PytorchOCR/torchocr/networks/necks/RecSequenceDecoder.py

# -*- coding: utf-8 -*-
# @Time    : 2020/5/15 17:46
# @Author  : zhoujun

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

from collections import OrderedDict

from torch import nn
import torch


class DecoderWithRNN(nn.Module):
    def __init__(self, in_channels, **kwargs):
        super().__init__()
        rnn_hidden_size = kwargs.get('hidden_size', 96)
        self.out_channels = rnn_hidden_size * 2
        self.layers = 2
        self.lstm = nn.LSTM(in_channels, rnn_hidden_size, bidirectional=True, batch_first=True, num_layers=self.layers)

    def forward(self, x):
        x = self.lstm(x)[0]
        return x


class Reshape(nn.Module):
    def __init__(self, in_channels, **kwargs):
        super().__init__()
        self.out_channels = in_channels

    def forward(self, x):
        B, C, H, W = x.shape
        x = x.reshape(B, C, H * W)
        x = x.permute((0, 2, 1))  # (NTC)(batch, width, channel)s
        return x


class SequenceDecoder(nn.Module):
    def __init__(self, in_channels, **kwargs):
        super().__init__()
        self.reshape = Reshape(in_channels)
        self.decoder = DecoderWithRNN(in_channels, **kwargs)
        self.out_channels = self.decoder.out_channels

    def forward(self, x):
        x = self.reshape(x)
        x = self.decoder(x)
        return x