import tensorflow as tf import numpy as np from board import Board import os import time import sys import random from eval import Eval import glob from operator import itemgetter class Network: # board_features_quack has size 28 # board_features_quack_fat has size 30 # board_features_tesauro has size 198 board_reps = { 'quack-fat' : (30, Board.board_features_quack_fat), 'quack' : (28, Board.board_features_quack), 'tesauro' : (198, Board.board_features_tesauro), 'quack-norm': (30, Board.board_features_quack_norm) } def custom_tanh(self, x, name=None): return tf.scalar_mul(tf.constant(2.00), tf.tanh(x, name)) def __init__(self, config, name): self.config = config self.checkpoint_path = os.path.join(config['model_storage_path'], config['model']) self.name = name # Set board representation from config self.input_size, self.board_trans_func = Network.board_reps[ self.config['board_representation'] ] self.output_size = 1 self.hidden_size = 40 self.max_learning_rate = 0.1 self.min_learning_rate = 0.001 self.global_step = tf.Variable(0, trainable=False, name="global_step") self.learning_rate = tf.maximum(self.min_learning_rate, tf.train.exponential_decay(self.max_learning_rate, self.global_step, 50000, 0.96, staircase=True), name="learning_rate") # Restore trained episode count for model episode_count_path = os.path.join(self.checkpoint_path, "episodes_trained") if os.path.isfile(episode_count_path): with open(episode_count_path, 'r') as f: self.episodes_trained = int(f.read()) else: self.episodes_trained = 0 self.x = tf.placeholder('float', [1, self.input_size], name='input') self.value_next = tf.placeholder('float', [1, self.output_size], name="value_next") xavier_init = tf.contrib.layers.xavier_initializer() W_1 = tf.get_variable("w_1", (self.input_size, self.hidden_size), initializer=xavier_init) W_2 = tf.get_variable("w_2", (self.hidden_size, self.output_size), initializer=xavier_init) b_1 = tf.get_variable("b_1", (self.hidden_size,), initializer=tf.zeros_initializer) b_2 = tf.get_variable("b_2", (self.output_size,), initializer=tf.zeros_initializer) value_after_input = tf.sigmoid(tf.matmul(self.x, W_1) + b_1, name='hidden_layer') self.value = tf.sigmoid(tf.matmul(value_after_input, W_2) + b_2, name='output_layer') # TODO: Alexander thinks that self.value will be computed twice (instead of once) difference_in_values = tf.reshape(tf.subtract(self.value_next, self.value, name='difference_in_values'), []) tf.summary.scalar("difference_in_values", tf.abs(difference_in_values)) trainable_vars = tf.trainable_variables() gradients = tf.gradients(self.value, trainable_vars) apply_gradients = [] global_step_op = self.global_step.assign_add(1) with tf.variable_scope('apply_gradients'): for gradient, trainable_var in zip(gradients, trainable_vars): backprop_calc = self.learning_rate * difference_in_values * gradient grad_apply = trainable_var.assign_add(backprop_calc) apply_gradients.append(grad_apply) with tf.control_dependencies([global_step_op]): self.training_op = tf.group(*apply_gradients, name='training_op') self.saver = tf.train.Saver(max_to_keep=1) def eval_state(self, sess, state): return sess.run(self.value, feed_dict={self.x: state}) def save_model(self, sess, episode_count, global_step): self.saver.save(sess, os.path.join(self.checkpoint_path, 'model.ckpt'), global_step=global_step) with open(os.path.join(self.checkpoint_path, "episodes_trained"), 'w+') as f: print("[NETWK] ({name}) Saving model to:".format(name=self.name), os.path.join(self.checkpoint_path, 'model.ckpt')) f.write(str(episode_count) + "\n") def restore_model(self, sess): """ Restore a model for a session, such that a trained model and either be further trained or used for evaluation :param sess: Current session :return: Nothing. It's a side-effect that a model gets restored for the network. """ if glob.glob(os.path.join(self.checkpoint_path, 'model.ckpt*.index')): latest_checkpoint = tf.train.latest_checkpoint(self.checkpoint_path) print("[NETWK] ({name}) Restoring model from:".format(name=self.name), str(latest_checkpoint)) self.saver.restore(sess, latest_checkpoint) variables_names = [v.name for v in tf.trainable_variables()] values = sess.run(variables_names) for k, v in zip(variables_names, values): print("Variable: ", k) print("Shape: ", v.shape) print(v) # Restore