Tensorflow使用MirroredStrategy（）恢复培训

Question

我在Linux操作系统上训练了我的模型，这样我就可以使用MirroredStrategy()并在2个GPU上训练。训练在公元610年停止。我想恢复训练，但当我加载模型并对其进行评估时，内核就死了。我用的是Jupyter笔记本。如果我减少我的训练数据集，代码将运行，但它将只在1个GPU上运行。我的分发策略是否保存在正在加载的模型中，还是必须再次包含它

更新

我试图包括MirroredStrategy()：

mirrored_strategy = tf.distribute.MirroredStrategy()
with mirrored_strategy.scope():

    new_model = load_model('\\models\\model_0610.h5', 
                custom_objects = {'dice_coef_loss': dice_coef_loss, 
                'dice_coef': dice_coef}, compile = True)
    new_model.evaluate(train_x,  train_y, batch_size = 2,verbose=1)

新错误

包含MirroredStrategy()时出错：

ValueError: 'handle' is not available outside the replica context or a 'tf.distribute.Stragety.update()' call.

源代码：

smooth = 1
def dice_coef(y_true, y_pred):
    y_true_f = K.flatten(y_true)
    y_pred_f = K.flatten(y_pred)
    intersection = K.sum(y_true_f * y_pred_f)
    return (2. * intersection + smooth) / (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)

def dice_coef_loss(y_true, y_pred):
    return (1. - dice_coef(y_true, y_pred))

new_model = load_model('\\models\\model_0610.h5', 
                       custom_objects = {'dice_coef_loss': dice_coef_loss, 'dice_coef': dice_coef}, compile = True)
new_model.evaluate(train_x,  train_y, batch_size = 2,verbose=1)

observe_var = 'dice_coef'
strategy = 'max' # greater dice_coef is better
model_resume_dir = '//models_resume//'

model_checkpoint = ModelCheckpoint(model_resume_dir + 'resume_{epoch:04}.h5', 
                                   monitor=observe_var, mode='auto', save_weights_only=False, 
                                   save_best_only=False, period = 2)

new_model.fit(train_x, train_y, batch_size = 2, epochs = 5000, verbose=1, shuffle = True, 
              validation_split = .15, callbacks = [model_checkpoint])

new_model.save(model_resume_dir + 'final_resume.h5')

Answer 1

加载模型时的new_model.evaluate()和compile = True是导致问题的原因。我设置了compile = False，并从原始脚本中添加了一个编译行

mirrored_strategy = tf.distribute.MirroredStrategy()
with mirrored_strategy.scope():

    new_model = load_model('\\models\\model_0610.h5', 
                custom_objects = {'dice_coef_loss': dice_coef_loss, 
                'dice_coef': dice_coef}, compile = False)
    new_model.compile(optimizer = Adam(learning_rate = 1e-4, loss = dice_coef_loss,
                metrics = [dice_coef])