class ToCharacters(object):
segmenter = Segmenter()
def __init__(self):
'''
constructor
'''
self.segmenter = Segmenter()
def segment(self, image):
'returns Number of available character segments in a line'
return self.segmenter.segmentHorizontally(image)
#returns the same but manipulated line Image
def splitDika(self, image):
segmentLine = ToLines()
totalLines = segmentLine.segment(image)
pixelCounter = PixelCounter()
for lineNum in range(totalLines):
lineImg = segmentLine.getLineSegment(lineNum)
lineRowRange = segmentLine.getCoordinatesOfLineSegment(lineNum)
countArr = pixelCounter.getCountArr_V(lineImg, 0)
#print countArr
min = pixelCounter.getMin(countArr)
#dika width is min vertically and it is most occurred horizontally
for x in pixelCounter.getOccurredPositionList(countArr, min):
for height in range(lineRowRange[0],
(lineRowRange[1] + lineRowRange[0]) / 2):
cv.Set2D(image, height, x, 255)
return image
def learn(data, K, resolution, epoch, device, checkpoint=None):
"""
Data is a tensor ith size (len(dataset), number of colours,resolution[0], resolution[1])
"""
model = Segmenter(resolution, K, device)
optimizer = optim.Adam(model.parameters(), lr=LR)
if checkpoint != None:
model.load_state_dict(checkpoint["model"])
optimizer.load_state_dict(checkpoint["optimizer"])
dataloader = torch.utils.data.DataLoader(data,
BATCH,
shuffle=True,
drop_last=True)
for i in tqdm(range(epoch), desc="Epochs"):
for batch in tqdm(dataloader, desc="Batchs"):
model.zero_grad()
Uenc = model(batch, True)
Nloss = Ncutsoft(Uenc, batch).to(device)
print("Loss on this batch: {}".format(float(Nloss)))
print("Accuracy on this batch : {}%".format(
round(100 * (1 - float(Nloss) / K), 2)))
Nloss.backward()
optimizer.step()
model.zero_grad()
Udec = model(batch, False)[1]
Jloss = Jrecons(Udec, batch).to(device)
Jloss.backward()
optimizer.step()
torch.save(
{
"model": model.state_dict(),
"optimizer": optimizer.state_dict()
}, path_saved_models + '{}epoch'.format(i))
return (model, optimizer)
class ToWords(object):
segmenter=Segmenter()
def __init__(self):
'''
constructor
'''
self.segmenter=Segmenter()
def segment(self,image):
transImg=cv.CreateImage((image.height,image.width), cv.IPL_DEPTH_8U, 1)
cv.Transpose(image, transImg)
'returns Number of available line segments'
return self.segmenter.segmentHorizontally(transImg)
def getWordSegment(self,wordNumber):
'returns line segment Image'
image=self.segmenter.getSegment(wordNumber)
return image
def getReadableWordSegment(self,wordNumber):
'returns transposed form of line segment Image'
image=self.segmenter.getSegment(wordNumber)
transImg=cv.CreateImage((image.height,image.width), cv.IPL_DEPTH_8U, 1)
cv.Transpose(image, transImg)
return transImg
class ToWords(object):
segmenter = Segmenter()
def __init__(self):
'''
constructor
'''
self.segmenter = Segmenter()
def segment(self, image):
transImg = cv.CreateImage((image.height, image.width), cv.IPL_DEPTH_8U,
1)
cv.Transpose(image, transImg)
'returns Number of available line segments'
return self.segmenter.segmentHorizontally(transImg)
def getWordSegment(self, wordNumber):
'returns line segment Image'
image = self.segmenter.getSegment(wordNumber)
return image
def getReadableWordSegment(self, wordNumber):
'returns transposed form of line segment Image'
image = self.segmenter.getSegment(wordNumber)
transImg = cv.CreateImage((image.height, image.width), cv.IPL_DEPTH_8U,
1)
cv.Transpose(image, transImg)
return transImg
def init():
"""
Initialize FeatureExtraction initializing Segmenter
"""
print("Initializing FeatureExtractor...")
