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
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
|
#include <pybind11/pybind11.h>
#include <pybind11/stl.h>
#include <torch/csrc/jit/python/module_python.h> // @manual
#include <torch/csrc/utils/pybind.h> // @manual
#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 <torchtext/csrc/vocab_factory.h> // @manual
#include <iostream>
namespace torchtext {
namespace py = pybind11;
namespace {
Vocab build_vocab_from_text_file(
const std::string& file_path,
const int64_t min_freq,
const int64_t num_cpus,
py::object fn) {
torch::jit::script::Module module(*torch::jit::as_module(fn));
return _build_vocab_from_text_file(file_path, min_freq, num_cpus, module);
}
} // namespace
// Registers our custom classes with pybind11.
PYBIND11_MODULE(_torchtext, m) {
// Classes
py::class_<Regex, c10::intrusive_ptr<Regex>>(m, "Regex")
.def(py::init<std::string>())
.def("Sub", &Regex::Sub)
.def("FindAndConsume", &Regex::FindAndConsume)
.def(py::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));
}));
py::class_<RegexTokenizer, c10::intrusive_ptr<RegexTokenizer>>(
m, "RegexTokenizer")
.def_readonly("patterns_", &RegexTokenizer::patterns_)
.def_readonly("replacements_", &RegexTokenizer::replacements_)
.def_readonly("to_lower_", &RegexTokenizer::to_lower_)
.def(py::init<std::vector<std::string>, std::vector<std::string>, bool>())
.def("forward", &RegexTokenizer::forward)
.def(py::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));
}));
py::class_<SentencePiece, c10::intrusive_ptr<SentencePiece>>(
m, "SentencePiece")
.def(py::init<std::string>())
.def(
"_return_content",
[](const SentencePiece& self) { return py::bytes(self.content_); })
.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(py::pickle(
// __getstate__
[](const c10::intrusive_ptr<SentencePiece>& self) -> py::bytes {
return py::bytes(self->content_);
},
// __setstate__
[](py::bytes state) -> c10::intrusive_ptr<SentencePiece> {
return c10::make_intrusive<SentencePiece>(std::string(state));
}));
py::class_<Vectors, c10::intrusive_ptr<Vectors>>(m, "Vectors")
.def(py::init<
std::vector<std::string>,
std::vector<int64_t>,
torch::Tensor,
torch::Tensor>())
.def_readonly("vectors_", &Vectors::vectors_)
.def_readonly("unk_tensor_", &Vectors::unk_tensor_)
.def("get_stoi", &Vectors::get_stoi)
.def("__getitem__", &Vectors::__getitem__)
.def("lookup_vectors", &Vectors::lookup_vectors)
.def("__setitem__", &Vectors::__setitem__)
.def("__len__", &Vectors::__len__)
.def(py::pickle(
// __getstate__
[](const c10::intrusive_ptr<Vectors>& self) -> VectorsStates {
return _serialize_vectors(self);
},
// __setstate__
[](VectorsStates states) -> c10::intrusive_ptr<Vectors> {
return _deserialize_vectors(states);
}));
py::class_<Vocab, c10::intrusive_ptr<Vocab>>(m, "Vocab")
.def(py::init<StringList, c10::optional<int64_t>>())
.def_readonly("itos_", &Vocab::itos_)
.def_readonly("default_index_", &Vocab::default_index_)
.def(
"__contains__",
[](c10::intrusive_ptr<Vocab>& self, const py::str& item) -> bool {
Py_ssize_t length;
const char* buffer = PyUnicode_AsUTF8AndSize(item.ptr(), &length);
return self->__contains__(c10::string_view{buffer, (size_t)length});
})
.def(
"__getitem__",
[](c10::intrusive_ptr<Vocab>& self, const py::str& item) -> int64_t {
Py_ssize_t length;
const char* buffer = PyUnicode_AsUTF8AndSize(item.ptr(), &length);
return self->__getitem__(c10::string_view{buffer, (size_t)length});
})
.def("insert_token", &Vocab::insert_token)
.def("set_default_index", &Vocab::set_default_index)
.def("get_default_index", &Vocab::get_default_index)
.def("__len__", &Vocab::__len__)
