, label in train_ds.take(1):
  print("Image shape: ", image.numpy().shape)
  print("Label: ", label.numpy())
Image shape:  (180, 180, 3)
Label:  4

Configure dataset for performance

To train a model with this dataset you will want the data:

  • To be well shuffled.
  • To be batched.
  • Batches to be available as soon as possible.

These features can be added using the tf.data API. For more details, visit the Input Pipeline Performance guide.

def configure_for_performance(ds):
  ds = ds.cache()
  ds = ds.shuffle(buffer_size=1000)
  ds = ds.batch(batch_size)
  ds = ds.prefetch(buffer_size=AUTOTUNE)
  return ds

train_ds = configure_for_performance(train_ds)
val_ds = configure_for_performance(val_ds)

Visualize the data

You can visualize this dataset similarly to the one you created previously:

image_batch, label_batch = next(iter(train_ds))

plt.figure(figsize=(10, 10))
for i in range(9):
  ax = plt.subplot(3, 3, i + 1)
  plt.imshow(image_batch[i].numpy().astype("uint8"))
  label = label_batch[i]
  plt.title(class_names[label])
  plt.axis("off")
2021-10-01 03:21:04.086862: 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.

png

Continue training the model

You have now manually built a similar tf.data.Dataset to the one created by tf.keras.utils.image_dataset_from_directory above. You can continue training the model with it. As before, you will train for just a few epochs to keep the running time short.

model.fit(
  train_ds,
  validation_data=val_ds,
  epochs=3
)
Epoch 1/3
92/92 [==============================] - 2s 20ms/step - loss: 0.7723 - accuracy: 0.7081 - val_loss: 0.9020 - val_accuracy: 0.6580
Epoch 2/3
92/92 [==============================] - 1s 12ms/step - loss: 0.5976 - accuracy: 0.7810 - val_loss: 0.7204 - val_accuracy: 0.7234
Epoch 3/3
92/92 [==============================] - 1s 12ms/step - loss: 0.4157 - accuracy: 0.8437 - val_loss: 0.9229 - val_accuracy: 0.7153
<keras.callbacks.History at 0x7fcb603f0590>

Using TensorFlow Datasets

So far, this tutorial has focused on loading data off disk. You can also find a dataset to use by exploring the large catalog of easy-to-download datasets at TensorFlow Datasets.

As you have previously loaded the Flowers dataset off disk, let's now import it with TensorFlow Datasets.

Download the Flowers dataset using TensorFlow Datasets:

(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,
)

The flowers dataset has five classes:

num_classes = metadata.features['label'].num_classes
print(num_classes)
5

Retrieve an image from the dataset:

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

image, label = next(iter(train_ds))
_ = plt.imshow(image)
_ = plt.title(get_label_name(label))
2021-10-01 03:21:12.800729: 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.

png

As before, remember to batch, shuffle, and configure the training, validation, and test sets for performance:

train_ds = configure_for_performance(train_ds)
val_ds = configure_for_performance(val_ds)
test_ds = configure_for_performance(test_ds)

You can find a complete example of working with the Flowers dataset and TensorFlow Datasets by visiting the Data augmentation tutorial.

Next steps

This tutorial showed two ways of loading images off disk. First, you learned how to load and preprocess an image dataset using Keras preprocessing layers and utilities. Next, you learned how to write an input pipeline from scratch using tf.data. Finally, you learned how to download a dataset from TensorFlow Datasets.

For your next steps: