TensorFlow2-手写数字问题体验(3)

超、凢脫俗 2022-11-29 03:04 57阅读 0赞

import os
    import tensorflow as tf
    from tensorflow import keras
    from tensorflow.keras import layers, optimizers, datasets

# 只使用cpu
    os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
    # 设置 TensorFlow 的日志级别,避免输出过多提示信息
    os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

# 使用 keras.datasets 加载 mnist 数据集,将会返回两组数据
    (x, y), (x_val, y_val) = datasets.mnist.load_data()

# 数据类型
    print("datasets type:", type(x))

datasets type: <class 'numpy.ndarray'>

# 显示数据维度
    print("datasets: ", x.shape, y.shape, x_val.shape, y_val.shape)

datasets:  (60000, 28, 28) (60000,) (10000, 28, 28) (10000,)

# 将 Numpy 类型的数据转换为 tensor
    # keras.datasets 加载到的是 Numpy 类型的数据,将其转换为 tensor
    # convert to tensor
    # x:[0~255] => [0.0~1.0]    图片最高的像素点是255
    x = tf.convert_to_tensor(x, dtype=tf.float32) / 255.
    y = tf.convert_to_tensor(y, dtype=tf.int32)

# 数据类型
    print("datasets type:", type(x))

datasets type: <class 'tensorflow.python.framework.ops.EagerTensor'>

# 返回一个热张量
    y = tf.one_hot(y, depth=10)

print(x.shape, y.shape)

(60000, 28, 28) (60000, 10)

# 查看数据分布情况,shape,dype,min,max
    print(x.shape, y.shape, x.dtype, y.dtype)
    print(tf.reduce_min(x), tf.reduce_max(y))

(60000, 28, 28) (60000, 10) <dtype: 'float32'> <dtype: 'float32'>
    tf.Tensor(0.0, shape=(), dtype=float32) tf.Tensor(1.0, shape=(), dtype=float32)

# 第3步: 将 tensor 转换为 datasets
    # tensorflow 推荐使用 tf.data.Dataset 来加载数据集
    # tensor => datasets
    # 200个为一个batch返回datasets
    train_dataset = tf.data.Dataset.from_tensor_slices((x, y))
    # 批量训练
    train_dataset = train_dataset.batch(200)

# 模型，3层模型
    # relu: 非线性
    # out=rule(X@W+b)
    model = keras.Sequential([
        layers.Dense(512, activation='relu'),
        layers.Dense(256, activation='relu'),
        layers.Dense(10)])
    
    # 优化器
    # w1 = w - lr*(d_loss/d_w)
    optimizer = optimizers.SGD(learning_rate=0.001)

def train_epoch(epoch):
        # Step4.loop
        for step, (x, y) in enumerate(train_dataset):
    
            with tf.GradientTape() as tape:
                # 转为二维数组
                # [b, 28, 28] => [b, 784]
                # print(x.shape) # 3维
                x = tf.reshape(x, (-1, 28 * 28))
                # print(x.shape) # 2维
                # 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)
            # w' = w - lr * grad
            optimizer.apply_gradients(zip(grads, model.trainable_variables))
    
            if step % 100 == 0:
                print(epoch, step, 'loss:', loss.numpy())

def train():
        for epoch in range(30):
            train_epoch(epoch)

train()

0 0 loss: 2.3555892
    0 100 loss: 0.9576269
    0 200 loss: 0.83611715
    1 0 loss: 0.68340844
    1 100 loss: 0.7010591
    1 200 loss: 0.6275456
    2 0 loss: 0.56274295
    2 100 loss: 0.61272794
    2 200 loss: 0.54326683
    3 0 loss: 0.5059351
    3 100 loss: 0.5606175
    3 200 loss: 0.49353546
    4 0 loss: 0.46889228
    4 100 loss: 0.5237827
    4 200 loss: 0.4590548
    5 0 loss: 0.44137704
    5 100 loss: 0.49544248
    5 200 loss: 0.43307236
    6 0 loss: 0.41934803
    6 100 loss: 0.47261453
    6 200 loss: 0.41206986
    7 0 loss: 0.40094435
    7 100 loss: 0.45364487
    7 200 loss: 0.39463615
    8 0 loss: 0.3851467
    8 100 loss: 0.4374189
    8 200 loss: 0.37994912
    9 0 loss: 0.37151062
    9 100 loss: 0.42325506
    9 200 loss: 0.36702237
    10 0 loss: 0.35947707
    10 100 loss: 0.4105991
    10 200 loss: 0.35564843
    11 0 loss: 0.3486898
    11 100 loss: 0.39929065
    11 200 loss: 0.3454248
    12 0 loss: 0.33890578
    12 100 loss: 0.38912597
    12 200 loss: 0.33617145
    13 0 loss: 0.32996795
    13 100 loss: 0.3799546
    13 200 loss: 0.3277427
    14 0 loss: 0.32176715
    14 100 loss: 0.37157205
    14 200 loss: 0.31990257
    15 0 loss: 0.3142795
    15 100 loss: 0.36378965
    15 200 loss: 0.3127023
    16 0 loss: 0.30731258
    16 100 loss: 0.35668594
    16 200 loss: 0.3060998
    17 0 loss: 0.30089006
    17 100 loss: 0.35007447
    17 200 loss: 0.2999802
    18 0 loss: 0.29485452
    18 100 loss: 0.34396407
    18 200 loss: 0.2941921
    19 0 loss: 0.28922617
    19 100 loss: 0.33832982
    19 200 loss: 0.28871384
    20 0 loss: 0.2839748
    20 100 loss: 0.33302107
    20 200 loss: 0.28360277
    21 0 loss: 0.2790264
    21 100 loss: 0.32806015
    21 200 loss: 0.2788163
    22 0 loss: 0.27433667
    22 100 loss: 0.32340646
    22 200 loss: 0.2743058
    23 0 loss: 0.26991957
    23 100 loss: 0.31906232
    23 200 loss: 0.27008367
    24 0 loss: 0.26571178
    24 100 loss: 0.31494868
    24 200 loss: 0.2660941
    25 0 loss: 0.2617136
    25 100 loss: 0.3110497
    25 200 loss: 0.2623037
    26 0 loss: 0.257893
    26 100 loss: 0.30739006
    26 200 loss: 0.25867736
    27 0 loss: 0.25424826
    27 100 loss: 0.30390614
    27 200 loss: 0.25520083
    28 0 loss: 0.25077572
    28 100 loss: 0.30059946
    28 200 loss: 0.25189033
    29 0 loss: 0.24750763
    29 100 loss: 0.29749864
    29 200 loss: 0.24871086