Source code for mmocr.models.textrecog.decoders.nrtr_decoder
# Copyright (c) OpenMMLab. All rights reserved.
import math
from typing import Dict, List, Optional, Sequence, Union
import torch
import torch.nn as nn
from mmengine.model import ModuleList
from mmocr.models.common import PositionalEncoding, TFDecoderLayer
from mmocr.models.common.dictionary import Dictionary
from mmocr.registry import MODELS
from mmocr.structures import TextRecogDataSample
from .base import BaseDecoder
[docs]@MODELS.register_module()
class NRTRDecoder(BaseDecoder):
"""Transformer Decoder block with self attention mechanism.
Args:
n_layers (int): Number of attention layers. Defaults to 6.
d_embedding (int): Language embedding dimension. Defaults to 512.
n_head (int): Number of parallel attention heads. Defaults to 8.
d_k (int): Dimension of the key vector. Defaults to 64.
d_v (int): Dimension of the value vector. Defaults to 64
d_model (int): Dimension :math:`D_m` of the input from previous model.
Defaults to 512.
d_inner (int): Hidden dimension of feedforward layers. Defaults to 256.
n_position (int): Length of the positional encoding vector. Must be
greater than ``max_seq_len``. Defaults to 200.
dropout (float): Dropout rate for text embedding, MHSA, FFN. Defaults
to 0.1.
module_loss (dict, optional): Config to build module_loss. Defaults
to None.
postprocessor (dict, optional): Config to build postprocessor.
Defaults to None.
dictionary (dict or :obj:`Dictionary`): The config for `Dictionary` or
the instance of `Dictionary`.
max_seq_len (int): Maximum output sequence length :math:`T`. Defaults
to 30.
init_cfg (dict or list[dict], optional): Initialization configs.
"""
def __init__(self,
n_layers: int = 6,
d_embedding: int = 512,
n_head: int = 8,
d_k: int = 64,
d_v: int = 64,
d_model: int = 512,
d_inner: int = 256,
n_position: int = 200,
dropout: float = 0.1,
module_loss: Optional[Dict] = None,
postprocessor: Optional[Dict] = None,
dictionary: Optional[Union[Dict, Dictionary]] = None,
max_seq_len: int = 30,
init_cfg: Optional[Union[Dict, List[Dict]]] = None) -> None:
super().__init__(
module_loss=module_loss,
postprocessor=postprocessor,
dictionary=dictionary,
init_cfg=init_cfg,
max_seq_len=max_seq_len)
self.padding_idx = self.dictionary.padding_idx
self.start_idx = self.dictionary.start_idx
self.max_seq_len = max_seq_len
self.trg_word_emb = nn.Embedding(
self.dictionary.num_classes,
d_embedding,
padding_idx=self.padding_idx)
self.position_enc = PositionalEncoding(
d_embedding, n_position=n_position)
self.dropout = nn.Dropout(p=dropout)
self.layer_stack = ModuleList([
TFDecoderLayer(
d_model, d_inner, n_head, d_k, d_v, dropout=dropout)
for _ in range(n_layers)
])
self.layer_norm = nn.LayerNorm(d_model, eps=1e-6)
pred_num_class = self.dictionary.num_classes
self.classifier = nn.Linear(d_model, pred_num_class)
self.softmax = nn.Softmax(dim=-1)
def _get_target_mask(self, trg_seq: torch.Tensor) -> torch.Tensor:
"""Generate mask for target sequence.
Args:
trg_seq (torch.Tensor): Input text sequence. Shape :math:`(N, T)`.
Returns:
Tensor: Target mask. Shape :math:`(N, T, T)`.
E.g.:
seq = torch.Tensor([[1, 2, 0, 0]]), pad_idx = 0, then
target_mask =
torch.Tensor([[[True, False, False, False],
[True, True, False, False],
[True, True, False, False],
[True, True, False, False]]])
"""
pad_mask = (trg_seq != self.padding_idx).unsqueeze(-2)
len_s = trg_seq.size(1)
subsequent_mask = 1 - torch.triu(
torch.ones((len_s, len_s), device=trg_seq.device), diagonal=1)
subsequent_mask = subsequent_mask.unsqueeze(0).bool()
return pad_mask & subsequent_mask
def _get_source_mask(self, src_seq: torch.Tensor,
valid_ratios: Sequence[float]) -> torch.Tensor:
"""Generate mask for source sequence.
Args:
src_seq (torch.Tensor): Image sequence. Shape :math:`(N, T, C)`.
valid_ratios (list[float]): The valid ratio of input image. For
example, if the width of the original image is w1 and the width
after padding is w2, then valid_ratio = w1/w2. Source mask is
used to cover the area of the padding region.
Returns:
Tensor or None: Source mask. Shape :math:`(N, T)`. The region of
padding area are False, and the rest are True.
"""
N, T, _ = src_seq.size()
mask = None
if len(valid_ratios) > 0:
mask = src_seq.new_zeros((N, T), device=src_seq.device)
for i, valid_ratio in enumerate(valid_ratios):
valid_width = min(T, math.ceil(T * valid_ratio))
mask[i, :valid_width] = 1
return mask
def _attention(self,
trg_seq: torch.Tensor,
src: torch.Tensor,
src_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
"""A wrapped process for transformer based decoder including text
embedding, position embedding, N x transformer decoder and a LayerNorm
operation.
Args:
trg_seq (Tensor): Target sequence in. Shape :math:`(N, T)`.
src (Tensor): Source sequence from encoder in shape
Shape :math:`(N, T, D_m)` where :math:`D_m` is ``d_model``.
src_mask (Tensor, Optional): Mask for source sequence.
Shape :math:`(N, T)`. Defaults to None.
Returns:
Tensor: Output sequence from transformer decoder.
Shape :math:`(N, T, D_m)` where :math:`D_m` is ``d_model``.
"""
trg_embedding = self.trg_word_emb(trg_seq)
trg_pos_encoded = self.position_enc(trg_embedding)
trg_mask = self._get_target_mask(trg_seq)
tgt_seq = self.dropout(trg_pos_encoded)
output = tgt_seq
for dec_layer in self.layer_stack:
output = dec_layer(
output,
src,
self_attn_mask=trg_mask,
dec_enc_attn_mask=src_mask)
output = self.layer_norm(output)
return output
[docs] def forward_train(self,
feat: Optional[torch.Tensor] = None,
out_enc: torch.Tensor = None,
data_samples: Sequence[TextRecogDataSample] = None
) -> torch.Tensor:
"""Forward for training. Source mask will be used here.
Args:
feat (Tensor, optional): Unused.
out_enc (Tensor): Encoder output of shape : math:`(N, T, D_m)`
where :math:`D_m` is ``d_model``. Defaults to None.
data_samples (list[TextRecogDataSample]): Batch of
TextRecogDataSample, containing gt_text and valid_ratio
information. Defaults to None.
Returns:
Tensor: The raw logit tensor. Shape :math:`(N, T, C)` where
:math:`C` is ``num_classes``.
"""
valid_ratios = []
for data_sample in data_samples:
valid_ratios.append(data_sample.get('valid_ratio'))
src_mask = self._get_source_mask(out_enc, valid_ratios)
trg_seq = []
for data_sample in data_samples:
trg_seq.append(
data_sample.gt_text.padded_indexes.to(out_enc.device))
trg_seq = torch.stack(trg_seq, dim=0)
attn_output = self._attention(trg_seq, out_enc, src_mask=src_mask)
outputs = self.classifier(attn_output)
return outputs
[docs] def forward_test(self,
feat: Optional[torch.Tensor] = None,
out_enc: torch.Tensor = None,
data_samples: Sequence[TextRecogDataSample] = None
) -> torch.Tensor:
"""Forward for testing.
Args:
feat (Tensor, optional): Unused.
out_enc (Tensor): Encoder output of shape:
math:`(N, T, D_m)` where :math:`D_m` is ``d_model``.
Defaults to None.
data_samples (list[TextRecogDataSample]): Batch of
TextRecogDataSample, containing gt_text and valid_ratio
information. Defaults to None.
Returns:
Tensor: Character probabilities. of shape
:math:`(N, self.max_seq_len, C)` where :math:`C` is
``num_classes``.
"""
valid_ratios = []
for data_sample in data_samples:
valid_ratios.append(data_sample.get('valid_ratio'))
src_mask = self._get_source_mask(out_enc, valid_ratios)
N = out_enc.size(0)
init_target_seq = torch.full((N, self.max_seq_len + 1),
self.padding_idx,
device=out_enc.device,
dtype=torch.long)
# bsz * seq_len
init_target_seq[:, 0] = self.start_idx
outputs = []
for step in range(0, self.max_seq_len):
decoder_output = self._attention(
init_target_seq, out_enc, src_mask=src_mask)
# bsz * seq_len * C
step_result = self.classifier(decoder_output[:, step, :])
# bsz * num_classes
outputs.append(step_result)
_, step_max_index = torch.max(step_result, dim=-1)
init_target_seq[:, step + 1] = step_max_index
outputs = torch.stack(outputs, dim=1)
return self.softmax(outputs)