BP神经网络的python实现，里面包含adam、RMSProp等多种优化算法，网络结构参数化实现.zip

import numpy as np from L_layer_model import * import matplotlib.pyplot as plt from predict import * plt.rcParams['font.sans-serif']=['SimHei'] plt.rcParams['axes.unicode_minus']=False def opt_iter_compare(train_x,train_y,hidden_layer_sizes,learning_rate,mini_batch_size,test_x,test_y): layers_dims=[train_x.shape[0],hidden_layer_sizes,train_y.shape[0]] parameters3 = L_layer_model(train_x, train_y, layers_dims, learning_rate=learning_rate, mini_batch_size=mini_batch_size, lambd=0, num_iterations=1000, beta=0.9, optimizer='gd', print_cost=True, isPlot=True, over_stop=True) parameters2 = L_layer_model(train_x, train_y, layers_dims, learning_rate=learning_rate, mini_batch_size=mini_batch_size, lambd=0, num_iterations=1000, beta=0.9, optimizer='momentum', print_cost=True, isPlot=True, over_stop=True) parameters1 = L_layer_model(train_x, train_y,layers_dims, learning_rate=learning_rate, mini_batch_size=mini_batch_size, lambd=0, num_iterations=1000, beta=0.9, optimizer='RMSprop', print_cost=True, isPlot=True, over_stop=True) parameters0 = L_layer_model(train_x, train_y, layers_dims, learning_rate=learning_rate, mini_batch_size=mini_batch_size, lambd=0, num_iterations=1000, beta=0.9, optimizer='adam', print_cost=True, isPlot=True, over_stop=True) cost, error1, testy2 = predict(train_x, train_y, parameters3[0], L=len(layers_dims), lambd=0) cost, error2, testy2 = predict(test_x, test_y, parameters3[0], L=len(layers_dims), lambd=0) print(error1,error2) cost, error1, testy2 = predict(train_x, train_y, parameters2[0], L=len(layers_dims), lambd=0) cost, error2, testy2 = predict(test_x, test_y, parameters2[0], L=len(layers_dims), lambd=0) print(error1, error2) cost, error1, testy2 = predict(train_x, train_y, parameters1[0], L=len(layers_dims), lambd=0) cost, error2, testy2 = predict(test_x, test_y, parameters1[0], L=len(layers_dims), lambd=0) print(error1, error2) cost, error1, testy2 = predict(train_x, train_y, parameters0[0], L=len(layers_dims), lambd=0) cost, error2, testy2 = predict(test_x, test_y, parameters0[0], L=len(layers_dims), lambd=0) print(error1, error2) print(len(parameters3[1])-11,len(parameters2[1])-11,len(parameters1[1])-11,len(parameters0[1])-11) def normal(X): X_mean=np.mean(X,axis=0) X_std=np.std(X,axis=0) X_normal=(X-X_mean)/(X_std+1e-8) return X_normal,X_mean,X_std def opt_iter_compare_impl(): dataset = np.loadtxt('test71.csv', delimiter=',') X1 = dataset[0:14000, 0:61] # 800x61 # 10dB X1[0:14000, 60:61] = dataset[0:14000, 70:71] # X1=np.hstack((X1,dataset[0:400,65:])) Y1 = dataset[0:14000, 65:66] # 800x5 X_normal, X_mean, X_std = normal(X1) train_x = X_normal.T # 61x800 train_y = Y1.T # 5x800 x2 = dataset[14000:20000, 0:61] x2[0:6000, 60:61] = dataset[14000:20000, 70:71] y2 = dataset[14000:20000, 65:66] x2 = (x2 - X_mean) / (X_std + 1e-8) test_x = x2.T # 61x800 test_y = y2.T # 5x800 opt_iter_compare(train_x, train_y, 8, 0.0007112048067781445, 32, test_x, test_y) plt.show() def dist_error_get(test_x,test_y,parameters,L): a = [] b = [] d = [] e = [] f = [] g = [] h = [] p = [] q = [] cost, error, testy2 = predict(test_x, test_y, parameters, L=L, lambd=0) pred_y = np.squeeze(testy2) true_y = np.squeeze(test_y) for i in range(0, 6000): if (true_y[i] > 5 and true_y[i] <= 