autoft.py

# coding:utf-8
# Copyright (c) 2019  PaddlePaddle Authors. All Rights Reserved.
#
# 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.
from multiprocessing.pool import ThreadPool
import cma
import copy
import json
import math
import numpy as np
import six
import time

from tb_paddle import SummaryWriter
from paddlehub.common.logger import logger
from paddlehub.common.utils import mkdir
from paddlehub.autofinetune.evaluator import REWARD_SUM

if six.PY3:
    INF = math.inf
else:
    INF = float("inf")


class BaseTuningStrategy(object):
    def __init__(
            self,
            evaluator,
            cudas=["0"],
            popsize=5,
            output_dir=None,
    ):
        self._num_thread = len(cudas)
        self._popsize = popsize
        self.cudas = cudas
        self.is_cuda_free = {"free": [], "busy": []}
        self.is_cuda_free["free"] = cudas
        self._round = 0

        self.evaluator = evaluator
        self.init_input = evaluator.get_init_params()
        self.num_hparam = len(self.init_input)
        self.best_hparams_all_pop = []
        self.best_reward_all_pop = INF
        self.current_hparams = [[0] * self.num_hparam] * self._popsize
        self.hparams_name_list = [
            param["name"] for param in evaluator.params['param_list']
        ]

        if output_dir is None:
            now = int(time.time())
            time_str = time.strftime("%Y%m%d%H%M%S", time.localtime(now))
            self._output_dir = "output_" + time_str
        else:
            self._output_dir = output_dir
        self.writer = SummaryWriter(logdir=self._output_dir + '/tb_paddle')

    @property
    def thread(self):
        return self._num_thread

    @property
    def popsize(self):
        return self._popsize

    @property
    def output_dir(self):
        return self._output_dir

    @property
    def iteration(self):
        return self._iteration

    @property
    def round(self):
        return self._round

    def set_output_dir(self, output_dir=None):
        if output_dir is not None:
            output_dir = output_dir
        else:
            output_dir = self._output_dir
        return output_dir

    def randomSolution(self):
        solut = [0] * self.num_hparm
        for i in range(self.num_hparm):
            ratio = (np.random.random_sample() - 0.5) * 2.0
            if ratio >= 0:
                solut[i] = (
                    1.0 - self.init_input[i]) * ratio + self.init_input[i]
            else:
                solut[i] = (
                    self.init_input[i] + 1.0) * ratio + self.init_input[i]
        return solut

    def smallPeturb(self):
        for i in range(self.popsize):
            for j in range(self.num_hparm):
                ratio = (np.random.random_sample() - 0.5) * 2.0
                if ratio >= 0:
                    self.current_hparams[i][j] = (
                        1.0 - self.current_hparams[i][j]
                    ) * ratio * self.epsilon + self.current_hparams[i][j]
                else:
                    self.current_hparams[i][j] = (
                        self.current_hparams[i][j] +
                        1.0) * ratio * self.epsilon + self.current_hparams[i][j]

    def estimatePopGradients(self):
        gradients = [[0] * self.num_hparm] * self.popsize
        for i in range(self.popsize):
            for j in range(self.num_hparm):
                gradients[i][j] = self.current_hparams[i][
                    j] - self.best_hparms_all_pop[j]
        return gradients

    def estimateLocalGradients(self):
        gradients = [[0] * self.num_hparm] * self.popsize
        for i in range(self.popsize):
            for j in range(self.num_hparm):
                gradients[i][j] = self.current_hparams[i][
                    j] - self.best_hparams_per_pop[i][j]
        return gradients

    def estimateMomemtum(self):
        popGrads = self.estimatePopGradients()
        localGrads = self.estimateLocalGradients()
        for i in range(self.popsize):
            for j in range(self.num_hparm):
                self.momentums[i][j] = (
                    1 - 3.0 * self.alpha / self.iteration
                ) * self.momentums[i][j] - self.alpha * localGrads[i][
                    j] - self.alpha * popGrads[i][j]

