def on_load(self):
self.sector_manager = Sector_Manager(SECTORS)
self.route_manager = Route_Manager(ROUTES,
test_routes=VISUALIZE,
draw_paths=VISUALIZE)
self.traffic_manager = Traffic_Manager(
max_ac=MAX_AC,
times=TIME_SEP,
max_spd=CONSTRAINTS["cas"]["max"],
min_spd=CONSTRAINTS["cas"]["min"],
max_alt=32000,
min_alt=32000,
network=self.route_manager)
self.memory = Memory()
self.agent = Agent(state_size=STATE_SHAPE,
action_size=ACTION_SHAPE,
value_size=VALUE_SHAPE)
try:
self.agent.load(path=FILE + "best.h5")
except:
try:
self.agent.load(path=FILE + ".h5")
except:
pass
self.initilized = True
print("ATC: READY")
string = "=================================\n UPDATE: RUNNING EPOCH {}\n=================================\n".format(
self.format_epoch())
self.print_all(string)
def main():
# Obtendo argumentos do terminal
args_length = len(sys.argv) - 1
args = sys.argv[1:]
if args_length != 2:
print_args_error()
return 0
process_filename = args[0]
files_filename = args[1]
# Inicializar sistemas
global GLOBAL_clock
global GLOBAL_processes
GLOBAL_processes = load_processes(process_filename, files_filename)
processes = processes_by_init_time(GLOBAL_processes)
memory = Memory()
# sistema de arquivos
# inicializar
# adicionar dados do arquivo
filesys = FileSystem(files_filename)
#recursos
resources = ResourceManager()
# escalonador
scheduler = Scheduler(resources, memory)
process_running = None
print(GLOBAL_clock)
# loop pelo clock ate lista de processos gerais ficar vazia
while GLOBAL_processes:
# timer (clock)
# retorna processos ou lista de processo
# pelo clock
proc_list = process_launcher(GLOBAL_clock, processes)
# atualiza escalonador
# checar bloqueados no escalonador
# atualizar escalonador (aging)
scheduler.update(process_running)
# forma pacote de processos
# envia pacote para o escalonador
scheduler.send_ready_process(proc_list)
# escalonador retorna processo a ser executado
# retorna sinal se eh pra trocar ou nao o processo atual
process_running = scheduler.get_process_to_execute(process_running)
# dispatcher
dispatch(process_running, memory, filesys, resources)
GLOBAL_clock += 1
#print(GLOBAL_clock)
print(filesys)
def _init_memory(self):
"""
Initialize memory class when it's needed by pymem.
"""
self.memory = Memory()
class Pymem(object):
"""Provides class instance methods for general process and memory
manipulations.
"""
def __init__(self):
"""
Initialize pymem objects
"""
self.process = None
self.memory = None
self.module = None
self.pid = None # refractor
self.process32 = None # refractor
self.process_handle = None
def _init_process(self):
"""
Initialize process class when it's needed by pymem.
"""
self.process = Process()
def _init_memory(self):
"""
Initialize memory class when it's needed by pymem.
"""
self.memory = Memory()
def _init_module(self):
"""
Initialize module class when it's needed by pymem.
"""
self.module = Module()
@is_init('process')
def open_process(self, process_id, debug=True):
"""
Opens a process for interaction.
"""
if debug:
if self.process.open_debug(process_id):
self.process_handle = self.process.h_process
self.pid = process_id
self.process32 = self.process.process32
return True
return False
return self.process.open(process_id)
@is_init('process')
def open_process_from_name(self, process_name, debug=True, number=0):
"""
Opens a process from its name for interaction.
"""
processes = process_from_name(process_name)
if processes is not None:
if len(processes) - 1 == number:
process = processes[len(processes) - 1]
return self.open_process(process.th32ProcessID, debug)
return False
@is_init('memory')
def read_offset(self, address, selected_type):
"""
Read memory from a process.
If the type <T> is supported, this method will provide the required
call in order to read, from the process. If either the type <T> is not
supported or process is not Open, the method will raise an Exception.
Supported types : float, int, uint, long, ulong, int64, uint64, byte
"""
if self.process.is_process_open:
if self.memory.is_process_set == False:
self.memory.set_process(self.process.h_process)
return self.memory.read_offset(address, selected_type)
return False
@is_init('module')
def list_module32(self):
"""
Return module (MODULEENTRY32) loaded by current process.
"""
if self.process32 is not None:
if self.module.is_process_set == False:
self.module.set_process32(self.process32)
return self.module.list_module32()
return []
@is_init('module')
def has_module32(self, module):
"""
Return True if current process has loaded given module (dll)
"""
if self.process32 is not None:
if self.module.is_process_set == False:
self.module.set_process32(self.process32)
return self.module.has_module32(module)
return False
def __init__(self,
neighbor_finder,
node_features,
edge_features,
device,
dropout=0.1,
memory_update_at_start=True,
message_dimension=100,
memory_dimension=200,
n_neighbors=None,
aggregator_type="last",
mean_time_shift_src=0,
std_time_shift_src=1,
mean_time_shift_dst=0,
std_time_shift_dst=1,
threshold=2):
super(DGNN, self).__init__()
self.neighbor_finder = neighbor_finder
self.device = device
self.logger = logging.getLogger(__name__)
self.node_raw_features = torch.from_numpy(
node_features.astype(np.float32)).to(device)
self.edge_raw_features = torch.from_numpy(
edge_features.astype(np.float32)).to(device)
self.n_node_features = self.node_raw_features.shape[1]
self.n_nodes = self.node_raw_features.shape[0]
self.n_edge_features = self.edge_raw_features.shape[1]
self.embedding_dimension = self.n_node_features
self.n_neighbors = n_neighbors
self.memory_s = None
self.memory_g = None
self.threshold = threshold
self.mean_time_shift_src = mean_time_shift_src
self.std_time_shift_src = std_time_shift_src
self.mean_time_shift_dst = mean_time_shift_dst
self.std_time_shift_dst = std_time_shift_dst
self.memory_dimension = memory_dimension
self.memory_update_at_start = memory_update_at_start
self.message_dimension = message_dimension
self.memory_merge = MemoryMerge(self.memory_dimension, self.device)
self.memory_s = Memory(n_nodes=self.n_nodes,
memory_dimension=self.memory_dimension,
message_dimension=message_dimension,
device=device)
self.memory_g = Memory(n_nodes=self.n_nodes,
memory_dimension=self.memory_dimension,
message_dimension=message_dimension,
device=device)
self.message_dim = message_dimension
self.message_aggregator = get_message_aggregator(
aggregator_type=aggregator_type, device=device)
self.message_function = MessageFunction(
memory_dimension=memory_dimension,
message_dimension=message_dimension,
edge_dimension=self.n_edge_features,
device=self.device)
self.memory_updater_s = MemoryUpdater(
memory=self.memory_s,
message_dimension=message_dimension,
memory_dimension=self.memory_dimension,
mean_time_shift_src=self.mean_time_shift_src / 2,
device=self.device)
self.memory_updater_g = MemoryUpdater(
