forked from apache/tvm
-
Notifications
You must be signed in to change notification settings - Fork 3
/
gradient.cc
266 lines (220 loc) · 8.07 KB
/
gradient.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
259
260
261
262
263
264
265
266
/*!
* Copyright (c) 2018 by Contributors
* \file gradient.cc
* \brief Automatic gradients for any operation
*/
#include <topi/nn.h>
#include <topi/broadcast.h>
#include <nnvm/op.h>
#include <nnvm/node.h>
#include <nnvm/op_attr_types.h>
#include <nnvm/compiler/op_attr_types.h>
#include <nnvm/top/tensor.h>
#include "../op_common.h"
#include <tvm/operation.h>
#include <tvm/ir_pass.h>
#include <tvm/ir.h>
#include <tvm/api_registry.h>
namespace nnvm {
namespace top {
/*!
* \brief A map from string to string but with an operator>> defined.
*/
class StrDict {
public:
std::unordered_map<std::string, std::string> dict;
/*!
* \brief Save StrDict to JSON.
* \param writer JSONWriter
*/
inline void Save(dmlc::JSONWriter* writer) const {
writer->Write(dict);
}
/*!
* \brief Load StrDict from JSON.
* \param reader JSONReader
*/
inline void Load(dmlc::JSONReader* reader) {
reader->Read(&dict);
}
/*!
* \brief allow output string of tuple to ostream
* \param os the output stream
* \param t the tuple
* \return the ostream
*/
friend std::ostream &operator<<(std::ostream &os, const StrDict &sd) {
dmlc::JSONWriter(&os).Write(sd);
return os;
}
/*!
* \brief read tuple from the istream
* \param is the input stream
* \param t The tuple
* \return the istream
*/
friend std::istream &operator>>(std::istream &is, StrDict &sd) {
dmlc::JSONReader(&is).Read(&sd);
return is;
}
};
using namespace nnvm::compiler;
struct GradientParam : public dmlc::Parameter<GradientParam> {
std::string original_op;
StrDict original_attrs;
int input_index;
DMLC_DECLARE_PARAMETER(GradientParam) {
DMLC_DECLARE_FIELD(original_op);
DMLC_DECLARE_FIELD(original_attrs);
DMLC_DECLARE_FIELD(input_index);
}
NodeAttrs original() const {
NodeAttrs res;
res.op = Op::Get(original_op);
for (const auto& key_val_pair : original_attrs.dict)
res.dict[key_val_pair.first] = key_val_pair.second;
if (res.op->attr_parser) {
res.op->attr_parser(&res);
}
return res;
}
};
DMLC_REGISTER_PARAMETER(GradientParam);
inline bool GradientShape(const nnvm::NodeAttrs& attrs,
std::vector<TShape> *in_attrs,
std::vector<TShape> *out_attrs) {
auto parsed = nnvm::get<GradientParam>(attrs.parsed);
NodeAttrs o_attrs = parsed.original();
uint32_t o_num_inputs = o_attrs.op->num_inputs;
if (o_attrs.op->get_num_inputs)
o_num_inputs = o_attrs.op->get_num_inputs(o_attrs);
uint32_t o_num_outputs = o_attrs.op->num_outputs;
if (o_attrs.op->get_num_outputs)
o_num_outputs = o_attrs.op->get_num_outputs(o_attrs);
CHECK_EQ(in_attrs->size(), o_num_inputs + o_num_outputs);
CHECK_EQ(out_attrs->size(), 1);
NNVM_ASSIGN_OUTPUT_SHAPE(attrs, *out_attrs, 0, (*in_attrs)[parsed.input_index]);
NNVM_ASSIGN_INPUT_SHAPE(attrs, *in_attrs, parsed.input_index, (*out_attrs)[0]);
return true;
}
inline bool GradientType(const nnvm::NodeAttrs& attrs,
std::vector<int> *in_attrs,
std::vector<int> *out_attrs) {
auto parsed = nnvm::get<GradientParam>(attrs.parsed);
NodeAttrs o_attrs = parsed.original();
uint32_t o_num_inputs = o_attrs.op->num_inputs;
if (o_attrs.op->get_num_inputs)
o_num_inputs = o_attrs.op->get_num_inputs(o_attrs);
uint32_t o_num_outputs = o_attrs.op->num_outputs;
if (o_attrs.op->get_num_outputs)
o_num_outputs = o_attrs.op->get_num_outputs(o_attrs);
CHECK_EQ(in_attrs->size(), o_num_inputs + o_num_outputs);
CHECK_EQ(out_attrs->size(), 1);
NNVM_ASSIGN_OUTPUT_TYPE(attrs, *out_attrs, 0, (*in_attrs)[parsed.input_index]);
NNVM_ASSIGN_INPUT_TYPE(attrs, *in_attrs, parsed.input_index, (*out_attrs)[0]);
return true;
}
Array<Tensor> GradientCompute(const NodeAttrs& attrs,
const Array<Tensor>& inputs,
const Array<Tensor>& out_info) {
static auto& ftvmcompute = nnvm::Op::GetAttr<FTVMCompute>("FTVMCompute");
