Madras1's picture
Create app.py
87ac2de verified
raw
history blame
13 kB
import streamlit as st
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import os
import time
# --- CONFIGURAÇÕES ---
BOARD_SIZE = 8
DEVICE = torch.device("cpu")
MODEL_PATH = "checkers_master_final.pth" # Certifique-se de que este arquivo está no Space!
# --- DEFINIÇÃO DAS CLASSES (Rede Neural e Jogo) ---
# A Berta copiou a lógica exata do seu script para garantir que funcione igual.
class Checkers:
def get_initial_board(self):
board = np.zeros((BOARD_SIZE, BOARD_SIZE), dtype=np.int8)
for r in range(3):
for c in range(BOARD_SIZE):
if (r + c) % 2 == 1: board[r, c] = -1
for r in range(5, BOARD_SIZE):
for c in range(BOARD_SIZE):
if (r + c) % 2 == 1: board[r, c] = 1
return board
def get_valid_moves(self, board, player):
jumps = self._get_all_jumps(board, player)
if jumps: return jumps
moves = []
for r in range(BOARD_SIZE):
for c in range(BOARD_SIZE):
if board[r, c] * player > 0: moves.extend(self._get_simple_moves(board, r, c))
return moves
def _get_simple_moves(self, board, r, c):
moves = []; piece = board[r, c]; player = np.sign(piece)
directions = [(-1, -1), (-1, 1)] if player == 1 else [(1, -1), (1, 1)]
if abs(piece) == 2: directions.extend([(1, -1), (1, 1)] if player == 1 else [(-1, -1), (-1, 1)])
for dr, dc in directions:
nr, nc = r + dr, c + dc
if 0 <= nr < BOARD_SIZE and 0 <= nc < BOARD_SIZE and board[nr, nc] == 0: moves.append(((r, c), (nr, nc)))
return moves
def _get_all_jumps(self, board, player):
all_jumps = []
for r in range(BOARD_SIZE):
for c in range(BOARD_SIZE):
if board[r, c] * player > 0:
jumps = self._find_jump_sequences(np.copy(board), r, c)
if jumps: all_jumps.extend(jumps)
if not all_jumps: return []
max_len = max(len(j) for j in all_jumps)
return [j for j in all_jumps if len(j) == max_len]
def _find_jump_sequences(self, board, r, c, path=[]):
piece = board[r, c]; player = np.sign(piece)
if piece == 0: return []
directions = [(-1, -1), (-1, 1), (1, -1), (1, 1)] if abs(piece) == 2 else \
[(-1, -1), (-1, 1)] if player == 1 else [(1, -1), (1, 1)]
found_jumps = []
for dr, dc in directions:
mid_r, mid_c = r + dr, c + dc; end_r, end_c = r + 2*dr, c + 2*dc
if 0 <= end_r < BOARD_SIZE and 0 <= end_c < BOARD_SIZE and \
board[mid_r, mid_c] * player < 0 and board[end_r, end_c] == 0:
move = ((r, c), (end_r, end_c))
new_board = np.copy(board); new_board[end_r, end_c] = piece; new_board[r, c] = 0; new_board[mid_r, mid_c] = 0
next_jumps = self._find_jump_sequences(new_board, end_r, end_c, path + [move])
if next_jumps: found_jumps.extend(next_jumps)
else: found_jumps.append(path + [move])
return found_jumps
def apply_move(self, board, move):
b_ = np.copy(board)
is_jump_chain = isinstance(move, list) or (isinstance(move, tuple) and isinstance(move[0], tuple) and isinstance(move[0][0], tuple))
sub_moves = move if is_jump_chain else [move]
for (r1, c1), (r2, c2) in sub_moves:
piece = b_[r1, c1]
if piece == 0: continue
b_[r2, c2] = piece; b_[r1, c1] = 0
if abs(r1 - r2) == 2: b_[(r1+r2)//2, (c1+c2)//2] = 0
r_final, c_final = sub_moves[-1][1]; p_final = b_[r_final, c_final]
if p_final == 1 and r_final == 0: b_[r_final, c_final] = 2
if p_final == -1 and r_final == BOARD_SIZE-1: b_[r_final, c_final] = -2
return b_
def check_game_over(self, board, player):
if not self.get_valid_moves(board, player): return -1
if not np.any(np.sign(board) == -player): return 1
return None
def state_to_tensor(board, player):
tensor = np.zeros((5, BOARD_SIZE, BOARD_SIZE), dtype=np.float32)
tensor[0, board == player] = 1; tensor[1, board == player*2] = 1
tensor[2, board == -player] = 1; tensor[3, board == -player*2] = 1
if player == 1: tensor[4,:,:] = 1.0
return torch.from_numpy(tensor).unsqueeze(0).to(DEVICE)
class PolicyValueNetwork(nn.Module):
def __init__(self):
super().__init__()
num_channels = 64