memory=self.memory_g,
message_dimension=message_dimension,
memory_dimension=self.memory_dimension,
mean_time_shift_src=self.mean_time_shift_dst / 2,
device=self.device)
self.propagater_s = Propagater(
memory=self.memory_s,
message_dimension=message_dimension,
memory_dimension=self.memory_dimension,
mean_time_shift_src=self.mean_time_shift_src / 2,
neighbor_finder=self.neighbor_finder,
n_neighbors=self.n_neighbors,
tau=self.threshold,
device=self.device)
self.propagater_g = Propagater(
memory=self.memory_g,
message_dimension=message_dimension,
memory_dimension=self.memory_dimension,
mean_time_shift_src=self.mean_time_shift_dst / 2,
neighbor_finder=self.neighbor_finder,
n_neighbors=self.n_neighbors,
tau=self.threshold,
device=self.device)
self.W_s = nn.Parameter(
torch.zeros(
(memory_dimension, memory_dimension // 2)).to(self.device))
#nn.xavier_
self.W_g = nn.Parameter(
torch.zeros(
(memory_dimension, memory_dimension // 2)).to(self.device))
class DGNN(nn.Module):
def __init__(self,
neighbor_finder,
node_features,
edge_features,
device,
dropout=0.1,
memory_update_at_start=True,
message_dimension=100,
memory_dimension=200,
n_neighbors=None,
aggregator_type="last",
mean_time_shift_src=0,
std_time_shift_src=1,
mean_time_shift_dst=0,
std_time_shift_dst=1,
threshold=2):
super(DGNN, self).__init__()
self.neighbor_finder = neighbor_finder
self.device = device
self.logger = logging.getLogger(__name__)
self.node_raw_features = torch.from_numpy(
node_features.astype(np.float32)).to(device)
self.edge_raw_features = torch.from_numpy(
edge_features.astype(np.float32)).to(device)
self.n_node_features = self.node_raw_features.shape[1]
self.n_nodes = self.node_raw_features.shape[0]
self.n_edge_features = self.edge_raw_features.shape[1]
self.embedding_dimension = self.n_node_features
self.n_neighbors = n_neighbors
self.memory_s = None
self.memory_g = None
self.threshold = threshold
self.mean_time_shift_src = mean_time_shift_src
self.std_time_shift_src = std_time_shift_src
self.mean_time_shift_dst = mean_time_shift_dst
self.std_time_shift_dst = std_time_shift_dst
self.memory_dimension = memory_dimension
self.memory_update_at_start = memory_update_at_start
self.message_dimension = message_dimension
self.memory_merge = MemoryMerge(self.memory_dimension, self.device)
self.memory_s = Memory(n_nodes=self.n_nodes,
memory_dimension=self.memory_dimension,
message_dimension=message_dimension,
device=device)
self.memory_g = Memory(n_nodes=self.n_nodes,
memory_dimension=self.memory_dimension,
message_dimension=message_dimension,
device=device)
self.message_dim = message_dimension
self.message_aggregator = get_message_aggregator(
aggregator_type=aggregator_type, device=device)
self.message_function = MessageFunction(
memory_dimension=memory_dimension,
message_dimension=message_dimension,
edge_dimension=self.n_edge_features,
device=self.device)
self.memory_updater_s = MemoryUpdater(
memory=self.memory_s,
message_dimension=message_dimension,
memory_dimension=self.memory_dimension,
mean_time_shift_src=self.mean_time_shift_src / 2,
device=self.device)
self.memory_updater_g = MemoryUpdater(
memory=self.memory_g,
message_dimension=message_dimension,
memory_dimension=self.memory_dimension,
mean_time_shift_src=self.mean_time_shift_dst / 2,
device=self.device)
self.propagater_s = Propagater(
memory=self.memory_s,
message_dimension=message_dimension,
memory_dimension=self.memory_dimension,
mean_time_shift_src=self.mean_time_shift_src / 2,
neighbor_finder=self.neighbor_finder,
n_neighbors=self.n_neighbors,
tau=self.threshold,
device=self.device)
self.propagater_g = Propagater(
memory=self.memory_g,
message_dimension=message_dimension,
memory_dimension=self.memory_dimension,
mean_time_shift_src=self.mean_time_shift_dst / 2,
neighbor_finder=self.neighbor_finder,
n_neighbors=self.n_neighbors,
tau=self.threshold,
device=self.device)
self.W_s = nn.Parameter(
torch.zeros(
(memory_dimension, memory_dimension // 2)).to(self.device))
#nn.xavier_
self.W_g = nn.Parameter(
torch.zeros(
(memory_dimension, memory_dimension // 2)).to(self.device))
def update_memory(self, source_nodes, destination_nodes, messages_s,
messages_g):
# Aggregate messages for the same nodes
unique_src_nodes, unique_src_messages, unique_src_timestamps = self.message_aggregator.aggregate(
source_nodes, messages_s)
unique_des_nodes, unique_des_messages, unique_des_timestamps = self.message_aggregator.aggregate(
destination_nodes, messages_g)
# Update the memory with the aggregated messages
self.memory_updater_s.update_memory(unique_src_nodes,
unique_src_messages,
timestamps=unique_src_timestamps)
self.memory_updater_g.update_memory(unique_des_nodes,
unique_des_messages,
timestamps=unique_des_timestamps)
def propagate(self, source_nodes, destination_nodes, messages_s,
messages_g):
unique_src_nodes, unique_src_messages, unique_src_timestamps = self.message_aggregator.aggregate(
source_nodes, messages_s)
unique_des_nodes, unique_des_messages, unique_des_timestamps = self.message_aggregator.aggregate(
destination_nodes, messages_g)
self.propagater_s(self.memory_s.memory,
unique_src_nodes,
unique_src_messages,
timestamps=unique_src_timestamps)
self.propagater_g(self.memory_g.memory,
unique_des_nodes,
unique_des_messages,
timestamps=unique_des_timestamps)
def compute_loss(self, memory_s, memory_g, source_nodes,
destination_nodes):
source_mem = self.memory_merge(memory_s[1][source_nodes],
memory_g[1][source_nodes])
destination_mem = self.memory_merge(memory_s[1][destination_nodes],
memory_g[1][destination_nodes])
source_emb = torch.matmul(source_mem, self.W_s)
destination_emb = torch.matmul(destination_mem, self.W_g)
score = torch.sum(source_emb * destination_emb, dim=1)
return score.sigmoid()
def forward(self,
source_nodes,
destination_nodes,
negative_nodes,
edge_times,
edge_idxs,
test=False):
n_samples = len(source_nodes)
nodes = np.concatenate(
[source_nodes, destination_nodes, negative_nodes])
positives = np.concatenate([source_nodes, destination_nodes])
timestamps = np.concatenate([edge_times, edge_times])
memory = None
time_diffs = None
memory_s, last_update_s, memory_g, last_update_g = \
self.get_updated_memory(list(range(self.n_nodes)), list(range(self.n_nodes)),
self.memory_s.messages, self.memory_g.messages)
pos_score = self.compute_loss(memory_s, memory_g, source_nodes,
destination_nodes)
neg_score = self.compute_loss(memory_s, memory_g, source_nodes,
negative_nodes)
self.update_memory(source_nodes, destination_nodes,
self.memory_s.messages, self.memory_g.messages)
self.propagate(source_nodes, destination_nodes, self.memory_s.messages,
self.memory_g.messages)
self.memory_s.clear_messages(positives)
self.memory_g.clear_messages(positives)
unique_sources, source_id_to_messages = self.get_messages(
source_nodes, destination_nodes, edge_times, edge_idxs)
unique_destinations, destination_id_to_messages = self.get_messages(
destination_nodes, source_nodes, edge_times, edge_idxs)
self.memory_s.store_raw_messages(unique_sources, source_id_to_messages)
self.memory_g.store_raw_messages(unique_destinations,
destination_id_to_messages)
if not test:
return pos_score, neg_score
else:
source_mem = self.memory_merge(memory_s[1][source_nodes],
memory_g[1][source_nodes])
destination_mem = self.memory_merge(memory_s[1][destination_nodes],
memory_g[1][destination_nodes])
return source_mem, destination_mem
def get_messages(self, source_nodes, destination_nodes, edge_times,
edge_idxs):
edge_times = torch.from_numpy(edge_times).float().to(self.device)
edge_features = self.edge_raw_features[edge_idxs]
source_memory = self.memory_merge(
self.memory_s.memory[1][source_nodes],
self.memory_g.memory[1][source_nodes])
destination_memory = self.memory_merge(
self.memory_s.memory[1][destination_nodes],
self.memory_g.memory[1][destination_nodes])
source_message = self.message_function.compute_message(
source_memory, destination_memory, edge_features)
messages = defaultdict(list)
unique_sources = np.unique(source_nodes)
for i in range(len(source_nodes)):
messages[source_nodes[i]].append(
(source_message[i], edge_times[i]))
return unique_sources, messages
def get_updated_memory(self, source_nodes, destination_nodes, message_s,
message_g):
unique_src_nodes, unique_src_messages, unique_src_timestamps = self.message_aggregator.aggregate(
source_nodes, message_s)
unique_des_nodes, unique_des_messages, unique_des_timestamps = self.message_aggregator.aggregate(
destination_nodes, message_g)
updated_src_memory, updated_src_last_update = self.memory_updater_s.update_memory(
unique_src_nodes,
unique_src_messages,
timestamps=unique_src_timestamps,
inplace=False)
updated_des_memory, updated_des_last_update = self.memory_updater_g.update_memory(
unique_des_nodes,
unique_des_messages,
timestamps=unique_des_timestamps,
inplace=False)
updated_src_memory = self.propagater_s(
updated_src_memory,
unique_src_nodes,
unique_src_messages,
timestamps=unique_src_timestamps,
inplace=False)
updated_des_memory = self.propagater_g(
updated_des_memory,
unique_des_nodes,
unique_des_messages,
timestamps=unique_des_timestamps,
inplace=False)
return updated_src_memory, updated_src_last_update, updated_des_memory, updated_des_last_update
def set_neighbor_finder(self, neighbor_finder):
self.neighbor_finder = neighbor_finder
self.propagater_s.neighbor_finder = neighbor_finder
self.propagater_g.neighbor_finder = neighbor_finder
def __init__(self,
neighbor_finder,
node_features,
edge_features,
device,
n_layers=2,
n_heads=2,
dropout=0.1,
use_memory=False,
memory_update_at_start=True,
message_dimension=100,
memory_dimension=500,
embedding_module_type="graph_attention",
message_function="mlp",
mean_time_shift_src=0,
std_time_shift_src=1,
mean_time_shift_dst=0,
std_time_shift_dst=1,
n_neighbors=None,
aggregator_type="last",
memory_updater_type="gru",
use_destination_embedding_in_message=False,
use_source_embedding_in_message=False,
dyrep=False):
super(TGN, self).__init__()
self.n_layers = n_layers
self.neighbor_finder = neighbor_finder
self.device = device
self.logger = logging.getLogger(__name__)
self.node_raw_features = torch.from_numpy(
node_features.astype(np.float32)).to(device)
self.edge_raw_features = torch.from_numpy(
edge_features.astype(np.float32)).to(device)
self.n_node_features = self.node_raw_features.shape[1]
self.n_nodes = self.node_raw_features.shape[0]
self.n_edge_features = self.edge_raw_features.shape[1]
self.embedding_dimension = self.n_node_features
self.n_neighbors = n_neighbors
self.embedding_module_type = embedding_module_type
self.use_destination_embedding_in_message = use_destination_embedding_in_message
self.use_source_embedding_in_message = use_source_embedding_in_message
self.dyrep = dyrep
self.use_memory = use_memory
self.time_encoder = TimeEncode(dimension=self.n_node_features)
self.memory = None
self.mean_time_shift_src = mean_time_shift_src
self.std_time_shift_src = std_time_shift_src
self.mean_time_shift_dst = mean_time_shift_dst
self.std_time_shift_dst = std_time_shift_dst
if self.use_memory:
self.memory_dimension = memory_dimension
self.memory_update_at_start = memory_update_at_start
raw_message_dimension = 2 * self.memory_dimension + self.n_edge_features + \
self.time_encoder.dimension
message_dimension = message_dimension if message_function != "identity" else raw_message_dimension
self.memory = Memory(n_nodes=self.n_nodes,
memory_dimension=self.memory_dimension,
input_dimension=message_dimension,
message_dimension=message_dimension,
device=device)
self.message_aggregator = get_message_aggregator(
aggregator_type=aggregator_type, device=device)
self.message_function = get_message_function(
module_type=message_function,
raw_message_dimension=raw_message_dimension,
message_dimension=message_dimension)
self.memory_updater = get_memory_updater(
module_type=memory_updater_type,
memory=self.memory,
message_dimension=message_dimension,
memory_dimension=self.memory_dimension,
device=device)
self.embedding_module_type = embedding_module_type
self.embedding_module = get_embedding_module(
module_type=embedding_module_type,
node_features=self.node_raw_features,
edge_features=self.edge_raw_features,
memory=self.memory,
neighbor_finder=self.neighbor_finder,
time_encoder=self.time_encoder,
n_layers=self.n_layers,
n_node_features=self.n_node_features,
n_edge_features=self.n_edge_features,
n_time_features=self.n_node_features,
embedding_dimension=self.embedding_dimension,
device=self.device,
n_heads=n_heads,
dropout=dropout,
use_memory=use_memory,
n_neighbors=self.n_neighbors)
# MLP to compute probability on an edge given two node embeddings
self.affinity_score = MergeLayer(self.n_node_features,
self.n_node_features,
self.n_node_features, 1)
class TGN(torch.nn.Module):
def __init__(self,
neighbor_finder,
node_features,
edge_features,
device,
n_layers=2,
n_heads=2,
dropout=0.1,
use_memory=False,
memory_update_at_start=True,
message_dimension=100,
memory_dimension=500,
embedding_module_type="graph_attention",
message_function="mlp",
mean_time_shift_src=0,
std_time_shift_src=1,
mean_time_shift_dst=0,
std_time_shift_dst=1,
n_neighbors=None,
aggregator_type="last",
memory_updater_type="gru",
use_destination_embedding_in_message=False,
use_source_embedding_in_message=False,
dyrep=False):
super(TGN, self).__init__()
self.n_layers = n_layers
self.neighbor_finder = neighbor_finder
self.device = device
self.logger = logging.getLogger(__name__)
self.node_raw_features = torch.from_numpy(
node_features.astype(np.float32)).to(device)
self.edge_raw_features = torch.from_numpy(
edge_features.astype(np.float32)).to(device)
self.n_node_features = self.node_raw_features.shape[1]
self.n_nodes = self.node_raw_features.shape[0]
self.n_edge_features = self.edge_raw_features.shape[1]
self.embedding_dimension = self.n_node_features
self.n_neighbors = n_neighbors
self.embedding_module_type = embedding_module_type
self.use_destination_embedding_in_message = use_destination_embedding_in_message
self.use_source_embedding_in_message = use_source_embedding_in_message
self.dyrep = dyrep
self.use_memory = use_memory
self.time_encoder = TimeEncode(dimension=self.n_node_features)
self.memory = None
self.mean_time_shift_src = mean_time_shift_src
self.std_time_shift_src = std_time_shift_src
self.mean_time_shift_dst = mean_time_shift_dst
self.std_time_shift_dst = std_time_shift_dst
if self.use_memory:
self.memory_dimension = memory_dimension
self.memory_update_at_start = memory_update_at_start
raw_message_dimension = 2 * self.memory_dimension + self.n_edge_features + \
self.time_encoder.dimension
message_dimension = message_dimension if message_function != "identity" else raw_message_dimension
self.memory = Memory(n_nodes=self.n_nodes,
memory_dimension=self.memory_dimension,
input_dimension=message_dimension,
message_dimension=message_dimension,
device=device)
self.message_aggregator = get_message_aggregator(
aggregator_type=aggregator_type, device=device)
self.message_function = get_message_function(
module_type=message_function,
raw_message_dimension=raw_message_dimension,
message_dimension=message_dimension)
self.memory_updater = get_memory_updater(
module_type=memory_updater_type,
memory=self.memory,
message_dimension=message_dimension,
memory_dimension=self.memory_dimension,
device=device)
self.embedding_module_type = embedding_module_type
self.embedding_module = get_embedding_module(
module_type=embedding_module_type,
node_features=self.node_raw_features,
edge_features=self.edge_raw_features,
memory=self.memory,
neighbor_finder=self.neighbor_finder,
time_encoder=self.time_encoder,
n_layers=self.n_layers,
n_node_features=self.n_node_features,
n_edge_features=self.n_edge_features,
n_time_features=self.n_node_features,
embedding_dimension=self.embedding_dimension,
device=self.device,
n_heads=n_heads,
dropout=dropout,
use_memory=use_memory,
n_neighbors=self.n_neighbors)
# MLP to compute probability on an edge given two node embeddings
self.affinity_score = MergeLayer(self.n_node_features,
self.n_node_features,
self.n_node_features, 1)
def compute_temporal_embeddings(self,
source_nodes,
destination_nodes,
negative_nodes,
edge_times,
edge_idxs,
n_neighbors=20):
"""
Compute temporal embeddings for sources, destinations, and negatively sampled destinations.
source_nodes [batch_size]: source ids.
:param destination_nodes [batch_size]: destination ids
:param negative_nodes [batch_size]: ids of negative sampled destination
:param edge_times [batch_size]: timestamp of interaction
:param edge_idxs [batch_size]: index of interaction
:param n_neighbors [scalar]: number of temporal neighbor to consider in each convolutional
layer
:return: Temporal embeddings for sources, destinations and negatives
"""
n_samples = len(source_nodes)
nodes = np.concatenate(
[source_nodes, destination_nodes, negative_nodes])
positives = np.concatenate([source_nodes, destination_nodes])
timestamps = np.concatenate([edge_times, edge_times, edge_times])
memory = None
time_diffs = None
# get memory from before the current timestep
if self.use_memory:
if self.memory_update_at_start:
# Update memory for all nodes with messages stored in previous batches
memory, last_update = self.get_updated_memory(
list(range(self.n_nodes)), self.memory.messages)
else:
memory = self.memory.get_memory(list(range(self.n_nodes)))
last_update = self.memory.last_update
# Compute differences between the time the memory of a node was last updated,
# and the time for which we want to compute the embedding of a node
source_time_diffs = torch.LongTensor(edge_times).to(
self.device) - last_update[source_nodes].long()
source_time_diffs = (source_time_diffs - self.mean_time_shift_src
) / self.std_time_shift_src
destination_time_diffs = torch.LongTensor(edge_times).to(
self.device) - last_update[destination_nodes].long()
destination_time_diffs = (
destination_time_diffs -
self.mean_time_shift_dst) / self.std_time_shift_dst
negative_time_diffs = torch.LongTensor(edge_times).to(
self.device) - last_update[negative_nodes].long()
negative_time_diffs = (
negative_time_diffs -
self.mean_time_shift_dst) / self.std_time_shift_dst
time_diffs = torch.cat([
source_time_diffs, destination_time_diffs, negative_time_diffs
],
dim=0)
#TODO where are the node / edge features? shouldn't they be input in the compute_embedding function?
# Compute the embeddings using the embedding module, combines memory + current batch of interactions
pdb.set_trace()
node_embedding = self.embedding_module.compute_embedding(
memory=memory,
source_nodes=nodes,
timestamps=timestamps,
n_layers=self.n_layers,
n_neighbors=n_neighbors,
time_diffs=time_diffs)
source_node_embedding = node_embedding[:n_samples]
destination_node_embedding = node_embedding[n_samples:2 * n_samples]
negative_node_embedding = node_embedding[2 * n_samples:]
# update memory with information from current batch
if self.use_memory:
if self.memory_update_at_start:
# Persist the updates to the memory only for sources and destinations (since now we have
# new messages for them)
self.update_memory(positives, self.memory.messages)
assert torch.allclose(memory[positives], self.memory.get_memory(positives), atol=1e-5), \
"Something wrong in how the memory was updated"
# Remove messages for the positives since we have already updated the memory using them
self.memory.clear_messages(positives)
unique_sources, source_id_to_messages = self.get_raw_messages(
source_nodes, source_node_embedding, destination_nodes,
destination_node_embedding, edge_times, edge_idxs)
unique_destinations, destination_id_to_messages = self.get_raw_messages(
destination_nodes, destination_node_embedding, source_nodes,
source_node_embedding, edge_times, edge_idxs)
if self.memory_update_at_start:
self.memory.store_raw_messages(unique_sources,
source_id_to_messages)
self.memory.store_raw_messages(unique_destinations,
destination_id_to_messages)
else:
self.update_memory(unique_sources, source_id_to_messages)
self.update_memory(unique_destinations,
destination_id_to_messages)
if self.dyrep:
source_node_embedding = memory[source_nodes]
destination_node_embedding = memory[destination_nodes]
negative_node_embedding = memory[negative_nodes]
return source_node_embedding, destination_node_embedding, negative_node_embedding
def compute_edge_probabilities(self,
source_nodes,
destination_nodes,
negative_nodes,
edge_times,
edge_idxs,
n_neighbors=20):
"""
Compute probabilities for edges between sources and destination and between sources and
negatives by first computing temporal embeddings using the TGN encoder and then feeding them
into the MLP decoder.
