多层感知器神经网络具有10%的准确度

时间:2016-12-10 04:41:43

标签: numpy machine-learning scipy scikit-learn neural-network

我正在使用CIFAR-10数据集来训练我的神经网络,我得到的分数约为0.10,就好像我的神经网络只是在猜测(正确答案的1/10几率) ,CIFAR-10有10种类型的东西(汽车,飞机,猫,狗等)的32x32图像。我怀疑我的代码有问题。

顺便说一下,文件unpickling不是问题

请帮助我!!

from sklearn import datasets
from sklearn.neural_network import MLPClassifier
import numpy as np
import time

labels = ["airplane", "automobile", "bird", "cat", "deer", "dog", "frog", "horse", "ship", "truck"]

def unpickle(f):
    import cPickle
    fo = open(f, "rb")
    dict = cPickle.load(fo)
    fo.close()
    return dict

def setup_mlp(unpickled_data):
    new_mlp = MLPClassifier(hidden_layer_sizes=3072, solver='sgd', batch_size=1000, max_iter=500, random_state=1, learning_rate_init=0.01)
    return new_mlp
#hidden_layer_sizes is the number of neurons in a layer. In this case I have one hidden layer & 3072 neurons in that layer
if __name__ == "__main__":
    unpickled_batch = unpickle("./data_batch_1")
    print int(unpickled_batch["data"].shape[0])
    #time.sleep(1000)
    X_train = unpickled_batch["data"][:1000]
    Y_train = unpickled_batch["labels"][:1000]
    print "Decoded batch, now training\n"
    mlp = setup_mlp(unpickled_batch)
    mlp.fit(X_train, Y_train)
    print "Score=" + str(mlp.score(X_train, Y_train))

1 个答案:

答案 0 :(得分:0)

昨天我遇到了类似的问题,所以也许这可以帮到你。

如果您将数据集标准化,MLPClassifier在Cifar-10上运行良好,即 添加

from sklearn import preprocessing X_train = preprocessing.scale(X_train)

加载数据后

但这还不够,因为你的测试集与训练集相同,你最终得分= 1.0,没有新图像预测的信息。

要解决此问题,您可以加载' test_batch'来自cifar-10文件或拆分unpickled_batch ["数据"]和unpickled_batch ["标签"]来训练i测试数据集。

之后你应该得到大约0.3分,这取决于测试集的大小。