def output_ports(self):
"""Returns definitions of module output ports.
input_ids:
indices of tokens which constitute batches of text segments
input_type_ids:
tensor with 0's and 1's to denote the text segment type
input_mask:
bool tensor with 0s in place of tokens to be masked
loss_mask:
used to mask and ignore tokens in the loss function
subtokens_mask:
used to ignore the outputs of unwanted tokens in
the inference and evaluation like the start and end tokens
intents:
intents labels
slots:
slots labels
"""
return {
"input_ids": NeuralType(('B', 'T'), ChannelType()),
"input_type_ids": NeuralType(('B', 'T'), ChannelType()),
"input_mask": NeuralType(('B', 'T'), ChannelType()),
"loss_mask": NeuralType(('B', 'T'), MaskType()),
"subtokens_mask": NeuralType(('B', 'T'), ChannelType()),
"intents": NeuralType(tuple('B'), LabelsType()),
"slots": NeuralType(('B', 'T'), LabelsType()),
}
def output_ports(self):
"""Returns definitions of module output ports.
src_ids: indices of tokens which correspond to source sentences
src_mask: bool tensor with 0s in place of source tokens to be masked
tgt_ids: indices of tokens which correspond to target sentences
tgt_mask: bool tensor with 0s in place of target tokens to be masked
labels: indices of tokens which should be predicted from each of the
corresponding target tokens in tgt_ids; for standard neural
machine translation equals to tgt_ids shifted by 1 to the right
sent_ids: indices of the sentences in a batch; important for
evaluation with external metrics, such as SacreBLEU
"""
return {
# "src_ids": NeuralType({0: AxisType(BatchTag), 1: AxisType(TimeTag)}),
# "src_mask": NeuralType({0: AxisType(BatchTag), 1: AxisType(TimeTag)}),
# "tgt_ids": NeuralType({0: AxisType(BatchTag), 1: AxisType(TimeTag)}),
# "tgt_mask": NeuralType({0: AxisType(BatchTag), 1: AxisType(TimeTag)}),
# "labels": NeuralType({0: AxisType(BatchTag), 1: AxisType(TimeTag)}),
# "sent_ids": NeuralType({0: AxisType(BatchTag)}),
"src_ids": NeuralType(('B', 'T'), ChannelType()),
"src_mask": NeuralType(('B', 'T'), ChannelType()),
"tgt_ids": NeuralType(('B', 'T'), ChannelType()),
"tgt_mask": NeuralType(('B', 'T'), ChannelType()),
"labels": NeuralType(('B', 'T'), LabelsType()),
"sent_ids": NeuralType(tuple('B'), ChannelType()),
}
def input_ports(self):
"""Returns definitions of module input ports.
"""
return {
# "logits": NeuralType({0: AxisType(BatchTag), 1: AxisType(TimeTag), 2: AxisType(ChannelTag)}),
# "start_positions": NeuralType({0: AxisType(BatchTag)}),
# "end_positions": NeuralType({0: AxisType(BatchTag)}),
"logits": NeuralType(('B', 'T', 'D'), LogitsType()),
"start_positions": NeuralType(tuple('B'), ChannelType()),
"end_positions": NeuralType(tuple('B'), ChannelType()),
}
def input_ports(self):
"""Returns definitions of module input ports.
"""
return {
# "intent_logits": NeuralType({0: AxisType(BatchTag), 1: AxisType(ChannelTag)}),
# "slot_logits": NeuralType({0: AxisType(BatchTag), 1: AxisType(TimeTag), 2: AxisType(ChannelTag)}),
# "loss_mask": NeuralType({0: AxisType(BatchTag), 1: AxisType(TimeTag)}),
# "intents": NeuralType({0: AxisType(BatchTag)}),
# "slots": NeuralType({0: AxisType(BatchTag), 1: AxisType(TimeTag)}),
"intent_logits": NeuralType(('B', 'D'), LogitsType()),
"slot_logits": NeuralType(('B', 'T', 'D'), LogitsType()),
"loss_mask": NeuralType(('B', 'T'), ChannelType()),
"intents": NeuralType(tuple('B'), ChannelType()),
"slots": NeuralType(('B', 'T'), ChannelType()),
}
