def create_placeholders(raw_graphs):
"""Creates placeholders for the model training and evaluation.
Args:
rand: A random seed (np.RandomState instance).
batch_size: Total number of graphs per batch.
min_max_nodes: A 2-tuple with the [lower, upper) number of nodes per
graph. The number of nodes for a graph is uniformly sampled within this
range.
geo_density: A `float` threshold parameters for the geographic threshold graph's
threshold. Default= the number of nodes.
Returns:
source_ph: The source graph's placeholders, as a graph namedtuple.
target_ph: The target graph's placeholders, as a graph namedtuple.
"""
# Create some example data for inspecting the vector sizes.
source_graphs = [source_from_raw(raw) for raw in raw_graphs]
source_ph = utils_tf.placeholders_from_networkxs(
source_graphs, force_dynamic_num_graphs=True)
target_graphs = [target_from_raw(raw) for raw in raw_graphs]
target_ph = utils_tf.placeholders_from_networkxs(
target_graphs, force_dynamic_num_graphs=True)
return source_ph, target_ph
def create_placeholders(rand, batch_size, num_nodes_min_max, theta):
"""Creates placeholders for the model training and evaluation.
Args:
rand: A random seed (np.RandomState instance).
batch_size: Total number of graphs per batch.
num_nodes_min_max: A 2-tuple with the [lower, upper) number of nodes per
graph. The number of nodes for a graph is uniformly sampled within this
range.
theta: A `float` threshold parameters for the geographic threshold graph's
threshold. Default= the number of nodes.
Returns:
input_ph: The input graph's placeholders, as a graph namedtuple.
target_ph: The target graph's placeholders, as a graph namedtuple.
"""
#Create some example data for inspecting the vector sizes.
input_graphs, target_graphs, _ = generate_networkx_graphs(
rand, batch_size, num_nodes_min_max, theta)
# input_graphs, target_graphs, _ = generate_surface_graphs(
# rand,
# batch_size,
# num_nodes_min_max,
# theta
# )
input_ph = utils_tf.placeholders_from_networkxs(input_graphs)
target_ph = utils_tf.placeholders_from_networkxs(target_graphs)
return input_ph, target_ph
def _create_placeholders(self, loader):
'''
Creates input and output placeholders for model. Loader is used to get the shapes of the data
:param loader: GraphTypleDataLoader instance. Used only to get the shapes of the graphs.
'''
# Create some example data for inspecting the vector sizes.
input_graphs, target_graphs = loader.next(raw_graphs=True)
self.input_ph = utils_tf.placeholders_from_networkxs(input_graphs)
self.target_ph = utils_tf.placeholders_from_networkxs(target_graphs)
def test_graph_network(tf_session):
logdir = os.path.join(TEST_FOLDER, 'test_logdir/with_trace')
os.makedirs(logdir, exist_ok=True)
starcluster_graphs_nx, gaia_graphs_nx = make_example_graphs(
2,
num_stars=10,
sc_edge_size=2,
sc_node_size=12,
sc_global_size=5,
g_edge_size=3,
g_node_size=13,
g_global_size=6)
with tf_session.graph.as_default():
sc_pl = placeholders_from_networkxs(starcluster_graphs_nx,
force_dynamic_num_graphs=False)
g_pl = placeholders_from_networkxs(gaia_graphs_nx,
force_dynamic_num_graphs=False)
sc_graphtuple = networkxs_to_graphs_tuple(starcluster_graphs_nx)
g_graphtuple = networkxs_to_graphs_tuple(gaia_graphs_nx)
encoded_size = 7
t_network = StarClusterTNetwork(encoded_size,
sc_encoder_latent_size=16,
sc_encoder_num_layers=2,
sc_decoder_latent_size=16,
sc_decoder_num_layers=2,
g_encoder_latent_size=16,
g_encoder_num_layers=2)
t_network_output = t_network(g_pl,
sc_pl,
num_samples=2,
num_processing_steps=1)
summary = tf.summary.merge_all()
writer = tf.compat.v1.summary.FileWriter(logdir,
tf_session.graph,
session=tf_session)
tf_session.run(tf.global_variables_initializer())
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
t_network_output_res, summary_eval = tf_session.run(
[t_network_output, summary],
feed_dict={
sc_pl: sc_graphtuple,
g_pl: g_graphtuple
},
options=run_options,
run_metadata=run_metadata)
writer.add_run_metadata(run_metadata, 'step%d' % 0)
writer.add_summary(summary_eval, 0)
def create_placeholders(rand, batch_size, raw_input_graphs, raw_target_graphs,
edge_permutations):
"""Creates placeholders for the model training and evaluation."""
