def plotImageData(data, imgShape, saveDir = None, prefix = 'imgdata', tileShape = (20,30), show = False, onlyRescaled = False):
isColor = (len(imgShape) > 2)
if not onlyRescaled:
image = Image.fromarray(tile_raster_images(
X = data.T, img_shape = imgShape,
tile_shape = tileShape, tile_spacing=(1,1),
scale_rows_to_unit_interval = False))
if saveDir:
image.save(os.path.join(saveDir, '%s.png' % prefix))
if show:
image.show()
image = Image.fromarray(tile_raster_images(
X = data.T, img_shape = imgShape,
tile_shape = tileShape, tile_spacing=(1,1),
scale_rows_to_unit_interval = True,
scale_colors_together = True))
if saveDir:
image.save(os.path.join(saveDir, '%s_rescale.png' % prefix))
if show:
image.show()
if isColor:
image = Image.fromarray(tile_raster_images(
X = data.T, img_shape = imgShape,
tile_shape = tileShape, tile_spacing=(1,1),
scale_rows_to_unit_interval = True,
scale_colors_together = False))
if saveDir:
image.save(os.path.join(saveDir, '%s_rescale_indiv.png' % prefix))
if show:
image.show()
def demo():
random.seed(0)
Nw = 15
xx, yy = randomCircleSampleMatrix(Nw=Nw, Nsamples=100)
saveImage = False
if saveImage:
image_data = numpy.ones(
((Nw + 1) * 10 - 1,
(Nw + 1) * 10 - 1), dtype='uint8') * 51 # dark gray
tmp = tile_raster_images(X=xx,
img_shape=(Nw, Nw),
tile_shape=(10, 10),
tile_spacing=(1, 1))
image_data = tmp
image = Image.fromarray(image_data)
image.save(os.path.join('demoCircleSamples.png'))
else:
pyplot.figure()
for ii in range(25):
ax = pyplot.subplot(5, 5, ii)
imagesc(xx[ii, :].reshape((Nw, Nw)), ax=ax)
#pyplot.title(repr(yy[ii,:]))
pyplot.title('%d, %d, %.2f' % tuple(yy[ii, :]))
pyplot.show()
def plotTopActivations(activations,
data,
imgShape,
saveDir=None,
nActivations=50,
nSamples=20,
prefix='topact',
show=False):
'''Plots top and bottom few activations for the first number activations.'''
sortIdx = argsort(activations, 1)
nActivations = min(nActivations, activations.shape[0])
plotData = zeros((prod(imgShape), nActivations * nSamples))
for ii in range(nActivations):
idx = sortIdx[ii, -1:-(nSamples + 1):-1]
plotData[:, (ii * nSamples):((ii + 1) * nSamples)] = data[:, idx]
image = Image.fromarray(
tile_raster_images(X=plotData.T,
img_shape=imgShape,
tile_shape=(nActivations, nSamples),
tile_spacing=(1, 1),
scale_rows_to_unit_interval=True))
if saveDir:
image.save(os.path.join(saveDir, '%s.png' % prefix))
if show:
image.show()
def makeData():
for Nw in (2, 4, 10, 16, 28, 50):
#for Nsamples in (50, 500, 5000):
for Nsamples in (50, 500, 5000, 50000):
random.seed(0)
for string in [
'train_sc_p5', 'test_sc_p5', 'train_sc_p1', 'test_sc_p1',
'train_4sc_p5', 'test_4sc_p5', 'train_4sc_p1',
'test_4sc_p1'
]:
probability = .5 if 'p5' in string else .1
function = random4SquaresCircles if '4sc' in string else randomSquareCircle
sampleXAndY = function(Nw, Nsamples, prob=probability)
saveToFile(
'../data/simpleShapes/%s_%02d_%d.pkl.gz' %
(string, Nw, Nsamples), sampleXAndY)
xx, yy = sampleXAndY
if Nsamples == 5000:
image = Image.fromarray(
tile_raster_images(X=xx,
img_shape=(Nw, Nw),
tile_shape=(10, 15),
tile_spacing=(1, 1),
scale_rows_to_unit_interval=False))
image.save('../data/simpleShapes/%s_%02d.png' %
(string, Nw))
def plotImageRicaWW(WW, imgShape, saveDir, tileShape = None, prefix = 'WW'):
imgIsColor = len(imgShape) > 2
nOutputs, nInputs = WW.shape
if tileShape is None: tileShape = getTileShape(nOutputs)
if saveDir:
image = Image.fromarray(tile_raster_images(
X = WW,
img_shape = imgShape, tile_shape = tileShape,
tile_spacing=(1,1),
scale_colors_together = True))
image.save(os.path.join(saveDir, '%s.png' % prefix))
if imgIsColor:
image = Image.fromarray(tile_raster_images(
X = WW,
img_shape = imgShape, tile_shape = tileShape,
tile_spacing=(1,1),
scale_colors_together = False))
image.save(os.path.join(saveDir, '%s_rescale_indiv.png' % prefix))
def plotImageData(data,
imgShape,
saveDir=None,
prefix='imgdata',
tileShape=(20, 30),
show=False,
onlyRescaled=False):
isColor = (len(imgShape) > 2)
if not onlyRescaled:
image = Image.fromarray(
tile_raster_images(X=data.T,
img_shape=imgShape,
tile_shape=tileShape,
tile_spacing=(1, 1),
scale_rows_to_unit_interval=False))
if saveDir:
image.save(os.path.join(saveDir, '%s.png' % prefix))
if show:
image.show()
image = Image.fromarray(
tile_raster_images(X=data.T,
img_shape=imgShape,
tile_shape=tileShape,
tile_spacing=(1, 1),
scale_rows_to_unit_interval=True,
scale_colors_together=True))
if saveDir:
image.save(os.path.join(saveDir, '%s_rescale.png' % prefix))
if show:
image.show()
if isColor:
image = Image.fromarray(
tile_raster_images(X=data.T,
img_shape=imgShape,
tile_shape=tileShape,
tile_spacing=(1, 1),
scale_rows_to_unit_interval=True,
scale_colors_together=False))
if saveDir:
image.save(os.path.join(saveDir, '%s_rescale_indiv.png' % prefix))
if show:
image.show()
def plotImageRicaWW(WW, imgShape, saveDir, tileShape=None, prefix='WW'):
imgIsColor = len(imgShape) > 2
nOutputs, nInputs = WW.shape
if tileShape is None: tileShape = getTileShape(nOutputs)
if saveDir:
image = Image.fromarray(
tile_raster_images(X=WW,
img_shape=imgShape,
tile_shape=tileShape,
tile_spacing=(1, 1),
scale_colors_together=True))
image.save(os.path.join(saveDir, '%s.png' % prefix))
if imgIsColor:
image = Image.fromarray(
tile_raster_images(X=WW,
img_shape=imgShape,
tile_shape=tileShape,
tile_spacing=(1, 1),
scale_colors_together=False))
image.save(os.path.join(saveDir, '%s_rescale_indiv.png' % prefix))
def demo():
random.seed(0)
Nw = 2
xx, yy = randomSquareCircle(Nw = Nw, Nsamples = 150)
image = Image.fromarray(tile_raster_images(
X = xx, img_shape = (Nw,Nw),
tile_shape = (10, 15), tile_spacing=(1,1),
scale_rows_to_unit_interval = False))
image.save('demo.png')
print 'saved as demo.png'
image.show()
def demo():
random.seed(0)
Nw = 2
xx, yy = randomSquareCircle(Nw=Nw, Nsamples=150)
image = Image.fromarray(
tile_raster_images(X=xx,
img_shape=(Nw, Nw),
tile_shape=(10, 15),
tile_spacing=(1, 1),
scale_rows_to_unit_interval=False))
image.save('demo.png')
print 'saved as demo.png'
image.show()
def makeData():
for Nw in (2, 4, 10, 16, 28, 50):
#for Nsamples in (50, 500, 5000):
for Nsamples in (50, 500, 5000, 50000):
random.seed(0)
for string in ['train_sc_p5', 'test_sc_p5', 'train_sc_p1', 'test_sc_p1',
'train_4sc_p5', 'test_4sc_p5', 'train_4sc_p1', 'test_4sc_p1']:
probability = .5 if 'p5' in string else .1
function = random4SquaresCircles if '4sc' in string else randomSquareCircle
sampleXAndY = function(Nw, Nsamples, prob = probability)
saveToFile('../data/simpleShapes/%s_%02d_%d.pkl.gz' % (string, Nw, Nsamples), sampleXAndY)
xx, yy = sampleXAndY
if Nsamples == 5000:
image = Image.fromarray(tile_raster_images(
X = xx, img_shape = (Nw,Nw),
tile_shape = (10, 15), tile_spacing=(1,1),
scale_rows_to_unit_interval = False))
image.save('../data/simpleShapes/%s_%02d.png' % (string, Nw))
def plotTopActivations(activations, data, imgShape, saveDir = None, nActivations = 50, nSamples = 20, prefix = 'topact', show = False):
'''Plots top and bottom few activations for the first number activations.'''
sortIdx = argsort(activations, 1)
nActivations = min(nActivations, activations.shape[0])
plotData = zeros((prod(imgShape), nActivations*nSamples))
for ii in range(nActivations):
idx = sortIdx[ii,-1:-(nSamples+1):-1]
plotData[:,(ii*nSamples):((ii+1)*nSamples)] = data[:,idx]
image = Image.fromarray(tile_raster_images(
X = plotData.T, img_shape = imgShape,
tile_shape = (nActivations, nSamples), tile_spacing=(1,1),
scale_rows_to_unit_interval = True))
if saveDir:
image.save(os.path.join(saveDir, '%s.png' % prefix))
if show:
image.show()
def demo():
random.seed(0)
Nw = 15
xx, yy = randomCircleSampleMatrix(Nw = Nw, Nsamples = 100)
saveImage = False
if saveImage:
image_data = numpy.ones(((Nw+1)*10-1,(Nw+1)*10-1), dtype='uint8') * 51 # dark gray
tmp = tile_raster_images(X = xx,
img_shape = (Nw,Nw),
tile_shape = (10,10),
tile_spacing = (1,1))
image_data = tmp
image = Image.fromarray(image_data)
image.save(os.path.join('demoCircleSamples.png'))
else:
pyplot.figure()
for ii in range(25):
ax = pyplot.subplot(5,5,ii)
imagesc(xx[ii,:].reshape((Nw,Nw)), ax=ax)
#pyplot.title(repr(yy[ii,:]))
pyplot.title('%d, %d, %.2f' % tuple(yy[ii,:]))
pyplot.show()
def test_rbm(learning_rate=0.1, training_epochs = 15,
datasets = None, batch_size = 20,
n_chains = 20, n_samples = 14, output_dir = 'rbm_plots',
img_dim = 28, n_input = None, n_hidden = 500, quickHack = False,
visibleModel = 'binary', initWfactor = 1.0,
imgPlotFunction = None):
'''
Demonstrate how to train an RBM.
This is demonstrated on MNIST.
:param learning_rate: learning rate used for training the RBM
:param training_epochs: number of epochs used for training
:param dataset: path the the pickled dataset
:param batch_size: size of a batch used to train the RBM
:param n_chains: number of parallel Gibbs chains to be used for sampling
:param n_samples: number of samples to plot for each chain
:param visibleModel: 'real' or 'binary'
:param initWfactor: Typicaly 1 for binary or .01 for real
XXX:param pcaDims: None to skip PCA or >0 to use PCA to reduce dimensionality of data first.
'''
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
if quickHack:
train_set_x = train_set_x[:2500,:]
if train_set_y is not None:
train_set_y = train_set_y[:2500]
print ('(%d, %d, %d) %d dimensional examples in (train, valid, test)' %
(train_set_x.shape[0], valid_set_x.shape[0], test_set_x.shape[0], train_set_x.shape[1]))
# compute number of minibatches for training, validation and testing
n_train_batches = train_set_x.shape[0] / batch_size
print 'n_train_batches is', n_train_batches
rng = numpy.random.RandomState(1)
if n_input is None:
n_input = train_set_x.shape[1]
# construct the RBM class
rbm = RBM(nVisible=n_input, nHidden = n_hidden, numpyRng = rng,
visibleModel = visibleModel, initWfactor = initWfactor)
#################################
# Training the RBM #
#################################
print 'starting training.'
if not os.path.isdir(output_dir):
os.makedirs(output_dir)
plotting_time = 0.
start_time = time.clock()
# go through training epochs
meanCosts = []
ii = -1
metrics = array([])
plotEvery = 100
for epoch in xrange(training_epochs):
# go through the training set
for batch_index in xrange(n_train_batches):
#print 'about to train using exemplars %d to %d.' % (batch_index*batch_size, (batch_index+1)*batch_size)
ii += 1
if ii % plotEvery == 0:
plotWeights = '%03i_%05i' % (epoch, batch_index)
calcMetrics = True
else:
plotWeights = False
calcMetrics = False
# metric is xEntropyCost, reconError
metric = rbm.train(train_set_x[batch_index*batch_size:(batch_index+1)*batch_size],
lr = learning_rate, metrics = calcMetrics, plotWeights = plotWeights,
output_dir = output_dir)
if calcMetrics:
if len(metrics) == 0:
metrics = array([metric])
else:
metrics = vstack((metrics, metric))
if ii % plotEvery == 0:
# Plot filters after each single step
plotting_start = time.clock()
# Construct image from the weight matrix
image = Image.fromarray(tile_raster_images(
X = imgPlotFunction(rbm.W.T) if imgPlotFunction else rbm.W.T,
img_shape = (img_dim,img_dim),tile_shape = (10,10),
tile_spacing=(1,1)))
image.save(os.path.join(output_dir, 'filters_at_epoch_batch_%03i_%05i.png' % (epoch, batch_index)))
plotting_stop = time.clock()
plotting_time += (plotting_stop - plotting_start)
#print ' Training epoch %d batch %d, xEntropyCost is ' % (epoch, batch_index), numpy.mean(mean_cost),
print ' Training epoch %d batch %d, xEntropyCost is ' % (epoch, batch_index), metrics[-1,0],
print '\trecon error ', metrics[-1,1]
thisEpochStart = epoch *n_train_batches/plotEvery
thisEpochEnd = (epoch+1)*n_train_batches/plotEvery
epochMeanXEnt = mean(metrics[thisEpochStart:thisEpochEnd,0])
epochMeanRecon = mean(metrics[thisEpochStart:thisEpochEnd,1])
print 'Training epoch %d mean xEntropyCost is ' % (epoch), epochMeanXEnt, '\trecon error ', epochMeanRecon
meanCosts.append(epochMeanXEnt)
# Plot filters after each training epoch
plotting_start = time.clock()
# Construct image from the weight matrix
image = Image.fromarray(tile_raster_images(
X = imgPlotFunction(rbm.W.T) if imgPlotFunction else rbm.W.T,
img_shape = (img_dim,img_dim),tile_shape = (10,10),
tile_spacing=(1,1)))
image.save(os.path.join(output_dir, 'filters_at_epoch_%03i.png' % epoch))
plotting_stop = time.clock()
plotting_time += (plotting_stop - plotting_start)
plotting_start = time.clock()
pyplot.plot(metrics)
pyplot.savefig(os.path.join(output_dir, 'reconErr.png'))
plotting_time += (time.clock() - plotting_start)
end_time = time.clock()
pretraining_time = (end_time - start_time) - plotting_time
print ('Training took %f minutes' %(pretraining_time/60.))
print ('Plotting took %f minutes' %(plotting_time/60.))
#################################
# Plot some samples from RBM #
#################################
# find out the number of test samples
number_of_test_samples = test_set_x.shape[0]
plot_every = 1
# if imgPlotFunction is defined, then also plot before function if
# the data is of the same dimension (e.g. for ZCA, but not for
# PCA).
plotRawAlso = (imgPlotFunction and train_set_x.shape[0] == img_dim * img_dim)
# create a space to store the image for plotting ( we need to leave
# room for the tile_spacing as well)
image_data = numpy.ones(((img_dim+1)*n_samples-1,(img_dim+1)*n_chains-1), dtype='uint8') * 51 # dark gray
if plotRawAlso:
image_data_raw = numpy.ones(((img_dim+1)*n_samples-1,(img_dim+1)*n_chains-1), dtype='uint8') * 51 # dark gray
for ii in xrange(n_chains):
# generate `plot_every` intermediate samples that we discard, because successive samples in the chain are too correlated
test_idx = rng.randint(number_of_test_samples)
samples = numpy.zeros((n_chains, n_input))
visMean = test_set_x[test_idx,:]
visSample = visMean
for jj in xrange(n_samples):
samples[jj,:] = visMean # show the mean, but use the sample for gibbs steps
if jj == n_samples-1: break # skip the last for speed
plot_every = 2**jj # exponentially increasing number of gibbs samples. max for n_samples=14 is 2^12
for ss in xrange(plot_every):
visMean, visSample = rbm.gibbs_vhv(visSample)[4:6] # 4 for mean, 5 for sample
print ' ... plotting sample ', ii
image_data[:,(img_dim+1)*ii:(img_dim+1)*ii+img_dim] = tile_raster_images(
X = imgPlotFunction(samples) if imgPlotFunction else samples,
img_shape = (img_dim,img_dim),
tile_shape = (n_samples, 1),
tile_spacing = (1,1))
if plotRawAlso:
image_data_raw[:,(img_dim+1)*ii:(img_dim+1)*ii+img_dim] = tile_raster_images(
X = samples,
img_shape = (img_dim,img_dim),
tile_shape = (n_samples, 1),
tile_spacing = (1,1))
image = Image.fromarray(image_data)
image.save(os.path.join(output_dir, 'samples.png'))
if plotRawAlso:
image = Image.fromarray(image_data)
image.save(os.path.join(output_dir, 'samplesRaw.png'))
saveToFile(os.path.join(output_dir, 'rbm.pkl.gz'), rbm)
return rbm, meanCosts
def plotRicaReconstructions(rica,
data,
imgShape,
saveDir=None,
unwhitener=None,
tileShape=None,
number=50,
prefix='recon',
onlyHilights=False,
hilightCmap=None):
'''Plots reconstructions for some randomly chosen data points.'''
if saveDir:
print 'Plotting %d recon plots...' % number,
sys.stdout.flush()
imgIsColor = len(imgShape) > 2
nOutputs, nInputs = rica.WW.shape
if tileShape is None: tileShape = getTileShape(nOutputs)
tileRescaleFactor = 2
reconRescaleFactor = 3
font = ImageFont.load_default()
hidden = dot(rica.WW, data[:, :number])
reconstruction = dot(rica.WW.T, hidden)
if unwhitener:
#pdb.set_trace() DEBUG?
