def evaluate(self,
data_stream,
train=False,
file_pred=None,
file_targets=None):
loss = 0.
num_examples = 0
iterator = data_stream.get_epoch_iterator()
if train:
print 'Train evaluation started'
i = 0
for inputs in iterator:
inputs = dict(zip(data_stream.sources, inputs))
x_mask_val = inputs['features_mask']
x_val = inputs['features']
y_val = inputs['phonemes']
drops_forw_states_val = inputs['drops_forw_states']
drops_forw_cells_val = inputs['drops_forw_cells']
drops_forw_igates_val = inputs['drops_forw_igates']
drops_back_states_val = inputs['drops_back_states']
drops_back_cells_val = inputs['drops_back_cells']
drops_back_igates_val = inputs['drops_back_igates']
y_mask_val = inputs['phonemes_mask']
y_hat = self.prediction_func(
x_val, x_mask_val, drops_forw_states_val, drops_forw_cells_val,
drops_forw_igates_val, drops_back_states_val,
drops_back_cells_val, drops_back_igates_val)
y_predict = numpy.argmax(y_hat, axis=2)
for batch in xrange(inputs['features'].shape[1]):
y_val_cur = y_val[:sum(y_mask_val[:, batch]), batch]
predicted = y_predict[:sum(x_mask_val[:, batch]), batch]
predicted = ctc_strip(predicted)
predictions = [
self.phoneme_dict[phone_ind] for phone_ind in predicted
if self.phoneme_dict[phone_ind] not in self.black_list
]
targets = [
self.phoneme_dict[phone_ind] for phone_ind in y_val_cur
if self.phoneme_dict[phone_ind] not in self.black_list
]
predictions = [x[0] for x in groupby(predictions)]
targets = [x[0] for x in groupby(targets)]
i += 1
if file_pred:
file_pred.write(' '.join(predictions) + '(%d)\n' % i)
if file_targets:
file_targets.write(' '.join(targets) + '(%d)\n' % i)
loss += Evaluation.wer([predictions], [targets])
num_examples += 1
print '.. found sequence example:', ' '.join(predictions)
print '.. real output was: ', ' '.join(targets)
if train:
break
if train:
print 'Train evaluation finished'
per = loss.sum() / num_examples
return {'per': per}
def evaluate(self, data_stream, train=False, file_pred=None,
file_targets=None):
loss = 0.
num_examples = 0
iterator = data_stream.get_epoch_iterator()
if train:
print 'Train evaluation started'
i = 0
for inputs in iterator:
inputs = dict(zip(data_stream.sources, inputs))
x_mask_val = inputs['features_mask']
x_val = inputs['features']
y_val = inputs['phonemes']
y_mask_val = inputs['phonemes_mask']
for batch_ind in xrange(inputs['features'].shape[1]):
if x_val.ndim == 2:
input_beam = numpy.tile(x_val[:, batch_ind][:, None],
(1, self.beam_size))
else:
input_beam = numpy.tile(x_val[:, batch_ind, :][:, None, :],
(1, self.beam_size, 1))
input_mask_beam = numpy.tile(x_mask_val[:, batch_ind][:, None],
(1, self.beam_size))
predictions, _ = self.beam_search.search(
{self.x: input_beam,
self.x_mask: input_mask_beam},
self.eol_symbol, 100)
predictions = [self.phoneme_dict[phone_ind] for phone_ind
in predictions[0]
if self.phoneme_dict[phone_ind] not in
self.black_list][1:-1]
targets = y_val[:sum(y_mask_val[:, batch_ind]), batch_ind]
targets = [self.phoneme_dict[phone_ind] for phone_ind
in targets
if self.phoneme_dict[phone_ind] not in
self.black_list][1:-1]
predictions = [x[0] for x in groupby(predictions)]
targets = [x[0] for x in groupby(targets)]
i += 1
if file_pred:
file_pred.write(' '.join(predictions) + '(%d)\n' % i)
if file_targets:
file_targets.write(' '.join(targets) + '(%d)\n' %i)
loss += Evaluation.wer([predictions], [targets])
num_examples += 1
print '.. found sequence example:', ' '.join(predictions)
print '.. real output was: ', ' '.join(targets)
if train:
break
if train:
print 'Train evaluation finished'
per = loss.sum() / num_examples
return {'per': per}
def evaluate(self, data_stream, train=False, file_pred=None,
file_targets=None):
loss = 0.