FeatureExtractor.__feature_extractor_model = FeatureExtractor.__generate_model()
Segmenter.init()
print("Done.")
def pitchTableForBuffer(self, pitchBuffer):
filterer = Filterer(pitchBuffer, lowPassCutoffInHZ=settings.minimumPitchInHZ, highPassCutoffInHZ=settings.maximumPitchInHZ, gain=1)
buf = filterer.process()
segmenter = Segmenter(buf, windowWidth=2)
pitchTable = sp.zeros(segmenter.numberOfSegments())
for (buf, index) in segmenter.eachSegment():
pitchTable[index] = PitchEstimator.pitchForPeriod(buf)
return pitchTable
class ToCharacters(object):
segmenter=Segmenter()
def __init__(self):
'''
constructor
'''
self.segmenter=Segmenter()
def segment(self,image):
'returns Number of available character segments in a line'
return self.segmenter.segmentHorizontally(image)
#returns the same but manipulated line Image
def splitDika(self,image):
segmentLine=ToLines()
totalLines=segmentLine.segment(image)
pixelCounter = PixelCounter()
for lineNum in range(totalLines):
lineImg=segmentLine.getLineSegment(lineNum)
lineRowRange=segmentLine.getCoordinatesOfLineSegment(lineNum)
countArr=pixelCounter.getCountArr_V(lineImg, 0)
#print countArr
min=pixelCounter.getMin(countArr)
#dika width is min vertically and it is most occurred horizontally
for x in pixelCounter.getOccurredPositionList(countArr, min):
for height in range(lineRowRange[0],(lineRowRange[1]+lineRowRange[0])/2):
cv.Set2D(image,height,x,255)
return image
def POST(self):
# Open a temporary file
file = open("%s/in_%f.dat" % (directory, random()), 'wb')
inputs = web.input(video={}, type={}, start={}, end={}, faces={})
file.write(inputs['video'].value)
# Flush the file
file.flush()
os.fsync(file.fileno())
file.close()
start = None
end = None
if len(inputs["start"]) > 0:
start = int(inputs["start"])
if len(inputs["end"]) > 0:
end = int(inputs["end"])
client = Client(file.name, int(inputs['type']), start=start, end=end)
segmentNames = client.run(Segmenter(), inputs['faces'] == "on", "THEO",
"ogg", False)
return render.video(segments=segmentNames)
def testModel():
'Test the model on a directory of character images'
parser = argparse.ArgumentParser()
parser.add_argument('--model',
help='select which model to use',
default="model")
parser.add_argument('--test',
help="path to dir",
default="../../data/sample_data/IAM/")
args = parser.parse_args()
mf = ModelFactory(modelName=args.model)
model = mf.load()
CER = []
ground_truth = []
predicted = []
labels = [
"nominating", "any", "move", "meeting", "been", "texts", "ready", "one"
]
ctr = 0
imgFiles = []
gt = []
for root, _, files in os.walk(args.test):
for file in files:
# if file[-4:] != '.jpg':
# continue
imgFiles.append(os.path.join(root, file))
imgFiles = sorted(imgFiles)
print("Hello")
for file in imgFiles:
print(file)
img = cv2.imread(file, cv2.IMREAD_GRAYSCALE)
#print(img.shape)
#charImg = mf.preprocess(img, invert=True)
#print("Image processed")
#prediction = mf.predictChar(model, charImg)
segmenter = Segmenter()
pred = mf.predictWord(model, segmenter, img, show=False)
print(pred)
print(file)
gt = file.split('/')[-1]
print(gt)
gt = gt.split('.')[0]
print(gt)
gt = labels[ctr]
print("labels", gt)
ctr += 1
CER.append(mf.getCER(pred, gt))
ground_truth.append(gt)
predicted.append(pred)
for i in range(len(predicted)):
print(ground_truth[i], predicted[i], CER[i])
print(np.average(np.array(CER)))
def gold_data(self, sentence):
seg_sentence = SegSentence(sentence)
fwd_bigrams, unigrams = self.sentence_sequence(seg_sentence)
features = Segmenter.training_data(seg_sentence)
return {
'segSentence': seg_sentence,
'fwd_bigrams': fwd_bigrams,
'unigrams': unigrams,
'features': features,
}
class ToLines(object):
segmenter=Segmenter()
def __init__(self):
'''
constructor
'''
self.segmenter=Segmenter()
def segment(self,image):
'returns Number of available line segments'
return self.segmenter.segmentHorizontally(image)
def getLineSegment(self,lineNumber):
'returns line segment Image'
return self.segmenter.getSegment(lineNumber)
def getCoordinatesOfLineSegment(self,lineNumber):
