def transformer(
path,
s3_path,
class_name,
label,
model='bert',
quantized=False,
**kwargs,
):
check_file(path[model], s3_path[model], quantized=quantized, **kwargs)
if quantized:
model_path = 'quantized'
else:
model_path = 'model'
g = load_graph(path[model][model_path], **kwargs)
if len(label) > 2 or class_name == 'relevancy':
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
selected_class = MULTICLASS_BERT
selected_node = 'import/dense/BiasAdd:0'
if model in ['xlnet', 'alxlnet']:
selected_class = MULTICLASS_XLNET
selected_node = 'import/transpose_3:0'
else:
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
selected_class = BINARY_BERT
selected_node = 'import/dense/BiasAdd:0'
if model in ['xlnet', 'alxlnet']:
selected_class = BINARY_XLNET
selected_node = 'import/transpose_3:0'
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
if model in ['bert', 'tiny-bert']:
from malaya.transformers.bert import (
_extract_attention_weights_import, )
from malaya.transformers.bert import bert_num_layers
tokenizer = sentencepiece_tokenizer_bert(path[model]['tokenizer'],
path[model]['vocab'])
if model in ['albert', 'tiny-albert']:
from malaya.transformers.albert import (
_extract_attention_weights_import, )
from malaya.transformers.albert import bert_num_layers
from albert import tokenization
tokenizer = tokenization.FullTokenizer(
vocab_file=path[model]['vocab'],
do_lower_case=False,
spm_model_file=path[model]['tokenizer'],
)
return selected_class(
X=g.get_tensor_by_name('import/Placeholder:0'),
segment_ids=None,
input_masks=g.get_tensor_by_name('import/Placeholder_1:0'),
logits=g.get_tensor_by_name('import/logits:0'),
logits_seq=g.get_tensor_by_name('import/logits_seq:0'),
vectorizer=g.get_tensor_by_name(selected_node),
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
label=label,
attns=_extract_attention_weights_import(bert_num_layers[model], g),
class_name=class_name,
)
if model in ['xlnet', 'alxlnet']:
if model in ['xlnet']:
from malaya.transformers.xlnet import (
_extract_attention_weights_import, )
if model in ['alxlnet']:
from malaya.transformers.alxlnet import (
_extract_attention_weights_import, )
tokenizer = sentencepiece_tokenizer_xlnet(path[model]['tokenizer'])
return selected_class(
X=g.get_tensor_by_name('import/Placeholder:0'),
segment_ids=g.get_tensor_by_name('import/Placeholder_1:0'),
input_masks=g.get_tensor_by_name('import/Placeholder_2:0'),
logits=g.get_tensor_by_name('import/logits:0'),
logits_seq=g.get_tensor_by_name('import/logits_seq:0'),
vectorizer=g.get_tensor_by_name(selected_node),
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
label=label,
attns=_extract_attention_weights_import(g),
class_name=class_name,
)
def transformer(model: str = 'xlnet', **kwargs):
"""
Load Transformer toxicity model.
Parameters
----------
model : str, optional (default='bert')
Model architecture supported. Allowed values:
* ``'bert'`` - BERT architecture from google.
* ``'tiny-bert'`` - BERT architecture from google with smaller parameters.
* ``'albert'`` - ALBERT architecture from google.
* ``'tiny-albert'`` - ALBERT architecture from google with smaller parameters.
* ``'xlnet'`` - XLNET architecture from google.
* ``'alxlnet'`` - XLNET architecture from google + Malaya.
