import math
from math import sqrt

import torch
import torch.nn as nn
import torch.nn.functional as F

from modules.commons.common_layers import SinusoidalPosEmb, AdamWConv1d
from modules.commons.common_layers import KaimingNormalConv1d as Conv1d
from utils.hparams import hparams


class ResidualBlock(nn.Module):
    def __init__(self, encoder_hidden, residual_channels, dilation):
        super().__init__()
        self.residual_channels = residual_channels
        self.dilated_conv = nn.Conv1d(
            residual_channels,
            2 * residual_channels,
            kernel_size=3,
            padding=dilation,
            dilation=dilation
        )
        self.diffusion_projection = nn.Linear(residual_channels, residual_channels)
        self.conditioner_projection = nn.Conv1d(encoder_hidden, 2 * residual_channels, 1)
        self.output_projection = nn.Conv1d(residual_channels, 2 * residual_channels, 1)

    def forward(self, x, conditioner, diffusion_step):
        diffusion_step = self.diffusion_projection(diffusion_step).unsqueeze(-1)
        conditioner = self.conditioner_projection(conditioner)
        y = x + diffusion_step

        y = self.dilated_conv(y) + conditioner

        # Using torch.split instead of torch.chunk to avoid using onnx::Slice
        gate, filter = torch.split(y, [self.residual_channels, self.residual_channels], dim=1)
        y = torch.sigmoid(gate) * torch.tanh(filter)

        y = self.output_projection(y)

        # Using torch.split instead of torch.chunk to avoid using onnx::Slice
        residual, skip = torch.split(y, [self.residual_channels, self.residual_channels], dim=1)
        return (x + residual) / math.sqrt(2.0), skip


class WaveNet(nn.Module):
    def __init__(self, in_dims, n_feats, *, num_layers=20, num_channels=256, dilation_cycle_length=4):
        super().__init__()
        self.in_dims = in_dims
        self.n_feats = n_feats
        self.input_projection = Conv1d(in_dims * n_feats, num_channels, 1)
        self.diffusion_embedding = SinusoidalPosEmb(num_channels)
        self.mlp = nn.Sequential(
            nn.Linear(num_channels, num_channels * 4),
            nn.Mish(),
            nn.Linear(num_channels * 4, num_channels)
        )
        self.residual_layers = nn.ModuleList([
            ResidualBlock(
                encoder_hidden=hparams['hidden_size'],
                residual_channels=num_channels,
                dilation=2 ** (i % dilation_cycle_length)
            )
            for i in range(num_layers)
        ])
        self.skip_projection = Conv1d(num_channels, num_channels, 1)
        self.output_projection = AdamWConv1d(num_channels, in_dims * n_feats, 1)
        nn.init.zeros_(self.output_projection.weight)

    def forward(self, spec, diffusion_step, cond):
        """
        :param spec: [B, F, M, T]
        :param diffusion_step: [B, 1]
        :param cond: [B, H, T]
        :return:
        """
        if self.n_feats == 1:
            # Use indexing instead of squeeze to avoid emitting an onnx::If
            # whose branches have different rank, which breaks shape inference
            # for the downstream Conv on PyTorch >= 2.0.
            x = spec[:, 0]  # [B, M, T]
        else:
            x = spec.flatten(start_dim=1, end_dim=2)  # [B, F x M, T]
        x = self.input_projection(x)  # [B, C, T]

        x = F.relu(x)
        diffusion_step = self.diffusion_embedding(diffusion_step)
        diffusion_step = self.mlp(diffusion_step)
        skip = []
        for layer in self.residual_layers:
            x, skip_connection = layer(x, cond, diffusion_step)
            skip.append(skip_connection)

        x = torch.sum(torch.stack(skip), dim=0) / sqrt(len(self.residual_layers))
        x = self.skip_projection(x)
        x = F.relu(x)
        x = self.output_projection(x)  # [B, M, T]
        if self.n_feats == 1:
            x = x[:, None, :, :]
        else:
            # Using reshape instead of unflatten for ONNX export compatibility
            # x = x.unflatten(dim=1, sizes=(self.n_feats, self.in_dims))
            x = x.reshape(-1, self.n_feats, self.in_dims, x.shape[2])
        return x