def main():
config = tf.ConfigProto()
# config.gpu_options.allow_growth = True
# config.log_device_placement = True
with tf.Session(config=config) as sess:
agent = DeepDeterministicPolicyGradientAgent(env=env)
# setup saver util and either load latest ckpt or init variables
saver_util = None
if opts.ckpt_dir is not None:
saver_util = util.SaverUtil(sess, opts.ckpt_dir, opts.ckpt_freq)
else:
sess.run(tf.global_variables_initializer())
for v in tf.global_variables():
print(v.name, util.shape_and_product_of(v), file=sys.stderr)
# print >>sys.stderr, v.name, util.shape_and_product_of(v)
# now that we've either init'd from scratch, or loaded up a checkpoint,
# we can do any required post init work.
agent.post_var_init_setup()
#opts.num_eval = 100
# run either eval or training
if opts.num_eval > 0:
agent.run_eval(opts.num_eval, opts.eval_action_noise)
else:
agent.run_training(opts.max_num_actions, opts.max_run_time,
opts.batch_size, opts.batches_per_step,
saver_util)
if saver_util is not None:
saver_util.force_save()
env.reset() # just to flush logging, clumsy :/
def conv_net_on(self, input_layer, opts):
# TODO: reinclude batch_norm config
# if opts.use_batch_norm:
# normalizer_fn = slim.batch_norm
# normalizer_params = { 'is_training': IS_TRAINING }
# else:
normalizer_fn = None
normalizer_params = None
# whiten image, per channel, using batch_normalisation layer with
# params calculated directly from batch.
axis = list(range(input_layer.get_shape().ndims - 1))
batch_mean, batch_var = tf.nn.moments(input_layer, axis) # calcs moments per channel
whitened_input_layer = tf.nn.batch_normalization(input_layer, batch_mean, batch_var,
scale=None, offset=None,
variance_epsilon=1e-6)
# TODO: num_outputs here are really dependant on the incoming channels,
# which depend on the #repeats & cameras so they should be a param.
model = slim.conv2d(whitened_input_layer, num_outputs=32, kernel_size=[7, 7],
normalizer_fn=normalizer_fn,
normalizer_params=normalizer_params,
scope='conv1')
model = slim.max_pool2d(model, kernel_size=[2, 2], scope='pool1')
self.pool1 = model
print >>sys.stderr, "pool1", util.shape_and_product_of(model)
model = slim.conv2d(model, num_outputs=32, kernel_size=[5, 5],
normalizer_fn=normalizer_fn,
normalizer_params=normalizer_params,
scope='conv2')
model = slim.max_pool2d(model, kernel_size=[2, 2], scope='pool2')
self.pool2 = model
print >>sys.stderr, "pool2", util.shape_and_product_of(model)
model = slim.conv2d(model, num_outputs=16, kernel_size=[3, 3],
normalizer_fn=normalizer_fn,
normalizer_params=normalizer_params,
scope='conv3')
model = slim.max_pool2d(model, kernel_size=[2, 2], scope='pool2')
self.pool3 = model
print >>sys.stderr, "pool3", util.shape_and_product_of(model)
return slim.flatten(model, scope='flat')
def __init__(self, opts):
self.opts = opts
config = tf.ConfigProto()
#config.gpu_options.allow_growth = True
#config.log_device_placement = True
config.gpu_options.per_process_gpu_memory_fraction = 0.5 #opts.gpu_mem_fraction
self.sess = tf.Session(config=config)
render_shape = (opts.height, opts.width, 3)
self.replay_memory = replay_memory.ReplayMemory(
opts=opts, state_shape=render_shape, action_dim=2, load_factor=1.2)
if opts.event_log_in:
self.replay_memory.reset_from_event_log(opts.event_log_in,
opts.event_log_in_num)
# s1 and s2 placeholders
batched_state_shape = [None] + list(render_shape)
s1 = tf.placeholder(shape=batched_state_shape, dtype=tf.float32)
s2 = tf.placeholder(shape=batched_state_shape, dtype=tf.float32)
# initialise base models for value & naf networks. value subportion of net is
# explicitly created seperate because it has a target network note: in the case of
# --share-input-state-representation the input state network of the value_net will
# be reused by the naf.l_value and naf.output_actions net
self.value_net = models.ValueNetwork("value", s1, opts)
