def plot_multiple_vectors_as_images(dVectors, title=None):
""" Based on dictionary dVectors, multiple vectors are
plotted next to each other. """
numbVectors = len(dVectors)
# init figure
fig = plt.figure(figsize=(15, 4))
grid = axes_grid1.AxesGrid(fig,
111,
nrows_ncols=(1, numbVectors),
axes_pad=0.55,
share_all=True,
cbar_location="right",
cbar_mode="each",
cbar_size="5%",
cbar_pad="2%",)
if title is not None:
fig.suptitle(title)
# plot vectors from dVectors
for idx, (key, relVal) in enumerate(dVectors.items()):
im = grid[idx].imshow(render.vec2im(relVal), cmap='jet')
grid[idx].set_title(key)
grid[idx].axis('off')
grid.cbar_axes[idx].colorbar(im)
# finalize figure
if title is None:
fig.tight_layout()
else:
fig.tight_layout(rect=[0, 0, 0.95, .95])
def unique_shapes():
"""Finds the unique shapes w.r.t. rotation. """
# check whether calling function is appropriate
errorMessage = "Only meant for TrianglesAndSquares dataset"
assert settings.dataName == 'TrianglesAndSquares', errorMessage
dirPath = os.path.dirname(os.path.realpath(__file__))
#fileName = dirPath + r"\results_tools\uniqueShapes.pickle"
fileName = os.path.join(dirPath, "results_tools", "uniqueShapes.pickle")
try:
return pickle.load(open(fileName, "rb"), encoding='latin1') # the encoding was needed in order to load a Python 2.7 pickle in Python 3.6. Maybe this is not necessary anymore at a later stage.
except (OSError, IOError):
# load data
X, Y = data_loader.load_data()
imageDim = len(X['train'][0])
uniqueShapeRotSq = np.empty((0, imageDim))
uniqueShapeRotTr = np.empty((0, imageDim))
for idx in range(len(X['train'])):
x = X['train'][[idx]]
y = Y['train'][[idx]]
# find xMid in the middle of the image of x
xImage = render.vec2im(x)
halfIdx = int(math.floor(len(xImage)/2))
xMid = xImage[halfIdx, halfIdx]
# set shape color to 1 and background color to 0
x[0][x[0] == xMid] = 1
x[0][np.logical_not(x[0] == 1)] = 0
if y[0][0] == 1:
# a square
considered = any(np.equal(uniqueShapeRotSq, x[0]).all(1))
if not considered:
uniqueShapeRotSq = np.vstack([uniqueShapeRotSq, x[0]])
else:
# a triangle
considered = any(np.equal(uniqueShapeRotTr, x[0]).all(1))
if not considered:
uniqueShapeRotTr = np.vstack([uniqueShapeRotTr, x[0]])
# sort results
uniqueShapeRotSq = uniqueShapeRotSq[np.argsort([min(np.flatnonzero(im)) for im in uniqueShapeRotSq])]
uniqueShapeRotTr = uniqueShapeRotTr[np.argsort([min(np.flatnonzero(im)) for im in uniqueShapeRotTr])]
# save results
uniqShapeRot = {'square': uniqueShapeRotSq,
'triangle': uniqueShapeRotTr}
with open(fileName, 'wb') as handle:
pickle.dump(uniqShapeRot, handle, protocol=pickle.HIGHEST_PROTOCOL)
return uniqShapeRot
def union_shapes():
"""Returns the union of all squares and triangle, respectively. """
# check whether calling function is appropriate
errorMessage = "Only meant for TrianglesAndSquares dataset"
assert settings.dataName == 'TrianglesAndSquares', errorMessage
dirPath = os.path.dirname(os.path.realpath(__file__))
fileName = dirPath + r"\results_tools\unionShapes.pickle"
try:
return pickle.load(open(fileName, "rb"))
except (OSError, IOError):
# load data
X, Y = data_loader.load_data()
imageDim = len(X['train'][0])
# init images of inner circles
unionSq = np.zeros((imageDim,))
unionTr = np.zeros((imageDim,))
for idx in range(len(X['train'])):
x = X['train'][[idx]]
y = Y['train'][[idx]]
