def test_save_load(self): policy = CNNPolicy(["board", "liberties", "sensibleness", "capture_size"]) model_file = 'TESTPOLICY.json' weights_file = 'TESTWEIGHTS.h5' model_file2 = 'TESTPOLICY2.json' weights_file2 = 'TESTWEIGHTS2.h5' # test saving model/weights separately policy.save_model(model_file) policy.model.save_weights(weights_file, overwrite=True) # test saving them together policy.save_model(model_file2, weights_file2) copypolicy = CNNPolicy.load_model(model_file) copypolicy.model.load_weights(weights_file) copypolicy2 = CNNPolicy.load_model(model_file2) for w1, w2 in zip(copypolicy.model.get_weights(), copypolicy2.model.get_weights()): self.assertTrue(np.all(w1 == w2)) os.remove(model_file) os.remove(weights_file) os.remove(model_file2) os.remove(weights_file2)
def test_save_load(self):
policy = CNNPolicy(["board", "liberties", "sensibleness", "capture_size"])
model_file = "TESTPOLICY.json"
weights_file = "TESTWEIGHTS.h5"
model_file2 = "TESTPOLICY2.json"
weights_file2 = "TESTWEIGHTS2.h5"
# test saving model/weights separately
policy.save_model(model_file)
policy.model.save_weights(weights_file, overwrite=True)
# test saving them together
policy.save_model(model_file2, weights_file2)
copypolicy = CNNPolicy.load_model(model_file)
copypolicy.model.load_weights(weights_file)
copypolicy2 = CNNPolicy.load_model(model_file2)
for w1, w2 in zip(copypolicy.model.get_weights(), copypolicy2.model.get_weights()):
self.assertTrue(np.all(w1 == w2))
os.remove(model_file)
os.remove(weights_file)
os.remove(model_file2)
os.remove(weights_file2)
def test_save_load(self): policy = CNNPolicy(["board", "liberties", "sensibleness", "capture_size"]) model_file = 'TESTPOLICY.json' weights_file = 'TESTWEIGHTS.h5' policy.save_model(model_file) policy.model.save_weights(weights_file) copypolicy = CNNPolicy.load_model(model_file) copypolicy.model.load_weights(weights_file) os.remove(model_file) os.remove(weights_file)
def test_save_load(self):
policy = CNNPolicy(
["board", "liberties", "sensibleness", "capture_size"])
model_file = 'TESTPOLICY.json'
weights_file = 'TESTWEIGHTS.h5'
policy.save_model(model_file)
policy.model.save_weights(weights_file)
copypolicy = CNNPolicy.load_model(model_file)
copypolicy.model.load_weights(weights_file)
os.remove(model_file)
os.remove(weights_file)
from AlphaGo.training.reinforcement_policy_trainer import run_training
from AlphaGo.models.policy import CNNPolicy
import os
from cProfile import Profile
# make a miniature model for playing on a miniature 7x7 board
architecture = {'filters_per_layer': 32, 'layers': 4, 'board': 7}
features = ['board', 'ones', 'turns_since', 'liberties', 'capture_size', 'self_atari_size', 'liberties_after', 'sensibleness']
policy = CNNPolicy(features, **architecture)
datadir = os.path.join('benchmarks', 'data')
modelfile = os.path.join(datadir, 'mini_rl_model.json')
weights = os.path.join(datadir, 'init_weights.hdf5')
outdir = os.path.join(datadir, 'rl_output')
stats_file = os.path.join(datadir, 'reinforcement_policy_trainer.prof')
if not os.path.exists(datadir):
os.makedirs(datadir)
if not os.path.exists(weights):
policy.model.save_weights(weights)
policy.save_model(modelfile)
profile = Profile()
arguments = (modelfile, weights, outdir, '--learning-rate', '0.001', '--save-every', '2', '--game-batch', '20', '--iterations', '10', '--verbose')
profile.runcall(run_training, arguments)
profile.dump_stats(stats_file)
from AlphaGo.training.supervised_policy_trainer import run_training
from AlphaGo.models.policy import CNNPolicy
from cProfile import Profile
architecture = {'filters_per_layer': 128, 'layers': 12}
features = ['board', 'ones', 'turns_since']
policy = CNNPolicy(features, **architecture)
policy.save_model('model.json')
profile = Profile()
# --epochs 5 --minibatch 32 --learning-rate 0.01
arguments = ('model.json', 'debug_feature_planes.hdf5', 'training_results/', 5,
32, .01)
def run_supervised_policy_training():
run_training(*arguments)
profile.runcall(run_supervised_policy_training)
profile.dump_stats('supervised_policy_training_bench_results.prof')
from AlphaGo.training.reinforcement_policy_trainer import run_training
from AlphaGo.models.policy import CNNPolicy
import os
from cProfile import Profile
# make a miniature model for playing on a miniature 7x7 board
architecture = {'filters_per_layer': 32, 'layers': 4, 'board': 7}
features = ['board', 'ones', 'turns_since', 'liberties', 'capture_size',
'self_atari_size', 'liberties_after', 'sensibleness']
policy = CNNPolicy(features, **architecture)
datadir = os.path.join('benchmarks', 'data')
modelfile = os.path.join(datadir, 'mini_rl_model.json')
weights = os.path.join(datadir, 'init_weights.hdf5')
outdir = os.path.join(datadir, 'rl_output')
stats_file = os.path.join(datadir, 'reinforcement_policy_trainer.prof')
if not os.path.exists(datadir):
os.makedirs(datadir)
if not os.path.exists(weights):
policy.model.save_weights(weights)
policy.save_model(modelfile)
profile = Profile()
arguments = (modelfile, weights, outdir, '--learning-rate', '0.001', '--save-every', '2',
'--game-batch', '20', '--iterations', '10', '--verbose')
profile.runcall(run_training, arguments)
profile.dump_stats(stats_file)
from AlphaGo.training.supervised_policy_trainer import run_training
from AlphaGo.models.policy import CNNPolicy
from cProfile import Profile
architecture = {'filters_per_layer': 128, 'layers': 12}
features = ['board', 'ones', 'turns_since']
policy = CNNPolicy(features, **architecture)
policy.save_model('model.json')
profile = Profile()
# --epochs 5 --minibatch 32 --learning-rate 0.01
arguments = ('model.json', 'debug_feature_planes.hdf5', 'training_results/', 5, 32, .01)
def run_supervised_policy_training():
run_training(*arguments)
profile.runcall(run_supervised_policy_training)
profile.dump_stats('supervised_policy_training_bench_results.prof')