num_examples = 0
iterator = data_stream.get_epoch_iterator()
if train:
print 'Train evaluation started'
i = 0
for inputs in iterator:
inputs = dict(zip(data_stream.sources, inputs))
x_mask_val = inputs['features_mask']
x_val = inputs['features']
#transpose
y_val = inputs['phonemes']
y_mask_val = inputs['phonemes_mask']
y_hat = self.prediction_func(x_val, x_mask_val)
y_predict = numpy.argmax(y_hat, axis=2)
for batch in xrange(inputs['features'].shape[0]):
y_val_cur = y_val[:sum(y_mask_val[:, batch]), batch]
predicted = y_predict[:sum(x_mask_val[:, batch]), batch]
predicted = ctc_strip(predicted)
predictions = [self.phoneme_dict[phone_ind] for phone_ind
in predicted
if self.phoneme_dict[phone_ind] not in
self.black_list]
targets = [self.phoneme_dict[phone_ind] for phone_ind
in y_val_cur
if self.phoneme_dict[phone_ind] not in
self.black_list]
predictions = [x[0] for x in groupby(predictions)]
targets = [x[0] for x in groupby(targets)]
i += 1
if file_pred:
file_pred.write(' '.join(predictions) + '(%d)\n' % i)
if file_targets:
file_targets.write(' '.join(targets) + '(%d)\n' %i)
loss += Evaluation.wer([predictions], [targets])
num_examples += 1
print '.. found sequence example:', ' '.join(predictions)
print '.. real output was: ', ' '.join(targets)
if train:
break
if train:
print 'Train evaluation finished'
per = loss.sum() / num_examples
return {'per': per}
def evaluate(self,
data_stream,
train=False,
file_pred=None,
file_targets=None):
loss = 0.
num_examples = 0
iterator = data_stream.get_epoch_iterator()
if train:
print 'Train evaluation started'
i = 0
for inputs in iterator:
inputs = dict(zip(data_stream.sources, inputs))
x_mask_val = inputs['features_mask']
x_val = inputs['features']
y_val = inputs['triphones']
y_mask_val = inputs['triphones_mask']
y_hat = self.prediction_func(x_val, x_mask_val)
y_predict = numpy.argmax(y_hat, axis=2)
for batch in xrange(inputs['features'].shape[1]):
y_val_cur = y_val[:sum(y_mask_val[:, batch]), batch]
predicted = y_predict[:sum(x_mask_val[:, batch]), batch]
predictions = [str(trihone) for triphone in predicted]
targets = [str(triphone) for triphone in y_val_cur]
predictions = [x[0] for x in groupby(predictions)]
targets = [x[0] for x in groupby(targets)]
i += 1
if file_pred:
file_pred.write(' '.join(predictions) + '(%d)\n' % i)
if file_targets:
file_targets.write(' '.join(targets) + '(%d)\n' % i)
loss += (sum(numpy.not_equal(y_val_cur,
predicted))).astype(
'float32') /\
len(predicted)
num_examples += 1
#print '.. found sequence example:', ' '.join(predictions)
#print '.. real output was: ', ' '.join(targets)
if train:
break
if train:
print 'Train evaluation finished'
fer = loss / num_examples
return {'fer': fer}
def evaluate(self, data_stream, train=False, file_pred=None,
file_targets=None):
loss = 0.