dataOrig = unwhitener(data[:, :number])
reconstructionOrig = unwhitener(reconstruction[:, :number])
for ii in xrange(number):
# Hilighted tiled image
hilightAmount = abs(hidden[:, ii])
maxHilight = hilightAmount.max()
#hilightAmount -= hilightAmount.min() # Don't push to 0
hilightAmount /= maxHilight + 1e-6
if hilightCmap:
cmap = cm.get_cmap(hilightCmap)
hilights = cmap(hilightAmount)[:, :3] # chop off alpha channel
else:
# default black -> red colormap
hilights = outer(hilightAmount, array([1, 0, 0]))
tileImg = Image.fromarray(
tile_raster_images(X=rica.WW,
img_shape=imgShape,
tile_shape=tileShape,
tile_spacing=(2, 2),
scale_colors_together=True,
hilights=hilights,
onlyHilights=onlyHilights))
tileImg = tileImg.resize(
[x * tileRescaleFactor for x in tileImg.size])
# Input / Reconstruction image
if unwhitener:
rawReconErr = array([
dataOrig[:, ii], data[:, ii], reconstruction[:, ii],
reconstructionOrig[:, ii],
reconstruction[:, ii] - data[:, ii],
reconstructionOrig[:, ii] - dataOrig[:, ii]
])
# Scale data-raw and recon-raw together between 0 and 1
rawReconErr = scale_rows_together_to_unit_interval(
rawReconErr, [0, 3], anchor0=False)
# Scale data-white and recon-white together, map 0 -> 50% gray
rawReconErr = scale_rows_together_to_unit_interval(
rawReconErr, [1, 2], anchor0=True)
# Scale diffs independently to [0,1]
rawReconErr = scale_some_rows_to_unit_interval(
rawReconErr, [4, 5])
else:
rawReconErr = array([
data[:, ii], reconstruction[:, ii],
reconstruction[:, ii] - data[:, ii]
])
# Scale data-raw and recon-raw together between 0 and 1
rawReconErr = scale_rows_together_to_unit_interval(
rawReconErr, [0, 1], anchor0=False)
# Scale diffs independently to [0,1]
rawReconErr = scale_some_rows_to_unit_interval(
rawReconErr, [2])
rawReconErrImg = Image.fromarray(
tile_raster_images(X=rawReconErr,
img_shape=imgShape,
tile_shape=(rawReconErr.shape[0], 1),
tile_spacing=(1, 1),
scale_rows_to_unit_interval=False))
rawReconErrImg = rawReconErrImg.resize(
[x * reconRescaleFactor for x in rawReconErrImg.size])
# Add Red activation limit
redString = '%g' % maxHilight
fontSize = font.font.getsize(redString)
size = (max(tileImg.size[0],
fontSize[0]), tileImg.size[1] + fontSize[1])
tempImage = Image.new('RGBA', size, (51, 51, 51))
tempImage.paste(tileImg, (0, 0))
draw = ImageDraw.Draw(tempImage)
draw.text(((size[0] - fontSize[0]) / 2, size[1] - fontSize[1]),
redString,
font=font)
tileImg = tempImage
# Combined
costEtc = rica.cost(rica.WW, data[:, ii:ii + 1])
costString = rica.getReconPlotString(costEtc)
fontSize = font.font.getsize(costString)
size = (max(
tileImg.size[0] + rawReconErrImg.size[0] + reconRescaleFactor,
fontSize[0]),
max(tileImg.size[1], rawReconErrImg.size[1]) + fontSize[1])
wholeImage = Image.new('RGBA', size, (51, 51, 51))
wholeImage.paste(tileImg, (0, 0))
wholeImage.paste(rawReconErrImg,
(tileImg.size[0] + reconRescaleFactor, 0))
draw = ImageDraw.Draw(wholeImage)
draw.text(((size[0] - fontSize[0]) / 2, size[1] - fontSize[1]),
costString,
font=font)
wholeImage.save(os.path.join(saveDir,
'%s_%04d.png' % (prefix, ii)))
print 'done.'
def plotRicaReconstructions(rica, data, imgShape, saveDir = None, unwhitener = None, tileShape = None, number = 50, prefix = 'recon', onlyHilights = False, hilightCmap = None):
'''Plots reconstructions for some randomly chosen data points.'''
if saveDir:
print 'Plotting %d recon plots...' % number,
sys.stdout.flush()
imgIsColor = len(imgShape) > 2
nOutputs, nInputs = rica.WW.shape
if tileShape is None: tileShape = getTileShape(nOutputs)
tileRescaleFactor = 2
reconRescaleFactor = 3
font = ImageFont.load_default()
hidden = dot(rica.WW, data[:,:number])
reconstruction = dot(rica.WW.T, hidden)
if unwhitener:
#pdb.set_trace() DEBUG?
dataOrig = unwhitener(data[:,:number])
reconstructionOrig = unwhitener(reconstruction[:,:number])
for ii in xrange(number):
# Hilighted tiled image
hilightAmount = abs(hidden[:,ii])
maxHilight = hilightAmount.max()
#hilightAmount -= hilightAmount.min() # Don't push to 0
hilightAmount /= maxHilight + 1e-6
if hilightCmap:
cmap = cm.get_cmap(hilightCmap)
hilights = cmap(hilightAmount)[:,:3] # chop off alpha channel
else:
# default black -> red colormap
hilights = outer(hilightAmount, array([1,0,0]))
tileImg = Image.fromarray(tile_raster_images(
X = rica.WW,
img_shape = imgShape, tile_shape = tileShape,
tile_spacing=(2,2),
scale_colors_together = True,
hilights = hilights,
onlyHilights = onlyHilights))
tileImg = tileImg.resize([x*tileRescaleFactor for x in tileImg.size])
# Input / Reconstruction image
if unwhitener:
rawReconErr = array([dataOrig[:,ii], data[:,ii], reconstruction[:,ii], reconstructionOrig[:,ii],
reconstruction[:,ii]-data[:,ii], reconstructionOrig[:,ii]-dataOrig[:,ii]])
# Scale data-raw and recon-raw together between 0 and 1
rawReconErr = scale_rows_together_to_unit_interval(rawReconErr, [0, 3], anchor0 = False)
# Scale data-white and recon-white together, map 0 -> 50% gray
rawReconErr = scale_rows_together_to_unit_interval(rawReconErr, [1, 2], anchor0 = True)
# Scale diffs independently to [0,1]
rawReconErr = scale_some_rows_to_unit_interval(rawReconErr, [4, 5])
else:
rawReconErr = array([data[:,ii], reconstruction[:,ii],
reconstruction[:,ii]-data[:,ii]])
# Scale data-raw and recon-raw together between 0 and 1
rawReconErr = scale_rows_together_to_unit_interval(rawReconErr, [0, 1], anchor0 = False)
# Scale diffs independently to [0,1]
rawReconErr = scale_some_rows_to_unit_interval(rawReconErr, [2])
rawReconErrImg = Image.fromarray(tile_raster_images(
X = rawReconErr,
img_shape = imgShape, tile_shape = (rawReconErr.shape[0], 1),
tile_spacing=(1,1),
scale_rows_to_unit_interval = False))
rawReconErrImg = rawReconErrImg.resize([x*reconRescaleFactor for x in rawReconErrImg.size])
# Add Red activation limit
redString = '%g' % maxHilight
fontSize = font.font.getsize(redString)
size = (max(tileImg.size[0], fontSize[0]), tileImg.size[1] + fontSize[1])
tempImage = Image.new('RGBA', size, (51, 51, 51))
tempImage.paste(tileImg, (0, 0))
draw = ImageDraw.Draw(tempImage)
draw.text(((size[0]-fontSize[0])/2, size[1]-fontSize[1]), redString, font=font)
tileImg = tempImage
# Combined
costEtc = rica.cost(rica.WW, data[:,ii:ii+1])
costString = rica.getReconPlotString(costEtc)
fontSize = font.font.getsize(costString)
size = (max(tileImg.size[0] + rawReconErrImg.size[0] + reconRescaleFactor, fontSize[0]),
max(tileImg.size[1], rawReconErrImg.size[1]) + fontSize[1])
wholeImage = Image.new('RGBA', size, (51, 51, 51))
wholeImage.paste(tileImg, (0, 0))
wholeImage.paste(rawReconErrImg, (tileImg.size[0] + reconRescaleFactor, 0))
draw = ImageDraw.Draw(wholeImage)
draw.text(((size[0]-fontSize[0])/2, size[1]-fontSize[1]), costString, font=font)
wholeImage.save(os.path.join(saveDir, '%s_%04d.png' % (prefix, ii)))
print 'done.'
