def on_batch_end(self, n_update, logs=None):
self.cum_update += 1
if self.epoch_count + self.reload_epoch < self.start_sampling_on_epoch:
return
elif self.cum_update % self.each_n_updates != 0:
return
# Evaluate on each set separately
for s in self.set_name:
if self.beam_search:
params_prediction = {'max_batch_size': self.batch_size,
'n_parallel_loaders': self.extra_vars[
'n_parallel_loaders'],
'predict_on_sets': [s],
'n_samples': self.n_samples,
'pos_unk': False}
params_prediction.update(checkDefaultParamsBeamSearch(self.extra_vars))
predictions, truths, sources = self.model_to_eval.predictBeamSearchNet(self.ds, params_prediction)
else:
params_prediction = {'batch_size': self.batch_size,
'n_parallel_loaders': self.extra_vars[
'n_parallel_loaders'],
'predict_on_sets': [s],
'n_samples': self.n_samples,
}
# Convert predictions
postprocess_fun = None
if self.is_3DLabel:
postprocess_fun = [self.ds.convert_3DLabels_to_bboxes,
self.extra_vars[s]['references_orig_sizes']]
predictions = self.model_to_eval.predictNet(self.ds,
params_prediction,
postprocess_fun=postprocess_fun)
if self.print_sources:
if self.in_pred_idx is not None:
sources = [srcs[self.in_pred_idx][0] for srcs in sources]
sources = decode_predictions_beam_search(sources,
self.index2word_x,
pad_sequences=True,
verbose=self.verbose)
if s in predictions:
if params_prediction['pos_unk']:
samples = predictions[s][0]
alphas = predictions[s][1]
heuristic = self.extra_vars['heuristic']
else:
samples = predictions[s]
alphas = None
heuristic = None
predictions = predictions[s]
if self.is_text:
if self.out_pred_idx is not None:
samples = samples[self.out_pred_idx]
# Convert predictions into sentences
if self.beam_search:
predictions = decode_predictions_beam_search(samples,
self.index2word_y,
alphas=alphas,
x_text=sources,
heuristic=heuristic,
mapping=self.extra_vars.get('mapping', None),
verbose=self.verbose)
else:
predictions = decode_predictions(samples,
1,
self.index2word_y,
self.sampling_type,
verbose=self.verbose)
truths = decode_predictions_one_hot(truths, self.index2word_y,
verbose=self.verbose)
# Apply detokenization function if needed
if self.extra_vars.get('apply_detokenization', False):
if self.print_sources:
sources = map(self.extra_vars['detokenize_f'], sources)
predictions = map(self.extra_vars['detokenize_f'],
predictions)
truths = map(self.extra_vars['detokenize_f'], truths)
# Write samples
if self.print_sources:
# Write samples
for i, (source, sample, truth) in list(enumerate(zip(sources, predictions, truths))):
if sys.version_info.major == 2:
source = str(source.encode('utf-8'))
sample = str(sample.encode('utf-8'))
truth = str(truth.encode('utf-8'))
print("Source (%d): %s" % (i, source))
print("Hypothesis (%d): %s" % (i, sample))
print("Reference (%d): %s" % (i, truth))
print("")
else:
for i, (sample, truth) in list(enumerate(zip(predictions, truths))):
if sys.version_info.major == 2:
sample = str(sample.encode('utf-8'))
truth = str(truth.encode('utf-8'))
print("Hypothesis (%d): %s" % (i, sample))
print("Reference (%d): %s" % (i, truth))
print("")
def on_batch_end(self, n_update, logs={}):
self.cum_update += 1
if self.epoch_count + self.reload_epoch < self.start_sampling_on_epoch:
return
elif self.cum_update % self.each_n_updates != 0:
return
# Evaluate on each set separately
for s in self.set_name:
# Apply model predictions
params_prediction = {
'batch_size': self.batch_size,
'n_parallel_loaders': self.extra_vars['n_parallel_loaders'],
'predict_on_sets': [s],
'n_samples': self.n_samples,
'pos_unk': False,
'heuristic': 0,
'mapping': None
}
if self.beam_search:
params_prediction.update(
checkDefaultParamsBeamSearch(self.extra_vars))
predictions, truths, sources = self.model_to_eval.predictBeamSearchNet(
self.ds, params_prediction)
else:
# Convert predictions
postprocess_fun = None
if self.is_3DLabel:
postprocess_fun = [
self.ds.convert_3DLabels_to_bboxes,
self.extra_vars[s]['references_orig_sizes']
]
predictions = self.model_to_eval.predictNet(
self.ds,
params_prediction,
postprocess_fun=postprocess_fun)
if self.print_sources:
if self.in_pred_idx is not None:
sources = [srcs for srcs in sources[0][self.in_pred_idx]]
sources = decode_predictions_beam_search(sources,
self.index2word_x,
pad_sequences=True,
verbose=self.verbose)
if s in predictions:
if params_prediction['pos_unk']:
samples = predictions[s][0]
alphas = predictions[s][1]
heuristic = params_prediction['heuristic']
else:
samples = predictions[s]
alphas = None
heuristic = None
predictions = predictions[s]
if self.is_text:
if self.out_pred_idx is not None:
samples = samples[self.out_pred_idx]
# Convert predictions into sentences
if self.beam_search:
predictions = decode_predictions_beam_search(
samples,
self.index2word_y,
alphas=alphas,
x_text=sources,
heuristic=heuristic,
mapping=params_prediction['mapping'],
verbose=self.verbose)
else:
predictions = decode_predictions(samples,
1,
self.index2word_y,
self.sampling_type,
verbose=self.verbose)
truths = decode_predictions_one_hot(truths,
self.index2word_y,
verbose=self.verbose)
# Write samples
if self.print_sources:
# Write samples
for i, (source, sample, truth) in enumerate(
zip(sources, predictions, truths)):
print("Source (%d): %s" % (i, source))
print("Hypothesis (%d): %s" % (i, sample))
print("Reference (%d): %s" % (i, truth))
print("")
else:
for i, (sample,
truth) in enumerate(zip(predictions, truths)):
print("Hypothesis (%d): %s" % (i, sample))
print("Reference (%d): %s" % (i, truth))
print("")
def on_batch_end(self, n_update, logs=None):
self.cum_update += 1
if self.epoch_count + self.reload_epoch < self.start_sampling_on_epoch:
return
elif self.cum_update % self.each_n_updates != 0:
return
# Evaluate on each set separately
for s in self.set_name:
if self.beam_search:
params_prediction = {'max_batch_size': self.batch_size,
'n_parallel_loaders': self.extra_vars['n_parallel_loaders'],
'predict_on_sets': [s],
'n_samples': self.n_samples,
'pos_unk': False}
params_prediction.update(checkDefaultParamsBeamSearch(self.extra_vars))
predictions, truths, sources = self.model_to_eval.predictBeamSearchNet(self.ds,
params_prediction)
else:
params_prediction = {'batch_size': self.batch_size,
'n_parallel_loaders': self.extra_vars['n_parallel_loaders'],
'predict_on_sets': [s],
'n_samples': self.n_samples,
'verbose': self.verbose,
}
# Convert predictions
postprocess_fun = None
if self.is_3DLabel:
postprocess_fun = [self.ds.convert_3DLabels_to_bboxes,
self.extra_vars[s]['references_orig_sizes']]
predictions = self.model_to_eval.predictNet(self.ds,
params_prediction,
postprocess_fun=postprocess_fun)
if self.print_sources:
if self.in_pred_idx is not None:
sources = [srcs[self.in_pred_idx][0] for srcs in sources]
sources = decode_predictions_beam_search(sources,
self.index2word_x,
pad_sequences=True,
verbose=self.verbose)
if s in predictions:
if params_prediction['pos_unk']:
samples = predictions[s][0]
alphas = predictions[s][1]
heuristic = self.extra_vars['heuristic']
else:
samples = predictions[s]
alphas = None
heuristic = None
predictions = predictions[s]
if self.is_text:
if self.out_pred_idx is not None:
samples = samples[self.out_pred_idx]
# Convert predictions into sentences
if self.beam_search:
predictions = decode_predictions_beam_search(samples,
self.index2word_y,
glossary=self.extra_vars.get('glossary', None),
alphas=alphas,
x_text=sources,
heuristic=heuristic,
mapping=self.extra_vars.get('mapping', None),
verbose=self.verbose)
else:
predictions = decode_predictions(samples,
1,
self.index2word_y,
self.sampling_type,
verbose=self.verbose)
truths = decode_predictions_one_hot(truths, self.index2word_y,
verbose=self.verbose)
# Apply detokenization function if needed
if self.extra_vars.get('apply_detokenization', False):
if self.print_sources:
sources = list(map(self.extra_vars['detokenize_f'], sources))
predictions = list(map(self.extra_vars['detokenize_f'], predictions))
truths = list(map(self.extra_vars['detokenize_f'], truths))
# Write samples
if self.print_sources:
# Write samples
for i, (source, sample, truth) in list(enumerate(zip(sources, predictions, truths))):
if sys.version_info.major == 2:
source = str(source.encode('utf-8'))
sample = str(sample.encode('utf-8'))
truth = str(truth.encode('utf-8'))
print("Source (%d): %s" % (i, source))
print("Hypothesis (%d): %s" % (i, sample))
print("Reference (%d): %s" % (i, truth))
print("")
else:
for i, (sample, truth) in list(enumerate(zip(predictions, truths))):
if sys.version_info.major == 2:
sample = str(sample.encode('utf-8'))
truth = str(truth.encode('utf-8'))
print("Hypothesis (%d): %s" % (i, sample))
print("Reference (%d): %s" % (i, truth))
print("")