Is there some way to save best model only with tensorflow.estimator.train_and_evaluate()?

Question

I try retrain TF Object Detection API model from checkpoint with already .config file for training pipeline with tf.estimator.train_and_evaluate() method like in models/research/object_detection/model_main.py. And it saves checkpoints every N steps or every N seconds.

But I want to save only one best model like in Keras. Is there some way to do it with TF Object Detection API model? Maybe some options/callbacks for tf.Estimator.train or some way to use Detection API with Keras?

Tumer

I have to say that both solutions from Sharky and from prouast are correct. I tried up both of them and it works fine. So if somebody have the same issue as I had, you can use any of this options. Thanks a lot to Sharky and prouast for clear and useful answers! – Tumer

Thanks for this - but where do I have to put that code from thse answers? I am using the default model_main_tf2.py script from the Object Detection API for training. Do I need to modify the OD scripts themselves? – Circumstance

Answer 1

You can try using BestExporter. As far as I know, it's the only option for what you're trying to do.

exporter = tf.estimator.BestExporter(
      compare_fn=_loss_smaller,
      exports_to_keep=5)

eval_spec = tf.estimator.EvalSpec(
    input_fn,
    steps,
    exporters)

https://www.tensorflow.org/api_docs/python/tf/estimator/BestExporter

Forrer

I see that the exporter exports files in a slightly different format saved_model.pb.. and the meta file (example: model.ckpt-3073.meta) is missing... are they equivalent to use? I remember *.pb is generated after exporting frozen inference graph. – Bicarbonate

BestExporter also saves ckpt, despite that it is not copied to a new folder, which is doomed to be cleared if new ckpts are generated (say keep_checkpoint_max=5 in tf.estimator.RunConfig), where @Ponderable 's answer will fit. – Ringster

Answer 2

I have been using https://github.com/bluecamel/best_checkpoint_copier which works well for me.

Example:

best_copier = BestCheckpointCopier(
   name='best', # directory within model directory to copy checkpoints to
   checkpoints_to_keep=10, # number of checkpoints to keep
   score_metric='metrics/total_loss', # metric to use to determine "best"
   compare_fn=lambda x,y: x.score < y.score, # comparison function used to determine "best" checkpoint (x is the current checkpoint; y is the previously copied checkpoint with the highest/worst score)
   sort_key_fn=lambda x: x.score,
   sort_reverse=False) # sort order when discarding excess checkpoints

pass it to your eval_spec:

eval_spec = tf.estimator.EvalSpec(
   ...
   exporters=best_copier,
   ...)

Ponderable

Thanks for this - but where do I have to put that code? I am using the default model_main_tf2.py script from the Object Detection API for training. Do I need to modify the OD scripts themselves? – Circumstance

Answer 3

If you are training using the models repo of tensorflow/models. models/research/object_detection/model_lib.py file create_train_and_eval_specs function can be modified to include the best exporter:

final_exporter = tf.estimator.FinalExporter(
    name=final_exporter_name, serving_input_receiver_fn=predict_input_fn)

best_exporter = tf.estimator.BestExporter(
    name="best_exporter",
    serving_input_receiver_fn=predict_input_fn,
    event_file_pattern='eval_eval/*.tfevents.*',
    exports_to_keep=5)
exporters = [final_exporter, best_exporter]

train_spec = tf.estimator.TrainSpec(
    input_fn=train_input_fn, max_steps=train_steps)

eval_specs = [
    tf.estimator.EvalSpec(
        name=eval_spec_name,
        input_fn=eval_input_fn,
        steps=eval_steps,
        exporters=exporters)
]

Bicarbonate