modeling_stacked.py

import torch
import numpy as np
import random
import os

# 1. Set random seeds
seed = 2025
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
os.environ["PYTHONHASHSEED"] = str(seed)

# 2. Disable dropout & training randomness
torch.use_deterministic_algorithms(True, warn_only=True)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False

from transformers.modeling_outputs import TokenClassifierOutput
import torch
import torch.nn as nn
from transformers import PreTrainedModel, AutoModel, AutoConfig, BertConfig
from torch.nn import CrossEntropyLoss
from typing import Optional, Tuple, Union
import logging, json, os

from .configuration_stacked import ImpressoConfig

logger = logging.getLogger(__name__)


def get_info(label_map):
    num_token_labels_dict = {task: len(labels) for task, labels in label_map.items()}
    return num_token_labels_dict


class ExtendedMultitaskTimeModelForTokenClassification(PreTrainedModel):
    config_class = ImpressoConfig
    _keys_to_ignore_on_load_missing = [r"position_ids"]

    def __init__(self, config, temporal_fusion_strategy="baseline", num_years=327):
        super().__init__(config)
        self.num_token_labels_dict = get_info(config.label_map)
        self.config = config
        self.temporal_fusion_strategy = temporal_fusion_strategy
        self.model = AutoModel.from_pretrained(
            config.pretrained_config["_name_or_path"], config=config.pretrained_config
        )
        self.model.config.use_cache = False
        self.model.config.pretraining_tp = 1
        self.num_years = num_years

        classifier_dropout = getattr(config, "classifier_dropout", 0.1) or config.hidden_dropout_prob
        self.dropout = nn.Dropout(classifier_dropout)

        self.temporal_fusion = TemporalFusion(config.hidden_size, strategy=self.temporal_fusion_strategy,
                                              num_years=num_years)

        # Additional transformer layers
        self.transformer_encoder = nn.TransformerEncoder(
            nn.TransformerEncoderLayer(
                d_model=config.hidden_size, nhead=config.num_attention_heads
            ),
            num_layers=2,
        )
        self.token_classifiers = nn.ModuleDict({
            task: nn.Linear(config.hidden_size, num_labels)
            for task, num_labels in self.num_token_labels_dict.items()
        })

        self.post_init()

    def forward(
            self,
            input_ids: Optional[torch.Tensor] = None,
            attention_mask: Optional[torch.Tensor] = None,
            token_type_ids: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.Tensor] = None,
            head_mask: Optional[torch.Tensor] = None,
            labels: Optional[torch.Tensor] = None,
            inputs_embeds: Optional[torch.Tensor] = None,
            token_labels: Optional[dict] = None,
            date_indices: Optional[torch.Tensor] = None,
            year_index: Optional[torch.Tensor] = None,
            decade_index: Optional[torch.Tensor] = None,
            century_index: Optional[torch.Tensor] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
    ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:

        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if inputs_embeds is None:
            inputs_embeds = self.model.embeddings(input_ids)

        # Early cross-attention fusion
        if self.temporal_fusion_strategy == "early-cross-attention":
            year_emb = self.temporal_fusion.compute_time_embedding(year_index)  # (B, H)
            inputs_embeds = self.temporal_fusion.cross_attn(inputs_embeds, year_emb)

        bert_kwargs = {
            "inputs_embeds": inputs_embeds if self.temporal_fusion_strategy == "early-cross-attention" else None,
            "input_ids": input_ids if self.temporal_fusion_strategy != "early-cross-attention" else None,
            "attention_mask": attention_mask,
            "token_type_ids": token_type_ids,
            "position_ids": position_ids,
            "head_mask": head_mask,
            "output_attentions": output_attentions,
            "output_hidden_states": output_hidden_states,
            "return_dict": return_dict,
        }

        if any(keyword in self.config.name_or_path.lower() for keyword in ["llama", "deberta"]):
            bert_kwargs.pop("token_type_ids", None)
            bert_kwargs.pop("head_mask", None)

        outputs = self.model(**bert_kwargs)
        token_output = self.dropout(outputs[0])  # (B, T, H)
        hidden_states = list(outputs.hidden_states) if output_hidden_states else None

