def evaluate():
"""Train ring_net for a number of steps."""
with tf.Graph().as_default():
# make inputs
state, reward, action = ring_net.inputs(1, 15)
# possible input dropout
input_keep_prob = tf.placeholder("float")
state_drop = tf.nn.dropout(state, input_keep_prob)
# possible dropout inside
keep_prob_encoding = tf.placeholder("float")
keep_prob_lstm = tf.placeholder("float")
# unwrap
reward_2_o = []
# first step
x_2, reward_2, hidden_state = ring_net.encode_compress_decode(state[:,0,:,:,:], action[:,1,:], None, keep_prob_encoding, keep_prob_lstm)
tf.get_variable_scope().reuse_variables()
# unroll for 9 more steps
for i in xrange(9):
x_2, reward_2, hidden_state = ring_net.encode_compress_decode(state[:,i+1,:,:,:], action[:,i+2,:], hidden_state, keep_prob_encoding, keep_prob_lstm)
# now collect values
reward_2_o.append(reward_2)
for i in xrange(4):
x_2, reward_2, hidden_state = ring_net.encode_compress_decode(x_2, action[:,i+11,:], hidden_state, keep_prob_encoding, keep_prob_lstm)
reward_2_o.append(reward_2)
reward_2_o = tf.pack(reward_2_o)
reward_2_o = tf.transpose(reward_2_o, perm=[1,0,2])
# restore network
variables_to_restore = tf.all_variables()
saver = tf.train.Saver(variables_to_restore)
sess = tf.Session()
ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
#ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
print("restored file from " + ckpt.model_checkpoint_path)
else:
print("no chekcpoint file found from " + FLAGS.checkpoint_dir + ", this is an error")
# Start que runner
tf.train.start_queue_runners(sess=sess)
reward_g_o, reward_o = sess.run([reward_2_o, reward],feed_dict={keep_prob_encoding:1.0, keep_prob_lstm:1.0, input_keep_prob:1.0})
print(reward_g_o.shape)
print(reward_o.shape)
plt.figure(0)
plt.plot(reward_g_o[0,:,0], label= "predicted reward")
plt.plot(reward_o[0,10:,0], label= "reward")
plt.title("reward vs step")
plt.xlabel("step")
plt.ylabel("reward")
plt.legend()
plt.savefig("paper_reward.png")
def evaluate():
""" Eval the system"""
with tf.Graph().as_default():
# make inputs
state_start, reward_start, action_start, = ring_net.inputs(1, 15)
action_size = int(action_start.get_shape()[2])
action = tf.placeholder(tf.float32, (1, action_size))
# unwrap
x_2_o = []
# first step
x_2, reward_2, hidden_state = ring_net.encode_compress_decode(state_start[:,0,:,:,:], action_start[:,1,:], None, 1.0, 1.0)
tf.get_variable_scope().reuse_variables()
# unroll for 9 more steps
for i in xrange(9):
x_2, reward_2, hidden_state = ring_net.encode_compress_decode(state_start[:,i+1,:,:,:], action_start[:,i+2,:], hidden_state, 1.0, 1.0)
# rename output_t
x_1 = x_2
print(hidden_state)
hidden_state_1 = hidden_state
x_2, reward_2, hidden_state_2 = ring_net.encode_compress_decode(x_1, action, hidden_state_1, 1.0, 1.0)
# restore network
variables_to_restore = tf.all_variables()
saver = tf.train.Saver(variables_to_restore)
sess = tf.Session()
ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
#ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
print("restored file from " + ckpt.model_checkpoint_path)
else:
print("no chekcpoint file found from " + FLAGS.checkpoint_dir + ", this is an error")
# get frame
tf.train.start_queue_runners(sess=sess)
play_action = random_action(action_size)
x_2_g, hidden_2_g = sess.run([x_1, hidden_state_1], feed_dict={})
# Play!!!!
for step in xrange(100):
print(step)
#time.sleep(.5)
# calc generated frame from t
play_action = random_action(action_size)
x_2_g, hidden_2_g = sess.run([x_2, hidden_state_2],feed_dict={x_1:x_2_g, hidden_state_1:hidden_2_g, action:play_action})
frame = np.uint8(np.minimum(np.maximum(0, x_2_g*255.0), 255))
frame = frame[0, :, :, :]
video.write(frame)
#frame = cv2.resize(frame, (500, 500))
#cv2.imshow('frame', frame)
#cv2.waitKey(0)
#if cv2.waitKey(1) & 0xFF == ord('q'):
# break
video.release()
cv2.destroyAllWindows()
def train():
"""Train ring_net for a number of steps."""
with tf.Graph().as_default():
# make inputs
state, reward, action = ring_net.inputs(4, 15)
# possible input dropout
input_keep_prob = tf.placeholder("float")
state_drop = tf.nn.dropout(state, input_keep_prob)
# possible dropout inside
keep_prob_encoding = tf.placeholder("float")
keep_prob_lstm = tf.placeholder("float")
# unwrap
x_2_o = []
# first step
x_2, reward_2, hidden_state = ring_net.encode_compress_decode(state[:,0,:,:,:], action[:,1,:], None, keep_prob_encoding, keep_prob_lstm)
tf.get_variable_scope().reuse_variables()
# unroll for 9 more steps
for i in xrange(9):
x_2, reward_2, hidden_state = ring_net.encode_compress_decode(state[:,i+1,:,:,:], action[:,i+2,:], hidden_state, keep_prob_encoding, keep_prob_lstm)
# now collect values
x_2_o.append(x_2)
for i in xrange(4):
x_2, reward_2, hidden_state = ring_net.encode_compress_decode(x_2, action[:,i+11,:], hidden_state, keep_prob_encoding, keep_prob_lstm)
x_2_o.append(x_2)
tf.image_summary('images_gen_' + str(i), x_2)
x_2_o = tf.pack(x_2_o)
x_2_o = tf.transpose(x_2_o, perm=[1,0,2,3,4])
# error
error = tf.nn.l2_loss(state[:,10:15,:,:,:] - x_2_o)
tf.scalar_summary('loss', error)
# train (hopefuly)
train_op = ring_net.train(error, 1e-5)
# List of all Variables
variables = tf.all_variables()
# Build a saver
saver = tf.train.Saver(tf.all_variables(), max_to_keep=1)
# Summary op
summary_op = tf.merge_all_summaries()
# Start running operations on the Graph.
sess = tf.Session()
# init from seq 1 model
print("init from " + RESTORE_DIR)
saver_restore = tf.train.Saver(variables)
ckpt = tf.train.get_checkpoint_state(RESTORE_DIR)
saver_restore.restore(sess, ckpt.model_checkpoint_path)
# Start que runner
tf.train.start_queue_runners(sess=sess)
# Summary op
graph_def = sess.graph.as_graph_def(add_shapes=True)
summary_writer = tf.train.SummaryWriter(SAVE_DIR, graph_def=graph_def)
for step in xrange(500000):
t = time.time()
_ , loss_value = sess.run([train_op, error],feed_dict={keep_prob_encoding:1.0, keep_prob_lstm:1.0, input_keep_prob:1.0})
elapsed = time.time() - t
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step%100 == 0:
print("loss value at " + str(loss_value))
print("time per batch is " + str(elapsed))
summary_str = sess.run(summary_op, feed_dict={keep_prob_encoding:1.0, keep_prob_lstm:1.0, input_keep_prob:1.0})
summary_writer.add_summary(summary_str, step)
if step%1000 == 0:
checkpoint_path = os.path.join(SAVE_DIR, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
print("saved to " + SAVE_DIR)
def evaluate():
"""Train ring_net for a number of steps."""
