src/data/collator.py

from dataclasses import dataclass
from typing import Any, Dict, List

import torch


@dataclass
class MyCollator:
    pad_token_id: int
    attention_pad_value: int = 0
    label_pad_value: int = -100

    def __call__(self, features: List[Dict[str, Any]]) -> Dict[str, Any]:
        # * extract data from features, and format them from dict to list
        input_ids = [f["input_ids"] for f in features]  # List[List[int]]
        placeholder_indices = [f["placeholder_indices"] for f in features]  # List[List[int]]
        super_input_ids = [f["super_input_ids"] for f in features]  # List[List[List[int]]]
        super_token_indices = [f["super_token_indices"] for f in features]  # List[List[List[int]]]
        labels = [f["labels"] for f in features] if "labels" in features[0] else None  # List[List[int]]

        # * process model input
        (
            input_ids,
            attention_mask,
            placeholder_indices,
            labels,
        ) = self.process_model_inputs(
            input_ids,
            placeholder_indices,
            labels,
        )

        # * process super_tokenizer input
        (
            super_input_ids,
            super_attention_mask,
            super_token_indices,
        ) = self.process_super_tokenizer_inputs(
            super_input_ids, 
            super_token_indices,
        )

        # * to torch tensor
        input_ids = torch.tensor(input_ids)
        attention_mask = torch.tensor(attention_mask)
        super_input_ids = torch.tensor(super_input_ids)
        super_attention_mask = torch.tensor(super_attention_mask)
        labels = torch.tensor(labels) if labels else None

        # * format 
        res = {
            "input_ids": input_ids,
            "attention_mask": attention_mask,
            "super_input_ids": super_input_ids,
            "super_attention_mask": super_attention_mask,
            "placeholder_indices": placeholder_indices,
            "super_token_indices": super_token_indices,
            "labels": labels,
        }

        return res
    
    def process_model_inputs(self, input_ids, placeholder_indices, labels):
        # * get attention mask
        max_len = get_max_length_in_nested_lists(input_ids)
        attention_mask = get_attention_mask_from_nested_lists(input_ids)

        # * get new placeholder_indices since padding side is left
        placeholder_indices = [
            [idx + max_len - len(input_ids[i]) for idx in placeholder_indices[i]]
            for i in range(len(placeholder_indices))
        ]

        # * pad
        input_ids = pad_nested_lists(input_ids, max_len, self.pad_token_id, "left")
        attention_mask = pad_nested_lists(attention_mask, max_len, self.attention_pad_value, "left")
        if labels:
            labels = pad_nested_lists(labels, max_len, self.label_pad_value, "left")

        return input_ids, attention_mask, placeholder_indices, labels

    def process_super_tokenizer_inputs(self, input_ids, super_token_indices):
        # * 3D -> 2D
        input_ids = sum(input_ids, [])  # List[List[int]]
        super_token_indices = sum(super_token_indices, [])  # List[List[int]]
        
        # * filter empty item
        new_input_ids = []
        new_super_token_indices = []
        for i in range(len(input_ids)):
            if len(super_token_indices[i]) != 0:
                new_input_ids.append(input_ids[i])
                new_super_token_indices.append(super_token_indices[i])
        input_ids = new_input_ids
        super_token_indices = new_super_token_indices
        
        if len(input_ids) == 0:
            return [], [], []
        
        # * get attention mask and pad
        max_len = get_max_length_in_nested_lists(input_ids)
        attention_mask = get_attention_mask_from_nested_lists(input_ids)
        
        input_ids = pad_nested_lists(input_ids, max_len, self.pad_token_id)
        attention_mask = pad_nested_lists(attention_mask, max_len, self.attention_pad_value)

        return input_ids, attention_mask, super_token_indices

def get_max_length_in_nested_lists(lst):
    if isinstance(lst[0], list):
        lengths = []
        for elem in lst:
            length = get_max_length_in_nested_lists(elem)
            lengths.append(length)
        max_length = max(lengths)
        return max_length
    else:
        return len(lst)


def get_attention_mask_from_nested_lists(lst):
    if isinstance(lst[0], list):
        attention_mask = []
        for elem in lst:
            mask = get_attention_mask_from_nested_lists(elem)
            attention_mask.append(mask)
        return attention_mask
    else:
        return [1] * len(lst)


def pad_nested_lists(lst, max_length, padding_value, padding_side="right"):
    if isinstance(lst, list) and len(lst) and isinstance(lst[0], list):
        for i, elem in enumerate(lst):
            lst[i] = pad_nested_lists(elem, max_length, padding_value, padding_side)
        return lst
    elif isinstance(lst, list):
        if padding_side == "right":
            return lst + [padding_value for _ in range(max_length - len(lst))]
        else:
            return [padding_value for _ in range(max_length - len(lst))] + lst
    else:
        raise NotImplementedError(f"Unrecognized type {lst}")