import numpy as np

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


class ResidualBlock(nn.Module):
    def __init__(self, channels):
        super().__init__()
        self.conv1 = nn.Conv2d(channels, channels, kernel_size=3, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(channels)
        self.conv2 = nn.Conv2d(channels, channels, kernel_size=3, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(channels)

    def forward(self, x):
        residual = x
        x = F.relu(self.bn1(self.conv1(x)))
        x = self.bn2(self.conv2(x))
        return F.relu(x + residual)


class UltimateTicTacToeModel(nn.Module):
    def __init__(self, board_size, action_size, device, channels=64, num_blocks=6):
        super().__init__()

        self.action_size = action_size
        self.input_shape = board_size
        self.input_channels = board_size[0]
        self.board_height = board_size[1]
        self.board_width = board_size[2]
        self.device = torch.device(device)

        self.stem = nn.Sequential(
            nn.Conv2d(self.input_channels, channels, kernel_size=3, padding=1, bias=False),
            nn.BatchNorm2d(channels),
            nn.ReLU(inplace=True),
        )
        self.residual_tower = nn.Sequential(*(ResidualBlock(channels) for _ in range(num_blocks)))

        self.policy_head = nn.Sequential(
            nn.Conv2d(channels, 32, kernel_size=1, bias=False),
            nn.BatchNorm2d(32),
            nn.ReLU(inplace=True),
        )
        self.policy_fc = nn.Linear(32 * self.board_height * self.board_width, self.action_size)

        self.value_head = nn.Sequential(
            nn.Conv2d(channels, 32, kernel_size=1, bias=False),
            nn.BatchNorm2d(32),
            nn.ReLU(inplace=True),
        )
        self.value_fc1 = nn.Linear(32 * self.board_height * self.board_width, 128)
        self.value_fc2 = nn.Linear(128, 1)

        self.to(self.device)

    def forward(self, x):
        x = x.view(-1, *self.input_shape)
        x = self.stem(x)
        x = self.residual_tower(x)

        policy = self.policy_head(x)
        policy = torch.flatten(policy, 1)
        policy = self.policy_fc(policy)

        value = self.value_head(x)
        value = torch.flatten(value, 1)
        value = F.relu(self.value_fc1(value))
        value = torch.tanh(self.value_fc2(value))

        return F.softmax(policy, dim=1), value

    def predict(self, board):
        board = torch.as_tensor(board, dtype=torch.float32, device=self.device)
        board = board.view(1, *self.input_shape)
        self.eval()
        with torch.no_grad():
            pi, v = self.forward(board)

        return pi.detach().cpu().numpy()[0], float(v.item())