Final Up to date on July 20, 2022
Once we work on a machine studying downside associated to photographs, not solely we have to accumulate some photographs as coaching information, but in addition have to make use of augmentation to create variations within the picture. It’s very true for extra advanced object recognition issues.
There are numerous methods for picture augmentation. You could use some exterior libraries or write your individual capabilities for that. There are some modules in TensorFlow and Keras for augmentation, too. On this submit you’ll uncover how we are able to use the Keras preprocessing layer in addition to tf.picture module in TensorFlow for picture augmentation.
After studying this submit, you’ll know:
- What are the Keras preprocessing layers and easy methods to use them
- What are the capabilities supplied by
tf.picturemodule for picture augmentation - The best way to use augmentation along with
tf.informationdataset
Let’s get began.
Picture Augmentation with Keras Preprocessing Layers and tf.picture.
Picture by Steven Kamenar. Some rights reserved.
Overview
This text is cut up into 5 sections; they’re:
- Getting Photographs
- Visualizing the Photographs
- Keras Preprocessing Layesr
- Utilizing tf.picture API for Augmentation
- Utilizing Preprocessing Layers in Neural Networks
Getting Photographs
Earlier than we see how we are able to do augmentation, we have to get the pictures. Finally, we want the pictures to be represented as arrays, for instance, in HxWx3 in 8-bit integers for the RGB pixel worth. There are numerous methods to get the pictures. Some might be downloaded as a ZIP file. Should you’re utilizing TensorFlow, chances are you’ll get some picture dataset from the tensorflow_datasets library.
On this tutorial, we’re going to use the citrus leaves photographs, which is a small dataset in lower than 100MB. It may be downloaded from tensorflow_datasets as follows:
|
import tensorflow_datasets as tfds ds, meta = tfds.load(‘citrus_leaves’, with_info=True, cut up=‘prepare’, shuffle_files=True) |
Working this code the primary time will obtain the picture dataset into your pc with the next output:
|
Downloading and getting ready dataset 63.87 MiB (obtain: 63.87 MiB, generated: 37.89 MiB, whole: 101.76 MiB) to ~/tensorflow_datasets/citrus_leaves/0.1.2… Extraction accomplished…: 100%|██████████████████████████████| 1/1 [00:06<00:00,  6.54s/ file] Dl Dimension…: 100%|██████████████████████████████████████████| 63/63 [00:06<00:00,  9.63 MiB/s] Dl Accomplished…: 100%|███████████████████████████████████████| 1/1 [00:06<00:00,  6.54s/ url] Dataset citrus_leaves downloaded and ready to ~/tensorflow_datasets/citrus_leaves/0.1.2. Subsequent calls will reuse this information. |
The operate above returns the pictures as a tf.information dataset object and the metadata. It is a classification dataset. We are able to print the coaching labels with the next:
|
... for i in vary(meta.options[‘label’].num_classes): Â Â Â Â print(meta.options[‘label’].int2str(i)) |
and this prints:
|
Black spot canker greening wholesome |
Should you run this code once more at a later time, you’ll reuse the downloaded picture. However the different technique to load the downloaded photographs right into a tf.information dataset is to the image_dataset_from_directory() operate.
As we are able to see the display screen output above, the dataset is downloaded into the listing ~/tensorflow_datasets. Should you have a look at the listing, you see the listing construction as follows:
|
…/Citrus/Leaves ├── Black spot ├── Melanose ├── canker ├── greening └── wholesome |
The directories are the labels and the pictures are recordsdata saved beneath their corresponding listing. We are able to let the operate to learn the listing recursively right into a dataset:
|
import tensorflow as tf from tensorflow.keras.utils import image_dataset_from_listing  # set to mounted picture measurement 256×256 PATH = “…/Citrus/Leaves” ds = image_dataset_from_directory(PATH,                                   validation_split=0.2, subset=“coaching”,                                   image_size=(256,256), interpolation=“bilinear”,                                   crop_to_aspect_ratio=True,                                   seed=42, shuffle=True, batch_size=32) |
You could wish to set batch_size=None if you do not need the dataset to be batched. Normally we want the dataset to be batched for coaching a neural community mannequin.
Visualizing the Photographs
You will need to visualize the augmentation consequence so we are able to confirm the augmentation result’s what we wish it to be. We are able to use matplotlib for this.
In matplotlib, we have now the imshow() operate to show a picture. Nonetheless, for the picture to be displayed accurately, the picture ought to be offered as an array of 8-bit unsigned integer (uint8).
