Pownie/backgammon
1
0
Fork 0
Code Issues 6 Pull Requests Packages Projects Releases Wiki Activity

fix wrongful mergings

Browse Source
This commit is contained in:
Christoffer Müller Madsen 2018-03-27 13:02:36 +02:00
parent 28b82e8228
commit 2654006222
Signed by: christoffer
GPG Key ID: 337BA5A95E686EFD
1 changed files with 22 additions and 53 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

63
network.py
View File

@ -152,8 +152,8 @@ class Network:
# print("Found the best state, being:", np.array(move_scores).argmax())
return best_move_pair
def eval(self, trained_eps=0):
def do_eval(sess, method, episodes=1000, trained_eps=trained_eps):
def eval(self, episode_count, trained_eps = 0, tf_session = None):
def do_eval(sess, method, episodes = 1000, trained_eps = 0):
start_time = time.time()
def print_time_estimate(eps_completed):
@ -291,15 +291,24 @@ class Network:
sys.stderr.write("[EVAL ] Evaluation method '{}' is not defined\n".format(method))
return [0]
with tf.Session() as session:
if tf_session == None:
with tf.Session():
session.run(tf.global_variables_initializer())
self.restore_model(session)
outcomes = [(method, do_eval(session,
outcomes = [ (method, do_eval(session,
method,
self.config['episode_count'],
trained_eps=trained_eps))
episode_count,
trained_eps = trained_eps))
for method
in self.config['eval_methods']]
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):
@ -401,43 +410,3 @@ class Network:
# save the current state again, so we can continue running backprop based on the "previous" turn.
# NOTE: We need to make a method so that we can take a single turn or at least just pick the next best move, so we know how to evaluate according to TD-learning. Right now, our game just continues in a while loop without nothing to stop it!
def eval(self, episode_count, trained_eps = 0, tf_session = None):
def do_eval(sess, method, episodes = 1000, trained_eps = 0):
start_time = time.time()
writer.close()
return outcomes
# take turn, which finds the best state and picks it, based on the current network
# save current state
# run training operation (session.run(self.training_op, {x:x, value_next, value_next})),
# (something which does the backprop, based on the state after having taken a turn,
# found before, and the state we saved in the beginning and from now we'll
# save it at the end of the turn
# save the current state again, so we can continue running backprop based on the "previous" turn.
if tf_session == None:
with tf.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