In the video under ‘Retraining with the Object detection API’, the first video under this series - ‘Loading a RetinaNet model’ seems to make a mistake at timestamp 2:03 -2:05 when talking about loading the model configuration. Laurence mentioned loading a ResNet configuration, instead of a RetinaNet model configuration. Is there a reason for this, or it is a mistake?
Also, can someone explain why we use a leading underscore for the parameters of tf.v2.compat.Checkpoint. For instance
tf.compat.v2.Checkpoint(_base_tower_layers_for_heads = detection_model._box_predictor)
Why not just base_tower_layers?
Some applications and computer scripts may not recognize spaces or will process your files differently when using spaces . A best practice is to replace spaces in file names with an underline (_) or hyphen (-)
That is how the original layers are named in the Model, and the reason they use this naming I think is the same as for file naming above.
I can not find that timestamp can you point it out with a link here…
Thank you @gent.spah , In that week assignment, I found something that might relate to the answer also.
I discovered that was how they defined the class attributes that resulted from using the model_builder.build.
Also, about the timestamp issue, I think I am cleared.
Thank you for your response…
1 Like