Model parser in helpers.py has been fixed. It now takes the model-specific hyperparameters and feeds them together with the model architecture into the instatiated model class.
Relating to above bullet point, the .yaml configs have been updated for beta-vae with a category, which specifies the model_hyperparams. This approach was taken, as dictionaries are not easily subsettable.
The adidas specific data loading code snippets were commented out, such that mnist works again
The replearner() class was gutted, such that the functions _sample and _embedding can be found in the respective VAE model class.
Implemented CatVAE: CatVAE works. Annealing of temperature does not work yet, as the function requires an inherited attributed
batch_idx
. Currently handled with try/except: https://github.com/MoritzWag/Representation-Learning/blob/11bef10f8c66eadc6052a3c3186f669feb8f73a8/library/library/models2/cat_vae.py#L141-L150 Furthermore the traversals in experiment.py have been commented out, as the traversals do not work with the current implementation. As a result, the_embedding()
function in line 119 has not been updated yet.Model parser in helpers.py has been fixed. It now takes the model-specific hyperparameters and feeds them together with the model architecture into the instatiated model class.
Relating to above bullet point, the .yaml configs have been updated for beta-vae with a category, which specifies the
model_hyperparams
. This approach was taken, as dictionaries are not easily subsettable.The adidas specific data loading code snippets were commented out, such that mnist works again
The
replearner()
class was gutted, such that the functions_sample
and_embedding
can be found in the respective VAE model class.TO-DO's: Fix annealing and traversals in CatVAE