模型权重不改变Keras

时间:2017-05-26 05:56:35

标签: python random neural-network keras

我正在尝试在keras中创建具有随机权重的神经网络。我正在使用模型的set_weights()函数来分配随机权重。但是,无论权重如何,model.predict()都会在某个输入上提供相同的输出。每次运行程序时输出都不同,但在程序运行时它是相同的。这是代码:

ConnectFourAI.py:

from keras.models import Sequential
from keras.layers import Dense
from minimax2 import ConnectFour
import numpy as np
from time import sleep
import itertools
import random
import time

def get_model():

    model = Sequential()
    model.add(Dense(630, input_dim=84, kernel_initializer='uniform', activation='relu'))
    model.add(Dense(630,kernel_initializer='normal', activation='relu'))
    model.add(Dense(7, kernel_initializer='normal', activation='relu'))
    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
    return model

map = {
    'x':[1,0],
    ' ':[0,0],
    'o':[0,1]
}

model = get_model()

def get_AI_move(grid):
    global model
    inp = np.array(list(itertools.chain.from_iterable([map[t] for t in np.array(grid).reshape(42)]))).reshape(1,84)
    nnout = model.predict(inp)
    # print(list(nnout[0]))
    out = np.argmax(nnout)
    while grid[0][out] != " ":
        out = np.random.randint(7)
    print("out = %d"%out)
    return out

shapes = [(w.shape) for w in model.get_weights()]

print(list(model.get_weights()[0][0][0:5]))
def score_func(x, win):
        if win == "x":
            return 10000
        elif win == " ":
            return 2000
        else:
            return x**2




if __name__=="__main__":

    for i in range(100):
        weights = [np.random.randn(*s) for s in shapes]
        # print(list(weights[0][0][0:5]))
        model.set_weights(weights)
        print(list(model.get_weights()[0][0][0:5]))
        game = ConnectFour()
        game.start_new()
        rounds = game._round
        win = game._winner
        score = score_func(rounds, win)
        print("%dth game scored %.3f"%(i+1,score))

        seed = int(time.time()* 10**6)%(2**32)+1
        np.random.seed(seed)

要重新创建此错误,您需要一个额外的文件。在这个文件中一切都很好,但对random的唯一调用总是给出相同的值。这是file

1 个答案:

答案 0 :(得分:0)

我不知道到底出了什么问题,但我想出了一个解决方法。显然,随机模块中存在一些问题,因为当从2个不同的文件调用随机模块时会发生这种行为。所以我使用了一个文件而不是两个文件,并得到了我期望的结果。