fixed dumb bugs; still messy
This commit is contained in:
parent
08481e508a
commit
2fc7a2a09c
10
bot.py
10
bot.py
|
|
@ -7,15 +7,21 @@ import numpy as np
|
|||
import random
|
||||
|
||||
class Bot:
|
||||
def __init__(self, sym, config = None):
|
||||
def __init__(self, sym, config = None, name = "unnamed"):
|
||||
self.config = config
|
||||
self.cup = Cup()
|
||||
self.sym = sym
|
||||
self.graph = tf.Graph()
|
||||
with self.graph.as_default():
|
||||
self.session = tf.Session()
|
||||
self.network = Network(self.session, config)
|
||||
self.network = Network(self.session, config, name)
|
||||
self.network.restore_model()
|
||||
variables_names = [v.name for v in tf.trainable_variables()]
|
||||
values = self.session.run(variables_names)
|
||||
for k, v in zip(variables_names, values):
|
||||
print("Variable: ", k)
|
||||
print("Shape: ", v.shape)
|
||||
print(v)
|
||||
|
||||
def restore_model(self):
|
||||
with self.graph.as_default():
|
||||
|
|
|
|||
10
eval.py
10
eval.py
|
|
@ -14,8 +14,14 @@ class Eval:
|
|||
# TODO: Test this, the score results should be deterministic
|
||||
@staticmethod
|
||||
def make_pubeval_move(board, sym, roll):
|
||||
legal_moves = Board.calculate_legal_states(tuple(board), sym, roll)
|
||||
moves_and_scores = [(board, pubeval.eval(False, Board.board_features_to_pubeval(board, sym))) for board in legal_moves]
|
||||
legal_moves = Board.calculate_legal_states(board, sym, roll)
|
||||
moves_and_scores = [ ( board,
|
||||
pubeval.eval(False, Board.board_features_to_pubeval(board, sym)))
|
||||
for board
|
||||
in legal_moves ]
|
||||
scores = [ x[1] for x in moves_and_scores ]
|
||||
best_move_pair = moves_and_scores[np.array(scores).argmax()]
|
||||
|
||||
return best_move_pair
|
||||
|
||||
|
||||
|
|
|
|||
48
game.py
48
game.py
|
|
@ -8,6 +8,7 @@ from eval import Eval
|
|||
import numpy as np
|
||||
import sys
|
||||
import time
|
||||
import os # for path join
|
||||
|
||||
class Game:
|
||||
|
||||
|
|
@ -21,8 +22,8 @@ class Game:
|
|||
self.cup = Cup()
|
||||
|
||||
def set_up_bots(self):
|
||||
self.p1 = Bot(1, config = self.config)
|
||||
self.p2 = Bot(1, config = self.config)
|
||||
self.p1 = Bot(1, config = self.config, name = "p1")
|
||||
self.p2 = Bot(1, config = self.config, name = "p2")
|
||||
|
||||
def roll(self):
|
||||
return self.cup.roll()
|
||||
|
|
@ -169,7 +170,7 @@ class Game:
|
|||
roll = self.roll()
|
||||
self.board = (self.p1.make_move(self.board, self.p1.get_sym(), roll))[0]
|
||||
roll = self.roll()
|
||||
self.board = Board.flip(Eval.make_random_move(Board.flip(self.board), self.p2.get_sym(), roll))
|
||||
self.board = Board.flip(Eval.make_random_move(Board.flip(self.board), 1, roll))
|
||||
sys.stderr.write("\t outcome {}".format(Board.outcome(self.board)[1]))
|
||||
outcomes.append(Board.outcome(self.board)[1])
|
||||
sys.stderr.write("\n")
|
||||
|
|
@ -183,11 +184,28 @@ class Game:
|
|||
for i in range(1, episodes + 1):
|
||||
sys.stderr.write("[EVAL ] Episode {}".format(i))
|
||||
self.board = Board.initial_state
|
||||
#print("init:", self.board, sep="\n")
|
||||
while Board.outcome(self.board) is None:
|
||||
#print("-"*30)
|
||||
roll = self.roll()
|
||||
#print(roll)
|
||||
|
||||
prev_board = tuple(self.board)
|
||||
self.board = (self.p1.make_move(self.board, self.p1.get_sym(), roll))[0]
|
||||
#print("post p1:", self.board, sep="\n")
|
||||
|
||||
#print("."*30)
|
||||
roll = self.roll()
|
||||
self.board = Board.flip(Eval.make_pubeval_move(self.board, self.p2.get_sym(), roll)[0][0:26])
|
||||
#print(roll)
|
||||
|
||||
prev_board = tuple(self.board)
|
||||
self.board = Eval.make_pubeval_move(self.board, -1, roll)[0][0:26]
|
||||
#print("post pubeval:", self.board, sep="\n")
|
||||
|
||||
|
||||
#print("*"*30)
|
||||
#print(self.board)
|
||||
#print("+"*30)
|
||||
