Spaces:
Runtime error
Runtime error
File size: 9,508 Bytes
6a62ffb |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 |
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import torch
from fairseq.utils import new_arange
# -------------- Helper Functions --------------------------------------------------- #
def load_libnat():
try:
from fairseq import libnat_cuda
return libnat_cuda, True
except ImportError as e:
print(str(e) + "... fall back to CPU version")
try:
from fairseq import libnat
return libnat, False
except ImportError as e:
import sys
sys.stderr.write(
"ERROR: missing libnat_cuda. run `python setup.py build_ext --inplace`\n"
)
raise e
def _get_ins_targets(in_tokens, out_tokens, padding_idx, unk_idx):
libnat, use_cuda = load_libnat()
def _get_ins_targets_cuda(in_tokens, out_tokens, padding_idx, unk_idx):
in_masks = in_tokens.ne(padding_idx)
out_masks = out_tokens.ne(padding_idx)
mask_ins_targets, masked_tgt_masks = libnat.generate_insertion_labels(
out_tokens.int(),
libnat.levenshtein_distance(
in_tokens.int(),
out_tokens.int(),
in_masks.sum(1).int(),
out_masks.sum(1).int(),
),
)
masked_tgt_masks = masked_tgt_masks.bool() & out_masks
mask_ins_targets = mask_ins_targets.type_as(in_tokens)[
:, 1 : in_masks.size(1)
].masked_fill_(~in_masks[:, 1:], 0)
masked_tgt_tokens = out_tokens.masked_fill(masked_tgt_masks, unk_idx)
return masked_tgt_masks, masked_tgt_tokens, mask_ins_targets
def _get_ins_targets_cpu(in_tokens, out_tokens, padding_idx, unk_idx):
in_seq_len, out_seq_len = in_tokens.size(1), out_tokens.size(1)
in_tokens_list = [
[t for t in s if t != padding_idx] for i, s in enumerate(in_tokens.tolist())
]
out_tokens_list = [
[t for t in s if t != padding_idx]
for i, s in enumerate(out_tokens.tolist())
]
full_labels = libnat.suggested_ed2_path(
in_tokens_list, out_tokens_list, padding_idx
)
mask_inputs = [
[len(c) if c[0] != padding_idx else 0 for c in a[:-1]] for a in full_labels
]
# generate labels
masked_tgt_masks = []
for mask_input in mask_inputs:
mask_label = []
for beam_size in mask_input[1:-1]: # HACK 1:-1
mask_label += [0] + [1 for _ in range(beam_size)]
masked_tgt_masks.append(
mask_label + [0 for _ in range(out_seq_len - len(mask_label))]
)
mask_ins_targets = [
mask_input[1:-1]
+ [0 for _ in range(in_seq_len - 1 - len(mask_input[1:-1]))]
for mask_input in mask_inputs
]
# transform to tensor
masked_tgt_masks = torch.tensor(
masked_tgt_masks, device=out_tokens.device
).bool()
mask_ins_targets = torch.tensor(mask_ins_targets, device=in_tokens.device)
masked_tgt_tokens = out_tokens.masked_fill(masked_tgt_masks, unk_idx)
return masked_tgt_masks, masked_tgt_tokens, mask_ins_targets
if use_cuda:
return _get_ins_targets_cuda(in_tokens, out_tokens, padding_idx, unk_idx)
return _get_ins_targets_cpu(in_tokens, out_tokens, padding_idx, unk_idx)
def _get_del_targets(in_tokens, out_tokens, padding_idx):
libnat, use_cuda = load_libnat()
def _get_del_targets_cuda(in_tokens, out_tokens, padding_idx):
in_masks = in_tokens.ne(padding_idx)
out_masks = out_tokens.ne(padding_idx)
word_del_targets = libnat.generate_deletion_labels(
in_tokens.int(),
libnat.levenshtein_distance(
in_tokens.int(),
out_tokens.int(),
in_masks.sum(1).int(),
out_masks.sum(1).int(),
),
)
word_del_targets = word_del_targets.type_as(in_tokens).masked_fill_(
~in_masks, 0
)
return word_del_targets
def _get_del_targets_cpu(in_tokens, out_tokens, padding_idx):
out_seq_len = out_tokens.size(1)
with torch.cuda.device_of(in_tokens):
in_tokens_list = [
[t for t in s if t != padding_idx]
for i, s in enumerate(in_tokens.tolist())
]
out_tokens_list = [
[t for t in s if t != padding_idx]
for i, s in enumerate(out_tokens.tolist())
]
full_labels = libnat.suggested_ed2_path(
in_tokens_list, out_tokens_list, padding_idx
)
word_del_targets = [b[-1] for b in full_labels]
word_del_targets = [
labels + [0 for _ in range(out_seq_len - len(labels))]
for labels in word_del_targets
]
# transform to tensor
word_del_targets = torch.tensor(word_del_targets, device=out_tokens.device)
return word_del_targets
if use_cuda:
return _get_del_targets_cuda(in_tokens, out_tokens, padding_idx)
return _get_del_targets_cpu(in_tokens, out_tokens, padding_idx)
def _apply_ins_masks(
in_tokens, in_scores, mask_ins_pred, padding_idx, unk_idx, eos_idx
):
in_masks = in_tokens.ne(padding_idx)
in_lengths = in_masks.sum(1)