.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 py::list& items) {
std::vector<int64_t> indices(items.size());
int64_t counter = 0;
for (const auto& item : items) {
Py_ssize_t length;
const char* buffer = PyUnicode_AsUTF8AndSize(item.ptr(), &length);
indices[counter++] =
self->__getitem__(c10::string_view{buffer, (size_t)length});
}
return indices;
})
.def("get_stoi", &Vocab::get_stoi)
.def("get_itos", &Vocab::get_itos)
.def(py::pickle(
// __getstate__
[](const c10::intrusive_ptr<Vocab>& self) -> VocabStates {
return _serialize_vocab(self);
},
// __setstate__
[](VocabStates states) -> c10::intrusive_ptr<Vocab> {
return _deserialize_vocab(states);
}));
py::class_<GPT2BPEEncoder, c10::intrusive_ptr<GPT2BPEEncoder>>(
m, "GPT2BPEEncoder")
.def(py::init<
std::unordered_map<std::string, int64_t>,
std::unordered_map<std::string, int64_t>,
std::string,
std::unordered_map<int64_t, std::string>,
bool>())
.def_property_readonly("bpe_encoder_", &GPT2BPEEncoder::GetBPEEncoder)
.def_property_readonly(
"bpe_merge_ranks_", &GPT2BPEEncoder::GetBPEMergeRanks)
.def_readonly("seperator_", &GPT2BPEEncoder::seperator_)
.def_property_readonly("byte_encoder_", &GPT2BPEEncoder::GetByteEncoder)
.def("encode", &GPT2BPEEncoder::Encode)
.def("tokenize", &GPT2BPEEncoder::Tokenize)
.def(py::pickle(
// __getstate__
[](const c10::intrusive_ptr<GPT2BPEEncoder>& self)
-> GPT2BPEEncoderStatesPybind {
return _serialize_gpt2_bpe_encoder_pybind(self);
},
// __setstate__
[](GPT2BPEEncoderStatesPybind states)
-> c10::intrusive_ptr<GPT2BPEEncoder> {
return _deserialize_gpt2_bpe_encoder_pybind(states);
}));
py::class_<CLIPEncoder, c10::intrusive_ptr<CLIPEncoder>>(m, "CLIPEncoder")
.def(py::init<
std::unordered_map<std::string, int64_t>,
std::unordered_map<std::string, int64_t>,
std::string,
std::unordered_map<int64_t, std::string>,
bool>())
.def_property_readonly("bpe_encoder_", &CLIPEncoder::GetBPEEncoder)
.def_property_readonly("bpe_merge_ranks_", &CLIPEncoder::GetBPEMergeRanks)
.def_readonly("seperator_", &CLIPEncoder::seperator_)
.def_property_readonly("byte_encoder_", &CLIPEncoder::GetByteEncoder)
.def("encode", &CLIPEncoder::Encode)
.def("tokenize", &CLIPEncoder::Tokenize)
.def(py::pickle(
// __getstate__
[](const c10::intrusive_ptr<CLIPEncoder>& self)
-> CLIPEncoderStatesPybind {
return _serialize_clip_encoder_pybind(self);
},
// __setstate__
[](CLIPEncoderStatesPybind states)
-> c10::intrusive_ptr<CLIPEncoder> {
return _deserialize_clip_encoder_pybind(states);
}));
py::class_<BERTEncoder, c10::intrusive_ptr<BERTEncoder>>(m, "BERTEncoder")
.def(py::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 py::list& items) {
std::vector<std::string> input;
for (const auto& item : items) {
Py_ssize_t length;
const char* buffer = PyUnicode_AsUTF8AndSize(item.ptr(), &length);
input.push_back(std::string(buffer));
}
return self->BatchEncode(input);
})
.def(
"batch_tokenize",
[](const c10::intrusive_ptr<BERTEncoder>& self,
const py::list& items) {
std::vector<std::string> input;
for (const auto& item : items) {
Py_ssize_t length;
const char* buffer = PyUnicode_AsUTF8AndSize(item.ptr(), &length);
input.push_back(std::string(buffer));
}
return self->BatchTokenize(input);
})
.def(py::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);
}));
// Functions
m.def(
"_load_token_and_vectors_from_file", &_load_token_and_vectors_from_file);
m.def("_load_vocab_from_file", &_load_vocab_from_file);
m.def("_build_vocab_from_text_file", &build_vocab_from_text_file);
m.def(
"_build_vocab_from_text_file_using_python_tokenizer",
&_build_vocab_from_text_file_using_python_tokenizer);
}
} // namespace torchtext
|
