Difficulties with `nn.sparse_softmax_cross_entropy_with_logits` about tensorflow.jl HOT 6 CLOSED

I will take a look soonish

from tensorflow.jl.

I checked; there is a genuine issue with taking gradients of SparseSoftmaxCrossEntryWithLogits operations. Not sure of the cause yet though.

from tensorflow.jl.

This is fixed by JuliaIO/ProtoBuf.jl#87

from tensorflow.jl.

This is expected, is it not? The docs for nn.sparse_softmax_cross_entropy_with_logits state that the labels and predicted labels need to have the same shape. Take a look at the shape of logits and the shape of labels.

Nope, I'm dead wrong, there's no strict relationship defined between the number of valid label values and the number of output nodes in the net. I was confused between the docs of sparse_softmax_cross_entropy_with_logits and softmax_cross_entropy_with_logits.

from tensorflow.jl.

If you alter the shape of your weights from [3, 3] to [3, 2], it works.

from tensorflow.jl.

Closed by JuliaLang/METADATA.jl#7811 (comment)

from tensorflow.jl.