keraflow.models.Model Class Reference

A model is a directed Kensor graph. More...

Inheritance diagram for keraflow.models.Model:

Public Member Functions
def	__init__
def	compile
	Configure the model and prepare inner utilized tensors. More...
def	fit
	Adjust model parameters to fit input data and true answers. More...
def	predict
	Returns the output values given the input data. More...
def	evaluate
	Returns the loss value and metrics values given the input data and true answer. More...
def	save_to_file
	Save the model to disk for later use. More...
def	save_arch
def	save_weights

Static Public Member Functions
def	load_from_file
	Load model from disk. More...

Detailed Description

A model is a directed Kensor graph.

Kensor connects with one another by Layer.__call__ and users specify the input kensors and out kensors to build the graph.

See Also

An Overview of Model

Constructor & Destructor Documentation

def keraflow.models.Model.__init__	(		self,
			inputs,
			outputs,
			name = None
	)

Parameters

inputs	single/list of Input.
outputs	single/list of Kensor.
name	str. Name of the model for easy debugging.

Member Function Documentation

def keraflow.models.Model.compile	(		self,
			optimizer,
			loss,
			metrics = []
	)

Configure the model and prepare inner utilized tensors.

Parameters

optimizer	str(name of optimizer class)/optimizer object. See keraflow.optimizers
loss	dict/list/single objective function(s)/str(name of objective function(s)). Objective(s) for output(s). See Objectives for a list of predefined objective functions.
metrics	dict/list/single objective function(s)/str(name of objective function(s)). Objective(s) for outputs(s). See Objectives for a list of predefined objective functions.

def keraflow.models.Model.evaluate	(		self,
			x,
			y,
			sample_weights = None,
			batch_size = 32,
			train_mode = False
	)

Returns the loss value and metrics values given the input data and true answer.

Parameters

x	dict/list/single numpy array(s) or an Nd list. Input data. Rules for assigning x is the same as in fit.
y	dict/list/single numpy array(s) or an Nd list. True answer. Rules for assigning y is the same as in fit.
sample_weights	list/numpy array. Weights of each sample.
batch_size	int. The batch size to predict the testing data. Might cause memorage error if set too large.
train_mode	boolean. For debugging usage. Do not use this flag in your code.

def keraflow.models.Model.fit	(		self,
			x,
			y,
			batch_size = 32,
			nb_epoch = 10,
			verbose = 1,
			callbacks = [],
			sample_weights = None,
			validation_split = 0.0,
			validation_data = None,
			shuffle = True
	)

Adjust model parameters to fit input data and true answers.

Parameters

x	dict/list/single numpy array(s) or an Nd list. Input data.
y	dict/list/single numpy array(s) or an Nd list. True answer.
batch_size	int. Number of samples per gradient update. Report error if unequal to (not None) input kensors' batch_size.
nb_epoch	int. The number of epochs to train the model.
verbose	0 for no logging, 1 for logging to stdout.
callbacks	list of keraflow.callbacks.Callback instances.
sample_weights	list/numpy array. Weights of each sample.
validation_split	float between 0 and 1. Fraction of the training data to be used as validation data.
validation_data	tuple (val_x, val_y), or (val_x, val_y, val_sample_weight) to be used as held-out validation data. val_x, val_y, and val_sample_weights accept the same type of data as x, y, and sample_weights respectively. This argument override validation_split argument.
shuffle	boolean. Whether to shuffle the samples at each epoch.

Returns

list of dict. Each dict in the list describes the loss (and validation loss if applicable) and metrics for one epoch.

Note

Rules for assigning x, y:

1. If a dict is passed (and the corresponding input kensors and output layers are named), the values are assigned according to the keys.
2. If a list of numpy array is passed, the arrays are assigned to the kensor in the order you construct the model.
3. If there's only a single input/output kensor, a single numpy array (or Nd list, will be casted into numpy array) is acceptable.

def keraflow.models.Model.load_from_file (   arch_fname,   weight_fname = None  )

static

Load model from disk.

Parameters

arch_fname	str. The file name storing the model architecture.
weight_fname	str. The file name storing weights of the model. If None, the weights are not loaded.

Returns

Model. The Model instance reconstructed from the file. Note that you should compile it before using it.

def keraflow.models.Model.predict	(		self,
			x,
			batch_size = 32,
			train_mode = False
	)

Returns the output values given the input data.

Parameters

x	dict/list/single numpy array(s) or an Nd list. Input data. Rules for assigning x is the same as in fit.
batch_size	int. The batch size to predict the testing data. Might cause memorage error if set too large.
train_mode	boolean. For debugging usage. Do not use this flag in your code.

Returns

list of numpy arrays. The length of the list is equal to the number of output channels.

def keraflow.models.Model.save_arch	(		self,
			arch_fname,
			overwrite = False,
			kwargs
	)

Parameters

arch_fname	str. The file name for storing the model architecture. Should ends with either 'json' or 'yaml' for different storing types.
kwargs	keyword arguments. The keyword arguments for json.dump or yaml.dump.
overwrite	Whether to overwrite existing files.

def keraflow.models.Model.save_to_file	(		self,
			arch_fname,
			weight_fname = None,
			overwrite = False,
			kwargs
	)

Save the model to disk for later use.

Combining operations done by save_arch and save_weights.

Parameters

arch_fname	str. The file name for storing the model architecture. Should ends with either 'json' or 'yaml' for different storing types.
weight_fname	str. The file name for storing weights of the model. If None, the weights are not saved.
overwrite	Whether to overwrite existing files.
kwargs	keyword arguments. The keyword arguments for json.dump or yaml.dump.

def keraflow.models.Model.save_weights	(		self,
			weight_fname,
			overwrite = False
	)

Parameters

weight_fname	str. The file name for storing weights of the model. If None, the weights are not saved.
overwrite	Whether to overwrite existing files.