-
Notifications
You must be signed in to change notification settings - Fork 71
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Merge pull request #46 from dietmarwo/ADD_CR_FM_NES_JAX
Add a Python/JAX port of CR-FM-NES
- Loading branch information
Showing
14 changed files
with
361 additions
and
8 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,227 @@ | ||
| # Copyright 2022 The EvoJAX Authors. | ||
| # | ||
| # 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. | ||
|
|
||
| """JAX port of Fast Moving Natural Evolution Strategy | ||
| for High-Dimensional Problems (CR-FM-NES), see https://arxiv.org/abs/2201.11422 . | ||
| Derived from https://github.com/nomuramasahir0/crfmnes""" | ||
|
|
||
| import math | ||
| import numpy as np | ||
| from typing import Union | ||
| import logging | ||
| import jax | ||
| import jax.numpy as jnp | ||
|
|
||
| from evojax.algo.base import NEAlgorithm | ||
| from evojax.util import create_logger | ||
|
|
||
| class CRFMNES(NEAlgorithm): | ||
| """A wrapper of CR-FM-NES jax port.""" | ||
|
|
||
| def __init__(self, | ||
| param_size: int, | ||
| pop_size: int, | ||
| init_stdev: float = 0.1, | ||
| seed: int = 0, | ||
| logger: logging.Logger = None): | ||
| if logger is None: | ||
| self.logger = create_logger(name='FCRFM') | ||
| else: | ||
| self.logger = logger | ||
| self.pop_size = pop_size | ||
|
|
||
| self.crfm = CRFM(param_size, pop_size, init_stdev, jax.random.PRNGKey(seed)) | ||
|
|
||
| self.params = None | ||
| self._best_params = None | ||
|
|
||
| self.jnp_array = jax.jit(jnp.array) | ||
| self.jnp_stack = jax.jit(jnp.stack) | ||
|
|
||
| def ask(self) -> jnp.ndarray: | ||
| self.params = self.crfm.ask() | ||
| return self.jnp_stack(self.params) | ||
|
|
||
| def tell(self, fitness: jnp.ndarray) -> None: | ||
| self.crfm.tell(-np.array(fitness)) | ||
| self._best_params = self.crfm.x_best | ||
|
|
||
| @property | ||
| def best_params(self) -> jnp.ndarray: | ||
| return self.jnp_array(self._best_params) | ||
|
|
||
| @best_params.setter | ||
| def best_params(self, params: Union[np.ndarray, jnp.ndarray]) -> None: | ||
| self._best_params = jnp.array(params) | ||
| self.crfm.set_m(self._best_params.copy()) | ||
|
|
||
| class CRFM(): | ||
| def __init__(self, num_dims: int, popsize: int, input_sigma: float, rng: jax.random.PRNGKey): | ||
| """Fast Moving Natural Evolution Strategy | ||
| for High-Dimensional Problems (CR-FM-NES), see https://arxiv.org/abs/2201.11422 . | ||
| Derived from https://github.com/nomuramasahir0/crfmnes""" | ||
| if popsize % 2 == 1: # requires even popsize | ||
| popsize += 1 | ||
| self.lamb = popsize | ||
| self.dim = num_dims | ||
| self.sigma = input_sigma | ||
| self.rng = rng | ||
| self.m = jnp.full((self.dim, 1), 0) | ||
| self.v = jax.random.normal(rng, (self.dim, 1)) / jnp.sqrt(self.dim) | ||
| self.D = jnp.ones([self.dim, 1]) | ||
|
|
||
| self.w_rank_hat = (jnp.log(self.lamb / 2 + 1) - jnp.log(jnp.arange(1, self.lamb + 1))).reshape(self.lamb, 1) | ||
| self.w_rank_hat = self.w_rank_hat.at[jnp.where(self.w_rank_hat < 0)].set(0) | ||
| self.w_rank = self.w_rank_hat / sum(self.w_rank_hat) - (1. / self.lamb) | ||
| self.mueff = float(1 / jnp.dot((self.w_rank + (1 / self.lamb)).T, (self.w_rank + (1 / self.lamb)))[0][0]) | ||
|
|
||
| self.cs = (self.mueff + 2.) / (self.dim + self.mueff + 5.) | ||
