这次作业比较简单按照公式打就行:
首先导包,读取数据并查看数据
import numpy as np
import matplotlib.pyplot as plt
import opt_utils
import time
x_train, y_train = opt_utils.load_dataset()
# print(x_train.shape)
# print(y_train.shape)
构建神经网络
def init(x):
first_num = 9
second_num = 5
third_num = 2
np.random.seed(5)
w1 = np.random.randn(first_num, x.shape[0]) * np.sqrt(2 / x.shape[0])
b1 = np.zeros((first_num, 1))
w2 = np.random.randn(second_num, first_num) * np.sqrt(2 / first_num)
b2 = np.zeros((second_num, 1))
w3 = np.random.randn(third_num, second_num) * np.sqrt(2 / second_num)
b3 = np.zeros((third_num, 1))
w4 = np.random.randn(1, third_num) * np.sqrt(2 / third_num)
b4 = np.zeros((1, 1))
ini_param = {
"w1": w1,
"b1": b1,
"w2": w2,
"b2": b2,
"w3": w3,
"b3": b3,
"w4": w4,
"b4": b4
}
return ini_param
# 构建前向
def cal_z(w, a, b): # a维度:上层节点数*例子数 w维度:该层节点数*上层节点数
return np.dot(w, a) + b
def cal_sigma(z): # z维度:该层节点数*例子数
return 1 / (1 + np.exp(-z))
def cal_tan_h(z): # z1维度hidden_num*x.shape[1]
return np.tanh(z)
def cal_relu(z): # z维度:该层节点数*例子数
return np.maximum(0, z)
def forward_f(x, p):
z1 = cal_z(p["w1"], x, p["b1"])
a1 = cal_relu(z1)
z2 = cal_z(p["w2"], a1, p["b2"])
a2 = cal_relu(z2)
z3 = cal_z(p["w3"], a2, p["b3"])
a3 = cal_relu(z3)
z4 = cal_z(p["w4"], a3, p["b4"])
a4 = cal_sigma(z4)
forward_param = {
"z1": z1,
"a1": a1,
"z2": z2,
"a2": a2,
"z3": z3,
"a3": a3,
"z4": z4,
"a4": a4
}
return forward_param
# 计算损失函数
def cost_f(a, y): # a维度: 1*例子数 y维度:1*例子数
m = y.shape[1]
return -np.sum(y * np.log(a) + (1 - y) * np.log(1 - a)) / m
# 后向
def cal_dz_last(a, y): # ai维度:第i层节点数*例子数 y维度:1*例子数
return a - y
def cal_dw_db(dz, a, m): # dzi维度:第i层节点数*例子数 ai维度:第i层节点数*例子数
return np.dot(dz, a.T) / m, np.sum(dz, axis=1, keepdims=True) / m
def cal_da(dz, w): # dzi维度:第i层节点数*例子数 wi维度:第i层节点数*第i-1层节点数
return np.dot(w.T, dz)
def cal_drelu(da, z): # dai维度:第i层节点数*例子数 zi维度:第i层节点数*例子数
t = np.ones(z.shape)
t[z <= 0] = 0
return da * t
def back_f(p, f_p, x, y):
dz4 = cal_dz_last(f_p["a4"], y)
dw4, db4 = cal_dw_db(dz4, f_p["a3"], y.shape[1])
da3 = cal_da(dz4, p["w4"])
dz3 = cal_drelu(da3, f_p["z3"])
dw3, db3 = cal_dw_db(dz3, f_p["a2"], y.shape[1])
da2 = cal_da(dz3, p["w3"])
dz2 = cal_drelu(da2, f_p["z2"])
dw2, db2 = cal_dw_db(dz2, f_p["a1"], y.shape[1])
da1 = cal_da(dz2, p["w2"])
dz1 = cal_drelu(da1, f_p["z1"])
dw1, db1 = cal_dw_db(dz1, x, y.shape[1])
back_param = {
"dw4": dw4,
"db4": db4,
"dw3": dw3,
"db3": db3,
"dw2": dw2,
"db2": db2,
"dw1": dw1,
"db1": db1
}
return back_param
# 更新参数
def update_p(p, b_p, learning_rate):
upd_p = {
"w1": p["w1"] - learning_rate * b_p["dw1"],
"b1": p["b1"] - learning_rate * b_p["db1"],
"w2": p["w2"] - learning_rate * b_p["dw2"],
"b2": p["b2"] - learning_rate * b_p["db2"],
"w3": p["w3"] - learning_rate * b_p["dw3"],
"b3": p["b3"] - learning_rate * b_p["db3"],
"w4": p["w4"] - learning_rate * b_p["dw4"],
"b4": p["b4"] - learning_rate * b_p["db4"]
}
return upd_p
这边建模,训练。要改成批处理梯度下降和min_batch梯度下降可以通用的:
# 建模
def model(x, y, learning_rate, loop_num, p):