return self.segmenter.getRowRange(lineNumber)
def getPixelCountsOfLineSegment(self,lineNumber):
#print self.segmenter.getPixelCountsOfSegment(lineNumber)
return self.segmenter.getPixelCountsOfSegment(lineNumber)
def __init__(self, buf):
self.mainBuffer = buf
self.pitchTable = None
self.pitchBuffer = Buffer.copy(buf)
if settings.preEmphasis:
PreEmphasizer.processBuffer(buf)
self.pitchTable = {}
wrappedPitch = False
if settings.overridePitch:
wrappedPitch = settings.pitchValue
else:
self.pitchTable = self.pitchTableForBuffer(self.pitchBuffer)
coefficients = sp.zeros(11)
segmenter = Segmenter(buf=self.mainBuffer, windowWidth=settings.windowWidth)
frames = []
for (cur_buf, i) in segmenter.eachSegment():
HammingWindow.processBuffer(cur_buf)
coefficients = cur_buf.getCoefficientsFor()
reflector = Reflector.translateCoefficients(coefficients, cur_buf.size)
if wrappedPitch:
pitch = int(wrappedPitch)
else:
pitch = self.pitchTable[i]
frameData = FrameData(reflector, pitch, repeat=False)
frames.append(frameData)
if settings.includeExplicitStopFrame:
frames.append(FrameData.stopFrame())
self.frames = frames
def extract_all_features(im):
"""
Given an image (as np tensor) extract several features
:param im: a np tensor 1x256x256x3
:return: lbp, rgb histogram, ycbcr histogram, ycbcr statistics, rgb statistics, sift descriptors and neural features
"""
im_copy = im.copy()
segmented_im, mask = Segmenter.segment_image(im_copy)
if np.sum(mask) == 0:
raise FeatureExtractionException()
rgb_hist = FeatureExtractor.extract_rgb_hist(segmented_im, mask)
ycbcr_hist = FeatureExtractor.extract_ycbcr_hist(segmented_im, mask)
neural = FeatureExtractor.extract_neural_features(im_copy)
lbp = FeatureExtractor.compute_LBP_rgb(segmented_im, mask)
ycbcr_statistics = FeatureExtractor.extract_statistics_ycbcr(segmented_im, mask)
rgb_statistics = FeatureExtractor.extract_statistics_rgb(segmented_im, mask)
sift_kp = FeatureExtractor.extract_sift_kp(segmented_im, mask)
if sift_kp is None or len(sift_kp.shape) == 0:
raise FeatureExtractionException()
return lbp, rgb_hist, ycbcr_hist, ycbcr_statistics, rgb_statistics, sift_kp, neural
def predict(filepath, classifier):
"""
Input: filepath to the image file (string)
Output: MusicXML (string) and midi (base 64 string)
"""
basename = osp.basename(filepath)
filename, extension = osp.splitext(basename)
cached_file = Path(osp.join("cached_stream", filename + '.pkl'))
if cached_file.is_file():
song_stream = converter.thaw(cached_file)
else:
symbols, seg = Segmenter.symbols_from_File(filepath, use_cache=True)
y = classifier.predict_symbols(symbols, use_class_numbers=True)
for i in range(len(y)):
symbols[i].work_out_type(y[i])
song = Song()
song.add_symbols(symbols)
song.parse_symbols()
converter.freeze(song.stream,
fmt='pickle',
fp=osp.join(os.getcwd(), "cached_stream",
filename + ".pkl"))
song_stream = song.stream
# get musicxml as a string
out = musicxml.m21ToXml.GeneralObjectExporter(song_stream).parse()
outStr = out.decode('utf-8')
# save midi file, then convert into b64 format
song_stream.write('midi', osp.join('resources', 'mid', filename + '.mid'))
file = open(osp.join('resources', 'mid', filename + '.mid'), 'rb').read()
b64 = base64.b64encode(file)
return outStr.replace('\n', ''), b64
def main():
sheet_path = './resources/sheets/training/'
output_path = './resources/sheets/training/trimmed/'
reject_parth = './resources/sheets/training/rejected/'
for filename in os.listdir(sheet_path):
if filename.endswith('.jpg') or filename.endswith(
'.jepg') or filename.endswith('.png'):
print(filename)
segmenter = Segmenter(os.path.join(sheet_path, filename))
staff_removed = segmenter.remove_staff_lines()
boxed = segmenter.getSymbols(merge_overlap=True)
cv2.imwrite(output_path + 'boxed{}.png'.format(filename[:-6]),
boxed)
segmenter.saveSymbols(save_origonal=False,
format=filename[:-6] + '_note{:03}.png',
width=150,
path=output_path,
reject_path=reject_parth,
dirty_times=0)
if args.model is None:
print('Must specify --model or (or --write-vocab) parameter.')
