How to get the class names in a tensorflow dataset?

user_5645165 0
Viewed 1924

I am trying to understand how to work with tensorflow datasets, tfds.

The dataset is a directory of this kind

-dataset
  -train
    -class_name1
      -files...
    -class_name2
      -files...
   -val
    -class_name1
      -files...
    -class_name2
      -files...
  -test
    -class_name1
      -files...
    -class_name2
      -files...

Here is some code:

import tensorflow_datasets as tfds
builder = tfds.ImageFolder('/content/dataset')
train_ds, val_ds, test_ds = builder.as_dataset(split=['train', 'val', 'test'], shuffle_files=True, as_supervised=True)

print(builder.info)

Output:
tfds.core.DatasetInfo(
    name='image_folder',
    version=1.0.0,
    description='Generic image classification dataset.',
    homepage='https://www.tensorflow.org/datasets/catalog/image_folder',
    features=FeaturesDict({
        'image': Image(shape=(None, None, 3), dtype=tf.uint8),
        'image/filename': Text(shape=(), dtype=tf.string),
        'label': ClassLabel(shape=(), dtype=tf.int64, num_classes=243),
    }),
    total_num_examples=73090,
    splits={
        'test': 5849,
        'train': 58469,
        'val': 8772,
    },
    supervised_keys=('image', 'label'),
    citation="""""",
    redistribution_info=,
)

When i am plotting, doing classification reports, confusion matrix etc i want to be able to use the class_names not the integer labels.

Is there some easy command that give me access to the class_names? (There are 243 classes, not 2)

tensorflow-datasets tensorflow deep-learning machine-learning python
3 Answers

Your initial builder:

builder = tfds.ImageFolder(path_to_data)

About extracting class names. they are kindda hidden. Not available inside train_ds.class_names with this builder.

You have to use directly your builder instead.

Getting builder.info (https://www.tensorflow.org/datasets/api_docs/python/tfds/core/DatasetInfo)

Then .features (https://www.tensorflow.org/datasets/api_docs/python/tfds/features/FeaturesDict)

And then you have a ClassLabel (https://www.tensorflow.org/datasets/api_docs/python/tfds/features/ClassLabel) inside 'label' key of FeaturesDict,

print(builder.info.features['label'])
ClassLabel(shape=(), dtype=tf.int64, num_classes=99)

and inside it, you got a names args/attribute describes as:

names: list< str >, string names for the integer classes. The order in which the names are provided is kept.

So, at the end you got your answer directly by:

class_names = builder.info.features['label'].names
user_29603010

You have access to train_ds, val_ds, test_ds, and they have class_names, which you could access using the label as an index, as mentioned by fpierrem

Here's am example of a plot:

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(train_ds.class_names[labels[i]])
        plt.axis("off")
    
plt.show()

see. https://www.tensorflow.org/tutorials/keras/classification#import_the_fashion_mnist_dataset and https://www.tensorflow.org/tutorials/load_data/images#visualize_the_data

user_20484540

You can use train_ds.class_names.

user_173356950
Related