def load(module, model, model_class, quantized=False, **kwargs):
try:
import tensorflow_text
except BaseException:
raise ModuleNotFoundError(
'tensorflow-text not installed. Please install it by `pip install tensorflow-text` and try again. Also, make sure tensorflow-text version same as tensorflow version.'
)
path = check_file(
file=model,
module=module,
keys={'model': 'model.pb', 'vocab': VOCAB_MODEL.get(module, MS_EN_BPE_MODEL)},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], t5_graph=True, **kwargs)
tokenizer = SentencePieceBatchEncoder(vocab_file=path['vocab'])
inputs = ['inputs']
outputs = []
input_nodes, output_nodes = nodes_session(
g, inputs, outputs, extra={'decode': 'import/SelectV2_3:0'}
)
return model_class(
input_nodes=input_nodes, output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
)
def load(module, model, encoder, model_class, quantized=False, **kwargs):
path = check_file(
file=model,
module=module,
keys={
'model': 'model.pb',
'vocab': LM_VOCAB[module]
},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
encoder = ENCODER_MODEL[encoder](vocab_file=path['vocab'], id_mode=True)
inputs = ['Placeholder']
outputs = ['greedy', 'beam']
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
return model_class(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
encoder=encoder,
)
def _transformer(model,
bert_class,
xlnet_class,
quantized=False,
siamese=False,
**kwargs):
model = model.lower()
if model not in _transformer_availability:
raise ValueError(
'model not supported, please check supported models from `malaya.similarity.available_transformer()`.'
)
path = check_file(
file=model,
module='similarity',
keys={
'model': 'model.pb',
'vocab': MODEL_VOCAB[model],
'tokenizer': MODEL_BPE[model],
},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
if model in ['bert', 'tiny-bert']:
tokenizer = sentencepiece_tokenizer_bert(path['tokenizer'],
path['vocab'])
if model in ['albert', 'tiny-albert']:
tokenizer = AlbertTokenizer(vocab_file=path['vocab'],
spm_model_file=path['tokenizer'])
selected_class = bert_class
if siamese:
selected_node = 'import/bert/pooler/dense/BiasAdd:0'
if model in ['xlnet', 'alxlnet']:
tokenizer = sentencepiece_tokenizer_xlnet(path['tokenizer'])
selected_class = xlnet_class
if siamese:
selected_node = 'import/model_1/sequnece_summary/summary/BiasAdd:0'
if not siamese:
selected_node = _vectorizer_mapping[model]
inputs = ['Placeholder', 'Placeholder_1', 'Placeholder_2']
outputs = ['logits']
input_nodes, output_nodes = nodes_session(
g, inputs, outputs, extra={'vectorizer': selected_node})
return selected_class(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
label=['not similar', 'similar'],
)
def load(module, model, model_class, maxlen, quantized=False, **kwargs):
path = check_file(
file=model,
module=module,
keys={
'model': 'model.pb',
'vocab': TRANSLATION_BPE_MODEL
},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
inputs = ['Placeholder']
outputs = ['logits']
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
encoder = SentencePieceEncoder(vocab_file=path['vocab'])
return model_class(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
encoder=encoder,
maxlen=maxlen,
)
def load(module, model, encoder, model_class, quantized=False, **kwargs):
path = check_file(
file=model,
module=module,
keys={
'model': 'model.pb',
'vocab': LM_VOCAB[module]
},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
if encoder == 'subword':
encoder = text_encoder.SubwordTextEncoder(path['vocab'])
if encoder == 'yttm':
bpe, subword_mode = load_yttm(path['vocab'], True)
encoder = YTTMEncoder(bpe, subword_mode)
inputs = ['Placeholder']
outputs = ['greedy', 'beam']
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
return model_class(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
encoder=encoder,
)
def load_lm(path, s3_path, model, model_class, 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)
X = g.get_tensor_by_name('import/Placeholder:0')
top_p = g.get_tensor_by_name('import/Placeholder_2:0')
greedy = g.get_tensor_by_name('import/greedy:0')
beam = g.get_tensor_by_name('import/beam:0')
nucleus = g.get_tensor_by_name('import/nucleus:0')
tokenizer = SentencePieceEncoder(path[model]['vocab'])
return model_class(
X,
top_p,
greedy,
beam,
nucleus,
generate_session(graph=g, **kwargs),
tokenizer,
)
def _transformer(model,
bert_class,
xlnet_class,
quantized=False,
siamese=False,
**kwargs):
model = model.lower()
if model not in _transformer_availability:
raise ValueError(
'model not supported, please check supported models from `malaya.similarity.available_transformer()`.'