print(' (Use -h or --help flag for full option list.)')
sys.exit()
if args.test is not None:
from seg_sentence import SegSentence
from network import Network
from Segmenter import Segmenter
test_sentence = SegSentence.load_sentence_file(args.test)
print('Loaded test trees from {}'.format(args.test))
network = Network.load(args.model)
print('Loaded model from: {}'.format(args.model))
accuracy = Segmenter.evaluate_corpus(test_sentence,fm,network)
#Segmenter.write_predicted('data/predicted',test_sentence,fm,network)
print('Accuracy: {}'.format(accuracy))
elif args.train is not None:
from network import Network
if args.np_seed is not None:
import numpy as np
np.random.seed(args.np_seed)
network = Network.train(
corpus = fm,
bigrams_dims = args.bigrams_dims,
unigrams_dims = args.unigrams_dims,
lstm_units = args.lstm_units,
from ImageProcessor import ImageProcessor from Segmenter import Segmenter import matplotlib.pyplot as plt import numpy as np path = "lena.bmp" #Read and process image IP = ImageProcessor() img = IP.getImg(path)/256. imgSize = IP.getSize(path) imgSpace = IP.process(path) #Segment image S = Segmenter() K = 300 #number of clusters errorThreshold = 0.01 clusters = S.segment(K, imgSpace, imgSize, errorThreshold) #Get back clusters into image shape width, height = imgSize[0],imgSize[1] clusters = np.flipud(clusters.reshape((width,height))) #Plot image #http://stackoverflow.com/questions/15822934/matplotlib-imshow-how-to-plot-points-instead-of-image # Reduce the data by a factor of 4 (so that we can see the points) im = img # generate coordinates for the image. Note that the image is "top down", so the y coordinate goes from high to low. ys, xs = np.mgrid[im.shape[0]:0:-1, 0:im.shape[1]]
from Segmenter import Segmenter
import json
import sys
# from gensim.corpora import Dictionary
from gensim.models.phrases import Phrases
seg = Segmenter()
# vocab = Dictionary()
phrases = Phrases()
text_path = sys.argv[1]
def get_data(text_path):
for line in open(text_path, "r"):
line = line.strip()
if line:
data = json.loads(line)
yield data['abstract']
for ind, text in enumerate(get_data(text_path)):
segments = seg(text, segment_len=1, segment_overlap=0)
phrases.add_vocab(segments)
# vocab.add_documents(segments, prune_at=2000000)
if ind % 10000:
print(f"\rProcessed:{ind}", end = "")
# this is a dummy file to run other scripts
# it contains examples on how to run the main Segmenter class
from lib.Helpers import load_text, load_file
from Segmenter import Segmenter
from Evaluation import Evaluation
# loading data
file_path = 'data/br-phono-train.txt'
#file_path = 'data/small.txt'
(text, word_freq, char_freq) = load_text(file_path)
segmenter = Segmenter(text=text,
char_freq=char_freq,
p=2,
alpha=20,
p_hash=0.5)
print "Start segmenting \n"
segm_text = segmenter.run(200)
filetext = load_file(file_path)
print "Start evaluating \n"
evaluation = Evaluation(filetext, segm_text)
P, R, F, BP, BR, BF, LP, LR, LF = evaluation.run()
print "Boundary evaluation: \n Precision: %.2f, Recall: %.2f, F-measure: %.2f \n" % (
P, R, F)
print "Ambigious boundary evaluation: \n Precision: %.2f, Recall: %.2f, F-measure: %.2f \n" % (
BP, BR, BF)
print "Lexicon evaluation: \n Precision: %.2f, Recall: %.2f, F-measure: %.2f \n" % (
LP, LR, LF)
def train(corpus, bigrams_dims, unigrams_dims, lstm_units, hidden_units,
epochs, batch_size, train_data_file, dev_data_file,
model_save_file, droprate, unk_params, alpha, beta):
start_time = time.time()
fm = corpus
bigrams_size = corpus.total_bigrams()
unigrams_size = corpus.total_unigrams()
network = Network(
bigrams_size=bigrams_size,
unigrams_size=unigrams_size,
