A VarStore is used to store all the variables used by a given model. The model creates variables by calling Var_store.new_var for which it has to provide a name. Var_store.sub creates a sub-directory in the var store which is useful to group some variables together.

create ?frozen ?device ~name () creates a new variable store on the specified device (defaulting to cpu).

sub t subname returns a var-store corresponding to path subname in t.

subi is similar to sub but uses an integer for the subname.

Same as sub.

Same as subi.

num_trainable_vars t returns the number trainable variables stored in t.

iter_trainable_vars t ~f applies f to all trainable variables stored in t.

all_vars t returns all the variables stored in t.

name t returns the var-store name.

device t returns the device used to store variables hold by t.

new_var ?trainable t ~shape ~init ~name creates a new variable in t with shape shape and the specified initialization. The tensor associated with the variable is returned.

new_var_copy ?trainable t ~src ~name creates a new variable in t by copying tensor src, so using the same shape, element kind and element values.

freeze t freezes all variables in the var-store, none of the variables are trainable anymore and their gradients are not tracked.

unfreeze t unfreezes the 'trainable' variables from t.