)
check_file(PATH_SIMILARITY[model],
S3_PATH_SIMILARITY[model],
quantized=quantized,
**kwargs)
if quantized:
model_path = 'quantized'
else:
model_path = 'model'
g = load_graph(PATH_SIMILARITY[model][model_path], **kwargs)
path = PATH_SIMILARITY
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
if model in ['bert', 'tiny-bert']:
tokenizer = sentencepiece_tokenizer_bert(path[model]['tokenizer'],
path[model]['vocab'])
if model in ['albert', 'tiny-albert']:
tokenizer = tokenization.FullTokenizer(
vocab_file=path[model]['vocab'],
do_lower_case=False,
spm_model_file=path[model]['tokenizer'],
)
selected_class = bert_class
if siamese:
selected_node = 'import/bert/pooler/dense/BiasAdd:0'
if model in ['xlnet', 'alxlnet']:
tokenizer = sentencepiece_tokenizer_xlnet(path[model]['tokenizer'])
selected_class = xlnet_class
if siamese:
selected_node = 'import/model_1/sequnece_summary/summary/BiasAdd:0'
if not siamese:
selected_node = _vectorizer_mapping[model]
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'),
vectorizer=g.get_tensor_by_name(selected_node),
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
label=['not similar', 'similar'],
)
def load(path,
s3_path,
model,
encoder,
model_class,
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 encoder == 'subword':
encoder = text_encoder.SubwordTextEncoder(path[model]['vocab'])
if encoder == 'yttm':
bpe, subword_mode = load_yttm(path[model]['vocab'], True)
encoder = YTTMEncoder(bpe, subword_mode)
return model_class(
g.get_tensor_by_name('import/Placeholder:0'),
g.get_tensor_by_name('import/greedy:0'),
g.get_tensor_by_name('import/beam:0'),
generate_session(graph=g, **kwargs),
encoder,
)
def load_tatabahasa(module, model, model_class, quantized=False, **kwargs):
path = check_file(
file=model,
module=module,
keys={
'model': 'model.pb',
'vocab': T2T_BPE_MODEL
},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
tokenizer = SentencePieceEncoder(vocab_file=path['vocab'])
inputs = ['x_placeholder']
outputs = ['greedy', 'tag_greedy']
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
return model_class(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
)
def load_pegasus(module,
model,
model_class,
maxlen,
quantized=False,
**kwargs):
path = check_file(
file=model,
module=module,
keys={
'model': 'model.pb',
'vocab': PEGASUS_BPE_MODEL
},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
inputs = ['Placeholder', 'top_p', 'temperature']
outputs = ['logits']
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
tokenizer = WordPieceTokenizer(vocab_file=path['vocab'])
return model_class(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
maxlen=maxlen,
)
def load_lm(module, model, model_class, quantized=False, **kwargs):
path = check_file(
file=model,
module=module,
keys={
'model': 'model.pb',
'vocab': T2T_BPE_MODEL
},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
X = g.get_tensor_by_name('import/Placeholder:0')
top_p = g.get_tensor_by_name('import/Placeholder_2:0')
greedy = g.get_tensor_by_name('import/greedy:0')
beam = g.get_tensor_by_name('import/beam:0')
nucleus = g.get_tensor_by_name('import/nucleus:0')
tokenizer = SentencePieceEncoder(path['vocab'])
inputs = ['Placeholder', 'Placeholder_2']
outputs = ['greedy', 'beam', 'nucleus']
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
return model_class(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
)
def transformer(model: str = 'xlnet', quantized: bool = False, **kwargs):
"""
Load Transformer Constituency Parsing model, transfer learning Transformer + self attentive parsing.
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.