bigrams_dims=bigrams_dims,
unigrams_dims=unigrams_dims,
lstm_units=lstm_units,
hidden_units=hidden_units,
label_size=fm.total_labels(),
span_nums=fm.total_span_nums(),
droprate=droprate,
)
network.init_params()
print('Hidden units : {} ,per LSTM units : {}'.format(
hidden_units,
lstm_units,
))
print('Embeddings: bigrams = {}, unigrams = {}'.format(
(bigrams_size, bigrams_dims), (unigrams_size, unigrams_dims)))
print('Dropout rate : {}'.format(droprate))
print('Parameters initialized in [-0.01,0.01]')
print('Random UNKing parameter z = {}'.format(unk_params))
training_data = corpus.gold_data_from_file(train_data_file)
num_batched = -(-len(training_data) // batch_size)
print('Loaded {} training sentences ({} batches of size {})!'.format(
len(training_data),
num_batched,
batch_size,
))
parse_every = -(-num_batched // 4)
dev_sentences = SegSentence.load_sentence_file(dev_data_file)
print('Loaded {} validation sentences!'.format(len(dev_sentences)))
best_acc = FScore()
for epoch in xrange(1, epochs + 1):
print('............ epoch {} ............'.format(epoch))
total_cost = 0.0
total_states = 0
training_acc = FScore()
np.random.shuffle(training_data)
for b in xrange(num_batched):
batch = training_data[(b * batch_size):(b + 1) * batch_size]
explore = [
Segmenter.exploration(example,
fm,
network,
alpha=alpha,
beta=beta) for example in batch
]
for (_, acc) in explore:
training_acc += acc
batch = [example for (example, _) in explore]
dynet.renew_cg()
network.prep_params()
errors = []
for example in batch:
## random UNKing ##
for (i, uni) in enumerate(example['unigrams']):
if uni <= 2:
continue
u_freq = fm.unigrams_freq_list[uni]
drop_prob = unk_params / (unk_params + u_freq)
r = np.random.random()
if r < drop_prob:
example['unigrams'][i] = 0
for (i, bi) in enumerate(example['fwd_bigrams']):
if bi <= 2:
continue
b_freq = fm.bigrams_freq_list[bi]
drop_prob = unk_params / (unk_params + b_freq)
r = np.random.random()
if r < drop_prob:
example['fwd_bigrams'][i] = 0
fwd, back = network.evaluate_recurrent(
example['fwd_bigrams'],
example['unigrams'],
)
for (left,
right), correct in example['label_data'].items():
# correct = example['label_data'][(left,right)]
scores = network.evaluate_labels(
fwd, back, left, right)
probs = dynet.softmax(scores)
loss = -dynet.log(dynet.pick(probs, correct))
errors.append(loss)
total_states += len(example['label_data'])
batch_error = dynet.esum(errors)
total_cost += batch_error.scalar_value()
batch_error.backward()
network.trainer.update()
mean_cost = total_cost / total_states
print(
'\rBatch {} Mean Cost {:.4f} [Train: {}]'.format(
b,
mean_cost,
training_acc,
),
end='',
)
sys.stdout.flush()
if ((b + 1) % parse_every) == 0 or b == (num_batched - 1):
dev_acc = Segmenter.evaluate_corpus(
dev_sentences,
fm,
network,
)
print(' [Val: {}]'.format(dev_acc))
if dev_acc.fscore() > best_acc.fscore():
best_acc = dev_acc
network.save(model_save_file)
print(' [saved model : {}]'.format(model_save_file))
current_time = time.time()
runmins = (current_time - start_time) / 60
print(' Elapsed time: {:.2f}m'.format(runmins))
return network
from ImageProcessor import ImageProcessor from Segmenter import Segmenter import matplotlib.pyplot as plt import numpy as np path = "lena.bmp" #Read and process image IP = ImageProcessor() img = IP.getImg(path) / 256. imgSize = IP.getSize(path) imgSpace = IP.process(path) #Segment image S = Segmenter() K = 300 #number of clusters errorThreshold = 0.01 clusters = S.segment(K, imgSpace, imgSize, errorThreshold) #Get back clusters into image shape width, height = imgSize[0], imgSize[1] clusters = np.flipud(clusters.reshape((width, height))) #Plot image #http://stackoverflow.com/questions/15822934/matplotlib-imshow-how-to-plot-points-instead-of-image # Reduce the data by a factor of 4 (so that we can see the points) im = img # generate coordinates for the image. Note that the image is "top down", so the y coordinate goes from high to low. ys, xs = np.mgrid[im.shape[0]:0:-1, 0:im.shape[1]] # Scatter plots take 1d arrays of xs and ys, and the colour takes a 2d array,