def testIca(datasets, savedir = None, smallImgHack = False, quickHack = False):
'''Test ICA on a given dataset.'''
random.seed(1)
# 0. Get data
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
if quickHack:
print '!!! Using quickHack !!!'
train_set_x = train_set_x[:2500,:]
if train_set_y is not None:
train_set_y = train_set_y[:2500]
if smallImgHack:
print '!!! Using smallImgHack !!! (images will be misaligned)'
train_set_x = train_set_x[:,:100]
print ('(%d, %d, %d) %d dimensional examples in (train, valid, test)' %
(train_set_x.shape[0], valid_set_x.shape[0], test_set_x.shape[0], train_set_x.shape[1]))
nDim = train_set_x.shape[1]
imgDim = int(round(sqrt(nDim))) # Might not always be true...
randIdxRaw = random.randint(0, nDim, 100)
randIdxWhite = random.randint(0, nDim, 100)
randIdxSource = random.randint(0, nDim, 100)
image = Image.fromarray(tile_raster_images(
X = train_set_x,
img_shape = (imgDim,imgDim), tile_shape = (10,15),
tile_spacing=(1,1)))
if savedir: image.save(os.path.join(savedir, 'data_raw.png'))
image.show()
pyplot.figure()
for ii in range(20):
idx = randIdxRaw[ii]
pyplot.subplot(4,5,ii+1)
pyplot.title('raw dim %d' % idx)
pyplot.hist(train_set_x[:,idx])
if savedir: pyplot.savefig(os.path.join(savedir, 'data_raw_hist.png'))
# 1. Whiten data
print 'Whitening data with pca...'
pca = PCA(train_set_x)
xWhite = pca.toZca(train_set_x)
print ' done.'
pyplot.figure()
for ii in range(20):
idx = randIdxWhite[ii]
pyplot.subplot(4,5,ii+1)
pyplot.title('data white dim %d' % idx)
pyplot.hist(xWhite[:,idx])
if savedir: pyplot.savefig(os.path.join(savedir, 'data_white_hist.png'))
image = Image.fromarray(tile_raster_images(
X = xWhite,
img_shape = (imgDim,imgDim), tile_shape = (10,15),
tile_spacing=(1,1)))
if savedir: image.save(os.path.join(savedir, 'data_white.png'))
image.show()
# 1.1 plot hist
pyplot.figure()
pyplot.hold(True)
pyplot.title('data white 20 random dims')
histMax = 0
histMin = 1e10
for ii in range(20):
idx = randIdxWhite[ii]
hist, binEdges = histogram(xWhite[:,idx], bins = 20, density = True)
histMax = max(histMax, max(hist))
histMin = min(histMin, min(hist[hist != 0])) # min non-zero entry
binMiddles = binEdges[:-1] + (binEdges[1] - binEdges[0])/2
#print ' %d from %f to %f' % (ii, min(binMiddles), max(binMiddles))
pyplot.semilogy(binMiddles, hist, '.-')
pyplot.axis('tight')
ax = looser(pyplot.axis(), semilogy = True)
xAbsMax = max(fabs(ax[0:2]))
xx = linspace(-xAbsMax, xAbsMax, 100)
pyplot.semilogy(xx, mlab.normpdf(xx, 0, 1), 'k', linewidth = 3)
pyplot.axis((-xAbsMax, xAbsMax, ax[2], ax[3]))
if savedir: pyplot.savefig(os.path.join(savedir, 'data_white_log_hist.png'))
# 1.2 plot points
pyplot.figure()
pyplot.hold(True)
pyplot.title('data white 20 random dims')
nSamples = min(xWhite.shape[0], 1000)
print 'data_white_log_points plotted with', nSamples, 'samples.'
for ii in range(10):
idx = randIdxWhite[ii]
pyplot.plot(xWhite[:nSamples,idx],
ii + random.uniform(-.25, .25, nSamples), 'o')
pyplot.axis('tight')
if savedir: pyplot.savefig(os.path.join(savedir, 'data_white_log_points.png'))
# 2. Fit ICA
rng = random.RandomState(1)
ica = FastICA(random_state = rng, whiten = False)
print 'Fitting ICA...'
ica.fit(xWhite)
print ' done.'
if savedir: saveToFile(os.path.join(savedir, 'ica.pkl.gz'), ica)
print 'Geting sources and mixing matrix...'
sourcesWhite = ica.transform(xWhite) # Estimate the sources
#S_fica /= S_fica.std(axis=0) # (should already be done)
mixingMatrix = ica.get_mixing_matrix()
print ' done.'