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.tf.Constituency class
"""
model = model.lower()
if model not in _transformer_availability:
raise ValueError(
'model not supported, please check supported models from `malaya.constituency.available_transformer()`.'
)
path = check_file(
file=model,
module='constituency',
keys={
'model': 'model.pb',
'vocab': MODEL_VOCAB[model],
'tokenizer': MODEL_BPE[model],
},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
inputs = ['input_ids', 'word_end_mask']
outputs = ['charts', 'tags']
tokenizer = SentencePieceTokenizer(vocab_file=path['vocab'],
spm_model_file=path['tokenizer'])
input_nodes, output_nodes = nodes_session(
g, inputs, outputs, extra={'vectorizer': _vectorizer_mapping[model]})
mode = 'bert' if 'bert' in model else 'xlnet'
return Constituency(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
dictionary=settings.constituency,
mode=mode,
)
def load(path,
s3_path,
model,
model_class,
compressed=True,
quantized=False,
**kwargs):
try:
import tensorflow_text
import tf_sentencepiece
except:
raise ModuleNotFoundError(
'tensorflow-text and tf-sentencepiece not installed. Please install it by `pip install tensorflow-text==1.15.0 tf-sentencepiece==0.1.86` and try again. Also, make sure tensorflow-text version same as tensorflow version.'
)
if compressed and not quantized:
path = path['t5-compressed']
s3_path = s3_path['t5-compressed']
check_file(path[model]['model'], s3_path[model], **kwargs)
if not os.path.exists(path[model]['directory'] + 'saved_model.pb'):
import tarfile
with tarfile.open(path[model]['model']['model']) as tar:
tar.extractall(path=path[model]['path'])
X = None
decode = None
sess = generate_session(graph=None, **kwargs)
meta_graph_def = tf.compat.v1.saved_model.load(
sess, ['serve'], path[model]['directory'])
signature_def = meta_graph_def.signature_def['serving_default']
pred = lambda x: sess.run(
fetches=signature_def.outputs['outputs'].name,
feed_dict={signature_def.inputs['input'].name: x},
)
else:
path = path['t5']
s3_path = s3_path['t5']
check_file(path[model],
s3_path[model],
quantized=quantized,
optimized=True,
**kwargs)
if quantized:
model_path = 'quantized'
else:
model_path = 'model'
g = load_graph(path[model][model_path], **kwargs)
X = g.get_tensor_by_name('import/inputs:0')
decode = g.get_tensor_by_name(
'import/SentenceTokenizer_1/SentenceTokenizer/SentencepieceDetokenizeOp:0'
)
sess = generate_session(graph=g, **kwargs)
pred = None
return model_class(X=X, decode=decode, sess=sess, pred=pred)
def transformer(path,
s3_path,
class_name,
model='xlnet',
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)
try:
with open(path[model]['setting']) as fopen:
nodes = json.load(fopen)
except:
raise Exception(
f"model corrupted due to some reasons, please run malaya.clear_cache('{class_name}/{model}/{size}') and try again"
)
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
if model in ['bert', 'tiny-bert']:
tokenizer = sentencepiece_tokenizer_bert(path[model]['tokenizer'],
path[model]['vocab'])
if model in ['albert', 'tiny-albert']:
from albert import tokenization
tokenizer = tokenization.FullTokenizer(
vocab_file=path[model]['vocab'],
do_lower_case=False,
spm_model_file=path[model]['tokenizer'],
)
return TAGGING_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'),
vectorizer=g.get_tensor_by_name('import/dense/BiasAdd:0'),
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
settings=nodes,
)
if model in ['xlnet', 'alxlnet']:
tokenizer = sentencepiece_tokenizer_xlnet(path[model]['tokenizer'])
return TAGGING_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'),
vectorizer=g.get_tensor_by_name('import/transpose_3:0'),
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
settings=nodes,
)
def transformer(model: str = 'xlnet', **kwargs):
"""
Load Transformer Constituency Parsing model, transfer learning Transformer + self attentive parsing.
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.