# Create some example data for inspecting the vector sizes.
input_graphs, target_graphs = generate_networkx_graphs(
rand, batch_size, raw_input_graphs, raw_target_graphs,
edge_permutations, np.arange(2), True)
input_ph = utils_tf.placeholders_from_networkxs(input_graphs)
target_ph = utils_tf.placeholders_from_networkxs(target_graphs)
weight_ph = tf.zeros_like(target_ph.edges[:, :1])
return input_ph, target_ph, weight_ph
def create_placeholders(input_graphs, target_graphs):
"""
Creates placeholders for the model training and evaluation.
Returns:
input_ph: The input graph's placeholders, as a graph namedtuple.
target_ph: The target graph's placeholders, as a graph namedtuple.
"""
input_ph = utils_tf.placeholders_from_networkxs(
input_graphs, name="input_placeholders_from_networksx")
target_ph = utils_tf.placeholders_from_networkxs(
target_graphs, name="target_placeholders_from_networkxs")
return input_ph, target_ph
def create_placeholders(graphcache,rand,batch_size,all_db):
"""Creates placeholders for the model training and evaluation.
Args:
rand: A random seed (np.RandomState instance).
batch_size: Total number of graphs per batch.
Returns:
input_ph: The input graph's placeholders, as a graph namedtuple.
target_ph: The target graph's placeholders, as a graph namedtuple.
"""
# Create some example data for inspecting the vector sizes.
input_graphs, target_graphs, _, _ = generate_networkx_graphs(graphcache,all_db, rand, batch_size,'train')
input_ph = utils_tf.placeholders_from_networkxs(input_graphs)
target_ph = utils_tf.placeholders_from_networkxs(target_graphs)
return input_ph, target_ph
def test_feed_data(self):
networkx = [_generate_graph(batch_index) for batch_index in range(16)]
placeholders = utils_tf.placeholders_from_networkxs(
networkx, force_dynamic_num_graphs=True)
# Does not need to be the same size
networkxs = [_generate_graph(batch_index) for batch_index in range(2)]
with self.test_session() as sess:
output = sess.run(
placeholders,
utils_tf.get_feed_dict(placeholders,
utils_np.networkxs_to_graphs_tuple(networkxs)))
self.assertAllEqual(
np.array([[0, 0], [1, 0], [2, 0], [3, 0], [0, 1], [1, 1], [2, 1],
[3, 1]]), output.nodes)
self.assertEqual(np.float32, output.nodes.dtype)
self.assertAllEqual(np.array([[0], [1]]), output.globals)
self.assertEqual(np.float32, output.globals.dtype)
sorted_edges_content = sorted(
[(x, y, z)
for x, y, z in zip(output.receivers, output.senders, output.edges)])
self.assertAllEqual([0, 0, 1, 4, 4, 5],
[x[0] for x in sorted_edges_content])
self.assertAllEqual([1, 2, 3, 5, 6, 7],
[x[1] for x in sorted_edges_content])
self.assertEqual(np.float64, output.edges.dtype)
self.assertAllEqual(
np.array([[0, 1, 0], [1, 2, 0], [2, 3, 0], [0, 1, 1], [1, 2, 1],
[2, 3, 1]]), [x[2] for x in sorted_edges_content])
def create_placeholders(rand, batch_size, min_max_nodes, geo_density):
