How to sample batch from only one class at each iteration

Question

I'd like to train a classifier on one ImageNet dataset (1000 classes each with around 1300 images). For some reason, I need each batch to contain 64 images from the same class, and consecutive batches from different classes. Is it possible (and efficient) with the latest TensorFlow?

tf.contrib.data.sample_from_datasets in TF 1.9 allows sampling from a list of tf.data.Dataset objects, with weights indicating the probabilities. I wonder if the following idea makes sense:

• Save data of each class as a separate tfrecord file.
• Pass a tf.data.Dataset.from_generator object as the weights . The object samples from a Categorical distribution such that each sample looks like [0,...,0,1,0,...,0] with 999 0 s and 1 1 ;
• Create 1000 tf.data.Dataset objects, each linked a tfrecord file.

I thought, in this way, maybe at each iteration, sample_from_datasets will first sample a sparse weight vector that indicates which tf.data.Dataset to sample from, then same from that class.

Is it correct? Are there any other efficient ways?

Update

As kindly suggested by P-Gn, one way to sample data from one class would be:

dataset = tf.data.TFRecordDataset(filenames)
dataset = dataset.map(some_parser_fun)  # parse one datum from tfrecord
dataset = dataset.shuffle(buffer_size)

if sample_same_class:
    group_fun = tf.contrib.data.group_by_window(
        key_func=lambda data_x, data_y: data_y,
        reduce_func=lambda key, d: d.batch(batch_size),
        window_size=batch_size)
    dataset = dataset.apply(group_fun)
else:
    dataset = dataset.batch(batch_size)

dataset = dataset.repeat()
data_batch = dataset.make_one_shot_iterator().get_next()

A follow-up question can be found at How to sample batch from a specific class?

Answer 1

I don't think your solution could work, if I understand it correctly, because sample_from_dataset expects a list of values for its weights , not a Tensor .

However if you don't mind having 1000 Dataset s as in your proposed solution, then I would suggest to simply

• create one Dataset per class,
• batch each of these datasets — each batch has samples from a single class,
• zip all of them into one big Dataset of batches,
• shuffle this Dataset — the shuffling will occur on the batches, not on the samples, so it won't change the fact that batches are single class.

A more sophisticated way is to rely on tf.contrib.data.group_by_window . Let me illustrate that with a synthetic example.

import numpy as np
import tensorflow as tf

def gen():
  while True:
    x = np.random.normal()
    label = np.random.randint(10)
    yield x, label

batch_size = 4
batch = (tf.data.Dataset
  .from_generator(gen, (tf.float32, tf.int64), (tf.TensorShape([]), tf.TensorShape([])))
  .apply(tf.contrib.data.group_by_window(
    key_func=lambda x, label: label,
    reduce_func=lambda key, d: d.batch(batch_size),
    window_size=batch_size))
  .make_one_shot_iterator()
  .get_next())

sess = tf.InteractiveSession()
sess.run(batch)
# (array([ 0.04058843,  0.2843775 , -1.8626076 ,  1.1154234 ], dtype=float32),
# array([6, 6, 6, 6], dtype=int64))
sess.run(batch)
# (array([ 1.3600663,  0.5935658, -0.6740045,  1.174328 ], dtype=float32),
# array([3, 3, 3, 3], dtype=int64))

Answer 2

P-Gn's solution to create separate datasets for each class is probably optimal. However, this can be avoided as follows

# Init some dataset
num_classes = 10
label = tf.range(num_classes, dtype=tf.int32)
features = tf.cast(label * 10, dtype=tf.float32) + tf.random.uniform(shape=[tf.shape(label)[0]], maxval=0.01)
dataset = tf.data.Dataset.from_tensor_slices({'label':label, 'features': features})
dataset = dataset.repeat()

# Split to buckets by label
batch_size = 4
dataset = dataset.apply(tf.data.experimental.bucket_by_sequence_length(
    element_length_func=lambda s: s['label'],
    bucket_boundaries=list(range(1, num_classes)),
    bucket_batch_sizes=[batch_size] * num_classes,
))

# Show result
iterator = dataset.as_numpy_iterator()
for i in range(5):
    print(next(iterator))

# {'label': array([0, 0, 0, 0], dtype=int32), 'features': array([0.00370963, 0.00370963, 0.00370963, 0.00370963], dtype=float32)}
# {'label': array([1, 1, 1, 1], dtype=int32), 'features': array([10.009371, 10.009371, 10.009371, 10.009371], dtype=float32)}
# {'label': array([2, 2, 2, 2], dtype=int32), 'features': array([20.001854, 20.001854, 20.001854, 20.001854], dtype=float32)}
# {'label': array([3, 3, 3, 3], dtype=int32), 'features': array([30.005934, 30.005934, 30.005934, 30.005934], dtype=float32)}
# {'label': array([4, 4, 4, 4], dtype=int32), 'features': array([40.001686, 40.001686, 40.001686, 40.001686], dtype=float32)}