def label_symbols(self,
min_x_pos=500,
edit_last=False,
dont_use_cache=False):
measures = self.xmldoc.getElementsByTagName('Measure')
xml_objects = []
while len(measures) is not 0:
current_measure = measures.pop(0)
current_measure_number = current_measure.getAttribute('number')
for child in current_measure.childNodes:
if child.localName is not None:
if child.nodeName == 'Clef':
xml_objects.append(('clef', current_measure_number))
elif child.nodeName == 'TimeSig':
xml_objects.append(('timeSig', current_measure_number))
elif child.nodeName == 'Rest':
xml_objects.append(('rest_{}'.format(
child.childNodes[1].childNodes[0].data),
current_measure_number))
elif child.nodeName == 'Chord':
if child.childNodes[1].nodeName == 'BeamMode':
xml_objects.append(('tied_{}'.format(
child.childNodes[3].childNodes[0].data),
current_measure_number))
elif child.childNodes[3].nodeName != 'Beam':
xml_objects.append(('{}'.format(
child.childNodes[1].childNodes[0].data),
current_measure_number))
xml_objects.append(('barline', current_measure_number))
first_pic = True
for filename in self.picture_filenames:
segmenter = Segmenter.load_segmenter_from_file(filename)
if (segmenter is None) or dont_use_cache:
segmenter = Segmenter(filename)
if not first_pic:
if edit_last or filename != self.picture_filenames[-1]:
segmenter.blockout_markings()
segmenter.remove_staff_lines()
segmenter.getSymbols(True)
segmenter.save_to_file()
symbols = segmenter.symbols
i = -1
saveCount = 0
while len(symbols) != 0:
i += 1
saveCount = saveCount - 1 if saveCount > 0 else 0
if not is_barline(
symbols[0]) and xml_objects[0][0] == 'barline':
saveCount = 0
print(
'error not enough xmls found at bar {} line {} in file {}'
.format(self.matched_pairs[-1][0][1],
symbols[0].line_num, filename[-20:-4]))
popped = self.matched_pairs.pop()
while popped[0][0] != 'barline':
popped = self.matched_pairs.pop()
self.matched_pairs.append(popped)
popped = symbols.pop(0)
while not is_barline(popped):
popped = symbols.pop(0)
symbols.insert(0, popped)
if is_barline(symbols[0]) and xml_objects[0][0] != 'barline':
saveCount = 0
print(
'error too many xmls found at bar {} line {} in file {}'
.format(self.matched_pairs[-1][0][1],
symbols[0].line_num, filename[-20:-4]))
# save = True
popped = self.matched_pairs.pop()
while popped[0][0] != 'barline':
popped = self.matched_pairs.pop()
self.matched_pairs.append(popped)
popped = xml_objects.pop(0)
while popped[0] != 'barline':
popped = xml_objects.pop(0)
xml_objects.insert(0, popped)
if symbols[0].x < min_x_pos and (not first_pic or i != 0):
self.matched_pairs.append((('clef', -1), symbols.pop(0)))
else:
self.matched_pairs.append(
(xml_objects.pop(0), symbols.pop(0)))
if saveCount > 0:
thumbnail_name = os.path.join(
self.output_path, '{}_{}_line{}__{}.png'.format(
filename[-13:-10], self.matched_pairs[-1][0][0],
self.matched_pairs[-1][1].line_num, i))
cv2.imwrite(thumbnail_name, self.matched_pairs[-1][1].im)
first_pic = False
import numpy as np
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.metrics import roc_auc_score
import functools
import sys
import math
from Segmenter import Segmenter
from Tokenizer import Tokenizer
import TextNormalizer
import codecs