Given we have now a dataset created utilizing image_dataset_from_directory(), we are able to get the primary batch (of 32 photographs) and show a number of of them utilizing imshow(), as follows:
|
... import matplotlib.pyplot as plt  fig, ax = plt.subplots(3, 3, sharex=True, sharey=True, figsize=(5,5))  for photographs, labels in ds.take(1):     for i in vary(3):         for j in vary(3):             ax[i][j].imshow(photographs[i*3+j].numpy().astype(“uint8”))             ax[i][j].set_title(ds.class_names[labels[i*3+j]]) plt.present() |
Right here we show 9 photographs in a grid, and label the pictures with their corresponding classification label, utilizing ds.class_names. The photographs ought to be transformed to NumPy array in uint8 for show. This code shows a picture like the next:
The whole code from loading the picture to show is as follows.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 |
from tensorflow.keras.utils import image_dataset_from_directory import matplotlib.pyplot as plt  # use image_dataset_from_directory() to load photographs, with picture measurement scaled to 256×256 PATH=‘…/Citrus/Leaves’  # modify to your path ds = image_dataset_from_directory(PATH,                                   validation_split=0.2, subset=“coaching”,                                   image_size=(256,256), interpolation=“mitchellcubic”,                                   crop_to_aspect_ratio=True,                                   seed=42, shuffle=True, batch_size=32)  # Take one batch from dataset and show the pictures fig, ax = plt.subplots(3, 3, sharex=True, sharey=True, figsize=(5,5))  for photographs, labels in ds.take(1):     for i in vary(3):         for j in vary(3):             ax[i][j].imshow(photographs[i*3+j].numpy().astype(“uint8”))             ax[i][j].set_title(ds.class_names[labels[i*3+j]]) plt.present() |
Word that, should you’re utilizing tensorflow_datasets to get the picture, the samples are offered as a dictionary as an alternative of a tuple of (picture,label). You need to change your code barely into the next:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 |
import tensorflow_datasets as tfds import matplotlib.pyplot as plt  # use tfds.load() or image_dataset_from_directory() to load photographs ds, meta = tfds.load(‘citrus_leaves’, with_info=True, cut up=‘prepare’, shuffle_files=True) ds = ds.batch(32)  # Take one batch from dataset and show the pictures fig, ax = plt.subplots(3, 3, sharex=True, sharey=True, figsize=(5,5))  for pattern in ds.take(1):     photographs, labels = pattern[“image”], pattern[“label”]     for i in vary(3):         for j in vary(3):             ax[i][j].imshow(photographs[i*3+j].numpy().astype(“uint8”))             ax[i][j].set_title(meta.options[‘label’].int2str(labels[i*3+j])) plt.present() |
In the remainder of this submit, we assume the dataset is created utilizing image_dataset_from_directory(). You could have to tweak the code barely in case your dataset is created otherwise.
Keras Preprocessing Layers
Keras comes with many neural community layers similar to convolution layers that we have to prepare. There are additionally layers with no parameters to coach, similar to flatten layers to transform an array similar to a picture right into a vector.
The preprocessing layers in Keras are particularly designed to make use of in early levels in a neural community. We are able to use them for picture preprocessing, similar to to resize or rotate the picture or to regulate the brightness and distinction. Whereas the preprocessing layers are purported to be half of a bigger neural community, we are able to additionally use them as capabilities. Under is how we are able to use the resizing layer as a operate to rework some photographs and show them side-by-side with the unique:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 |
...  # create a resizing layer out_height, out_width = 128,256 resize = tf.keras.layers.Resizing(out_height, out_width)  # present unique vs resized fig, ax = plt.subplots(2, 3, figsize=(6,4))  for photographs, labels in ds.take(1):     for i in vary(3):         ax[0][i].imshow(photographs[i].numpy().astype(“uint8”))         ax[0][i].set_title(“unique”)         # resize         ax[1][i].imshow(resize(photographs[i]).numpy().astype(“uint8”))         ax[1][i].set_title(“resize”) plt.present() |
Our photographs are in 256×256 pixels and the resizing layer will make them into 256×128 pixels. The output of the above code is as follows:
For the reason that resizing layer is a operate itself, we are able to chain them to the dataset itself. For instance,
|
... def increase(picture, label):     return resize(picture), label  resized_ds = ds.map(increase)  for picture, label in resized_ds:   ... |
The dataset ds has samples within the type of (picture, label). Therefore we created a operate that takes in such tuple and preprocess the picture with the resizing layer. We assigned this operate as an argument for map() within the dataset. Once we draw a pattern from the brand new dataset created with the map() operate, the picture can be a reworked one.
