# Copyright (c) ModelScope Contributors. All rights reserved.
import torch.nn
from collections import namedtuple
from functools import partial

from swift.loss import loss_map
from swift.metrics import eval_metrics_map
from swift.utils import get_logger
from .base import BaseMegatronTrainer

logger = get_logger()
ModelOutputs = namedtuple('ModelOutputs', ['logits'])


class MegatronRerankerTrainer(BaseMegatronTrainer):

    def __init__(self, args, template):
        super().__init__(args, template)
        self._loss_func = loss_map[args.loss_type](args, self)
        self.eval_metrics = eval_metrics_map['reranker'](args, self)

    @staticmethod
    def _get_listwise_reranker_preds(logits, labels):
        positive_indices = torch.nonzero(labels == 1, as_tuple=False).squeeze(-1).tolist()
        positive_indices.append(labels.shape[0])
        preds = []
        for i in range(len(positive_indices) - 1):
            start, end = positive_indices[i], positive_indices[i + 1]
            preds.append(logits[start:end].argmax())
        preds = torch.stack(preds)
        labels = torch.tensor([0] * (len(positive_indices) - 1), device=preds.device)
        return preds, labels

    def loss_func(self,
                  output_tensor: torch.Tensor,
                  *,
                  labels: torch.Tensor,
                  packed_seq_params=None,
                  attention_mask=None):
        training = self.unwrapped_models[0].training
        logits = self.get_last_tokens(output_tensor, packed_seq_params, attention_mask)
        loss = self._loss_func(ModelOutputs(logits=logits), labels)
        args = self.args
        logits_detach = logits.detach().squeeze(-1)
        if not training:
            self.eval_metrics.update(logits_detach, labels)
        if args.loss_type == 'listwise_reranker':
            preds, labels = self._get_listwise_reranker_preds(logits_detach, labels)
        else:
            preds = (logits_detach.detach() > 0).long()
        acc = (preds == labels).float().mean()
        metric = {'loss': loss.detach().clone(), 'acc': acc}
        metric = self._all_reduce_metric(metric)
        return loss, metric

    def prepare_model(self):
        super().prepare_model()
        for model in self.unwrapped_models:
            lm_model = model.language_model if hasattr(model, 'language_model') else model
            lm_model.tokenizer = self.template.tokenizer

    def forward_step(self, data_iterator, model):
        vp_stage = model.module.module.vp_stage
        data = self.get_batch(data_iterator, vp_stage)
        labels = data.pop('labels', None)
        output_tensor = model(**data)
        loss_func = partial(
            self.loss_func,
            labels=labels,
            packed_seq_params=data.get('packed_seq_params'),
            attention_mask=data.get('attention_mask')
            if data.get('attention_mask') is not None else data.get('attention_mask_2d'))
        return output_tensor, loss_func