with tf.Graph().as_default():
# make inputs
state, reward, action = ring_net.inputs(1, 15)
# possible input dropout
input_keep_prob = tf.placeholder("float")
state_drop = tf.nn.dropout(state, input_keep_prob)
# possible dropout inside
keep_prob_encoding = tf.placeholder("float")
keep_prob_lstm = tf.placeholder("float")
# unwrap
reward_2_o = []
# first step
x_2, reward_2, hidden_state = ring_net.encode_compress_decode(
state[:, 0, :, :, :], action[:, 1, :], None, keep_prob_encoding,
keep_prob_lstm)
tf.get_variable_scope().reuse_variables()
# unroll for 9 more steps
for i in xrange(9):
x_2, reward_2, hidden_state = ring_net.encode_compress_decode(
state[:, i + 1, :, :, :], action[:, i + 2, :], hidden_state,
keep_prob_encoding, keep_prob_lstm)
# now collect values
reward_2_o.append(reward_2)
for i in xrange(4):
x_2, reward_2, hidden_state = ring_net.encode_compress_decode(
x_2, action[:, i + 11, :], hidden_state, keep_prob_encoding,
keep_prob_lstm)
reward_2_o.append(reward_2)
reward_2_o = tf.pack(reward_2_o)
reward_2_o = tf.transpose(reward_2_o, perm=[1, 0, 2])
# restore network
variables_to_restore = tf.all_variables()
saver = tf.train.Saver(variables_to_restore)
sess = tf.Session()
ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
#ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
print("restored file from " + ckpt.model_checkpoint_path)
else:
print("no chekcpoint file found from " + FLAGS.checkpoint_dir +
", this is an error")
# Start que runner
tf.train.start_queue_runners(sess=sess)
reward_g_o, reward_o = sess.run([reward_2_o, reward],
feed_dict={
keep_prob_encoding: 1.0,
keep_prob_lstm: 1.0,
input_keep_prob: 1.0
})
print(reward_g_o.shape)
print(reward_o.shape)
plt.figure(0)
plt.plot(reward_g_o[0, :, 0], label="predicted reward")
plt.plot(reward_o[0, 10:, 0], label="reward")
plt.title("reward vs step")
plt.xlabel("step")
plt.ylabel("reward")
plt.legend()
plt.savefig("paper_reward.png")
def train():
"""Train ring_net for a number of steps."""
with tf.Graph().as_default():
# make inputs
state, reward, action = ring_net.inputs(4, 15)
# possible input dropout
input_keep_prob = tf.placeholder("float")
state_drop = tf.nn.dropout(state, input_keep_prob)
# possible dropout inside
keep_prob_encoding = tf.placeholder("float")
keep_prob_lstm = tf.placeholder("float")
# unwrap
x_2_o = []
# first step
x_2, reward_2, hidden_state = ring_net.encode_compress_decode(
state[:, 0, :, :, :], action[:, 1, :], None, keep_prob_encoding,
keep_prob_lstm)
tf.get_variable_scope().reuse_variables()
# unroll for 9 more steps
for i in xrange(9):
x_2, reward_2, hidden_state = ring_net.encode_compress_decode(
state[:, i + 1, :, :, :], action[:, i + 2, :], hidden_state,
keep_prob_encoding, keep_prob_lstm)
# now collect values
x_2_o.append(x_2)
for i in xrange(4):
x_2, reward_2, hidden_state = ring_net.encode_compress_decode(
x_2, action[:, i + 11, :], hidden_state, keep_prob_encoding,
keep_prob_lstm)
x_2_o.append(x_2)
tf.image_summary('images_gen_' + str(i), x_2)
x_2_o = tf.pack(x_2_o)
x_2_o = tf.transpose(x_2_o, perm=[1, 0, 2, 3, 4])
# error
error = tf.nn.l2_loss(state[:, 10:15, :, :, :] - x_2_o)
tf.scalar_summary('loss', error)
# train (hopefuly)
train_op = ring_net.train(error, 1e-5)
# List of all Variables
variables = tf.all_variables()
# Build a saver
saver = tf.train.Saver(tf.all_variables(), max_to_keep=1)
# Summary op
summary_op = tf.merge_all_summaries()
# Start running operations on the Graph.
sess = tf.Session()
# init from seq 1 model
print("init from " + RESTORE_DIR)
saver_restore = tf.train.Saver(variables)
ckpt = tf.train.get_checkpoint_state(RESTORE_DIR)
saver_restore.restore(sess, ckpt.model_checkpoint_path)
# Start que runner
tf.train.start_queue_runners(sess=sess)
# Summary op
graph_def = sess.graph.as_graph_def(add_shapes=True)
summary_writer = tf.train.SummaryWriter(SAVE_DIR, graph_def=graph_def)
for step in xrange(500000):
t = time.time()
_, loss_value = sess.run([train_op, error],
feed_dict={
keep_prob_encoding: 1.0,
keep_prob_lstm: 1.0,
input_keep_prob: 1.0
})
elapsed = time.time() - t
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step % 100 == 0:
print("loss value at " + str(loss_value))
print("time per batch is " + str(elapsed))
summary_str = sess.run(summary_op,
feed_dict={
keep_prob_encoding: 1.0,
keep_prob_lstm: 1.0,
input_keep_prob: 1.0
})
summary_writer.add_summary(summary_str, step)
if step % 1000 == 0:
checkpoint_path = os.path.join(SAVE_DIR, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
print("saved to " + SAVE_DIR)
def evaluate():
""" Eval the system"""
with tf.Graph().as_default():
# make inputs
state_start, reward_start, action_start, = ring_net.inputs(1, 15)
action_size = int(action_start.get_shape()[2])
action = tf.placeholder(tf.float32, (1, action_size))
# unwrap
x_2_o = []
# first step
x_2, reward_2, hidden_state = ring_net.encode_compress_decode(
state_start[:, 0, :, :, :], action_start[:, 1, :], None, 1.0, 1.0)
tf.get_variable_scope().reuse_variables()
# unroll for 9 more steps
for i in xrange(9):
x_2, reward_2, hidden_state = ring_net.encode_compress_decode(
state_start[:, i + 1, :, :, :], action_start[:, i + 2, :],
hidden_state, 1.0, 1.0)
# rename output_t
x_1 = x_2
print(hidden_state)
hidden_state_1 = hidden_state
x_2, reward_2, hidden_state_2 = ring_net.encode_compress_decode(
x_1, action, hidden_state_1, 1.0, 1.0)
# restore network
variables_to_restore = tf.all_variables()
saver = tf.train.Saver(variables_to_restore)
sess = tf.Session()
ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
#ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
print("restored file from " + ckpt.model_checkpoint_path)
else:
print("no chekcpoint file found from " + FLAGS.checkpoint_dir +
", this is an error")
# get frame
tf.train.start_queue_runners(sess=sess)
play_action = random_action(action_size)
x_2_g, hidden_2_g = sess.run([x_1, hidden_state_1], feed_dict={})
# Play!!!!
for step in xrange(100):
print(step)
#time.sleep(.5)
# calc generated frame from t
play_action = random_action(action_size)
x_2_g, hidden_2_g = sess.run([x_2, hidden_state_2],
feed_dict={
x_1: x_2_g,
hidden_state_1: hidden_2_g,
action: play_action
})
frame = np.uint8(np.minimum(np.maximum(0, x_2_g * 255.0), 255))
frame = frame[0, :, :, :]
video.write(frame)
#frame = cv2.resize(frame, (500, 500))
#cv2.imshow('frame', frame)
#cv2.waitKey(0)
#if cv2.waitKey(1) & 0xFF == ord('q'):
# break
video.release()
cv2.destroyAllWindows()