File size: 6,320 Bytes
30c8b41
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')


class CTCLabelConverter(object):
    """ Convert between text-label and text-index """

    def __init__(self, character):
        # character (str): set of the possible characters.
        dict_character = list(character)

        self.dict = {}
        for i, char in enumerate(dict_character):
            # NOTE: 0 is reserved for 'CTCblank' token required by CTCLoss
            self.dict[char] = i + 1

        self.character = ['[CTCblank]'] + dict_character  # dummy '[CTCblank]' token for CTCLoss (index 0)

    def encode(self, text, batch_max_length=25):
        """convert text-label into text-index.
        input:
            text: text labels of each image. [batch_size]
            batch_max_length: max length of text label in the batch. 25 by default

        output:
            text: text index for CTCLoss. [batch_size, batch_max_length]
            length: length of each text. [batch_size]
        """
        length = [len(s) for s in text]

        # The index used for padding (=0) would not affect the CTC loss calculation.
        batch_text = torch.LongTensor(len(text), batch_max_length).fill_(0)
        for i, t in enumerate(text):
            text = list(t)
            text = [self.dict[char] for char in text]
            batch_text[i][:len(text)] = torch.LongTensor(text)
        return (batch_text.to(device), torch.IntTensor(length).to(device))

    def decode(self, text_index, length):
        """ convert text-index into text-label. """
        texts = []
        for index, l in enumerate(length):
            t = text_index[index, :]

            char_list = []
            for i in range(l):
                if t[i] != 0 and (not (i > 0 and t[i - 1] == t[i])):  # removing repeated characters and blank.
                    char_list.append(self.character[t[i]])
            text = ''.join(char_list)

            texts.append(text)
        return texts


class CTCLabelConverterForBaiduWarpctc(object):
    """ Convert between text-label and text-index for baidu warpctc """

    def __init__(self, character):
        # character (str): set of the possible characters.
        dict_character = list(character)

        self.dict = {}
        for i, char in enumerate(dict_character):
            # NOTE: 0 is reserved for 'CTCblank' token required by CTCLoss
            self.dict[char] = i + 1

        self.character = ['[CTCblank]'] + dict_character  # dummy '[CTCblank]' token for CTCLoss (index 0)

    def encode(self, text, batch_max_length=25):
        """convert text-label into text-index.
        input:
            text: text labels of each image. [batch_size]
        output:
            text: concatenated text index for CTCLoss.
                    [sum(text_lengths)] = [text_index_0 + text_index_1 + ... + text_index_(n - 1)]
            length: length of each text. [batch_size]
        """
        length = [len(s) for s in text]
        text = ''.join(text)
        text = [self.dict[char] for char in text]

        return (torch.IntTensor(text), torch.IntTensor(length))

    def decode(self, text_index, length):
        """ convert text-index into text-label. """
        texts = []
        index = 0
        for l in length:
            t = text_index[index:index + l]

            char_list = []
            for i in range(l):
                if t[i] != 0 and (not (i > 0 and t[i - 1] == t[i])):  # removing repeated characters and blank.
                    char_list.append(self.character[t[i]])
            text = ''.join(char_list)

            texts.append(text)
            index += l
        return texts


class AttnLabelConverter(object):
    """ Convert between text-label and text-index """

    def __init__(self, character):
        # character (str): set of the possible characters.
        # [GO] for the start token of the attention decoder. [s] for end-of-sentence token.
        list_token = ['[GO]', '[s]']  # ['[s]','[UNK]','[PAD]','[GO]']
        list_character = list(character)
        self.character = list_token + list_character

        self.dict = {}
        for i, char in enumerate(self.character):
            # print(i, char)
            self.dict[char] = i

    def encode(self, text, batch_max_length=25):
        """ convert text-label into text-index.
        input:
            text: text labels of each image. [batch_size]
            batch_max_length: max length of text label in the batch. 25 by default

        output:
            text : the input of attention decoder. [batch_size x (max_length+2)] +1 for [GO] token and +1 for [s] token.
                text[:, 0] is [GO] token and text is padded with [GO] token after [s] token.
            length : the length of output of attention decoder, which count [s] token also. [3, 7, ....] [batch_size]
        """
        length = [len(s) + 1 for s in text]  # +1 for [s] at end of sentence.
        # batch_max_length = max(length) # this is not allowed for multi-gpu setting
        batch_max_length += 1
        # additional +1 for [GO] at first step. batch_text is padded with [GO] token after [s] token.
        batch_text = torch.LongTensor(len(text), batch_max_length + 1).fill_(0)
        for i, t in enumerate(text):
            text = list(t)
            text.append('[s]')
            text = [self.dict[char] for char in text]
            batch_text[i][1:1 + len(text)] = torch.LongTensor(text)  # batch_text[:, 0] = [GO] token
        return (batch_text.to(device), torch.IntTensor(length).to(device))

    def decode(self, text_index, length):
        """ convert text-index into text-label. """
        texts = []
        for index, l in enumerate(length):
            text = ''.join([self.character[i] for i in text_index[index, :]])
            texts.append(text)
        return texts


class Averager(object):
    """Compute average for torch.Tensor, used for loss average."""

    def __init__(self):
        self.reset()

    def add(self, v):
        count = v.data.numel()
        v = v.data.sum()
        self.n_count += count
        self.sum += v

    def reset(self):
        self.n_count = 0
        self.sum = 0

    def val(self):
        res = 0
        if self.n_count != 0:
            res = self.sum / float(self.n_count)
        return res