sys.stderr.write("\t outcome {}".format(Board.outcome(self.board)[1]))
|
||||
outcomes.append(Board.outcome(self.board)[1])
|
||||
sys.stderr.write("\n")
|
||||
|
|
@ -196,6 +214,28 @@ class Game:
|
|||
print_time_estimate(i)
|
||||
|
||||
return outcomes
|
||||
elif method == 'dumbmodel':
|
||||
config_prime = self.config.copy()
|
||||
config_prime['model_path'] = os.path.join(config_prime['model_storage_path'], 'dumbmodel')
|
||||
eval_bot = Bot(1, config = config_prime, name = "dumbmodel")
|
||||
#print(self.config, "\n", config_prime)
|
||||
outcomes = []
|
||||
for i in range(1, episodes + 1):
|
||||
sys.stderr.write("[EVAL ] Episode {}".format(i))
|
||||
self.board = Board.initial_state
|
||||
while Board.outcome(self.board) is None:
|
||||
roll = self.roll()
|
||||
self.board = (self.p1.make_move(self.board, self.p1.get_sym(), roll))[0]
|
||||
|
||||
roll = self.roll()
|
||||
self.board = Board.flip(eval_bot.make_move(Board.flip(self.board), self.p1.get_sym(), roll)[0])
|
||||
sys.stderr.write("\t outcome {}".format(Board.outcome(self.board)[1]))
|
||||
outcomes.append(Board.outcome(self.board)[1])
|
||||
sys.stderr.write("\n")
|
||||
|
||||
if i % 50 == 0:
|
||||
print_time_estimate(i)
|
||||
return outcomes
|
||||
else:
|
||||
sys.stderr.write("[EVAL ] Evaluation method '{}' is not defined\n".format(method))
|
||||
return [0]
|
||||
|
|
|
|||
3
main.py
3
main.py
|
|
@ -73,7 +73,8 @@ config = {
|
|||
'eval': args.eval,
|
||||
'eval_after_train': args.eval_after_train,
|
||||
'start_episode': args.start_episode,
|
||||
'train_perpetually': args.train_perpetually
|
||||
'train_perpetually': args.train_perpetually,
|
||||
'model_storage_path': model_storage_path
|
||||
}
|
||||
|
||||
# Make sure directories exist
|
||||
|
|
|
|||
64
network.py
64
network.py
|
|
@ -15,15 +15,13 @@ class Network:
|
|||
#os.environ["TF_CPP_MIN_LOG_LEVEL"]="2"
|
||||
|
||||
def custom_tanh(self, x, name=None):
|
||||
a = tf.Variable(2.00, tf.float32)
|
||||
return tf.scalar_mul(a, tf.tanh(x, name))
|
||||
|
||||
return tf.scalar_mul(tf.constant(2.00), tf.tanh(x, name))
|
||||
|
||||
def __init__(self, session, config = None):
|
||||
def __init__(self, session, config, name):
|
||||
self.config = config
|
||||
self.session = session
|
||||
self.checkpoint_path = config['model_path']
|
||||
|
||||
self.name = name
|
||||
|
||||
# Restore trained episode count for model
|
||||
episode_count_path = os.path.join(self.checkpoint_path, "episodes_trained")
|
||||
|
|
@ -37,18 +35,23 @@ class Network:
|
|||
|
||||
xavier_init = tf.contrib.layers.xavier_initializer()
|
||||
|
||||
W_1 = tf.Variable(xavier_init((Network.input_size, Network.hidden_size)))
|
||||
W_2 = tf.Variable(xavier_init((Network.hidden_size, Network.output_size)))
|
||||
W_1 = tf.get_variable("w_1", (Network.input_size, Network.hidden_size),
|
||||
initializer=xavier_init)
|
||||
W_2 = tf.get_variable("w_2", (Network.hidden_size, Network.output_size),
|
||||
initializer=xavier_init)
|
||||
|
||||
b_1 = tf.zeros(Network.hidden_size,)
|
||||
b_2 = tf.zeros(Network.output_size,)
|
||||
b_1 = tf.get_variable("b_1", (Network.hidden_size,),
|
||||
initializer=tf.zeros_initializer)
|
||||
b_2 = tf.get_variable("b_2", (Network.output_size,),
|
||||
initializer=tf.zeros_initializer)
|
||||
|
||||
value_after_input = self.custom_tanh(tf.matmul(self.x, W_1) + b_1, name='hidden_layer')
|
||||
|
||||
# TODO: Remember to make this tanh * 2
|
||||
self.value = self.custom_tanh(tf.matmul(value_after_input, W_2) + b_2, name='output_layer')
|
||||
|
||||
# tf.reduce_sum basically finds the sum of it's input, so this gives the difference between the two values, in case they should be lists, which they might be if our input changes
|
||||
# tf.reduce_sum basically finds the sum of its input, so this gives the
|
||||
