print(f"Labels mapping: {LABELS_TO_ID}")
print()
metrics = {"validation": [AccuracyMeter]}
model_config = config.SenseModelParameters(model_name=args.config_name,
hidden_size=args.hidden_size,
num_classes=len(LABELS_TO_ID),
freeze_weights=False,
context_layers=(-1, ))
configuration = config.Configuration(
model_parameters=model_config,
model=args.model,
save_path=args.save_path,
sequence_max_len=args.seq_len,
batch_size=args.batch_size,
epochs=args.epochs,
device=torch.device(args.device),
tokenizer=transformers.AutoTokenizer.from_pretrained(args.model),
)
"""
model = TransformerWrapper.load_pretrained(
args.pretrained_model,
params=configuration,
pooler = BertPoolingStrategy(configuration),
loss = SoftmaxLoss(configuration))
"""
model = DistilBertForSequenceClassification.from_pretrained(
args.pretrained_model, num_labels=len(LABELS_TO_ID))
train_split, valid_split = dataset.split_dataset(test_perc=0.1)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--name', type=str, dest="config_name")
parser.add_argument(
'--model',
type=str,
dest="model",
default="sentence-transformers/quora-distilbert-multilingual")
parser.add_argument('--save_path',
dest="save_path",
type=str,
default="./output")
args = parser.parse_args()
tokenizer = AutoTokenizer.from_pretrained(args.model)
model_config = config.ModelParameters(model_name=args.config_name)
configuration = config.Configuration(model_parameters=model_config,
model=args.model,
save_path=args.save_path,
tokenizer=tokenizer)
sentence_model = OnnxSentenceTransformerWrapper.load_pretrained(
path=args.model, params=configuration)
convert_to_onnx(sentence_model, configuration, quantize=True)
model_config = config.SenseModelParameters(
model_name=args.config_name,
hidden_size=args.hidden_size,
num_classes=2,
use_pretrained_embeddings=args.use_pretrained_embeddings,
freeze_weights=args.freeze_weights,
context_layers=(-1, -2, -3, -4))
configuration = config.Configuration(
model_parameters=model_config,
model=args.model,
save_path=args.save_path,
sequence_max_len=args.seq_len,
dropout_prob=args.dropout,
lr=args.lr,
batch_size=args.batch_size,
epochs=args.epochs,
device=torch.device(args.device),
embedding_map=config.CONFIG.embedding_map,
bnids_map=config.CONFIG.bnids_map,
tokenizer=transformers.AutoTokenizer.from_pretrained(args.model),
pretrained_embeddings_dim=config.DIMENSIONS_MAP[
args.sense_embeddings_type],
senses_as_features=args.senses_as_features)
model = WordEncoderModel(
params=configuration,
loss=SoftmaxLoss,
pooling_strategy=POOLING_STRATEGIES[args.pooling_strategy],
)
num_train_steps = len(train_data_loader) * args.epochs
tokenizer = DistilBertTokenizer.from_pretrained(args.pretrained_path)
model_config = config.ModelParameters(
model_name = "args.config_name",
hidden_size = 768,
num_classes = 10,
use_pretrained_embeddings = False,
freeze_weights = False,
context_layers = (-1,)
)
configuration = config.Configuration(
model_parameters=model_config,
model = "",
save_path = "args.save_path",
sequence_max_len = 128,
dropout_prob = 0.1,
lr = 2e-5,
batch_size = 16,
epochs = 1,
device = torch.device("cpu"),
tokenizer = tokenizer,
)
model = transformers.DistilBertForSequenceClassification.from_pretrained(args.pretrained_path)
model.eval()
input = "福岡市は公立小中学校などの教員採用で、筆記試験と面接を省く新たな採用方式を2022年から導入する"
convert_to_torchscript(model, tokenizer, input, os.path.join(args.save_path, args.config_name))
train_split, valid_split = dataset.split_dataset(test_perc=0.1)
#train_dataset = Dataset(train_split)
valid_dataset = Dataset(valid_split)
LABELS_TO_ID = dataset.label_to_id
model_config = config.SenseModelParameters(model_name=args.config_name,
hidden_size=args.embed_dim,
num_classes=len(LABELS_TO_ID),
freeze_weights=False,
context_layers=(-1, ))
configuration = config.Configuration(
model_parameters=model_config,
model=args.model,
save_path=args.output_dir,
sequence_max_len=args.seq_len,
batch_size=args.batch_size,
epochs=args.epochs,
device=torch.device(args.device),
tokenizer=tokenizer,
)
valid_data_loader = SmartParaphraseDataloader.build_batches(
valid_dataset, 16, mode="sequence", config=configuration)
autoconfig = AutoConfig.from_pretrained(
args.pretrained_model_path,
output_attentions=True,
)
autoconfig.num_labels = len(LABELS_TO_ID)
model = AutoModelForSequenceClassification.from_pretrained(
args.pretrained_model_path, config=autoconfig)
"""
type=float,
dest="corpus_percentage",
default=0.01)
parser.add_argument('--nq', type=int, dest="num_queries", default=10)
parser.add_argument('--topk', type=int, dest="topk", default=3)
parser.add_argument('--sbert', type=bool, dest="use_sbert", default=False)
args = parser.parse_args()
random.seed(43)
model_config = config.ModelParameters(model_name="eval_sentence_mining")
configuration_teacher = config.Configuration(
model_parameters=model_config,
model=args.teacher_model,
save_path="./results",
batch_size=16,
device=torch.device(args.device),
tokenizer=transformers.AutoTokenizer.from_pretrained(
args.teacher_model, use_fast=True))
configuration_student = config.Configuration(
model_parameters=model_config,
model=args.student_model,
save_path="./results",
batch_size=16,
device=torch.device(args.device),
tokenizer=transformers.AutoTokenizer.from_pretrained(
args.teacher_model, use_fast=True))
teacher_model = SentenceTransformerWrapper.load_pretrained(
args.teacher_model, params=configuration_teacher)