trained episode count for model episode_count_path = os.path.join(self.checkpoint_path, "episodes_trained") if os.path.isfile(episode_count_path): with open(episode_count_path, 'r') as f: self.config['start_episode'] = int(f.read()) elif glob.glob(os.path.join(os.path.join(self.config['model_storage_path'], "baseline_model"), 'model.ckpt*.index')): checkpoint_path = os.path.join(self.config['model_storage_path'], "baseline_model") latest_checkpoint = tf.train.latest_checkpoint(checkpoint_path) print("[NETWK] ({name}) Restoring model from:".format(name=self.name), str(latest_checkpoint)) self.saver.restore(sess, latest_checkpoint) variables_names = [v.name for v in tf.trainable_variables()] values = sess.run(variables_names) for k, v in zip(variables_names, values): print("Variable: ", k) print("Shape: ", v.shape) print(v) elif not self.config['force_creation']: print("You need to have baseline_model inside models") exit() def make_move(self, sess, board, roll, player): """ Find the best move given a board, roll and a player, by finding all possible states one can go to and then picking the best, by using the network to evaluate each state. The highest score is picked for the 1-player and the max(1-score) is picked for the -1-player. :param sess: :param board: Current board :param roll: Current roll :param player: Current player :return: A pair of the best state to go to, together with the score of that state """ legal_moves = Board.calculate_legal_states(board, player, roll) moves_and_scores = [(move, self.eval_state(sess, self.board_trans_func(move, player))) for move in legal_moves] scores = [x[1] if np.sign(player) > 0 else 1-x[1] for x in moves_and_scores] best_score_index = np.array(scores).argmax() best_move_pair = moves_and_scores[best_score_index] return best_move_pair def calculate_2_ply(self, sess, board, roll, player): """ Find the best move based on a 2-ply look-ahead. First the best move is found for a single ply and then an exhaustive search is performed on the best 15 moves from the single ply. :param sess: :param board: :param roll: The original roll :param player: The current player :return: Best possible move based on 2-ply look-ahead """ # find all legal states from the given board and the given roll init_legal_states = Board.calculate_legal_states(board, player, roll) # find all values for the above boards zero_ply_moves_and_scores = [(move, self.eval_state(sess, self.board_trans_func(move, player))) for move in init_legal_states] # pythons reverse is in place and I can't call [:15] on it, without applying it to an object like so. Fuck. best_fifteen = sorted(zero_ply_moves_and_scores, key=itemgetter(1)) # They're sorted from smallest to largest, therefore we wan't to reverse if the current player is 1, since # player 1 wishes to maximize. It's not needed for player -1, since that player seeks to minimize. if player == 1: best_fifteen.reverse() best_fifteen_boards = [x[0] for x in best_fifteen[:15]] all_rolls_scores = self.do_ply(sess, best_fifteen_boards, player) best_score_index = np.array(all_rolls_scores).argmax() best_board = best_fifteen_boards[best_score_index] return [best_board, max(all_rolls_scores)] def n_ply(self, n_init, sess, boards_init, player_init): def ply(n, boards, player): def calculate_possible_states(board): possible_rolls = [ (1, 1), (1, 2), (1, 3), (1, 4), (1, 5), (1, 6), (2, 2), (2, 3), (2, 4), (2, 5), (2, 6), (3, 3), (3, 4), (3, 5), (3, 6), (4, 4), (4, 5), (4, 6), (5, 5), (5, 6), (6, 6) ] return [ Board.calculate_legal_states(board, player, roll) for roll in possible_rolls ] def find_best_state_score(boards): score_pairs = [ (board, self.eval_state(sess, self.board_trans_func(board, player))) for board in boards ] scores = [ pair[1] for pair in score_pairs ] best_score_pair = score_pairs[np.array(scores).argmax()] return best_score_pair def average_score(boards): return sum(boards)/len(boards) def average_ply_score(board): states_for_rolls = calculate_possible_states(board) best_state_score_for_each_roll = [ find_best_state_score(states) for states in states_for_rolls ] best_score_for_each_roll = [ x[1] for x in best_state_score_for_each_roll ] average_score_var = average_score(best_score_for_each_roll) return average_score_var if n == 1: print("blalhlalha") average_score_pairs = [ (board, average_ply_score(board)) for board in boards ] return average_score_pairs elif n > 1: # n != 1 def average_for_score_pairs(score_pairs): scores = [ pair[1] for pair in score_pairs ] return sum(scores)/len(scores) def average_plain(scores): return sum(scores)/len(scores) print("+"*20) print(n) print(type(boards)) print(boards) possible_states_for_boards = [ (board, calculate_possible_states(board)) for board in boards ] average_score_pairs = [ (inner_boards[0], average_plain([ average_for_score_pairs(ply(n - 1, inner_board, player * -1)) for inner_board in inner_boards[1] ])) for inner_boards in possible_states_for_boards ] return average_score_pairs else: assert False if n_init < 1: print("Unexpected argument n = {}".format(n_init)); exit() boards_with_scores = ply(n_init, boards_init, -1 * player_init) print(boards_with_scores) scores = [ ( pair[1] if player_init == 1 else (1 - pair[1]) ) for pair in boards_with_scores ] best_score_pair = boards_with_scores[np.array(scores).argmax()] return best_score_pair[0] def do_ply(self, sess, boards, player): """ Calculates a single extra ply, resulting in a larger search space for our best move. This is somewhat hardcoded to only do a single ply, seeing that it calls max on all scores, rather than allowing the function to search deeper, which could result in an even larger search space. If we wish to have more than 2-ply, this should be fixed, so we could extend this method to allow for 3-ply. :param sess: :param boards: The boards to try all rolls on :param player: The player of the previous ply :return: An array of scores where each index describes one of the boards which was given as param to this function. """ def gen_21_rolls(): """ Calculate all possible rolls, [[1,1], [1,2] ..] :return: All possible rolls """ a = [] for x in range(1, 7): for y in range(1, 7): if not [x, y] in a and not [y, x] in a: a.append([x, y]) return a all_rolls = gen_21_rolls() all_rolls_scores = [] # loop over boards for a_board in boards: a_board_scores = [] # loop over all rolls, for each board for roll in all_rolls: # find all states we can get to, given the board and roll and the opposite player all_rolls_boards = Board.calculate_legal_states(a_board, player*-1, roll) # find scores for each board found above spec_roll_scores = [self.eval_state(sess, self.board_trans_func(new_board, player*-1)) for new_board in all_rolls_boards] # if the original player is the -1 player, then we need to find (1-value) spec_roll_scores = [x if player == 1 else (1-x) for x in spec_roll_scores] # find the best score best_score = max(spec_roll_scores) # append the best score to a_board_scores, where we keep track of the best score for each board a_board_scores.append(best_score) # save the expected average of board scores all_rolls_scores.append(sum(a_board_scores)/len(a_board_scores)) # return all the average scores return all_rolls_scores def eval(self, episode_count, trained_eps = 0, tf_session = None): """ Used to evaluate a model. Can either use pubeval, a model playing at an intermediate level, or dumbeval a model which has been given random weights, so it acts deterministically random. :param episode_count: The amount of episodes to run :param trained_eps: The amount of episodes the model we want to evaluate, has trained :param tf_session: :return: outcomes: The outcomes of the evaluation session """ def do_eval(sess, method, episodes = 1000, trained_eps = 0): """ Do the actual evaluation :param sess: :param method: Either pubeval or dumbeval :param episodes: Amount of episodes to use in the evaluation :param trained_eps: :return: outcomes : Described above """ start_time = time.time() def print_time_estimate(eps_completed): cur_time = time.time() time_diff = cur_time - start_time eps_per_sec = eps_completed / time_diff secs_per_ep = time_diff / eps_completed eps_remaining = (episodes - eps_completed) sys.stderr.write( "[EVAL ] Averaging {per_sec} episodes per second\n".format(per_sec=round(eps_per_sec, 2))) sys.stderr.write( "[EVAL ] {eps_remaining} episodes remaining; approx. {time_remaining} seconds remaining\n".format( eps_remaining=eps_remaining, time_remaining=int(eps_remaining * secs_per_ep))) sys.stderr.write( "[EVAL ] Evaluating {eps} episode(s) with method '{method}'\n".format(eps=episodes, method=method)) if