There are extra preprocessing layers obtainable. In under, we show some.
As we noticed above, we are able to resize the picture. We are able to additionally randomly enlarge or shrink the peak or width of a picture. Equally, we are able to zoom in or zoom out on a picture. Under is an instance to govern the picture measurement in numerous methods for a most of 30% improve or lower:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 |
...  # Create preprocessing layers out_height, out_width = 128,256 resize = tf.keras.layers.Resizing(out_height, out_width) top = tf.keras.layers.RandomHeight(0.3) width = tf.keras.layers.RandomWidth(0.3) zoom = tf.keras.layers.RandomZoom(0.3)  # Visualize photographs and augmentations fig, ax = plt.subplots(5, 3, figsize=(6,14))  for photographs, labels in ds.take(1):     for i in vary(3):         ax[0][i].imshow(photographs[i].numpy().astype(“uint8”))         ax[0][i].set_title(“unique”)         # resize         ax[1][i].imshow(resize(photographs[i]).numpy().astype(“uint8”))         ax[1][i].set_title(“resize”)         # top         ax[2][i].imshow(top(photographs[i]).numpy().astype(“uint8”))         ax[2][i].set_title(“top”)         # width         ax[3][i].imshow(width(photographs[i]).numpy().astype(“uint8”))         ax[3][i].set_title(“width”)         # zoom         ax[4][i].imshow(zoom(photographs[i]).numpy().astype(“uint8”))         ax[4][i].set_title(“zoom”) plt.present() |
This code exhibits photographs as follows:
Whereas we specified a hard and fast dimension in resize, we have now a random quantity of manipulation in different augmentations.
We are able to additionally do flipping, rotation, cropping, and geometric translation utilizing preprocessing layers:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 |
... # Create preprocessing layers flip = tf.keras.layers.RandomFlip(“horizontal_and_vertical”) # or “horizontal”, “vertical” rotate = tf.keras.layers.RandomRotation(0.2) crop = tf.keras.layers.RandomCrop(out_height, out_width) translation = tf.keras.layers.RandomTranslation(height_factor=0.2, width_factor=0.2)  # Visualize augmentations fig, ax = plt.subplots(5, 3, figsize=(6,14))  for photographs, labels in ds.take(1):     for i in vary(3):         ax[0][i].imshow(photographs[i].numpy().astype(“uint8”))         ax[0][i].set_title(“unique”)         # flip         ax[1][i].imshow(flip(photographs[i]).numpy().astype(“uint8”))         ax[1][i].set_title(“flip”)         # crop         ax[2][i].imshow(crop(photographs[i]).numpy().astype(“uint8”))         ax[2][i].set_title(“crop”)         # translation         ax[3][i].imshow(translation(photographs[i]).numpy().astype(“uint8”))         ax[3][i].set_title(“translation”)         # rotate         ax[4][i].imshow(rotate(photographs[i]).numpy().astype(“uint8”))         ax[4][i].set_title(“rotate”) plt.present() |
This code exhibits the next photographs:
And at last, we are able to do augmentations on coloration changes as nicely:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 |
... brightness = tf.keras.layers.RandomBrightness([–0.8,0.8]) distinction = tf.keras.layers.RandomContrast(0.2)  # Visualize augmentation fig, ax = plt.subplots(3, 3, figsize=(6,7))  for photographs, labels in ds.take(1):     for i in vary(3):         ax[0][i].imshow(photographs[i].numpy().astype(“uint8”))         ax[0][i].set_title(“unique”)         # brightness         ax[1][i].imshow(brightness(photographs[i]).numpy().astype(“uint8”))         ax[1][i].set_title(“brightness”)         # distinction         ax[2][i].imshow(distinction(photographs[i]).numpy().astype(“uint8”))         ax[2][i].set_title(“distinction”) plt.present() |
This exhibits the pictures as follows:
For completeness, under is the code to show the results of numerous augmentations:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 |