self.target_value_net = models.ValueNetwork("target_value", s2, opts)
self.network = models.NafNetwork("naf",
s1,
s2,
self.value_net,
self.target_value_net,
action_dim=2,
opts=opts)
with self.sess.as_default():
# setup saver util and either load latest ckpt or init variables
self.saver_util = None
if opts.ckpt_dir is not None:
self.saver_util = util.SaverUtil(self.sess, opts.ckpt_dir,
opts.ckpt_freq)
else:
self.sess.run(tf.initialize_all_variables())
for v in tf.all_variables():
print >> sys.stderr, v.name, util.shape_and_product_of(v)
# setup target network
self.target_value_net.set_as_target_network_for(
self.value_net, 0.01)
def __init__(self, opts):
# self.opts = opts
self.network = models.NafNetwork("naf", action_dim=2, opts=opts)
config = tf.ConfigProto()
#config.gpu_options.allow_growth = True
#config.log_device_placement = True
config.gpu_options.per_process_gpu_memory_fraction = opts.gpu_mem_fraction
self.sess = tf.Session(config=config)
with self.sess.as_default():
# setup saver to load first set of ckpts. block until some are available
self.loader = ckpt_util.AgentCkptLoader(self.sess, opts.ckpt_dir)
self.loader.blocking_load_ckpt()
# dump info on vars
for v in tf.all_variables():
if '/biases:' not in v.name:
print >> sys.stderr, v.name, util.shape_and_product_of(v)
def simple_conv_net_on(self, input_layer, opts):
if opts.use_batch_norm:
normalizer_fn = slim.batch_norm
normalizer_params = {'is_training': IS_TRAINING}
else:
normalizer_fn = None
normalizer_params = None
# optionally drop blue channel, in a simple cart pole env we only need r/g
#if opts.drop_blue_channel:
# input_layer = input_layer[:,:,:,0:2,:,:]
# state is (batch, height, width, rgb, camera_idx, repeat)
# rollup rgb, camera_idx and repeat into num_channels
# i.e. (batch, height, width, rgb*camera_idx*repeat)
height, width = map(int, input_layer.get_shape()[1:3])
num_channels = input_layer.get_shape()[3:].num_elements()
input_layer = tf.reshape(input_layer,
[-1, height, width, num_channels])
print("input_layer",
util.shape_and_product_of(input_layer),
file=sys.stderr)
# whiten image, per channel, using batch_normalisation layer with
# params calculated directly from batch.
axis = list(range(input_layer.get_shape().ndims - 1))
batch_mean, batch_var = tf.nn.moments(
input_layer, axis) # gives moments per channel
whitened_input_layer = tf.nn.batch_normalization(input_layer,
batch_mean,
batch_var,
scale=None,
offset=None,
variance_epsilon=1e-6)
# TODO: num_outputs here are really dependant on the incoming channels,
# which depend on the #repeats & cameras so they should be a param.
model = slim.conv2d(whitened_input_layer,
num_outputs=10,
kernel_size=[5, 5],
normalizer_fn=normalizer_fn,
normalizer_params=normalizer_params,
scope='conv1')
model = slim.max_pool2d(model, kernel_size=[2, 2], scope='pool1')
self.pool1 = model
print("pool1", util.shape_and_product_of(model), file=sys.stderr)
model = slim.conv2d(model,
num_outputs=10,
kernel_size=[5, 5],
normalizer_fn=normalizer_fn,
normalizer_params=normalizer_params,
scope='conv2')
model = slim.max_pool2d(model, kernel_size=[2, 2], scope='pool2')
self.pool2 = model
print("pool2", util.shape_and_product_of(model), file=sys.stderr)
model = slim.conv2d(model,
num_outputs=10,
kernel_size=[3, 3],
normalizer_fn=normalizer_fn,
normalizer_params=normalizer_params,
scope='conv3')
model = slim.max_pool2d(model, kernel_size=[2, 2], scope='pool2')
self.pool3 = model
print("pool3", util.shape_and_product_of(model), file=sys.stderr)
return model
def conv_net_on(self, input_layer, opts):
