test_tt.py

#!/usr/bin/env python

"""
Usage example employing Lasagne and TT-layer on the MNIST dataset.

This is a simplified version of
https://github.com/Lasagne/Lasagne/blob/master/examples/mnist.py
with using the TT-layer.
"""

from __future__ import print_function

import sys
import os
import time

import numpy as np
np.set_printoptions(threshold=np.nan)
import theano
import theano.tensor as T

import lasagne

from ttlayer import TTLayer

np.random.seed(1234)
size = 256

def load_dataset():
  X = np.random.rand(size)
  X_train = np.array([X]*1)
  return X_train

def build_simple_mlp(input_var=None):
  l_in = lasagne.layers.InputLayer(shape=(None, size), input_var=input_var)

  # Another 16-unit layer:
  l_hid1 = lasagne.layers.DenseLayer(
          l_in, num_units=size, W = np.random.rand(size, size),
          nonlinearity=lasagne.nonlinearities.identity)
  return l_hid1

def full_main(num_epochs=500):
  np.random.seed(1234)
  X_train = load_dataset()
  input_var = T.matrix('inputs')
  lhid_1 = build_simple_mlp(input_var)

  output = lasagne.layers.get_output(lhid_1)
  print_fn = theano.function([input_var], output)

  res = print_fn(X_train).reshape(size)
  print("\nFULL:\n", res)
  return res 

def build_mlp(input_var=None):
  l_in = lasagne.layers.InputLayer(shape=(None, size), input_var=input_var)
  l_hid1 = TTLayer(l_in, tt_input_shape=[4, 4, 4, 4], tt_output_shape=[4, 4, 4, 4], 
                   tt_ranks=[1, 10, 100, 10, 1], nonlinearity=lasagne.nonlinearities.identity)
  return l_hid1 

def main(num_epochs=500):
  np.random.seed(1234)
  X_train = load_dataset()
  input_var = T.matrix('inputs')
  lhid_1 = build_mlp(input_var)

  output = lasagne.layers.get_output(lhid_1)
  print_fn = theano.function([input_var], output)

  res = print_fn(X_train).reshape(size)

  print("\nTT:\n", res)
  return res

if __name__ == '__main__':
  np.random.seed(1234)
  X_tt = main() 
  X_fc = full_main()
  print('\nNORM:', np.linalg.norm(X_tt - X_fc))