Question

我正在尝试在 TensorFlow 中编写一个简单的神经网络，以了解其具体细节。旨在简单地拟合非线性函数，而无需明确使用模型、层或优化器。

但我觉得训练太早饱和了，有人可以建议修改以使其表现更好吗？

import tensorflow as tf
import numpy as np

class SimpleDenseLayer(tf.keras.layers.Layer):
    def __init__(self, in_length=0, in_dim=0, out_dim=0):
        super(SimpleDenseLayer,self).__init__()
        print(in_length,in_dim,out_dim)
        self.kernel = tf.Variable(shape=(in_dim,out_dim),trainable=True,
                                initial_value=tf.random.uniform((in_dim,out_dim)))
        self.bias = tf.Variable(shape=(in_length,out_dim),trainable=True,
                                initial_value=tf.random.uniform((in_length,out_dim)))
    
    def call(self,x):
        return tf.nn.relu(tf.matmul(x,self.kernel) + self.bias)


x = tf.random.uniform((100,2))
y = x[0]**3 - 3*x[1]

layer1 = SimpleDenseLayer(100,2,80)
layer2 = SimpleDenseLayer(100,80,1)
loss = lambda y1, y2: tf.reduce_sum(tf.math.abs(y1-y2))


for i in range(50):
    with tf.GradientTape() as gt:
        nn = layer2(layer1(x))
        los = loss(nn,y)
        grad1,grad2 = gt.gradient(los,[layer1.variables, layer2.variables])
    print(i,los)
    for grad,var in zip(grad1,layer1.variables):
        var.assign_sub(grad*0.0001)
    for grad,var in zip(grad2,layer2.variables):
        var.assign_sub(grad*0.0001)

Tensorflow 简单的训练循环

0 个答案: