import torch
import torch.nn as nn
import torch.nn.functional as F
from layers.Transformer_EncDec import Encoder, EncoderLayer
from layers.SelfAttention_Family import ReformerLayer
from layers.Embed import DataEmbedding


class Model(nn.Module):
    """
    Reformer with O(LlogL) complexity
    Paper link: https://openreview.net/forum?id=rkgNKkHtvB
    """

    def __init__(self, configs, bucket_size=4, n_hashes=4):
        """
        bucket_size: int, 
        n_hashes: int, 
        """
        super(Model, self).__init__()
        self.pred_len = configs.pred_len
        self.seq_len = configs.seq_len

        if configs.channel_independence:
            self.enc_in = 1
            self.dec_in = 1
            self.c_out = 1
        else:
            self.enc_in = configs.enc_in
            self.dec_in = configs.dec_in
            self.c_out = configs.c_out

        self.enc_embedding = DataEmbedding(self.enc_in, configs.d_model, configs.embed, configs.freq,
                                           configs.dropout)
        # Encoder
        self.encoder = Encoder(
            [
                EncoderLayer(
                    ReformerLayer(None, configs.d_model, configs.n_heads,
                                  bucket_size=bucket_size, n_hashes=n_hashes),
                    configs.d_model,
                    configs.d_ff,
                    dropout=configs.dropout,
                    activation=configs.activation
                ) for l in range(configs.e_layers)
            ],
            norm_layer=torch.nn.LayerNorm(configs.d_model)
        )

        self.projection = nn.Linear(
            configs.d_model, configs.c_out, bias=True)

    def long_forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec):
        # add placeholder
        x_enc = torch.cat([x_enc, x_dec[:, -self.pred_len:, :]], dim=1)
        if x_mark_enc is not None:
            x_mark_enc = torch.cat(
                [x_mark_enc, x_mark_dec[:, -self.pred_len:, :]], dim=1)

        enc_out = self.enc_embedding(x_enc, x_mark_enc)  # [B,T,C]
        enc_out, attns = self.encoder(enc_out, attn_mask=None)
        dec_out = self.projection(enc_out)

        return dec_out  # [B, L, D]


    def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):
        dec_out = self.long_forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)
        return dec_out[:, -self.pred_len:, :]  # [B, L, D]