A problem on zs.Concrete about zhusuan HOT 6 CLOSED

@miskcoo Can you have a look? btw, We don't recommend to use zs.Concrete outside BayesianNet. This is a historical problem and we will forbid this behavior in future. Use zs.distributions.Concrete instead.

from zhusuan.

@rtz19970824 I think this is right. Note that zs.Concrete is a multivariate distribution similar to zs.OnehotCategorical, so the input shape is like batch_shape + input_shape, where input_shape is the last dimension of [10, 10], and batch_shape is the first dimension. However, you group last 1 dimension of batch_shape, log_prob will sum out this dim.

from zhusuan.

@miskcoo I'm still a little confused. You mean the whole [10, 10] tensor is viewed as one sample so the output is only the probability of the whole sample? If I want to treat every single [1, 10] tensor as logits of one distribution, how can I do?

from zhusuan.

Don't pass group_ndims=1, this parameter behaves like

prob = tf.reduce_sum(prob, axis=0)

from zhusuan.

Does it work for you? @rtz19970824

from zhusuan.

Yes, thanks.

from zhusuan.