# Create some example data for inspecting the vector sizes.
raw_graphs = generate_raw_graphs(rand, batch_size, min_max_nodes, geo_density)
source_graphs = [source_from_raw(raw) for raw in raw_graphs]
source_ph = utils_tf.placeholders_from_networkxs(
source_graphs,
force_dynamic_num_graphs=True
)
target_graphs = [target_from_raw(raw) for raw in raw_graphs]
# print_graphs(target_graphs)
target_ph = utils_tf.placeholders_from_networkxs(
target_graphs,
force_dynamic_num_graphs=True
)
return source_ph, target_ph
def test_placeholders_from_networkxs(self):
num_graphs = 16
networkxs = [
_generate_graph(batch_index) for batch_index in range(num_graphs)
]
placeholders = utils_tf.placeholders_from_networkxs(
networkxs, force_dynamic_num_graphs=False)
self._assert_expected_shapes(placeholders, num_graphs=num_graphs)
self.assertEqual(tf.float32, placeholders.nodes.dtype)
self.assertEqual(tf.float64, placeholders.edges.dtype)
def __init__(self,gtemp):
tf.reset_default_graph()
with tf.variable_scope("defscope"):
self.SEED = 3
self.GAMMA = 0.99
self.BATCH = 5
random.seed(a=self.SEED)
self.random = np.random.RandomState(seed=self.SEED)
np.random.seed(self.SEED)
tf.set_random_seed(self.SEED)
""" replay memory """
self.D = deque()
""" initialize and setup the network """
self.inputPh = utils_tf.placeholders_from_networkxs(
[self._gtmp2intmp(gtemp)])
self.targetPh = utils_tf.placeholders_from_networkxs(
[self._gtmp2ttmp(gtemp)])
self.numProcessingSteps = 10
self.model = models.EncodeProcessDecode(edge_output_size=1,
node_output_size=0)
self.output_ops_tr = self.model(self.inputPh, self.numProcessingSteps)
self.loss_ops_tr = self._create_loss_ops(self.targetPh, self.output_ops_tr)
self.loss_op_tr = sum(self.loss_ops_tr) / self.numProcessingSteps
self.learning_rate = 1e-3
self.optimizer = tf.train.AdamOptimizer(self.learning_rate)
self.step_op = self.optimizer.minimize(self.loss_op_tr)
self.inputPh, self.targetPh = self._make_all_runnable_in_session(self.inputPh,
self.targetPh)
self.sess = tf.Session()
self.sess.run(tf.global_variables_initializer())
self.epsilon = 0.01
self.outg = None
def test_get_feed_dict_raises(self, none_fields):
networkxs = [_generate_graph(batch_index) for batch_index in range(16)]
placeholders = utils_tf.placeholders_from_networkxs(networkxs)
feed_values = utils_np.networkxs_to_graphs_tuple(networkxs)
with self.assertRaisesRegexp(ValueError, ""):
utils_tf.get_feed_dict(
placeholders.map(lambda _: None, none_fields), feed_values)
with self.assertRaisesRegexp(ValueError, ""):
utils_tf.get_feed_dict(placeholders,
feed_values.map(lambda _: None, none_fields))
def test_placeholders_from_networkxs_missing_edges(self):
num_graphs = 16
networkxs = [
_generate_graph(batch_index, add_edges=False)
for batch_index in range(num_graphs)
]
placeholders = utils_tf.placeholders_from_networkxs(
networkxs, force_dynamic_num_graphs=False)
self.assertEqual(None, placeholders.edges)
self._assert_expected_shapes(
placeholders, but_for=["edges"], num_graphs=num_graphs)
def create_placeholders(config, batch_data, batch_processing=True):
""" if gripper_as_global = False, this function will still return 3 values (input_ph, target_ph, input_ctrl_ph) but the last will
(input_ctrl_ph) will be None, caller needs to check this """
input_graphs, target_graphs, _ = create_singulation_graphs(
config,
batch_data,
initial_pos_vel_known=config.initial_pos_vel_known,
batch_processing=batch_processing)
if batch_processing:
""" in this case we get a list of chunk lists (each representing a full batch)"""
input_graphs = input_graphs[0]
target_graphs = target_graphs[0]
else:
input_graphs = [input_graphs[0]]
target_graphs = [target_graphs[0]]
input_ph = utils_tf.placeholders_from_networkxs(
input_graphs, force_dynamic_num_graphs=True)
target_ph = utils_tf.placeholders_from_networkxs(
target_graphs, force_dynamic_num_graphs=True)
return input_ph, target_ph
def __init__(self, name, gtemp):
self.SEED = 3
self.gtemp = gtemp
random.seed(a=self.SEED)
tf.set_random_seed(self.SEED)
self.name = name
self.inputPh = utils_tf.placeholders_from_networkxs(
[self._gtmp2intmp(gtemp)])
# the scenrio: 1 def 1 att 1 uav: N_edges + N_edges + 2(uav action dim)
# def act dim: N_edges
N_edges = len(gtemp.edges)
# question: split actions or concat all together?
# answer: need to split actions
# self.allActPh = tf.placeholder(tf.float64, shape=[1,N_edges*2+2])
self.defActPh = tf.placeholder(tf.float64, shape=[1, N_edges])
self.attActPh = tf.placeholder(tf.float64, shape=[1, N_edges])
self.uav2ActPh = tf.placeholder(tf.float64, shape=[1, 2])
""" setting up the actor network """
self.actor_out_graphs, self.actor_out_edge = self.actor_network(
gtemp, name + "_actor")
""" setting up the critic network """
self.critic_out_ph = self.critic_network(
gtemp, name + "_critic",
tf.concat([self.defActPh, self.attActPh, self.uav2ActPh], axis=1))
""" structure to compute loss and optimize actor network """
self.actor_loss_op = -tf.reduce_mean(
self.critic_network(
gtemp,
name + "_critic",
tf.concat([self.actor_out_edge, self.attActPh, self.uav2ActPh],
axis=1),
reuse=True))
self.actor_optimizer = tf.train.AdamOptimizer(1e-3)
# step operation to be placed in feed_dict
self.actor_step_op = self.actor_optimizer.minimize(self.actor_loss_op)
""" structure to compute loss and optimize critic network """
self.targetQ_ph = tf.placeholder(shape=[1, 1], dtype=tf.float64)
self.critic_loss_op = tf.reduce_mean(
tf.square(self.targetQ_ph - self.critic_out_ph))
self.critic_optimizer = tf.train.AdamOptimizer(1e-3)
# step operation to be placed in feed_dict
self.critic_step_op = self.critic_optimizer.minimize(
self.critic_loss_op)
def test_placeholders_from_networkxs_hints(self):
num_graphs = 16
networkxs = [
_generate_graph(batch_index, n_nodes=0, add_edges=False)
for batch_index in range(num_graphs)
]
placeholders = utils_tf.placeholders_from_networkxs(
networkxs,
node_shape_hint=[14],
edge_shape_hint=[17],
data_type_hint=tf.float64,
force_dynamic_num_graphs=False)
self.assertAllEqual([None, 14], placeholders.nodes.shape.as_list())
self.assertAllEqual([None, 17], placeholders.edges.shape.as_list())
self._assert_expected_shapes(
placeholders, but_for=["nodes", "edges"], num_graphs=num_graphs)
self.assertEqual(tf.float64, placeholders.nodes.dtype)
self.assertEqual(tf.float64, placeholders.edges.dtype)
def test_feed_data_no_nodes(self):
networkx = [
_generate_graph(batch_index, n_nodes=0, add_edges=False)
for batch_index in range(16)
]
placeholders = utils_tf.placeholders_from_networkxs(
networkx, force_dynamic_num_graphs=True)
# Does not need to be the same size
networkxs = [
_generate_graph(batch_index, n_nodes=0, add_edges=False)
for batch_index in range(2)
]
self.assertEqual(None, placeholders.nodes)
self.assertEqual(None, placeholders.edges)
with self.test_session() as sess:
output = sess.run(
placeholders.replace(nodes=tf.no_op(), edges=tf.no_op()),
utils_tf.get_feed_dict(placeholders,
utils_np.networkxs_to_graphs_tuple(networkxs)))
self.assertAllEqual(np.array([[0], [1]]), output.globals)
self.assertEqual(np.float32, output.globals.dtype)
def create_placeholders(input_graph, target_graph):
input_ph = utils_tf.placeholders_from_networkxs([input_graph])
target_ph = utils_tf.placeholders_from_networkxs([target_graph])
return input_ph, target_ph
# add ops to save and restore all the variables
output_dir = os.path.join(config['output_dir'], prod_name)
print("trained model will be save at:", output_dir)
if not os.path.exists(output_dir):
os.makedirs(output_dir)
## start to build tensorflow sessions
tf.reset_default_graph()
#from nx_graph.model_mlp import SegmentClassifier
#model = SegmentClassifier()
from nx_graph import get_model
model = get_model(config['model']['name'])
input_graphs, target_graphs = generate_input_target(n_graphs)
input_ph = utils_tf.placeholders_from_networkxs(
input_graphs, force_dynamic_num_graphs=True)
target_ph = utils_tf.placeholders_from_networkxs(
target_graphs, force_dynamic_num_graphs=True)
output_ops_tr = model(input_ph, num_processing_steps_tr)
