Question

我正在尝试从上一次培训中恢复，但是我能够保存模型但无法还原。我有以下代码，其运行没有错误。我知道它并没有恢复，因为当我重新开始训练时，损失值又回到了很大的值。

有帮助吗？

ckpt_path = os.path.abspath(os.path.dirname(__file__)) + '/weights/'

labels_net, loss = vgg16(crop_size)
optimizer = tf.train.AdamOptimizer(learning_rate=0.0001).minimize(loss)
saver = tf.train.Saver(max_to_keep=3)

# Train
with tf.Session() as sess:

    # Load previous weights
    if os.listdir(ckpt_path) ==[]:
        sess.run(tf.global_variables_initializer())
    else:
        for file in os.listdir(ckpt_path):
            if 'vgg16' in file:
                try:
                    saver = tf.train.import_meta_graph(os.path.join(ckpt_path+file))
                    saver.restore(sess, ckpt_path+'vgg16-2')
                    print('Resuming training....')
                except:
                    sess.run(tf.global_variables_initializer())
            else:
                sess.run(tf.global_variables_initializer())

    print('Epoch', 'Training loss')
    for epoch_i in range(epochs):
        for batch_i in range(batches):

            batch_crops = getBatch(crops_train, batch_i, batch_size)
            batch_labels = getBatch(labels_train, batch_i, batch_size)
            x = sess.graph.get_tensor_by_name('x:0')
            y = sess.graph.get_tensor_by_name('y:0')
            sess.run(optimizer, feed_dict={x: batch_crops, y: batch_labels})#, options=run_options, run_metadata=run_metadata)

        train_loss = sess.run(loss, feed_dict={x: batch_crops, y: batch_labels})   
        print(epoch_i+1, train_loss)
        saver.save(sess, ckpt_path+'vgg16', global_step=2)

Answer 1

我对张量流了解不多，但是。我不认为您正在加载要保存的文件。

您的负载线为saver.restore(sess, ckpt_path+'vgg16-2')

因此您要保存到vgg16并从vgg16-2加载

从上次训练中恢复训练后的变量

1 个答案: