capsa_tf.vote¶

class capsa_tf.vote.Wrapper¶

__init__(n_voters=4, alpha=3, weight_noise=0.1, use_bias=False, inline=False, finetune=False)¶

Initialize a Vote Wrapper with configs.

Parameters:

n_voters (int) – More voters will give a more diverse set of opinions and quality of uncertainty, but will also cost more memory and compute time. (default: 4)
alpha (int) – Approximate multiple votes with a shared internal representation. Smaller alpha (i.e., alpha=1) indicates more sharing between voters (faster runtime and less memory requirement). (default: 3)
weight_noise (float) – How much noise to use when initializing new weights. Suggested range [0., 0.3] (default: 0.1)
use_bias (bool) – If True - train a bias node per voter. If False - remove the bias. (default: False)
inline (bool) – Whether to inline inner functions. Defaults to inline=False which disallows inner functions. (default: False)
finetune (bool) – Setting this flag to True freezes the model’s original weights. Defaults to False (default: False)

__call__(fn)¶

Applies wrapper to tf.function decorated function.

Parameters:: fn (Callable) – The function to wrap
Return type:: WrappedFunction
Returns:: The wrapped function

Example Usage

Wrapping a tf.function¶

from capsa_tf import sample # or sculpt, vote
wrapper = sample.Wrapper(n_samples=3) # Initialize a wrapper object with your config options

def forward_pass(x):
    ...
    return y

wrapped_forward_pass = wrapper(forward_pass) # wrap your function

y = wrapped_forward_pass(x) # Use the wrapped function as usual
y, risk = wrapped_forward_pass(x, return_risk=True) # Use the wrapped function to obtain risk values

Decorator approach¶

from capsa_tf import sample # or sculpt, vote

@sample.Wrapper(n_samples=3) # Initialize a wrapper object with your config options
def forward_pass(x):
    ...
    return y

y = forward_pass(x) # Use the wrapped function as usual
y = forward_pass(x,return_risk=True) # Use the wrapped function to obtain risk values