def test_dataset_to_pen(self):
inputDataset = TestData.getDataset()
outputDataset = ConvertedDataset(inputDataset)
for sampleId, rawSample in enumerate(inputDataset.samples):
sample = inputDataset.undo_preprocess(rawSample)
penPositions = sample_to_penpositions(
sample, inputDataset.char_labels[sampleId],
inputDataset.eoc_labels[sampleId],
inputDataset.bow_labels[sampleId])
outputDataset.addSample(penPositions, inputDataset.texts[sampleId])
outputDataset.applyPreProcessing()
with tempfile.TemporaryDirectory() as tmpDir:
tmpFile = os.path.join(tmpDir, 'dataset.npz')
outputDataset.save(tmpFile)
verificationDataset = HandWritingDatasetConditional(tmpFile)
for sampleId, _ in enumerate(inputDataset.samples):
inputSample = inputDataset.samples[
0] * inputDataset.norm_std + inputDataset.norm_mean
inputSample *= inputDataset.scale_max - inputDataset.scale_min
verificationSample = (verificationDataset.samples[0] *
verificationDataset.norm_std +
verificationDataset.norm_mean)
verificationSample *= verificationDataset.scale_max - verificationDataset.scale_min
inputSample[:, 2] = inputDataset.samples[0][:, 2]
verificationSample[:, 2] = verificationDataset.samples[0][:, 2]
np.testing.assert_array_almost_equal(inputSample,
verificationSample,
decimal=4)
np.testing.assert_array_equal(
inputDataset.char_labels[sampleId],
verificationDataset.char_labels[sampleId])
np.testing.assert_array_almost_equal(
inputDataset.eoc_labels[sampleId],
verificationDataset.eoc_labels[sampleId])
np.testing.assert_array_almost_equal(
inputDataset.bow_labels[sampleId],
verificationDataset.bow_labels[sampleId])
self.assertEqual(inputDataset.texts[sampleId],
verificationDataset.texts[sampleId])
np.testing.assert_array_equal(inputDataset.alphabet,
verificationDataset.alphabet)
def renderDataset(inputDataset, outputFolder):
if not os.path.exists(outputFolder):
os.makedirs(outputFolder)
preprocessingIsIncorrect = find_scaling_errors(inputDataset)
for sampleId, rawSample in enumerate(inputDataset.samples):
if sampleId % 50 == 0:
print()
sys.stdout.write('Rendering sample ' + str(sampleId) + ' / ' +
str(len(inputDataset.samples)) + ' ')
sys.stdout.flush()
else:
sys.stdout.write('.')
sys.stdout.flush()
sample = inputDataset.undo_preprocess(rawSample)
if preprocessingIsIncorrect[sampleId]:
sample = fix_scaling_error(sample, inputDataset.scale_max,
inputDataset.scale_min)
penPositions = sample_to_penpositions(
sample, inputDataset.char_labels[sampleId],
inputDataset.eoc_labels[sampleId],
inputDataset.bow_labels[sampleId])
skeletonImage, skeletonCharImage, skeletonEocImage, skeletonBowImage, skeletonMetadata = penpositions_to_skeletonimages(
penPositions)
img = Image.fromarray(255 - skeletonImage.astype('uint8') * 255,
mode='L')
img.save(os.path.join(outputFolder, str(sampleId) + '.png'), 'PNG')
text = penpositions_to_text(penPositions)
