MDN

Public Member Functions
def	__init__ (self, n_mix=5, hidden=[100] *5, lr=1e-3, l2=1e-3, n_iter=1e4, batch=128, imputations=5, epsilon=1e-3, activation='relu', scalerx=None, scalery=None, model_path='Weights', model_name='MDN', no_load=False, no_save=False, seed=None, verbose=False, debug=False, **kwargs)
def	predict (self, X, chunk_size=1e5, return_coefs=False, **kwargs)
def	extract_predictions (self, coefs, confidence_interval=None, threshold=None, avg_est=False)
def	fit (self, X, Y, output_slices=None, **kwargs)
def	build (self)
def	loss (self, y, output)
def	__call__ (self, inputs)
def	get_config (self)
def	set_config (self, config, *args, **kwargs)
def	update_config (self, config, keys=None)
def	save (self)
def	load (self)
def	get_coefs (self, output)

Public Attributes
	model
	config

Static Public Attributes
string	distribution = 'MultivariateNormalTriL'

Detailed Description

Mixture Density Network which handles multi-output, full (symmetric) covariance.

Parameters
----------
n_mix : int, optional (default=5)
    Number of mixtures used in the gaussian mixture model.

hidden : list, optional (default=[100, 100, 100, 100, 100])
    Number of layers and hidden units per layer in the neural network.

lr : float, optional (default=1e-3)
    Learning rate for the model.

l2 : float, optional (default=1e-3)
    L2 regularization scale for the model weights.

n_iter : int, optional (default=1e4)
    Number of iterations to train the model for 

batch : int, optional (default=128)
    Size of the minibatches for stochastic optimization.

imputations : int, optional (default=5)
    Number of samples used in multiple imputation when handling NaN
    target values during training. More samples results in a higher
    accuracy for the likelihood estimate, but takes longer and may
    result in overfitting. Assumption is that any missing data is 
    MAR / MCAR, in order to allow a multiple imputation approach.

epsilon : float, optional (default=1e-3)
    Normalization constant added to diagonal of the covariance matrix.

activation : str, optional (default=relu)
    Activation function applied to hidden layers.

scalerx : transformer, optional (default=IdentityTransformer)
    Transformer which has fit, transform, and inverse_transform methods
    (i.e. follows the format of sklearn transformers). Scales the x 
    values prior to training / prediction. Stored along with the saved
    model in order to have consistent inputs to the model.

scalery : transformer, optional (default=IdentityTransformer)
    Transformer which has fit, transform, and inverse_transform methods
    (i.e. follows the format of sklearn transformers). Scales the y 
    values prior to training, and the output values after prediction. 
    Stored along with the saved model in order to have consistent 
    outputs from the model.

model_path : pathlib.Path, optional (default=./Weights)
    Folder location to store saved models.

model_name : str, optional (default=MDN)
    Name to assign to the model. 

no_load : bool, optional (default=False)
    If true, train a new model rather than loading a previously 
    trained one.

no_save : bool, optional (default=False)
    If true, do not save the model when training is completed.

seed : int, optional (default=None)
    Random seed. If set, ensure consistent output.

verbose : bool, optional (default=False)
    If true, print various information while loading / training.

debug : bool, optional (default=False)
    If true, use control flow dependencies to determine where NaN
    values are entering the model. Model runs slower with this 
    parameter set to true.

Definition at line 21 of file MDN.py.

Constructor & Destructor Documentation

__init__()

def __init__	(		self,
			n_mix = 5,
			hidden = [100]*5,
			lr = 1e-3,
			l2 = 1e-3,
			n_iter = 1e4,
			batch = 128,
			imputations = 5,
			epsilon = 1e-3,
			activation = 'relu',
			scalerx = None,
			scalery = None,
			model_path = 'Weights',
			model_name = 'MDN',
			no_load = False,
			no_save = False,
			seed = None,
			verbose = False,
			debug = False,
		**	kwargs
	)

Definition at line 97 of file MDN.py.

Member Function Documentation

__call__()

def __call__	(		self,
			inputs
	)

Definition at line 351 of file MDN.py.

build()

def build

(

self

)

Definition at line 294 of file MDN.py.

extract_predictions()

def extract_predictions	(		self,
			coefs,
			confidence_interval = None,
			threshold = None,
			avg_est = False
	)

Function used to extract model predictions from the given set of 
coefficients. Users should call the predict() method instead, if
predictions from input data are needed. 

confidence_interval : float, optional (default=None)
    If a confidence interval value is given, then this function
    returns (along with the predictions) the upper and lower 
    {confidence_interval*100}% confidence bounds around the prediction.

threshold : float, optional (default=None)
    If set, the model outputs the maximum prior estimate when the prior
    probability is above this threshold; and outputs the average estimate
    when below the threshold. Any passed value should be in the range (0, 1],
    though the sign of the threshold can be negative in order to switch the
    estimates (i.e. negative threshold would output average estimate when prior
    is greater than the (absolute) value).  

avg_est : bool, optional (default=False)
    If true, model outputs the prior probability weighted mean as the
    estimate. Otherwise, model outputs the maximum prior estimate.

Definition at line 183 of file MDN.py.

fit()

def fit	(		self,
			X,
			Y,
			output_slices = None,
		**	kwargs
	)

Definition at line 223 of file MDN.py.

get_coefs()

def get_coefs	(		self,
			output
	)

Definition at line 386 of file MDN.py.

get_config()

def get_config

(

self

)

Definition at line 355 of file MDN.py.

load()

def load

(

self

)

Definition at line 379 of file MDN.py.

loss()

def loss	(		self,
			y,
			output
	)

Definition at line 331 of file MDN.py.

predict()

def predict	(		self,
			X,
			chunk_size = 1e5,
			return_coefs = False,
		**	kwargs
	)

Top level interface to get predictions for a given dataset, which wraps _predict_chunk 
to generate estimates in smaller chunks. See the docstring of extract_predictions() for 
a description of other keyword parameters that can be given. 
    
chunk_size : int, optional (default=1e5)
    Controls the size of chunks which are estimated by the model. If None is passed,
    chunking is not used and the model is given all of the X dataset at once. 

return_coefs : bool, optional (default=False)
    If True, return the estimated coefficients (prior, mu, sigma) along with the 
    other requested outputs. Note that rescaling the coefficients using scalerx/y
    is left up to the user, as calculations involving sigma must be performed in 
    the basis learned by the model.

Definition at line 150 of file MDN.py.

save()

def save

(

self

)

Definition at line 373 of file MDN.py.

set_config()

def set_config	(		self,
			config,
		*	args,
		**	kwargs
	)

Definition at line 359 of file MDN.py.

update_config()

def update_config	(		self,
			config,
			keys = None
	)

Definition at line 364 of file MDN.py.

Member Data Documentation

config

config

Definition at line 360 of file MDN.py.

distribution

string distribution = 'MultivariateNormalTriL'

static

Definition at line 95 of file MDN.py.

model

model

Definition at line 326 of file MDN.py.

---

The documentation for this class was generated from the following file:

• /gfs-oceanweb/web/ocssw/ocssw_src/src/aquaverse/MDN/model/MDN.py