Returns
-------
result : malaya.model.tf.CONSTITUENCY class
"""
model = model.lower()
if model not in _transformer_availability:
raise ValueError(
'model not supported, please check supported models from malaya.constituency.available_transformer()'
)
check_file(PATH_CONSTITUENCY[model], S3_PATH_CONSTITUENCY[model], **kwargs)
g = load_graph(PATH_CONSTITUENCY[model]['model'], **kwargs)
with open(PATH_CONSTITUENCY[model]['dictionary']) as fopen:
dictionary = json.load(fopen)
if model in ['bert', 'tiny-bert', 'albert', 'tiny-albert']:
tokenizer = sentencepiece_tokenizer_bert(
PATH_CONSTITUENCY[model]['tokenizer'],
PATH_CONSTITUENCY[model]['vocab'],
)
mode = 'bert'
if model in ['xlnet']:
tokenizer = sentencepiece_tokenizer_xlnet(
PATH_CONSTITUENCY[model]['tokenizer'])
mode = 'xlnet'
from malaya.model.tf import CONSTITUENCY
return CONSTITUENCY(
input_ids=g.get_tensor_by_name('import/input_ids:0'),
word_end_mask=g.get_tensor_by_name('import/word_end_mask:0'),
charts=g.get_tensor_by_name('import/charts:0'),
tags=g.get_tensor_by_name('import/tags:0'),
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
dictionary=dictionary,
mode=mode,
)
def deep_model(quantized: bool = False, **kwargs):
"""
Load deep learning language detection model.
Original size is 51.2MB, Quantized size 12.8MB.
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.tf.DeepLang class
"""
path = check_file(
file = 'lang-32',
module = 'language-detection',
keys = {
'model': 'model.pb',
'vector': LANGUAGE_DETECTION_BOW,
'bpe': LANGUAGE_DETECTION_VOCAB,
},
quantized = quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
bpe, subword_mode = load_yttm(path['bpe'])
try:
with open(path['vector'], 'rb') as fopen:
vector = pickle.load(fopen)
except:
raise ValueError(
"model corrupted due to some reasons, please run `malaya.clear_cache('language-detection/lang-32')` and try again"
)
inputs = [
'X_Placeholder/shape',
'X_Placeholder/values',
'X_Placeholder/indices',
'W_Placeholder/shape',
'W_Placeholder/values',
'W_Placeholder/indices',
]
outputs = ['logits']
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
return DeepLang(
input_nodes = input_nodes,
output_nodes = output_nodes,
sess = generate_session(graph = g, **kwargs),
vectorizer = vector,
bpe = bpe,
type = subword_mode,
label = lang_labels,
)
def deep_model(quantized: bool = False, **kwargs):
"""
Load LSTM + Bahdanau Attention stemming model, this also include lemmatization.
Original size 41.6MB, quantized size 10.6MB .
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.stem.DeepStemmer class
"""
if check_tf_version() > 1:
raise Exception(
f'Tensorflow 2.0 and above not able to use `deep_model` for stemmer, use Tensorflow 1.15 instead.'
