Add layers to do the BiGRU index selecting for the question-evidence common word feature in the Gated Attention Reader, and also an Overlap layer.
Overlap has some hacky casting in the mask to get around the bug in the tensorflow backend of keras that casts unmasked input to uint8 before concatenation.
Add layers to do the BiGRU index selecting for the question-evidence common word feature in the Gated Attention Reader, and also an
Overlap
layer.Overlap has some hacky casting in the mask to get around the bug in the tensorflow backend of keras that casts unmasked input to
uint8
before concatenation.built docs: biGRU index selector in api summary: https://137-47793051-gh.circle-artifacts.com/0/home/ubuntu/deep_qa/doc/_build/html/layers/core_layers.html#bigruindexselector biGRU index selector in full api reference: https://137-47793051-gh.circle-artifacts.com/0/home/ubuntu/deep_qa/doc/_build/html/api_doc/deep_qa.layers.html#module-deep_qa.layers.bigru_index_selector
overlap in api summary: https://137-47793051-gh.circle-artifacts.com/0/home/ubuntu/deep_qa/doc/_build/html/layers/core_layers.html#overlap overlap in full api reference: https://137-47793051-gh.circle-artifacts.com/0/home/ubuntu/deep_qa/doc/_build/html/api_doc/deep_qa.layers.html#module-deep_qa.layers.overlap