from tensorflow.keras.utils import image_dataset_from_directory import tensorflow as tf import matplotlib.pyplot as plt  # use image_dataset_from_directory() to load photographs, with picture measurement scaled to 256×256 PATH=‘…/Citrus/Leaves’  # modify to your path ds = image_dataset_from_directory(PATH,                                   validation_split=0.2, subset=“coaching”,                                   image_size=(256,256), interpolation=“mitchellcubic”,                                   crop_to_aspect_ratio=True,                                   seed=42, shuffle=True, batch_size=32)  # Create preprocessing layers out_height, out_width = 128,256 resize = tf.keras.layers.Resizing(out_height, out_width) top = tf.keras.layers.RandomHeight(0.3) width = tf.keras.layers.RandomWidth(0.3) zoom = tf.keras.layers.RandomZoom(0.3)  flip = tf.keras.layers.RandomFlip(“horizontal_and_vertical”) rotate = tf.keras.layers.RandomRotation(0.2) crop = tf.keras.layers.RandomCrop(out_height, out_width) translation = tf.keras.layers.RandomTranslation(height_factor=0.2, width_factor=0.2)  brightness = tf.keras.layers.RandomBrightness([–0.8,0.8]) distinction = tf.keras.layers.RandomContrast(0.2)  # Visualize photographs and augmentations fig, ax = plt.subplots(5, 3, figsize=(6,14)) for photographs, labels in ds.take(1):     for i in vary(3):         ax[0][i].imshow(photographs[i].numpy().astype(“uint8”))         ax[0][i].set_title(“unique”)         # resize         ax[1][i].imshow(resize(photographs[i]).numpy().astype(“uint8”))         ax[1][i].set_title(“resize”)         # top         ax[2][i].imshow(top(photographs[i]).numpy().astype(“uint8”))         ax[2][i].set_title(“top”)         # width         ax[3][i].imshow(width(photographs[i]).numpy().astype(“uint8”))         ax[3][i].set_title(“width”)         # zoom         ax[4][i].imshow(zoom(photographs[i]).numpy().astype(“uint8”))         ax[4][i].set_title(“zoom”) plt.present()  fig, ax = plt.subplots(5, 3, figsize=(6,14)) for photographs, labels in ds.take(1):     for i in vary(3):         ax[0][i].imshow(photographs[i].numpy().astype(“uint8”))         ax[0][i].set_title(“unique”)         # flip         ax[1][i].imshow(flip(photographs[i]).numpy().astype(“uint8”))         ax[1][i].set_title(“flip”)         # crop         ax[2][i].imshow(crop(photographs[i]).numpy().astype(“uint8”))         ax[2][i].set_title(“crop”)         # translation         ax[3][i].imshow(translation(photographs[i]).numpy().astype(“uint8”))         ax[3][i].set_title(“translation”)         # rotate         ax[4][i].imshow(rotate(photographs[i]).numpy().astype(“uint8”))         ax[4][i].set_title(“rotate”) plt.present()  fig, ax = plt.subplots(3, 3, figsize=(6,7)) for photographs, labels in ds.take(1):     for i in vary(3):         ax[0][i].imshow(photographs[i].numpy().astype(“uint8”))         ax[0][i].set_title(“unique”)         # brightness         ax[1][i].imshow(brightness(photographs[i]).numpy().astype(“uint8”))         ax[1][i].set_title(“brightness”)         # distinction         ax[2][i].imshow(distinction(photographs[i]).numpy().astype(“uint8”))         ax[2][i].set_title(“distinction”) plt.present() |
Lastly, you will need to level out that the majority neural community mannequin can work higher if the enter photographs are scaled. Whereas we normally use 8-bit unsigned integer for the pixel values in a picture (e.g., for show utilizing imshow() as above), neural community prefers the pixel values to be between 0 and 1, or between -1 and +1. This may be carried out with a preprocessing layers, too. Under is how we are able to replace considered one of our instance above so as to add the scaling layer into the augmentation:
|
... out_height, out_width = 128,256 resize = tf.keras.layers.Resizing(out_height, out_width) rescale = tf.keras.layers.Rescaling(1/127.5, offset=–1)  # rescale pixel values to [-1,1]  def increase(picture, label):     return rescale(resize(picture)), label  rescaled_resized_ds = ds.map(increase)  for picture, label in rescaled_resized_ds:   ... |
Utilizing tf.picture API for Augmentation
Moreover the preprocessing layer, the tf.picture module additionally supplied some capabilities for augmentation. In contrast to the preprocessing layer, these capabilities are meant for use in a user-defined operate and assigned to a dataset utilizing map() as we noticed above.
The capabilities supplied by tf.picture will not be duplicates of the preprocessing layers, though there are some overlap. Under is an instance of utilizing the tf.picture capabilities to resize and crop photographs:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 |
...  fig, ax = plt.subplots(5, 3, figsize=(6,14))  for photographs, labels in ds.take(1):     for i in vary(3):         # unique         ax[0][i].imshow(photographs[i].numpy().astype(“uint8”))         ax[0][i].set_title(“unique”)         # resize         h = int(256 * tf.random.uniform([], minval=0.8, maxval=1.2))         w = int(256 * tf.random.uniform([], minval=0.8, maxval=1.2))         ax[1][i].imshow(tf.picture.resize(photographs[i], [h,w]).numpy().astype(“uint8”))         ax[1][i].set_title(“resize”)         # crop         y, x, h, w = (128 * tf.random.uniform((4,))).numpy().astype(“uint8”)         ax[2][i].imshow(tf.picture.crop_to_bounding_box(photographs[i], y, x, h, w).numpy().astype(“uint8”))         ax[2][i].set_title(“crop”)         # central crop         x = tf.random.uniform([], minval=0.4, maxval=1.0)         ax[3][i].imshow(tf.picture.central_crop(photographs[i], x).numpy().astype(“uint8”))         ax[3][i].set_title(“central crop”)         # crop to (h,w) at random offset         h, w = (256 * tf.random.uniform((2,))).numpy().astype(“uint8”)         seed = tf.random.uniform((2,), minval=0, maxval=65536).numpy().astype(“int32”)         ax[4][i].imshow(tf.picture.stateless_random_crop(photographs[i], [h,w,3], seed).numpy().astype(“uint8”))         ax[4][i].set_title(“random crop”) plt.present() |
Under is the output of the above code:
Whereas the show of photographs match what we might count on from the code, the usage of tf.picture capabilities is kind of completely different from that of the preprocessing layers. Each tf.picture operate is completely different. Subsequently, we are able to see the crop_to_bounding_box() operate takes pixel coordinates however the central_crop() operate assumes a fraction ratio as argument.
These capabilities are additionally completely different in the best way randomness is dealt with. A few of these operate doesn’t assume random conduct. Subsequently, the random resize ought to have the precise output measurement generated utilizing a random quantity generator individually earlier than calling the resize operate. Another operate, similar to stateless_random_crop(), can do augmentation randomly however a pair of random seed in int32 must be specified explicitly.
To proceed the instance, there are the capabilities for flipping a picture and extracting the Sobel edges:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 |
... fig, ax = plt.subplots(5, 3, figsize=(6,14))  for photographs, labels in ds.take(1):     for i in vary(3):         ax[0][i].imshow(photographs[i].numpy().astype(“uint8”))         ax[0][i].set_title(“unique”)         # flip         seed = tf.random.uniform((2,), minval=0, maxval=65536).numpy().astype(“int32”)         ax[1][i].imshow(tf.picture.stateless_random_flip_left_right(photographs[i], seed).numpy().astype(“uint8”))         ax[1][i].set_title(“flip left-right”)         # flip         seed = tf.random.uniform((2,), minval=0, maxval=65536).numpy().astype(“int32”)         ax[2][i].imshow(tf.picture.stateless_random_flip_up_down(photographs[i], seed).numpy().astype(“uint8”))         ax[2][i].set_title(“flip up-down”)         # sobel edge         sobel = tf.picture.sobel_edges(photographs[i:i+1])         ax[3][i].imshow(sobel[0, ..., 0].numpy().astype(“uint8”))         ax[3][i].set_title(“sobel y”)         # sobel edge         ax[4][i].imshow(sobel[0, ..., 1].numpy().astype(“uint8”))         ax[4][i].set_title(“sobel x”) plt.present() |
which exhibits the next:
And the next are the capabilities to govern the brightness, distinction, and colours:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 |
... fig, ax = plt.subplots(5, 3, figsize=(6,14))  for photographs, labels in ds.take(1):     for i in vary(3):         ax[0][i].imshow(photographs[i].numpy().astype(“uint8”))         ax[0][i].set_title(“unique”)         # brightness         seed = tf.random.uniform((2,), minval=0, maxval=65536).numpy().astype(“int32”)         ax[1][i].imshow(tf.picture.stateless_random_brightness(photographs[i], 0.3, seed).numpy().astype(“uint8”))         ax[1][i].set_title(“brightness”)         # distinction         ax[2][i].imshow(tf.picture.stateless_random_contrast(photographs[i], 0.7, 1.3, seed).numpy().astype(“uint8”))         ax[2][i].set_title(“distinction”)         # saturation         ax[3][i].imshow(tf.picture.stateless_random_saturation(photographs[i], 0.7, 1.3, seed).numpy().astype(“uint8”))         ax[3][i].set_title(“saturation”)         # hue         ax[4][i].imshow(tf.picture.stateless_random_hue(photographs[i], 0.3, seed).numpy().astype(“uint8”))         ax[4][i].set_title(“hue”) plt.present() |
