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
|
#include <torch/script.h>
#include <torchtext/csrc/bert_tokenizer.h> // @manual
#include <torchtext/csrc/clip_tokenizer.h> // @manual
#include <torchtext/csrc/gpt2_bpe_tokenizer.h> // @manual
#include <torchtext/csrc/regex.h>
#include <torchtext/csrc/regex_tokenizer.h> // @manual
#include <torchtext/csrc/sentencepiece.h> // @manual
#include <torchtext/csrc/vectors.h> // @manual
#include <torchtext/csrc/vocab.h> // @manual
#include <iostream>
namespace torchtext {
TORCH_LIBRARY_FRAGMENT(torchtext, m) {
m.class_<Regex>("Regex")
.def(torch::init<std::string>())
.def("Sub", &Regex::Sub)
.def_pickle(
// __getstate__
[](const c10::intrusive_ptr<Regex>& self) -> std::string {
return _serialize_regex(self);
},
// __setstate__
[](std::string state) -> c10::intrusive_ptr<Regex> {
return _deserialize_regex(std::move(state));
});
m.class_<RegexTokenizer>("RegexTokenizer")
.def(torch::
init<std::vector<std::string>, std::vector<std::string>, bool>())
.def("forward", &RegexTokenizer::forward)
.def_pickle(
// __getstate__
[](const c10::intrusive_ptr<RegexTokenizer>& self)
-> RegexTokenizerStates {
return _serialize_regex_tokenizer(self);
},
// __setstate__
[](RegexTokenizerStates states)
-> c10::intrusive_ptr<RegexTokenizer> {
return _deserialize_regex_tokenizer(std::move(states));
});
m.class_<SentencePiece>("SentencePiece")
.def(torch::init<std::string>())
.def("Encode", &SentencePiece::Encode)
.def("EncodeAsIds", &SentencePiece::EncodeAsIds)
.def("DecodeIds", &SentencePiece::DecodeIds)
.def("EncodeAsPieces", &SentencePiece::EncodeAsPieces)
.def("DecodePieces", &SentencePiece::DecodePieces)
.def("GetPieceSize", &SentencePiece::GetPieceSize)
.def("unk_id", &SentencePiece::unk_id)
.def("PieceToId", &SentencePiece::PieceToId)
.def("IdToPiece", &SentencePiece::IdToPiece)
.def_pickle(
// The underlying content of SentencePiece contains byte string,
// and returing it as std::string cause UTF8 decoding error.
// Since TorchScript does not support byte string, we use byte Tensor
// to pass around the data.
// __getstate__
[](const c10::intrusive_ptr<SentencePiece>& self) -> torch::Tensor {
auto* data =
static_cast<void*>(const_cast<char*>(self->content_.data()));
auto numel = static_cast<int64_t>(self->content_.size());
return torch::from_blob(data, {numel}, {torch::kUInt8}).clone();
},
// __setstate__
[](torch::Tensor state) -> c10::intrusive_ptr<SentencePiece> {
auto* data = static_cast<char*>(state.data_ptr());
auto numel = state.size(0);
return c10::make_intrusive<SentencePiece>(std::string(data, numel));
});
m.class_<Vectors>("Vectors")
.def(torch::init<
std::vector<std::string>,
std::vector<std::int64_t>,
torch::Tensor,
torch::Tensor>())
.def("__getitem__", &Vectors::__getitem__)
.def("lookup_vectors", &Vectors::lookup_vectors)
.def("__setitem__", &Vectors::__setitem__)
.def("__len__", &Vectors::__len__)
.def_pickle(
// __getstate__
[](const c10::intrusive_ptr<Vectors>& self) -> VectorsStates {
return _serialize_vectors(self);
},
// __setstate__
[](VectorsStates states) -> c10::intrusive_ptr<Vectors> {
return _deserialize_vectors(states);
});
m.class_<Vocab>("Vocab")
.def(torch::init<StringList, c10::optional<int64_t>>())
.def(
"__contains__",
[](const c10::intrusive_ptr<Vocab>& self, const std::string& item)
