SGDMPS

SGDMPS is a subclass of MPS used in place of the MPS for stochastic gradient descent optimization. A usage example is as follows:

import MNISTpreprocessing
from trmps import *

# Parameters
input_size = 196
shrink = True
d_feature = 2
d_output = 10
permuted = False

# Initialise with linear regression
data_source = MNISTpreprocessing.MNISTDatasource(shrink, permuted = permuted)
network = SimpleMPS(d_feature, d_output, input_size)
network.prepare(data_source=data_source)
feature, label = data_source.test_data
network.test(feature, label)

Properties

d_feature: int (read only)

The sizes of the feature vectors. (Referred to as ‘Local dimension’ in the paper). Unlike with when MPSs are optimized via DMRG, the stochastic gradient descent optimization algorithm does not scale as (d_feature)^3 so it is not neccesary to be as careful with this.

feature_reg: float (read only)

Parameter for regularising the weights. Set it larger than 1 for better performance.

reg: float (read only)

The ratio between loss and the regularisation penalty. Set it smaller than 1 for better performance.

start_node: tf.Tensor (read only)

Untrainable tf.Variable, the first node of the MPS. Is a vector. See Notes for developement for further details, but shouldn’t need to interface with this directly.

end_node: tf.Tensor (read only)

Untrainable tf.Variable, the last node of the MPS. Is a vector. See Notes for developement for further details, but shouldn’t need to interface with this directly.

nodes: list of tf.Variable

List of trainable tf.Variable, the middle nodes of the MPS.

Creating a new SGDMPS

__init__(d_feature, d_output, input_size, feature_reg=1.1, reg=0.001, special_node_loc=None):

Initialises the MPS. The prepare method must be called after this before anything else can be done.

d_feature: integer

The sizes of the feature vectors. (Referred to as ‘Local dimension’ in the paper)

d_output: integer

The size of the output. e.g. with 10-class classification, expressed as a one-hot vector, this would be 10

input_size: int

The input size, i.e. the number of matrices composing the matrix product state

feature_reg: float

Parameter for regularising the weights. Set it larger than 1 for better performance.

reg: float

Ratio between loss and the regularisation penalty. Set it smaller than 1 for better performance.