cost_t = []
for i in range(loop_num):
f_p = forward_f(x, p)
b_p = back_f(p, f_p, x, y)
p = update_p(p, b_p, learning_rate)
if i % 10 == 0:
cost_t.append(cost_f(f_p["a4"], y))
return p, np.array(cost_t)
def model_min_batch(x, y, min_batch_size, learning_rate, loop_num):
# 洗牌
size = x.shape[1]
temp = np.random.permutation(size) # 获取一个0-m之间的随机排列序列数组
shuttle_x = x[:, temp]
shuttle_y = y[:, temp]
# 划分
p = init(x)
cost = []
num = int(size / min_batch_size)
if size % min_batch_size != 0:
start = int(num * min_batch_size)
x_min_batch = shuttle_x[:, start:size - 1]
y_min_batch = shuttle_y[:, start:size - 1]
p, cost_t = model(x_min_batch, y_min_batch, learning_rate, loop_num, p)
cost.extend(cost_t)
for i in range(num):
start = int(i * min_batch_size)
end = int((i + 1) * min_batch_size - 1)
x_min_batch = shuttle_x[:, start:end]
y_min_batch = shuttle_y[:, start:end]
p, cost_t = model(x_min_batch, y_min_batch, learning_rate, loop_num, p)
cost.extend(cost_t)
return p, cost
首先,不分块只需将min_batch_size设为数据集大小就行。
model_p_gd, cost_gd = model_min_batch(x_train, y_train, x_train.shape[1], 0.003, 800)
time_start_gd = time.time()
model_p_gd, cost_gd = model_min_batch(x_train, y_train, x_train.shape[1], 0.003, 800)
time_end_gd = time.time()
print("整体运时间:", time_end_gd - time_start_gd)
train_f_p_gd=forward_f(x_train,model_p_gd)
print("整个运行准确度:", 100 * (1 - np.sum(np.abs(np.round(train_f_p_gd["a4"]) - y_train)) / y_train.shape[1]), '%')
print_figure(x_train, y_train, model_p_gd, cost_gd,"batch_gd")
分批运行修改传入参数:
odel_p_mb, cost_mb = model_min_batch(x_train, y_train, 64, 0.003, 800)
time_start_mb = time.time()
model_p_mb, cost_mb = model_min_batch(x_train, y_train, 64, 0.003, 800) # 分成2^n每份最佳
time_end_mb = time.time()
print("分批:", time_end_mb - time_start_mb)
train_f_p_mb=forward_f(x_train,model_p_mb)
print("分批运行准确度:", 100 * (1 - np.sum(np.abs(np.round(train_f_p_mb["a4"]) - y_train)) / y_train.shape[1]), '%')
print_figure(x_train, y_train, model_p_mb, cost_mb,"minbatch")
使用动量梯度下降,要修改以下三个部分:
1.后向加入vdw和vdb
2.更新w和b的值,需要将传入后向参数改为传入v值
3.修改训练的函数
修改后向:
def back_f_mementum(p, f_p, x, y, v, beta1):
dz4 = cal_dz_last(f_p["a4"], y)
dw4, db4 = cal_dw_db(dz4, f_p["a3"], y.shape[1])
v["vdw4"] = beta1 * v["vdw4"] + (1 - beta1) * dw4
v["vdb4"] = beta1 * v["vdb4"] + (1 - beta1) * db4
da3 = cal_da(dz4, p["w4"])
dz3 = cal_drelu(da3, f_p["z3"])
dw3, db3 = cal_dw_db(dz3, f_p["a2"], y.shape[1])
v["vdw3"] = beta1 * v["vdw3"] + (1 - beta1) * dw3
v["vdb3"] = beta1 * v["vdb3"] + (1 - beta1) * db3
da2 = cal_da(dz3, p["w3"])
dz2 = cal_drelu(da2, f_p["z2"])
dw2, db2 = cal_dw_db(dz2, f_p["a1"], y.shape[1])
v["vdw2"] = beta1 * v["vdw2"] + (1 - beta1) * dw2
v["vdb2"] = beta1 * v["vdb2"] + (1 - beta1) * db2
da1 = cal_da(dz2, p["w2"])
dz1 = cal_drelu(da1, f_p["z1"])
dw1, db1 = cal_dw_db(dz1, x, y.shape[1])
v["vdw1"] = beta1 * v["vdw1"] + (1 - beta1) * dw1
v["vdb1"] = beta1 * v["vdb1"] + (1 - beta1) * db1
back_param = {
"dw4": dw4,
"db4": db4,