Returns
-------
result : malaya.model.bert.SIGMOID_BERT class
"""
model = model.lower()
if model not in _availability:
raise Exception(
'model not supported, please check supported models from malaya.sentiment.available_transformer()'
)
check_file(PATH_TOXIC[model], S3_PATH_TOXIC[model], **kwargs)
g = load_graph(PATH_TOXIC[model]['model'])
path = PATH_TOXIC
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
if model in ['bert', 'tiny-bert']:
from malaya.transformers.bert import (
_extract_attention_weights_import, )
from malaya.transformers.bert import bert_num_layers
tokenizer = sentencepiece_tokenizer_bert(path[model]['tokenizer'],
path[model]['vocab'])
if model in ['albert', 'tiny-albert']:
from malaya.transformers.albert import (
_extract_attention_weights_import, )
from malaya.transformers.albert import bert_num_layers
from albert import tokenization
tokenizer = tokenization.FullTokenizer(
vocab_file=path[model]['vocab'],
do_lower_case=False,
spm_model_file=path[model]['tokenizer'],
)
return SIGMOID_BERT(
X=g.get_tensor_by_name('import/Placeholder:0'),
segment_ids=None,
input_masks=g.get_tensor_by_name('import/Placeholder_1:0'),
logits=g.get_tensor_by_name('import/logits:0'),
logits_seq=g.get_tensor_by_name('import/logits_seq:0'),
sess=generate_session(graph=g),
tokenizer=tokenizer,
label=label,
attns=_extract_attention_weights_import(bert_num_layers[model], g),
class_name='toxic',
)
if model in ['xlnet', 'alxlnet']:
if model in ['xlnet']:
from malaya.transformers.xlnet import (
_extract_attention_weights_import, )
if model in ['alxlnet']:
from malaya.transformers.alxlnet import (
_extract_attention_weights_import, )
tokenizer = sentencepiece_tokenizer_xlnet(path[model]['tokenizer'])
return SIGMOID_XLNET(
X=g.get_tensor_by_name('import/Placeholder:0'),
segment_ids=g.get_tensor_by_name('import/Placeholder_1:0'),
input_masks=g.get_tensor_by_name('import/Placeholder_2:0'),
logits=g.get_tensor_by_name('import/logits:0'),
logits_seq=g.get_tensor_by_name('import/logits_seq:0'),
sess=generate_session(graph=g),
tokenizer=tokenizer,
label=label,
attns=_extract_attention_weights_import(g),
class_name='toxic',
)
def transformer(model: str = 'xlnet', quantized: bool = False, **kwargs):
"""
Load Transformer toxicity model.
Parameters
----------
model : str, optional (default='bert')
Model architecture supported. Allowed values:
* ``'bert'`` - Google BERT BASE parameters.
* ``'tiny-bert'`` - Google BERT TINY parameters.
* ``'albert'`` - Google ALBERT BASE parameters.
* ``'tiny-albert'`` - Google ALBERT TINY parameters.
* ``'xlnet'`` - Google XLNET BASE parameters.
* ``'alxlnet'`` - Malaya ALXLNET BASE parameters.
quantized : bool, optional (default=False)
if True, will load 8-bit quantized model.
Quantized model not necessary faster, totally depends on the machine.
Returns
-------
result : malaya.model.bert.SIGMOID_BERT class
"""
model = model.lower()
if model not in _transformer_availability:
raise Exception(
'model not supported, please check supported models from `malaya.toxicity.available_transformer()`.'
)
check_file(
PATH_TOXIC[model], S3_PATH_TOXIC[model], quantized = quantized, **kwargs
)
if quantized:
model_path = 'quantized'
else:
model_path = 'model'
g = load_graph(PATH_TOXIC[model][model_path], **kwargs)
path = PATH_TOXIC
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
if model in ['bert', 'tiny-bert']:
from malaya.transformers.bert import (
_extract_attention_weights_import,
)
from malaya.transformers.bert import bert_num_layers
tokenizer = sentencepiece_tokenizer_bert(
path[model]['tokenizer'], path[model]['vocab']
)
if model in ['albert', 'tiny-albert']:
from malaya.transformers.albert import (
_extract_attention_weights_import,
)
from malaya.transformers.albert import bert_num_layers
from albert import tokenization
tokenizer = tokenization.FullTokenizer(
vocab_file = path[model]['vocab'],
do_lower_case = False,
spm_model_file = path[model]['tokenizer'],
)