# find xMid in the middle of the image of x
xImage = render.vec2im(x)
halfIdx = int(math.floor(len(xImage)/2))
xMid = xImage[halfIdx, halfIdx]
# set shape color to 1 and background color to 0
x[0][x[0] == xMid] = 1
x[0][np.logical_not(x[0] == 1)] = 0
if y[0][0] == 1:
# a square
unionSq += x[0]
else:
# a triangle
unionTr += x[0]
# rescale result
unionSq[unionSq > 0] = 1
unionTr[unionTr > 0] = 1
# save results
res = (unionSq, unionTr)
with open(fileName, 'wb') as handle:
pickle.dump(res, handle, protocol=pickle.HIGHEST_PROTOCOL)
# plot shapes
plot_vector_as_image(unionSq)
plot_vector_as_image(unionTr)
return res
def plot_vector_as_image(vector, title=None):
""" Plots a vector as image with colorbar. """
fig = plt.figure()
ax = fig.add_subplot(111)
if title is not None:
ax.set_title(title)
cax = ax.matshow(render.vec2im(vector))
fig.colorbar(cax)
ax.scatter(train_features[1].numpy()[filter][:, 0],
train_features[1].numpy()[filter][:, 1],
marker=label_symbols[label_idx],
alpha=0.3)
ax.legend()
ax.grid(True)
# inspect weights to first layer
import mpl_toolkits.axes_grid1 as axes_grid1
import render
W1 = model.layers[0].weights[0].numpy()
B = model.layers[0].weights[1].numpy()
dim = int(W1.shape[1])
W11 = W1[:, 0]
W12 = W1[:, 1]
W11Im = render.vec2im(W11)
W12Im = render.vec2im(W12)
# init figure
fig = plt.figure()
grid = axes_grid1.AxesGrid(
fig,
111,
nrows_ncols=(1, dim),
axes_pad=0.55,
share_all=True,
cbar_location="right",
cbar_mode="each",
cbar_size="5%",
cbar_pad="2%",
)
def plot_NN_layer(self, layer_idx, nr_to_plot=None):
"""Plots of nr_to_plot nearest neighbors (NN) of a layer in a grid. """
if nr_to_plot is None:
nr_to_plot = self.k + 1 # plot input data and all neighbors found
# init data
NN = self.NN_layers[layer_idx]
confirmation_perc_NN_pred = 100 * np.mean(
np.argmax(NN['Y'], 1) == np.argmax(self.Y_input_predict, 1)[0])
confirmation_perc_NN_true_label = 100 * np.mean(
np.argmax(NN['Y'], 1) == np.argmax(self.Y_input, 1)[0])
# init figure
nr_of_rows = math.ceil(math.sqrt(nr_to_plot))
fig, ax = plt.subplots(nr_of_rows,
nr_of_rows,
sharex='col',
sharey='row',
figsize=(16, 8))
fig.suptitle(
f'Plot of nearest neighbors (NN) in layer {NN["layer name"]} with {round(confirmation_perc_NN_pred,2)}% and {round(confirmation_perc_NN_true_label,2)}% of NN confirm with prediction and true label, resp.'
)
# plot input image and all nearest neighbors
nr_NN_plotted = 0
for i in range(nr_of_rows):
for j in range(nr_of_rows):
# format subplot
ax[i, j].set_xticklabels([])
ax[i, j].set_yticklabels([])
if nr_NN_plotted + 1 >= nr_to_plot:
continue # all intended images plotted
if i + j == 0:
# plot input image
cax = ax[0, 0].imshow(render.vec2im(self.X_input),
vmin=0,
vmax=1)
# ax[0, 0].matshow(render.vec2im(self.X_input))
ax[0, 0].set_title(
f'Input image; label {self.Y_input}; predict {self.Y_input_predict.round(2)}',
fontsize=8)
continue
ax[i, j].imshow(render.vec2im(NN['X'][nr_NN_plotted]),
vmin=0,
vmax=1)
# ax[i, j].matshow(render.vec2im(NN['X'][nr_NN_plotted]))
ax[i, j].set_title(
f'NN {nr_NN_plotted}; label = {NN["Y"][nr_NN_plotted]}; distance = {round(NN["distances"][nr_NN_plotted], 4)}',
fontsize=8)
nr_NN_plotted += 1
# finalize figure
fig.subplots_adjust(right=0.8)
# put colorbar at desire position
cbar_ax = fig.add_axes([0.85, 0.15, 0.05, 0.7])
fig.colorbar(cax, cax=cbar_ax)