tf dense layer两种创建方式的对比和numpy实现

最新推荐文章于 2024-08-07 16:11:09 发布

007在学算法

最新推荐文章于 2024-08-07 16:11:09 发布

阅读量683

点赞数 1

本文链接：https://blog.csdn.net/hvgdfx/article/details/113848221

版权

本文详细对比了在TensorFlow中使用`add`方法和继承`Layer`类来构建Dense层的区别。通过实例展示了如何初始化权重和偏置，以及自定义权重矩阵。同时，还探讨了使用numpy实现类似功能的方法。

摘要由CSDN通过智能技术生成

文章目录

1 Dense Layer

目的：构建Dense层，对比两种方法实现的区别

2 对比原始的add layer方法和继承方法的不同

2.1 global config

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras.layers import Input, LSTM, Dense
import numpy as np

np.random.seed(1)

rows = 10000  	# 样本数
columns = 100	# 特征数

train_x1 = np.random.random(size=(int(rows/2), columns))
train_y1 = np.random.choice([0], size=(int(rows/2), 1))
train_x2 = np.random.random(size=(int(rows/2), columns))+1
train_y2 = np.random.choice([1], size=(int(rows/2), 1))

train_x = np.vstack((train_x1, train_x2))
train_y = np.vstack((train_y1, train_y2))

units = 5	# 自定义cell个数

2.1 用add实现

tf.random.set_seed(1)		# 固定随机值

model1 = keras.Sequential()
model1.add(Input(shape=(columns,)))
model1.add(Dense(units=units))

model1.compile(optimizer="adam", loss="mse", metrics=["accuracy"])

model1.fit(train_x, train_y, epochs=10)
model1.predict(train_x)[-1][-1]

l1 = model1.layers[0]
w1, b1 = l1.get_weights()

api中的参数分几部分

initializer 初始项，初始化参数
regularizer 正则项，选择不同正则模式L1L2
constraint 约束项，非负约束或者最大模约束

2.2 用继承实现

tf.random.set_seed(1)		# 固定随机值

class MyDenseLayer(keras.layers.Layer):
	def __init__(self, num_outputs):
		super(MyDenseLayer, self).__init__()
		self.num_outputs = num_outputs
	def build(self, input_shape):
		self.kernel = self.add_weight(name="kernel", shape=[int(input_shape[-1]), self.num_outputs])
		self.bias = self.add_weight(name="bias", shape=[self.num_outputs, ], initializer=keras.initializers.zeros)
		self.build = True
	def call(self, input):
		return tf.matmul(input, self.kernel) + self.bias

model2 = keras.Sequential()
model2.add(Input(shape=(columns,)))
model2.add(MyDenseLayer(units))		

model2.compile(loss="mse", optimizer="adam", metrics=['accuracy'])

model2.fit(train_x, train_y, epochs=10)
model2.predict(train_x)[-1][-1]

l2 = model2.layers[0]
w2, b2 = l2.get_weights()

3 有权重的对比

3.1 用自带add_weight方法自定义权重

自定义初始的权重和偏移项。为什么要定义这两个方法？为了用add_weight方法初始权重和偏移项

def w_init(shape, dtype=tf.float32):
	return tf.random.normal(shape=shape, dtype=dtype)

def b_init(shape, dtype=tf.float32):
	return tf.zeros(shape=shape, dtype=dtype)

3.2 用add实现

tf.random.set_seed(1)		# 固定随机值

model3 = keras.Sequential()
model3.add(Input(shape=(columns,)))
model3.add(Dense(units=units, kernel_initializer=w_init, bias_initializer=b_init))	# 需要固定weights
		
model3.compile(optimizer="adam", loss="mse", metrics=["accuracy"])

model3.fit(train_x, train_y, epochs=10)
model3.predict(train_x)[-1][-1]

l3 = model3.layers[0]
w3, b3 = l3.get_weights()

3.3 用继承实现

tf.random.set_seed(1)		# 固定随机值

class MyDenseLayer(keras.layers.Layer):
	def __init__(self, num_outputs):
		super(MyDenseLayer, self).__init__()
		self.num_outputs = num_outputs
	def build(self, input_shape):
		self.kernel = self.add_weight(initializer=w_init, shape=(input_shape[-1], self.num_outputs), dtype=tf.float32)	# 自定义权重
		self.bias = self.add_weight(initializer=b_init, shape=(self.num_outputs,), dtype=tf.float32)		# 自定义偏移项
	def call(self, input):
		return tf.matmul(input, self.kernel) + self.bias

model4 = keras.Sequential()
model4.add(Input(shape=(columns,)))
model4.add(MyDenseLayer(units))		

model4.compile(loss="mse", optimizer="adam", metrics=['accuracy'])

model4.fit(train_x, train_y, epochs=10)
model4.predict(train_x)[-1][-1]

l4 = model4.layers[0]
w4, b4 = l4.get_weights()

