| from dataclasses import dataclass, field
|
| from typing import List, Optional, Any
|
| import os
|
|
|
|
|
|
|
| @dataclass
|
| class Arguments:
|
| train_data: str = field(
|
| default=None, metadata={"help": "Path to training data"}
|
| )
|
| val_data: str = field(
|
| default=None, metadata={"help": "Path to validation data"}
|
| )
|
| test_data: str = field(
|
| default=None, metadata={"help": "Path to test data"}
|
| )
|
| syntactic_file: str = field(
|
| default=None, metadata={"help": "Path to syntactic_file data"}
|
| )
|
|
|
| num_labels: int = field(default=2, metadata={"help": "Number of labels"})
|
|
|
| batch_size: int = field(default=8)
|
| val_batch_size: int = field(default=32)
|
|
|
|
|
| max_len: int = field(
|
| default=256,
|
| metadata={
|
| "help": "The maximum total input sequence length after tokenization for passage. Sequences longer "
|
| "than this will be truncated, sequences shorter will be padded."
|
| },
|
| )
|
|
|
| pad_to_multiple_of: int = field(default=2, metadata={"help": ""})
|
|
|
| temperature: Optional[float] = field(default=2.0)
|
| distill_temperature: Optional[float] = field(default=2.0)
|
|
|
| knowledge_distillation: bool = field(default=True, metadata={"help": "Use knowledge distillation"})
|
| finetune_hidden_states: bool = field(default=True)
|
| output_attentions: bool = field(default=True)
|
|
|
| teach_device: str = field(default='cuda:1')
|
| student_device: str = field(default='cuda:0')
|
|
|
| num_train_epochs: int = field(default=1)
|
|
|
| learning_rate: float = field(default=1e-4)
|
| weight_decay: float = field(default=0.01)
|
| warmup_ratio: float = field(default=0.1)
|
|
|
| geom_loss_weight: float = field(default=50)
|
| hard_label_loss_weight: float = field(default=1.0)
|
|
|
| teacher_layers_mapping: List[int] = field(default=list)
|
| student_encoder_layers_finetuned: List[int] = field(default=list)
|
| n_encoder_finetuned: int = field(default=6)
|
| finetune_embedding: bool = field(default=False)
|
|
|
| orthogonal: bool = field(default=True)
|
| span_loss: bool = field(default=True)
|
|
|
| span_weight_pooling: bool = field(default=True)
|
| span_loss_weight: bool = field(default=True)
|
|
|
| p: float = field(default=1.0)
|
|
|
| hidden_loss_weights: List[float] = field(default=None)
|
| teacher_embedding_dimension: int = field(default=1024)
|
|
|
|
|
| output_dir: Optional[str] = field(default=None, metadata={"help": "Where to store the final model"})
|
|
|
| teacher_model: str = field(default='')
|
| teacher_tokenizer: str = field(default='')
|
| student_model: str = field(default='google-bert/bert-base-uncased')
|
| student_tokenizer: str = field(default='google-bert/bert-base-uncased')
|
| hf_token: str = field(default='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')
|
|
|
| load_student_tokenizer_kwargs: dict = field(default_factory=dict)
|
| load_teacher_tokenizer_kwargs: dict = field(default_factory=dict)
|
|
|
| def __post_init__(self):
|
| if not os.path.exists(self.train_data):
|
| raise FileNotFoundError(f"cannot find file: {self.train_data}, please set a true path")
|
|
|
| if len(self.teacher_layers_mapping) != len(self.student_encoder_layers_finetuned):
|
| raise ValueError("teacher_layers_mapping and student_encoder_layers_finetuned should have the same length")
|
|
|
|
|
