def update_config_with_metadata(output_feature, feature_metadata, *args,
**kwargs):
output_feature["vocab_size"] = feature_metadata["{}_vocab_size".format(
output_feature["level"])]
output_feature["max_sequence_length"] = feature_metadata[
"{}_max_sequence_length".format(output_feature["level"])]
if isinstance(output_feature[LOSS]["class_weights"], (list, tuple)):
# [0, 0] for UNK and PAD
output_feature[LOSS]["class_weights"] = [
0, 0
] + output_feature[LOSS]["class_weights"]
if len(output_feature[LOSS]
["class_weights"]) != output_feature["vocab_size"]:
raise ValueError(
"The length of class_weights ({}) is not compatible with "
"the number of classes ({})".format(
len(output_feature[LOSS]["class_weights"]),
output_feature["vocab_size"]))
if output_feature[LOSS]["class_similarities_temperature"] > 0:
if "class_similarities" in output_feature:
distances = output_feature["class_similarities"]
temperature = output_feature[LOSS][
"class_similarities_temperature"]
for i in range(len(distances)):
distances[i, :] = softmax(distances[i, :],
temperature=temperature)
output_feature[LOSS]["class_similarities"] = distances
else:
raise ValueError("class_similarities_temperature > 0,"
"but no class similarities are provided "
"for feature {}".format(
output_feature[COLUMN]))
def update_config_with_metadata(
output_feature,
feature_metadata,
*args,
**kwargs
):
output_feature['num_classes'] = feature_metadata[
'{}_vocab_size'.format(output_feature['level'])
]
output_feature['max_sequence_length'] = feature_metadata[
'{}_max_sequence_length'.format(output_feature['level'])
]
if isinstance(output_feature[LOSS]['class_weights'], (list, tuple)):
# [0, 0] for UNK and PAD
output_feature[LOSS]['class_weights'] = (
[0, 0] + output_feature[LOSS]['class_weights']
)
if (len(output_feature[LOSS]['class_weights']) !=
output_feature['num_classes']):
raise ValueError(
'The length of class_weights ({}) is not compatible with '
'the number of classes ({})'.format(
len(output_feature[LOSS]['class_weights']),
output_feature['num_classes']
)
)
if output_feature[LOSS]['class_similarities_temperature'] > 0:
if 'class_similarities' in output_feature:
distances = output_feature['class_similarities']
temperature = output_feature[LOSS][
'class_similarities_temperature']
for i in range(len(distances)):
distances[i, :] = softmax(
distances[i, :],
temperature=temperature
)
output_feature[LOSS]['class_similarities'] = distances
else:
raise ValueError(
'class_similarities_temperature > 0,'
'but no class similarities are provided '
'for feature {}'.format(output_feature[NAME])
)
if output_feature[LOSS][TYPE] == 'sampled_softmax_cross_entropy':
level_str2freq = '{}_str2freq'.format(output_feature['level'])
level_idx2str = '{}_idx2str'.format(output_feature['level'])
output_feature[LOSS]['class_counts'] = [
feature_metadata[level_str2freq][cls]
for cls in feature_metadata[level_idx2str]
]
def update_model_definition_with_metadata(
output_feature,
feature_metadata,
*args,
**kwargs
):
output_feature['num_classes'] = feature_metadata['vocab_size']
output_feature['max_sequence_length'] = (
feature_metadata['max_sequence_length']
)
if isinstance(output_feature[LOSS]['class_weights'], (list, tuple)):
output_feature[LOSS]['class_weights'] = (
[0, 0] + output_feature[LOSS]['class_weights']
)
# for UNK and PAD
if (len(output_feature[LOSS]['class_weights']) !=
output_feature['num_classes']):
raise ValueError(
'The length of class_weights ({}) is not compatible with '
'the number of classes ({})'.format(
len(output_feature[LOSS]['class_weights']),
output_feature['num_classes']
)
)
if output_feature[LOSS]['class_distance_temperature'] > 0:
if 'distances' in feature_metadata:
distances = feature_metadata['distances']
temperature = output_feature[LOSS]['class_distance_temperature']
for i in range(len(distances)):
distances[i, :] = softmax(
distances[i, :],
temperature=temperature
)
output_feature[LOSS]['distances'] = distances
else:
raise ValueError(
'No class distance metadata available '
'for feature {}'.format(output_feature['name'])
)
if output_feature[LOSS]['type'] == 'sampled_softmax_cross_entropy':
output_feature[LOSS]['class_counts'] = [(
feature_metadata['str2freq'][cls]
for cls in feature_metadata['idx2str']
)]
def update_config_with_metadata(output_feature, feature_metadata, *args,
**kwargs):
output_feature['num_classes'] = feature_metadata['vocab_size']
output_feature['top_k'] = min(output_feature['num_classes'],
output_feature['top_k'])
if isinstance(output_feature[LOSS]['class_weights'], (list, tuple)):
if (len(output_feature[LOSS]['class_weights']) !=
output_feature['num_classes']):
raise ValueError(
'The length of class_weights ({}) is not compatible with '
'the number of classes ({}) for feature {}. '
'Check the metadata JSON file to see the classes '
'and their order and consider there needs to be a weight '
'for the <UNK> class too.'.format(
len(output_feature[LOSS]['class_weights']),
output_feature['num_classes'], output_feature[COLUMN]))
if isinstance(output_feature[LOSS]['class_weights'], dict):
if (feature_metadata['str2idx'].keys() !=
output_feature[LOSS]['class_weights'].keys()):
raise ValueError(
