# -*- coding: utf-8 -*-
# ELEKTRONN - Neural Network Toolkit
#
# Copyright (c) 2014 - now
# Max-Planck-Institute for Medical Research, Heidelberg, Germany
# Authors: Marius Killinger, Gregor Urban
"""
Supplementary functions to plot various CNN states
"""
import numpy as np
import matplotlib.pyplot as plt
[docs]def plotFilters(cnn, layer=0, channel=None, normalize=False, savename='filters_layer0.png'):
W = cnn.layers[layer].W.get_value()
if W.shape[1] == 3: # RGB
r = embedMatricesInGray(W[:, 0, :, :], normalize=normalize)
g = embedMatricesInGray(W[:, 1, :, :], normalize=normalize)
b = embedMatricesInGray(W[:, 2, :, :], normalize=normalize)
mat = np.dstack((r, g, b))
lower = min(0, mat.min())
spread = mat.max() - lower
mat = mat - lower
mat = 255.0 / spread * mat
mat = mat.astype('uint8')
elif W.shape[1] == 1 or channel is None: # gray-scale
W = W[:, 0, :, :]
mat = embedMatricesInGray(W, normalize=normalize)
elif W.shape[1] == 1 or channel is not None: # gray-scale
W = W[:, channel, :, :]
mat = embedMatricesInGray(W, normalize=normalize)
plt.figure()
plt.imshow(mat, interpolation='None', cmap='gray') # Note that cmap is ignored by matplotlib for RGB-img
plt.show()
print "Saving filter image as as %s" % (savename)
plt.savefig(savename, bbox_inches='tight')
[docs]def showActivations(cnn, data, show_first_class_prob=False, no_show=False):
"""
Plots activation maps given data. It requires that cnn.debug_functions contains a list of functions
that return the activations (i.e. cnn.compileDebugFunctions must have been called
Parameters
----------
cnn:
instance of MixedConvNN
data:
input to cnn for which activations should be shown
show_first_class_prob:
True/False whether to additionally show the probability map for the first class
no_show:
True/False whether to pop up plots or silently return a list of image arrays
"""
n = len(cnn.layers)
all_act = []
myfig = plt.figure(figsize=(15, 14))
if len(cnn.debug_functions) == 0:
print "No debug functions found in cnn! Compile them first"
for layer, dbgf, i in zip(cnn.layers, cnn.debug_functions, range(n)):
activity = dbgf(data)[
0
] # only first image in batch (nof, x-dim, y-dim) or (z-dim, nof, x-dim, y-dim)
sp = activity.shape
if len(sp) != 3:
if len(sp) == 4:
activity = activity.reshape(sp[0] * sp[1], sp[2], sp[3]) #np.swapaxes(np.swapaxes(activity,2,0),2,4)
else:
continue
if len(sp) == 3:
image = embedMatricesInGray(activity, 1, True)
elif len(sp) == 5:
image = embedMatricesInGray(activity, 1, True, fixed_n_horizontal=sp[1])
all_act.append(image)
showMultipleFiguresAdd(myfig, n, i, image, "Layer-" + str(i) + " Activity")
plt.tight_layout()
if show_first_class_prob:
prob_pred = cnn.class_probabilities(data)
plt.figure()
plt.gray()
plt.title('prediction map of first class')
plt.imshow(prob_pred[:, 0].reshape(np.sqrt(np.shape(prob_pred)[0]), -1), interpolation='nearest')
if not no_show:
plt.show()
return all_act
[docs]def showParamHistogram(cnn, no_show=False, onlyW=True):
"""
Plots histograms of parameter/weight values.