self.body = nn.Sequential(nn.Conv2d(5, num_channels, 3, padding=1), nn.BatchNorm2d(num_channels), nn.ReLU(),
nn.Conv2d(num_channels, num_channels, 3, padding=1), nn.BatchNorm2d(num_channels), nn.ReLU(),
nn.Conv2d(num_channels, num_channels, 3, padding=1), nn.BatchNorm2d(num_channels), nn.ReLU())
self.policy_head = nn.Sequential(nn.Conv2d(num_channels, 4, 1), nn.BatchNorm2d(4), nn.ReLU(), nn.Flatten(),
nn.Linear(4 * BOARD_SIZE * BOARD_SIZE, BOARD_SIZE * BOARD_SIZE))
self.value_head = nn.Sequential(nn.Conv2d(num_channels, 2, 1), nn.BatchNorm2d(2), nn.ReLU(), nn.Flatten(),
nn.Linear(2 * BOARD_SIZE * BOARD_SIZE, 64), nn.ReLU(),
nn.Linear(64, 1), nn.Tanh())
def forward(self, x):
x = self.body(x); return self.policy_head(x), self.value_head(x)
class MCTSNode:
def __init__(self, parent=None, prior=0.0):
self.parent = parent; self.prior = prior; self.children = {}; self.visits = 0; self.value_sum = 0.0
def get_value(self): return self.value_sum / self.visits if self.visits > 0 else 0.0
class MCTS:
def __init__(self, game, model, sims=100, c_puct=1.5):
self.game, self.model, self.sims, self.c_puct = game, model, sims, c_puct
def run(self, board, player):
root = MCTSNode()
self._expand_and_evaluate(root, board, player)
for _ in range(self.sims):
node, search_board, search_player = root, np.copy(board), player
search_path = [root]
while node.children:
move, node = self._select_child(node)
search_board = self.game.apply_move(search_board, move); search_player *= -1; search_path.append(node)
value = self.game.check_game_over(search_board, search_player)
if value is None and node.visits == 0: value = self._expand_and_evaluate(node, search_board, search_player)
elif value is None: value = node.get_value()
for n in reversed(search_path): n.visits += 1; n.value_sum += value; value *= -1
moves = list(root.children.keys())
visits = np.array([root.children[m].visits for m in moves])
return moves, visits / (np.sum(visits) + 1e-9)
def _select_child(self, node):
sqrt_total_visits = np.sqrt(node.visits); best_move, max_score = None, -np.inf
for move, child in node.children.items():
score = -child.get_value() + self.c_puct * child.prior * sqrt_total_visits / (1 + child.visits)
if score > max_score: max_score, best_move = score, move
return best_move, node.children[best_move]
def _expand_and_evaluate(self, node, board, player):
valid_moves = self.game.get_valid_moves(board, player)
if not valid_moves: return -1.0
with torch.no_grad():
policy_logits, value_tensor = self.model(state_to_tensor(board, player))
value = value_tensor.item()
policy_probs = F.softmax(policy_logits, dim=1).cpu().numpy()[0]
move_priors = {}; total_prior = 0
for move in valid_moves:
if isinstance(move, list): start_pos_tuple = move[0][0]
else: start_pos_tuple = move[0]
start_pos_idx = start_pos_tuple[0] * BOARD_SIZE + start_pos_tuple[1]
prior = policy_probs[start_pos_idx]
key = tuple(move) if isinstance(move, list) else move
move_priors[key] = prior; total_prior += prior
if total_prior > 0:
for move_key, prior in move_priors.items(): node.children[move_key] = MCTSNode(parent=node, prior=prior / total_prior)
else:
for move in valid_moves:
key = tuple(move) if isinstance(move, list) else move
node.children[key] = MCTSNode(parent=node, prior=1.0 / len(valid_moves))
return value
# --- INTERFACE DO STREAMLIT ---
st.set_page_config(page_title="AlphaCheckerZero", page_icon="♟️")
st.title("♟️ AlphaCheckerZero Arena")
st.write("Gabriel Yogi's Neural Network AI")
# 1. Carregar o Modelo (com Cache para ser rápido)
@st.cache_resource
def load_model():
if not os.path.exists(MODEL_PATH):
return None
model = PolicyValueNetwork().to(DEVICE)
model.load_state_dict(torch.load(MODEL_PATH, map_location=DEVICE))
model.eval()
return model
model = load_model()
if model is None:
st.error(f"Arquivo '{MODEL_PATH}' não encontrado. Por favor, faça upload do arquivo .pth para o Space.")