with open(os.path.join(outputFolder,
str(sampleId) + '.txt'), 'w') as fil:
fil.write(text)
def test_pen_to_strokes(self):
inputDataset = TestData.getDataset()
outputDataset = ConvertedDataset(inputDataset)
for sampleId, rawSample in enumerate(inputDataset.samples):
sample = inputDataset.undo_preprocess(rawSample)
penPositions = sample_to_penpositions(
sample, inputDataset.char_labels[sampleId],
inputDataset.eoc_labels[sampleId],
inputDataset.bow_labels[sampleId])
strokes = penpositions_to_strokes(penPositions)
penPositions2 = strokes_to_penpositions(strokes)
self.assertEqual(len(penPositions), len(penPositions2))
penPos1 = [pos1.pos for pos1 in penPositions]
penPos2 = [pos2.pos for pos2 in penPositions2]
np.testing.assert_array_almost_equal(penPos1, penPos2)
penUp1 = [pos1.penUp for pos1 in penPositions[:-1]
] # don't test last penUp, seems to be random
penUp2 = [pos2.penUp for pos2 in penPositions2[:-1]]
np.testing.assert_array_almost_equal(penUp1, penUp2)
penChar1 = [pos1.charLabel for pos1 in penPositions[:]]
penChar2 = [pos2.charLabel for pos2 in penPositions2[:]]
np.testing.assert_array_equal(penChar1, penChar2)
penEoc1 = [pos1.eocLabel for pos1 in penPositions[:]]
penEoc2 = [pos2.eocLabel for pos2 in penPositions2[:]]
np.testing.assert_array_almost_equal(penEoc1, penEoc2)
penBow1 = [pos1.bowLabel for pos1 in penPositions[:]]
penBow2 = [pos2.bowLabel for pos2 in penPositions2[:]]
np.testing.assert_array_almost_equal(penBow1, penBow2)
def convertDataset(inputDataset, DRAW_STEPS=False):
outputDataset = createOutputDataset(inputDataset)
preprocessingIsIncorrect = find_scaling_errors(inputDataset)
for sampleId, rawSample in enumerate(inputDataset.samples):
if sampleId % 50 == 0:
print()
sys.stdout.write('Adding sample ' + str(sampleId) + ' / ' +
str(len(inputDataset.samples)) + ' ')
sys.stdout.flush()
else:
sys.stdout.write('.')
sys.stdout.flush()
sample = inputDataset.undo_preprocess(rawSample)
if preprocessingIsIncorrect[sampleId]:
sample = fix_scaling_error(sample, inputDataset.scale_max,
inputDataset.scale_min)
penPositions = sample_to_penpositions(
sample, inputDataset.char_labels[sampleId],
inputDataset.eoc_labels[sampleId],
inputDataset.bow_labels[sampleId])
skeletonImage, skeletonCharImage, skeletonEocImage, skeletonBowImage, skeletonMetadata = penpositions_to_skeletonimages(
penPositions)
if False:
figure = plt.figure('SkeletonImages')
plt.subplot(4, 1, 1)
plt.imshow(skeletonImage)
plt.subplot(4, 1, 2)
plt.imshow(skeletonCharImage,
cmap='nipy_spectral',
vmin=40,
vmax=70)
plt.subplot(4, 1, 3)
plt.imshow(skeletonEocImage, cmap='nipy_spectral')
plt.subplot(4, 1, 4)
plt.imshow(skeletonBowImage, cmap='nipy_spectral')
plt.show()
exit(1)
thinnedImage = skeletonize(skeletonImage)
graph = skeleton_to_graph(thinnedImage)
if DRAW_STEPS:
print("Drawing ...")
