Tensorflow object detection config files documentation

Question

I am using the object detection api in tensorflow. I noticed that practically all parameters pass through the config file. I could not find any documentation or tutorial on the options for these config files though.

I know that in the official git they provide a list of config files for their pretrained models which could be very helpful but it does not cover every case and of course does not provide any explanation if needed.

For example in train_config section there are some data augmentation options which are quite self explanatory but the potential existence of other options is unclear:

  data_augmentation_options {
    random_horizontal_flip {
    }
  }
  data_augmentation_options {
    ssd_random_crop {
    }
  }

Is there a source I could refer to? For example in this tutorial two extra options (batch_queue_capacity and prefetch_queue_capacity) I did not know about appear. Where could I find a decent list of options I have? I know that it's model specific but some of them are universal and really helpful.

javidcf · Accepted Answer · 2018-03-07 11:08:23Z

28

As mentioned in the configuration documentation, configuration files are just Protocol Buffers objects described in the .proto files under research/object_detection/protos. The top level object is a TrainEvalPipelineConfig defined in pipeline.proto, and different files describe each of the elements. For example, data_augmentation_options are PreprocessingStep objects, defined in preprocessor.proto (which in turn can include a range of other possible objects for different preprocessing tasks). The meaning of each object and field may or may not be obvious or well-documented, but you can always refer to the source code to see exactly how each value is being used (for example, check preprocessor.py to understand how data augmentation is done).

answered Mar 7, 2018 at 11:08

javidcf

59.9k7 gold badges87 silver badges134 bronze badges

Sign up to request clarification or add additional context in comments.

7 Comments

user4028648 Over a year ago

So why don´t they extend them with .pbtxt instead of .config? This is just confusing, right?

javidcf Over a year ago

@EmielBoss Well in principle you should be able to use the framework without even knowing what protocol buffers are, so I suppose .config may be more meaningful for many people. However, since there is no proper documentation for the settings besides the .proto files, that "complexity hiding" does not quite work so well.

user4028648 Over a year ago

Just to check whether I have the terminology right: protocol buffers are serialized structured (according to the protocol), like .pbtxt and .pb, right? What would the .proto file itself be called? Just the "protocol"?

javidcf Over a year ago

@EmielBoss I'm not sure if there's a very specific terminology about it, but at least what you say makes sense to me fwiw (that is, matches my understanding and I think I'd understand it if I hear it).

Ashish Johnson Over a year ago

Hi, is there any performance boost if config files are used for model building in contrast to normal way of using tf keras API while building model ?

|

Collectives™ on Stack Overflow

Tensorflow object detection config files documentation

1 Answer 1

7 Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

7 Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related