    def is_stop(self):
        return False

    def get_current_hparams(self):
        return self.current_hparams

    def feedback(self, params_list, reward_list):
        return NotImplementedError

    def get_best_hparams(self):
        return self.best_hparams_all_pop

    def get_best_eval_value(self):
        return REWARD_SUM - self.best_reward_all_pop

    def step(self, output_dir):
        solutions = self.get_current_hparams()

        params_cudas_dirs = []
        solution_results = []
        cnt = 0
        solutions_ckptdirs = {}
        mkdir(output_dir)

        for idx, solution in enumerate(solutions):
            cuda = self.is_cuda_free["free"][0]
            ckptdir = output_dir + "/ckpt-" + str(idx)
            log_file = output_dir + "/log-" + str(idx) + ".info"
            params_cudas_dirs.append([solution, cuda, ckptdir, log_file])
            solutions_ckptdirs[tuple(solution)] = ckptdir
            self.is_cuda_free["free"].remove(cuda)
            self.is_cuda_free["busy"].append(cuda)
            if len(params_cudas_dirs
                   ) == self.thread or idx == len(solutions) - 1:
                tp = ThreadPool(len(params_cudas_dirs))
                solution_results += tp.map(self.evaluator.run,
                                           params_cudas_dirs)
                cnt += 1
                tp.close()
                tp.join()
                for param_cuda in params_cudas_dirs:
                    self.is_cuda_free["free"].append(param_cuda[1])
                    self.is_cuda_free["busy"].remove(param_cuda[1])
                params_cudas_dirs = []

        self.feedback(solutions, solution_results)

        return solutions_ckptdirs


class HAZero(BaseTuningStrategy):
    def __init__(
            self,
            evaluator,
            cudas=["0"],
            popsize=1,
            output_dir=None,
            sigma=0.2,
    ):
        super(HAZero, self).__init__(evaluator, cudas, popsize, output_dir)

        self._sigma = sigma

        self.evolution_stratefy = cma.CMAEvolutionStrategy(
            self.init_input, sigma, {
                'popsize': self.popsize,
                'bounds': [-1, 1],
                'AdaptSigma': True,
                'verb_disp': 1,
                'verb_time': 'True',
            })

    @property
    def sigma(self):
        return self._sigma

    def get_current_hparams(self):
        return self.evolution_stratefy.ask()

    def is_stop(self):
        return self.evolution_stratefy.stop()

    def feedback(self, params_list, reward_list):
        self._round = self._round + 1

        local_min_reward = min(reward_list)
        local_min_reward_index = reward_list.index(local_min_reward)
        local_hparams = self.evaluator.convert_params(
            params_list[local_min_reward_index])
        print("The local best eval value in the %s-th round is %s." %
              (self._round - 1, REWARD_SUM - local_min_reward))
        print("The local best hyperparameters are as:")
        for index, hparam_name in enumerate(self.hparams_name_list):
            print("%s=%s" % (hparam_name, local_hparams[index]))

        for i in range(self.popsize):
            if reward_list[i] < self.best_reward_all_pop:
                self.best_hparams_all_pop = self.current_hparams[i]
                self.best_reward_all_pop = reward_list[i]

        best_hparams = self.evaluator.convert_params(self.best_hparams_all_pop)
        for index, name in enumerate(self.hparams_name_list):
            self.writer.add_scalar(
                tag="hyperparameter_tuning/" + name,
                scalar_value=best_hparams[index],
                global_step=self.round)
        self.writer.add_scalar(
            tag="hyperparameter_tuning/best_eval_value",
            scalar_value=self.get_best_eval_value(),
            global_step=self.round)

        self.evolution_stratefy.tell(params_list, reward_list)
        self.evolution_stratefy.disp()

    def get_best_hparams(self):
        return list(self.evolution_stratefy.result.xbest)


class PSHE2(BaseTuningStrategy):
    def __init__(
            self,
            evaluator,
            cudas=["0"],
            popsize=1,
            output_dir=None,
            alpha=0.5,
            epsilon=0.2,
    ):
        super(PSHE2, self).__init__(evaluator, cudas, popsize, output_dir)

        self._alpha = alpha
        self._epsilon = epsilon

        self.best_hparams_per_pop = [[0] * self.num_hparam] * self._popsize
        self.best_reward_per_pop = [INF] * self._popsize
        self.momentums = [[0] * self.num_hparam] * self._popsize
        for i in range(self.popsize):
            self.current_hparams[i] = self.set_random_hparam()

