forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 1
/
data_filler.cc
258 lines (235 loc) · 8.68 KB
/
data_filler.cc
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
#include "caffe2/predictor/emulator/data_filler.h"
#include "caffe2/predictor/emulator/utils.h"
namespace caffe2 {
namespace emulator {
void DataNetFiller::fill_parameter(Workspace* ws) const {
// As we use initial parameter initialization for this BenchmarkState,
// we can just run the init_net
CAFFE_ENFORCE(
ws->RunNetOnce(init_net_),
"Failed running the init_net: ",
ProtoDebugString(init_net_));
}
void DataNetFiller::fill_input_internal(TensorList_t* input_data) const {
Workspace ws;
CAFFE_ENFORCE(ws.RunNetOnce(data_net_));
for (const auto& name : input_names_) {
input_data->emplace_back(
BlobGetMutableTensor(ws.GetBlob(name), CPU)->Clone());
}
}
void fill_with_type(
const TensorFiller& filler,
const std::string& type,
TensorCPU* output) {
CPUContext context;
if (type == "float") {
filler.Fill<float>(output, &context);
} else if (type == "double") {
filler.Fill<double>(output, &context);
} else if (type == "uint8_t" || type == "unsigned char") {
filler.Fill<uint8_t>(output, &context);
} else if (type == "uint16_t") {
filler.Fill<uint16_t>(output, &context);
} else if (type == "int8_t") {
filler.Fill<int8_t>(output, &context);
} else if (type == "int16_t") {
filler.Fill<int16_t>(output, &context);
} else if (type == "int32_t" || type == "int") {
filler.Fill<int32_t>(output, &context);
} else if (type == "int64_t" || type == "long") {
filler.Fill<int64_t>(output, &context);
} else if (type == "bool") {
auto mutable_filler = filler;
mutable_filler.Min(0).Max(2).Fill<uint8_t>(output, &context);
} else {
throw std::invalid_argument("filler does not support type " + type);
}
}
DataRandomFiller::DataRandomFiller(
const NetDef& run_net,
const std::vector<std::vector<std::vector<int64_t>>>& input_dims,
const std::vector<std::vector<std::string>>& input_types) {
// parse dimensions
CAFFE_ENFORCE_EQ(input_dims.size(), run_net.op_size());
CAFFE_ENFORCE_EQ(input_types.size(), run_net.op_size());
// load op inputs and outputs
std::unordered_set<std::string> output_names;
for (size_t i = 0; i < run_net.op_size(); ++i) {
const auto& op = run_net.op(i);
const auto& op_dims = input_dims[i];
const auto& op_types = input_types[i];
CAFFE_ENFORCE(
op_dims.size() == op.input_size(),
op.name() + " has " + c10::to_string(op.input_size()) +
" inputs; while the input dimension size is " +
c10::to_string(op_dims.size()));
CAFFE_ENFORCE(
op_types.size() == op.input_size(),
op.name() + " has " + c10::to_string(op.input_size()) +
" inputs; while the input type size is " +
c10::to_string(op_types.size()));
for (size_t j = 0; j < op.input_size(); ++j) {
inputs_[op.input(j)] =
std::make_pair(get_tensor_filler(op, j, op_dims), op_types[j]);
}
// Hack, we normal have a path of
// length -> LengthsiRangeFill -> Gather -> w -> SparseLengthsWeighted*
// \---------------------------------------/
// So when we generate the value of length, we need to bound it to the size
// of weight input of Gather too
if (op.type().find("SparseLengthsWeighted") == 0 && i > 0) {
const auto& prev_op = run_net.op(i - 1);
if (prev_op.type() == "Gather") {
const auto& prev_dims = input_dims[i - 1];
VLOG(1) << "Setting max length value to " << prev_dims[0].front()
<< " for " << op.input(3);
inputs_[op.input(3)].first.Max(prev_dims[0].front());
}
}
for (size_t j = 0; j < op.output_size(); ++j) {
output_names.emplace(op.output(j));
}
}
// load parameters
std::unordered_set<std::string> parameters;
for (size_t i = 0; i < run_net.arg_size(); ++i) {
const auto& arg = run_net.arg(i);
// TODO: replace "PredictorParameters" with the constant in OSS bbp
if (arg.has_name() && arg.name() == "PredictorParameters") {
parameters.reserve(arg.strings_size());
for (size_t j = 0; j < arg.strings_size(); ++j) {
parameters.emplace(arg.strings(j));
}
break;
}
}
if (parameters.size() == 0) {
VLOG(1) << "Fail to find any parameters";
}
for (const auto& param : parameters) {
// remove unused parameters
if (inputs_.find(param) != inputs_.end()) {
// inputs_[param] will be erase from inputs_ in the next step
parameters_.emplace(param, inputs_[param]);
}
}
for (const auto& param : parameters_) {
inputs_.erase(param.first);
}
for (const auto& name : output_names) {
inputs_.erase(name);
}
CAFFE_ENFORCE(inputs_.size() > 0, "Empty input for run net");
// generate input names
for (const auto& input : inputs_) {
input_names_.push_back(input.first);
}
}
void DataRandomFiller::fill_parameter(Workspace* ws) const {
for (auto& param : parameters_) {
Blob* blob = ws->CreateBlob(param.first);
fill_with_type(
param.second.first,
param.second.second,
BlobGetMutableTensor(blob, CPU));
CAFFE_ENFORCE(ws->GetBlob(param.first)->GetRaw());
}
}
void DataRandomFiller::fill_input_internal(TensorList_t* input_data) const {
for (auto& name : input_names_) {
input_data->emplace_back(CPU);
const auto& it = inputs_.find(name);
CAFFE_ENFORCE(it != inputs_.end());
fill_with_type(it->second.first, it->second.second, &input_data->back());
}
}
TestDataRandomFiller::TestDataRandomFiller(
const NetDef& net,
const std::vector<std::vector<std::vector<int64_t>>>& inputDims,
const std::vector<std::vector<std::string>>& inputTypes)
: DataRandomFiller() {
std::unordered_set<std::string> outputNames;
// Determine blobs that are outputs of some ops (intermediate blobs).