figure = plt.figure("Graphs")
plt.subplot(4, 1, 1)
plt.imshow(thinnedImage, cmap='binary', vmax=10)
graph.plot()
resolve_strokes(graph)
strokes = graph_to_strokes(graph)
strokes.sort()
annotateStrokes(strokes, skeletonCharImage, skeletonEocImage,
skeletonBowImage)
smoothStrokes = resample_strokes_smooth(strokes)
if DRAW_STEPS:
plt.subplot(4, 1, 2)
plt.imshow(thinnedImage, cmap='binary', vmax=10)
graph.plot()
plt.subplot(4, 1, 3)
plt.imshow(thinnedImage, cmap='binary', vmax=10)
strokes.plot()
plt.subplot(4, 1, 4)
plt.imshow(thinnedImage, cmap='binary', vmax=10)
smoothStrokes.plot()
plt.show()
exit(1)
fakePenPositions = strokes_to_penpositions(smoothStrokes)
addSampleToDataset(outputDataset, fakePenPositions,
inputDataset.texts[sampleId])
if False:
figure = plt.figure('PenPositionsImages')
plt.subplot(2, 1, 1)
plt.imshow(thinnedImage, cmap='binary', vmax=10)
smoothStrokes.plot()
plt.subplot(2, 1, 2)
plt.imshow(skeletonImage, cmap='binary', vmax=10)
currentStrokeX = list()
currentStrokeY = list()
for penPosition in fakePenPositions:
currentStrokeX.append(penPosition.pX)
currentStrokeY.append(penPosition.pY)
if penPosition.penUp:
plt.plot(currentStrokeX, currentStrokeY, '.-')
currentStrokeX = list()
currentStrokeY = list()
plt.show()
exit(1)
if False:
sample = outputDataset.get('samples')[0]
penPositions = sample_to_penpositions(
sample,
outputDataset.get('char_labels')[sampleId],
outputDataset.get('eoc_labels')[sampleId],
outputDataset.get('sow_labels')[sampleId])
skeletonImage, skeletonCharImage, skeletonEocImage, skeletonBowImage, skeletonMetadata = penpositions_to_skeletonimages(
penPositions)
figure = plt.figure('OutputSkeletonImages')
plt.subplot(4, 1, 1)
plt.imshow(skeletonImage)
plt.subplot(4, 1, 2)
plt.imshow(skeletonCharImage,
cmap='nipy_spectral',
vmin=40,
vmax=70)
plt.subplot(4, 1, 3)
plt.imshow(skeletonEocImage, cmap='nipy_spectral')
plt.subplot(4, 1, 4)
plt.imshow(skeletonBowImage, cmap='nipy_spectral')
plt.show()
exit(1)
print()
return outputDataset
def convertDataset(inputDataset, resample=False):
outputDataset = ConvertedDataset(inputDataset)
preprocessingIsIncorrect = find_scaling_errors(inputDataset)
for sampleId, rawSample in enumerate(inputDataset.samples):
if sampleId % 50 == 0:
print()
sys.stdout.write('Adding sample ' + str(sampleId) + ' / ' +
str(len(inputDataset.samples)) + ' ')
sys.stdout.flush()
else:
sys.stdout.write('.')
sys.stdout.flush()
sample = inputDataset.undo_preprocess(rawSample)
if preprocessingIsIncorrect[sampleId]:
sample = fix_scaling_error(sample, inputDataset.scale_max,
inputDataset.scale_min)
penPositions = sample_to_penpositions(
sample, inputDataset.char_labels[sampleId],
inputDataset.eoc_labels[sampleId],
inputDataset.bow_labels[sampleId])
strokes = penpositions_to_strokes(penPositions)
skeletonImage, _, _, _, _ = penpositions_to_skeletonimages(
penPositions)
if False:
strokeAccelerations = analyse_strokes_acceleration(strokes)
figure = plt.figure('StrokeAccelerations')
plt.hist(strokeAccelerations, bins=30)
plt.show()
#exit(1)
smoothStrokes = resample_strokes_smooth(strokes)
fakePenPositions = strokes_to_penpositions(smoothStrokes)
outputDataset.addSample(fakePenPositions, inputDataset.texts[sampleId])
if True:
figure = plt.figure('PenPositionsImages')
plt.subplot(3, 1, 1)
plt.imshow(skeletonImage, cmap='binary', vmax=10)
strokes.plot()
plt.subplot(3, 1, 2)
plt.imshow(skeletonImage, cmap='binary', vmax=10)
smoothStrokes.plot()
plt.subplot(3, 1, 3)
plt.imshow(skeletonImage, cmap='binary', vmax=10)
currentStrokeX = list()
currentStrokeY = list()
for penPosition in fakePenPositions:
currentStrokeX.append(penPosition.pos[0])
currentStrokeY.append(penPosition.pos[1])
if penPosition.penUp:
plt.plot(currentStrokeX, currentStrokeY, '.-')
currentStrokeX = list()
currentStrokeY = list()
if currentStrokeX:
plt.plot(currentStrokeX, currentStrokeY, '.-')
plt.show()
exit(1)
print()
return outputDataset