sources = pca.fromZca(sourcesWhite)
# 3. Show independent components and inferred sources
image = Image.fromarray(tile_raster_images(
X = mixingMatrix,
img_shape = (imgDim,imgDim), tile_shape = (10,15),
tile_spacing=(1,1)))
if savedir: image.save(os.path.join(savedir, 'ic_white.png'))
image.show()
image = Image.fromarray(tile_raster_images(
X = mixingMatrix.T,
img_shape = (imgDim,imgDim), tile_shape = (10,15),
tile_spacing=(1,1)))
if savedir: image.save(os.path.join(savedir, 'ic_white.T.png'))
image.show()
image = Image.fromarray(tile_raster_images(
X = pca.fromZca(mixingMatrix),
img_shape = (imgDim,imgDim), tile_shape = (10,15),
tile_spacing=(1,1)))
if savedir: image.save(os.path.join(savedir, 'ic_raw.png'))
image.show()
image = Image.fromarray(tile_raster_images(
X = pca.fromZca(mixingMatrix.T),
img_shape = (imgDim,imgDim), tile_shape = (10,15),
tile_spacing=(1,1)))
if savedir: image.save(os.path.join(savedir, 'ic_raw.T.png'))
image.show()
pyplot.figure()
for ii in range(20):
idx = randIdxSource[ii]
pyplot.subplot(4,5,ii+1)
pyplot.title('sourceWhite %d' % idx)
pyplot.hist(sourcesWhite[:,idx])
if savedir: pyplot.savefig(os.path.join(savedir, 'sources_white_hist.png'))
image = Image.fromarray(tile_raster_images(
X = sourcesWhite,
img_shape = (imgDim,imgDim), tile_shape = (10,15),
tile_spacing=(1,1)))
if savedir: image.save(os.path.join(savedir, 'sources_white.png'))
image.show()
# 3.1 plot hist
pyplot.figure()
pyplot.hold(True)
pyplot.title('sources white 20 random dims')
histMax = 0
histMin = 1e10
for ii in range(20):
idx = randIdxSource[ii]
hist, binEdges = histogram(sourcesWhite[:,idx], bins = 20, density = True)
histMax = max(histMax, max(hist))
histMin = min(histMin, min(hist[hist != 0])) # min non-zero entry
binMiddles = binEdges[:-1] + (binEdges[1] - binEdges[0])/2
#print ' %d from %f to %f' % (ii, min(binMiddles), max(binMiddles))
pyplot.semilogy(binMiddles, hist, '.-')
pyplot.axis('tight')
ax = looser(pyplot.axis(), semilogy = True)
xAbsMax = max(fabs(ax[0:2]))
xx = linspace(-xAbsMax, xAbsMax, 100)
pyplot.semilogy(xx, mlab.normpdf(xx, 0, 1), 'k', linewidth = 3)
pyplot.axis((-xAbsMax, xAbsMax, ax[2], ax[3]))
if savedir: pyplot.savefig(os.path.join(savedir, 'sources_white_log_hist.png'))
# 3.2 plot points
pyplot.figure()
pyplot.hold(True)
pyplot.title('sources white 20 random dims')
nSamples = min(sourcesWhite.shape[0], 1000)
print 'sources_white_log_points plotted with', nSamples, 'samples.'
for ii in range(10):
idx = randIdxWhite[ii]
pyplot.plot(sourcesWhite[:nSamples,idx],
ii + random.uniform(-.25, .25, nSamples), 'o')
pyplot.axis('tight')
if savedir: pyplot.savefig(os.path.join(savedir, 'sources_white_log_points.png'))
image = Image.fromarray(tile_raster_images(
X = sources,
img_shape = (imgDim,imgDim), tile_shape = (10,15),
tile_spacing=(1,1)))
if savedir: image.save(os.path.join(savedir, 'sources_raw.png'))
image.show()
if savedir:
print 'plots saved in', savedir
else:
import ipdb; ipdb.set_trace()
def testIca(datasets, savedir=None, smallImgHack=False, quickHack=False):
'''Test ICA on a given dataset.'''
random.seed(1)
# 0. Get data
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
if quickHack:
print '!!! Using quickHack !!!'
train_set_x = train_set_x[:2500, :]
if train_set_y is not None:
train_set_y = train_set_y[:2500]
if smallImgHack:
print '!!! Using smallImgHack !!! (images will be misaligned)'
train_set_x = train_set_x[:, :100]
print('(%d, %d, %d) %d dimensional examples in (train, valid, test)' %
(train_set_x.shape[0], valid_set_x.shape[0], test_set_x.shape[0],
train_set_x.shape[1]))
nDim = train_set_x.shape[1]
imgDim = int(round(sqrt(nDim))) # Might not always be true...
randIdxRaw = random.randint(0, nDim, 100)
randIdxWhite = random.randint(0, nDim, 100)
randIdxSource = random.randint(0, nDim, 100)
image = Image.fromarray(
tile_raster_images(X=train_set_x,
img_shape=(imgDim, imgDim),
tile_shape=(10, 15),
tile_spacing=(1, 1)))
if savedir: image.save(os.path.join(savedir, 'data_raw.png'))
image.show()
pyplot.figure()
for ii in range(20):
idx = randIdxRaw[ii]
pyplot.subplot(4, 5, ii + 1)
pyplot.title('raw dim %d' % idx)
pyplot.hist(train_set_x[:, idx])
if savedir: pyplot.savefig(os.path.join(savedir, 'data_raw_hist.png'))
# 1. Whiten data
print 'Whitening data with pca...'
pca = PCA(train_set_x)
xWhite = pca.toZca(train_set_x)
print ' done.'
pyplot.figure()
for ii in range(20):
idx = randIdxWhite[ii]
pyplot.subplot(4, 5, ii + 1)
pyplot.title('data white dim %d' % idx)
pyplot.hist(xWhite[:, idx])
if savedir: pyplot.savefig(os.path.join(savedir, 'data_white_hist.png'))
image = Image.fromarray(
tile_raster_images(X=xWhite,
img_shape=(imgDim, imgDim),
tile_shape=(10, 15),
tile_spacing=(1, 1)))
if savedir: image.save(os.path.join(savedir, 'data_white.png'))
image.show()
# 1.1 plot hist
pyplot.figure()
pyplot.hold(True)
pyplot.title('data white 20 random dims')
histMax = 0
histMin = 1e10
for ii in range(20):
idx = randIdxWhite[ii]
hist, binEdges = histogram(xWhite[:, idx], bins=20, density=True)
histMax = max(histMax, max(hist))
histMin = min(histMin, min(hist[hist != 0])) # min non-zero entry
binMiddles = binEdges[:-1] + (binEdges[1] - binEdges[0]) / 2
#print ' %d from %f to %f' % (ii, min(binMiddles), max(binMiddles))
pyplot.semilogy(binMiddles, hist, '.-')
pyplot.axis('tight')
ax = looser(pyplot.axis(), semilogy=True)
xAbsMax = max(fabs(ax[0:2]))
xx = linspace(-xAbsMax, xAbsMax, 100)
pyplot.semilogy(xx, mlab.normpdf(xx, 0, 1), 'k', linewidth=3)
pyplot.axis((-xAbsMax, xAbsMax, ax[2], ax[3]))
if savedir:
pyplot.savefig(os.path.join(savedir, 'data_white_log_hist.png'))
# 1.2 plot points
pyplot.figure()
pyplot.hold(True)
pyplot.title('data white 20 random dims')
nSamples = min(xWhite.shape[0], 1000)
print 'data_white_log_points plotted with', nSamples, 'samples.'
