PSO/pso/particle.py

import tensorflow as tf
from tensorflow import keras

# import cupy as cp
import numpy as np
import gc


class Particle:
    def __init__(self, model: keras.models, loss, negative: bool = False):
        self.model = model
        self.loss = loss
        init_weights = self.model.get_weights()
        i_w_, s_, l_ = self._encode(init_weights)
        i_w_ = np.random.rand(len(i_w_)) / 2 - 0.25
        self.velocities = self._decode(i_w_, s_, l_)
        self.negative = negative
        self.best_score = 0
        self.best_weights = init_weights

        del i_w_, s_, l_
        del init_weights
        gc.collect()
        
    def __del__(self):
        del self.model
        del self.loss
        del self.velocities
        del self.negative
        del self.best_score
        del self.best_weights
        gc.collect()

    """
    Returns:
        (cupy array) : 가중치 - 1차원으로 풀어서 반환
        (list) : 가중치의 원본 shape
        (list) : 가중치의 원본 shape의 길이
    """

    def _encode(self, weights: list):
        # w_gpu = cp.array([])
        w_gpu = np.array([])
        lenght = []
        shape = []
        for layer in weights:
            shape.append(layer.shape)
            w_ = layer.reshape(-1)
            lenght.append(len(w_))
            # w_gpu = cp.append(w_gpu, w_)
            w_gpu = np.append(w_gpu, w_)

        return w_gpu, shape, lenght

    """
    Returns:
        (list) : 가중치 원본 shape으로 복원
    """

    def _decode(self, weight: list, shape, lenght):
        weights = []
        start = 0
        for i in range(len(shape)):
            end = start + lenght[i]
            w_ = weight[start:end]
            # w_ = weight[start:end].get()
            w_ = np.reshape(w_, shape[i])
            # w_ = w_.reshape(shape[i])
            weights.append(w_)
            start = end
        del start, end, w_
        del shape, lenght
        del weight

        return weights

    def get_score(self, x, y, renewal: str = "acc"):
        self.model.compile(loss=self.loss, optimizer="sgd", metrics=["accuracy"])
        score = self.model.evaluate(x, y, verbose=0)
        # print(score)
        if renewal == "acc":
            if score[1] > self.best_score:
                self.best_score = score[1]
                self.best_weights = self.model.get_weights()
        elif renewal == "loss":
            if score[0] < self.best_score:
                self.best_score = score[0]
                self.best_weights = self.model.get_weights()

        return score

    def _update_velocity(self, local_rate, global_rate, w, g_best):
        encode_w, w_sh, w_len = self._encode(weights=self.model.get_weights())
        encode_v, v_sh, v_len = self._encode(weights=self.velocities)
        encode_p, p_sh, p_len = self._encode(weights=self.best_weights)
        encode_g, g_sh, g_len = self._encode(weights=g_best)
        r0 = np.random.rand()
        r1 = np.random.rand()
        if self.negative:
            new_v = (
                w * encode_v
                + -1 * local_rate * r0 * (encode_p - encode_w)
                + -1 * global_rate * r1 * (encode_g - encode_w)
            )
        else:
            new_v = (
                w * encode_v
                + local_rate * r0 * (encode_p - encode_w)
                + global_rate * r1 * (encode_g - encode_w)
            )
        self.velocities = self._decode(new_v, w_sh, w_len)
        del encode_w, w_sh, w_len
        del encode_v, v_sh, v_len
        del encode_p, p_sh, p_len
        del encode_g, g_sh, g_len
        del r0, r1

    def _update_velocity_w(self, local_rate, global_rate, w, w_p, w_g, g_best):
        encode_w, w_sh, w_len = self._encode(weights=self.model.get_weights())
        encode_v, v_sh, v_len = self._encode(weights=self.velocities)
        encode_p, p_sh, p_len = self._encode(weights=self.best_weights)
        encode_g, g_sh, g_len = self._encode(weights=g_best)
        r0 = np.random.rand()
        r1 = np.random.rand()
        if self.negative:
            new_v = (
                w * encode_v
                + -1 * local_rate * r0 * (w_p * encode_p - encode_w)
                + -1 * global_rate * r1 * (w_g * encode_g - encode_w)
            )
        else:
            new_v = (
                w * encode_v
                + local_rate * r0 * (w_p * encode_p - encode_w)
                + global_rate * r1 * (w_g * encode_g - encode_w)
            )
        self.velocities = self._decode(new_v, w_sh, w_len)
        del encode_w, w_sh, w_len
        del encode_v, v_sh, v_len
        del encode_p, p_sh, p_len
        del encode_g, g_sh, g_len
        del r0, r1

    def _update_weights(self):
        encode_w, w_sh, w_len = self._encode(weights=self.model.get_weights())
        encode_v, v_sh, v_len = self._encode(weights=self.velocities)
        new_w = encode_w + encode_v
        self.model.set_weights(self._decode(new_w, w_sh, w_len))
        del encode_w, w_sh, w_len
        del encode_v, v_sh, v_len

    def f(self, x, y, weights):
        self.model.set_weights(weights)
        score = self.model.evaluate(x, y, verbose=0)[1]
        if score > 0:
            return 1 / (1 + score)
        else:
            return 1 + np.abs(score)

    def step(self, x, y, local_rate, global_rate, w, g_best, renewal: str = "acc"):
        self._update_velocity(local_rate, global_rate, w, g_best)
        self._update_weights()
        return self.get_score(x, y, renewal)

    def step_w(
        self, x, y, local_rate, global_rate, w, g_best, w_p, w_g, renewal: str = "acc"
    ):
        self._update_velocity_w(local_rate, global_rate, w, w_p, w_g, g_best)
        self._update_weights()
        return self.get_score(x, y, renewal)

    def get_best_score(self):
        return self.best_score

    def get_best_weights(self):
        return self.best_weights