# ------------------------------------------------------------------------
df = pd.read_csv("../data/dftrain17.csv", encoding='utf-8')
segmenter = Segmenter()
#test_str = u'Сразу после возвращения Фрама Нансен стал главным специалистом по полярным исследованиям в мире, по выражению Р. Хантфорда — оракулом для всех исследователей полярных широт Севера и Юга [188]. Нансен консультировал бельгийского барона Адриена де Жерлаша, который планировал в 1898 году свою экспедицию в Антарктиду, одним из участников команды был Руаль Амундсен[189]. Известнейший исследователь Гренландии Кнуд Расмуссен сравнил посещение Нансена с посвящением в рыцари[190]. В то же время Нансен категорически отказался встречаться со своим соотечественником Карстеном Борхгревинком, сочтя его мошенником, хотя именно он совершил первую успешную зимовку на побережье Антарктиды[191]. В 1900 году в Норвегию приехал за консультациями Роберт Скотт со своим покровителем Клементом Маркхэмом — давним другом Нансена, готовившим британскую экспедицию в Антарктиду. Несмотря на то, что англичане практически проигнорировали все советы, Нансен и Скотт остались в хороших отношениях[192].'
#for l in segmenter.split(test_str):
# print(l)
# -------------------------------------------------------------------------
wrt0 = codecs.open('../data/rows(y=0).txt', 'w', 'utf-8')
wrt1 = codecs.open('../data/rows(y=1).txt', 'w', 'utf-8')
for index, row in tqdm.tqdm(df.iterrows(), total=df.shape[0]):
question = row.question
paragraph = row.paragraph
wrt = wrt0 if row['target'] == 0 else wrt1
if opts.frameLengths is None or opts.frameLengths == []:
opts.frameLengths = unspecifiedFrameLengths
for vid in opts.videos:
print("*** Video : ", vid)
for length in opts.frameLengths:
try:
print(" --- Length : " + str(opts.start) + " -> " +
str(opts.start + length))
client = Client(vid,
SplitType.ON_FACE_CLUSTERS,
start=opts.start,
end=opts.start + length)
t1 = timeit(lambda: client.run(Segmenter(),
True,
"THEO",
"OGG",
False,
options={
"verbose": 0,
"write": False
}),
number=1)
t2 = timeit(lambda: client.run(Segmenter(),
True,
"THEO",
"OGG",
False,
options={
def __init__(self):
'''
constructor
'''
self.segmenter=Segmenter()
from Dataset import Dataset
from Segmenter import Segmenter
from BlobAnalyzer import BlobAnalyzer
# reading dataset
dataset = Dataset("./immagini/*.BMP")
segmenter = Segmenter()
analyzer = BlobAnalyzer()
# set true: show the computation of all the process
show = False
for image in dataset.images:
binary, labels, n_labels = segmenter.run(image, show=show)
stats = analyzer.run(binary, labels, n_labels, show=show)
analyzer.show(image, stats)
'''
Created on Nov 16, 2014
@author: Etienne
'''
from ImageProcessor import ImageProcessor
from Segmenter import Segmenter
import matplotlib.pyplot as plt
IP = ImageProcessor()
imgSize = IP.getSize("5.jpg")
imgSpace = IP.process("5.jpg")
# print imgSpace[:766,3:5]
S = Segmenter()
K = 108
#Test pick centers
#Should give S points evenly spaced across the space
gridInterval = S.calcGridInterval(imgSpace.shape, K)
centers = S.pickCenters(K, gridInterval, imgSpace, imgSize)
print centers.shape
# print centers
X0 = centers[:,3]
Y0 = centers[:,4]
# print X,Y
plt.plot(X0,Y0,"rx")
class Modifiers:
segmenter = Segmenter()
pixelCounter = PixelCounter()
def __init__(self):
'Constructor'
#returns horizontal starting row and ending row of dika
def locateDika(self, lineImg):
countArr_V = self.pixelCounter.getCountArr_V(lineImg, 0)
countArr_H = self.pixelCounter.getCountArr_H(lineImg, 0)
dikaWidth = self.pixelCounter.getMin(countArr_V)
max = self.pixelCounter.getMax(countArr_H)
#print max
maxAt = self.pixelCounter.getOccurredPositionList(countArr_H, max)
