title: "Python實現bp神經網絡識別MNIST數據集"
date: 2018-06-18T14:01:49+08:00
tags: [""]
categories: ["python"]
前言
訓練時讀入的是.mat格式的訓練集，測試正確率時用的是png格式的圖片
代碼
#!/usr/bin/env python3
# coding=utf-8
import math
import sys
import os
import numpy as np
from PIL import Image
import scipy.io as sio
def sigmoid(x):
return np.array(list(map(lambda i: 1 / (1 + math.exp(-i)), x)))
def get_train_pattern():
# 返回訓練集的特征和標簽
# current_dir = os.getcwd()
current_dir = "/home/lxp/F/developing_folder/intelligence_system/bpneuralnet/"
train = sio.loadmat(current_dir + "mnist_train.mat")["mnist_train"]
train_label = sio.loadmat(
current_dir + "mnist_train_labels.mat")["mnist_train_labels"]
train = np.where(train > 180, 1, 0) # 二值化
return train, train_label
def get_test_pattern():
# 返回測試集
# base_url = os.getcwd() + "/test/"
base_url = "/home/lxp/F/developing_folder/intelligence_system/bpneuralnet/mnist_test/"
test_img_pattern = []
for i in range(10):
img_url = os.listdir(base_url + str(i))
t = []
for url in img_url:
img = Image.open(base_url + str(i) + "/" + url)
img = img.convert('1') # 二值化
img_array = np.asarray(img, 'i') # 轉化為int數組
img_vector = img_array.reshape(
img_array.shape[0] * img_array.shape[1]) # 展開成一維數組
t.append(img_vector)
test_img_pattern.append(t)
return test_img_pattern
class BPNetwork:
# 神經網絡類
def __init__(self, in_count, hiden_count, out_count, in_rate, hiden_rate):
"""
:param in_count: 輸入層數
:param hiden_count: 隱藏層數
:param out_count: 輸出層數
:param in_rate: 輸入層學習率
:param hiden_rate: 隱藏層學習率
"""
# 各個層的節點數量
self.in_count = in_count
self.hiden_count = hiden_count
self.out_count = out_count
# 輸入層到隱藏層連線的權重隨機初始化
self.w1 = 0.2 * \
np.random.random((self.in_count, self.hiden_count)) - 0.1
# 隱藏層到輸出層連線的權重隨機初始化
self.w2 = 0.2 * \
np.random.random((self.hiden_count, self.out_count)) - 0.1
# 隱藏層偏置向量
self.hiden_offset = np.zeros(self.hiden_count)
# 輸出層偏置向量
self.out_offset = np.zeros(self.out_count)
# 輸入層學習率
self.in_rate = in_rate
# 隱藏層學習率
self.hiden_rate = hiden_rate
def train(self, train_img_pattern, train_label):
if self.in_count != len(train_img_pattern[0]):
sys.exit("輸入層維數與樣本維數不等")
# for num in range(10):
# for num in range(10):
for i in range(len(train_img_pattern)):
if i % 5000 == 0:
print(i)
# 生成目標向量
target = [0] * 10
target[train_label[i][0]] = 1
# for t in range(len(train_img_pattern[num])):
# 前向傳播
# 隱藏層值等於輸入層*w1+隱藏層偏置
hiden_value = np.dot(
train_img_pattern[i], self.w1) + self.hiden_offset
hiden_value = sigmoid(hiden_value)
# 計算輸出層的輸出
out_value = np.dot(hiden_value, self.w2) + self.out_offset
out_value = sigmoid(out_value)
# 反向更新
error = target - out_value
# 計算輸出層誤差
out_error = out_value * (1 - out_value) * error
# 計算隱藏層誤差
hiden_error = hiden_value * \
(1 - hiden_value) * np.dot(self.w2, out_error)
# 更新w2，w2是j行k列的矩陣，存儲隱藏層到輸出層的權值
for k in range(self.out_count):
# 更新w2第k列的值，連接隱藏層所有節點到輸出層的第k個節點的邊
# 隱藏層學習率×輸入層誤差×隱藏層的輸出值
self.w2[:, k] += self.hiden_rate * out_error[k] * hiden_value
# 更新w1
for j in range(self.hiden_count):
self.w1[:, j] += self.in_rate * \
hiden_error[j] * train_img_pattern[i]
# 更新偏置向量
self.out_offset += self.hiden_rate * out_error
self.hiden_offset += self.in_rate * hiden_error
def test(self, test_img_pattern):
"""
測試神經網絡的正確率
:param test_img_pattern[num][t]表示數字num的第t張圖片
:return:
"""
right = np.zeros(10)
test_sum = 0
for num in range(10): # 10個數字
# print("正在識別", num)
num_count = len(test_img_pattern[num])
test_sum += num_count
for t in range(num_count): # 數字num的第t張圖片
hiden_value = np.dot(
test_img_pattern[num][t], self.w1) + self.hiden_offset
hiden_value = sigmoid(hiden_value)
out_value = np.dot(hiden_value, self.w2) + self.out_offset
out_value = sigmoid(out_value)
# print(out_value)
if np.argmax(out_value) == num:
# 識別正確
right[num] += 1
print("數字%d的識別正確率%f" % (num, right[num] / num_count))
# 平均識別率
print("平均識別率為：", sum(right) / test_sum)
"""
def test1:
"""
def run():
# 讀入訓練集
train, train_label = get_train_pattern()
# 讀入測試圖片
test_pattern = get_test_pattern()
# 神經網絡配置參數
in_count = 28 * 28
hiden_count = 6
out_count = 10
in_rate = 0.1
hiden_rate = 0.1
bpnn = BPNetwork(in_count, hiden_count, out_count, in_rate, hiden_rate)
bpnn.train(train, train_label)
bpnn.test(test_pattern)
# 單張測試
# 識別單獨一張圖片，返回識別結果
"""
while True:
img_name = input("輸入要識別的圖片\n")
base_url = "/home/lxp/F/developing_folder/intelligence_system/bpneuralnet/"
img_url = base_url + img_name
img = Image.open(img_url)
img = img.convert('1') # 二值化
img_array = np.asarray(img, 'i') # 轉化為int數組
# 得到圖片的特征向量
img_v = img_array.reshape(img_array.shape[0] * img_array.shape[1]) # 展開成一維數組
bpnn.test1(img_v)
"""
if __name__ == "__main__":
run()
# train, train_label = get_train_pattern()
# print(train_label[5][0])
# test = get_test_pattern()
數據集下載：
鏈接: https://pan.baidu.com/s/1ldWTSqVUm6l1cc4EDOzHpQ 提取碼: mm93