Repository navigation
Expand file tree
/
Copy pathdual_net.cc
More file actions
190 lines (166 loc) · 6.7 KB
/
Copy pathdual_net.cc
File metadata and controls
190 lines (166 loc) · 6.7 KB
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
// Copyright 2018 Google LLC
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "cc/dual_net.h"
#include "cc/check.h"
#include "cc/constants.h"
#include "cc/symmetries.h"
#include "tensorflow/core/framework/graph.pb.h"
#include "tensorflow/core/lib/core/status.h"
using tensorflow::DT_FLOAT;
using tensorflow::Env;
using tensorflow::GraphDef;
using tensorflow::NewSession;
using tensorflow::ReadBinaryProto;
using tensorflow::Session;
using tensorflow::SessionOptions;
using tensorflow::Status;
using tensorflow::string;
using tensorflow::Tensor;
using tensorflow::TensorShape;
using tensorflow::error::CANCELLED;
namespace minigo {
constexpr int DualNet::kNumStoneFeatures;
constexpr int DualNet::kNumBoardFeatures;
void DualNet::InitializeFeatures(const Position& position,
BoardFeatures* features) {
const auto my_color = position.to_play();
const auto their_color = OtherColor(my_color);
const float to_play = my_color == Color::kBlack ? 1 : 0;
for (int i = 0; i < kN * kN; ++i) {
int j = i * kNumStoneFeatures;
auto stone_color = position.stones()[i].color();
auto my_stone = stone_color == my_color ? 1 : 0;
auto their_stone = stone_color == their_color ? 1 : 0;
for (int plane = 0; plane < kPlayerFeature; plane += 2) {
(*features)[j++] = my_stone;
(*features)[j++] = their_stone;
}
(*features)[j++] = to_play;
}
}
// The update loop here is a little tricky.
//
// The chart below shows, for each move, how the stones from the last 8 moves
// should be distributed through the input planes.
//
// planes
// move | to play | 0 1 2 3 4 5 ... 16
// ------+---------+-----------------------------------------
// 1 | B | B_1 W_1 - - - - ... 1
// 2 | W | W_2 B_2 W_1 B_1 - - ... 0
// 3 | B | B_3 W_3 B_2 W_2 B_1 W_1 ... 1
// 4 | W | W_4 B_4 W_3 B_3 W_2 B_2 ... 0
// ... | ... | ... ... ... ... ... ... ... ...
//
// For example: on move 3, planes 0 & 1 hold the black & white stones that are
// on the board before move 3 is played, planes 2 & 3 hold the stones that were
// on the board before move 2 was played, planes 4 & 5 hold the stones that
// were on the board before move 1 was played, etc.
//
// So... to update the features, we need to do four things:
// 1) Shuffle the planes for moves t .. t-6 over to the planes for moves
// t-1 .. t-7.
// 2) Swap the black and white planes for moves t-1 .. t-7.
// 3) Write the new black and white stones into planes 0 & 1 (or planes 1 & 0
// depending on who is to play first).
// 4) Write the "to play" feature into plane 16.
//
// Steps 3 and 4 are trivial.
//
// Steps 1 and 2 can be accomplished in one by the following:
// 1) Copy even planes from plane N to plane N + 3.
// 2) Copy odd planes from plane N to plane N + 1.
//
// The code below does this slightly differently, updated the planes in the
// reverse order because that allows old_features and new_features to point to
// the same array, but the end result is the same.
void DualNet::UpdateFeatures(const BoardFeatures& old_features,
const Position& position,
BoardFeatures* new_features) {
const auto my_color = position.to_play();
const auto their_color = OtherColor(my_color);
const float to_play = my_color == Color::kBlack ? 1 : 0;
for (int i = 0; i < kN * kN; ++i) {
auto stone_color = position.stones()[i].color();
const auto* src = old_features.data() + i * kNumStoneFeatures;
auto* dst = new_features->data() + i * kNumStoneFeatures;
dst[kPlayerFeature] = to_play;
for (int j = kPlayerFeature - 2; j > 0; j -= 2) {
dst[j + 1] = src[j - 2];
dst[j] = src[j - 1];
}
dst[1] = stone_color == their_color ? 1 : 0;
dst[0] = stone_color == my_color ? 1 : 0;
}
}
DualNet::DualNet() = default;
DualNet::~DualNet() {
if (session_ != nullptr) {
session_->Close();
}
}
void DualNet::Initialize(const std::string& graph_path) {
GraphDef graph_def;
TF_CHECK_OK(ReadBinaryProto(Env::Default(), graph_path, &graph_def));
session_.reset(NewSession(SessionOptions()));
TF_CHECK_OK(session_->Create(graph_def));
inputs_.clear();
inputs_.emplace_back(
"pos_tensor",
Tensor(DT_FLOAT, TensorShape({1, kN, kN, kNumStoneFeatures})));
output_names_.clear();
output_names_.push_back("policy_output");
output_names_.push_back("value_output");
}
void DualNet::RunMany(absl::Span<const BoardFeatures* const> features,
absl::Span<Output> outputs, Random* rnd) {
MG_DCHECK(features.size() == outputs.size());
int batch_size = static_cast<int>(features.size());
auto& feature_tensor = inputs_[0].second;
if (feature_tensor.dim_size(0) != batch_size) {
feature_tensor =
Tensor(DT_FLOAT, TensorShape({batch_size, kN, kN, kNumStoneFeatures}));
}
// Select symmetry operations to apply.
symmetries_used_.clear();
if (rnd != nullptr) {
symmetries_used_.reserve(batch_s
5053
ize);
for (int i = 0; i < batch_size; ++i) {
symmetries_used_.push_back(static_cast<symmetry::Symmetry>(
rnd->UniformInt(0, symmetry::kNumSymmetries - 1)));
}
} else {
symmetries_used_.resize(batch_size, symmetry::kIdentity);
}
// Copy the features into the input tensor.
for (int i = 0; i < batch_size; ++i) {
symmetry::ApplySymmetry<float, kN, kNumStoneFeatures>(
symmetries_used_[i], features[i]->data(),
feature_tensor.flat<float>().data() + i * kNumBoardFeatures);
}
// Run the model.
TF_CHECK_OK(session_->Run(inputs_, output_names_, {}, &outputs_));
// Copy the policy and value out of the output tensors.
const auto& policy_tensor = outputs_[0].flat<float>();
const auto& value_tensor = outputs_[1].flat<float>();
for (int i = 0; i < batch_size; ++i) {
const auto* policy_tensor_data = policy_tensor.data() + i * kNumMoves;
symmetry::ApplySymmetry<float, kN, 1>(
symmetry::Inverse(symmetries_used_[i]), policy_tensor_data,
outputs[i].policy.data());
outputs[i].policy[Coord::kPass] = policy_tensor_data[Coord::kPass];
outputs[i].value = value_tensor.data()[i];
}
}
} // namespace minigo