)
path = check_file(
file='lstm-bahdanau',
module='stem',
keys={
'model': 'model.pb',
'vocab': STEMMER_VOCAB
},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
bpe, subword_mode = load_yttm(path['vocab'], id_mode=True)
inputs = ['Placeholder']
outputs = []
input_nodes, output_nodes = nodes_session(
g,
inputs,
outputs,
extra={
'greedy': 'import/decode_1/greedy:0',
'beam': 'import/decode_2/beam:0',
},
)
tokenizer = Tokenizer().tokenize
return DeepStemmer(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
bpe=bpe,
subword_mode=subword_mode,
tokenizer=tokenizer,
)
def transformer(class_name, model='xlnet', quantized=False, **kwargs):
path = check_file(
file=model,
module=class_name,
keys={
'model': 'model.pb',
'vocab': MODEL_VOCAB[model],
'tokenizer': MODEL_BPE[model],
'setting': TAGGING_SETTING[class_name],
},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
try:
with open(path['setting']) as fopen:
nodes = json.load(fopen)
except:
raise Exception(
f"model corrupted due to some reasons, please run `malaya.clear_cache('{class_name}/{model}/{size}')` and try again"
)
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
if model in ['bert', 'tiny-bert']:
tokenizer = sentencepiece_tokenizer_bert(path['tokenizer'],
path['vocab'])
if model in ['albert', 'tiny-albert']:
tokenizer = AlbertTokenizer(vocab_file=path['vocab'],
spm_model_file=path['tokenizer'])
inputs = ['Placeholder', 'Placeholder_1']
vectorizer = {'vectorizer': 'import/dense/BiasAdd:0'}
Model = TaggingBERT
if model in ['xlnet', 'alxlnet']:
tokenizer = sentencepiece_tokenizer_xlnet(path['tokenizer'])
inputs = ['Placeholder', 'Placeholder_1', 'Placeholder_2']
vectorizer = {'vectorizer': 'import/transpose_3:0'}
Model = TaggingXLNET
outputs = ['logits']
input_nodes, output_nodes = nodes_session(g,
inputs,
outputs,
extra=vectorizer)
return Model(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
settings=nodes,
)
def transformer(module, model='xlnet', quantized=False, tok=None, **kwargs):
path = check_file(
file=model,
module=module,
keys={
'model': 'model.pb',
'vocab': MODEL_VOCAB[model],
'tokenizer': MODEL_BPE[model],
},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
inputs = ['Placeholder', 'Placeholder_1']
vectorizer = {'vectorizer': 'import/dense/BiasAdd:0'}
selected_model = TaggingBERT
selected_tokenizer = SentencePieceTokenizer
elif model in ['xlnet', 'alxlnet']:
inputs = ['Placeholder', 'Placeholder_1', 'Placeholder_2']
vectorizer = {'vectorizer': 'import/transpose_3:0'}
selected_model = TaggingXLNET
selected_tokenizer = SentencePieceTokenizer
elif model in ['fastformer', 'tiny-fastformer']:
inputs = ['Placeholder']
vectorizer_nodes = {'fastformer': 'import/fast_transformer/add_24:0',
'tiny-fastformer': 'import/fast_transformer/add_8:0'}
vectorizer = {'vectorizer': vectorizer_nodes[model]}
selected_model = TaggingFastFormer
selected_tokenizer = WordPieceTokenizer
outputs = ['logits']
tokenizer = selected_tokenizer(vocab_file=path['vocab'], spm_model_file=path['tokenizer'])
input_nodes, output_nodes = nodes_session(
g, inputs, outputs, extra=vectorizer
)
return selected_model(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
settings=TAGGING_SETTING[module],
tok=tok
)
def transformer_squad(class_name, model = 'bert', 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)
inputs = ['Placeholder', 'Placeholder_1', 'Placeholder_2', 'Placeholder_3']
if model in ['bert', 'tiny-bert']:
tokenizer = sentencepiece_tokenizer_bert(
path['tokenizer'], path['vocab']
)
if model in ['albert', 'tiny-albert']:
tokenizer = AlbertTokenizer(
vocab_file = path['vocab'], spm_model_file = path['tokenizer']
)
if model in ['xlnet', 'alxlnet']:
tokenizer = sentencepiece_tokenizer_xlnet(path['tokenizer'])
inputs.append('Placeholder_4')
outputs = [
'start_top_log_probs',
'start_top_index',
'end_top_log_probs',
'end_top_index',
'cls_logits',
'logits_vectorize',
]
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
mode = 'bert' if 'bert' in model else 'xlnet'
return SQUAD(
input_nodes = input_nodes,
output_nodes = output_nodes,
sess = generate_session(graph = g, **kwargs),
tokenizer = tokenizer,
class_name = class_name,
mode = mode,
length = LENGTHS[mode],
)
def load(module, model, model_class, quantized=False, **kwargs):
try:
import tensorflow_text
except BaseException:
raise ModuleNotFoundError(
'tensorflow-text not installed. Please install it by `pip install tensorflow-text` and try again. Also, make sure tensorflow-text version same as tensorflow version.'
