def test_from_pretrained_with_tuple_values(self):
# For the auto model mapping, FunnelConfig has two models: FunnelModel and FunnelBaseModel
model = TFAutoModel.from_pretrained("sgugger/funnel-random-tiny")
self.assertIsInstance(model, TFFunnelModel)
config = copy.deepcopy(model.config)
config.architectures = ["FunnelBaseModel"]
model = TFAutoModel.from_config(config)
self.assertIsInstance(model, TFFunnelBaseModel)
with tempfile.TemporaryDirectory() as tmp_dir:
model.save_pretrained(tmp_dir)
model = TFAutoModel.from_pretrained(tmp_dir)
self.assertIsInstance(model, TFFunnelBaseModel)
def __init__(self, tokenizer):
super().__init__()
self.tokenizer = tokenizer
config = AutoConfig.from_pretrained(TINY_MODEL_CHECKPOINT)
self.bert = TFAutoModel.from_config(config)
def instantiate_layer(
self,
load_pretrained_weights=True
) -> Union[object, TFAutoModel, TFBertModel]:
"""
Instanciate a transformer to be loaded in a Keras layer using the availability method of the pre-trained transformer.
"""
if self.loading_method == LOADING_METHOD_HUGGINGFACE_NAME:
if load_pretrained_weights:
transformer_model = TFAutoModel.from_pretrained(self.name,
from_pt=True)
self.transformer_config = transformer_model.config
return transformer_model
else:
config_path = os.path.join(".", self.local_dir_path,
TRANSFORMER_CONFIG_FILE_NAME)
self.transformer_config = AutoConfig.from_pretrained(
config_path)
return TFAutoModel.from_config(self.transformer_config)
elif self.loading_method == LOADING_METHOD_LOCAL_MODEL_DIR:
if load_pretrained_weights:
transformer_model = TFAutoModel.from_pretrained(
self.local_dir_path, from_pt=True)
self.transformer_config = transformer_model.config
return transformer_model
else:
config_path = os.path.join(".", self.local_dir_path,
TRANSFORMER_CONFIG_FILE_NAME)
self.transformer_config = AutoConfig.from_pretrained(
config_path)
#self.transformer_config = AutoConfig.from_pretrained(self.local_dir_path)
return TFAutoModel.from_config(self.transformer_config)
elif self.loading_method == LOADING_METHOD_PLAIN_MODEL:
if load_pretrained_weights:
self.transformer_config = AutoConfig.from_pretrained(
self.local_config_file)
# transformer_model = TFBertModel.from_pretrained(self.local_weight_file, from_tf=True)
raise NotImplementedError(
"The load of TF weights from huggingface automodel classes is not yet implemented. \
Please use load from Hugging Face Hub or from directory for the initial loading of the transformers weights."
)
else:
config_path = os.path.join(".", self.local_dir_path,
TRANSFORMER_CONFIG_FILE_NAME)
self.transformer_config = AutoConfig.from_pretrained(
config_path)
return TFBertModel.from_config(self.transformer_config)
else:
# TODO: revise this
if load_pretrained_weights:
transformer_model = TFAutoModel.from_pretrained(
self.local_dir_path, from_pt=True)
self.transformer_config = transformer_model.config
return transformer_model
else:
config_path = os.path.join(".", self.local_dir_path,
TRANSFORMER_CONFIG_FILE_NAME)
self.transformer_config = AutoConfig.from_pretrained(
config_path)
return TFAutoModel.from_config(self.transformer_config)