for (auto opIdx = 0; opIdx < net.op_size(); ++opIdx) {
const auto& op = net.op(opIdx);
for (auto outputIdx = 0; outputIdx < op.output_size(); ++outputIdx) {
outputNames.emplace(op.output(outputIdx));
}
}
// Determine ops that have non-intermediate inputs.
std::unordered_set<size_t> opWithRequiredInputs;
for (auto opIdx = 0; opIdx < net.op_size(); ++opIdx) {
const auto& op = net.op(opIdx);
for (auto inputIdx = 0; inputIdx < op.input_size(); ++inputIdx) {
if (!outputNames.count(op.input(inputIdx))) {
opWithRequiredInputs.emplace(opIdx);
break;
}
}
}
CAFFE_ENFORCE_EQ(inputDims.size(), opWithRequiredInputs.size());
CAFFE_ENFORCE_EQ(inputTypes.size(), opWithRequiredInputs.size());
int counter = 0;
for (auto opIdx = 0; opIdx < net.op_size(); ++opIdx) {
if (!opWithRequiredInputs.count(opIdx)) {
// Skip intermediate ops.
continue;
}
const auto& op = net.op(opIdx);
const auto& op_dims = inputDims[counter];
const auto& op_types = inputTypes[counter];
++counter;
int countRequiredInputs = 0;
for (auto inputIdx = 0; inputIdx < op.input_size(); ++inputIdx) {
if (!outputNames.count(op.input(inputIdx))) {
++countRequiredInputs;
}
}
CAFFE_ENFORCE(
op_dims.size() == countRequiredInputs,
op.name() + " has " + c10::to_string(op.input_size()) +
" (required) inputs; while the input dimension size is " +
c10::to_string(op_dims.size()));
CAFFE_ENFORCE(
op_types.size() == countRequiredInputs,
op.name() + " has " + c10::to_string(op.input_size()) +
" (required) inputs; while the input type size is " +
c10::to_string(op_types.size()));
int dimCounter = 0;
for (auto inputIdx = 0; inputIdx < op.input_size(); ++inputIdx) {
auto inputName = op.input(inputIdx);
if (outputNames.count(inputName)) {
// Skip intermediate inputs.
continue;
}
inputs_[inputName] = std::make_pair(
get_tensor_filler(op, dimCounter, op_dims), op_types[dimCounter]);
++dimCounter;
}
}
CAFFE_ENFORCE(inputs_.size() > 0, "Empty input for run net");
// generate input names
for (const auto& input : inputs_) {
input_names_.push_back(input.first);
}
}
void TestDataRandomFiller::fillInputToWorkspace(Workspace* workspace) const {
for (auto& name : input_names_) {
const auto& it = inputs_.find(name);
CAFFE_ENFORCE(it != inputs_.end());
auto* tensor =
BlobGetMutableTensor(workspace->CreateBlob(name), caffe2::CPU);
fill_with_type(it->second.first, it->second.second, tensor);
}
}
void fillRandomNetworkInputs(
const NetDef& net,
const std::vector<std::vector<std::vector<int64_t>>>& inputDims,
const std::vector<std::vector<std::string>>& inputTypes,
Workspace* workspace) {
TestDataRandomFiller(net, inputDims, inputTypes)
.fillInputToWorkspace(workspace);
}
} // namespace emulator
} // namespace caffe2