hongym7 / early_stopping Goto Github PK

Hello,

your README is not very long. Therefore, I want to ask on this way. Is this function an implementation of the Early Stopping feature for the Tensorflow Object-Detection? Do I only have to copy this function in the ../models/research/object_detection/ folder and execute it as the provided model_main.py by the official Tensorflow OD API?

How should I use it? As far as I understood it, the function searches for an convergence in the first 100k iterations (default value), am I right? What is the criterion to abort the training process?

Kind Regards

EDIT:

If I copy the function inside the folder and execute it, I get the following error message:
TypeError: create_train_and_eval_specs() missing 2 required positional arguments: 'predict_input_fn' and 'train_steps'

Obviously, train_input_fn and train_steps are missing. Please clarify why they are missing.

Hi, I see that in your code there is this part:

early_stopping1 = tf.contrib.estimator.stop_if_no_decrease_hook(
        estimator,
        metric_name='loss_1',
        max_steps_without_decrease=5,
        min_steps=0
    )

    early_stopping2 = tf.contrib.estimator.stop_if_no_decrease_hook(
        estimator,
        metric_name='loss_2',
        max_steps_without_decrease=5,
        min_steps=0
    )

    early_stopping3 = tf.contrib.estimator.stop_if_lower_hook(
        estimator,
        metric_name='loss_1',
        threshold=10,
        eval_dir=None,
        min_steps=0,
        run_every_secs=60,

    run_every_steps=None
)

May I ask what is it used for ? Because I do not see those variables being used anywhere in the code. Thank you very much.