4 commits, mostly self explanatory. Most significant is Expand loss function flexibility:
This changes the list of command line arguments use_kl, use_mse etc to a dictionary
loss_fns = {
"similarity": True, # default is True
"l1_l2_reg": True, # default is True
...
}
and implements the necessary changes to DAMTrainer to work with this, making all components of the loss functional optional. Also, added lambda scaling coefficients to all the new loss function components (tuned roughly by my first few experiments).
4 commits, mostly self explanatory. Most significant is Expand loss function flexibility: This changes the list of command line arguments use_kl, use_mse etc to a dictionary loss_fns = { "similarity": True, # default is True "l1_l2_reg": True, # default is True ... } and implements the necessary changes to DAMTrainer to work with this, making all components of the loss functional optional. Also, added lambda scaling coefficients to all the new loss function components (tuned roughly by my first few experiments).