import torch
import torch.nn as nn
from datasets import load_dataset, load_from_disk, concatenate_datasets
from transformers import AutoTokenizer, TrainingArguments, Trainer, DataCollatorWithPadding, XLMRobertaPreTrainedModel, XLMRobertaModel
from transformers.modeling_outputs import SequenceClassifierOutput
import evaluate
import numpy as np



class HierarchicalXLMRoberta(XLMRobertaPreTrainedModel):
    def __init__(self, config, num_labels_level1, num_labels_level2):
        super().__init__(config)
        self.num_labels_level1 = num_labels_level1
        self.num_labels_level2 = num_labels_level2
        self.roberta = XLMRobertaModel(config)
        self.classifier_level1 = nn.Linear(config.hidden_size, num_labels_level1)
        # Level2 classifier takes concatenated pooled_output + level1_logits
        self.classifier_level2 = nn.Linear(config.hidden_size + num_labels_level1, num_labels_level2)
        self.init_weights()

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        level1_encoded=None,
        level2_encoded=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.roberta(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        pooled_output = outputs.pooler_output

        logits_level1 = self.classifier_level1(pooled_output)
        
        # Concatenate pooled_output with level1 logits for level2
        combined_for_level2 = torch.cat([pooled_output, logits_level1], dim=-1)
        logits_level2 = self.classifier_level2(combined_for_level2)

        loss = None
        if level1_encoded is not None and level2_encoded is not None:
            loss_fct = nn.CrossEntropyLoss()
            loss_level1 = loss_fct(logits_level1.view(-1, self.num_labels_level1), level1_encoded.view(-1))
            loss_level2 = loss_fct(logits_level2.view(-1, self.num_labels_level2), level2_encoded.view(-1))
            loss = loss_level1 + loss_level2  # Joint loss (you can weight them if needed, e.g., 0.5 * each)

        if not return_dict:
            output = (logits_level1, logits_level2) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return SequenceClassifierOutput(
            loss=loss,
            logits=(logits_level1, logits_level2),
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )