How long have CMA-ES cost on training?

[yrj88351]

How long have CMA-ES cost on training?
I wanna use tensorflow to test the CMA-ES on MNIST, as the same networks as yours, the training time become slower as the iteration increase.....
And the loss_val (cross-entropy) and accuracy have no obvious change after 50 iteration....(it have cost about 5 hours)
I use the same hyper-parameter as your setting...(population:101,sigma:0.01)

[yrj88351]

import os
import numpy as np
import cma
from tensorflow.examples.tutorials.mnist import input_data
import tensorflow as tf
mnist = input_data.read_data_sets('./MNIST',
one_hot=True)

os.environ['CUDA_VISIBLE_DEVICES'] = '2'
config = tf.ConfigProto()
config.gpu_options.allow_growth = True

sess=tf.Session(config=config)

def cal_loss(par):

batch_xs,batch_ys=mnist.train.next_batch(2000)
conv_1=np.array(par[0:200])
conv_1=np.reshape(conv_1,[5,5,1,8])
conv_b1=np.array(par[200:208])
conv_2=np.array(par[208:3408])
conv_2=np.reshape(conv_2,[5,5,8,16])
conv_b2=np.array(par[3408:3424])
w1=np.array(par[3424:11264])
w1=np.reshape(w1,[784,10])
b1=np.array(par[11264:11274])

with sess.as_default():
    
    x=tf.placeholder(dtype=tf.float32,shape=[None,784])
    x_image=tf.reshape(x,[-1,28,28,1])
    y_=tf.placeholder(dtype=tf.float32,shape=[None,10])
    
    kernel_1=tf.placeholder(dtype=tf.float32,shape=[5,5,1,8])
    kernel_b1=tf.placeholder(dtype=tf.float32,shape=[8])

    kernel_2=tf.placeholder(dtype=tf.float32,shape=[5,5,8,16])
    kernel_b2=tf.placeholder(dtype=tf.float32,shape=[16])

    fc_w1=tf.placeholder(dtype=tf.float32,shape=[7*7*16,10])
    fc_b1=tf.placeholder(dtype=tf.float32,shape=[10])
    
    conv1=tf.nn.conv2d(x_image,kernel_1,strides=[1,1,1,1],padding='SAME')
    output_conv1=tf.nn.relu(tf.nn.bias_add(conv1,kernel_b1))
    output_pool1=tf.nn.max_pool(output_conv1,ksize=[1,2,2,1],strides=[1,2,2,1],
                                 padding='SAME')
    
    conv2=tf.nn.conv2d(output_pool1,kernel_2,strides=[1,1,1,1],padding='SAME')
    output_conv2=tf.nn.relu(tf.nn.bias_add(conv2,kernel_b2))
    output_pool2=tf.nn.max_pool(output_conv2,ksize=[1,2,2,1],strides=[1,2,2,1],
                                padding='SAME')
    
    output_pool2_flatten=tf.reshape(output_pool2,[-1,7*7*16])
    y=tf.nn.softmax(tf.matmul(output_pool2_flatten,fc_w1)+fc_b1)

    cross_entropy=tf.reduce_mean(-tf.reduce_sum(y_*tf.log(y+1e-8),reduction_indices=[1]))
    
    init=tf.global_variables_initializer()
    sess.run(init)
    
    loss=sess.run(cross_entropy,feed_dict={
                                   kernel_1:conv_1,
                                   kernel_b1:conv_b1,
                                   kernel_2:conv_2,
                                   kernel_b2:conv_b2,
                                   fc_w1:w1,
                                   fc_b1:b1,
                                   x:(batch_xs),
                                   y_:(batch_ys)
                                   })
    
    return loss

def cal_accuracy(par):

batch_xs=mnist.test.images
batch_ys=mnist.test.labels
conv_1=np.array(par[0:200])
conv_1=np.reshape(conv_1,[5,5,1,8])
conv_b1=np.array(par[200:208])
conv_2=np.array(par[208:3408])
conv_2=np.reshape(conv_2,[5,5,8,16])
conv_b2=np.array(par[3408:3424])
w1=np.array(par[3424:11264])
w1=np.reshape(w1,[784,10])
b1=np.array(par[11264:11274])

with sess.as_default():    
    
    x=tf.placeholder(dtype=tf.float32,shape=[None,784])
    x_image=tf.reshape(x,[-1,28,28,1])
    y_=tf.placeholder(dtype=tf.float32,shape=[None,10])
    
    kernel_1=tf.placeholder(dtype=tf.float32,shape=[5,5,1,8])
    kernel_b1=tf.placeholder(dtype=tf.float32,shape=[8])

    kernel_2=tf.placeholder(dtype=tf.float32,shape=[5,5,8,16])
    kernel_b2=tf.placeholder(dtype=tf.float32,shape=[16])

    fc_w1=tf.placeholder(dtype=tf.float32,shape=[7*7*16,10])
    fc_b1=tf.placeholder(dtype=tf.float32,shape=[10])
    
    conv1=tf.nn.conv2d(x_image,kernel_1,strides=[1,1,1,1],padding='SAME')
    output_conv1=tf.nn.relu(tf.nn.bias_add(conv1,kernel_b1))
    output_pool1=tf.nn.max_pool(output_conv1,ksize=[1,2,2,1],strides=[1,2,2,1],
                                 padding='SAME')
    
    conv2=tf.nn.conv2d(output_pool1,kernel_2,strides=[1,1,1,1],padding='SAME')
    output_conv2=tf.nn.relu(tf.nn.bias_add(conv2,kernel_b2))
    output_pool2=tf.nn.max_pool(output_conv2,ksize=[1,2,2,1],strides=[1,2,2,1],
                                padding='SAME')
    
    output_pool2_flatten=tf.reshape(output_pool2,[-1,7*7*16])
    y=tf.nn.softmax(tf.matmul(output_pool2_flatten,fc_w1)+fc_b1)
    
    correct_prediction=tf.equal(tf.argmax(y,1),tf.argmax(y_,1))
    accuracy=tf.reduce_mean(tf.cast(correct_prediction,tf.float32))

    
    init=tf.global_variables_initializer()
    sess.run(init)
    
    loss=sess.run(accuracy,feed_dict={
                                   kernel_1:conv_1,
                                   kernel_b1:conv_b1,
                                   kernel_2:conv_2,
                                   kernel_b2:conv_b2,
                                   fc_w1:w1,
                                   fc_b1:b1,
                                   x:(batch_xs),
                                   y_:(batch_ys)
                                   })
    
    return loss

def main():

es = cma.CMAEvolutionStrategy(11274 * [0.0],0.01,{'verb_disp':1,'popsize':101})
while not es.stop():
    solutions = es.ask()
    print(np.array(solutions).shape)
    es.tell(solutions, [cal_loss(x) for x in solutions])
    print('loss:',end='')
    print(es.result[1])
    print('accuracy:',end='')
    print(cal_accuracy(es.result[0]))
    es.disp()
    if es.result[1]<0.8:
        break

es.result_pretty()

if name == 'main':
main()

[yrj88351]

I put my code there....
I think i really need your advice ...