-> bool { return self->__contains__(c10::string_view{item}); })
.def(
"__getitem__",
[](const c10::intrusive_ptr<Vocab>& self, const std::string& item)
-> int64_t { return self->__getitem__(c10::string_view{item}); })
.def("insert_token", &Vocab::insert_token)
.def("__len__", &Vocab::__len__)
.def("set_default_index", &Vocab::set_default_index)
.def("get_default_index", &Vocab::get_default_index)
.def("append_token", &Vocab::append_token)
.def("lookup_token", &Vocab::lookup_token)
.def("lookup_tokens", &Vocab::lookup_tokens)
.def(
"lookup_indices",
[](const c10::intrusive_ptr<Vocab>& self,
const std::vector<std::string>& items) {
std::vector<int64_t> indices(items.size());
int64_t counter = 0;
for (const auto& item : items) {
indices[counter++] = self->__getitem__(c10::string_view{item});
}
return indices;
})
.def("get_stoi", &Vocab::get_stoi)
.def("get_itos", &Vocab::get_itos)
.def_pickle(
// __getstate__
[](const c10::intrusive_ptr<Vocab>& self) -> VocabStates {
return _serialize_vocab(self);
},
// __setstate__
[](VocabStates states) -> c10::intrusive_ptr<Vocab> {
return _deserialize_vocab(states);
});
m.class_<GPT2BPEEncoder>("GPT2BPEEncoder")
.def(torch::init<
c10::Dict<std::string, int64_t>,
c10::Dict<std::string, int64_t>,
std::string,
c10::Dict<int64_t, std::string>,
bool>())
.def("encode", &GPT2BPEEncoder::Encode)
.def("tokenize", &GPT2BPEEncoder::Tokenize)
.def_pickle(
// __getstate__
[](const c10::intrusive_ptr<GPT2BPEEncoder>& self)
-> GPT2BPEEncoderStatesTorchbind {
return _serialize_gpt2_bpe_encoder_torchbind(self);
},
// __setstate__
[](GPT2BPEEncoderStatesTorchbind states)
-> c10::intrusive_ptr<GPT2BPEEncoder> {
return _deserialize_gpt2_bpe_encoder_torchbind(states);
});
m.class_<CLIPEncoder>("CLIPEncoder")
.def(torch::init<
c10::Dict<std::string, int64_t>,
c10::Dict<std::string, int64_t>,
std::string,
c10::Dict<int64_t, std::string>,
bool>())
.def("encode", &CLIPEncoder::Encode)
.def("tokenize", &CLIPEncoder::Tokenize)
.def_pickle(
// __getstate__
[](const c10::intrusive_ptr<CLIPEncoder>& self)
-> CLIPEncoderStatesTorchbind {
return _serialize_clip_encoder_torchbind(self);
},
// __setstate__
[](CLIPEncoderStatesTorchbind states)
-> c10::intrusive_ptr<CLIPEncoder> {
return _deserialize_clip_encoder_torchbind(states);
});
m.class_<BERTEncoder>("BERTEncoder")
.def(torch::init<
const std::string,
bool,
c10::optional<bool>,
std::vector<std::string>>())
.def("encode", &BERTEncoder::Encode)
.def("tokenize", &BERTEncoder::Tokenize)
.def(
"batch_encode",
[](const c10::intrusive_ptr<BERTEncoder>& self,
const std::vector<std::string>& items) {
return self->BatchEncode(items);
})
.def(
"batch_tokenize",
[](const c10::intrusive_ptr<BERTEncoder>& self,
const std::vector<std::string>& items) {
return self->BatchTokenize(items);
})
.def_pickle(
// __getstate__
[](const c10::intrusive_ptr<BERTEncoder>& self) -> BERTEncoderStates {
return _serialize_bert_encoder(self);
},
// __setstate__
[](BERTEncoderStates states) -> c10::intrusive_ptr<BERTEncoder> {
return _deserialize_bert_encoder(states);
});
;
m.def("torchtext::generate_sp_model", &generate_sp_model);
m.def("torchtext::load_sp_model", &load_sp_model);
m.def("torchtext::load_sp_model_string", &load_sp_model_string);
m.def("torchtext::gpt2_bpe_pre_tokenizer", &gpt2_bpe_pre_tokenizer);
}
} // namespace torchtext
|
