def predictString(self, line):
"""
Predicts a string from an input image.
Args:
line (numpy.array): Input image
Returns:
A unicode string containing the recognition result.
"""
line = line.reshape(-1, self.rnn.ninput(), 1)
self.rnn.inputs.aset(line.astype('float32'))
self.rnn.forward()
self.outputs = self.rnn.outputs.array().reshape(
line.shape[0], self.rnn.noutput())
codes = [
x[0]
for x in kraken.lib.lstm.translate_back_locations(self.outputs)
]
cls = clstm.Classes()
cls.resize(len(codes))
for i, v in enumerate(codes):
cls[i] = int(v)
res = self.rnn.decode(cls)
return res
def trainString(self, line, s, update=1):
"""
Trains the network using an input numpy array and a unicode string.
Strings are assumed to be in ``display`` order as produced as the
result of the BiDi algorithm.
Args:
line (numpy.array): Input image
s (str): Expected output string
update (bool): Switch to disable weight updates
Returns:
An unicode string containing the recognized sequence.
"""
labels = clstm.Classes()
self.rnn.encode(labels, s)
cls = self.trainSequence(line, labels)
return self.rnn.decode(cls)
def trainSequence(self, line, labels, update=1):
"""
Trains the network using an input numpy array and a series of labels.
Args:
line (numpy.array): Input image
labels (clstm.Classes): Label sequence
update (bool): Switch to disable weight updates
Returns:
clstm.Classes containing the recognized label sequence.
"""
line = line.reshape(-1, self.rnn.ninput(), 1)
self.rnn.inputs.aset(line.astype('float32'))
self.rnn.forward()
self.outputs = self.rnn.outputs.array().reshape(
line.shape[0], self.rnn.noutput())
# build CTC alignment
targets = clstm.Sequence()
aligned = clstm.Sequence()
clstm.mktargets(targets, labels, self.rnn.noutput())
clstm.seq_ctc_align(aligned, self.rnn.outputs, targets)
# calculate deltas, backpropagate and update weights
deltas = aligned.array() - self.rnn.outputs.array()
self.rnn.d_outputs.aset(deltas)
self.rnn.backward()
if update:
self.rnn.update()
codes = kraken.lib.lstm.translate_back(self.outputs)
cls = clstm.Classes()
cls.resize(len(codes))
for i, v in enumerate(codes):
cls[i] = v
return cls