num_examples = 0
iterator = data_stream.get_epoch_iterator()
if train:
print 'Train evaluation started'
i = 0
for inputs in iterator:
inputs = dict(zip(data_stream.sources, inputs))
x_mask_val = inputs['features_mask']
x_val = inputs['features']
y_val = inputs['triphones']
y_mask_val = inputs['triphones_mask']
y_hat = self.prediction_func(x_val, x_mask_val)
y_predict = numpy.argmax(y_hat, axis=2)
for batch in xrange(inputs['features'].shape[1]):
y_val_cur = y_val[:sum(y_mask_val[:, batch]), batch]
predicted = y_predict[:sum(x_mask_val[:, batch]), batch]
predictions = [str(trihone) for triphone in predicted]
targets = [str(triphone) for triphone in y_val_cur]
predictions = [x[0] for x in groupby(predictions)]
targets = [x[0] for x in groupby(targets)]
i += 1
if file_pred:
file_pred.write(' '.join(predictions) + '(%d)\n' % i)
if file_targets:
file_targets.write(' '.join(targets) + '(%d)\n' %i)
loss += (sum(numpy.not_equal(y_val_cur,
predicted))).astype(
'float32') /\
len(predicted)
num_examples += 1
#print '.. found sequence example:', ' '.join(predictions)
#print '.. real output was: ', ' '.join(targets)
if train:
break
if train:
print 'Train evaluation finished'
fer = loss / num_examples
return {'fer': fer}
def verify_groupby(*args, **kwargs):
if 'n' in kwargs:
if 'm' not in kwargs:
raise ValueError('got n without m')
pickle = True
n = kwargs.pop('n')
m = kwargs.pop('m')
elif 'm' in kwargs:
raise ValueError('got m without n')
else:
pickle = False
n = m = None
if 'pickle_outer' in kwargs:
pickle_outer = kwargs.pop('pickle_outer')
else:
pickle_outer = None
reference = itertools.groupby(*args, **kwargs)
actual = groupby(*args, **kwargs)
outer_iters = 0
while True:
if outer_iters == pickle_outer:
actual = cPickle.loads(cPickle.dumps(actual))
try:
ref_key, ref_grouper = next(reference)
except StopIteration:
check_stops(actual)
break
try:
actual_key, actual_grouper = next(actual)
except StopIteration:
assert False, "prematurely exhausted; expected {}".format(ref_key)
if pickle:
this_n = n[0]
n = n[1:]
this_m = m[0]
m = m[1:]
verify_pickle(partial(_identity, actual_grouper),
partial(_identity, ref_grouper), this_n, this_m)
else:
verify_same(partial(_identity, actual_grouper),
partial(_identity, ref_grouper), None)
outer_iters += 1
def verify_groupby(*args, **kwargs):
if 'n' in kwargs:
if 'm' not in kwargs:
raise ValueError('got n without m')
pickle = True
n = kwargs.pop('n')
m = kwargs.pop('m')
elif 'm' in kwargs:
raise ValueError('got m without n')
else:
pickle = False
n = m = None
if 'pickle_outer' in kwargs:
pickle_outer = kwargs.pop('pickle_outer')
else:
pickle_outer = None
reference = itertools.groupby(*args, **kwargs)
actual = groupby(*args, **kwargs)
outer_iters = 0
while True:
if outer_iters == pickle_outer:
actual = cPickle.loads(cPickle.dumps(actual))
try:
ref_key, ref_grouper = next(reference)
except StopIteration:
check_stops(actual)
break
try:
actual_key, actual_grouper = next(actual)
except StopIteration:
assert False, "prematurely exhausted; expected {}".format(ref_key)
if pickle:
this_n = n[0]
n = n[1:]
this_m = m[0]
m = m[1:]
verify_pickle(partial(_identity, actual_grouper),
partial(_identity, ref_grouper), this_n, this_m)
else:
verify_same(partial(_identity, actual_grouper),
partial(_identity, ref_grouper), None)
outer_iters += 1
def network_evaluation(predict_fn,
data_stream,
phoneme_dict,
black_list):
#####################
# get data iterator #
#####################
data_iterator = data_stream.get_epoch_iterator()
#########################
# set evaluation result #
#########################
total_ctc = 0.
total_per = 0.
total_sample = 0.
total_batch = 0.