| self.cc = (4. + self.mueff / self.dim) / (self.dim + 4. + 2. * self.mueff / self.dim) | ||
| self.c1_cma = 2. / (math.pow(self.dim + 1.3, 2) + self.mueff) | ||
| # initialization | ||
| self.chiN = math.sqrt(self.dim) * (1. - 1. / (4. * self.dim) + 1. / (21. * self.dim * self.dim)) | ||
| self.pc = jnp.zeros((self.dim, 1)) | ||
| self.ps = jnp.zeros((self.dim, 1)) | ||
| # distance weight parameter | ||
| self.h_inv = get_h_inv(self.dim) | ||
| self.alpha_dist = lambda lambF: self.h_inv * min(1., math.sqrt(self.lamb / self.dim)) * math.sqrt( | ||
| lambF / self.lamb) | ||
| self.w_dist_hat = lambda z, lambF: exp(self.alpha_dist(lambF) * jnp.linalg.norm(z)) | ||
| # learning rate | ||
| self.eta_m = 1.0 | ||
| self.eta_move_sigma = 1. | ||
| self.eta_stag_sigma = lambda lambF: math.tanh((0.024 * lambF + 0.7 * self.dim + 20.) / (self.dim + 12.)) | ||
| self.eta_conv_sigma = lambda lambF: 2. * math.tanh((0.025 * lambF + 0.75 * self.dim + 10.) / (self.dim + 4.)) | ||
| self.c1 = lambda lambF: self.c1_cma * (self.dim - 5) / 6 * (lambF / self.lamb) | ||
| self.eta_B = lambda lambF: jnp.tanh((min(0.02 * lambF, 3 * jnp.log(self.dim)) + 5) / (0.23 * self.dim + 25)) | ||
|
|
||
| self.g = 0 | ||
| self.no_of_evals = 0 | ||
| self.iteration = 0 | ||
| self.stop = 0 | ||
|
|
||
| self.idxp = jnp.arange(self.lamb / 2, dtype=int) | ||
| self.idxm = jnp.arange(self.lamb / 2, self.lamb, dtype=int) | ||
| self.z = jnp.zeros([self.dim, self.lamb]) | ||
|
|
||
| self.f_best = float('inf') | ||
| self.x_best = jnp.empty(self.dim) | ||
|
|
||
| def set_m(self, params: jnp.ndarray): | ||
| self.m = jnp.array(params).reshape((self.dim, 1)) | ||
|
|
||
| def ask(self) -> jnp.ndarray: | ||
| key, self.rng = jax.random.split(self.rng) | ||
| zhalf = jax.random.normal(key, (self.dim, int(self.lamb / 2))) | ||
| self.z = self.z.at[:, self.idxp].set(zhalf) | ||
| self.z = self.z.at[:, self.idxm].set(-zhalf) | ||
| self.normv = jnp.linalg.norm(self.v) | ||
| self.normv2 = self.normv ** 2 | ||
| self.vbar = self.v / self.normv | ||
| self.y = self.z + ((jnp.sqrt(1 + self.normv2) - 1) * jnp.dot(self.vbar, jnp.dot(self.vbar.T, self.z))) | ||
| self.x = self.m + (self.sigma * self.y) * self.D | ||
| return self.x.T | ||
|
|
||
| def tell(self, evals_no_sort: np.ndarray) -> None: | ||
| sorted_indices = sort_indices_by(evals_no_sort, self.z) | ||
| best_eval_id = sorted_indices[0] | ||
| f_best = evals_no_sort[best_eval_id] | ||
| x_best = self.x[:, best_eval_id] | ||
| self.z = self.z[:, sorted_indices] | ||
| y = self.y[:, sorted_indices] | ||
| x = self.x[:, sorted_indices] | ||
| self.no_of_evals += self.lamb | ||
| self.g += 1 | ||
| if f_best < self.f_best: | ||
| self.f_best = f_best | ||
| self.x_best = x_best | ||
| # This operation assumes that if the solution is infeasible, infinity comes in as input. | ||
| lambF = jnp.sum(evals_no_sort < jnp.finfo(float).max) | ||
| # evolution path p_sigma | ||
| self.ps = (1 - self.cs) * self.ps + jnp.sqrt(self.cs * (2. - self.cs) * self.mueff) * jnp.dot(self.z, self.w_rank) | ||
| ps_norm = jnp.linalg.norm(self.ps) | ||
| # distance weight | ||
| f1 = self.h_inv * min(1., math.sqrt(self.lamb / self.dim)) * math.sqrt(lambF / self.lamb) | ||
| w_tmp = self.w_rank_hat * jnp.exp(jnp.linalg.norm(self.z, axis = 0) * f1).reshape((self.lamb,1)) | ||
| weights_dist = w_tmp / sum(w_tmp) - 1. / self.lamb | ||
| # switching weights and learning rate | ||