'The class_weights keys ({}) are not compatible with '
'the classes ({}) of feature {}. '
'Check the metadata JSON file to see the classes '
'and consider there needs to be a weight '
'for the <UNK> class too.'.format(
output_feature[LOSS]['class_weights'].keys(),
feature_metadata['str2idx'].keys(),
output_feature[COLUMN]))
else:
class_weights = output_feature[LOSS]['class_weights']
idx2str = feature_metadata['idx2str']
class_weights_list = [class_weights[s] for s in idx2str]
output_feature[LOSS]['class_weights'] = class_weights_list
if output_feature[LOSS]['class_similarities_temperature'] > 0:
if 'class_similarities' in output_feature[LOSS]:
similarities = output_feature[LOSS]['class_similarities']
temperature = output_feature[LOSS][
'class_similarities_temperature']
curr_row = 0
first_row_length = 0
is_first_row = True
for row in similarities:
if is_first_row:
first_row_length = len(row)
is_first_row = False
curr_row += 1
else:
curr_row_length = len(row)
if curr_row_length != first_row_length:
raise ValueError(
'The length of row {} of the class_similarities '
'of {} is {}, different from the length of '
'the first row {}. All rows must have '
'the same length.'.format(
curr_row, output_feature[COLUMN],
curr_row_length, first_row_length))
else:
curr_row += 1
all_rows_length = first_row_length
if all_rows_length != len(similarities):
raise ValueError('The class_similarities matrix of {} has '
'{} rows and {} columns, '
'their number must be identical.'.format(
output_feature[COLUMN],
len(similarities), all_rows_length))
if all_rows_length != output_feature['num_classes']:
raise ValueError(
'The size of the class_similarities matrix of {} is '
'{}, different from the number of classe ({}). '
'Check the metadata JSON file to see the classes '
'and their order and '
'consider <UNK> class too.'.format(
output_feature[COLUMN], all_rows_length,
output_feature['num_classes']))
similarities = np.array(similarities, dtype=np.float32)
for i in range(len(similarities)):
similarities[i, :] = softmax(similarities[i, :],
temperature=temperature)
output_feature[LOSS]['class_similarities'] = similarities
else:
raise ValueError('class_similarities_temperature > 0, '
'but no class_similarities are provided '
'for feature {}'.format(
output_feature[COLUMN]))
if output_feature[LOSS][TYPE] == 'sampled_softmax_cross_entropy':
output_feature[LOSS]['class_counts'] = [
feature_metadata['str2freq'][cls]
for cls in feature_metadata['idx2str']
]
def update_config_with_metadata(output_feature, feature_metadata, *args,
**kwargs):
output_feature["vocab_size"] = feature_metadata["vocab_size"]
output_feature["max_sequence_length"] = feature_metadata[
"max_sequence_length"]
if isinstance(output_feature[LOSS]["class_weights"], (list, tuple)):
if len(output_feature[LOSS]
["class_weights"]) != output_feature["vocab_size"]:
raise ValueError(
"The length of class_weights ({}) is not compatible with "
"the number of classes ({}) for feature {}. "
"Check the metadata JSON file to see the classes "
"and their order and consider there needs to be a weight "
"for the <UNK> and <PAD> class too.".format(
len(output_feature[LOSS]["class_weights"]),
output_feature["vocab_size"],
output_feature[COLUMN],
))
if output_feature[LOSS]["class_similarities_temperature"] > 0:
if "class_similarities" in output_feature[LOSS]:
similarities = output_feature[LOSS]["class_similarities"]
temperature = output_feature[LOSS][
"class_similarities_temperature"]
curr_row = 0
first_row_length = 0
is_first_row = True
for row in similarities:
if is_first_row:
first_row_length = len(row)
is_first_row = False
curr_row += 1
else:
curr_row_length = len(row)
if curr_row_length != first_row_length:
raise ValueError(
"The length of row {} of the class_similarities "
"of {} is {}, different from the length of "
"the first row {}. All rows must have "
"the same length.".format(
curr_row, output_feature[COLUMN],
curr_row_length, first_row_length))
else:
curr_row += 1
all_rows_length = first_row_length
if all_rows_length != len(similarities):
raise ValueError("The class_similarities matrix of {} has "
"{} rows and {} columns, "
"their number must be identical.".format(
output_feature[COLUMN],
len(similarities), all_rows_length))
if all_rows_length != output_feature["vocab_size"]:
raise ValueError(
"The size of the class_similarities matrix of {} is "
"{}, different from the number of classes ({}). "
"Check the metadata JSON file to see the classes "
"and their order and "
"consider <UNK> and <PAD> class too.".format(
output_feature[COLUMN], all_rows_length,
output_feature["vocab_size"]))
similarities = np.array(similarities, dtype=np.float32)
for i in range(len(similarities)):
similarities[i, :] = softmax(similarities[i, :],
temperature=temperature)
output_feature[LOSS]["class_similarities"] = similarities
else:
raise ValueError("class_similarities_temperature > 0, "
"but no class_similarities are provided "
"for feature {}".format(
output_feature[COLUMN]))
def _predictions_training(self, inputs, training=None): # not executed
# inputs == logits
probs = softmax(inputs)
preds = tf.argmax(inputs)
return {PREDICTIONS: preds, PROBABILITIES: probs}