# Training loss.
loss_ops_tr = create_loss_ops(target_ph, output_ops_tr)
# Loss across processing steps.
loss_op_tr = sum(loss_ops_tr) / num_processing_steps_tr
# Optimizer
learning_rate = config_tr['learning_rate']
optimizer = tf.train.AdamOptimizer(learning_rate)
step_op = optimizer.minimize(loss_op_tr)
def create_placeholders(self, input_graphs, batchsize=20):
#input_ph = utils_tf.placeholders_from_networkxs(list(input_graphs[0:batchsize]), force_dynamic_num_graphs=True)
input_ph = utils_tf.placeholders_from_networkxs(
[input_graphs[0]], force_dynamic_num_graphs=True)
output_ph = tf.placeholder(tf.float64, shape=(None, 1))
return input_ph, output_ph
ttmp.graph["features"] = [0.0]
return ttmp
def make_all_runnable_in_session(*args):
"""Lets an iterable of TF graphs be output from a session as NP graphs."""
return [utils_tf.make_runnable_in_session(a) for a in args]
gtemp, _ = getDefaultGraph5x5()
g0 = _gtmp2intmp(gtemp)
t0 = _gtmp2ttmp(gtemp)
inputPh = utils_tf.placeholders_from_networkxs([_gtmp2intmp(gtemp)])
targetPh = utils_tf.placeholders_from_networkxs([_gtmp2ttmp(gtemp)])
# input state
input_graphs = utils_np.networkxs_to_graphs_tuple([g0])
# target
target_graphs = utils_np.networkxs_to_graphs_tuple([t0])
feed_dict = {inputPh: input_graphs, targetPh: target_graphs}
# for critic network
criticModel = models.EncodeProcessDecode(edge_output_size=1,
node_output_size=0)
def create_trained_model(config_name, input_ckpt=None):
"""
@config: configuration for train_nx_graph
"""
# load configuration file
config = load_config(config_name)
config_tr = config['train']
log_every_seconds = config_tr['time_lapse']
batch_size = n_graphs = config_tr['batch_size'] # need optimization
num_processing_steps_tr = config_tr['n_iters'] ## level of message-passing
prod_name = config['prod_name']
if input_ckpt is None:
input_ckpt = os.path.join(config['output_dir'], prod_name)
# generate inputs
base_dir = os.path.join(config['data']['input_dir'],
'event00000{}_g{:03d}.npz')
generate_input_target = inputs_generator(base_dir)
# build TF graph
tf.reset_default_graph()
model = get_model(config['model']['name'])
input_graphs, target_graphs = generate_input_target(n_graphs)
input_ph = utils_tf.placeholders_from_networkxs(
input_graphs, force_dynamic_num_graphs=True)
target_ph = utils_tf.placeholders_from_networkxs(
target_graphs, force_dynamic_num_graphs=True)
output_ops_tr = model(input_ph, num_processing_steps_tr)
def evaluator(iteration, n_test_graphs=10):
try:
sess.close()
except NameError:
pass
sess = tf.Session()
saver = tf.train.Saver()
saver.restore(sess,
os.path.join(input_ckpt, ckpt_name.format(iteration)))
odds = []
tdds = []
for _ in range(n_test_graphs):
feed_dict = create_feed_dict(generate_input_target,
batch_size,
input_ph,
target_ph,
is_trained=False)
predictions = sess.run(
{
"outputs": output_ops_tr,