auto parsed = nnvm::get<GradientParam>(attrs.parsed);
NodeAttrs o_attrs = parsed.original();
uint32_t o_num_inputs = o_attrs.op->num_inputs;
if (o_attrs.op->get_num_inputs)
o_num_inputs = o_attrs.op->get_num_inputs(o_attrs);
Array<Tensor> o_inputs(inputs.begin(), inputs.begin() + o_num_inputs);
Array<Tensor> head_grads(inputs.begin() + o_num_inputs, inputs.end());
Array<Tensor> input_placeholders;
std::unordered_map<Tensor, Tensor> placeholders_to_inputs;
for (const Tensor& input : o_inputs) {
Tensor place =
tvm::PlaceholderOpNode::make(input->op->name, input->shape, input->dtype).output(0);
input_placeholders.push_back(place);
placeholders_to_inputs[place] = input;
}
Array<Tensor> forward = ftvmcompute[o_attrs.op](o_attrs, input_placeholders, head_grads);
Array<Tensor> results;
const Tensor& place = input_placeholders[parsed.input_index];
Tensor res;
auto head_grads_iter = head_grads.begin();
for (const Tensor& out : forward) {
Tensor part = tvm::ir::JacobianRecursive(out, {place}, *head_grads_iter)[0];
part = tvm::TensorNode::make(part->shape, part->dtype,
part->op->ReplaceInputs(part->op, placeholders_to_inputs), part->value_index);
if (res.operator->()) {
res = topi::add(res, part);
}
else
res = part;
++head_grads_iter;
}
return {res};
}
NNVM_REGISTER_OP(gradient)
.describe(R"doc(Gradients for any specified operation.
)doc" NNVM_ADD_FILELINE)
//.set_support_level(1)
.set_num_inputs(([](const NodeAttrs& attrs) {
NodeAttrs o_attrs = nnvm::get<GradientParam>(attrs.parsed).original();
uint32_t o_num_inputs = o_attrs.op->num_inputs;
if (o_attrs.op->get_num_inputs)
o_num_inputs = o_attrs.op->get_num_inputs(o_attrs);
uint32_t o_num_outputs = o_attrs.op->num_outputs;
if (o_attrs.op->get_num_outputs)
o_num_outputs = o_attrs.op->get_num_outputs(o_attrs);
return o_num_inputs + o_num_outputs;
}))
.set_num_outputs(1)
.set_attr_parser(ParamParser<GradientParam>)
.set_attr<FGetAttrDict>("FGetAttrDict", ParamGetAttrDict<GradientParam>)
.add_arguments(GradientParam::__FIELDS__())
.set_attr<FInferShape>("FInferShape", GradientShape)
.set_attr<FInferType>("FInferType", GradientType)
//.set_attr<FCorrectLayout>("FCorrectLayout", DotCorrectLayout)
.set_attr<FTVMCompute>("FTVMCompute", GradientCompute)
.set_attr<FTVMSchedule>("FTVMSchedule",
[](const NodeAttrs& attrs, const Array<Tensor>& outs, const std::string& target) {
Array<tvm::Operation> out_ops;
for (auto t : outs)
out_ops.push_back(t->op);
return create_schedule(out_ops);
});
std::vector<NodeEntry> AutomaticFGradient(const NodePtr& n, const std::vector<NodeEntry>& ograds) {
std::vector<NodeEntry> grad_inputs(n->inputs);
grad_inputs.insert(grad_inputs.end(), ograds.begin(), ograds.end());
std::vector<NodeEntry> result;
for (uint32_t i = 0; i < n->num_inputs(); ++i) {
std::unordered_map<std::string, std::string> grad_attrs;
std::ostringstream orig_attrs;
orig_attrs << StrDict{n->attrs.dict};
grad_attrs["original_op"] = n->op()->name;
grad_attrs["original_attrs"] = orig_attrs.str();
grad_attrs["input_index"] = std::to_string(i);
result.push_back(nnvm::MakeNode("gradient", n->attrs.name + "_grad", grad_inputs, grad_attrs));
}
return result;
}
void MakeDifferentiable(const std::string& op_name, int plevel = 100) {
Op& op = dmlc::Registry<Op>::Get()->__REGISTER_OR_GET__(op_name);
if (op.num_inputs == kVarg)
std::cerr << op_name <<
" accepts variable number of arguments, automatic gen of gradients is not supported\n";
else
op.set_attr<FGradient>("FGradient", AutomaticFGradient, plevel);
}
void MakeDifferentiableAll(int plevel = 100) {
for (auto op : dmlc::Registry<Op>::List()) {
MakeDifferentiable(op->name, plevel);
}
}
TVM_REGISTER_API("MakeDifferentiable")
.set_body([](tvm::TVMArgs args, tvm::TVMRetValue *ret) {
MakeDifferentiable(args[0]);
});
TVM_REGISTER_API("MakeDifferentiableAll")
.set_body([](tvm::TVMArgs args, tvm::TVMRetValue *ret) {
MakeDifferentiableAll();
});
} // namespace top
} // namespace nnvm