# TODO: reinclude batch_norm config, hasn't been helping at all...
# convert input_layer from uint8 (0, 255) to float32 (0.0, 1.0)
input_layer = tf.to_float(input_layer) / 255
# whiten image, per channel, using batch_normalisation layer with
# params calculated directly from batch.
axis = list(range(input_layer.get_shape().ndims - 1))
batch_mean, batch_var = tf.nn.moments(
input_layer, axis) # calcs moments per channel
whitened_input_layer = tf.nn.batch_normalization(input_layer,
batch_mean,
batch_var,
scale=None,
offset=None,
variance_epsilon=1e-6)
model = slim.conv2d(whitened_input_layer,
num_outputs=8,
kernel_size=[5, 5],
scope='conv1a')
# model = slim.conv2d(whitened_input_layer, num_outputs=8, kernel_size=[5, 5], scope='conv1b')
model = slim.max_pool2d(model, kernel_size=[2, 2], scope='pool1')
self.pool1 = model
print >> sys.stderr, "pool1", util.shape_and_product_of(model)
model = slim.conv2d(model,
num_outputs=16,
kernel_size=[5, 5],
scope='conv2a')
# model = slim.conv2d(model, num_outputs=16, kernel_size=[5, 5], scope='conv2b')
model = slim.max_pool2d(model, kernel_size=[2, 2], scope='pool2')
self.pool2 = model
print >> sys.stderr, "pool2", util.shape_and_product_of(model)
model = slim.conv2d(model,
num_outputs=32,
kernel_size=[3, 3],
scope='conv3a')
# model = slim.conv2d(model, num_outputs=32, kernel_size=[3, 3], scope='conv3b')
model = slim.max_pool2d(model, kernel_size=[2, 2], scope='pool3')
self.pool3 = model
print >> sys.stderr, "pool3", util.shape_and_product_of(model)
# a final unpooled conv net just to drop params down. maybe pool here too actually?
# model = slim.conv2d(model, num_outputs=32, kernel_size=[3, 3], scope='conv4a')
# model = slim.conv2d(model, num_outputs=32, kernel_size=[3, 3], scope='conv3b')
# model = slim.max_pool2d(model, kernel_size=[2, 2], scope='pool4')
# self.pool3 = model
# print >>sys.stderr, "pool4", util.shape_and_product_of(model)
# do simple maxout on output to reduce dimensionality down for the upcoming
# fully connected layers. see https://arxiv.org/abs/1302.4389
# model = tf.reshape(model, (-1, 15, 20, 8, 4)) # (?, 15, 20, 32) -> (?, 15, 20, 8, 4)
# model = tf.reduce_max(model, reduction_indices=4) # (?, 15, 20, 8)
# print >>sys.stderr, "maxout", util.shape_and_product_of(model)
model = slim.flatten(model, scope='flat')
if opts.use_dropout:
model = slim.dropout(model,
is_training=IS_TRAINING,
scope="drop" % i)
return model
def run_trainer(episodes, opts):
# init replay memory
render_shape = (opts.height, opts.width, 3)
replay_memory = rm.ReplayMemory(opts=opts,
state_shape=render_shape,
action_dim=2,
load_factor=1.1)
if opts.event_log_in:
replay_memory.reset_from_event_logs(opts.event_log_in,
opts.event_log_in_num,
opts.reset_smooth_reward_factor)
# init network for training
config = tf.ConfigProto()
#config.gpu_options.allow_growth = True
#config.log_device_placement = True
config.gpu_options.per_process_gpu_memory_fraction = opts.gpu_mem_fraction
sess = tf.Session(config=config)
network = models.NafNetwork("naf", action_dim=2, opts=opts)
with sess.as_default():
# setup saver util and either load saved ckpt or init variables
saver = ckpt_util.TrainerCkptSaver(sess, opts.ckpt_dir,
opts.ckpt_save_freq)
for v in tf.all_variables():
if '/biases:' not in v.name:
print >> sys.stderr, v.name, util.shape_and_product_of(v)
network.setup_target_network()
# while true process episodes from run_agents
print util.dts(), "waiting for episodes"
while True:
start_time = time.time()
episode = episodes.get()
wait_time = time.time() - start_time
start_time = time.time()
replay_memory.add_episode(
episode, smooth_reward_factor=opts.smooth_reward_factor)
losses = []
if replay_memory.burnt_in():
for _ in xrange(opts.batches_per_new_episode):
batch = replay_memory.batch(opts.batch_size)
batch_losses = network.train(batch).T[
0] # .T[0] => (B, 1) -> (B,)
replay_memory.update_priorities(batch.idxs, batch_losses)
network.target_value_net.update_target_weights()
losses.extend(batch_losses)
saver.save_if_required()
process_time = time.time() - start_time
stats = {
"wait_time": wait_time,
"process_time": process_time,
"pending": episodes.qsize(),
"replay_memory": replay_memory.stats
}
if losses:
stats['loss'] = {
"min": float(np.min(losses)),
"median": float(np.median(losses)),
"mean": float(np.mean(losses)),
"max": float(np.max(losses))
}
print "STATS\t%s\t%s" % (util.dts(), json.dumps(stats))