    @property
    def alpha(self):
        return self._alpha

    @property
    def epsilon(self):
        return self._epsilon

    def set_random_hparam(self):
        solut = [0] * self.num_hparam
        for i in range(self.num_hparam):
            ratio = (np.random.random_sample() - 0.5) * 2.0
            if ratio >= 0:
                solut[i] = (
                    1.0 - self.init_input[i]) * ratio + self.init_input[i]
            else:
                solut[i] = (
                    self.init_input[i] + 1.0) * ratio + self.init_input[i]
        return solut

    def small_peturb(self):
        for i in range(self.popsize):
            for j in range(self.num_hparam):
                ratio = (np.random.random_sample() - 0.5) * 2.0
                if ratio >= 0:
                    self.current_hparams[i][j] = (
                        1.0 - self.current_hparams[i][j]
                    ) * ratio * self.epsilon + self.current_hparams[i][j]
                else:
                    self.current_hparams[i][j] = (
                        self.current_hparams[i][j] +
                        1.0) * ratio * self.epsilon + self.current_hparams[i][j]

    def estimate_popgradients(self):
        gradients = [[0] * self.num_hparam] * self.popsize
        for i in range(self.popsize):
            for j in range(self.num_hparam):
                gradients[i][j] = self.current_hparams[i][
                    j] - self.best_hparams_all_pop[j]
        return gradients

    def estimate_local_gradients(self):
        gradients = [[0] * self.num_hparam] * self.popsize
        for i in range(self.popsize):
            for j in range(self.num_hparam):
                gradients[i][j] = self.current_hparams[i][
                    j] - self.best_hparams_per_pop[i][j]
        return gradients

    def estimate_momemtum(self):
        popGrads = self.estimate_popgradients()
        localGrads = self.estimate_local_gradients()
        for i in range(self.popsize):
            for j in range(self.num_hparam):
                self.momentums[i][j] = (
                    1 - 3.0 * self.alpha / self.round
                ) * self.momentums[i][j] - self.alpha * localGrads[i][
                    j] - self.alpha * popGrads[i][j]

    def is_stop(self):
        return False

    def feedback(self, params_list, reward_list):
        self._round = self._round + 1

        local_min_reward = min(reward_list)
        local_min_reward_index = reward_list.index(local_min_reward)
        local_hparams = self.evaluator.convert_params(
            params_list[local_min_reward_index])
        print("The local best eval value in the %s-th round is %s." %
              (self._round - 1, REWARD_SUM - local_min_reward))
        print("The local best hyperparameters are as:")
        for index, hparam_name in enumerate(self.hparams_name_list):
            print("%s=%s" % (hparam_name, local_hparams[index]))

        for i in range(self.popsize):
            if reward_list[i] < self.best_reward_per_pop[i]:
                self.best_hparams_per_pop[i] = copy.deepcopy(
                    self.current_hparams[i])
                self.best_reward_per_pop[i] = reward_list[i]
            if reward_list[i] < self.best_reward_all_pop:
                self.best_hparams_all_pop = self.current_hparams[i]
                self.best_reward_all_pop = reward_list[i]

        best_hparams = self.evaluator.convert_params(self.best_hparams_all_pop)
        for index, name in enumerate(self.hparams_name_list):
            self.writer.add_scalar(
                tag="hyperparameter_tuning/" + name,
                scalar_value=best_hparams[index],
                global_step=self.round)
        self.writer.add_scalar(
            tag="hyperparameter_tuning/best_eval_value",
            scalar_value=self.get_best_eval_value(),
            global_step=self.round)

        self.estimate_momemtum()
        for i in range(self.popsize):
            for j in range(len(self.init_input)):
                self.current_hparams[i][j] = self.current_hparams[i][
                    j] + self.alpha * self.momentums[i][j]
        self.small_peturb()