Initializers

ActivationInitializer

class numpy_ml.neural_nets.initializers.ActivationInitializer(param=None)

A class for initializing activation functions. Valid param values are:

1. __str__ representations of an ActivationBase instance
2. ActivationBase instance

If param is None, return the identity function: f(X) = X

init_from_str(act_str)

Initialize activation function from the param string

OptimizerInitializer

class numpy_ml.neural_nets.initializers.OptimizerInitializer(param=None)

A class for initializing optimizers. Valid param values are:

1. __str__ representations of OptimizerBase instances
2. OptimizerBase instances
3. Parameter dicts (e.g., as produced via the summary method in LayerBase instances)

If param is None, return the SGD optimizer with default parameters.

init_from_str()

Initialize optimizer from the param string

init_from_dict()

Initialize optimizer from the param dictonary

SchedulerInitializer

class numpy_ml.neural_nets.initializers.SchedulerInitializer(param=None, lr=None)

A class for initializing learning rate schedulers. Valid param values are:

1. __str__ representations of SchedulerBase instances
2. SchedulerBase instances
3. Parameter dicts (e.g., as produced via the summary method in LayerBase instances)

If param is None, return the ConstantScheduler with learning rate equal to lr.

init_from_str()

Initialize scheduler from the param string

init_from_dict()

Initialize scheduler from the param dictionary

WeightInitializer

class numpy_ml.neural_nets.initializers.WeightInitializer(act_fn_str, mode='glorot_uniform')

A factory for weight initializers.

Parameters:

• act_fn_str (str) – The string representation for the layer activation function
• mode (str (default: 'glorot_uniform')) – The weight initialization strategy. Valid entries are {“he_normal”, “he_uniform”, “glorot_normal”, glorot_uniform”, “std_normal”, “trunc_normal”}