for ii in range(10):
idx = randIdxWhite[ii]
pyplot.plot(xWhite[:nSamples, idx],
ii + random.uniform(-.25, .25, nSamples), 'o')
pyplot.axis('tight')
if savedir:
pyplot.savefig(os.path.join(savedir, 'data_white_log_points.png'))
# 2. Fit ICA
rng = random.RandomState(1)
ica = FastICA(random_state=rng, whiten=False)
print 'Fitting ICA...'
ica.fit(xWhite)
print ' done.'
if savedir: saveToFile(os.path.join(savedir, 'ica.pkl.gz'), ica)
print 'Geting sources and mixing matrix...'
sourcesWhite = ica.transform(xWhite) # Estimate the sources
#S_fica /= S_fica.std(axis=0) # (should already be done)
mixingMatrix = ica.get_mixing_matrix()
print ' done.'
sources = pca.fromZca(sourcesWhite)
# 3. Show independent components and inferred sources
image = Image.fromarray(
tile_raster_images(X=mixingMatrix,
img_shape=(imgDim, imgDim),
tile_shape=(10, 15),
tile_spacing=(1, 1)))
if savedir: image.save(os.path.join(savedir, 'ic_white.png'))
image.show()
image = Image.fromarray(
tile_raster_images(X=mixingMatrix.T,
img_shape=(imgDim, imgDim),
tile_shape=(10, 15),
tile_spacing=(1, 1)))
if savedir: image.save(os.path.join(savedir, 'ic_white.T.png'))
image.show()
image = Image.fromarray(
tile_raster_images(X=pca.fromZca(mixingMatrix),
img_shape=(imgDim, imgDim),
tile_shape=(10, 15),
tile_spacing=(1, 1)))
if savedir: image.save(os.path.join(savedir, 'ic_raw.png'))
image.show()
image = Image.fromarray(
tile_raster_images(X=pca.fromZca(mixingMatrix.T),
img_shape=(imgDim, imgDim),
tile_shape=(10, 15),
tile_spacing=(1, 1)))
if savedir: image.save(os.path.join(savedir, 'ic_raw.T.png'))
image.show()
pyplot.figure()
for ii in range(20):
idx = randIdxSource[ii]
pyplot.subplot(4, 5, ii + 1)
pyplot.title('sourceWhite %d' % idx)
pyplot.hist(sourcesWhite[:, idx])
if savedir: pyplot.savefig(os.path.join(savedir, 'sources_white_hist.png'))
image = Image.fromarray(
tile_raster_images(X=sourcesWhite,
img_shape=(imgDim, imgDim),
tile_shape=(10, 15),
tile_spacing=(1, 1)))
if savedir: image.save(os.path.join(savedir, 'sources_white.png'))
image.show()
# 3.1 plot hist
pyplot.figure()
pyplot.hold(True)
pyplot.title('sources white 20 random dims')
histMax = 0
histMin = 1e10
for ii in range(20):
idx = randIdxSource[ii]
hist, binEdges = histogram(sourcesWhite[:, idx], bins=20, density=True)
histMax = max(histMax, max(hist))
histMin = min(histMin, min(hist[hist != 0])) # min non-zero entry
binMiddles = binEdges[:-1] + (binEdges[1] - binEdges[0]) / 2
#print ' %d from %f to %f' % (ii, min(binMiddles), max(binMiddles))
pyplot.semilogy(binMiddles, hist, '.-')
pyplot.axis('tight')
ax = looser(pyplot.axis(), semilogy=True)
xAbsMax = max(fabs(ax[0:2]))
xx = linspace(-xAbsMax, xAbsMax, 100)
pyplot.semilogy(xx, mlab.normpdf(xx, 0, 1), 'k', linewidth=3)
pyplot.axis((-xAbsMax, xAbsMax, ax[2], ax[3]))
if savedir:
pyplot.savefig(os.path.join(savedir, 'sources_white_log_hist.png'))
# 3.2 plot points
pyplot.figure()
pyplot.hold(True)
pyplot.title('sources white 20 random dims')
nSamples = min(sourcesWhite.shape[0], 1000)
print 'sources_white_log_points plotted with', nSamples, 'samples.'
for ii in range(10):
idx = randIdxWhite[ii]
pyplot.plot(sourcesWhite[:nSamples, idx],
ii + random.uniform(-.25, .25, nSamples), 'o')
pyplot.axis('tight')
if savedir:
pyplot.savefig(os.path.join(savedir, 'sources_white_log_points.png'))
image = Image.fromarray(
tile_raster_images(X=sources,
img_shape=(imgDim, imgDim),
tile_shape=(10, 15),
tile_spacing=(1, 1)))
if savedir: image.save(os.path.join(savedir, 'sources_raw.png'))
image.show()
if savedir:
print 'plots saved in', savedir
else:
import ipdb
ipdb.set_trace()
def test_rbm(learning_rate=0.1,
training_epochs=15,
datasets=None,
batch_size=20,
n_chains=20,
n_samples=14,
output_dir='rbm_plots',
img_dim=28,
n_input=None,
n_hidden=500,
quickHack=False,
visibleModel='binary',
initWfactor=1.0,
imgPlotFunction=None):
'''
Demonstrate how to train an RBM.
This is demonstrated on MNIST.
:param learning_rate: learning rate used for training the RBM
:param training_epochs: number of epochs used for training
:param dataset: path the the pickled dataset
:param batch_size: size of a batch used to train the RBM
:param n_chains: number of parallel Gibbs chains to be used for sampling
:param n_samples: number of samples to plot for each chain
:param visibleModel: 'real' or 'binary'
:param initWfactor: Typicaly 1 for binary or .01 for real
XXX:param pcaDims: None to skip PCA or >0 to use PCA to reduce dimensionality of data first.
