Summary: 28 instances, 23 unique

Text	Count
# TODO: return unnormalized proper score	1
assert not prevent_unk, "Greedy Search does not support prevention of unknown tokens"  # TODO: add support	2
# TODO: while we still have both transformer.TransformerConfig and transformer_pt.TransformerConfig,	1
# TODO: possibly support decoding on multiple GPUs	1
# TODO: consider a dictionary mapping as return value	1
self._reduction = 'mean'  # TODO: consider sum reduction and normalization outside of loss for reporting	1
# TODO: This is a legacy step from the bucketing module version of Sockeye.	1
# TODO: consider automatically selecting GreedySearch if flags to this method are compatible.	2
# TODO: scoring should support multiple devices	1
# TODO: check for missing parameters somehow (we allowed scaling to be missing)	1
# TODO: check for disabled predicted output length	2
# TODO: make this configurable in the model, separately per target factor.	1
# TODO: Consider joint tensor for all target factors	1
# TODO: deprecate mxnet CLIs	1
# TODO: figure out int8 quantization of OutputLayer, supporting weight tying & vocabulary selection	1
# TODO also consider weight tying with target factor input embeddings	2
# TODO (fhieber): Consider making inference classes regular modules (or move logic into the model module)	1
# overwriting here. TODO: make this better...	1
# TODO replace options list (e.g ENCODERS, DECODERS, ...) with Enum classes	1
score = np.array([-1.])  # TODO: return unnormalized proper score	1
# TODO: consider using gain=1 / math.sqrt(2)	1
# TODO DEPRECATE, NO LONGER USED	1
# TODO: port NVIDIAs implementation to PT C++ custom op	2