Parameters
----------
:type no_show: object
cnn:
instance of MixedConvNN
onlyW:
True/False whether to ignore the biases b
no_show:
True/False whether to pop up plots or silently return a list of image arrays
"""
n = len(cnn.layers)
for layer, i in zip(cnn.layers, range(n)):
for par in layer.params if onlyW == False else [layer.W]:
pa = par.get_value()
print 'Layer {0:2d}: mean={1:3.6e}, median={2:3.6e}, mean(abs)={3:2.6e}, median(abs)={4:2.6e}'.format(
1 + i, np.mean(pa), np.median(pa), np.mean(abs(pa)), np.median(abs(pa)))
print 'Layer ' + str(i + 1) + ': filter_shape =', np.shape(pa)
hist, bins = np.histogram(pa.flatten(), bins=100)
width = 0.7 * (bins[1] - bins[0])
center = (bins[:-1] + bins[1:]) / 2
plt.figure()
plt.bar(center, hist, align='center', width=width)
plt.xlabel(str(np.shape(pa)) + ' parameters ', fontsize=16)
plt.ylabel('# (Layer ' + str(i + 1) + ')' + str(), fontsize=16)
if not no_show:
plt.show()
[docs]def showActivityHistogram(cnn, data, no_show=False):
"""
Plots histograms of activation maps given data. It requires that cnn.debug_functions contains a list
of functions that return the activations (i.e. cnn.compileDebugFunctions must have been called
Parameters
----------
cnn:
instance of MixedConvNN
data:
input to cnn for which activations should be shown
no_show:
True/False whether to pop up plots or silently return a list of image arrays
"""
n = len(cnn.layers)
for layer, dbgf, i in zip(cnn.layers, cnn.debug_functions, range(n)):
activity = dbgf(data).flatten()
hist, bins = np.histogram(activity, bins=100)
width = 0.7 * (bins[1] - bins[0])
center = (bins[:-1] + bins[1:]) / 2
plt.figure()
plt.bar(center, hist, align='center', width=width)
plt.xlabel(str(len(activity)) + ' activations ', fontsize=16)
plt.ylabel('# (Layer ' + str(i + 1) + ')' + str(), fontsize=16)
if not no_show:
plt.show()
[docs]def embedMatricesInGray(mat, border_width=1, normalize=False, output_ratio=1.7, fixed_n_horizontal=0):
"""
Creates a big matrix out of smaller ones (mat) assumed format of mat:(index, ch, i_vert,i_horiz)
"""
# sh = np.shape(mat)
# if sh[1]==3: # interpret 3 channels as rgb
# mat = np.swapaxes(mat, 1, 3) # swap RGB channels at back
# if sh[1]==1: # work directly on this channel
# mat = mat[:,0]
# else: # create an image for each channel
# channels = []
# for i in xrange(sh[1]):
# ch = embedMatricesInGray(mat[:,i], border_width, normalize, output_ratio, fixed_n_horizontal)
# channels.append(ch)
# return channels
sh = np.shape(mat)
assert len(sh) == 3
n = sh[0]
if fixed_n_horizontal > 0:
nhor = fixed_n_horizontal
else:
nhor = int(np.sqrt(n * output_ratio)) # aim: ratio 16:9
nvert = int(n * 1.0 / nhor + 1) #warning: too big: nvert*nhor >= n
ret = np.zeros((nvert * (border_width + sh[1]), nhor * (border_width + sh[2])), dtype=np.float32)
if normalize:
maxs = [np.max(mat[i, :, :]) + 1e-8 for i in range(n)]
mins = [np.min(mat[i, :, :]) for i in range(n)]
else:
maxs = [1] * n
mins = [0] * n
for j in range(nvert):
for i in range(nhor):
if i + j * nhor >= n:
return ret
#print (j*(border_width+sh[1]),j*(border_width+sh[1])+sh[1], i*(border_width+sh[2]),i*(border_width+sh[2])+sh[2])
#print shape(mat[i+j*nhor,:,:])
ret[j * (border_width + sh[1]):j * (border_width + sh[1]) + sh[1],
i * (border_width + sh[2]):i * (border_width + sh[2]) + sh[2]
] = (mat[i + j * nhor, :, :] - mins[i + j * nhor]) / (maxs[i + j * nhor] - mins[i + j * nhor])
return ret
