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
|
// Note(jiayq): the import_array function is done inside
// caffe2_python.cc. Read
// http://docs.scipy.org/doc/numpy-1.10.1/reference/c-api.array.html#miscellaneous
// for more details.
#define NO_IMPORT_ARRAY
#include "pybind_state.h"
#include <pybind11/pybind11.h>
#include <pybind11/stl.h>
#include <caffe2/ideep/ideep_utils.h>
#include "caffe2/ideep/operators/operator_fallback_ideep.h"
namespace caffe2 {
namespace python {
USE_IDEEP_DEF_ALIASES();
class IDeepFetcher;
class IDeepFeeder;
REGISTER_IDEEP_OPERATOR(Python, IDEEPFallbackOp<PythonOp<CPUContext, false>>);
REGISTER_BLOB_FETCHER((TypeMeta::Id<itensor>()), IDeepFetcher);
REGISTER_BLOB_FEEDER(IDEEP, IDeepFeeder);
class IDeepFetcher : public BlobFetcherBase {
TypeMeta type_transform(const itensor& atensor) {
switch (atensor.get_data_type()) {
case itensor::data_type::f32:
return TypeMeta::Make<float>();
case itensor::data_type::s32:
return TypeMeta::Make<int>();
case itensor::data_type::s8:
return TypeMeta::Make<int8_t>();
case itensor::data_type::u8:
return TypeMeta::Make<uint8_t>();
default:
// Should we throw exception?
return TypeMeta();
}
}
public:
pybind11::object Fetch(const Blob& blob) override {
try {
return FetchTensor(blob.Get<itensor>(), true).obj;
} catch (ideep::error& e) {
LOG(ERROR) << "IDEEP error: " << e.message;
throw;
}
}
FetchedBlob FetchTensor(const itensor& atensor, bool force_copy) {
#ifdef USE_NUMPY
FetchedBlob result;
CAFFE_ENFORCE(
(atensor.ndims() != 0) &&
(atensor.get_nelems() == 0 || atensor.get_data_handle() != nullptr),
"Trying to fetch uninitialized tensor");
// NOTE: Only support float so far.
const int numpy_type = NPY_FLOAT;
CAFFE_ENFORCE(
numpy_type != -1,
"Unsupported ideep memory data type? This usually should not happen "
"since ideep memory usually only do float and double.");
itensor::dims dims;
bool need_reorder = atensor.need_reorder();
if (atensor.get_data_type() == idtype::f32 && !atensor.has_scale()) {
// For FP32 path, only support NCHW format input, so if atensor
// has NHWC format, we need reorder it to NCHW format.
dims = atensor.get_dims();
need_reorder = need_reorder || atensor.get_desc().is_nhwc();
} else {
dims = atensor.get_public_format_dims();
}
std::vector<npy_intp> npy_dims(dims.begin(), dims.end());
result.copied = force_copy || need_reorder;
// NOLINTNEXTLINE(cppcoreguidelines-init-variables)
void* outPtr;
if (result.copied) {
result.obj = py::reinterpret_steal<py::object>(
PyArray_SimpleNew(atensor.ndims(), npy_dims.data(), numpy_type));
outPtr = static_cast<void*>(
PyArray_DATA(reinterpret_cast<PyArrayObject*>(result.obj.ptr())));
} else {
outPtr = atensor.get_data_handle();
result.obj = py::reinterpret_steal<py::object>(PyArray_SimpleNewFromData(
atensor.ndims(), npy_dims.data(), numpy_type, outPtr));
}
if (numpy_type == NPY_OBJECT) {
CAFFE_THROW("We don't support strings.");
}
if (result.copied) {
if (atensor.get_data_type() == idtype::f32 && !atensor.has_scale()) {
itensor temp_ten(atensor.get_desc().to_default_format(), outPtr);
atensor.reorder_to(temp_ten);
} else {
atensor.to_public(outPtr);
}
}
return result;
#else
CAFFE_THROW("Caffe2 was compiled without NumPy support.");
#endif // USE_NUMPY
}
};
class IDeepFeeder : public BlobFeederBase {
itensor::data_type type_transform(const TypeMeta meta) {
if (meta == TypeMeta::Make<float>())
return itensor::data_type::f32;
else if (meta == TypeMeta::Make<int>())
return itensor::data_type::s32;
else if (meta == TypeMeta::Make<int8_t>())
return itensor::data_type::s8;
else if (meta == TypeMeta::Make<uint8_t>())
return itensor::data_type::u8;
else
return itensor::data_type::undef;
}
public:
void FeedTensor(
const DeviceOption& option,
PyArrayObject* original_array,
itensor* tensor) {
#ifdef USE_NUMPY
PyArrayObject* array = PyArray_GETCONTIGUOUS(original_array);
auto g = MakeGuard([&]() { Py_XDECREF(array); });
const auto npy_type = PyArray_TYPE(array);
const TypeMeta meta = NumpyTypeToCaffe(npy_type);
CAFFE_ENFORCE_NE(
meta,
ScalarType::Undefined,
"This numpy data type is not supported: ",
PyArray_TYPE(array),
".");
int ndim = PyArray_NDIM(array);
npy_intp* npy_dims = PyArray_DIMS(array);
itensor::dims adims;
for (int i = 0; i < ndim; i++) {
adims.push_back(static_cast<itensor::dims::value_type>(npy_dims[i]));
}
switch (npy_type) {
case NPY_OBJECT:
case NPY_UNICODE:
CAFFE_THROW("IDeep doesn't support string");
break;
default:
auto type = type_transform(meta);
if (tensor->get_dims() != adims || type != tensor->get_data_type()) {
tensor->resize(adims, type);
}
tensor->feed_from(adims, type, static_cast<void*>(PyArray_DATA(array)));
}
#else
CAFFE_THROW("Caffe2 was compiled without NumPy support.");
#endif // USE_NUMPY
}
bool ZeroDim(PyArrayObject* array) {
#ifdef USE_NUMPY
int ndim = PyArray_NDIM(array);
return ndim == 0;
#else
CAFFE_THROW("Caffe2 was compiled without NumPy support.");
#endif
}
void Feed(
const DeviceOption& option,
PyArrayObject* original_array,
Blob* blob,
bool in_place) override {
#ifdef USE_NUMPY
try {
PyArrayObject* array = PyArray_GETCONTIGUOUS(original_array);
auto g = MakeGuard([&]() { Py_XDECREF(array); });
const auto npy_type = PyArray_TYPE(array);
const TypeMeta meta = NumpyTypeToCaffe(npy_type);
// TODO: if necessary, use dispatcher.
if ((in_place && blob->IsType<itensor>()) ||
(meta.Match<float>() && !ZeroDim(original_array))) {
FeedTensor(option, original_array, blob->GetMutable<itensor>());
} else {
DeviceOption cpu_option(option);
cpu_option.set_device_type(DeviceTypeProto::PROTO_CPU);
TensorFeeder<CPUContext> cpu_tensor_feeder;
if (in_place) {
cpu_tensor_feeder.FeedTensor(
cpu_option,
original_array,
BlobGetMutableTensor(blob, OptionToDevice(cpu_option).type()),
true);
} else {
blob->Reset<Tensor>(new Tensor(
cpu_tensor_feeder.FeedTensor(cpu_option, original_array)));
}
}
} catch (ideep::error& e) {
LOG(ERROR) << "IDEEP error: " << e.message;
throw;
}
#else
CAFFE_THROW("Caffe2 was compiled without NumPy support.");
#endif
}
};
} // namespace python
} // namespace caffe2
|