return SIGMOID_BERT(
X = g.get_tensor_by_name('import/Placeholder:0'),
segment_ids = None,
input_masks = g.get_tensor_by_name('import/Placeholder_1:0'),
logits = g.get_tensor_by_name('import/logits:0'),
logits_seq = g.get_tensor_by_name('import/logits_seq:0'),
vectorizer = g.get_tensor_by_name('import/dense/BiasAdd:0'),
sess = generate_session(graph = g, **kwargs),
tokenizer = tokenizer,
label = label,
attns = _extract_attention_weights_import(
bert_num_layers[model], g
),
class_name = 'toxic',
)
if model in ['xlnet', 'alxlnet']:
if model in ['xlnet']:
from malaya.transformers.xlnet import (
_extract_attention_weights_import,
)
if model in ['alxlnet']:
from malaya.transformers.alxlnet import (
_extract_attention_weights_import,
)
tokenizer = sentencepiece_tokenizer_xlnet(path[model]['tokenizer'])
return SIGMOID_XLNET(
X = g.get_tensor_by_name('import/Placeholder:0'),
segment_ids = g.get_tensor_by_name('import/Placeholder_1:0'),
input_masks = g.get_tensor_by_name('import/Placeholder_2:0'),
logits = g.get_tensor_by_name('import/logits:0'),
logits_seq = g.get_tensor_by_name('import/logits_seq:0'),
vectorizer = g.get_tensor_by_name('import/transpose_3:0'),
sess = generate_session(graph = g, **kwargs),
tokenizer = tokenizer,
label = label,
attns = _extract_attention_weights_import(g),
class_name = 'toxic',
)
def transformer(
class_name,
label,
model='bert',
sigmoid=False,
quantized=False,
**kwargs,
):
path = check_file(
file=model,
module=class_name,
keys={
'model': 'model.pb',
'vocab': MODEL_VOCAB[model],
'tokenizer': MODEL_BPE[model],
},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
if sigmoid:
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
selected_class = SigmoidBERT
if model in ['xlnet', 'alxlnet']:
selected_class = SigmoidXLNET
else:
if len(label) > 2 or class_name == 'relevancy':
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
selected_class = MulticlassBERT
if model in ['xlnet', 'alxlnet']:
selected_class = MulticlassXLNET
if model in ['bigbird', 'tiny-bigbird']:
selected_class = MulticlassBigBird
else:
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
selected_class = BinaryBERT
if model in ['xlnet', 'alxlnet']:
selected_class = BinaryXLNET
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
if model in ['bert', 'tiny-bert']:
from malaya.transformers.bert import (
_extract_attention_weights_import, )
from malaya.transformers.bert import bert_num_layers
tokenizer = sentencepiece_tokenizer_bert(path['tokenizer'],
path['vocab'])
if model in ['albert', 'tiny-albert']:
from malaya.transformers.albert import (
_extract_attention_weights_import, )
from malaya.transformers.albert import bert_num_layers
from malaya.transformers.albert import tokenization
tokenizer = tokenization.FullTokenizer(
vocab_file=path['vocab'],
do_lower_case=False,
spm_model_file=path['tokenizer'],
)
inputs = ['Placeholder', 'Placeholder_1']
vectorizer = {'vectorizer': 'import/dense/BiasAdd:0'}
attention = _extract_attention_weights_import(bert_num_layers[model],
g)
if model in ['xlnet', 'alxlnet']:
if model in ['xlnet']:
from malaya.transformers.xlnet import (
_extract_attention_weights_import, )
if model in ['alxlnet']:
from malaya.transformers.alxlnet import (
_extract_attention_weights_import, )
inputs = ['Placeholder', 'Placeholder_1', 'Placeholder_2']
tokenizer = sentencepiece_tokenizer_xlnet(path['tokenizer'])
vectorizer = {'vectorizer': 'import/transpose_3:0'}
attention = _extract_attention_weights_import(g)
if model in ['bigbird', 'tiny-bigbird']:
inputs = ['Placeholder']
tokenizer = sentencepiece_tokenizer_bert(path['tokenizer'],
path['vocab'])
vectorizer = {'vectorizer': 'import/dense/BiasAdd:0'}
attention = None
outputs = ['logits', 'logits_seq']
input_nodes, output_nodes = nodes_session(
g,
inputs,
outputs,
extra=vectorizer,
attention={'attention': attention},
)
return selected_class(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
label=label,
class_name=class_name,
)
