def test(self):
#self.merge()
#self.compress()
#return
embedding_size = 100
for CLUSTER_MIN_SIZE in range(4,19,2):
for dsname in ['webkb','er']:
mln = MLN(dsname)
db = DBManager(dsname,mln)
print('merge db dom sizes:')
dom_obj_map = db.get_dom_objs_map(mln,db.merge_db_file)
cf = common_f()
#cf.delete_files(mln.pickle_location)
#cf.remove_irrelevant_atoms()
embedding_size += 100
embedding_size = embedding_size%1000
db.set_atoms()
bmf = bmf_cluster(dsname)
bmf.cluster(db,1,mln.pdm,dom_obj_map)
print('original db dom sizes(after compression):')
orig_dom_objs_map = db.get_dom_objs_map(mln,mln.orig_db_file)
CLUSTER_MIN_SIZE = 10
w2v = word2vec(dsname,db,CLUSTER_MIN_SIZE,embedding_size)
print('w2v cluster dom sizes:')
w2v_dom_objs_map = db.get_dom_objs_map(mln,w2v.w2v__cluster_db_file)
cr = cf.calculate_cr(orig_dom_objs_map,w2v_dom_objs_map)
print('cr : ' + str(cr))
rc = random_cluster(dsname)
rc.generate_random_db(db,w2v.pred_atoms_reduced_numbers,mln,w2v_dom_objs_map)
print('random cluster dom sizes')
db.get_dom_objs_map(mln,mln.random__cluster_db_file)
kmc = kmeans_cluster(dsname)
kmc.cluster(db,str(cr),mln.pdm,w2v_dom_objs_map,mln.dom_pred_map)
print('kmeans cluster dom sizes:')
kmeans_dom_objs_map = db.get_dom_objs_map(mln,kmc.kmeans__cluster_db_file)
mln.create_magician_mln()
#magician(dsname,mln)
tuffy(dsname)
orig_meta_map = {}
orig_meta_map['bmf'] = bmf.bmf_orig_meta_map
orig_meta_map['w2v'] = w2v.w2v_orig_meta_map
orig_meta_map['random'] = rc.rand_orig_meta_map
orig_meta_map['kmeans'] = kmc.kmeans_orig_meta_map
print('Dataset : ' + dsname + '; CR : ' + str(cr))
p = performance(dsname,embedding_size)
p.compare_marginal(mln,orig_meta_map,cr)
p.compare_map(mln,orig_meta_map,cr)
break
def __init__(self, dsname):
MLN.__init__(self, dsname)
self.set_pred_pairs()
def test(self):
#self.merge()
#self.compress()
#return
embedding_size = 100
for CLUSTER_MIN_SIZE in range(4, 19, 2):
for dsname in ['webkb']:
mln = MLN(dsname)
db = DBManager(dsname, mln)
print('merge db dom sizes:')
db.set_doms_atoms(mln, db.merge_db_file)
cf = common_f()
#cf.delete_files(mln.pickle_location)
if dsname == 'er':
cf.remove_irrelevant_atoms()
embedding_size = 300
print('generating sentences')
start = time.time()
cnn_atoms, ntn_atoms = db.pred_atoms, db.pred_atoms
while True:
#cnn_atoms = self.embed(cnn_atoms,mln.pdm,mln.dom_sizes_map,True)
ntn_atoms = self.embed(ntn_atoms, mln.pdm,
mln.dom_sizes_map, False)
sg = None
if dsname == 'review':
return
#end = time.time()
#print('Time : ',end-start)
else:
sg = sentence_generator(mln.pdm, db.pred_atoms,
db.TEST_SIZE, db)
#print('calling w2v')
#wv = word2vec_cnn()
#print('making images')
#wv.make_images(sg.sentences,mln.pdm,db.pred_atoms,mln.dom_sizes_map,dsname,sg.train_atoms,sg.test_atoms,db.TEST_SIZE)
cor = corrupt(dsname, db.pred_atoms, mln.pdm, db.dom_objs_map,
sg.sentences)
return
bmf = bmf_cluster(dsname)
bmf.cluster(db, 1, mln.pdm, dom_obj_map)
print('original db dom sizes(after compression):')
orig_dom_objs_map = db.get_dom_objs_map(mln, mln.orig_db_file)
CLUSTER_MIN_SIZE = 10
w2v = word2vec(dsname, db, CLUSTER_MIN_SIZE, embedding_size)
print('w2v cluster dom sizes:')
w2v_dom_objs_map = db.get_dom_objs_map(
mln, w2v.w2v__cluster_db_file)
cr = cf.calculate_cr(orig_dom_objs_map, w2v_dom_objs_map)
print('cr : ' + str(cr))
rc = random_cluster(dsname)
rc.generate_random_db(db, w2v.pred_atoms_reduced_numbers, mln,
w2v_dom_objs_map)
print('random cluster dom sizes')
db.get_dom_objs_map(mln, mln.random__cluster_db_file)
kmc = kmeans_cluster(dsname)
kmc.cluster(db, str(cr), mln.pdm, w2v_dom_objs_map,
mln.dom_pred_map)