        # Pass through additional transformer layers
        token_output = self.transformer_encoder(token_output.transpose(0, 1)).transpose(
            0, 1
        )
        # Apply fusion after transformer if needed
        if self.temporal_fusion_strategy not in ["baseline", "early-cross-attention"]:
            token_output = self.temporal_fusion(token_output, year_index)
            if output_hidden_states:
                hidden_states.append(token_output)  # add the final fused state

        task_logits = {}
        total_loss = 0
        for task, classifier in self.token_classifiers.items():
            logits = classifier(token_output)
            task_logits[task] = logits
            if token_labels and task in token_labels:
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(
                    logits.view(-1, self.num_token_labels_dict[task]),
                    token_labels[task].view(-1),
                )
                total_loss += loss

        if not return_dict:
            output = (task_logits,) + outputs[2:]
            return ((total_loss,) + output) if total_loss != 0 else output

        return TokenClassifierOutput(
            loss=total_loss,
            logits=task_logits,
            hidden_states=tuple(hidden_states) if hidden_states is not None else None,
            attentions=outputs.attentions if output_attentions else None,
        )


class TemporalFusion(nn.Module):
    def __init__(self, hidden_size, strategy="add", num_years=327, min_year=1700):
        super().__init__()
        self.strategy = strategy
        self.hidden_size = hidden_size
        self.min_year = min_year
        self.max_year = min_year + num_years - 1

        self.year_emb = nn.Embedding(num_years, hidden_size)

        if strategy == "concat":
            self.concat_proj = nn.Linear(hidden_size * 2, hidden_size)
        elif strategy == "film":
            self.film_gamma = nn.Linear(hidden_size, hidden_size)
            self.film_beta = nn.Linear(hidden_size, hidden_size)
        elif strategy == "adapter":
            self.adapter = nn.Sequential(
                nn.Linear(hidden_size, hidden_size),
                nn.ReLU(),
                nn.Linear(hidden_size, hidden_size),
            )
        elif strategy == "relative":
            self.relative_encoder = nn.Sequential(
                nn.Linear(hidden_size, hidden_size),
                nn.SiLU(),
                nn.LayerNorm(hidden_size),
            )
            self.film_gamma = nn.Linear(hidden_size, hidden_size)
            self.film_beta = nn.Linear(hidden_size, hidden_size)
        elif strategy == "multiscale":
            self.decade_emb = nn.Embedding(1000, hidden_size)
            self.century_emb = nn.Embedding(100, hidden_size)
        elif strategy in ["early-cross-attention", "late-cross-attention"]:
            self.year_encoder = nn.Sequential(
                nn.Linear(hidden_size, hidden_size),
                nn.SiLU()
            )
            self.cross_attn = TemporalCrossAttention(hidden_size)

    def compute_time_embedding(self, year_index):
        if self.strategy in ["early-cross-attention", "late-cross-attention"]:
            return self.year_encoder(self.year_emb(year_index))
        elif self.strategy == "multiscale":
            year_index = year_index.long()
            year = year_index + self.min_year
            decade = (year // 10).long()
            century = (year // 100).long()
            return (
                    self.year_emb(year_index) +
                    self.decade_emb(decade) +
                    self.century_emb(century)
            )
        else:
            return self.year_emb(year_index)

    def forward(self, token_output, year_index):
        B, T, H = token_output.size()

        if self.strategy == "baseline":
            return token_output

        year_emb = self.compute_time_embedding(year_index)

        if self.strategy == "concat":
            expanded_year = year_emb.unsqueeze(1).repeat(1, T, 1)
            fused = torch.cat([token_output, expanded_year], dim=-1)
            return self.concat_proj(fused)

        elif self.strategy == "film":
            gamma = self.film_gamma(year_emb).unsqueeze(1)
            beta = self.film_beta(year_emb).unsqueeze(1)
            return gamma * token_output + beta

        elif self.strategy == "adapter":
            return token_output + self.adapter(year_emb).unsqueeze(1)

        elif self.strategy == "add":
            expanded_year = year_emb.unsqueeze(1).repeat(1, T, 1)
            return token_output + expanded_year

        elif self.strategy == "relative":
            encoded = self.relative_encoder(year_emb)
            gamma = self.film_gamma(encoded).unsqueeze(1)
            beta = self.film_beta(encoded).unsqueeze(1)
            return gamma * token_output + beta

        elif self.strategy == "multiscale":
            expanded_year = year_emb.unsqueeze(1).expand(-1, T, -1)
            return token_output + expanded_year

        elif self.strategy == "late-cross-attention":
            return self.cross_attn(token_output, year_emb)

        else:
            raise ValueError(f"Unknown fusion strategy: {self.strategy}")


class TemporalCrossAttention(nn.Module):
    def __init__(self, hidden_size, num_heads=4):
        super().__init__()
        self.attn = nn.MultiheadAttention(embed_dim=hidden_size, num_heads=num_heads, batch_first=True)

    def forward(self, token_output, time_embedding):
        # token_output: (B, T, H), time_embedding: (B, H)
        time_as_seq = time_embedding.unsqueeze(1)  # (B, 1, H)
        attn_output, _ = self.attn(token_output, time_as_seq, time_as_seq)
        return token_output + attn_output