'''
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
if quickHack:
train_set_x = train_set_x[:2500, :]
if train_set_y is not None:
train_set_y = train_set_y[:2500]
print('(%d, %d, %d) %d dimensional examples in (train, valid, test)' %
(train_set_x.shape[0], valid_set_x.shape[0], test_set_x.shape[0],
train_set_x.shape[1]))
# compute number of minibatches for training, validation and testing
n_train_batches = train_set_x.shape[0] / batch_size
print 'n_train_batches is', n_train_batches
rng = numpy.random.RandomState(1)
if n_input is None:
n_input = train_set_x.shape[1]
# construct the RBM class
rbm = RBM(nVisible=n_input,
nHidden=n_hidden,
numpyRng=rng,
visibleModel=visibleModel,
initWfactor=initWfactor)
#################################
# Training the RBM #
#################################
print 'starting training.'
if not os.path.isdir(output_dir):
os.makedirs(output_dir)
plotting_time = 0.
start_time = time.clock()
# go through training epochs
meanCosts = []
ii = -1
metrics = array([])
plotEvery = 100
for epoch in xrange(training_epochs):
# go through the training set
for batch_index in xrange(n_train_batches):
#print 'about to train using exemplars %d to %d.' % (batch_index*batch_size, (batch_index+1)*batch_size)
ii += 1
if ii % plotEvery == 0:
plotWeights = '%03i_%05i' % (epoch, batch_index)
calcMetrics = True
else:
plotWeights = False
calcMetrics = False
# metric is xEntropyCost, reconError
metric = rbm.train(
train_set_x[batch_index * batch_size:(batch_index + 1) *
batch_size],
lr=learning_rate,
metrics=calcMetrics,
plotWeights=plotWeights,
output_dir=output_dir)
if calcMetrics:
if len(metrics) == 0:
metrics = array([metric])
else:
metrics = vstack((metrics, metric))
if ii % plotEvery == 0:
# Plot filters after each single step
plotting_start = time.clock()
# Construct image from the weight matrix
image = Image.fromarray(
tile_raster_images(X=imgPlotFunction(rbm.W.T)
if imgPlotFunction else rbm.W.T,
img_shape=(img_dim, img_dim),
tile_shape=(10, 10),
tile_spacing=(1, 1)))
image.save(
os.path.join(
output_dir, 'filters_at_epoch_batch_%03i_%05i.png' %
(epoch, batch_index)))
plotting_stop = time.clock()
plotting_time += (plotting_stop - plotting_start)
#print ' Training epoch %d batch %d, xEntropyCost is ' % (epoch, batch_index), numpy.mean(mean_cost),
print ' Training epoch %d batch %d, xEntropyCost is ' % (
epoch, batch_index), metrics[-1, 0],
print '\trecon error ', metrics[-1, 1]
thisEpochStart = epoch * n_train_batches / plotEvery
thisEpochEnd = (epoch + 1) * n_train_batches / plotEvery
epochMeanXEnt = mean(metrics[thisEpochStart:thisEpochEnd, 0])
epochMeanRecon = mean(metrics[thisEpochStart:thisEpochEnd, 1])
print 'Training epoch %d mean xEntropyCost is ' % (
epoch), epochMeanXEnt, '\trecon error ', epochMeanRecon
meanCosts.append(epochMeanXEnt)
# Plot filters after each training epoch
plotting_start = time.clock()
# Construct image from the weight matrix
image = Image.fromarray(
tile_raster_images(
X=imgPlotFunction(rbm.W.T) if imgPlotFunction else rbm.W.T,
img_shape=(img_dim, img_dim),
tile_shape=(10, 10),
tile_spacing=(1, 1)))
image.save(
os.path.join(output_dir, 'filters_at_epoch_%03i.png' % epoch))
plotting_stop = time.clock()
plotting_time += (plotting_stop - plotting_start)
plotting_start = time.clock()
pyplot.plot(metrics)
pyplot.savefig(os.path.join(output_dir, 'reconErr.png'))
plotting_time += (time.clock() - plotting_start)
end_time = time.clock()
pretraining_time = (end_time - start_time) - plotting_time
print('Training took %f minutes' % (pretraining_time / 60.))
print('Plotting took %f minutes' % (plotting_time / 60.))
#################################
# Plot some samples from RBM #
#################################
# find out the number of test samples
number_of_test_samples = test_set_x.shape[0]
plot_every = 1
# if imgPlotFunction is defined, then also plot before function if
# the data is of the same dimension (e.g. for ZCA, but not for
# PCA).
plotRawAlso = (imgPlotFunction
and train_set_x.shape[0] == img_dim * img_dim)
# create a space to store the image for plotting ( we need to leave
# room for the tile_spacing as well)
image_data = numpy.ones(
((img_dim + 1) * n_samples - 1, (img_dim + 1) * n_chains - 1),
dtype='uint8') * 51 # dark gray
if plotRawAlso:
image_data_raw = numpy.ones(
((img_dim + 1) * n_samples - 1, (img_dim + 1) * n_chains - 1),
dtype='uint8') * 51 # dark gray
for ii in xrange(n_chains):
# generate `plot_every` intermediate samples that we discard, because successive samples in the chain are too correlated
test_idx = rng.randint(number_of_test_samples)
samples = numpy.zeros((n_chains, n_input))
visMean = test_set_x[test_idx, :]
visSample = visMean
for jj in xrange(n_samples):
samples[
jj, :] = visMean # show the mean, but use the sample for gibbs steps
if jj == n_samples - 1: break # skip the last for speed
plot_every = 2**jj # exponentially increasing number of gibbs samples. max for n_samples=14 is 2^12
for ss in xrange(plot_every):
visMean, visSample = rbm.gibbs_vhv(visSample)[
4:6] # 4 for mean, 5 for sample
print ' ... plotting sample ', ii
image_data[:, (img_dim + 1) * ii:(img_dim + 1) * ii +
img_dim] = tile_raster_images(X=imgPlotFunction(samples) if
imgPlotFunction else samples,
img_shape=(img_dim, img_dim),
tile_shape=(n_samples, 1),
tile_spacing=(1, 1))
if plotRawAlso:
image_data_raw[:, (img_dim + 1) * ii:(img_dim + 1) * ii +
img_dim] = tile_raster_images(X=samples,
img_shape=(img_dim,
img_dim),
tile_shape=(n_samples,
1),
tile_spacing=(1, 1))
image = Image.fromarray(image_data)
image.save(os.path.join(output_dir, 'samples.png'))
if plotRawAlso:
image = Image.fromarray(image_data)
image.save(os.path.join(output_dir, 'samplesRaw.png'))
saveToFile(os.path.join(output_dir, 'rbm.pkl.gz'), rbm)
return rbm, meanCosts