def run(data_dir, export_dir, spec='bert_classifier', **kwargs):
"""Runs demo."""
# Chooses model specification that represents model.
spec = model_spec.get(spec)
# Gets training data and validation data.
train_data = TextClassifierDataLoader.from_csv(filename=os.path.join(
os.path.join(data_dir, 'train.tsv')),
text_column='sentence',
label_column='label',
model_spec=spec,
delimiter='\t',
is_training=True)
validation_data = TextClassifierDataLoader.from_csv(filename=os.path.join(
os.path.join(data_dir, 'dev.tsv')),
text_column='sentence',
label_column='label',
model_spec=spec,
delimiter='\t',
is_training=False)
# Fine-tunes the model.
model = text_classifier.create(train_data,
model_spec=spec,
validation_data=validation_data,
**kwargs)
# Gets evaluation results.
_, acc = model.evaluate(validation_data)
print('Eval accuracy: %f' % acc)
# Exports to TFLite format.
model.export(export_dir)
def _test_model_without_training(self, model_spec):
# Test without retraining.
model = text_classifier.create(self.train_data,
model_spec=model_spec,
do_train=False)
self._test_accuracy(model, threshold=0.0)
self._test_export_to_tflite(model, threshold=0.0)
def test_average_wordvec_model_create_v1_incompatible(self):
with self.assertRaisesRegex(ValueError, 'Incompatible versions'):
model_spec = ms.AverageWordVecModelSpec(seq_len=2)
all_data = text_dataloader.TextClassifierDataLoader.from_folder(
self.text_dir, model_spec=model_spec)
_ = text_classifier.create(
all_data,
model_spec=model_spec,
)
def test_bert_model(self):
model_spec = ms.BertClassifierModelSpec(seq_len=2, trainable=False)
all_data = text_dataloader.TextClassifierDataLoader.from_folder(
self.text_dir, model_spec=model_spec)
# Splits data, 90% data for training, 10% for testing
self.train_data, self.test_data = all_data.split(0.9)
model = text_classifier.create(self.train_data,
model_spec=model_spec,
epochs=1,
batch_size=1,
shuffle=True)
self._test_accuracy(model, 0.5)
def test_mobilebert_model(self):
model_spec = ms.mobilebert_classifier_spec(seq_len=2, trainable=False)
all_data = text_dataloader.TextClassifierDataLoader.from_folder(
self.tiny_text_dir, model_spec=model_spec)
# Splits data, 50% data for training, 50% for testing
self.train_data, self.test_data = all_data.split(0.5)
model = text_classifier.create(self.train_data,
model_spec=model_spec,
epochs=1,
batch_size=1,
shuffle=True)
self._test_accuracy(model, 0.0)
self._test_export_to_tflite(model, threshold=0.0, atol=1e-2)
self._test_export_to_tflite_quant(model)
def test_average_wordvec_model(self):
model_spec = ms.AverageWordVecModelSpec(seq_len=2)
all_data = text_dataloader.TextClassifierDataLoader.from_folder(
self.text_dir, model_spec=model_spec)
# Splits data, 90% data for training, 10% for testing
self.train_data, self.test_data = all_data.split(0.9)
model = text_classifier.create(self.train_data,
mef.ModelExportFormat.TFLITE,
model_spec=model_spec,
epochs=2,
batch_size=4,
shuffle=True)
self._test_accuracy(model)
self._test_export_to_tflite(model)
self._test_predict_top_k(model)
def test_mobilebert_model(self):
model_spec = ms.mobilebert_classifier_spec
model_spec.seq_len = 2
model_spec.trainable = False
all_data = text_dataloader.TextClassifierDataLoader.from_folder(
self.text_dir, model_spec=model_spec)
# Splits data, 90% data for training, 10% for testing
self.train_data, self.test_data = all_data.split(0.9)
model = text_classifier.create(self.train_data,
model_spec=model_spec,
epochs=1,
batch_size=1,
