In the original paper there are some clarifying statements:
The four inputs to a unit in S2 are added, then multiplied by a trainable coefficient, and added to a trainable bias. The result is passed through a sigmoidal function. (p.7, col.1)
Here, sigmoidal function is generic.
As in classical neural networks, units in layers up to F6 compute.. This ...
Yes, it is totally possible.
Any suggestions or ideas?
You will need to train your own model because you might not find any pre-trained models for the same. (but do check if they are available)
Pick any state of the art model and I will suggest choosing the pose estimation models.
Collect images and annotate your dataset in the appropriate format using ...
In your code by default training=False, set training=True it will work right.
Also, since you are already subclassing keras.Model you don't need to again call keras.Model. Remove the model() method, pass the input directly to call() method by setting training=True.
I agree the metrics between your test set and validation set are quite close, but looking at your code it seems you may have run for the full 100 epochs.
keras supports early stopping, i.e. when scores fail to improve meaningfully you can have the model revert to the best scores it has seen to date:
You should ...