File size: 4,993 Bytes
ad16788 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 |
"""ScorerInterface implementation for CTC."""
import numpy as np
import torch
from espnet.nets.ctc_prefix_score import CTCPrefixScore
from espnet.nets.ctc_prefix_score import CTCPrefixScoreTH
from espnet.nets.scorer_interface import BatchPartialScorerInterface
class CTCPrefixScorer(BatchPartialScorerInterface):
"""Decoder interface wrapper for CTCPrefixScore."""
def __init__(self, ctc: torch.nn.Module, eos: int):
"""Initialize class.
Args:
ctc (torch.nn.Module): The CTC implementaiton.
For example, :class:`espnet.nets.pytorch_backend.ctc.CTC`
eos (int): The end-of-sequence id.
"""
self.ctc = ctc
self.eos = eos
self.impl = None
def init_state(self, x: torch.Tensor):
"""Get an initial state for decoding.
Args:
x (torch.Tensor): The encoded feature tensor
Returns: initial state
"""
logp = self.ctc.log_softmax(x.unsqueeze(0)).detach().squeeze(0).cpu().numpy()
# TODO(karita): use CTCPrefixScoreTH
self.impl = CTCPrefixScore(logp, 0, self.eos, np)
return 0, self.impl.initial_state()
def select_state(self, state, i, new_id=None):
"""Select state with relative ids in the main beam search.
Args:
state: Decoder state for prefix tokens
i (int): Index to select a state in the main beam search
new_id (int): New label id to select a state if necessary
Returns:
state: pruned state
"""
if type(state) == tuple:
if len(state) == 2: # for CTCPrefixScore
sc, st = state
return sc[i], st[i]
else: # for CTCPrefixScoreTH (need new_id > 0)
r, log_psi, f_min, f_max, scoring_idmap = state
s = log_psi[i, new_id].expand(log_psi.size(1))
if scoring_idmap is not None:
return r[:, :, i, scoring_idmap[i, new_id]], s, f_min, f_max
else:
return r[:, :, i, new_id], s, f_min, f_max
return None if state is None else state[i]
def score_partial(self, y, ids, state, x):
"""Score new token.
Args:
y (torch.Tensor): 1D prefix token
next_tokens (torch.Tensor): torch.int64 next token to score
state: decoder state for prefix tokens
x (torch.Tensor): 2D encoder feature that generates ys
Returns:
tuple[torch.Tensor, Any]:
Tuple of a score tensor for y that has a shape `(len(next_tokens),)`
and next state for ys
"""
prev_score, state = state
presub_score, new_st = self.impl(y.cpu(), ids.cpu(), state)
tscore = torch.as_tensor(
presub_score - prev_score, device=x.device, dtype=x.dtype
)
return tscore, (presub_score, new_st)
def batch_init_state(self, x: torch.Tensor):
"""Get an initial state for decoding.
Args:
x (torch.Tensor): The encoded feature tensor
Returns: initial state
"""
logp = self.ctc.log_softmax(x.unsqueeze(0)) # assuming batch_size = 1
xlen = torch.tensor([logp.size(1)])
self.impl = CTCPrefixScoreTH(logp, xlen, 0, self.eos)
return None
def batch_score_partial(self, y, ids, state, x):
"""Score new token.
Args:
y (torch.Tensor): 1D prefix token
ids (torch.Tensor): torch.int64 next token to score
state: decoder state for prefix tokens
x (torch.Tensor): 2D encoder feature that generates ys
Returns:
tuple[torch.Tensor, Any]:
Tuple of a score tensor for y that has a shape `(len(next_tokens),)`
and next state for ys
"""
batch_state = (
(
torch.stack([s[0] for s in state], dim=2),
torch.stack([s[1] for s in state]),
state[0][2],
state[0][3],
)
if state[0] is not None
else None
)
return self.impl(y, batch_state, ids)
def extend_prob(self, x: torch.Tensor):
"""Extend probs for decoding.
This extention is for streaming decoding
as in Eq (14) in https://arxiv.org/abs/2006.14941
Args:
x (torch.Tensor): The encoded feature tensor
"""
logp = self.ctc.log_softmax(x.unsqueeze(0))
self.impl.extend_prob(logp)
def extend_state(self, state):
"""Extend state for decoding.
This extention is for streaming decoding
as in Eq (14) in https://arxiv.org/abs/2006.14941
Args:
state: The states of hyps
Returns: exteded state
"""
new_state = []
for s in state:
new_state.append(self.impl.extend_state(s))
return new_state
|