'target': target_ph
},
feed_dict=feed_dict)
output = predictions['outputs'][-1]
target = predictions['target']
odd, tdd = eval_output(target, output)
odds.append(odd)
tdds.append(tdd)
return np.concatenate(odds), np.concatenate(tdds)
return evaluator
datapath = u'../../mac_yr_local/v280/pursueAsk_graph_v280_190723/tmp_sym/pkls'
gene = data_generator(datapath)
gllx, glly = [], []
for x_y in gene:
gllx = [p['x'] for p in x_y]
#glly=[p['y'] for p in x_y]
glly = [p['yid'] for p in x_y]
break
tf.reset_default_graph()
######
#### placeholder
input_ph = utils_tf.placeholders_from_networkxs(
gllx, force_dynamic_num_graphs=True)
# target_ph = utils_tf.placeholders_from_networkxs(
# glly, force_dynamic_num_graphs=True)
target_ph = tf.placeholder(dtype=tf.int32, shape=[None, None, 1,
1]) # [batch step 1 1]
##
print('')
seed = 1
rand = np.random.RandomState(seed=seed)
# Model parameters.
# Number of processing (message-passing) steps.
#num_processing_steps_tr = 1 if debug_flag==True else 12
#num_processing_steps_ge = 16
t1.add_edge(1,0,features=[1.0])
t1.add_edge(1,2,features=[1.0])
t1.add_edge(2,1,features=[1.0])
t1.add_edge(2,3,features=[1.0])
t1.add_edge(3,2,features=[1.0])
t1.add_edge(0,3,features=[0.0])
t1.add_edge(3,0,features=[0.0])
t1.add_edge(0,2,features=[0.0])
t1.add_edge(2,0,features=[0.0])
t1.add_edge(0,0,features=[0.0])
t1.add_edge(1,1,features=[0.0])
t1.add_edge(2,2,features=[0.0])
t1.add_edge(3,3,features=[0.0])
input_ph = utils_tf.placeholders_from_networkxs([g0],force_dynamic_num_graphs=False,node_shape_hint=[4],edge_shape_hint=[9])
target_ph = utils_tf.placeholders_from_networkxs([t0],force_dynamic_num_graphs=False,node_shape_hint=[4],edge_shape_hint=[9])
input_graphs = utils_np.networkxs_to_graphs_tuple([g0])
target_graphs = utils_np.networkxs_to_graphs_tuple([t0])
feed_dict = {input_ph: input_graphs, target_ph: target_graphs}
num_processing_steps_tr = 1
model = models.EncodeProcessDecode(edge_output_size=1, node_output_size=0)
output_ops_tr = model(input_ph, num_processing_steps_tr)
t1.add_edge(1,0,features=[1.0])
t1.add_edge(1,2,features=[1.0])
t1.add_edge(2,1,features=[1.0])
t1.add_edge(2,3,features=[1.0])
t1.add_edge(3,2,features=[1.0])
t1.add_edge(0,3,features=[0.0])
t1.add_edge(3,0,features=[0.0])
t1.add_edge(0,2,features=[0.0])
t1.add_edge(2,0,features=[0.0])
t1.add_edge(0,0,features=[0.0])
t1.add_edge(1,1,features=[0.0])
t1.add_edge(2,2,features=[0.0])
t1.add_edge(3,3,features=[0.0])
input_ph = utils_tf.placeholders_from_networkxs([g0])
target_ph = utils_tf.placeholders_from_networkxs([t0])
input_graphs = utils_np.networkxs_to_graphs_tuple([g0])
target_graphs = utils_np.networkxs_to_graphs_tuple([t0])
feed_dict = {input_ph: input_graphs, target_ph: target_graphs}
num_processing_steps_tr = 10
num_processing_steps_ge = 10
model = models.EncodeProcessDecode(edge_output_size=1, node_output_size=0)