shuffle=True)
self._test_accuracy(model, 0.5)
self._test_export_to_tflite(model, test_predict_accuracy=False)
self._test_export_to_tflite_quant(model)
def test_bert_model(self):
model_spec = ms.BertClassifierModelSpec(seq_len=2, trainable=False)
all_data = text_dataloader.TextClassifierDataLoader.from_folder(
self.tiny_text_dir, model_spec=model_spec)
# Splits data, 50% data for training, 50% for testing
self.train_data, self.test_data = all_data.split(0.5)
model = text_classifier.create(self.train_data,
model_spec=model_spec,
epochs=1,
batch_size=1,
shuffle=True)
self._test_accuracy(model, 0.0)
self._test_export_to_tflite(
model,
threshold=0.0,
expected_json_file='bert_classifier_metadata.json')
self._test_model_without_training(model_spec)
def test_average_wordvec_model(self):
model_spec = ms.AverageWordVecModelSpec(seq_len=2)
all_data = text_dataloader.TextClassifierDataLoader.from_folder(
self.text_dir, model_spec=model_spec)
# Splits data, 90% data for training, 10% for testing
self.train_data, self.test_data = all_data.split(0.5)
model = text_classifier.create(self.train_data,
model_spec=model_spec,
epochs=1,
batch_size=1,
shuffle=True)
self._test_accuracy(model, threshold=0.0)
self._test_predict_top_k(model)
self._test_export_to_tflite(model, threshold=0.0)
self._test_export_to_saved_model(model)
self._test_export_labels(model)
self._test_export_vocab(model)
self._test_model_without_training(model_spec)
def main(_):
logging.set_verbosity(logging.INFO)
data_path = tf.keras.utils.get_file(
fname='sst.tar.gz',
origin=
'https://firebasestorage.googleapis.com/v0/b/mtl-sentence-representations.appspot.com/o/data%2FSST-2.zip?alt=media&token=aabc5f6b-e466-44a2-b9b4-cf6337f84ac8',
extract=True)
data_path = os.path.join(os.path.dirname(data_path), 'SST-2')
# Chooses model specification that represents model.
model_spec = BertModelSpec()
# Gets training data and validation data.
train_data = TextClassifierDataLoader.from_csv(filename=os.path.join(
os.path.join(data_path, 'train.tsv')),
text_column='sentence',
label_column='label',
model_spec=model_spec,
delimiter='\t')
validation_data = TextClassifierDataLoader.from_csv(filename=os.path.join(
os.path.join(data_path, 'dev.tsv')),
text_column='sentence',
label_column='label',
model_spec=model_spec,
delimiter='\t')
# Fine-tunes the model.
model = text_classifier.create(train_data,
model_spec=model_spec,
validation_data=validation_data)
# Gets evaluation results.
_, acc = model.evaluate(validation_data)
print('Eval accuracy: %f' % acc)
# Exports to TFLite format.
model.export(FLAGS.tflite_filename, FLAGS.label_filename,
FLAGS.vocab_filename)
def test_mobilebert_model(self):
model_spec = ms.mobilebert_classifier_spec
model_spec.seq_len = 2
model_spec.trainable = False
all_data = text_dataloader.TextClassifierDataLoader.from_folder(
self.text_dir, model_spec=model_spec)
# Splits data, 90% data for training, 10% for testing
self.train_data, self.test_data = all_data.split(0.9)
model = text_classifier.create(self.train_data,
model_spec=model_spec,
epochs=1,
batch_size=1,
shuffle=True)
self._test_accuracy(model, 0.5)
error_message = 'Couldn\'t convert MobileBert to TFLite for now.'
with self.assertRaises(ValueError) as error:
self._test_export_to_tflite(model, test_predict_accuracy=False)
self.assertEqual(error_message, str(error.exception))
with self.assertRaises(ValueError) as error:
self._test_export_to_tflite_quant(model)
self.assertEqual(error_message, str(error.exception))