)
if model.split('-')[-1] == '4k':
default_vocab = MS_EN_4k_BPE_MODEL
else:
default_vocab = MS_EN_BPE_MODEL
path = check_file(
file=model,
module=module,
keys={
'model': 'model.pb',
'vocab': VOCAB_MODEL.get(module, default_vocab)
},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], t5_graph=True, **kwargs)
tokenizer = SentencePieceBatchEncoder(vocab_file=path['vocab'])
inputs = ['inputs']
outputs = []
input_nodes, output_nodes = nodes_session(
g, inputs, outputs, extra={'decode': 'import/SelectV2_3:0'})
if module == 'kesalahan-tatabahasa':
word_tokenizer = Tokenizer(date=False, time=False).tokenize
elif module == 'spelling-correction':
word_tokenizer = Tokenizer(duration=False, date=False).tokenize
else:
word_tokenizer = None
return model_class(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
word_tokenizer=word_tokenizer,
)
def wiki_load_model():
if not os.path.isfile(PATH_SUMMARIZE['wiki']['model']):
print('downloading SUMMARIZE wikipedia frozen model')
download_file(S3_PATH_SUMMARIZE['wiki']['model'],
PATH_SUMMARIZE['wiki']['model'])
if not os.path.isfile(PATH_SUMMARIZE['wiki']['setting']):
print('downloading SUMMARIZE wikipedia dictionary')
download_file(
S3_PATH_SUMMARIZE['wiki']['setting'],
PATH_SUMMARIZE['wiki']['setting'],
)
g = load_graph(PATH_SUMMARIZE['wiki']['model'])
x = g.get_tensor_by_name('import/Placeholder_1:0')
logits = g.get_tensor_by_name('import/logits:0')
attention = g.get_tensor_by_name('import/attention:0')
sess = tf.InteractiveSession(graph=g)
with open(PATH_SUMMARIZE['wiki']['setting']) as fopen:
dictionary = json.load(fopen)
return sess, x, logits, attention, dictionary, 50
def transformer(
path, s3_path, class_name, model = 'xlnet', size = 'base', **kwargs
):
check_file(path[model][size], s3_path[model][size], **kwargs)
try:
with open(path[model][size]['setting']) as fopen:
nodes = json.load(fopen)
g = load_graph(path[model][size]['model'])
except:
raise Exception(
f"model corrupted due to some reasons, please run malaya.clear_cache('{class_name}/{model}/{size}') and try again"
)
if model in ['albert', 'bert']:
tokenizer, cls, sep = sentencepiece_tokenizer_bert(
path[model][size]['tokenizer'], path[model][size]['vocab']
)
return TAGGING_BERT(
X = g.get_tensor_by_name('import/Placeholder:0'),
segment_ids = None,
input_masks = None,
logits = g.get_tensor_by_name('import/logits:0'),
sess = generate_session(graph = g),
tokenizer = tokenizer,
cls = cls,
sep = sep,
settings = nodes,
)
if model in ['xlnet']:
tokenizer = sentencepiece_tokenizer_xlnet(
path[model][size]['tokenizer']
)
return TAGGING_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'),
sess = generate_session(graph = g),
tokenizer = tokenizer,
settings = nodes,
)
def load_char(module, model, left_dict, cleaning, quantized=False, **kwargs):
path = check_file(
file=model,
module=module,
keys={'model': 'model.pb'},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
inputs = ['x_placeholder']
outputs = ['greedy', 'tag_greedy']
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
return TransformerChar(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
left_dict=left_dict,
cleaning=cleaning,
)
def transformer(model='base', **kwargs):
"""
Load transformer encoder-decoder model to translate MS-to-EN.
Parameters
----------
model : str, optional (default='base')
Model architecture supported. Allowed values:
* ``'small'`` - transformer Small parameters.
* ``'base'`` - transformer Base parameters.
* ``'large'`` - transformer Large parameters.