method == 'pubeval': outcomes = [] for i in range(1, episodes + 1): sys.stderr.write("[EVAL ] Episode {}".format(i)) board = Board.initial_state while Board.outcome(board) is None: roll = (random.randrange(1, 7), random.randrange(1, 7)) board = (self.make_move(sess, board, roll, 1))[0] roll = (random.randrange(1, 7), random.randrange(1, 7)) board = Eval.make_pubeval_move(board, -1, roll)[0][0:26] sys.stderr.write("\t outcome {}".format(Board.outcome(board)[1])) outcomes.append(Board.outcome(board)[1]) sys.stderr.write("\n") if i % 10 == 0: print_time_estimate(i) return outcomes elif method == 'dumbeval': outcomes = [] for i in range(1, episodes + 1): sys.stderr.write("[EVAL ] Episode {}".format(i)) board = Board.initial_state while Board.outcome(board) is None: roll = (random.randrange(1, 7), random.randrange(1, 7)) board = (self.make_move(sess, board, roll, 1))[0] roll = (random.randrange(1, 7), random.randrange(1, 7)) board = Eval.make_dumbeval_move(board, -1, roll)[0][0:26] sys.stderr.write("\t outcome {}".format(Board.outcome(board)[1])) outcomes.append(Board.outcome(board)[1]) sys.stderr.write("\n") if i % 10 == 0: print_time_estimate(i) return outcomes else: sys.stderr.write("[EVAL ] Evaluation method '{}' is not defined\n".format(method)) return [0] if tf_session == None: with tf.Session() as session: session.run(tf.global_variables_initializer()) self.restore_model(session) outcomes = [ (method, do_eval(session, method, episode_count, trained_eps = trained_eps)) for method in self.config['eval_methods'] ] return outcomes else: outcomes = [ (method, do_eval(tf_session, method, episode_count, trained_eps = trained_eps)) for method in self.config['eval_methods'] ] return outcomes def train_model(self, episodes=1000, save_step_size=100, trained_eps=0): with tf.Session() as sess: writer = tf.summary.FileWriter("/tmp/log/tf", sess.graph) sess.run(tf.global_variables_initializer()) self.restore_model(sess) variables_names = [v.name for v in tf.trainable_variables()] values = sess.run(variables_names) for k, v in zip(variables_names, values): print("Variable: ", k) print("Shape: ", v.shape) print(v) start_time = time.time() def print_time_estimate(eps_completed): cur_time = time.time() time_diff = cur_time - start_time eps_per_sec = eps_completed / time_diff secs_per_ep = time_diff / eps_completed eps_remaining = (episodes - eps_completed) sys.stderr.write( "[TRAIN] Averaging {per_sec} episodes per second\n".format(per_sec=round(eps_per_sec, 2))) sys.stderr.write( "[TRAIN] {eps_remaining} episodes remaining; approx. {time_remaining} seconds remaining\n".format( eps_remaining=eps_remaining, time_remaining=int(eps_remaining * secs_per_ep))) sys.stderr.write("[TRAIN] Training {} episodes and save_step_size {}\n".format(episodes, save_step_size)) outcomes = [] for episode in range(1, episodes + 1): sys.stderr.write("[TRAIN] Episode {}".format(episode + trained_eps)) # TODO decide which player should be here player = 1 prev_board = Board.initial_state i = 0 while Board.outcome(prev_board) is None: i += 1 cur_board, cur_board_value = self.make_move(sess, prev_board, (random.randrange(1, 7), random.randrange(1, 7)), player) # adjust weights sess.run(self.training_op, feed_dict={self.x: self.board_trans_func(prev_board, player), self.value_next: cur_board_value}) player *= -1 prev_board = cur_board final_board = prev_board sys.stderr.write("\t outcome {}\t turns {}".format(Board.outcome(final_board)[1], i)) outcomes.append(Board.outcome(final_board)[1]) final_score = np.array([Board.outcome(final_board)[1]]) scaled_final_score = ((final_score + 2) / 4) with tf.name_scope("final"): merged = tf.summary.merge_all() global_step, summary, _ = sess.run([self.global_step, merged, self.training_op], feed_dict={self.x: self.board_trans_func(prev_board, player), self.value_next: scaled_final_score.reshape((1, 1))}) writer.add_summary(summary, episode + trained_eps) sys.stderr.write("\n") if episode % min(save_step_size, episodes) == 0: sys.stderr.write("[TRAIN] Saving model...\n") self.save_model(sess, episode + trained_eps, global_step) if episode % 50 == 0: print_time_estimate(episode) sys.stderr.write("[TRAIN] Saving model for final episode...\n") self.save_model(sess, episode+trained_eps, global_step) writer.close() return outcomes