# difference between the two values, in case they should be lists, which
|
||||
# they might be if our input changes
|
||||
difference_in_values = tf.reduce_sum(self.value_next - self.value, name='difference')
|
||||
|
||||
trainable_vars = tf.trainable_variables()
|
||||
|
|
@ -71,32 +74,53 @@ class Network:
|
|||
def eval_state(self, state):
|
||||
# Run state through a network
|
||||
|
||||
# Remember to create placeholders for everything because wtf tensorflow and graphs
|
||||
# Remember to create placeholders for everything because wtf tensorflow
|
||||
# and graphs
|
||||
|
||||
# Remember to create the dense layers
|
||||
|
||||
# Figure out a way of giving a layer a custom activiation function (we want something which gives [-2,2]. Naively tahn*2, however I fell this is wrong.
|
||||
# Figure out a way of giving a layer a custom activiation function (we
|
||||
# want something which gives [-2,2]. Naively tahn*2, however I fell this
|
||||
# is wrong.
|
||||
|
||||
# tf.group, groups a bunch of actions, so calculate the different gradients for the different weights, by using tf.trainable_variables() to find all variables and tf.gradients(current_value, trainable_variables) to find all the gradients. We can then loop through this and calculate the trace for each gradient and variable pair (note, zip can be used to combine the two lists found before), and then we can calculate the overall change in weights, based on the formula listed in tesauro (learning_rate * difference_in_values * trace), this calculation can be assigned to a tf variable and put in a list and then this can be grouped into a single operation, essentially building our own backprop function.
|
||||
# Grouping them is done by tf.group(*the_gradients_from_before_we_want_to_apply, name="training_op")
|
||||
# tf.group, groups a bunch of actions, so calculate the different
|
||||
# gradients for the different weights, by using tf.trainable_variables()
|
||||
# to find all variables and tf.gradients(current_value,
|
||||
# trainable_variables) to find all the gradients. We can then loop
|
||||
# through this and calculate the trace for each gradient and variable
|
||||
# pair (note, zip can be used to combine the two lists found before),
|
||||
# and then we can calculate the overall change in weights, based on the
|
||||
# formula listed in tesauro (learning_rate * difference_in_values *
|
||||
# trace), this calculation can be assigned to a tf variable and put in a
|
||||
# list and then this can be grouped into a single operation, essentially
|
||||
# building our own backprop function.
|
||||
|
||||
# If we remove the eligibily trace to begin with, we only have
|
||||
# to implement learning_rate * (difference_in_values) * gradients (the before-mentioned calculation.
|
||||
# Grouping them is done by
|
||||
# tf.group(*the_gradients_from_before_we_want_to_apply,
|
||||
# name="training_op")
|
||||
|
||||
# If we remove the eligibily trace to begin with, we only have to
|
||||
# implement learning_rate * (difference_in_values) * gradients (the
|
||||
# before-mentioned calculation.
|
||||
|
||||
|
||||
# print("Network is evaluating")
|
||||
# print("Network is evaluating")
|
||||
val = self.session.run(self.value, feed_dict={self.x: state})
|
||||
|
||||
#print("eval ({})".format(self.name), state, val, sep="\n")
|
||||
return val
|
||||
|
||||
def save_model(self, episode_count):
|
||||
self.saver.save(self.session, os.path.join(self.checkpoint_path, 'model.ckpt'))
|
||||
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):
|
||||
if os.path.isfile(self.checkpoint_path):
|
||||
if os.path.isfile(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(self.session, latest_checkpoint)
|
||||
|
||||
# Have a circular dependency, #fuck, need to rewrite something
|
||||
|
|
|
|||
Loading…
Reference in New Issue
Block a user