| weights = weights_dist if ps_norm >= self.chiN else self.w_rank | ||
| eta_sigma = self.eta_move_sigma if ps_norm >= self.chiN else self.eta_stag_sigma( | ||
| lambF) if ps_norm >= 0.1 * self.chiN else self.eta_conv_sigma(lambF) | ||
| # update pc, m | ||
| wxm = jnp.dot(x - self.m, weights) | ||
| self.pc = (1. - self.cc) * self.pc + jnp.sqrt(self.cc * (2. - self.cc) * self.mueff) * wxm / self.sigma | ||
| self.m += self.eta_m * wxm | ||
| normv4 = self.normv2 ** 2 | ||
| exY = jnp.append(y, self.pc / self.D, axis=1) # dim x lamb+1 | ||
| yy = exY * exY # dim x lamb+1 | ||
| ip_yvbar = jnp.dot(self.vbar.T, exY) | ||
| yvbar = exY * self.vbar # dim x lamb+1. exYのそれぞれの列にvbarがかかる | ||
| gammav = 1. + self.normv2 | ||
| vbarbar = self.vbar * self.vbar | ||
| alphavd = min( | ||
| [1, math.sqrt(normv4 + (2 * gammav - math.sqrt(gammav)) / jnp.max(vbarbar)) / (2 + self.normv2)]) # scalar | ||
| t = exY * ip_yvbar - self.vbar * (ip_yvbar ** 2 + gammav) / 2 # dim x lamb+1 | ||
| b = -(1 - alphavd ** 2) * normv4 / gammav + 2 * alphavd ** 2 | ||
| H = jnp.ones([self.dim, 1]) * 2 - (b + 2 * alphavd ** 2) * vbarbar # dim x 1 | ||
| invH = H ** (-1) | ||
| s_step1 = yy - self.normv2 / gammav * (yvbar * ip_yvbar) - jnp.ones([self.dim, self.lamb + 1]) # dim x lamb+1 | ||
| ip_vbart = jnp.dot(self.vbar.T, t) # 1 x lamb+1 | ||
| s_step2 = s_step1 - alphavd / gammav * ((2 + self.normv2) * (t * self.vbar) - self.normv2 * jnp.dot(vbarbar, ip_vbart)) # dim x lamb+1 | ||
| invHvbarbar = invH * vbarbar | ||
| ip_s_step2invHvbarbar = jnp.dot(invHvbarbar.T, s_step2) # 1 x lamb+1 | ||
| div = 1 + b * jnp.dot(vbarbar.T, invHvbarbar) | ||
| if jnp.amin(abs(div)) == 0: | ||
| self.logger.info('error: div is zero') | ||
| return | ||
| s = (s_step2 * invH) - b / div * jnp.dot(invHvbarbar, ip_s_step2invHvbarbar) # dim x lamb+1 | ||
| ip_svbarbar = jnp.dot(vbarbar.T, s) # 1 x lamb+1 | ||
| t = t - alphavd * ((2 + self.normv2) * (s * self.vbar) - jnp.dot(self.vbar, ip_svbarbar)) # dim x lamb+1 | ||
| # update v, D | ||
| exw = jnp.append(self.eta_B(lambF) * weights, jnp.full((1, 1), self.c1(lambF)), axis=0) # lamb+1 x 1 | ||
| self.v = self.v + jnp.dot(t, exw) / self.normv | ||
| self.D = self.D + jnp.dot(s, exw) * self.D | ||
| # calculate detA | ||
| if jnp.amin(self.D) < 0: | ||
| self.logger.info('error: invalid D') | ||
| return | ||
| nthrootdetA = exp(jnp.sum(jnp.log(self.D)) / self.dim + jnp.log(1 + jnp.dot(self.v.T, self.v)[0][0]) / (2 * self.dim)) | ||
| self.D = self.D / nthrootdetA | ||
| # update sigma | ||
| G_s = jnp.sum( jnp.dot( (self.z * self.z - jnp.ones([self.dim, self.lamb])), weights )) / self.dim | ||
| self.sigma = self.sigma * exp(eta_sigma / 2 * G_s) | ||
|
|
||
| def get_h_inv(dim: int) -> float: | ||
| f = lambda a, b: ((1. + a * a) * exp(a * a / 2.) / 0.24) - 10. - dim | ||
| f_prime = lambda a: (1. / 0.24) * a * exp(a * a / 2.) * (3. + a * a) | ||
| h_inv = 1.0 | ||
| while abs(f(h_inv, dim)) > 1e-10: | ||
| h_inv = h_inv - 0.5 * (f(h_inv, dim) / f_prime(h_inv)) | ||
| return h_inv | ||
|
|
||
| def exp(a: float) -> float: | ||
| return math.exp(min(100, a)) # avoid overflow | ||
|
|
||
| def sort_indices_by(evals: np.ndarray, z: jnp.ndarray) -> jnp.ndarray: | ||
| lam = len(evals) | ||
| sorted_indices = np.argsort(evals) | ||
| sorted_evals = evals[sorted_indices] | ||
| no_of_feasible_solutions = np.where(sorted_evals != jnp.inf)[0].size | ||
| if no_of_feasible_solutions != lam: | ||
| infeasible_z = z[:, np.where(evals == jnp.inf)[0]] | ||
| distances = np.sum(infeasible_z ** 2, axis=0) | ||
| infeasible_indices = sorted_indices[no_of_feasible_solutions:] | ||
| indices_sorted_by_distance = np.argsort(distances) | ||
| sorted_indices = sorted_indices.at[no_of_feasible_solutions:].set(infeasible_indices[indices_sorted_by_distance]) | ||
| return sorted_indices |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,15 @@ | ||
| es_name: "CRFMNES" | ||
| problem_type: "brax" | ||
| env_name: "ant" | ||
| normalize: true | ||
| es_config: | ||
| pop_size: 128 | ||
| init_stdev: 0.05 | ||
| num_tests: 128 | ||
| n_repeats: 16 | ||
| max_iter: 2000 | ||
| test_interval: 100 | ||
| log_interval: 20 | ||
| seed: 42 | ||
| gpu_id: [0, 1, 2, 3] | ||
| debug: false |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,15 @@ | ||
| es_name: "CRFMNES" | ||
| problem_type: "cartpole_easy" | ||
| normalize: false | ||
| es_config: | ||
| pop_size: 64 | ||
| init_stdev: 0.3 | ||
| hidden_size: 64 | ||
| num_tests: 100 | ||
| n_repeats: 16 | ||
| max_iter: 1000 | ||
| test_interval: 100 | ||
| log_interval: 50 | ||
| seed: 42 | ||
| gpu_id: 0 | ||
| debug: false |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,15 @@ | ||
| es_name: "CRFMNES" | ||
| problem_type: "cartpole_hard" | ||
| normalize: false | ||
| es_config: | ||
| pop_size: 64 | ||
| init_stdev: 0.3 | ||
| hidden_size: 64 | ||
| num_tests: 100 | ||
| n_repeats: 16 | ||
| max_iter: 1000 | ||
| test_interval: 100 | ||
| log_interval: 50 | ||
| seed: 42 | ||
| gpu_id: 0 | ||
| debug: false |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,16 @@ | ||
| es_name: "CRFMNES" | ||
| problem_type: "mnist" | ||
| normalize: false | ||
| es_config: | ||
| pop_size: 128 | ||
| init_stdev: 0.1 | ||
| hidden_size: 100 | ||
| batch_size: 1024 | ||
| max_iter: 2000 | ||
| test_interval: 500 | ||
| log_interval: 100 | ||
| num_tests: 1 | ||
| n_repeats: 1 | ||
| seed: 42 | ||
| gpu_id: [0, 1, 2, 3] | ||
| debug: false |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,16 @@ | ||
| es_name: "CRFMNES" | ||
| problem_type: "slimevolley" | ||
| env_name: "ant" | ||
| normalize: true | ||
| es_config: | ||
| pop_size: 128 | ||
| init_stdev: 0.3 | ||
| hidden_size: 100 | ||
| num_tests: 128 | ||
| n_repeats: 16 | ||
| max_iter: 6000 | ||
| test_interval: 100 | ||
| log_interval: 20 | ||
| seed: 42 | ||
| gpu_id: [0, 1, 2, 3] | ||
| debug: false |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,15 @@ | ||
| es_name: "CRFMNES" | ||
| problem_type: "waterworld" | ||
| normalize: false | ||
| es_config: | ||
| pop_size: 256 | ||
| init_stdev: 0.05 | ||
| hidden_size: 100 | ||
| num_tests: 100 | ||
| n_repeats: 32 | ||
| max_iter: 1000 | ||
| test_interval: 50 | ||
| log_interval: 10 | ||
| seed: 42 | ||
| gpu_id: [0, 1, 2, 3] | ||
| debug: false |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,15 @@ | ||
| es_name: "CRFMNES" | ||
| problem_type: "waterworld_ma" | ||
| normalize: false | ||
| es_config: | ||
| pop_size: 16 | ||
| init_stdev: 0.09 | ||
| hidden_size: 100 | ||
| num_tests: 16 | ||
| n_repeats: 64 | ||
| max_iter: 2000 | ||
| test_interval: 100 | ||
| log_interval: 10 | ||
| seed: 97 | ||
| gpu_id: [0, 1, 2, 3] | ||
| debug: false |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.