import torch
import torch.nn as nn
import torch.nn.init as init
import math


class TwoStreamLipNet(torch.nn.Module):
    def __init__(self, dropout_p=0.5, coord_input_dim=40, coord_hidden_dim=128):
        super(TwoStreamLipNet, self).__init__()
        self.conv1 = nn.Conv3d(3, 32, (3, 5, 5), (1, 2, 2), (1, 2, 2))
        self.pool1 = nn.MaxPool3d((1, 2, 2), (1, 2, 2))

        self.conv2 = nn.Conv3d(32, 64, (3, 5, 5), (1, 1, 1), (1, 2, 2))
        self.pool2 = nn.MaxPool3d((1, 2, 2), (1, 2, 2))

        self.conv3 = nn.Conv3d(64, 96, (3, 3, 3), (1, 1, 1), (1, 1, 1))
        self.pool3 = nn.MaxPool3d((1, 2, 2), (1, 2, 2))

        self.gru1 = nn.GRU(96 * 4 * 8, 256, 1, bidirectional=True)
        self.gru2 = nn.GRU(512, 256, 1, bidirectional=True)

        self.FC = nn.Linear(512 + 2 * coord_hidden_dim, 27 + 1)
        self.dropout_p = dropout_p

        self.relu = nn.ReLU(inplace=True)
        self.dropout = nn.Dropout(self.dropout_p)
        self.dropout3d = nn.Dropout3d(self.dropout_p)

        # New GRU layer for lip coordinates
        self.coord_gru = nn.GRU(
            coord_input_dim, coord_hidden_dim, 1, bidirectional=True
        )

        self._init()

    def _init(self):
        init.kaiming_normal_(self.conv1.weight, nonlinearity="relu")
        init.constant_(self.conv1.bias, 0)

        init.kaiming_normal_(self.conv2.weight, nonlinearity="relu")
        init.constant_(self.conv2.bias, 0)

        init.kaiming_normal_(self.conv3.weight, nonlinearity="relu")
        init.constant_(self.conv3.bias, 0)

        init.kaiming_normal_(self.FC.weight, nonlinearity="sigmoid")
        init.constant_(self.FC.bias, 0)

        for m in (self.gru1, self.gru2):
            stdv = math.sqrt(2 / (96 * 3 * 6 + 256))
            for i in range(0, 256 * 3, 256):
                init.uniform_(
                    m.weight_ih_l0[i : i + 256],
                    -math.sqrt(3) * stdv,
                    math.sqrt(3) * stdv,
                )
                init.orthogonal_(m.weight_hh_l0[i : i + 256])
                init.constant_(m.bias_ih_l0[i : i + 256], 0)
                init.uniform_(
                    m.weight_ih_l0_reverse[i : i + 256],
                    -math.sqrt(3) * stdv,
                    math.sqrt(3) * stdv,
                )
                init.orthogonal_(m.weight_hh_l0_reverse[i : i + 256])
                init.constant_(m.bias_ih_l0_reverse[i : i + 256], 0)

    def forward(self, x, coords):
        # branch 1
        x = self.conv1(x)
        x = self.relu(x)
        x = self.dropout3d(x)
        x = self.pool1(x)

        x = self.conv2(x)
        x = self.relu(x)
        x = self.dropout3d(x)
        x = self.pool2(x)

        x = self.conv3(x)
        x = self.relu(x)
        x = self.dropout3d(x)
        x = self.pool3(x)

        # (B, C, T, H, W)->(T, B, C, H, W)
        x = x.permute(2, 0, 1, 3, 4).contiguous()
        # (B, C, T, H, W)->(T, B, C*H*W)
        x = x.view(x.size(0), x.size(1), -1)

        self.gru1.flatten_parameters()
        self.gru2.flatten_parameters()

        x, h = self.gru1(x)
        x = self.dropout(x)
        x, h = self.gru2(x)
        x = self.dropout(x)

        # branch 2
        # Process lip coordinates through GRU
        self.coord_gru.flatten_parameters()

        # (B, T, N, C)->(T, B, C, N, C)
        coords = coords.permute(1, 0, 2, 3).contiguous()
        # (T, B, C, N, C)->(T, B, C, N*C)
        coords = coords.view(coords.size(0), coords.size(1), -1)
        coords, _ = self.coord_gru(coords)
        coords = self.dropout(coords)

        # combine the two branches
        combined = torch.cat((x, coords), dim=2)

        x = self.FC(combined)
        x = x.permute(1, 0, 2).contiguous()
        return x