1343464520 / am_softmax

This is a pytorch implementation of the am_softmax, this softmax layer includes the class assignment fully connected layer, as it is required for it to be normalized.

Pytorch implementation of additive margin softmax as seen on original paper
https://arxiv.org/pdf/1801.05599.pdf
This is a loss function designed for embedding learning as it encourages intra-class distances to be small and inter-class distances to be larger, so the input to the layer should be a batchXembedding tensor and the output will be softmax logits which can then be passed to a NLL implementation, eg. input size 64X512 means 64 embeddings of size 512 each.