4 用自定义矩阵为权重矩阵

4.1 初始化权重和偏移项矩阵

tf.random.set_seed(1)
w = tf.random.normal(shape=(columns, units), dtype=tf.float32)
b = tf.zeros(shape=(units,), dtype=tf.float32)

def w_init(shape, dtype=tf.float32):
	return w

def b_init(shape, dtype=tf.float32):
	return b

4.2 用add实现

tf.random.set_seed(1)		# 固定随机值

model5 = keras.Sequential()
model5.add(Input(shape=(columns,)))
model5.add(Dense(units=units, kernel_initializer=w_init, bias_initializer=b_init))		

model5.compile(loss="mse", optimizer="adam", metrics=['accuracy'])

model5.fit(train_x, train_y, epochs=10)
model5.predict(train_x)[-1][-1]

4.3 用继承实现

不用Layer.add_weight方法，自己实现一个权重矩阵，然后用权重矩阵作为初始化，进行训练

tf.random.set_seed(1)		# 固定随机值

class MyDenseLayer(keras.layers.Layer):
	def __init__(self, num_outputs):
		super(MyDenseLayer, self).__init__()
		self.num_outputs = num_outputs
	def build(self, input_shape):
		self.kernel = tf.Variable(w, trainable=True)
		self.bias = tf.Variable(b, trainable=True)
	def call(self, input):
		return tf.matmul(input, self.kernel) + self.bias

model6 = keras.Sequential()
model6.add(Input(shape=(columns,)))
model6.add(MyDenseLayer(units))		

model6.compile(loss="mse", optimizer="adam", metrics=['accuracy'])

model6.fit(train_x, train_y, epochs=10)
model6.predict(train_x)[-1][-1]

5 用numpy实现

5.1 初始化权重和偏移项矩阵

tf.random.set_seed(1)

train_x = np.ones(shape=(rows, columns), dtype="float32")	# 这里一定要dtype一致，否则numpy与keras计算结果会有差异，我这里统一使用float32
train_y = np.vstack([np.ones(shape=(int(rows/2), 1), dtype="float32"), np.zeros(shape=(int(rows/2),1), dtype="float32")])

w = tf.random.normal(shape=(columns, 1), dtype=tf.float32)
b = tf.zeros(shape=(1,), dtype=tf.float32)

def w_init(shape, dtype=tf.float32):
	return tf.convert_to_tensor(w, dtype=tf.float32)

def b_init(shape, dtype=tf.float32):
	return tf.convert_to_tensor(b, dtype=tf.float32)

5.2 add实现，有激活函数

tf.random.set_seed(1)		# 固定随机值

model7 = keras.Sequential()
model7.add(Input(shape=(columns,)))
model7.add(Dense(units=1, kernel_initializer=w_init, bias_initializer=b_init, activation="sigmoid"))

h1 = model7.predict(train_x)

model7.compile(loss="mse", optimizer=tf.keras.optimizers.SGD(learning_rate=learning_rate), metrics=['accuracy'])

model7.fit(train_x, train_y, epochs=1, batch_size=rows)	#这里要注意batch_size要用BatchGD，因为numpy实现时没有用batch，用的是全量数据更新

w1, b1 = model7.layers[0].weights

5.3 tf 实现

tf.random.set_seed(1)		# 固定随机值

x = tf.Variable(train_x, dtype=tf.float32)
w2 = w
b2 = b
with tf.GradientTape(persistent=True) as tape:
	tape.watch([w2, b2])
	y_pred = 1/(1+tf.math.exp(-1*tf.matmul(x, w2)+b2))
	loss = tf.math.reduce_mean(tf.math.square(tf.subtract(y_pred, train_y)))

dw2 = tape.gradient(target=loss, sources=w2)
db2 = tape.gradient(target=loss, sources=b2)

w2 = w2 - dw2*learning_rate
b2 = b2 - db2*learning_rate

5.4 numpy实现

import numpy as np

class MyModel:
    def __init__(self, w, b, learning_rate):
        self.w = w
        self.b = b
        self.learning_rate = learning_rate
    def fit(self, train_x, train_y, epochs, batch_size):
        self.x = train_x
        self.y = train_y
        for epoch in range(epochs):
            print(f"epoch {epoch}")
            self.forward()  # 正向传播
            self.get_loss()
            self.backward()
    def forward(self):
        self.h3 = self.sigmoid(np.dot(self.x, self.w) + self.b)
    def backward(self):
        learning_rate = 0.01
        dw3 = np.dot(self.x.T, 2*(self.h3 - self.y)*self.h3*(1-self.h3)/train_x.shape[0])  # loss对w的求导
        db3 = np.dot(np.ones(shape=(1, rows)), 2*(self.h3 - self.y)*self.h3*(1-self.h3)/train_x.shape[0])  # loss对b的求导
        self.w -= dw3 * learning_rate
        self.b -= db3 * learning_rate
    def sigmoid(self, x):
        return 1 / (1 + np.exp(-x))
    def get_loss(self):
        loss = np.sum((np.square(self.h3-self.y)), axis=0)/rows
        print(f"loss {loss}")
    def predict(self):
        pass

model8 = MyModel(w, b, learning_rate)
model8.fit(train_x, train_y, epochs=1, batch_size=rows)

w3 = model8.w
b3 = model8.b