Spaces:
Sleeping
Sleeping
| import os | |
| from pathlib import Path | |
| from src.utils.common import read_yaml | |
| from dotenv import find_dotenv, load_dotenv | |
| from src.models.entity import (ModelTrainerConfig, | |
| LoraParameters, | |
| BitsAndBytesParameters, | |
| TrainingArgumentsParameters, | |
| ModelPredictionConfig, | |
| ModelPredictionParameters) | |
| _ = load_dotenv(find_dotenv()) # read local .env file | |
| MODEL_CONFIG_FILE_PATH = os.environ['MODEL_CONFIG_FILE_PATH'] | |
| MODEL_PARAMS_FILE_PATH = os.environ['MODEL_PARAMS_FILE_PATH'] | |
| class ConfigurationManager: | |
| def __init__(self, | |
| model_config_filepath = MODEL_CONFIG_FILE_PATH, | |
| model_params_filepath = MODEL_PARAMS_FILE_PATH): | |
| self.config = read_yaml(Path(model_config_filepath)) | |
| self.params = read_yaml(Path(model_params_filepath)) | |
| def get_model_trainer_config(self) -> ModelTrainerConfig: | |
| config = self.config.model_trainer | |
| model_trainer_config = ModelTrainerConfig( | |
| root_dir = Path(config.root_dir), | |
| data_path = Path(config.data_path), | |
| base_model = config.base_model, | |
| training_name = config.training_name, | |
| upload_from_hf = config.upload_from_hf, | |
| hf_model_name = config.hf_model_name | |
| ) | |
| return model_trainer_config | |
| def get_lora_params(self) -> LoraParameters: | |
| params = self.params.lora_parameters | |
| lora_parameters = LoraParameters( | |
| r = params.r, | |
| target_modules = params.target_modules, | |
| lora_alpha = params.lora_alpha, | |
| lora_dropout = params.lora_dropout, | |
| bias = params.bias, | |
| task_type = params.task_type | |
| ) | |
| return lora_parameters | |
| def get_bits_and_bytes_params(self) -> BitsAndBytesParameters: | |
| params = self.params.bits_and_bytes_parameters | |
| bits_and_bytes_parameters = BitsAndBytesParameters( | |
| load_in_4bit = params.load_in_4bit, | |
| bnb_4bit_quant_type = params.bnb_4bit_quant_type, | |
| bnb_4bit_use_double_quant = params.bnb_4bit_use_double_quant | |
| ) | |
| return bits_and_bytes_parameters | |
| def get_training_args(self) -> TrainingArgumentsParameters: | |
| params = self.params.training_arguments | |
| training_args = TrainingArgumentsParameters( | |
| output_dir = params.output_dir, | |
| evaluation_strategy = params.evaluation_strategy, | |
| save_strategy = params.save_strategy, | |
| num_train_epochs = params.num_train_epochs, | |
| per_device_train_batch_size = params.per_device_train_batch_size, | |
| gradient_accumulation_steps = params.gradient_accumulation_steps, | |
| optim = params.optim, | |
| learning_rate = params.learning_rate, | |
| fp16 = params.fp16, | |
| max_grad_norm = params.max_grad_norm, | |
| warmup_ratio = params.warmup_ratio, | |
| group_by_length = params.group_by_length, | |
| lr_scheduler_type = params.lr_scheduler_type | |
| ) | |
| return training_args | |
| def get_model_prediction_config(self) -> ModelPredictionConfig: | |
| config = self.config.model_prediction | |
| model_prediction_config = ModelPredictionConfig( | |
| data_path=config.data_path, | |
| base_model = config.base_model, | |
| adapters_path = config.adapters_path | |
| ) | |
| return model_prediction_config | |
| def get_model_prediction_parameters(self) -> ModelPredictionParameters: | |
| config = self.params.prediction_parameters | |
| model_prediction_parameters = ModelPredictionParameters( | |
| length_penalty=config.length_penalty, | |
| num_beams = config.num_beams, | |
| max_length = config.max_length | |
| ) | |
| return model_prediction_parameters |