Creating a tensorflow dataset that outputs a dict

Question

I have a dict with "metadata" for my dataset, of sort {'m1': array_1, 'm2': array_2, ...}. Each of the arrays has shape (N, ...), where N is the number of samples.

The question: Is it possible to create a tf.data.Dataset that outputs a dictionary {'meta_1': sub_array_1, 'meta_2': sub_array_2, ...} for each iteration of the datasets iterator.get_next()? Here, sub_array_i should contain the ith metadata for one batch, so should have shape (batch_sz, ...).

What I tried so far is using tf.data.Dataset.from_generator(), like this:

N = 100
# dictionary of arrays:
metadata = {'m1': np.zeros(shape=(N,2)), 'm2': np.ones(shape=(N,3,5))} 
num_samples = N

def meta_dict_gen():
    for i in range(num_samples):
        ls = {}
        for key, val in metadata.items():
            ls[key] = val[i]
        yield ls

dataset = tf.data.Dataset.from_generator(meta_dict_gen, output_types=(dict))

The problem with this seems to be in output_types=(dict). The code above throws at me a

TypeError: Expected DataType for argument 'Tout' not < class 'dict'>.

---

I'm using tensorflow 1.8 and python 3.6.

Answer 1

So actually it is possible to do what you intend, you just have to be specific about the contents of the dict:

import tensorflow as tf
import numpy as np

N = 100
# dictionary of arrays:
metadata = {'m1': np.zeros(shape=(N,2)), 'm2': np.ones(shape=(N,3,5))}
num_samples = N

def meta_dict_gen():
    for i in range(num_samples):
        ls = {}
        for key, val in metadata.items():
            ls[key] = val[i]
        yield ls

dataset = tf.data.Dataset.from_generator(
    meta_dict_gen,
    output_types={k: tf.float32 for k in metadata},
    output_shapes={'m1': (2,), 'm2': (3, 5)})
iter = dataset.make_one_shot_iterator()
next_elem = iter.get_next()
print(next_elem)

Output:

{'m1': <tf.Tensor 'IteratorGetNext:0' shape=(2,) dtype=float32>,
 'm2': <tf.Tensor 'IteratorGetNext:1' shape=(3, 5) dtype=float32>}