tensorflow2.0 最詳細的模型訓練步驟 直接上代碼

import  os
import  tensorflow as tf
from    tensorflow import keras
from    tensorflow.keras import layers, optimizers, datasets
import numpy as np
import matplotlib.pyplot as plt
import datetime

def prepare_mnist_features_and_labels(x, y):
  x = tf.cast(x, tf.float32) / 255.0
  y = tf.cast(y, tf.int64)

  return x, y

def mnist_dataset():
  (x, y), (x_val, y_val) = datasets.fashion_mnist.load_data()
  x, x_val = np.expand_dims(x, axis=3), np.expand_dims(x_val, axis=3)

  y = tf.one_hot(y, depth=10)
  y_val = tf.one_hot(y_val, depth=10)
  ds = tf.data.Dataset.from_tensor_slices((x, y))
  ds = ds.map(prepare_mnist_features_and_labels)
  ds = ds.shuffle(60000).batch(100)


  ds_val = tf.data.Dataset.from_tensor_slices((x_val, y_val))
  ds_val = ds_val.map(prepare_mnist_features_and_labels)
  ds_val = ds_val.shuffle(10000).batch(100)
  return ds,ds_val

class LearningRateExponentialDecay:
    def __init__(self,initial_learning_rate,decay_epochs,decay_rate):
        self.initial_learning_rate=initial_learning_rate
        self.decay_epochs=decay_epochs
        self.decay_rate=decay_rate
    def __call__(self,epoch):
        dtype =type(self.initial_learning_rate)
        decay_epochs=np.array(self.decay_epochs).astype(dtype)
        decay_rate=np.array(self.decay_rate).astype(dtype)
        epoch = np.array(epoch).astype(dtype)
        p = epoch/decay_epochs
        lr = self.initial_learning_rate*np.power(decay_rate,p)
        return lr

def save_log():
    log_dir = os.path.join( "C:/Users/byroot/Desktop/test1/model_output", 'logs_{}'.format(datetime.datetime.now().strftime("%Y%m%d-%H%M%S")))
    if not os.path.exists(log_dir):
        os.makedirs(log_dir)
    return log_dir

def callbackfunc():
    # 模型儲存格式預設是saved_model,可以自己定義更改原有類來儲存hdf5
    ckpt = tf.keras.callbacks.ModelCheckpoint(filepath=path, monitor='val_loss',
                                              save_best_only=False, save_weights_only=False)

    # 當模型訓練不符合我們要求時停止訓練，連續5個epoch驗證集精度沒有提高0.001%停
    earlystop = tf.keras.callbacks.EarlyStopping(monitor='val_accuracy', min_delta=0.0001, patience=52)

    lr_schedule = LearningRateExponentialDecay(initial_learning_rate=0.001, decay_epochs=1, decay_rate=0.96)
    lr = tf.keras.callbacks.LearningRateScheduler(lr_schedule)

    # 定義當loss出現nan或inf時停止訓練的callback
    terminate = tf.keras.callbacks.TerminateOnNaN()

    # 降低學習率（要比學習率自動周期變化有更大變化和更長時間監控）
    reduce_lr = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=3,
                                                     min_delta=0.0001, min_lr=0)

    # 儲存訓練過程中大數标量名額，與tensorboard同一個檔案
    csv_logger = tf.keras.callbacks.CSVLogger(os.path.join(log_dir, 'logs.log'), separator=',')

    # 還要加入tensorboard的使用,這種方法記錄的内容有限
    tensorboard = tf.keras.callbacks.TensorBoard(log_dir=log_dir,
                                                 histogram_freq=1,  # 對參數和激活做直方圖，一定要有測試集
                                                 write_graph=True,  # 模型結構圖
                                                 write_images=True,  # 把模型參數做為圖檔形式存到
                                                 update_freq='epoch',  # epoch,batch,整數，太頻的話會減慢速度
                                                 profile_batch=2,  # 記錄模型性能
                                                 embeddings_freq=1,
                                                 embeddings_metadata=None  # 這個還不太清楚如何使用
                                                 )

    callback = [ckpt, earlystop, lr, tensorboard, terminate, reduce_lr, csv_logger]
    return callback

def saveModel():
    model_json = model.to_json()
    with open(os.path.join(log_dir, 'model_json.json'), 'w') as json_file:
        json_file.write(model_json)

if __name__ == '__main__':
    path = "C:/Users/byroot/Desktop/test1/" + "ckpt_epoch{epoch:02d}_val_acc{val_loss:.2f}.hdf5"

    """1.參數配置"""
    log_dir = save_log()
    os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'  # or any {'0', '1', '2'}

    """2.資料處理"""
    train_dataset, test_dataset = mnist_dataset()
    """3.設定模型"""
    model = keras.Sequential([
        layers.Reshape(target_shape=(28 * 28,), input_shape=(28, 28, 1 )),
        layers.Dense(200, activation='relu'),
        layers.Dense(200, activation='relu'),
        layers.Dense(200, activation='relu'),
        layers.Dense(10)])
    # no need to use compile if you have no loss/optimizer/metrics involved here.
    model.compile(optimizer=optimizers.Adam(0.001),
                  loss=tf.losses.CategoricalCrossentropy(from_logits=True),
                  metrics=['accuracy'])


    """4.畫出模型結構圖并儲存"""
    tf.keras.utils.plot_model(model,to_file=os.path.join('C:/Users/byroot/Desktop/test1/','model.png'),show_shapes=True,show_layer_names=True)
    """5.配置回調函數"""
    callback = callbackfunc()
    """6.開始訓練"""
    model.fit(train_dataset.repeat(), epochs=1, steps_per_epoch=10,
              validation_data=test_dataset.repeat(),
              validation_steps=2,
              callbacks = callback
              )
    """7.列印模型結構"""
    model.summary()
    """8.儲存模型結構及配置參數"""
    saveModel()
    """9.對模型在測試集上進行評估"""
    metrics = model.evaluate(test_dataset)
    print("val_loss:", metrics[0], "val_accuracy:", metrics[1])