def scheduleIsValid(network: STN, schedule: dict) -> STN:
## Check that the schedule is actually defined on all relevant vertices
# This number is arbitrary - any sufficiently small, positive constant works
epsilon = 0.001
vertices = network.getAllVerts()
for vertex in vertices:
vertexID = vertex.nodeID
assert vertexID in schedule
# Check that the schedule is valid
edges = network.getAllEdges()
for edge in edges:
# Loop through the constraints
start = edge.i
fin = edge.j
uBound = edge.Cij
lBound = -edge.Cji
boundedAbove = (schedule[fin] - schedule[start]) <= uBound + epsilon
boundedBelow = (schedule[fin] - schedule[start]) >= lBound - epsilon
# Check if constraint is not satisfied
if ((not boundedAbove) or (not boundedBelow)):
return False
return True
def sample(STN, success='default', LP='original', gauss=False):
if LP == 'original':
_, bounds, epsilons = originalLP(STN.copy(), naiveObj=False)
elif LP == 'proportion':
_, _, bounds, epsilons = proportionLP(STN.copy())
else:
_, _, bounds, epsilons = maxminLP(STN.copy())
original, shrinked = newInterval(STN, epsilons)
if not gauss:
degree = calculateMetric(original, shrinked)[2]
else:
degree = calculateMetric(original, shrinked, gauss)
schedule = {}
for i in list(STN.verts.keys()):
if i not in STN.uncontrollables:
time = (bounds[(i, '-')].varValue + bounds[(i, '+')].varValue) / 2
schedule[i] = time
# Collect the sample data.
count = 0
for i in range(10000):
if success == 'default':
result = sampleOnce(original, shrinked, gauss)
else:
result = altSampleOnce(STN, schedule.copy(), gauss)
if result:
count += 1
success = float(count / 10000)
return degree, success
def processOptimal():
json_folder = input("Please input folder with json file:\n")
json_list = glob.glob(os.path.join(json_folder, '*.json'))
result = {}
for fname in json_list:
p, f = os.path.split(fname)
print("Processing: ", f)
STN = loadSTNfromJSONfile(fname)
new_STN, count = relaxSearch(STN.copy())
new, orig, degree = dynamicMetric(STN.copy(), new_STN.copy())
result[f] = {}
result[f]['shrinked'] = new
result[f]['original'] = orig
result[f]['degree'] = degree
output_folder = input("Please input output folder:\n")
filename = os.path.join(output_folder, 'result_optimal.json')
with open(filename, 'w') as f:
json.dump(result, f)
return result
def _read_stn(self, file):
network = STN()
state = ""
for line in file:
if line.startswith('#'):
if "points" in line.lower() and "num" in line.lower():
state = "NO_POINTS"
elif "edges" in line.lower() and "num" in line.lower():
state = "NO_EDGES"
elif "links" in line.lower():
state = "NO_LINKS"
elif "names" in line.lower():
state = "NAMES"
elif "edges" in line.lower():
state = "EDGES"
edge_counter = 0
elif "links" in line.lower():
state = "LINKS"
else:
pass
else:
if state == 'NO_POINTS':
num_points = int(line)
network.length = num_points
network.successor_edges = [{} for i in range(num_points)]
network.names_list = ["0" for i in range(num_points)]
elif state == 'NO_EDGES':
num_edges = int(line)
elif state == 'NO_LINKS':
# for testing, throw an error
raise Exception(
"Simple Temporal Networks do not have contingent links."
)
elif state == 'NAMES':
list_of_nodes = line.split()
if len(list_of_nodes) != num_points:
raise Exception(
"Number of names does not match the number of nodes provided"
)
for idx, node_name in enumerate(list_of_nodes):
network.names_dict[node_name] = idx
network.names_list[idx] = node_name
elif state == 'EDGES':
weights = line.split()
edge_counter += 1
# make a list of list of tuples
idx_node = network.names_dict[weights[0]]
idx_successor = network.names_dict[weights[2]]
network.successor_edges[idx_node][idx_successor] = int(
weights[1])
elif state == 'LINKS':
raise Exception(
"Simple Temporal Networks do not have contingent links."
)
else:
pass
if num_edges != edge_counter:
raise Exception(
"Number of edges does not match the number given above")
return network
def readNeos(filename, json_folder):
f = open(filename, 'r')
for i in range(3):
line = f.readline()
obj_value = float(line[17:])
actual = math.exp(obj_value)
p, f = os.path.split(filename)
fname = f[:-4] + '.json'
json_file = os.path.join(json_folder, fname)
STN = loadSTNfromJSONfile(json_file)
result, conflicts, bounds, weight = DC_Checker(STN.copy(), report=False)
contingent = bounds['contingent']
total = 1
for (i, j) in list(STN.contingentEdges.keys()):
edge = STN.contingentEdges[(i, j)]
length = edge.Cij + edge.Cji
total *= length
if (i, j) not in contingent:
actual *= length
return actual, total, float(actual / total)
def early_execution(network: STN, realization: dict) -> bool:
## Bookkeeping for events
all_uncontrollables = set(network.uncontrollables)
unused_events = set(network.verts.keys())
not_scheduled = PriorityQueue()
final_schedule = {}
# Mapping from contingent sources to uncontrollables
contingent_pairs = network.contingentEdges.keys()
disabled_uncontrollables = {src: sink for (src, sink) in contingent_pairs}
# Initialize bounds for simulation - starts off with just controllables
# and zero time point
controllable_bounds = find_bounds(network)
true_weight = {}
for event in controllable_bounds:
not_scheduled.push(event, controllable_bounds[0])
true_weight[event] = controllable_bounds[0]
not_scheduled.addOrDecKey(ZERO_ID, 0)
# Run simulation
old_time = 0
while len(unused_events) > 0:
current_time, activated_event = not_scheduled.pop()
# This check ensures that we popped out a valid time_point
# A better way to deal with this would be to just figure out a way to
# increase priorities of elements in a heap
if activated_event in true_weight:
if true_weight[activated_event] > current_time:
continue
unused_events.remove(activated_event)
final_schedule[activated_event] = current_time
assert old_time < current_time, "Chronology violated!"
if activated_event in disabled_uncontrollables:
# If this is a contingent source, we add the associated uncontrollable sink
# to the queue
uncontrollable = disabled_uncontrollables[activated_event]
delay = realization[uncontrollable]
not_scheduled.push(uncontrollable, current_time + delay)
# Update the bounds for all other timepoints
# We only care about events being moved later in time
relevant_edges = network.getEdges(activated_event)
for edge in relevant_edges:
if (edge.j == activated_event) and (edge.i
not in all_uncontrollables):
if needs_early_update(edge, activated_event, current_time,
true_weight):
lower_bound = current_time - edge.Cij
true_weight[edge.i] = lower_bound
# Keep track of this for next iteration of loop
old_time = current_time
# Check if we dispatched succesfully
return emp.scheduleIsValid(network, final_schedule)
def find_bounds(network: STN) -> dict:
# Add zero timepoint
if ZERO_ID not in network.verts:
network.addVertex(ZERO_ID)
# Add bounds relative to zero timepoint
adjacent_to_zero = set(network.getAdjacent(ZERO_ID))
events = network.verts.keys()
bounds = {}
for event in events:
if (event != ZERO_ID) and (event not in adjacent_to_zero):
return 0
else:
return 0
# To make sure zero timepoint starts first
bounds[ZERO_ID] = (-1.0, 0.0)
return bounds
def _read_stnu(self, file):
network = STN()
state = ""
for line in file:
if line.startswith('#'):
if "points" in line.lower() and "num" in line.lower():
state = "NO_POINTS"
elif "edges" in line.lower() and "num" in line.lower():
state = "NO_EDGES"
elif "links" in line.lower():
state = "NO_LINKS"
elif "names" in line.lower():
state = "NAMES"
elif "edges" in line.lower():
state = "EDGES"
elif "links" in line.lower():
state = "LINKS"
else:
raise Exception("Invalid Network Type")
else:
if state == 'NO_POINTS':
num_points = int(line)
network.length = num_points
network.successor_edges = [{} for i in range(num_points)]
network.names_list = ["0" for i in range(num_points)]
elif state == 'NO_EDGES':
num_edges = int(line)
elif state == 'NO_LINKS':
no_links = int(line)
elif state == 'NAMES':
list_of_nodes = line.split()
if len(list_of_nodes) != num_points:
raise Exception(
"Number of names does not match the number of nodes provided"
)
for idx, node_name in enumerate(list_of_nodes):
network.names_dict[node_name] = idx
network.names_list[idx] = node_name
elif state == 'EDGES':
weights = line.split()
# make a list of list of tuples
idx_node = network.names_dict[weights[0]]
idx_successor = network.names_dict[weights[2]]
network.successor_edges[idx_node][idx_successor] = int(
weights[1])
elif state == 'LINKS':
# deal with contingent links later
pass
else:
pass
def loadSTNfromJSONobj(jsonSTN, using_PSTN=False):
stn = STN()
# Add the root vertex and put it in the T_x set
stn.addVertex(0)
# Add the vertices
for v in jsonSTN['nodes']:
stn.addVertex(v['node_id'])
# Add the edges
for e in jsonSTN['constraints']:
if using_PSTN and 'distribution' in e:
stn.addEdge(e['first_node'], e['second_node'],
float(e['min_duration']), float(e['max_duration']),
e['type'], e['distribution']['name'])
else:
stn.addEdge(e['first_node'], e['second_node'],
float(e['min_duration']), float(e['max_duration']),
e['type'])
return stn
def computeDynamic(nlp=False):
uncertain_folder = input("Please input uncertain STNUs folder:\n")
chain_folder = input("Please input chain STNUs folde:\n")
listOfFile = []
listOfFile += glob.glob(os.path.join(uncertain_folder, '*.json'))
listOfFile += glob.glob(os.path.join(chain_folder, '*.json'))
degree = {}
for fname in listOfFile:
p, f = os.path.split(fname)
print("Processing: ", f)
STN = loadSTNfromJSONfile(fname)
new_STN, count = relaxSearch(STN.copy(), nlp=nlp)
if not new_STN:
degree[f] = 0
else:
degree[f] = dynamicMetric(STN.copy(), new_STN.copy())[2]
return degree
def EncoderImage(data_name, img_dim, embed_size, n_attention, n_detectors, pretrained_alex, rectangle, precomp_enc_type='basic',
no_imgnorm=False, net="alex", div_transform=False):
"""A wrapper to image encoders. Chooses between an different encoders
that uses precomputed image features.
"""
# TEST: use spatial transformers
if precomp_enc_type == "trans":
img_enc = STN(n_detectors, embed_size, pretrained_alex, rectangle, net)
# basic SCAN image encoder
elif precomp_enc_type == 'basic':
img_enc = EncoderImagePrecomp(
img_dim, embed_size, no_imgnorm)
# basic SCAN image encoder with weight normalization
elif precomp_enc_type == 'weight_norm':
img_enc = EncoderImageWeightNormPrecomp(
img_dim, embed_size, no_imgnorm)
# TEST: use of attention module
elif precomp_enc_type == "attention":
img_enc = EncoderImageAttention(
img_dim, embed_size, n_attention, no_imgnorm)
# TEST: train one CNN end2end with spatial segmentations
elif precomp_enc_type == "cnn":
img_enc = CNN_end2end(img_dim, embed_size)
# use of Layers-SCAN with different models
elif precomp_enc_type == "layers":
if net == "alex":
img_enc = LayersModel(img_dim, embed_size)
elif net == "res":
img_enc = Layers_resnest(img_dim, embed_size)
elif net == "res_deep":
img_enc = LayersScanResDeep()
# TEST: use layers model without padding but create equal dimension with fc-layer
elif precomp_enc_type == "layers_same":
img_enc = LayersModelSame(img_dim, embed_size)
# attention module Layers-attention-SCAN
elif precomp_enc_type == "layers_attention":
img_enc = LayerAttention2(img_dim, embed_size, n_attention, no_imgnorm, net)
# TEST: attention module Layers-attention-SCAN with residual image
elif precomp_enc_type == "layers_attention_res":
img_enc = LayerAttention3( img_dim, embed_size, n_attention, no_imgnorm)
# TEST: with fusing Layers-attention-SCAN with prediction of clothing item
elif precomp_enc_type == "layers_attention_im":
img_enc = LayerAttention4(img_dim, embed_size, n_attention, no_imgnorm=False, net='alex')
# use n CNNs in parrarel with diversity loss
elif precomp_enc_type == "cnn_layers":
img_enc = CNN_layers(n_detectors, embed_size, pretrained_alex, net, div_transform)
else:
raise ValueError("Unknown precomp_enc_type: {}".format(precomp_enc_type))
return img_enc
def simulation(network: STN, size: int, verbose=False, gauss=False, relaxed=False) -> float:
# Collect useful data from the original network
contingent_pairs = network.contingentEdges.keys()
contingents = {src: sink for (src, sink) in contingent_pairs}
uncontrollables = set(contingents.values())
if relaxed:
dispatching_network, count, cycles, weights = relaxSearch(getMinLossBounds(network.copy(), 2))
if dispatching_network == None:
dispatching_network = network
else:
dispatching_network = network
total_victories = 0
dc_network = STNtoDCSTN(dispatching_network)
dc_network.addVertex(ZERO_ID)
controllability = dc_network.is_DC()
if verbose:
print("Finished checking DC...")
# Detect if the network has an inconsistency in a fixed edge
verts = dc_network.verts.keys()
for vert in verts:
if (vert, vert) in dc_network.edges:
if verbose:
print("Checking", vert)
edge = dc_network.edges[vert, vert][0]
if edge.weight < 0:
dc_network.edges[(vert, vert)].remove(edge)
dc_network.verts[vert].outgoing_normal.remove(edge)
dc_network.verts[vert].incoming_normal.remove(edge)
del dc_network.normal_edges[(vert, vert)]
# Run the simulation
for j in range(size):
realization = generate_realization(network, gauss)
copy = dc_network.copy()
result = dispatch(dispatching_network, copy, realization, contingents,
uncontrollables, verbose)
if verbose:
print("Completed a simulation.")
if result:
total_victories += 1
goodie = float(total_victories / size)
if verbose:
print(f"Worked {100*goodie}% of the time.")
return goodie
def safely_scheduled(network: STN, partial: dict, event) -> bool:
assert event in partial, "Event not in partial schedule!"
epsilon = 0.001
edges = network.getEdges(event)
for edge in edges:
if (edge.i in partial) and (edge.j in partial):
start, end = (edge.i, edge.j)
lBound, uBound = (-edge.Cji, edge.Cij)
boundedAbove = (partial[end] - partial[start]) <= uBound + epsilon
boundedBelow = (partial[end] - partial[start]) >= lBound - epsilon
if ((not boundedAbove) or (not boundedBelow)):
#print("Violated constraint", edge)
return False
return True
def train():
"""Train STN"""
params = Params()
with tf.Graph().as_default():
model = STN(params.gpu, params)
saver = tf.train.Saver(tf.global_variables())
tf_config = tf.ConfigProto(allow_soft_placement=True)
tf_config.gpu_options.allow_growth = True
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess = tf.Session(config=tf_config)
sess.run(init_op)
initial_step = 0
global_step = model.global_step
ckpt = tf.train.get_checkpoint_state(params.train_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
initial_step = global_step.eval(session=sess)
tf.train.start_queue_runners(sess=sess)
summary_writer = tf.summary.FileWriter(params.train_dir, sess.graph)
summary_op = tf.summary.merge_all()
for step in xrange(initial_step, params.max_steps):
if step % params.summary_step == 0:
op_list = [model.train_op, model.loss, summary_op]
_, loss_value, summary_str = sess.run(op_list)
summary_writer.add_summary(summary_str, step)
print('loss: {}'.format(loss_value))
else:
_, loss_value = sess.run([model.train_op, model.loss])
if step % 10 == 0:
format_str = ('%s: step %d, loss = %.2f')
print(format_str % (datetime.now(), step, loss_value))
sys.stdout.flush()
# Save the model checkpoint periodically.
if step % params.checkpoint_step == 0 or (step +
1) == params.max_steps:
checkpoint_path = os.path.join(params.train_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
def set_dynamic_zeropoint(network: STN):
network = network.copy()
largish = 1000000.0
if ZERO_ID not in network.verts:
network.addVertex(ZERO_ID)
adjacent_events = set(network.getAdjacent(ZERO_ID))
for event in network.verts:
if (event not in adjacent_events) and (event != ZERO_ID):
network.addEdge(ZERO_ID, event, 0.0, largish)
return network
def prepareDynamic(STN):
epsilons = {}
result, conflicts, bounds, weight = DC_Checker(STN.copy())
contingent = bounds['contingent']
constraint = ''
Obj = ''
for i, j in list(contingent.keys()):
edge, bound = contingent[(i, j)]
length = edge.Cij + edge.Cji
epsilons[j] = ('EPS_%i' % j, 0, length)
constraint += epsilons[j][0] + ' + '
Obj += 'log(' + str(length) + ' - ' + epsilons[j][0] + ') + '
constraint = constraint[:-3] + ' >= ' + str(-weight)
Obj = Obj[:-3]
return epsilons, constraint, Obj
def modelObjDynamic(STN, fname):
epsilons, constraint, Obj = prepareDynamic(STN.copy())
f = open(fname, 'w')
for v, l, h in list(epsilons.values()):
line = 'var ' + v + ' >= ' + str(l)
if h != None:
line += ', <= ' + str(h)
line += ';'
f.write(line + '\n')
Obj_line = '\nmaximize VOLUME: ' + Obj + ';\n\n'
f.write(Obj_line)
constraint_line = 'subject to WEIGHT: ' + constraint + ';\n'
f.write(constraint_line)
f.close()
def compare(actual_Dict):
dynamic_folder = input("Please input directory with DC STNUs:\n")
uncertain_folder = input("Please input directory with uncertain STNUs:\n")
compare_Dict = {}
for x in list(actual_Dict.keys()):
actual_volume = actual_Dict[x]
if x[:7] == 'dynamic':
fname = os.path.join(dynamic_folder, x + '.json')
else:
fname = os.path.join(uncertain_folder, x + '.json')
STN = loadSTNfromJSONfile(fname)
_, _, epsilons = originalLP(STN.copy())
original, shrinked = newInterval(STN, epsilons)
old, new, degree = calculateMetric(original, shrinked)
actual = float(actual_volume / old)
compare_Dict[x] = (degree, actual)
return compare_Dict
def train():
"""Train STN"""
# load data
print("loading MNIST dataset...")
trainData, validData, testData = loadMNIST("data/MNIST.npz")
batch_size = 50
with tf.Graph().as_default():
model = STN(FLAGS.gpu)
saver = tf.train.Saver(tf.global_variables())
tfConfig = tf.ConfigProto(allow_soft_placement=True)
tfConfig.gpu_options.allow_growth = True
init = tf.global_variables_initializer()
sess = tf.Session(config=tfConfig)
sess.run(init)
initial_step = 0
global_step = model.global_step
ckpt = tf.train.get_checkpoint_state(FLAGS.train_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
initial_step = global_step.eval(session=sess)
tf.train.start_queue_runners(sess=sess)
summary_writer = tf.summary.FileWriter(FLAGS.train_dir, sess.graph)
summary_op = tf.summary.merge_all()
for step in xrange(initial_step, FLAGS.max_steps):
start_time = time.time()
# generate training data
rand_idx = np.random.randint(len(trainData["image"]),
size=batch_size)
image_per_batch = trainData["image"][rand_idx]
label_per_batch = trainData["label"][rand_idx]
image_per_batch = np.reshape(image_per_batch,
[batch_size, 28, 28, 1])
feed_dict = {
model.image_input: image_per_batch,
model.labels: label_per_batch,
}
if step % FLAGS.summary_step == 0:
op_list = [model.train_op, model.loss, summary_op]
_, loss_value, summary_str = sess.run(op_list,
feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
print('loss: {}'.format(loss_value))
else:
_, loss_value = sess.run([model.train_op, model.loss],
feed_dict=feed_dict)
duration = time.time() - start_time
if step % 10 == 0:
num_images_per_step = batch_size
images_per_sec = num_images_per_step / duration
sec_per_batch = float(duration)
format_str = (
'%s: step %d, loss = %.2f (%.1f images/sec; %.3f '
'sec/batch)')
print(format_str % (datetime.now(), step, loss_value,
images_per_sec, sec_per_batch))
sys.stdout.flush()
# Save the model checkpoint periodically.
if step % FLAGS.checkpoint_step == 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 loadSTNfromJSONobj(jsonSTN, reduction=True, using_PSTN=True):
stn = STN()
# Add the root vertex and put it in the T_x set
stn.addVertex(0, 0, None)
# TODO: wtf? Why are we executing a point outside of a simulation in the
# first place?
stn.execute(0)
agents = []
# Add the vertices
for v in jsonSTN['nodes']:
# Accumulate a list of all the owners to retrieve the agents.
if not v['owner_id'] in agents:
agents.append(v['owner_id'])
# We don't necessarily need a location, just set it to None if we don't.
if not ('location' in v):
v['location'] = None
stn.addVertex(v['node_id'], v['local_id'], v['owner_id'],
v['location'])
# TODO: Change edge adding to allow integers to refer to vertecies,
# rather than actually connecting integers together. (silly legacy support)
stn.addEdge(0, v['node_id'], float(v['min_domain']),
float(v['max_domain']))
if 'executed' in v:
if v['executed']:
stn.execute(v['node_id'])
# Add the edges
for e in jsonSTN['constraints']:
if 'distribution' in e and using_PSTN:
stn.addEdge(e['first_node'], e['second_node'],
float(e['min_duration']), float(e['max_duration']),
e['distribution']['name'])
else:
stn.addEdge(e['first_node'], e['second_node'],
float(e['min_duration']), float(e['max_duration']))
#numAgents = jsonSTN['num_agents']
stn.numAgents = len(agents) # NOTE: deprecated, agents replaces numAgents.
stn.agents = agents
if reduction:
# Triangulate the STN
stnCopy = stn.copy()
agentWait = []
# Set up the wait timer for agent load balancing
for a in stn.agents:
agentWait.append(0)
# Perform the reduction
while len(stnCopy.verts) > 1:
for a in stn.agents:
if agentWait[a] == 0:
created = stnreduce(stnCopy, a, stn)
agentWait[a] = created
else:
agentWait[a] -= 1
# Return back an dictionary for easy labelling of return types.
# FIXME: Remove the deprecated numAgents value
output_dict = {'stn': stn, 'agent_count': stn.numAgents}
# Ideally, this would return a class, however, this is already butchering
# quite a bit of legacy support, and a class would end up being even more
# damaging.
return output_dict
def loadSTNfromJSONobj(jsonSTN, using_PSTN=True):
stn = STN()
# Add the root vertex and put it in the T_x set
stn.addVertex(0)
# Add the vertices
for v in jsonSTN['nodes']:
stn.addVertex(v['node_id'])
if 'min_domain' in v:
stn.addEdge(0, v['node_id'], float(v['min_domain']),
float(v['max_domain']))
else:
if not stn.edgeExists(0, v['node_id']):
stn.addEdge(0,v['node_id'], float(0), float('inf'))
# Add the edges
for e in jsonSTN['constraints']:
if stn.edgeExists(e['first_node'], e['second_node']):
stn.updateEdge(e['first_node'], e['second_node'],float(e['max_duration']))
stn.updateEdge(e['second_node'], e['first_node'],float(e['min_duration']))
else:
if using_PSTN and 'distribution' in e:
stn.addEdge(e['first_node'], e['second_node'],
float(max(0,e['min_duration'])), float(e['max_duration']),
e['distribution']['type'], e['distribution']['name'])
elif 'type' in e:
if e['type'] == 'stcu':
dist = "U_"+str(e['min_duration']) + "_" + str(e['max_duration'])
stn.addEdge(e['first_node'], e['second_node'],
float(max(0,e['min_duration'])), float(e['max_duration']),
e['type'], dist)
else:
stn.addEdge(e['first_node'], e['second_node'],
float(e['min_duration']), float(e['max_duration']),
e['type'])
else:
stn.addEdge(e['first_node'], e['second_node'],
float(e['min_duration']), float(e['max_duration']))
return stn
def prepare(STN):
bounds = {}
epsilons = {}
constraints = {}
for i in STN.verts:
if STN.getEdgeWeight(0, i) == float('inf'):
STN.setMakespan(MAX_FLOAT)
break
for i in list(STN.verts.keys()):
bounds[(i, '+')] = ('T%iHI' % i, 0, STN.getEdgeWeight(0, i))
low = 0 if STN.getEdgeWeight(i,0) == float('inf') else\
-STN.getEdgeWeight(i,0)
bounds[(i, '-')] = ('T%iLO' % i, low, None)
if i in STN.uncontrollables:
epsilons[(i, '-')] = ('EPS%iLO' % i, 0, None)
epsilons[(i, '+')] = ('EPS%iHI' % i, 0, None)
constraints[i] = bounds[(i, '-')][0] + " <= " + bounds[(i, '+')][0]
elif i == 0:
constraints[(i, '-')] = bounds[(i, '-')][0] + " == 0"
constraints[(i, '+')] = bounds[(i, '+')][0] + " == 0"
else:
constraints[i] = bounds[(i, '-')][0] + " == " + bounds[(i, '+')][0]
for i, j in STN.edges:
if (i, j) in STN.contingentEdges:
constraints[(i, j, '+')] = bounds[(j, '+')][0] + ' - ' + \
bounds[(i, '+')][0] + ' == ' + str(STN.getEdgeWeight(i,j)) \
+ ' - ' + epsilons[(j, '+')][0]
constraints[(i, j, '-')] = bounds[(j, '-')][0] + ' - ' + \
bounds[(i, '-')][0] + ' == ' + str(-STN.getEdgeWeight(j,i))\
+ ' + ' + epsilons[(j, '-')][0]
else:
# NOTE: We need to handle the infinite weight edges. Otherwise
# the LP would be infeasible
high = MAX_FLOAT if STN.getEdgeWeight(i,j) == float('inf') \
else STN.getEdgeWeight(i,j)
low = MAX_FLOAT if STN.getEdgeWeight(j,i) == float('inf') \
else STN.getEdgeWeight(j,i)
constraints[(i, j, '+')] = bounds[(j, '+')][0] + ' - ' + \
bounds[(i, '-')][0] + ' <= ' + str(high)
constraints[(i, j, '-')] = bounds[(i, '+')][0] + ' - ' + \
bounds[(j, '-')][0] + ' <= ' + str(low)
Obj = ''
for i in STN.uncontrollables:
Obj += 'log(' + bounds[(i, '+')][0] + ' - ' + bounds[(i, '-')][0] \
+ ') + '
Obj = Obj[:-3]
return bounds, epsilons, constraints, Obj
def generateParallelChain(agent, task):
total_event = ((2 * task) + 1) * agent + 1
while True:
new = STN()
new.addVertex(0)
for i in range(total_event):
new.addVertex(i + 1)
contingent = True
for i in range(agent):
start = ((2 * task) + 1) * i + 1
end = ((2 * task) + 1) * (i + 1)
new.addEdge(0, start, 0, 15)
for j in range(start, end):
type = 'stcu' if contingent else 'stc'
contingent = not contingent
if type == 'stcu':
# low = round(random.uniform(10, 20), 2)
# high = round(random.uniform(30, 40), 2)
low = random.randint(10, 20)
high = random.randint(30, 40)
new.addEdge(j, j + 1, low, high, type='stcu')
else:
# low = round(random.uniform(5, 10), 2)
# high = round(random.uniform(30, 35), 2)
low = random.randint(5, 10)
high = random.randint(35, 40)
new.addEdge(j, j + 1, low, high)
new.addEdge(end, total_event, -10, 10)
num_activity = (2 * task) + 1
max_length = max([e.Cij + e.Cji for e in list(new.edges.values())])
up_bound = max_length * num_activity
# low = round(random.uniform(0.35*up_bound, 0.45*up_bound), 2)
# high = round(random.uniform(0.5*up_bound, 0.6*up_bound), 2)
low = random.randint(int(0.45 * up_bound), int(0.53 * up_bound))
high = random.randint(int(0.55 * up_bound), int(0.65 * up_bound))
new.addEdge(0, total_event, low, high)
print("\n\nChecking consistensy...")
if not new.isConsistent():
continue
print("Checking Dynamic Controllability...")
try:
result, conflicts, bounds, weight = DC_Checker(new.copy(),
report=False)
except Exception:
continue
if result:
return new
def generateChain(task, free):
totalEvent = 2 * (task + 1)
while True:
new = STN()
for i in range(totalEvent):
new.addVertex(i)
L = [random.randint(0, 100) for i in range(task)]
s = sum(L)
L = [int(x / s * free) for x in L]
diff = free - sum(L)
L[-1] += diff
bounds = []
for i in range(totalEvent - 1):
type = 'stcu' if i % 2 == 0 else 'stc'
if type == 'stcu':
lowBound = random.randint(0, 50)
length = random.randint(1, 50)
bounds.append((lowBound, lowBound + length))
new.addEdge(i, i + 1, lowBound, lowBound + length, type='stcu')
else:
lowBound = random.randint(0, 100)
length = L[int((i - 1) / 2)]
bounds.append((lowBound, lowBound + length))
new.addEdge(i, i + 1, lowBound, lowBound + length)
low = sum([x[0] for x in bounds])
high = sum([x[1] for x in bounds])
S = sum([e.Cij + e.Cji for e in list(new.contingentEdges.values())])
# makespan = random.randint(int(0.5*low), low)
makespan = low + int(0.6 * S)
print(low, makespan, high)
new.addEdge(0, task * 2 + 1, 0, makespan)
if new.isConsistent():
return new
round(sum(rmse) / 8.0, 4))
print()
if __name__ == '__main__':
cv2.setNumThreads(0)
opt = parse_args()
print(opt)
Path(opt.ckpt_path).mkdir(parents=True, exist_ok=True)
train_set = StereoDataset(opt.data_path, opt.list_path, opt.train_split)
train_loader = DataLoader(dataset=train_set, num_workers=opt.threads, batch_size=opt.batch_size, shuffle=True)
test_set = StereoDataset(opt.data_path, opt.list_path, opt.test_split)
test_loader = DataLoader(dataset=test_set, num_workers=opt.threads, batch_size=opt.batch_size, shuffle=False)
dpnet = DPN(in_shape=(train_set.height, train_set.width))
stnet = STN(in_shape=(test_set.height, test_set.width), filt=not opt.no_filt)
dpnet = dpnet.cuda()
stnet = stnet.cuda()
dpnet = nn.DataParallel(dpnet, device_ids=[0, 1])
stnet = nn.DataParallel(stnet, device_ids=[0, 1])
dpn_optim = optim.Adam(dpnet.parameters(), lr=opt.lr, weight_decay=opt.decay)
stn_optim = optim.Adam(stnet.parameters(), lr=opt.lr, weight_decay=opt.decay)
dpn_sched = optim.lr_scheduler.StepLR(dpn_optim, opt.step, gamma=opt.gamma)
stn_sched = optim.lr_scheduler.StepLR(stn_optim, opt.step, gamma=opt.gamma)
# vis = Visualizer(server=opt.server, env=opt.env)
if opt.resume:
