Spaces:
Running
Running
File size: 7,316 Bytes
9bf4bd7 |
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 |
# Copyright (c) OpenMMLab. All rights reserved.
from typing import Dict, Optional, Sequence, Tuple, Union
import torch
import torch.nn as nn
from mmocr.models.common.dictionary import Dictionary
from mmocr.registry import MODELS
from mmocr.structures import TextRecogDataSample
from .base import BaseDecoder
@MODELS.register_module()
class ASTERDecoder(BaseDecoder):
"""Implement attention decoder.
Args:
in_channels (int): Number of input channels.
emb_dims (int): Dims of char embedding. Defaults to 512.
attn_dims (int): Dims of attention. Both hidden states and features
will be projected to this dims. Defaults to 512.
hidden_size (int): Dims of hidden state for GRU. Defaults to 512.
dictionary (dict or :obj:`Dictionary`): The config for `Dictionary` or
the instance of `Dictionary`. Defaults to None.
max_seq_len (int): Maximum output sequence length :math:`T`. Defaults
to 25.
module_loss (dict, optional): Config to build loss. Defaults to None.
postprocessor (dict, optional): Config to build postprocessor.
Defaults to None.
init_cfg (dict or list[dict], optional): Initialization configs.
Defaults to None.
"""
def __init__(self,
in_channels: int,
emb_dims: int = 512,
attn_dims: int = 512,
hidden_size: int = 512,
dictionary: Union[Dictionary, Dict] = None,
max_seq_len: int = 25,
module_loss: Dict = None,
postprocessor: Dict = None,
init_cfg=dict(type='Xavier', layer='Conv2d')):
super().__init__(
init_cfg=init_cfg,
dictionary=dictionary,
module_loss=module_loss,
postprocessor=postprocessor,
max_seq_len=max_seq_len)
self.start_idx = self.dictionary.start_idx
self.num_classes = self.dictionary.num_classes
self.in_channels = in_channels
self.embedding_dim = emb_dims
self.att_dims = attn_dims
self.hidden_size = hidden_size
# Projection layers
self.proj_feat = nn.Linear(in_channels, attn_dims)
self.proj_hidden = nn.Linear(hidden_size, attn_dims)
self.proj_sum = nn.Linear(attn_dims, 1)
# Decoder input embedding
self.embedding = nn.Embedding(self.num_classes, self.att_dims)
# GRU
self.gru = nn.GRU(
input_size=self.in_channels + self.embedding_dim,
hidden_size=self.hidden_size,
batch_first=True)
# Prediction layer
self.fc = nn.Linear(hidden_size, self.dictionary.num_classes)
self.softmax = nn.Softmax(dim=-1)
def _attention(self, feat: torch.Tensor, prev_hidden: torch.Tensor,
prev_char: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Implement the attention mechanism.
Args:
feat (Tensor): Feature map from encoder of shape :math:`(N, T, C)`.
prev_hidden (Tensor): Previous hidden state from GRU of shape
:math:`(1, N, self.hidden_size)`.
prev_char (Tensor): Previous predicted character of shape
:math:`(N, )`.
Returns:
tuple(Tensor, Tensor):
- output (Tensor): Predicted character of current time step of
shape :math:`(N, 1)`.
- state (Tensor): Hidden state from GRU of current time step of
shape :math:`(N, self.hidden_size)`.
"""
# Calculate the attention weights
B, T, _ = feat.size()
feat_proj = self.proj_feat(feat) # [N, T, attn_dims]
hidden_proj = self.proj_hidden(prev_hidden) # [1, N, attn_dims]
hidden_proj = hidden_proj.squeeze(0).unsqueeze(1) # [N, 1, attn_dims]
hidden_proj = hidden_proj.expand(B, T,
self.att_dims) # [N, T, attn_dims]
sum_tanh = torch.tanh(feat_proj + hidden_proj) # [N, T, attn_dims]
sum_proj = self.proj_sum(sum_tanh).squeeze(-1) # [N, T]
attn_weights = torch.softmax(sum_proj, dim=1) # [N, T]
# GRU forward
context = torch.bmm(attn_weights.unsqueeze(1), feat).squeeze(1)
char_embed = self.embedding(prev_char.long()) # [N, emb_dims]
output, state = self.gru(
torch.cat([char_embed, context], 1).unsqueeze(1), prev_hidden)
output = output.squeeze(1)
output = self.fc(output)
return output, state
def forward_train(
self,
feat: torch.Tensor = None,
out_enc: Optional[torch.Tensor] = None,
data_samples: Optional[Sequence[TextRecogDataSample]] = None
) -> torch.Tensor:
"""
Args:
feat (Tensor): Feature from backbone. Unused in this decoder.
out_enc (torch.Tensor, optional): Encoder output. Defaults to None.
data_samples (list[TextRecogDataSample], optional): Batch of
TextRecogDataSample, containing gt_text information. Defaults
to None.
Returns:
Tensor: The raw logit tensor. Shape :math:`(N, T, C)` where
:math:`C` is ``num_classes``.
"""
B = out_enc.shape[0]
state = torch.zeros(1, B, self.hidden_size).to(out_enc.device)
padded_targets = [
data_sample.gt_text.padded_indexes for data_sample in data_samples
]
padded_targets = torch.stack(padded_targets, dim=0).to(out_enc.device)
outputs = []
for i in range(self.max_seq_len):
prev_char = padded_targets[:, i].to(out_enc.device)
output, state = self._attention(out_enc, state, prev_char)
outputs.append(output)
outputs = torch.cat([_.unsqueeze(1) for _ in outputs], 1)
return outputs
def forward_test(
self,
feat: Optional[torch.Tensor] = None,
out_enc: Optional[torch.Tensor] = None,
data_samples: Optional[Sequence[TextRecogDataSample]] = None
) -> torch.Tensor:
"""
Args:
feat (Tensor): Feature from backbone. Unused in this decoder.
out_enc (torch.Tensor, optional): Encoder output. Defaults to None.
data_samples (list[TextRecogDataSample], optional): Batch of
TextRecogDataSample, containing gt_text information. Defaults
to None. Unused in this decoder.
Returns:
Tensor: The raw logit tensor. Shape :math:`(N, T, C)` where
:math:`C` is ``num_classes``.
"""
B = out_enc.shape[0]
predicted = []
state = torch.zeros(1, B, self.hidden_size).to(out_enc.device)
outputs = []
for i in range(self.max_seq_len):
if i == 0:
prev_char = torch.zeros(B).fill_(self.start_idx).to(
out_enc.device)
else:
prev_char = predicted
output, state = self._attention(out_enc, state, prev_char)
outputs.append(output)
_, predicted = output.max(-1)
outputs = torch.cat([_.unsqueeze(1) for _ in outputs], 1)
return self.softmax(outputs)
|