10): d.append(abs(pred_y[i] - true_y[i])) if (true_y[i] > 10 and true_y[i] <= 15): e.append(abs(pred_y[i] - true_y[i])) if (true_y[i] > 15 and true_y[i] <= 20): f.append(abs(pred_y[i] - true_y[i])) if (true_y[i] > 20 and true_y[i] <= 25): g.append(abs(pred_y[i] - true_y[i])) if (true_y[i] > 25 and true_y[i] <= 30): h.append(abs(pred_y[i] - true_y[i])) if (true_y[i] > 30 and true_y[i] <= 35): p.append(abs(pred_y[i] - true_y[i])) if (true_y[i] > 35 and true_y[i] <= 40): q.append(abs(pred_y[i] - true_y[i])) # print(len(d), len(e), len(f), len(g), len(h), len(p), len(q)) d = np.mean(d) e = np.mean(e) f = np.mean(f) g = np.mean(g) h = np.mean(h) p = np.mean(p) q = np.mean(q) a = [d, e, f, g, h, p, q] return a ''' 获取训练集和验证集 输入1，代表信噪比为10.输入为2，代表信噪比为20，以此类推 ''' def get_Xy(i): dataset = np.loadtxt('test71.csv', delimiter=',') X1 = dataset[0:14000, 0:61] # 800x61 # 10dB X1[0:14000, 60:61] = dataset[0:14000, 66+i:67+i] # X1=np.hstack((X1,dataset[0:400,65:])) Y1 = dataset[0:14000, 65:66] # 800x5 X_normal, X_mean, X_std = normal(X1) train_x = X_normal.T # 61x800 train_y = Y1.T # 5x800 x2 = dataset[14000:20000, 0:61] x2[0:6000, 60:61] = dataset[14000:20000, 66+i:67+i] y2 = dataset[14000:20000, 65:66] x2 = (x2 - X_mean) / (X_std + 1e-8) test_x = x2.T # 61x800 test_y = y2.T # 5x800 return train_x,train_y,test_x,test_y def dist_snr_compare(): train_x, train_y, test_x, test_y=get_Xy(1) layers_dims = [61, 9, 1] parameters3 = L_layer_model(train_x, train_y, layers_dims, learning_rate=0.002567161349434759, mini_batch_size=64, lambd=0, num_iterations=1000, beta=0.9, optimizer='adam', print_cost=True, isPlot=True, over_stop=False) a=dist_error_get(test_x,test_y,parameters3[0],3) train_x, train_y, test_x, test_y = get_Xy(2) layers_dims = [61, 9, 1] parameters3 = L_layer_model(train_x, train_y, layers_dims, learning_rate=0.008882819966208407, mini_batch_size=128, lambd=0, num_iterations=1000, beta=0.9, optimizer='adam', print_cost=True, isPlot=True, over_stop=False) a1 = dist_error_get(test_x, test_y, parameters3[0], 3) train_x, train_y, test_x, test_y = get_Xy(3) layers_dims = [61, 12, 1] parameters3 = L_layer_model(train_x, train_y, layers_dims, learning_rate=0.0018417745354362648, mini_batch_size=32, lambd=0, num_iterations=1000, beta=0.9, optimizer='adam', print_cost=True, isPlot=True, over_stop=False) a2 = dist_error_get(test_x, test_y, parameters3[0], 3) train_x, train_y, test_x, test_y = get_Xy(4) layers_dims = [61, 8, 1] parameters3 = L_layer_model(train_x, train_y, layers_dims, learning_rate=0.007112048067781445, mini_batch_size=32, lambd=0, num_iterations=1000, beta=0.9, optimizer='adam', print_cost=True, isPlot=True, over_stop=False) a3 = dist_error_get(test_x, test_y, parameters3[0], 3) b = ['5-10', '10-15', '15-20', '20-25', '25-30', '30-35', '35-40'] fig = plt.figure() plt.plot(b, a, 'b*-') plt.plot(b, a1, 'rd-') plt.plot(b, a2, 'ys-') plt.plot(b, a3, 'go-') plt.ylabel('error/m') plt.xlabel('distance/m') plt.legend(['10dB','20dB','30dB','40dB']) c1 = [a[0], a1[0], a2[0], a3[0]] c2 = [a[1], a1[1], a2[1], a3[1]] c3 = [a[2], a1[2], a2[2], a3[2]] c4 = [a[3], a1[3], a2[3], a3[3]] c5 = [a[4], a1[4], a2[4], a3[4]] c6 = [a[5], a1[5], a2[5], a3[5]] c7 = [a[6], a1[6], a2[6], a3[6]] b=['10dB','20dB','30dB','40dB'] plt.savefig('1.png') plt.f