visualizedeeplayer

visualizedeeplayer.py

from numpy.random import *
import pickle
import numpy

pixnum = 784
pixnum2 = 100

alpha = 0.1
numIterations = 500

def sigmoid(x):
    return 1. / (1. + numpy.exp(-x))

with open('dae1.pkl', 'rb') as f:
    dae1 = pickle.load(f) 

with open('dae2.pkl', 'rb') as f2:
    dae2 = pickle.load(f2) 

W = dae1.get_weights()
vmatrix = dae2.get_weights()
print "W: " + str(numpy.shape(vmatrix))
W = W.T
xx = rand(pixnum,1)

final = numpy.zeros((pixnum,1))

for col in range (0,pixnum2):
    v = vmatrix[:,col]
    xx = rand(pixnum,1)
    print col
    for i in range(0, numIterations):
        tmp1 = sigmoid( numpy.dot(W,xx) ) * ( 1 - sigmoid(numpy.dot(W,xx)))
        tmp2 = numpy.dot(v.T,numpy.diag(tmp1.flat))
        gradient = numpy.dot(tmp2,W)
        gradient.shape = (pixnum,1) 
        xx = xx + alpha * gradient

    if col == 0 :
        final = xx.T 
    else :
        final = numpy.vstack((final,xx.T)) 

#print numpy.shape(final)

import gzip
from PIL import Image

def visualize_weights(weights, panel_shape, tile_size):

    def scale(x):
        eps = 1e-8
        x = x.copy()
        x -= x.min()
        x *= 1.0 / (x.max() + eps)
        return 255.0*x

    margin_y = numpy.zeros(tile_size[1])
    margin_x = numpy.zeros((tile_size[0] + 1) * panel_shape[0])
    image = margin_x.copy()

    for y in range(panel_shape[1]):
        foo =  [ numpy.c_[ scale( x.reshape(tile_size) ), margin_y ] for x in weights[y*panel_shape[0]:(y+1)*panel_shape[0]]]
        print foo
        tmp = numpy.hstack(foo)
        tmp = numpy.vstack([tmp, margin_x])

        image = numpy.vstack([image, tmp])

    img = Image.fromarray(image)
    img = img.convert('RGB')
    return img

panel_shape = (int(numpy.sqrt(final.shape[0])), int(numpy.sqrt(final.shape[0])))
tile_size = (int(numpy.sqrt(final.size)), int(numpy.sqrt(final.size)))
img = visualize_weights(final,(10,10),(28,28))
img.show()