Returns
-------
result: malaya.model.tf.TRANSLATION class
"""
model = model.lower()
if model not in _transformer_availability:
raise Exception(
'model not supported, please check supported models from malaya.translation.ms_en.available_transformer()'
)
path = PATH_TRANSLATION['ms-en']
s3_path = S3_PATH_TRANSLATION['ms-en']
check_file(path[model], s3_path[model], **kwargs)
g = load_graph(path[model]['model'], **kwargs)
from malaya.text.t2t import text_encoder
from malaya.model.tf import TRANSLATION
encoder = text_encoder.SubwordTextEncoder(path[model]['vocab'])
return TRANSLATION(
g.get_tensor_by_name('import/Placeholder:0'),
g.get_tensor_by_name('import/greedy:0'),
g.get_tensor_by_name('import/beam:0'),
generate_session(graph=g, **kwargs),
encoder,
)
def deep_model(quantized: bool = False, **kwargs):
"""
Load LSTM + Bahdanau Attention stemming model, this also include lemmatization.
Original size 41.6MB, quantized size 10.6MB .
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.stem.DEEP_STEMMER class
"""
from malaya.preprocessing import _tokenizer
check_file(PATH_STEM['deep'],
S3_PATH_STEM['deep'],
quantized=quantized,
**kwargs)
if quantized:
model_path = 'quantized'
else:
model_path = 'model'
g = load_graph(PATH_STEM['deep'][model_path], **kwargs)
bpe, subword_mode = load_yttm(PATH_STEM['deep']['bpe'], id_mode=True)
return DEEP_STEMMER(
g.get_tensor_by_name('import/Placeholder:0'),
g.get_tensor_by_name('import/decode_1/greedy:0'),
g.get_tensor_by_name('import/decode_2/beam:0'),
generate_session(graph=g, **kwargs),
bpe,
subword_mode,
_tokenizer,
)
def load_tatabahasa(path,
s3_path,
model,
model_class,
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)
tokenizer = SentencePieceEncoder(path[model]['vocab'])
return model_class(
X=g.get_tensor_by_name('import/x_placeholder:0'),
greedy=g.get_tensor_by_name('import/greedy:0'),
tag_greedy=g.get_tensor_by_name('import/tag_greedy:0'),
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
)
def deep_model(model: str = 'bahdanau', **kwargs):
"""
Load seq2seq stemmer deep learning model.
Returns
-------
DEEP_STEMMER: malaya.stemmer._DEEP_STEMMER class
"""
check_file(PATH_STEM[model], S3_PATH_STEM[model], **kwargs)
try:
with open(PATH_STEM[model]['setting'], 'r') as fopen:
dic_stemmer = json.load(fopen)
g = load_graph(PATH_STEM[model]['model'])
except:
raise Exception(
f"model corrupted due to some reasons, please run malaya.clear_cache('stem/{model}') and try again"
)
return _DEEP_STEMMER(
g.get_tensor_by_name('import/Placeholder:0'),
g.get_tensor_by_name('import/logits:0'),
generate_session(graph=g),
dic_stemmer,
)
def load(model, quantized=False, **kwargs):
path = check_file(
file=model,
module='gpt2',
keys={
'model': 'model.pb',
'encoder': GPT2_ENCODER,
'vocab': GPT2_VOCAB
},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
with open(path['encoder']) as f:
en = json.load(f)
with open(path['vocab'], encoding='utf-8') as f:
bpe_data = f.read()
bpe_merges = [
tuple(merge_str.split()) for merge_str in bpe_data.split('\n')[1:-1]
]
encoder = GPT2Encoder(
encoder=en,
bpe_merges=bpe_merges,
)
inputs = ['X', 'temp', 'top_k', 'top_p', 'maxlen', 'n_samples']
outputs = ['output']
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
return GPT2(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
encoder=encoder,
)
def deep_model(**kwargs):
"""
Load LSTM + Bahdanau Attention stemming model.
Returns
-------
DEEP_STEMMER: malaya.stem.DEEP_STEMMER class
"""
from malaya.preprocessing import _tokenizer
check_file(PATH_STEM['deep'], S3_PATH_STEM['deep'], **kwargs)
g = load_graph(PATH_STEM['deep']['model'])
bpe, subword_mode = load_yttm(PATH_STEM['deep']['bpe'], id_mode=True)
return DEEP_STEMMER(
g.get_tensor_by_name('import/Placeholder:0'),
g.get_tensor_by_name('import/decode_1/greedy:0'),
g.get_tensor_by_name('import/decode_2/beam:0'),
generate_session(graph=g),
bpe,
subword_mode,
_tokenizer,
)