Spaces:
Runtime error
Runtime error
File size: 8,603 Bytes
2366e36 |
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 |
# Copyright (c) OpenMMLab. All rights reserved.
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from mmocr.models.builder import DECODERS
from mmocr.models.textrecog.layers import DotProductAttentionLayer
from .base_decoder import BaseDecoder
@DECODERS.register_module()
class SequenceAttentionDecoder(BaseDecoder):
"""Sequence attention decoder for RobustScanner.
RobustScanner: `RobustScanner: Dynamically Enhancing Positional Clues for
Robust Text Recognition <https://arxiv.org/abs/2007.07542>`_
Args:
num_classes (int): Number of output classes :math:`C`.
rnn_layers (int): Number of RNN layers.
dim_input (int): Dimension :math:`D_i` of input vector ``feat``.
dim_model (int): Dimension :math:`D_m` of the model. Should also be the
same as encoder output vector ``out_enc``.
max_seq_len (int): Maximum output sequence length :math:`T`.
start_idx (int): The index of `<SOS>`.
mask (bool): Whether to mask input features according to
``img_meta['valid_ratio']``.
padding_idx (int): The index of `<PAD>`.
dropout (float): Dropout rate.
return_feature (bool): Return feature or logits as the result.
encode_value (bool): Whether to use the output of encoder ``out_enc``
as `value` of attention layer. If False, the original feature
``feat`` will be used.
init_cfg (dict or list[dict], optional): Initialization configs.
Warning:
This decoder will not predict the final class which is assumed to be
`<PAD>`. Therefore, its output size is always :math:`C - 1`. `<PAD>`
is also ignored by loss as specified in
:obj:`mmocr.models.textrecog.recognizer.EncodeDecodeRecognizer`.
"""
def __init__(self,
num_classes=None,
rnn_layers=2,
dim_input=512,
dim_model=128,
max_seq_len=40,
start_idx=0,
mask=True,
padding_idx=None,
dropout=0,
return_feature=False,
encode_value=False,
init_cfg=None):
super().__init__(init_cfg=init_cfg)
self.num_classes = num_classes
self.dim_input = dim_input
self.dim_model = dim_model
self.return_feature = return_feature
self.encode_value = encode_value
self.max_seq_len = max_seq_len
self.start_idx = start_idx
self.mask = mask
self.embedding = nn.Embedding(
self.num_classes, self.dim_model, padding_idx=padding_idx)
self.sequence_layer = nn.LSTM(
input_size=dim_model,
hidden_size=dim_model,
num_layers=rnn_layers,
batch_first=True,
dropout=dropout)
self.attention_layer = DotProductAttentionLayer()
self.prediction = None
if not self.return_feature:
pred_num_classes = num_classes - 1
self.prediction = nn.Linear(
dim_model if encode_value else dim_input, pred_num_classes)
def forward_train(self, feat, out_enc, targets_dict, img_metas):
"""
Args:
feat (Tensor): Tensor of shape :math:`(N, D_i, H, W)`.
out_enc (Tensor): Encoder output of shape
:math:`(N, D_m, H, W)`.
targets_dict (dict): A dict with the key ``padded_targets``, a
tensor of shape :math:`(N, T)`. Each element is the index of a
character.
img_metas (dict): A dict that contains meta information of input
images. Preferably with the key ``valid_ratio``.
Returns:
Tensor: A raw logit tensor of shape :math:`(N, T, C-1)` if
``return_feature=False``. Otherwise it would be the hidden feature
before the prediction projection layer, whose shape is
:math:`(N, T, D_m)`.
"""
valid_ratios = [
img_meta.get('valid_ratio', 1.0) for img_meta in img_metas
] if self.mask else None
targets = targets_dict['padded_targets'].to(feat.device)
tgt_embedding = self.embedding(targets)
n, c_enc, h, w = out_enc.size()
assert c_enc == self.dim_model
_, c_feat, _, _ = feat.size()
assert c_feat == self.dim_input
_, len_q, c_q = tgt_embedding.size()
assert c_q == self.dim_model
assert len_q <= self.max_seq_len
query, _ = self.sequence_layer(tgt_embedding)
query = query.permute(0, 2, 1).contiguous()
key = out_enc.view(n, c_enc, h * w)
if self.encode_value:
value = key
else:
value = feat.view(n, c_feat, h * w)
mask = None
if valid_ratios is not None:
mask = query.new_zeros((n, h, w))
for i, valid_ratio in enumerate(valid_ratios):
valid_width = min(w, math.ceil(w * valid_ratio))
mask[i, :, valid_width:] = 1
mask = mask.bool()
mask = mask.view(n, h * w)
attn_out = self.attention_layer(query, key, value, mask)
attn_out = attn_out.permute(0, 2, 1).contiguous()
if self.return_feature:
return attn_out
out = self.prediction(attn_out)
return out
def forward_test(self, feat, out_enc, img_metas):
"""
Args:
feat (Tensor): Tensor of shape :math:`(N, D_i, H, W)`.
out_enc (Tensor): Encoder output of shape
:math:`(N, D_m, H, W)`.
img_metas (dict): A dict that contains meta information of input
images. Preferably with the key ``valid_ratio``.
Returns:
Tensor: The output logit sequence tensor of shape
:math:`(N, T, C-1)`.
"""
seq_len = self.max_seq_len
batch_size = feat.size(0)
decode_sequence = (feat.new_ones(
(batch_size, seq_len)) * self.start_idx).long()
outputs = []
for i in range(seq_len):
step_out = self.forward_test_step(feat, out_enc, decode_sequence,
i, img_metas)
outputs.append(step_out)
_, max_idx = torch.max(step_out, dim=1, keepdim=False)
if i < seq_len - 1:
decode_sequence[:, i + 1] = max_idx
outputs = torch.stack(outputs, 1)
return outputs
def forward_test_step(self, feat, out_enc, decode_sequence, current_step,
img_metas):
"""
Args:
feat (Tensor): Tensor of shape :math:`(N, D_i, H, W)`.
out_enc (Tensor): Encoder output of shape
:math:`(N, D_m, H, W)`.
decode_sequence (Tensor): Shape :math:`(N, T)`. The tensor that
stores history decoding result.
current_step (int): Current decoding step.
img_metas (dict): A dict that contains meta information of input
images. Preferably with the key ``valid_ratio``.
Returns:
Tensor: Shape :math:`(N, C-1)`. The logit tensor of predicted
tokens at current time step.
"""
valid_ratios = [
img_meta.get('valid_ratio', 1.0) for img_meta in img_metas
] if self.mask else None
embed = self.embedding(decode_sequence)
n, c_enc, h, w = out_enc.size()
assert c_enc == self.dim_model
_, c_feat, _, _ = feat.size()
assert c_feat == self.dim_input
_, _, c_q = embed.size()
assert c_q == self.dim_model
query, _ = self.sequence_layer(embed)
query = query.permute(0, 2, 1).contiguous()
key = out_enc.view(n, c_enc, h * w)
if self.encode_value:
value = key
else:
value = feat.view(n, c_feat, h * w)
mask = None
if valid_ratios is not None:
mask = query.new_zeros((n, h, w))
for i, valid_ratio in enumerate(valid_ratios):
valid_width = min(w, math.ceil(w * valid_ratio))
mask[i, :, valid_width:] = 1
mask = mask.bool()
mask = mask.view(n, h * w)
# [n, c, l]
attn_out = self.attention_layer(query, key, value, mask)
out = attn_out[:, :, current_step]
if self.return_feature:
return out
out = self.prediction(out)
out = F.softmax(out, dim=-1)
return out
|