#print maxAt
return maxAt
#get horizontally occurred pixel count
#+1 or -1 or set some threshold
# OR
# get horizontally most occurred pixel count => say 'x'
# find width of dika
# ?? is this really needed?? position from x-widthDika to x+widthDika can be considered to be dika??
# or may be count the pixels in that row and decide according to certain threshold
''
def extractTop(self, lineImg):
''
dikaList = self.locateDika(lineImg)
rect = CvRect
rect.x = 0
rect.y = 0
rect.width = lineImg.width
rect.height = dikaList[
0] - 2 # -2 is to eliminate disturbances from bha tha dha etc
segmentImg = cv.CreateImage((rect.width, rect.height), cv.IPL_DEPTH_8U,
1)
cv.SetImageROI(lineImg, (rect.x, rect.y, rect.width, rect.height))
cv.Resize(lineImg, segmentImg)
cv.ResetImageROI(lineImg)
cv.ShowImage("TOP", segmentImg)
# locate horizontal dika position
# set roi upper than dika
# extract the roi
def extractMainElement(self, lineImg):
''
#strip top
#strip bottom
#set roi to remaining
#extract roi
#OR
#using height of purnabiram extract all letters of that height
def extractBottom(self, lineImg):
''
dftrain = pd.read_csv("../data/dftrain.csv", encoding='utf-8')
dftest = pd.read_csv("../data/dftest.csv", encoding='utf-8')
known_words = set()
# русский лексикон
with codecs.open('../data/word2lemma.dat', 'r', 'utf-8') as rdr:
for line in rdr:
tx = line.strip().split(u'\t')
if len(tx) == 3:
word = tx[0].lower()
known_words.add(word)
# ------------------------------------------------------------------------
segmenter = Segmenter()
def is_oov(word):
return TextNormalizer.normalize_word(
word) not in known_words and len(word) > 0 and not word[0].isdigit()
for name, df in [('train', dftrain), ('test', dftest)]:
for index, row in tqdm.tqdm(df.iterrows(),
total=df.shape[0],
desc="Computing the tail matching for " +
name):
question = TextNormalizer.preprocess_question_str(row.question)
paragraph = row.paragraph
from Indexer import Indexer
from FeatureExtractor import FeatureExtractor
from Segmenter import Segmenter
from PendingIterator import PendingIterator
from PDFExtractor import PDFExtractor as PDF
from params import index_store
import sys
pending = PendingIterator(index_store)
segm = Segmenter()
# https: // docs.python.org / 3 / library / sqlite3.html
for p in pending:
segments = segm(PDF.read(p))
for s_id, s in segments:
print(s_id, s)
# -------------------------------------------------------------------
X_parag_train = None
X_parag1_train = None
X_quest_train = None
X_parag_test = None
X_parag1_test = None
X_quest_test = None
y_train = dftrain['target'].values
word2id = dict()
word2id[u''] = 0
segmenter = Segmenter()
for name, df in [('train', dftrain), ('test', dftest)]:
X_parag = np.zeros((df.shape[0], max_len), dtype=np.int32)
X_parag1 = np.zeros((df.shape[0], max_len), dtype=np.int32)
X_quest = np.zeros((df.shape[0], max_len), dtype=np.int32)
for index, row in tqdm.tqdm(df.iterrows(), total=df.shape[0], desc="Processing " + name):
question = TextNormalizer.preprocess_question_str(row.question)
paragraph = row.paragraph
quest_words = TextNormalizer.tokenize_words(question)[0:max_len]
parag_words = TextNormalizer.tokenize_words(paragraph)[0:max_len]
for word_pos,word in enumerate(quest_words):
def __init__(self):
'''
constructor
'''
self.segmenter = Segmenter()
