如何从 tf.estimator.Estimator 中获取最后一个 global_step

Question

如何在 train(...) 完成后从 tf.estimator.Estimator 中获取最后一个 global_step？例如，一个典型的基于 Estimator 的训练程序可能是这样设置的： n_epochs = 10 model_dir = '/path/to/model_dir'

def model_fn(features, labels, mode, params):
    # some code to build the model
    pass

def input_fn():
    ds = tf.data.Dataset()  # obviously with specifying a data source
    # manipulate the dataset
    return ds

run_config = tf.estimator.RunConfig(model_dir=model_dir)
estimator = tf.estimator.Estimator(model_fn=model_fn, config=run_config)

for epoch in range(n_epochs):
    estimator.train(input_fn=input_fn)
    # Now I want to do something which requires to know the last global step, how to get it?
    my_custom_eval_method(global_step)

只有 evaluate() 方法 returns 包含 global_step 作为字段的字典。如果出于某种原因我不能或不想使用此方法，我如何获得 global_step？

Answer 1

只需在训练循环之前创建一个钩子：

class GlobalStepHook(tf.train.SessionRunHook):
    def __init__(self):
        self._global_step_tensor = None
        self.value = None

    def begin(self):
        self._global_step_tensor = tf.train.get_global_step()

    def after_run(self, run_context, run_values):
        self.value = run_context.session.run(self._global_step_tensor)

    def __str__(self):
        return str(self.value)

global_step = GlobalStepHook()
for epoch in range(n_epochs):
    estimator.train(input_fn=input_fn, hooks=[global_step])
    # Now the global_step hook contains the latest value of global_step
    my_custom_eval_method(global_step.value)

Answer 2

最近，我发现估算器有 api get_variable_value

global_step = estimator.get_variable_value("global_step")

如何从 tf.estimator.Estimator 中获取最后一个 global_step

How to get the last global_step from an tf.estimator.Estimator

python

tensorflow

tensorflow-estimator