This code exhibits the next:
Under is the whole code to show the entire above:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 |
from tensorflow.keras.utils import image_dataset_from_directory import tensorflow as tf import matplotlib.pyplot as plt  # use image_dataset_from_directory() to load photographs, with picture measurement scaled to 256×256 PATH=‘…/Citrus/Leaves’  # modify to your path ds = image_dataset_from_directory(PATH,                                   validation_split=0.2, subset=“coaching”,                                   image_size=(256,256), interpolation=“mitchellcubic”,                                   crop_to_aspect_ratio=True,                                   seed=42, shuffle=True, batch_size=32)  # Visualize tf.picture augmentations  fig, ax = plt.subplots(5, 3, figsize=(6,14)) for photographs, labels in ds.take(1):     for i in vary(3):         # unique         ax[0][i].imshow(photographs[i].numpy().astype(“uint8”))         ax[0][i].set_title(“unique”)         # resize         h = int(256 * tf.random.uniform([], minval=0.8, maxval=1.2))         w = int(256 * tf.random.uniform([], minval=0.8, maxval=1.2))         ax[1][i].imshow(tf.picture.resize(photographs[i], [h,w]).numpy().astype(“uint8”))         ax[1][i].set_title(“resize”)         # crop         y, x, h, w = (128 * tf.random.uniform((4,))).numpy().astype(“uint8”)         ax[2][i].imshow(tf.picture.crop_to_bounding_box(photographs[i], y, x, h, w).numpy().astype(“uint8”))         ax[2][i].set_title(“crop”)         # central crop         x = tf.random.uniform([], minval=0.4, maxval=1.0)         ax[3][i].imshow(tf.picture.central_crop(photographs[i], x).numpy().astype(“uint8”))         ax[3][i].set_title(“central crop”)         # crop to (h,w) at random offset         h, w = (256 * tf.random.uniform((2,))).numpy().astype(“uint8”)         seed = tf.random.uniform((2,), minval=0, maxval=65536).numpy().astype(“int32”)         ax[4][i].imshow(tf.picture.stateless_random_crop(photographs[i], [h,w,3], seed).numpy().astype(“uint8”))         ax[4][i].set_title(“random crop”) plt.present()  fig, ax = plt.subplots(5, 3, figsize=(6,14)) for photographs, labels in ds.take(1):     for i in vary(3):         ax[0][i].imshow(photographs[i].numpy().astype(“uint8”))         ax[0][i].set_title(“unique”)         # flip         seed = tf.random.uniform((2,), minval=0, maxval=65536).numpy().astype(“int32”)         ax[1][i].imshow(tf.picture.stateless_random_flip_left_right(photographs[i], seed).numpy().astype(“uint8”))         ax[1][i].set_title(“flip left-right”)         # flip         seed = tf.random.uniform((2,), minval=0, maxval=65536).numpy().astype(“int32”)         ax[2][i].imshow(tf.picture.stateless_random_flip_up_down(photographs[i], seed).numpy().astype(“uint8”))         ax[2][i].set_title(“flip up-down”)         # sobel edge         sobel = tf.picture.sobel_edges(photographs[i:i+1])         ax[3][i].imshow(sobel[0, ..., 0].numpy().astype(“uint8”))         ax[3][i].set_title(“sobel y”)         # sobel edge         ax[4][i].imshow(sobel[0, ..., 1].numpy().astype(“uint8”))         ax[4][i].set_title(“sobel x”) plt.present()  fig, ax = plt.subplots(5, 3, figsize=(6,14)) for photographs, labels in ds.take(1):     for i in vary(3):         ax[0][i].imshow(photographs[i].numpy().astype(“uint8”))         ax[0][i].set_title(“unique”)         # brightness         seed = tf.random.uniform((2,), minval=0, maxval=65536).numpy().astype(“int32”)         ax[1][i].imshow(tf.picture.stateless_random_brightness(photographs[i], 0.3, seed).numpy().astype(“uint8”))         ax[1][i].set_title(“brightness”)         # distinction         ax[2][i].imshow(tf.picture.stateless_random_contrast(photographs[i], 0.7, 1.3, seed).numpy().astype(“uint8”))         ax[2][i].set_title(“distinction”)         # saturation         ax[3][i].imshow(tf.picture.stateless_random_saturation(photographs[i], 0.7, 1.3, seed).numpy().astype(“uint8”))         ax[3][i].set_title(“saturation”)         # hue         ax[4][i].imshow(tf.picture.stateless_random_hue(photographs[i], 0.3, seed).numpy().astype(“uint8”))         ax[4][i].set_title(“hue”) plt.present() |
These augmentation capabilities ought to be sufficient for many use. However you probably have some particular thought on augmentation, in all probability you would wish a greater picture processing library. OpenCV and Pillow are widespread however highly effective libraries that means that you can remodel photographs higher.
Utilizing Preprocessing Layers in Neural Networks
We used the Keras preprocessing layers as capabilities within the examples above. However they may also be used as layers in a neural community. It’s trivial to make use of. Under is an instance on how we are able to incorporate a preprocessing layer right into a classification community and prepare it utilizing a dataset:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 |
from tensorflow.keras.utils import image_dataset_from_directory import tensorflow as tf import matplotlib.pyplot as plt  # use image_dataset_from_directory() to load photographs, with picture measurement scaled to 256×256 PATH=‘…/Citrus/Leaves’  # modify to your path ds = image_dataset_from_directory(PATH,                                   validation_split=0.2, subset=“coaching”,                                   image_size=(256,256), interpolation=“mitchellcubic”,                                   crop_to_aspect_ratio=True,                                   seed=42, shuffle=True, batch_size=32)  AUTOTUNE = tf.information.AUTOTUNE ds = ds.cache().prefetch(buffer_size=AUTOTUNE)  num_classes = 5 mannequin = tf.keras.Sequential([   tf.keras.layers.RandomFlip(“horizontal_and_vertical”),   tf.keras.layers.RandomRotation(0.2),   tf.keras.layers.Rescaling(1/127.0, offset=–1),   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) ])  mannequin.compile(optimizer=‘adam’,               loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),               metrics=[‘accuracy’])   mannequin.match(ds, epochs=3) |
Working this code offers the next output:
|
Discovered 609 recordsdata belonging to five courses. Utilizing 488 recordsdata for coaching. Epoch 1/3 16/16 [==============================] – 5s 253ms/step – loss: 1.4114 – accuracy: 0.4283 Epoch 2/3 16/16 [==============================] – 4s 259ms/step – loss: 0.8101 – accuracy: 0.6475 Epoch 3/3 16/16 [==============================] – 4s 267ms/step – loss: 0.7015 – accuracy: 0.7111 |
Within the code above, we created the dataset with cache() and prefetch(). It is a efficiency approach to permit the dataset to arrange information asynchronously whereas the neural community is educated. This is able to be vital if the dataset has another augmentation assigned utilizing the map() operate.
You will notice some enchancment in accuracy should you eliminated the RandomFlip and RandomRotation layers since you make the issue simpler. Nonetheless, as we wish the community to foretell nicely on a large variations of picture high quality and properties, utilizing augmentation might help our ensuing community extra highly effective.
Additional Studying
Under are documentations from TensorFlow which might be associated to the examples above:
Abstract
On this submit, you may have seen how we are able to use the tf.information dataset with picture augmentation capabilities from Keras and TensorFlow.
Particularly, you discovered:
- The best way to use the preprocessing layers from Keras, each as a operate and as a part of a neural community
- The best way to create your individual picture augmentation operate and apply it to the dataset utilizing the
map()operate - The best way to use the capabilities supplied by the
tf.picturemodule for picture augmentation
Develop Deep Studying Tasks with Python!
 What If You Might Develop A Community in Minutes
…with only a few strains of Python
Uncover how in my new Book:
Deep Studying With Python
It covers end-to-end tasks on matters like:
Multilayer Perceptrons, Convolutional Nets and Recurrent Neural Nets, and extra…
Lastly Deliver Deep Studying To
Your Personal Tasks
Skip the Lecturers. Simply Outcomes.
