以梯度下降法为例
out=relu{relu{relu[X@W1+b1]@W2+b2}@W3+b3}
pred=argmax(out)
loss=MSE(out,label)
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
#x: [60k, 28, 28]
#y: [60k]
(x, y), (x_val, y_val) = datasets.mnist.load_data()
x = tf.convert_to_tensor(x, dtype = tf.float32) / 255.
#[1-255] > [0-1]
y = tf.convert_to_tensor(y, dtype = tf.int32)
train_db = tf.data.Dataset.from_tensor_slices((x,y)).batch(128)
train_iter = iter(train_db)
sample = next(train_iter)
#[b, 784] > [b, 256] > [b, 128] > [b,10]
#[dim_in, dim_out]
w1, b1 = tf.Variable(tf.random.truncated_normal([784, 256],stddev = 0.1)), tf.Variable(tf.zeros([256]))
w2, b2 = tf.Variable(tf.random.truncated_normal([256, 128],stddev = 0.1)), tf.Variable(tf.zeros([128]))
w3, b3 = tf.Variable(tf.random.truncated_normal([ 128, 10],stddev = 0.1)), tf.Variable(tf.zeros([10]))
#h1 =x@w1 +b1
lr = 1e-3
for epoch in range(10): #iteratedb for 10
for step, (x, y) in enumerate(train_db):
#x:[128,28, 28]
#y:[128]
x =tf.reshape(x, [-1, 28*28])
with tf.GradientTape() as tape:
h1 = x @ w1 + tf.broadcast_to(b1, [x.shape[0], 256])
h1 = tf.nn.relu(h1)
#[128, 784]@[784, 256] + [256] > [128,256] + [128, 256]
h2 = h1 @ w2 + tf.broadcast_to(b2, [x.shape[0], 128])
h2 = tf.nn.relu(h2)
#[128, 256] @ [256, 128] + [128] > [128, 128] + [128, 2128]
out = h2 @ w3 + b3
#[128, 128] @ [128,10] + [10] > [128, 10] + [128, 210]
#compute loss
#out :[128, 10]
y_onehot = tf.one_hot(y, depth = 10)
#y: [128, 10]
loss = tf.square(y_onehot -out)
#mean:scalar
loss = tf.reduce_mean(loss)
#compute gradients
grads = tape.gradient(loss, [w1, b1, w2, b2, w3, b3])
w1.assign_sub(lr * grads[0])
b1.assign_sub(lr * grads[1])
w2.assign_sub(lr * grads[2])
b2.assign_sub(lr * grads[3])
w3.assign_sub(lr * grads[4])
b3.assign_sub(lr * grads[5])
if step % 100 == 0:
print(epoch, step, 'loss:', float(loss))