"dw3": dw3,
"db3": db3,
"dw2": dw2,
"db2": db2,
"dw1": dw1,
"db1": db1
}
return back_param, v
修改更新w、b函数:
def update_p_momentum(p, v, learning_rate):
upd_p = {
"w1": p["w1"] - learning_rate * v["vdw1"],
"b1": p["b1"] - learning_rate * v["vdb1"],
"w2": p["w2"] - learning_rate * v["vdw2"],
"b2": p["b2"] - learning_rate * v["vdb2"],
"w3": p["w3"] - learning_rate * v["vdw3"],
"b3": p["b3"] - learning_rate * v["vdb3"],
"w4": p["w4"] - learning_rate * v["vdw4"],
"b4": p["b4"] - learning_rate * v["vdb4"]
}
return upd_p
修改训练函数:
def model_momentum(x, y, learning_rate, loop_num, p, v, beta1):
cost_t = []
for i in range(loop_num):
f_p = forward_f(x, p)
b_p, v = back_f_mementum(p, f_p, x, y, v, beta1)
p = update_p_momentum(p, v, learning_rate)
if i % 10 == 0:
cost_t.append(cost_f(f_p["a4"], y))
return p, np.array(cost_t), v
def model_min_batch_momentum(x, y, min_batch_size, learning_rate, loop_num, beta1=0.9):
# 洗牌
size = x.shape[1]
temp = np.random.permutation(size) # 获取一个0-m之间的随机排列序列数组
shuttle_x = x[:, temp]
shuttle_y = y[:, temp]
# 划分
p = init(x)
cost = []
num = int(size / min_batch_size)
v = {
"vdw4": 0,
"vdb4": 0,
"vdw3": 0,
"vdb3": 0,
"vdw2": 0,
"vdb2": 0,
"vdw1": 0,
"vdb1": 0
}
if size % min_batch_size != 0:
start = int(num * min_batch_size)
x_min_batch = shuttle_x[:, start:size - 1]
y_min_batch = shuttle_y[:, start:size - 1]
p, cost_t, v = model_momentum(x_min_batch, y_min_batch, learning_rate, loop_num, p, v, beta1)
cost.extend(cost_t)
for i in range(num):
start = int(i * min_batch_size)
end = int((i + 1) * min_batch_size - 1)
x_min_batch = shuttle_x[:, start:end]
y_min_batch = shuttle_y[:, start:end]
p, cost_t, v = model_momentum(x_min_batch, y_min_batch, learning_rate, loop_num, p, v, beta1)
cost.extend(cost_t)
return p, cost
用Adam算法需要修改:
1.后向在mementum基础上引入s
2.更新w和b的值
3修改训练的函数
修改后向:
def back_f_Adam(p, f_p, x, y, v, s, beta1, beta2):
dz4 = cal_dz_last(f_p["a4"], y)
dw4, db4 = cal_dw_db(dz4, f_p["a3"], y.shape[1])
v["vdw4"] = beta1 * v["vdw4"] + (1 - beta1) * dw4
v["vdb4"] = beta1 * v["vdb4"] + (1 - beta1) * db4
s["sdw4"] = beta2 * s["sdw4"] + (1 - beta2) * (dw4 ** 2)
s["sdb4"] = beta2 * s["sdb4"] + (1 - beta2) * (db4 ** 2)
da3 = cal_da(dz4, p["w4"])
dz3 = cal_drelu(da3, f_p["z3"])
dw3, db3 = cal_dw_db(dz3, f_p["a2"], y.shape[1])
v["vdw3"] = beta1 * v["vdw3"] + (1 - beta1) * dw3
v["vdb3"] = beta1 * v["vdb3"] + (1 - beta1) * db3
s["sdw3"] = beta2 * s["sdw3"] + (1 - beta2) * (dw3 ** 2)
s["sdb3"] = beta2 * s["sdb3"] + (1 - beta2) * (db3 ** 2)
da2 = cal_da(dz3, p["w3"])
dz2 = cal_drelu(da2, f_p["z2"])
dw2, db2 = cal_dw_db(dz2, f_p["a1"], y.shape[1])
v["vdw2"] = beta1 * v["vdw2"] + (1 - beta1) * dw2
v["vdb2"] = beta1 * v["vdb2"] + (1 - beta1) * db2
s["sdw2"] = beta2 * s["sdw2"] + (1 - beta2) * (dw2 ** 2)
s["sdb2"] = beta2 * s["sdb2"] + (1 - beta2) * (db2 ** 2)
da1 = cal_da(dz2, p["w2"])
dz1 = cal_drelu(da1, f_p["z1"])
dw1, db1 = cal_dw_db(dz1, x, y.shape[1])
v["vdw1"] = beta1 * v["vdw1"] + (1 - beta1) * dw1
v["vdb1"] = beta1 * v["vdb1"] + (1 - beta1) * db1
s["sdw1"] = beta2 * s["sdw1"] + (1 - beta2) * (dw1 ** 2)
s["sdb1"] = beta2 * s["sdb1"] + (1 - beta2) * (db1 ** 2)
back_param = {
"dw4": dw4,
"db4": db4,
"dw3": dw3,
"db3": db3,
"dw2": dw2,
"db2": db2,
"dw1": dw1,
"db1": db1
}
return back_param, v, s
修改更新函数:
def update_p_Adam(p, v, s, learning_rate, i, beta1, beta2):
e = 0.000000001
vbias_cor = (1 - beta1 ** (i+1)) + e
sbias_cor = (1 - beta2 ** (i+1)) + e
upd_p = {
"w1": p["w1"] - learning_rate * (v["vdw1"] / vbias_cor) / np.sqrt(s["sdw1"] / sbias_cor),
"b1": p["b1"] - learning_rate * (v["vdb1"] / vbias_cor) / np.sqrt(s["sdb1"] / sbias_cor),
"w2": p["w2"] - learning_rate * (v["vdw2"] / vbias_cor) / np.sqrt(s["sdw2"] / sbias_cor),
"b2": p["b2"] - learning_rate * (v["vdb2"] / vbias_cor) / np.sqrt(s["sdb2"] / sbias_cor),
"w3": p["w3"] - learning_rate * (v["vdw3"] / vbias_cor) / np.sqrt(s["sdw3"] / sbias_cor),
"b3": p["b3"] - learning_rate * (v["vdb3"] / vbias_cor) / np.sqrt(s["sdb3"] / sbias_cor),
"w4": p["w4"] - learning_rate * (v["vdw4"] / vbias_cor) / np.sqrt(s["sdw4"] / sbias_cor),
"b4": p["b4"] - learning_rate * (v["vdb4"] / vbias_cor) / np.sqrt(s["sdb4"] / sbias_cor)
}
return upd_p
修改训练函数:
def model_Adam(x, y, learning_rate, loop_num, p, v, s, beta1, beta2):
cost_t = []
for i in range(loop_num):
f_p = forward_f(x, p)
b_p, v, s = back_f_Adam(p, f_p, x, y, v, s, beta1, beta2)
p = update_p_Adam(p, v, s, learning_rate, i,beta1,beta2)
if i % 10 == 0:
cost_t.append(cost_f(f_p["a4"], y))
return p, np.array(cost_t), v, s
def model_min_batch_Adam(x, y, min_batch_size, learning_rate, loop_num, beta1=0.9, beta2=0.999):
# 洗牌
size = x.shape[1]
temp = np.random.permutation(size) # 获取一个0-m之间的随机排列序列数组
shuttle_x = x[:, temp]
shuttle_y = y[:, temp]
# 划分
p = init(x)
cost = []
num = int(size / min_batch_size)
v = {
"vdw4": 0,
"vdb4": 0,
"vdw3": 0,
"vdb3": 0,
"vdw2": 0,
"vdb2": 0,
"vdw1": 0,
"vdb1": 0
}
s = {
"sdw4": 0,
"sdb4": 0,
"sdw3": 0,
"sdb3": 0,
"sdw2": 0,
"sdb2": 0,
"sdw1": 0,
"sdb1": 0
}
if size % min_batch_size != 0:
start = int(num * min_batch_size)
x_min_batch = shuttle_x[:, start:size - 1]
y_min_batch = shuttle_y[:, start:size - 1]
p, cost_t, v, s = model_Adam(x_min_batch, y_min_batch, learning_rate, loop_num, p, v, s, beta1, beta2)
cost.extend(cost_t)
for i in range(num):
start = int(i * min_batch_size)
end = int((i + 1) * min_batch_size - 1)
x_min_batch = shuttle_x[:, start:end]
y_min_batch = shuttle_y[:, start:end]
p, cost_t, v, s = model_Adam(x_min_batch, y_min_batch, learning_rate, loop_num, p, v,s, beta1, beta2)
cost.extend(cost_t)
return p, cost
整体运行时间: 0.28107190132141113
整个运行准确度: 84.33333333333333 %
分批运行时间: 0.9800760746002197
分批运行准确度: 88.66666666666667 %
动量运行时间: 1.166001319885254
动量下降准确度: 88.66666666666667 %
Adam: 1.5370657444000244
Adam准确度: 92.0 %
4次训练采用相同epoch次数和学习率,根据loss图可以看出下降速率Adam>mementum=min_batch>batch gradient decent
min_batch牺牲了向量化,运行时间变长,可能数据太小mementum和min_batch没有明显区别