简单粗暴的tensorflow-变量保存与恢复

# train.py 模型训练阶段
model = MyModel()
# 实例化Checkpoint，指定保存对象为model（如果需要保存Optimizer的参数也可加入）
checkpoint = tf.train.Checkpoint(myModel=model)
# ...（模型训练代码）
# 模型训练完毕后将参数保存到文件（也可以在模型训练过程中每隔一段时间就保存一次）
checkpoint.save('./save/model.ckpt')    #保存模型参数，save_path_with_prefix目录+前缀
 
# test.py 模型使用阶段
model = MyModel()
checkpoint = tf.train.Checkpoint(myModel=model)             # 实例化Checkpoint，指定恢复对象为model
checkpoint.restore(tf.train.latest_checkpoint('./save'))    # 从文件恢复模型参数
# 模型使用代码

import tensorflow as tf
import numpy as np
import argparse
from zh.model.mnist.mlp import MLP
from zh.model.utils import MNISTLoader
 
parser = argparse.ArgumentParser(description='Process some integers.')
parser.add_argument('--mode', default='train', help='train or test')
parser.add_argument('--num_epochs', default=1)
parser.add_argument('--batch_size', default=50)
parser.add_argument('--learning_rate', default=0.001)
args = parser.parse_args()
data_loader = MNISTLoader()
 
 
def train():
    model = MLP()
    optimizer = tf.keras.optimizers.Adam(learning_rate=args.learning_rate)
    num_batches = int(data_loader.num_train_data // args.batch_size * args.num_epochs)
    checkpoint = tf.train.Checkpoint(myAwesomeModel=model)      # 实例化Checkpoint，设置保存对象为model
    for batch_index in range(1, num_batches+1):                 
        X, y = data_loader.get_batch(args.batch_size)
        with tf.GradientTape() as tape:
            y_pred = model(X)
            loss = tf.keras.losses.sparse_categorical_crossentropy(y_true=y, y_pred=y_pred)
            loss = tf.reduce_mean(loss)
            print("batch %d: loss %f" % (batch_index, loss.numpy()))
        grads = tape.gradient(loss, model.variables)
        optimizer.apply_gradients(grads_and_vars=zip(grads, model.variables))
        if batch_index % 100 == 0:                              # 每隔100个Batch保存一次
            path = checkpoint.save('./save/model.ckpt')         # 保存模型参数到文件
            print("model saved to %s" % path)
 
 
def test():
    model_to_be_restored = MLP()
    # 实例化Checkpoint，设置恢复对象为新建立的模型model_to_be_restored
    checkpoint = tf.train.Checkpoint(myAwesomeModel=model_to_be_restored)      
    checkpoint.restore(tf.train.latest_checkpoint('./save'))    # 从文件恢复模型参数
    y_pred = np.argmax(model_to_be_restored.predict(data_loader.test_data), axis=-1)
    print("test accuracy: %f" % (sum(y_pred == data_loader.test_label) / data_loader.num_test_data))
 
 
if __name__ == '__main__':
    if args.mode == 'train':
        train()
    if args.mode == 'test':
        test()