NCE or negative sampling layer

DL4J
1

Hi experts,
I have searched the dl4j doc and wonder if we have anything similar to tensorflow’s sampled softmax layer that implements NCE loss /negative loss. tf.nn.sampled_softmax_loss  |  TensorFlow v2.9.1
I know dl4j word2vec model is trained by negative sampling but not sure how to use it in a multiplayer FC network for a general high dimensional classification task.
Thanks in advance

1 Like
2

Unfortunately NCE loss isn’t implemented as standalone loss function for SameDiff yet. Mind filing an issue at github?

3

Thanks for the info. Do we have standalone negative sampling loss function?