Spaces:
Build error
Build error
File size: 8,892 Bytes
28451f7 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 |
/*
* SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* SPDX-License-Identifier: Apache-2.0
*
* 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.
*/
/** @file adam_optimizer.h
* @author Thomas Müller, NVIDIA
*/
#pragma once
#include <neural-graphics-primitives/common.h>
#include <neural-graphics-primitives/common_device.cuh>
#include <neural-graphics-primitives/json_binding.h>
#include <json/json.hpp>
namespace ngp {
class VarAdamOptimizer {
public:
VarAdamOptimizer(size_t size = 0, float learning_rate = 1e-3, float epsilon = 1e-08f, float beta1 = 0.9f, float beta2 = 0.99f) : m_state{size} {
m_hparams = { learning_rate, epsilon, beta1, beta2 };
}
VarAdamOptimizer& operator=(const VarAdamOptimizer& arg) {
m_state = arg.m_state;
m_hparams = arg.m_hparams;
return *this;
}
VarAdamOptimizer(const VarAdamOptimizer& arg) {
*this = arg;
}
void step(const std::vector<float>& gradient) {
++m_state.iter;
float actual_learning_rate = m_hparams.learning_rate * std::sqrt(1.0f - std::pow(m_hparams.beta2, (float)m_state.iter)) / (1.0f - std::pow(m_hparams.beta1, (float)m_state.iter));
for (size_t i = 0; i < m_state.first_moment.size(); ++i) {
m_state.first_moment[i] = m_hparams.beta1 * m_state.first_moment[i] + (1.0f - m_hparams.beta1) * gradient[i];
m_state.second_moment[i] = m_hparams.beta2 * m_state.second_moment[i] + (1.0f - m_hparams.beta2) * gradient[i] * gradient[i];
m_state.variable[i] -= actual_learning_rate * m_state.first_moment[i] / (std::sqrt(m_state.second_moment[i]) + m_hparams.epsilon);
}
}
uint32_t step() const {
return m_state.iter;
}
void set_learning_rate(float lr) {
m_hparams.learning_rate = lr;
}
std::vector<float>& variable() {
return m_state.variable;
}
const std::vector<float>& variable() const {
return m_state.variable;
}
void reset_state() {
m_state = State{m_state.first_moment.size()};
}
void to_json(nlohmann::json& j) const {
j["iter"] = m_state.iter;
j["first_moment"] = m_state.first_moment;
j["second_moment"] = m_state.second_moment;
j["variable"] = m_state.variable;
j["learning_rate"] = m_hparams.learning_rate;
j["epsilon"] = m_hparams.epsilon;
j["beta1"] = m_hparams.beta1;
j["beta2"] = m_hparams.beta2;
}
void from_json(const nlohmann::json& j) {
m_state.iter = j.at("iter");
m_state.first_moment = j.at("first_moment").get<std::vector<float>>();
m_state.second_moment = j.at("second_moment").get<std::vector<float>>();
m_state.variable = j.at("variable").get<std::vector<float>>();
m_hparams.learning_rate = j.at("learning_rate");
m_hparams.epsilon = j.at("epsilon");
m_hparams.beta1 = j.at("beta1");
m_hparams.beta2 = j.at("beta2");
}
private:
struct State {
State() = default;
State(const State&) = default;
State(size_t size) {
iter = 0;
first_moment = std::vector<float>(size, 0.0f);
second_moment = std::vector<float>(size, 0.0f);
variable = std::vector<float>(size, 0.0f);
}
uint32_t iter;
std::vector<float> first_moment;
std::vector<float> second_moment;
std::vector<float> variable;
} m_state;
struct Hyperparameters {
float learning_rate;
float epsilon;
float beta1;
float beta2;
} m_hparams;
};
inline void to_json(nlohmann::json& j, const VarAdamOptimizer& opt) {
opt.to_json(j);
}
inline void from_json(const nlohmann::json& j, VarAdamOptimizer& opt) {
opt.from_json(j);
}
template <typename T>
class AdamOptimizer {
public:
AdamOptimizer(float learning_rate = 1e-3, float epsilon = 1e-08f, float beta1 = 0.9f, float beta2 = 0.99f) {
m_hparams = { learning_rate, epsilon, beta1, beta2 };
}
AdamOptimizer& operator=(const AdamOptimizer& arg) {
m_state = arg.m_state;
m_hparams = arg.m_hparams;
return *this;
}
AdamOptimizer(const AdamOptimizer& arg) {
*this = arg;
}
void step(const T& gradient) {
++m_state.iter;
float actual_learning_rate = m_hparams.learning_rate * std::sqrt(1.0f - std::pow(m_hparams.beta2, (float)m_state.iter)) / (1.0f - std::pow(m_hparams.beta1, (float)m_state.iter));
