import os os.environ['TF_CPP_MIN_LOG_LEVEL']='2' #使tensorflow少打印一些不必要的信息 import tensorflow.compat.v1 as tf from tensorflow import keras from tensorflow.keras import layers, optimizers, datasets tf.enable_eager_execution() #保证sess.run()能够正常运行 #数据集加载 (x, y), (x_val, y_val) = datasets.mnist.load_data() x = tf.convert_to_tensor(x, dtype=tf.float32) / 255. y = tf.convert_to_tensor(y, dtype=tf.int32) y = tf.one_hot(y, depth=10) print(x.shape, y.shape) train_dataset = tf.data.Dataset.from_tensor_slices((x, y)) train_dataset = train_dataset.batch(200) #batch为200表示一次加载200张的图片 #降维 Dense是全连接 model = keras.Sequential([ layers.Dense(512, activation='relu'), #relu是非线性参数 layers.Dense(256, activation='relu'), layers.Dense(10)]) optimizer = optimizers.SGD(learning_rate=0.001) def train_epoch(epoch): # Step4.loop for step, (x, y) in enumerate(train_dataset): #循环300次 60kb/200等于大概300次 with tf.GradientTape() as tape: # [b, 28, 28] => [b, 784] x = tf.reshape(x, (-1, 28*28)) # Step1. compute output # [b, 784] => [b, 10] out = model(x) # Step2. compute loss loss = tf.reduce_sum(tf.square(out - y)) / x.shape[0] # Step3. optimize and update w1, w2, w3, b1, b2, b3 grads = tape.gradient(loss, model.trainable_variables) #grads里包含了对w1，w2，w3和b1,b2,b3的loss对其的求导 # w' = w - lr * grad optimizer.apply_gradients(zip(grads, model.trainable_variables)) if step % 100 == 0: print(epoch, step, 'loss:', loss.numpy()) def train(): #对整个数据集迭代次30次 for epoch in range(30): train_epoch(epoch) if __name__ == '__main__': train()
结果：
(60000, 28, 28) (60000, 10) 0 0 loss: 2.1289964 0 100 loss: 0.96601397 0 200 loss: 0.8044617 1 0 loss: 0.65632385 1 100 loss: 0.71072084 1 200 loss: 0.6174767 2 0 loss: 0.53884405 2 100 loss: 0.61792874 2 200 loss: 0.53729916 3 0 loss: 0.48332796 3 100 loss: 0.5644321 3 200 loss: 0.48922828 4 0 loss: 0.44779533 4 100 loss: 0.5270611 4 200 loss: 0.45555627 5 0 loss: 0.42214122 5 100 loss: 0.49914017 5 200 loss: 0.42974195 6 0 loss: 0.4022831 6 100 loss: 0.4767412 6 200 loss: 0.4090542 7 0 loss: 0.38604406 7 100 loss: 0.45791557 7 200 loss: 0.39167565 8 0 loss: 0.3723324 8 100 loss: 0.44173408 8 200 loss: 0.37691337 9 0 loss: 0.360519 9 100 loss: 0.42779246 9 200 loss: 0.36422646 10 0 loss: 0.35006583 10 100 loss: 0.41538823 10 200 loss: 0.3530626 11 0 loss: 0.3407312 11 100 loss: 0.40423894 11 200 loss: 0.34306836 12 0 loss: 0.3323893 12 100 loss: 0.3939416 12 200 loss: 0.3339965 13 0 loss: 0.3248109 13 100 loss: 0.38446128 13 200 loss: 0.32582656 14 0 loss: 0.31788555 14 100 loss: 0.37571213 14 200 loss: 0.3183561 15 0 loss: 0.3113761 15 100 loss: 0.3676333 15 200 loss: 0.31151268 16 0 loss: 0.30531833 16 100 loss: 0.36009517 16 200 loss: 0.30516908 17 0 loss: 0.2996593 17 100 loss: 0.35302532 17 200 loss: 0.29931957 18 0 loss: 0.29437816 18 100 loss: 0.34642395 18 200 loss: 0.2938386 19 0 loss: 0.2894483 19 100 loss: 0.34028184 19 200 loss: 0.2887537 20 0 loss: 0.28483075 20 100 loss: 0.3345565 20 200 loss: 0.28399432 21 0 loss: 0.2804789 21 100 loss: 0.3291541 21 200 loss: 0.27953643 22 0 loss: 0.27633134 22 100 loss: 0.32407936 22 200 loss: 0.27533495 23 0 loss: 0.27240857 23 100 loss: 0.3192857 23 200 loss: 0.27136424 24 0 loss: 0.26872116 24 100 loss: 0.31474534 24 200 loss: 0.26758516 25 0 loss: 0.2652039 25 100 loss: 0.31041327 25 200 loss: 0.26399314 26 0 loss: 0.26185223 26 100 loss: 0.30627567 26 200 loss: 0.2605623 27 0 loss: 0.25865546 27 100 loss: 0.3023752 27 200 loss: 0.25727862 28 0 loss: 0.2556298 28 100 loss: 0.29863724 28 200 loss: 0.25413704 29 0 loss: 0.25273502 29 100 loss: 0.29504693 29 200 loss: 0.2511155