了解 Tensorflow 对象检测 API，检查点的 kwargs class，什么是`_base_tower_layers_for_heads`？

Question

Currently, I've been learning how to use Object-Detection API from Tensorflow. I follow a quick start tutorial for training with custom data with this notebook as suggested by them. In the effort to understanding each line of the code, I stumbled upon this snippet code in the "Create Model and Restore Weight" part.

fake_box_predictor = tf.compat.v2.train.Checkpoint(
    _base_tower_layers_for_heads=detection_model._box_predictor._base_tower_layers_for_heads,
    # _prediction_heads=detection_model._box_predictor._prediction_heads,
    #    (i.e., the classification head that we *will not* restore)
    _box_prediction_head=detection_model._box_predictor._box_prediction_head,
    )

I don't really understand what are the keyword arguments that are available for the Checkpoint class in that particular snippet code. My question is; is there any documentation out there that shows the list of the keyword arguments? or at least explain what are _base_tower_layers_for_heads and _box_prediction_head ?

I've read the tf.train.Checkpoint documentation . It says that we can provide models or optimizers for the constructor's keyword argument. I am already familiar with this class to restore the weights to my model, however, I find it is alien to see _base_tower_layers_for_heads or _box_prediction_head for the keyword argument.

I do know about 'heads' and different types of 'heads' in the object detection architecture and their relation to transfer learning, what I don't understand is in the context of their data structure. How do I know, these keyword arguments exist? and is there any other else? I would really appreciate it if somebody could give me insights or at least tell me where can I find documentation that I can read to understand it more.

Answer 1

I found pretty useful information about Checkpoint class. It does not come from the documentation but from the tutorial, Training Checkpoints > Loading Mechanism .

This is what I understand so far:

1. The keyword arguments are basically the name of the attributes that are saved in our checkpoint file.
2. The variable that is passed to that particular keyword argument is the variable whose value we want to replace with checkpoint values, in other words, our model attributes.
3. The checkpoint architecture must match our model architecture in order to make this variable restoration works.

If we want to know what are the attributes that are saved in our checkpoint file, we can run this command

tf.train.list_variables(tf.train.latest_checkpoint('path_to_checkpoint'))