m_state.first_moment = m_hparams.beta1 * m_state.first_moment + (1.0f - m_hparams.beta1) * gradient;
m_state.second_moment = m_hparams.beta2 * m_state.second_moment + (1.0f - m_hparams.beta2) * gradient * gradient;
m_state.variable -= actual_learning_rate * m_state.first_moment / (sqrt(m_state.second_moment) + T(m_hparams.epsilon));
}
uint32_t step() const {
return m_state.iter;
}
void set_learning_rate(float lr) {
m_hparams.learning_rate = lr;
}
T& variable() {
return m_state.variable;
}
const T& variable() const {
return m_state.variable;
}
void reset_state() {
m_state = {};
}
void to_json(nlohmann::json& j) const {
j["iter"] = m_state.iter;
j["first_moment"] = m_state.first_moment;
j["second_moment"] = m_state.second_moment;
j["variable"] = m_state.variable;
j["learning_rate"] = m_hparams.learning_rate;
j["epsilon"] = m_hparams.epsilon;
j["beta1"] = m_hparams.beta1;
j["beta2"] = m_hparams.beta2;
}
void from_json(const nlohmann::json& j) {
m_state.iter = j.at("iter");
m_state.first_moment = j.at("first_moment");
m_state.second_moment = j.at("second_moment");
m_state.variable = j.at("variable");
m_hparams.learning_rate = j.at("learning_rate");
m_hparams.epsilon = j.at("epsilon");
m_hparams.beta1 = j.at("beta1");
m_hparams.beta2 = j.at("beta2");
}
private:
struct State {
uint32_t iter = 0;
T first_moment = T(0.0f);
T second_moment = T(0.0f);
T variable = T(0.0f);
} m_state = {};
struct Hyperparameters {
float learning_rate;
float epsilon;
float beta1;
float beta2;
} m_hparams = {};
};
template <typename T>
inline void to_json(nlohmann::json& j, const AdamOptimizer<T>& opt) {
opt.to_json(j);
}
template <typename T>
inline void from_json(const nlohmann::json& j, AdamOptimizer<T>& opt) {
opt.from_json(j);
}
class RotationAdamOptimizer {
public:
RotationAdamOptimizer(float learning_rate = 1e-3, float epsilon = 1e-08f, float beta1 = 0.9f, float beta2 = 0.99f) {
m_hparams = { learning_rate, epsilon, beta1, beta2 };
}
RotationAdamOptimizer& operator=(const RotationAdamOptimizer& arg) {
m_state = arg.m_state;
m_hparams = arg.m_hparams;
return *this;
}
RotationAdamOptimizer(const RotationAdamOptimizer& arg) {
*this = arg;
}
void step(const vec3& gradient) {
++m_state.iter;
float actual_learning_rate = m_hparams.learning_rate * std::sqrt(1 - std::pow(m_hparams.beta2, m_state.iter)) / (1 - std::pow(m_hparams.beta1, m_state.iter));
m_state.first_moment = m_hparams.beta1 * m_state.first_moment + (1 - m_hparams.beta1) * gradient;
m_state.second_moment = m_hparams.beta2 * m_state.second_moment + (1 - m_hparams.beta2) * gradient * gradient;
vec3 rot = actual_learning_rate * m_state.first_moment / (sqrt(m_state.second_moment) + m_hparams.epsilon);
m_state.variable = rotvec(rotmat(-rot) * rotmat(variable()));
}
uint32_t step() const {
return m_state.iter;
}
void set_learning_rate(float lr) {
m_hparams.learning_rate = lr;
}
const vec3& variable() const {
return m_state.variable;
}
void reset_state() {
m_state = {};
}
void to_json(nlohmann::json& j) const {
j["iter"] = m_state.iter;
j["first_moment"] = m_state.first_moment;
j["second_moment"] = m_state.second_moment;
j["variable"] = m_state.variable;
j["learning_rate"] = m_hparams.learning_rate;
j["epsilon"] = m_hparams.epsilon;
j["beta1"] = m_hparams.beta1;
j["beta2"] = m_hparams.beta2;
}
void from_json(const nlohmann::json& j) {
m_state.iter = j.at("iter");
m_state.first_moment = j.at("first_moment");
m_state.second_moment = j.at("second_moment");
m_state.variable = j.at("variable");
m_hparams.learning_rate = j.at("learning_rate");
m_hparams.epsilon = j.at("epsilon");
m_hparams.beta1 = j.at("beta1");
m_hparams.beta2 = j.at("beta2");
}
private:
struct State {
uint32_t iter = 0;
vec3 first_moment = vec3(0.0f);
vec3 second_moment = vec3(0.0f);
vec3 variable = vec3(0.0f);
} m_state;
struct Hyperparameters {
float learning_rate;
float epsilon;
float beta1;
float beta2;
} m_hparams;
};
inline void to_json(nlohmann::json& j, const RotationAdamOptimizer& opt) {
opt.to_json(j);
}
inline void from_json(const nlohmann::json& j, RotationAdamOptimizer& opt) {
opt.from_json(j);
}
}
|