# HACK: hacky way to shift all the paddings to eos first.
in_tokens.masked_fill_(~in_masks, eos_idx)
mask_ins_pred.masked_fill_(~in_masks[:, 1:], 0)
out_lengths = in_lengths + mask_ins_pred.sum(1)
out_max_len = out_lengths.max()
out_masks = new_arange(out_lengths, out_max_len)[None, :] < out_lengths[:, None]
reordering = (mask_ins_pred + in_masks[:, 1:].long()).cumsum(1)
out_tokens = (
in_tokens.new_zeros(in_tokens.size(0), out_max_len)
.fill_(padding_idx)
.masked_fill_(out_masks, unk_idx)
)
out_tokens[:, 0] = in_tokens[:, 0]
out_tokens.scatter_(1, reordering, in_tokens[:, 1:])
out_scores = None
if in_scores is not None:
in_scores.masked_fill_(~in_masks, 0)
out_scores = in_scores.new_zeros(*out_tokens.size())
out_scores[:, 0] = in_scores[:, 0]
out_scores.scatter_(1, reordering, in_scores[:, 1:])
return out_tokens, out_scores
def _apply_ins_words(in_tokens, in_scores, word_ins_pred, word_ins_scores, unk_idx):
word_ins_masks = in_tokens.eq(unk_idx)
out_tokens = in_tokens.masked_scatter(word_ins_masks, word_ins_pred[word_ins_masks])
if in_scores is not None:
out_scores = in_scores.masked_scatter(
word_ins_masks, word_ins_scores[word_ins_masks]
)
else:
out_scores = None
return out_tokens, out_scores
def _apply_del_words(
in_tokens, in_scores, in_attn, word_del_pred, padding_idx, bos_idx, eos_idx
):
# apply deletion to a tensor
in_masks = in_tokens.ne(padding_idx)
bos_eos_masks = in_tokens.eq(bos_idx) | in_tokens.eq(eos_idx)
max_len = in_tokens.size(1)
word_del_pred.masked_fill_(~in_masks, 1)
word_del_pred.masked_fill_(bos_eos_masks, 0)
reordering = new_arange(in_tokens).masked_fill_(word_del_pred, max_len).sort(1)[1]
out_tokens = in_tokens.masked_fill(word_del_pred, padding_idx).gather(1, reordering)
out_scores = None
if in_scores is not None:
out_scores = in_scores.masked_fill(word_del_pred, 0).gather(1, reordering)
out_attn = None
if in_attn is not None:
_mask = word_del_pred[:, :, None].expand_as(in_attn)
_reordering = reordering[:, :, None].expand_as(in_attn)
out_attn = in_attn.masked_fill(_mask, 0.0).gather(1, _reordering)
return out_tokens, out_scores, out_attn
def _skip(x, mask):
"""
Getting sliced (dim=0) tensor by mask. Supporting tensor and list/dict of tensors.
"""
if isinstance(x, int):
return x
if x is None:
return None
if isinstance(x, torch.Tensor):
if x.size(0) == mask.size(0):
return x[mask]
elif x.size(1) == mask.size(0):
return x[:, mask]
if isinstance(x, list):
return [_skip(x_i, mask) for x_i in x]
if isinstance(x, dict):
return {k: _skip(v, mask) for k, v in x.items()}
raise NotImplementedError
def _skip_encoder_out(encoder, encoder_out, mask):
if not mask.any():
return encoder_out
else:
return encoder.reorder_encoder_out(
encoder_out, mask.nonzero(as_tuple=False).squeeze()
)
def _fill(x, mask, y, padding_idx):
"""
Filling tensor x with y at masked positions (dim=0).
"""
if x is None:
return y
assert x.dim() == y.dim() and mask.size(0) == x.size(0)
assert x.dim() == 2 or (x.dim() == 3 and x.size(2) == y.size(2))
n_selected = mask.sum()
assert n_selected == y.size(0)
if n_selected == x.size(0):
return y
if x.size(1) < y.size(1):
dims = [x.size(0), y.size(1) - x.size(1)]
if x.dim() == 3:
dims.append(x.size(2))
x = torch.cat([x, x.new_zeros(*dims).fill_(padding_idx)], 1)
x[mask] = y
elif x.size(1) > y.size(1):
x[mask] = padding_idx
if x.dim() == 2:
x[mask, : y.size(1)] = y
else:
x[mask, : y.size(1), :] = y
else:
x[mask] = y
return x
|