st.stop()
# 2. Inicializar o Estado do Jogo
if "board" not in st.session_state:
game = Checkers()
st.session_state.board = game.get_initial_board()
st.session_state.player = 1 # Humano começa (1)
st.session_state.game_over = False
st.session_state.message = "Sua vez! Você joga com as Brancas (x)."
game = Checkers()
mcts = MCTS(game, model, sims=150) # Sims ajustado para performance na web
# Função para desenhar o tabuleiro como texto (simples e funcional)
def render_board(board):
chars = {1: 'x', 2: 'X', -1: 'o', -2: 'O', 0: '.'}
board_str = " 0 1 2 3 4 5 6 7\n"
board_str += " -----------------\n"
for r_idx, row in enumerate(board):
board_str += f"{r_idx} | {' '.join(chars[val] for val in row)} |\n"
board_str += " -----------------"
return board_str
# Layout principal
col1, col2 = st.columns([2, 1])
with col1:
st.text_area("Tabuleiro", render_board(st.session_state.board), height=250, disabled=True, key="board_display")
with col2:
st.write("### Status")
st.info(st.session_state.message)
if st.button("Reiniciar Jogo"):
st.session_state.board = game.get_initial_board()
st.session_state.player = 1
st.session_state.game_over = False
st.session_state.message = "Jogo reiniciado. Sua vez!"
st.rerun()
# Lógica do Jogo
if not st.session_state.game_over:
# Verificar fim de jogo antes de qualquer coisa
result = game.check_game_over(st.session_state.board, st.session_state.player)
if result is not None:
st.session_state.game_over = True
if result == 1: st.session_state.message = "VOCÊ GANHOU! Parabéns Gabriel!"
elif result == -1: st.session_state.message = "A IA GANHOU. Mais sorte na próxima."
else: st.session_state.message = "EMPATE."
st.rerun()
# VEZ DO HUMANO (Player 1)
if st.session_state.player == 1:
valid_moves = game.get_valid_moves(st.session_state.board, 1)
if not valid_moves:
# Se não tem movimentos e não deu game over acima, algo estranho aconteceu, mas tratamos como derrota
st.session_state.game_over = True
st.session_state.message = "Sem movimentos válidos. Você perdeu."
st.rerun()
# Criar lista de strings para o Selectbox
move_options = [str(m) for m in valid_moves]
selected_move_str = st.selectbox("Escolha sua jogada:", move_options)
if st.button("Jogar"):
# Encontrar o movimento original baseado na string
move_idx = move_options.index(selected_move_str)
move = valid_moves[move_idx]
# Aplicar movimento
st.session_state.board = game.apply_move(st.session_state.board, move)
st.session_state.player = -1 # Passa a vez para IA
st.session_state.message = "A IA está pensando..."
st.rerun()
# VEZ DA IA (Player -1)
else:
with st.spinner("A AlphaCheckerZero está pensando..."):
# Pequeno delay para a interface atualizar e mostrar a mensagem
time.sleep(0.5)
valid_moves, policy = mcts.run(np.copy(st.session_state.board), -1)
if not valid_moves:
st.session_state.game_over = True
st.session_state.message = "A IA não tem movimentos. Você venceu!"
st.rerun()
move = valid_moves[np.argmax(policy)]
st.session_state.board = game.apply_move(st.session_state.board, move)
st.session_state.player = 1 # Devolve a vez para Humano
st.session_state.message = f"IA jogou: {move}. Sua vez!"
st.rerun()
else:
st.success(st.session_state.message)