##############
# evaluation #
##############
# for each batch iteration
for i, data in enumerate(data_iterator):
# get input data
input_data = data[0]
input_mask = data[1]
# get target data
target_data = data[2]
target_mask = data[3]
# get prediction data
predict_output = predict_fn(input_data,
input_mask,
target_data,
target_mask)
predict_idx = predict_output[0]
predict_ctc_cost = predict_output[1]
pred_cost_per_char = predict_output[2]
# add up ctc cost
total_ctc += predict_ctc_cost
total_batch += 1.
# for each data, per evaluation
for j in range(input_data.shape[0]):
cur_target_data = target_data[j, :numpy.sum(target_mask[j])]
cur_predict_data = predict_idx[j, :numpy.sum(input_mask[j])]
cur_predict_data = ctc_strip(cur_predict_data)
cur_target_phoneme = [phoneme_dict[phone_ind] for phone_ind in cur_target_data if phoneme_dict[phone_ind] not in black_list]
cur_predict_phoneme = [phoneme_dict[phone_ind] for phone_ind in cur_predict_data if phoneme_dict[phone_ind] not in black_list]
targets = [x[0] for x in groupby(cur_target_phoneme)]
predictions = [x[0] for x in groupby(cur_predict_phoneme)]
total_per += Evaluation.wer([predictions], [targets])
total_sample += 1
total_ctc = total_ctc/total_batch
total_per = total_per.sum()/total_sample
return total_ctc, total_per
def evaluate(self,
data_stream,
train=False,
file_pred=None,
file_targets=None):
loss = 0.
num_examples = 0
iterator = data_stream.get_epoch_iterator()
if train:
print 'Train evaluation started'
i = 0
for inputs in iterator:
inputs = dict(zip(data_stream.sources, inputs))
x_mask_val = inputs['features_mask']
x_val = inputs['features']
y_val = inputs['phonemes']
y_mask_val = inputs['phonemes_mask']
for batch_ind in xrange(inputs['features'].shape[1]):
if x_val.ndim == 2:
input_beam = numpy.tile(x_val[:, batch_ind][:, None],
(1, self.beam_size))
else:
input_beam = numpy.tile(x_val[:, batch_ind, :][:, None, :],
(1, self.beam_size, 1))
input_mask_beam = numpy.tile(x_mask_val[:, batch_ind][:, None],
(1, self.beam_size))
predictions, _ = self.beam_search.search(
{
self.x: input_beam,
self.x_mask: input_mask_beam
}, self.eol_symbol, 100)
predictions = [
self.phoneme_dict[phone_ind]
for phone_ind in predictions[0]
if self.phoneme_dict[phone_ind] not in self.black_list
][1:-1]
targets = y_val[:sum(y_mask_val[:, batch_ind]), batch_ind]
targets = [
self.phoneme_dict[phone_ind] for phone_ind in targets
if self.phoneme_dict[phone_ind] not in self.black_list
][1:-1]
predictions = [x[0] for x in groupby(predictions)]
targets = [x[0] for x in groupby(targets)]
i += 1
if file_pred:
file_pred.write(' '.join(predictions) + '(%d)\n' % i)
if file_targets:
file_targets.write(' '.join(targets) + '(%d)\n' % i)
loss += Evaluation.wer([predictions], [targets])
num_examples += 1
print '.. found sequence example:', ' '.join(predictions)
print '.. real output was: ', ' '.join(targets)
if train:
break
if train:
print 'Train evaluation finished'
per = loss.sum() / num_examples
return {'per': per}