class GenerateAndSave(TrainExtension):
"""
Keeps track of what the generator in a (vanilla) GAN returns for a
particular set of noise values.
"""
def __init__(self, generator, save_prefix, batch_size=20, grid_shape=(5, 4)):
assert isinstance(generator, Generator)
self.batch_sym = T.matrix('generate_batch')
self.generate_f = theano.function([self.batch_sym],
generator.dropout_fprop(self.batch_sym)[0])
self.batch = generator.get_noise(batch_size).eval()
self.save_prefix = save_prefix
self.patch_viewer = PatchViewer(grid_shape=grid_shape, patch_shape=(32, 32),
is_color=True)
def on_monitor(self, model, dataset, algorithm):
samples = self.generate_f(self.batch).swapaxes(0, 3)
self.patch_viewer.clear()
for sample in samples:
self.patch_viewer.add_patch(sample, rescale=True)
fname = self.save_prefix + '.%05i.png' % model.monitor.get_epochs_seen()
self.patch_viewer.save(fname)
def plot(w):
nblocks = int(model.n_g / model.sparse_gmask.bw_g)
filters_per_block = model.sparse_gmask.bw_g * model.sparse_hmask.bw_h
block_viewer = PatchViewer((model.sparse_gmask.bw_g, model.sparse_hmask.bw_h),
(opts.height, opts.width),
is_color = opts.color,
pad=(2,2))
chan_viewer = PatchViewer(get_dims(nblocks),
(block_viewer.image.shape[0],
block_viewer.image.shape[1]),
is_color = opts.color,
pad=(5,5))
main_viewer = PatchViewer(get_dims(nplots),
(chan_viewer.image.shape[0],
chan_viewer.image.shape[1]),
is_color = opts.color,
pad=(10,10))
topo_shape = [opts.height, opts.width, opts.chans]
view_converter = DefaultViewConverter(topo_shape)
if opts.splitblocks:
os.makedirs('filters/')
for chan_i in xrange(nplots):
viewer_dims = slice(0, None) if opts.color else chan_i
for bidx in xrange(nblocks):
for fidx in xrange(filters_per_block):
fi = bidx * filters_per_block + fidx
topo_view = view_converter.design_mat_to_topo_view(w[fi:fi+1,:])
try:
block_viewer.add_patch(topo_view[0,:,:,viewer_dims])
except:
import pdb; pdb.set_trace()
if opts.splitblocks:
pl.imshow(block_viewer.image, interpolation='nearest')
pl.axis('off')
pl.title('Wv - block %i, chan %i' % (bidx, chan_i))
pl.savefig('filters/filters_chan%i_block%i.png' % (bidx, chan_i))
chan_viewer.add_patch(block_viewer.image[:,:,viewer_dims] - 0.5)
block_viewer.clear()
main_viewer.add_patch(chan_viewer.image[:,:,viewer_dims] - 0.5)
chan_viewer.clear()
return copy.copy(main_viewer.image)
else:
new_w = w_di
else:
new_w = numpy.zeros((len(w_di), opts.height * opts.width)) if di else w_di
for fi in xrange(len(w_di)):
if opts.k != -1:
# build "new_w" as a linear combination of "strongest" filters in layer below
if di > 0:
temp.fill(0.)
idx = numpy.argsort(w_di[fi])[-opts.k:]
for fi_m1 in idx:
new_w[fi:fi+1] += w_di[fi, fi_m1] * prev_w[fi_m1:fi_m1+1,:]
#for fi_m1 in xrange(len(w_di[fi])):
else:
temp = w_di[fi:fi+1,:]
topo_view = view_converter.design_mat_to_topo_view(new_w[fi:fi+1])
block_viewer.add_patch(topo_view[0])
main_viewer.add_patch(block_viewer.image[:,:,0] - 0.5)
block_viewer.clear()
prev_w = new_w
pl.imshow(main_viewer.image, interpolation=None)
pl.axis('off');
pl.savefig('weights.png')
pl.show()
# positive weights
posw = copy.copy(W[1])
posw[posw < 0] = 0.
probs = posw / numpy.max(posw, axis=1)[:, None]
r = numpy.random.random(probs.shape)
pos_w1 = numpy.dot(posw * (probs > r), W[0])
# negative weights
negw = copy.copy(-W[1])
negw[negw < 0] = 0.
probs = negw / numpy.max(negw, axis=1)[:, None]
r = numpy.random.random(probs.shape)
neg_w1 = numpy.dot(-negw * (probs > r), W[0])
block_viewer = make_viewer(pos_w1 + neg_w1, get_dims(max_filters),
(opts.height, opts.width))
main_viewer.add_patch(block_viewer.image[:, :, 0] - 0.5)
pl.imshow(main_viewer.image, interpolation=None)
pl.axis('off')
pl.savefig('weights.png')
if k == 0:
pl.show()
else:
pl.draw()
block_viewer.clear()
main_viewer.clear()
posw[posw < 0] = 0.
probs = posw / numpy.max(posw, axis=1)[:, None]
r = numpy.random.random(probs.shape)
pos_w1 = numpy.dot(posw * (probs > r), W[0])
# negative weights
negw = copy.copy(-W[1])
negw[negw < 0] = 0.
probs = negw / numpy.max(negw, axis=1)[:, None]
r = numpy.random.random(probs.shape)
neg_w1 = numpy.dot(-negw * (probs > r), W[0])
block_viewer = make_viewer(
pos_w1 + neg_w1,
get_dims(max_filters),
(opts.height, opts.width))
main_viewer.add_patch(block_viewer.image[:,:,0] - 0.5)
pl.imshow(main_viewer.image, interpolation=None)
pl.axis('off');
pl.savefig('weights.png')
if k == 0:
pl.show()
else:
pl.draw()
block_viewer.clear()
main_viewer.clear()