print('kmeans cluster dom sizes:')
kmeans_dom_objs_map = db.get_dom_objs_map(
mln, kmc.kmeans__cluster_db_file)
mln.create_magician_mln()
magician(dsname, mln)
#tuffy(dsname)
orig_meta_map = {}
orig_meta_map['bmf'] = bmf.bmf_orig_meta_map
orig_meta_map['w2v'] = w2v.w2v_orig_meta_map
orig_meta_map['random'] = rc.rand_orig_meta_map
orig_meta_map['kmeans'] = kmc.kmeans_orig_meta_map
print('Dataset : ' + dsname + '; CR : ' + str(cr))
p = performance(dsname, embedding_size)
p.compare_marginal(mln, orig_meta_map, cr)
def compress(self):
for dsname in ['er', 'protein', 'webkb']:
mln = MLN(dsname)
db = DBManager(dsname, mln)
db.compress(mln, .3)
def merge(self):
for dsname in ['protein']:
mln = MLN(dsname)
db = DBManager(dsname, mln)
db.merge()
from DBManager import DBManager,dbconfig from mln import MLN from word2vec import word2vec dsname = 'er' mln = MLN(dsname) db = DBManager(dsname,mln) #db.merge() db.compress(mln,.1) #db.calculate_dom_sizes(mln) #w2v = word2vec(dsname,db)
def train():
rnd_seed = 0
np.random.seed(rnd_seed)
tf.set_random_seed(rnd_seed)
sim = GymSim('CartPole-v0', 5000, seed=rnd_seed)
sim.act_sample_batch(
5000, FLAGS.sample_neg_ratio) # bootstrap with random actions
sim.print_stats()
#embed()
#sys.exit()
q_network = MLN(sim.INPUT_DIM, sim.ACTION_DIM)
target_network = MLN(sim.INPUT_DIM, sim.ACTION_DIM, name_scope='target')
with tf.Graph().as_default():
global_step = tf.Variable(0, trainable=False)
action_pl = tf.placeholder(tf.int64, name='action_pl')
reward_pl = tf.placeholder(tf.float32, name='reward_pl')
state_pl = tf.placeholder(tf.float32, (None, sim.INPUT_DIM),
name='state_pl')
observ_pl = tf.placeholder(tf.float32, (None, sim.INPUT_DIM),
name='observ_pl')
action_q = q_network.inference(state_pl)
target_q = tf.stop_gradient(target_network.inference(observ_pl))
target_q_pt = tf.Print(target_q, [target_q])
action_q_pt = tf.Print(action_q, [action_q])
loss = dqn.td_loss(action_pl, sim.ACTION_DIM, action_q, reward_pl,
target_q)
train_op = dqn.train(FLAGS.learning_rate, loss, global_step)
saver = tf.train.Saver(tf.all_variables())
summary_op = tf.merge_all_summaries()
action_op = tf.argmax(action_q, 1, name='action_op')
copy_var = q_network.copy_to('target')
init = tf.initialize_all_variables()
sess = tf.Session(config=tf.ConfigProto(
log_device_placement=FLAGS.log_device_placement))
#initialize variables
sess.run(init)
summary_writer = tf.train.SummaryWriter(
os.path.join(FLAGS.train_dir, 'logs'), sess.graph)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
if step % 4 == 0:
sess.run(copy_var)
feed = sim.feed_batch(state_pl, action_pl, reward_pl, observ_pl,
FLAGS.batch_size)
_, loss_value = sess.run([train_op, loss], feed_dict=feed)
duration = time.time() - start_time
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step % 10 == 0:
num_examples_per_step = FLAGS.batch_size
examples_per_sec = num_examples_per_step / duration
sec_per_batch = float(duration)
format_str = ('%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str % (datetime.now(), step, loss_value,
examples_per_sec, sec_per_batch))
if step > FLAGS.sample_after:
pred_act = sess.run(action_op, feed_dict={state_pl: sim.state})
pred_act = pred_act[0]
sim.act_sample_once(pred_act,
neg_ratio=FLAGS.sample_neg_ratio,
append_db=True)
# visualization
if step % 1000 == 0 and step != 0:
sim.reset()
survive = 0
for _ in range(200):
pred_act = sess.run(action_op, feed_dict={state_pl: sim.state})
pred_act = pred_act[0]
done = sim.act_demo(pred_act)
if not done:
survive += 1
else:
print('Survived for %i frame' % survive)
survive = 0
#if step % 100 == 0:
# summary_str = sess.run(summary_op)
# summary_writer.add_summary(summary_str, step)
# Save the model checkpoint periodically.
if step % 1000 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
def train():
rnd_seed = 0
np.random.seed(rnd_seed)
tf.set_random_seed(rnd_seed)
sim = GymSim('CartPole-v0', 5000, seed=rnd_seed)
sim.act_sample_batch(5000, FLAGS.sample_neg_ratio) # bootstrap with random actions
sim.print_stats()
#embed()
#sys.exit()
q_network = MLN(sim.INPUT_DIM, sim.ACTION_DIM)
target_network = MLN(sim.INPUT_DIM, sim.ACTION_DIM, name_scope='target')
with tf.Graph().as_default():
global_step = tf.Variable(0, trainable=False)
action_pl = tf.placeholder(tf.int64, name='action_pl')
reward_pl = tf.placeholder(tf.float32, name='reward_pl')
state_pl = tf.placeholder(tf.float32, (None, sim.INPUT_DIM), name='state_pl')
observ_pl = tf.placeholder(tf.float32, (None, sim.INPUT_DIM), name='observ_pl')
action_q = q_network.inference(state_pl)
target_q = tf.stop_gradient(target_network.inference(observ_pl))
target_q_pt = tf.Print(target_q, [target_q])
action_q_pt = tf.Print(action_q, [action_q])
loss = dqn.td_loss(action_pl, sim.ACTION_DIM, action_q, reward_pl, target_q)
train_op = dqn.train(FLAGS.learning_rate, loss, global_step)
saver = tf.train.Saver(tf.all_variables())
summary_op = tf.merge_all_summaries()
action_op = tf.argmax(action_q, 1, name='action_op')
copy_var = q_network.copy_to('target')
init = tf.initialize_all_variables()
sess = tf.Session(config=tf.ConfigProto(
log_device_placement=FLAGS.log_device_placement))
#initialize variables
sess.run(init)
summary_writer = tf.train.SummaryWriter(os.path.join(FLAGS.train_dir, 'logs')
, sess.graph)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
if step % 4 == 0:
sess.run(copy_var)
feed = sim.feed_batch(state_pl,
action_pl,
reward_pl,
observ_pl,
FLAGS.batch_size)
_, loss_value = sess.run([train_op, loss],
feed_dict = feed)
duration = time.time() - start_time
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step % 10 == 0:
num_examples_per_step = FLAGS.batch_size
examples_per_sec = num_examples_per_step / duration
sec_per_batch = float(duration)
format_str = ('%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print (format_str % (datetime.now(), step, loss_value,
examples_per_sec, sec_per_batch))
if step > FLAGS.sample_after:
pred_act = sess.run(action_op,
feed_dict={state_pl: sim.state})
pred_act = pred_act[0]
sim.act_sample_once(pred_act, neg_ratio=FLAGS.sample_neg_ratio,
append_db=True)
# visualization
if step % 1000 == 0 and step != 0:
sim.reset()
survive = 0
for _ in range(200):
pred_act = sess.run(action_op,
feed_dict={state_pl: sim.state})
pred_act = pred_act[0]
done = sim.act_demo(pred_act)
if not done:
survive += 1
else:
print('Survived for %i frame' % survive)
survive = 0
#if step % 100 == 0:
# summary_str = sess.run(summary_op)
# summary_writer.add_summary(summary_str, step)
# Save the model checkpoint periodically.
if step % 1000 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)