tf.contrib.distributions: undocumented behaviour in Multinomial and Categorical #10149

import tensorflow as tf
categorical = tf.contrib.distributions.Categorical(probs = [0.25, 0.5, 0.25])
multinomial = tf.contrib.distributions.Multinomial(total_count = 1., probs = [0.25, 0.5, 0.25])
mvn = tf.contrib.distributions.MultivariateNormalDiag(loc = [0., 0., 0.], scale_diag= [1., 1., 1.])

# expected values (points 1 and 2)

# The docs for Categorical say value should be float or double, but it expects an int
categorical.log_prob([0, 0, 1]) 
# <tf.Tensor 'Categorical_2/log_prob/Neg:0' shape=(3,) dtype=float32>
categorical.log_prob([0., 0., 1.])
# TypeError: Value passed to parameter 'labels' has DataType float32 not in list of allowed values: int32, int64

# The docs for Categorical say value should be float or double, which is how it behaves (though this is unlike categorical)
multinomial.log_prob([0, 0, 1]) 
# ValueError: Tensor conversion requested dtype int32 for Tensor with dtype float32: 'Tensor("Multinomial_2/log_prob/Log:0", shape=(3,), dtype=float32)'
multinomial.log_prob([0., 0., 1.])
# <tf.Tensor 'Multinomial_1/log_prob/sub:0' shape=() dtype=float32>

# output shape (points 3 and 4)

# The docs for both say that the output should be:
# "a Tensor of 'shape sample_shape(x) + self.batch_shape' with values of type self.dtype"
# though sample_shape doesn't seem to be relevant here, it's an argument to param_shapes() and sample()

# for Categorical (with int value), the result is a vector, matching the shape of value
categorical.log_prob([0, 0, 1]) 
# <tf.Tensor 'Categorical_2/log_prob/Neg:0' shape=(3,) dtype=float32>

# for Multinomial (with float value), the result is a scalar
multinomial.log_prob([0., 0., 1.])
# <tf.Tensor 'Multinomial_1/log_prob/sub:0' shape=() dtype=float32>

# for Multivariate Normal the result is a scalar
mvn.log_prob([0.1, 0.2, 0.3])
# <tf.Tensor 'MultivariateNormalDiag_2/log_prob/add:0' shape=() dtype=float32>