from abc import ABC

import torch
import torch.nn.functional as F

from matcha.models.components.decoder import Decoder
from matcha.utils.pylogger import get_pylogger

log = get_pylogger(__name__)


class BASECFM(torch.nn.Module, ABC):
    def __init__(
        self,
        n_feats,
        cfm_params,
        n_spks=1,
        spk_emb_dim=128,
    ):
        super().__init__()
        self.n_feats = n_feats
        self.n_spks = n_spks
        self.spk_emb_dim = spk_emb_dim
        self.solver = cfm_params.solver
        if hasattr(cfm_params, "sigma_min"):
            self.sigma_min = cfm_params.sigma_min
        else:
            self.sigma_min = 1e-4

        self.estimator = None

    @torch.inference_mode()
    def forward(self, mu, mask, n_timesteps, temperature=1.0, spks=None, cond=None):
        """Forward diffusion

        Args:
            z (_type_): mu + noise (we don't need this in this formulation), we will sample the noise again
            mask (_type_): output_mask
            mu (_type_): output of encoder
            n_timesteps (_type_): number of diffusion steps
            stoc (bool, optional): _description_. Defaults to False.
            spks (_type_, optional): _description_. Defaults to None.

        Returns:
            sample: _description_
        """
        z = torch.randn_like(mu) * temperature
        t_span = torch.linspace(0, 1, n_timesteps + 1, device=mu.device)
        return self.solve_euler(z, t_span=t_span, mu=mu, mask=mask, spks=spks, cond=cond)

    def solve_euler(self, x, t_span, mu, mask, spks, cond):
        """
        Fixed euler solver for ODEs.
        Args:
            x (_type_): _description_
            t (_type_): _description_
        """
        t, _, dt = t_span[0], t_span[-1], t_span[1] - t_span[0]
        sol = []

        steps = 1
        while steps <= len(t_span) - 1:
            dphi_dt = self.estimator(x, mask, mu, t, spks, cond)

            x = x + dt * dphi_dt
            t = t + dt
            sol.append(x)
            if steps < len(t_span) - 1:
                dt = t_span[steps + 1] - t
            steps += 1

        return sol[-1]

    def compute_loss(self, x1, mask, mu, spks=None, cond=None):
        """Computes diffusion loss

        Args:
            x1 (_type_): Target
            mask (_type_): target mask
            mu (_type_): output of encoder
            spks (_type_, optional): speaker embedding. Defaults to None.

        Returns:
            loss: diffusion loss
            y: conditional flow
        """
        b, _, t = mu.shape

        # random timestep
        t = torch.rand([b, 1, 1], device=mu.device, dtype=mu.dtype)
        # sample noise p(x_0)
        z = torch.randn_like(x1)

        y = (1 - (1 - self.sigma_min) * t) * z + t * x1
        u = x1 - (1 - self.sigma_min) * z

        loss = F.mse_loss(self.estimator(y, mask, mu, t.squeeze(), spks), u, reduction="sum") / (
            torch.sum(mask) * u.shape[1]
        )
        return loss, y


class CFM(BASECFM):
    def __init__(self, in_channels, out_channel, cfm_params, decoder_params, n_spks=1, spk_emb_dim=64):
        super().__init__(
            n_feats=in_channels,
            cfm_params=cfm_params,
            n_spks=n_spks,
            spk_emb_dim=spk_emb_dim,
        )

        in_channels = in_channels + (spk_emb_dim if n_spks > 1 else 0)
        # Just change the architecture of the estimator here
        self.estimator = Decoder(in_channels=in_channels, out_channels=out_channel, **decoder_params)