:param destination_nodes [batch_size]: destination ids
:param negative_nodes [batch_size]: ids of negative sampled destination
:param edge_times [batch_size]: timestamp of interaction
:param edge_idxs [batch_size]: index of interaction
:param n_neighbors [scalar]: number of temporal neighbor to consider in each convolutional
layer
:return: Probabilities for both the positive and negative edges
"""
n_samples = len(source_nodes)
source_node_embedding, destination_node_embedding, negative_node_embedding = self.compute_temporal_embeddings(
source_nodes, destination_nodes, negative_nodes, edge_times,
edge_idxs, n_neighbors)
score = self.affinity_score(
torch.cat([source_node_embedding, source_node_embedding], dim=0),
torch.cat([destination_node_embedding,
negative_node_embedding])).squeeze(dim=0)
pos_score = score[:n_samples]
neg_score = score[n_samples:]
return pos_score.sigmoid(), neg_score.sigmoid()
def update_memory(self, nodes, messages):
# Aggregate messages for the same nodes
unique_nodes, unique_messages, unique_timestamps = \
self.message_aggregator.aggregate(
nodes,
messages)
if len(unique_nodes) > 0:
unique_messages = self.message_function.compute_message(
unique_messages)
# Update the memory with the aggregated messages
self.memory_updater.update_memory(unique_nodes,
unique_messages,
timestamps=unique_timestamps)
def get_updated_memory(self, nodes, messages):
# Aggregate messages for the same nodes
unique_nodes, unique_messages, unique_timestamps = \
self.message_aggregator.aggregate(
nodes,
messages)
if len(unique_nodes) > 0:
unique_messages = self.message_function.compute_message(
unique_messages)
updated_memory, updated_last_update = self.memory_updater.get_updated_memory(
unique_nodes, unique_messages, timestamps=unique_timestamps)
return updated_memory, updated_last_update
def get_raw_messages(self, source_nodes, source_node_embedding,
destination_nodes, destination_node_embedding,
edge_times, edge_idxs):
edge_times = torch.from_numpy(edge_times).float().to(self.device)
edge_features = self.edge_raw_features[edge_idxs]
source_memory = self.memory.get_memory(source_nodes) if not \
self.use_source_embedding_in_message else source_node_embedding
destination_memory = self.memory.get_memory(destination_nodes) if \
not self.use_destination_embedding_in_message else destination_node_embedding
source_time_delta = edge_times - self.memory.last_update[source_nodes]
source_time_delta_encoding = self.time_encoder(
source_time_delta.unsqueeze(dim=1)).view(len(source_nodes), -1)
source_message = torch.cat([
source_memory, destination_memory, edge_features,
source_time_delta_encoding
],
dim=1)
messages = defaultdict(list)
unique_sources = np.unique(source_nodes)
for i in range(len(source_nodes)):
messages[source_nodes[i]].append(
(source_message[i], edge_times[i]))
return unique_sources, messages
def set_neighbor_finder(self, neighbor_finder):
self.neighbor_finder = neighbor_finder
self.embedding_module.neighbor_finder = neighbor_finder
class TGN(torch.nn.Module):
"""
TGN model
INIT INPUTS:
neighbor_finder: NeighborFinder instance
node_features: Nodes raw features of shape [n_nodes, node_feat_dim]
edge_features: Edges raw features of shape [n_interactinon, edge_feat_dim]
n_layers: 'L' in the paper
n_heads: Number of attention heads
dropout: For nn.MultiheadAttention()
use_memory: Bool variable, whether to augment the model with a node memory
memory_update_at_start: Bool variable, whether to update memory at the start of the batch
message_dimension: Node message dimension for m_i(t), default 100
memory_dimension: Node memory dimension for s_i(t), default 172
embedding_module_type: How to calculate embedding, default 'graph_attention'
message_function: How to calculate node message, default 'mlp'
mean_time_shift_src:
std_time_shift_src:
mean_time_shift_dst:
std_time_shift_dst:
n_neighbors: How many temporal neighbos to be extracted
aggregator_type: How to aggregate messages, default 'last'
memory_updater_type: How to update node memory
use_destination_embedding_in_message:
use_source_embedding_in_message:
"""
def __init__(self, neighbor_finder, node_features, edge_features, device, n_layers=2,
n_heads=2, dropout=0.1, use_memory=True, memory_update_at_start=True,
message_dimension=100, memory_dimension=172, embedding_module_type="graph_attention",
message_function="mlp", mean_time_shift_src=0, std_time_shift_src=1,
mean_time_shift_dst=0, std_time_shift_dst=1, n_neighbors=None, aggregator_type="last",
memory_updater_type="gru", use_destination_embedding_in_message=False,
use_source_embedding_in_message=False):
super(TGN, self).__init__()
self.n_layers = n_layers
self.neighbor_finder = neighbor_finder
self.device = device
self.logger = logging.getLogger(__name__)
self.node_raw_features = torch.from_numpy(node_features.astype(np.float32)).to(device) # node features to tensor
self.edge_raw_features = torch.from_numpy(edge_features.astype(np.float32)).to(device) # edge features to tensor
self.n_node_features = self.node_raw_features.shape[1] # node_feat_dim
self.n_nodes = self.node_raw_features.shape[0] # n_nodes
self.n_edge_features = self.edge_raw_features.shape[1] # edge_feat_dim
self.embedding_dimension = self.n_node_features # emb_dim = node_feat_dim
self.n_neighbors = n_neighbors
self.embedding_module_type = embedding_module_type
self.use_destination_embedding_in_message = use_destination_embedding_in_message
self.use_source_embedding_in_message = use_source_embedding_in_message
self.use_memory = use_memory
self.time_encoder = TimeEncode(dimension=self.n_node_features) # encodes time to shape [node_feat_dim]
self.memory = None
self.mean_time_shift_src = mean_time_shift_src
self.std_time_shift_src = std_time_shift_src
self.mean_time_shift_dst = mean_time_shift_dst
self.std_time_shift_dst = std_time_shift_dst
if self.use_memory:
self.memory_dimension = memory_dimension
self.memory_update_at_start = memory_update_at_start
# m_raw_i = (s_i || s_j || t || e)
raw_message_dimension = 2 * self.memory_dimension + self.n_edge_features + self.time_encoder.dimension # raw message dim
message_dimension = message_dimension if message_function != "identity" else raw_message_dimension # message dim
self.memory = Memory(n_nodes=self.n_nodes,
memory_dimension=self.memory_dimension,
input_dimension=message_dimension,
device=device)
self.message_function = get_message_function(module_type=message_function,
raw_message_dimension=raw_message_dimension,
message_dimension=message_dimension) # message function
self.message_aggregator = get_message_aggregator(aggregator_type=aggregator_type, device=device) # message aggregator
# self.memory_updater = GRUMemoryUpdater(memory=self.memory,
# message_dimension=message_dimension,
# memory_dimension=self.memory_dimension, device=device)
self.memory_updater = get_memory_updater(module_type=memory_updater_type,
memory=self.memory, message_dimension=message_dimension,
memory_dimension=self.memory_dimension, device=device) # memory updator
self.embedding_module_type = embedding_module_type
# self.embedding_module = get_embedding_module(module_type=embedding_module_type,
# node_features=self.node_raw_features,
# edge_features=self.edge_raw_features,
# neighbor_finder=self.neighbor_finder,
# time_encoder=self.time_encoder,
# n_layers=self.n_layers,
# n_node_features=self.n_node_features,
# n_edge_features=self.n_edge_features,
# n_time_features=self.n_node_features,
# embedding_dimension=self.embedding_dimension,
# device=self.device,
# n_heads=n_heads, dropout=dropout,
# use_memory=use_memory,
# n_neighbors=self.n_neighbors)
self.embedding_module = get_embedding_module(module_type=embedding_module_type,
node_features=self.node_raw_features,
edge_features=self.edge_raw_features,
neighbor_finder=self.neighbor_finder,
time_encoder=self.time_encoder,
n_layers=self.n_layers,
n_node_features=self.n_node_features,
n_edge_features=self.n_edge_features,
n_time_features=self.n_node_features,
embedding_dimension=self.embedding_dimension,
device=self.device,
n_heads=n_heads, dropout=dropout,
use_memory=use_memory,
n_neighbors=self.n_neighbors) # embedding module
# MLP to compute probability on an edge given two node embeddings
self.affinity_score = MergeLayer(self.n_node_features, self.n_node_features, self.n_node_features, 1)
def set_neighbor_finder(self, neighbor_finder):
self.neighbor_finder = neighbor_finder
self.embedding_module.neighbor_finder = neighbor_finder
def get_updated_memory(self, nodes, messages):
"""
Get (but not persist) updated nodes' memory by using messages (AGG-->MSG-->MEM, while in paper the order is MSG-->AGG-->MEM)
INPUTS:
nodes: A list of length n_nodes; Node ids
message: A dictionary {node_id:[([message_1], timestamp_1), ([message_2], timestamp_2), ...]}; Messages in previous batch
OUTPUTS:
updated_memory: A tensor of shape [unique_nodes, memory_dimension]
updated_last_update: A tensor of shape [unique_nodes]
"""
# Aggregate messages for the same nodes
unique_nodes, unique_messages, unique_timestamps = self.message_aggregator.aggregate(nodes, messages)
if len(unique_nodes) > 0:
unique_messages = self.message_function.compute_message(unique_messages)
updated_memory, updated_last_update = self.memory_updater.get_updated_memory(unique_nodes,
unique_messages,
timestamps=unique_timestamps)
return updated_memory, updated_last_update
def update_memory(self, nodes, messages):
"""
Updated nodes' memory by using messages (AGG-->MSG-->MEM, while in paper the order is MSG-->AGG-->MEM)
INPUTS:
nodes: A list of length len(nodes); Node ids
message: A dictionary {node_id:[([message_1], timestamp_1), ([message_2], timestamp_2), ...]}; Messages in previous batch
"""
# Aggregate messages for the same nodes
unique_nodes, unique_messages, unique_timestamps = self.message_aggregator.aggregate(nodes, messages)
if len(unique_nodes) > 0:
unique_messages = self.message_function.compute_message(unique_messages)
# Update nodes' memory with the aggregated messages
# Notice: update_memory() updates with no returns
self.memory_updater.update_memory(unique_nodes, unique_messages,
timestamps=unique_timestamps)
# def get_raw_messages(self, source_nodes, destination_nodes, edge_times, edge_idxs):
# """
# Get source_nodes' raw messages m_raw(t) = {[S(t-1), e(t)], t}
# INPUTS:
# source_nodes: Array of shape [batch_size]; Nodes' raw message to be calculated
# destination_nodes: Array of shape [batch_size];
# edge_times: Array of shape [batch_size]; Timestamps of interactions (i.e. Current timestamps) for source_nodes
# edge_idxs: Array of shape [batch_size]; Index of interactions (at edge_times) for source_nodes
# OUTPUTS:
# unique_sources: Array of shape [unique source nodes]
# messages: A dictionary {node_id:[([message_1], timestamp_1), ([message_2], timestamp_2), ...]}
# where [message_x] is [S_i(t-1), S_j(t-1), e_ij(t), Phi(t-(t-1))], timestamp_x is the timestamp for each message_x
# """
# edge_times = torch.from_numpy(edge_times).float().to(self.device)
# edge_features = self.edge_raw_features[edge_idxs] # e_ij(t), or e(t)
# source_memory = self.memory.get_memory(source_nodes) # S_i(t-1)
# destination_memory = self.memory.get_memory(destination_nodes) # S_j(t-1)
# source_time_delta = edge_times - self.memory.last_update[source_nodes]
# source_time_delta_encoding = self.time_encoder(source_time_delta.unsqueeze(dim=1)).view(len(source_nodes), -1) # Phi(t-t^wave)
# source_message = torch.cat([source_memory, destination_memory, edge_features, source_time_delta_encoding], dim=1)
# messages = defaultdict(list)
# unique_sources = np.unique(source_nodes)
# for i in range(len(source_nodes)):
# messages[source_nodes[i]].append((source_message[i], edge_times[i]))
# return unique_sources, messages
def get_raw_messages(self, source_nodes, source_node_embedding,
destination_nodes, destination_node_embedding,
edge_times, edge_idxs):
"""
Get source_nodes' raw messages m_raw_i(t) = {[S_i(t-1), S_j(t-1), e(t), Phi(t-(t_last)], t}
INPUTS:
source_nodes: Array of shape [batch_size]; Nodes' raw message to be calculated
destination_nodes: Array of shape [batch_size];
edge_times: Array of shape [batch_size]; Timestamps of interactions (i.e. Current timestamps) for source_nodes
edge_idxs: Array of shape [batch_size]; Index of interactions (at edge_times) for source_nodes
source_node_embedding: z_i(t) with shape [batch_size, emb_dim=node_dim=mem_dim]
destination_node_embedding: z_j(t) with shape [batch_size, emb_dim=node_dim=mem_dim]
OUTPUTS:
unique_sources: Array of shape [unique source nodes]
messages: A dictionary {node_id:[([message_1], timestamp_1), ([message_2], timestamp_2), ...]}
where [message_x] is [S_i(t-1), S_j(t-1), e_ij(t), Phi(t-(t-1))], timestamp_x is the timestamp for each message_x
"""
edge_times = torch.from_numpy(edge_times).float().to(self.device)
edge_features = self.edge_raw_features[edge_idxs] # e_ij(t), or e(t)
# s_i(t-1) or z_i(t)
source_memory = self.memory.get_memory(source_nodes) if not self.use_source_embedding_in_message else source_node_embedding
# s_j(t-1) or z_j(t)
destination_memory = self.memory.get_memory(destination_nodes) if not self.use_destination_embedding_in_message else destination_node_embedding
source_time_delta = edge_times - self.memory.last_update[source_nodes]
source_time_delta_encoding = self.time_encoder(source_time_delta.unsqueeze(dim=1)).view(len(source_nodes), -1) # Phi(t-t^wave)
unique_sources = np.unique(source_nodes)
source_message = torch.cat([source_memory, destination_memory, edge_features, source_time_delta_encoding], dim=1)
messages = defaultdict(list)
for i in range(len(source_nodes)):
messages[source_nodes[i]].append((source_message[i], edge_times[i]))
return unique_sources, messages
def compute_temporal_embeddings(self, source_nodes, destination_nodes, negative_nodes, edge_times, edge_idxs, n_neighbors=20):
"""
Compute temporal embeddings for sources, destinations, and negatively sampled destinations.
Corresponding to algorithm 1 and 2 in the paper.
INPUTS:
source_nodes: Array of shape [batch_size]; Source node ids.
destination_nodes: Array of shape [batch_size]; Destination node ids
negative_nodes: Array of shape [batch_size]; Ids of negative sampled destination
edge_times: Array of shape [batch_size]; Timestamps of interactions (i.e. Current timestamps) for those nodes (i.e. src, dest, neg)
edge_idxs: Array of shape [batch_size]; Index of interactions
n_neighbors: A number of temporal neighbor to consider in each layer (hop)
OUTPUTS: Temporal embeddings for sources, destinations and negatives
source_node_embedding: A tensor of shape [source_nodes, emb_dim]
destination_node_embedding: A tensor of shape [destination_nodes, emb_dim]
negative_node_embedding: A tensor of shape [negative_nodes, emb_dim]
"""
n_samples = len(source_nodes)
nodes = np.concatenate([source_nodes, destination_nodes, negative_nodes]) # all nodes
positives = np.concatenate([source_nodes, destination_nodes]) # positive pairs
timestamps = np.concatenate([edge_times, edge_times, edge_times]) # (current) timestamps for those nodes (i.e. V_2(t_1) and V_2(t_2))
memory = None
time_diffs = None
if self.use_memory:
### Line 5-7 in Algorithm 2: Update memory first with previous batch messages, and then calculate embeddings
if self.memory_update_at_start:
# update memory for ALL nodes with messages stored in previous batches
memory, last_update = self.get_updated_memory(list(range(self.n_nodes)), self.memory.messages)
### Line 3.5 in Algorithm 1: Use previous batch memory and calculate embeddings
else:
memory = self.memory.get_memory(list(range(self.n_nodes)))
last_update = self.memory.last_update
# Compute differences between the time the memory of a node was last updated,
# and the time for which we want to compute the embedding of a node
source_time_diffs = torch.LongTensor(edge_times).to(self.device) - last_update[source_nodes].long()
source_time_diffs = (source_time_diffs - self.mean_time_shift_src) / self.std_time_shift_src
destination_time_diffs = torch.LongTensor(edge_times).to(self.device) - last_update[destination_nodes].long()
destination_time_diffs = (destination_time_diffs - self.mean_time_shift_dst) / self.std_time_shift_dst
negative_time_diffs = torch.LongTensor(edge_times).to(self.device) - last_update[negative_nodes].long()
negative_time_diffs = (negative_time_diffs - self.mean_time_shift_dst) / self.std_time_shift_dst
# time_diffs, i.e. delta_t, is for TimeEmbedding method
time_diffs = torch.cat([source_time_diffs, destination_time_diffs, negative_time_diffs], dim=0)
# Compute the embeddings for [source_nodes, destination_nodes, negative_nodes]
# If memory_update_at_start is True: Line 8 in algorithm 2; The procedure is same as Figure 2 (right) in the paper
# If memory_update_at_start is False: Line 4 in algorithm 1; The procedure is same as Figure 2 (left) in the paper
node_embedding = self.embedding_module.compute_embedding(memory=memory,
source_nodes=nodes,
timestamps=timestamps,
n_layers=self.n_layers,
n_neighbors=n_neighbors,
time_diffs=time_diffs)
source_node_embedding = node_embedding[:n_samples]
destination_node_embedding = node_embedding[n_samples: 2 * n_samples]
negative_node_embedding = node_embedding[2 * n_samples:]
if self.use_memory:
### Line 12 in algorithm 2: If memory_update_at_start, we persist the update to memory (i.e. S(t-1)) here
if self.memory_update_at_start:
# Persist the updates to the memory only for sources and destinations
self.update_memory(positives, self.memory.messages)
# Remove messages for the positives, we have already updated the memory using positives old message
self.memory.clear_messages(positives)
### Line 7 in algorithm 1
### Line 11 in algorithm 2
# get raw message on source nodes
unique_sources, source_id_to_messages = self.get_raw_messages(source_nodes,
source_node_embedding,
destination_nodes,
destination_node_embedding,
edge_times, edge_idxs)
# get raw message on destination nodes
unique_destinations, destination_id_to_messages = self.get_raw_messages(destination_nodes,
destination_node_embedding,
source_nodes,
source_node_embedding,
edge_times, edge_idxs)
### Line 11 in Algorithm 2: If memory_update_at_start, we then store the new raw message
if self.memory_update_at_start:
self.memory.store_raw_messages(unique_sources, source_id_to_messages)
self.memory.store_raw_messages(unique_destinations, destination_id_to_messages)
### Line 7-9 in Algorithm 1: If not memory_update_at_start, we update memory here with new raw message
else:
self.update_memory(unique_sources, source_id_to_messages)
self.update_memory(unique_destinations, destination_id_to_messages)
return source_node_embedding, destination_node_embedding, negative_node_embedding
def compute_edge_probabilities(self, source_nodes, destination_nodes, negative_nodes, edge_times, edge_idxs, n_neighbors=20):
"""
Line 5 in algorithm 1; Line 9 in algorithm 2
Compute probabilities for edges between sources and destination and between sources and
negatives by first computing temporal embeddings using the TGN encoder and then feeding them
into the MLP decoder.
INPUTS:
source_nodes: Array of shape [batch_size]; Source node ids.
destination_nodes: Array of shape [batch_size]; Destination node ids.
negative_nodes: Array of shape [batch_size]; Negative node ids.
edge_times: Array of shape [batch_size]; Timestamps of interactions (i.e. Current timestamps) for those nodes (i.e. src, dest, neg)
edge_idxs: Array of shape [batch_size]; Index of interactions
n_neighbors: A number of temporal neighbor to consider in each layer (i.e. Each hop)
OUTPUTS:
Probabilities for both the positive and negative edges
"""
n_samples = len(source_nodes)
# get node embeddings for all nodes first
source_node_embedding, destination_node_embedding, negative_node_embedding = self.compute_temporal_embeddings(
source_nodes, destination_nodes, negative_nodes, edge_times, edge_idxs, n_neighbors)
# then calculate the P_pos and P_neg
score = self.affinity_score(torch.cat([source_node_embedding, source_node_embedding], dim=0),
torch.cat([destination_node_embedding, negative_node_embedding])).squeeze(dim=0)
pos_score = score[:n_samples]
neg_score = score[n_samples:]
return pos_score.sigmoid(), neg_score.sigmoid()
def _init_memory(self):
"""
Initialize memory class when it's needed by pymem.
"""
self.memory = Memory()
class Pymem(object):
"""Provides class instance methods for general process and memory
manipulations.
"""
def __init__(self):
"""
Initialize pymem objects
"""
self.process = None
self.memory = None
self.module = None
self.pid = None # refractor
self.process32 = None # refractor
self.process_handle = None
def _init_process(self):
"""
Initialize process class when it's needed by pymem.
"""
self.process = Process()
def _init_memory(self):
"""
Initialize memory class when it's needed by pymem.
"""
self.memory = Memory()
def _init_module(self):
"""
Initialize module class when it's needed by pymem.
"""
self.module = Module()
@is_init('process')
def open_process(self, process_id, debug=True):
"""
Opens a process for interaction.
"""
if debug:
if self.process.open_debug(process_id):
self.process_handle = self.process.h_process
self.pid = process_id
self.process32 = self.process.process32
return True
return False
return self.process.open(process_id)
@is_init('process')
def open_process_from_name(self, process_name, debug=True, number=0):
"""
Opens a process from its name for interaction.
"""
processes = process_from_name(process_name)
if processes is not None:
if len(processes) - 1 == number:
process = processes[len(processes) - 1]
return self.open_process(process.th32ProcessID, debug)
return False
@is_init('memory')
def read_offset(self, address, selected_type):
"""
Read memory from a process.
If the type <T> is supported, this method will provide the required
call in order to read, from the process. If either the type <T> is not
supported or process is not Open, the method will raise an Exception.
Supported types : float, int, uint, long, ulong, int64, uint64, byte
"""
if self.process.is_process_open:
if self.memory.is_process_set == False:
self.memory.set_process(self.process.h_process)
return self.memory.read_offset(address, selected_type)
return False
@is_init('module')
def list_module32(self):
"""
Return module (MODULEENTRY32) loaded by current process.
"""
if self.process32 is not None:
if self.module.is_process_set == False:
self.module.set_process32(self.process32)
return self.module.list_module32()
return []
@is_init('module')
def has_module32(self, module):
"""
Return True if current process has loaded given module (dll)
"""
if self.process32 is not None:
if self.module.is_process_set == False:
self.module.set_process32(self.process32)
return self.module.has_module32(module)
return False
class ATC(core.Entity):
''' Example new entity object for BlueSky. '''
def __init__(self):
super().__init__()
self.super_start = time.perf_counter()
self.initilized = False
self.epoch_counter = 0
# [Success, Fail]
self.results = np.zeros(2)
self.all_success = []
self.all_fail = []
self.mean_success = 0
self.all_mean_success, self.best = 0, 0
self.mean_rewards = []
self.epoch_actions = np.zeros(ACTION_SHAPE)
self.start = None
self.stop = None
self.dist = [0, -1]
self.spd = [0, -1]
self.trk = [0, 360]
self.vs = [0, -1]
self.last_observation = {}
self.last_reward_observation = {}
self.previous_action = {}
self.observation = {}
def on_load(self):
self.sector_manager = Sector_Manager(SECTORS)
self.route_manager = Route_Manager(ROUTES,
test_routes=VISUALIZE,
draw_paths=VISUALIZE)
self.traffic_manager = Traffic_Manager(
max_ac=MAX_AC,
times=TIME_SEP,
max_spd=CONSTRAINTS["cas"]["max"],
min_spd=CONSTRAINTS["cas"]["min"],
max_alt=32000,
min_alt=32000,
network=self.route_manager)
self.memory = Memory()
self.agent = Agent(state_size=STATE_SHAPE,
action_size=ACTION_SHAPE,
value_size=VALUE_SHAPE)
try:
self.agent.load(path=FILE + "best.h5")
except:
try:
self.agent.load(path=FILE + ".h5")
except:
pass
self.initilized = True
print("ATC: READY")
string = "=================================\n UPDATE: RUNNING EPOCH {}\n=================================\n".format(
self.format_epoch())
self.print_all(string)
# Functions that need to be called periodically can be indicated to BlueSky
# with the timed_function decorator
@core.timed_function(name='example', dt=12)
def update(self):
# Initilize system
if not self.initilized:
self.on_load()
# Start epoch timer
if not self.start:
self.start = time.perf_counter()
# Create aircraft
self.traffic_manager.spawn()
# Update Aircraft active sectors
self.traffic_manager.update_active(self.sector_manager.system_sectors)
# Generate a full distancematrix between each aircraft
full_dist_matrix = self.get_dist_martix()
# Get nearest ac in a matrix
nearest_ac = self.get_nearest_ac(dist_matrix=full_dist_matrix)
# Get goal distances for each aircraft
g_distance = self.get_goal_distances()
# Get an array of terminal aircraft
terminal_ac, terminal_id = self.get_terminal(nearest_ac, g_distance)
self.handle_terminal(terminal_id)
if self.traffic_manager.check_done():
self.epoch_reset()
return
if not TRAIN and (self.traffic_manager.total % 50 == 0):
string = "Success: {} | Fail: {} | Mean Success: {:.3f}%".format(
int(self.results[0]), int(self.results[1]),
(self.results[0] / MAX_AC) * 100)
self.print_all(string)
if len(traf.id) <= 0:
return
if not len(traf.id) == 0:
policy, normal_state, normal_context = self.get_actions(
terminal_ac, g_distance, full_dist_matrix)
if len(policy) > 0:
idx = 0
new_actions = {}
for i in range(len(traf.id)):
if terminal_ac[i] == 0 and len(
self.traffic_manager.active_sectors[i]) > 0:
if not np.any(np.isnan(policy[idx])):
_id = traf.id[i]
if not _id in self.last_observation.keys():
self.last_observation[_id] = [
normal_state[idx], normal_context[idx]
]
action = np.random.choice(
ACTION_SHAPE, 1, p=policy[idx].flatten())[0]
# print(policy[idx], action)
self.epoch_actions[action] += 1
if not _id in self.observation.keys(
) and _id in self.previous_action.keys():
self.observation[_id] = [
normal_state[idx], normal_context[idx]
]
self.memory.store(_id,
self.last_observation[_id],
self.previous_action[_id],
nearest_ac[idx])
self.last_observation[_id] = self.observation[
_id]
del self.observation[_id]
self.perform_action(i, action)
new_actions[_id] = action
self.previous_action = new_actions
idx += 1
# Act
def get_actions(self, terminal_ac, g_dists, dist_matrix):
ids = []
new_actions = {}
state = self.get_state()
normal_state, normal_context = self.normalise_all(
state, terminal_ac, g_dists, dist_matrix)
policy = []
if not len(normal_state) == 0:
policy = self.agent.act(normal_state, normal_context)
return policy, normal_state, normal_context
# For an aircraft perform an action
def perform_action(self, i, action):
if action < 3:
traf_alt = int(traf.alt[i] / ft)
new_alt = int(round((traf_alt + ACTIONS[action])))
alt = max(CONSTRAINTS["alt"]["min"],
min(CONSTRAINTS["alt"]["max"], new_alt))
# print(traf_alt, alt)
stack.stack("{} alt {}".format(traf.id[i], alt))
elif action == 4:
traf_alt = traf.alt[i] / ft
new_alt = int(round((traf_alt)))
# Get the current state
def get_state(self):
state = np.zeros((len(traf.id), 6))
start_ids, end_ids = self.get_all_nodes()
state[:, 0] = traf.lat
state[:, 1] = traf.lon
state[:, 2] = traf.trk
state[:, 3] = traf.alt
state[:, 4] = traf.tas
state[:, 5] = traf.vs
return state
# Get all nodes for each aircraft
def get_all_nodes(self):
start_ids = np.zeros(len(traf.id), dtype=int)
end_ids = np.zeros(len(traf.id), dtype=int)
for i in range(len(traf.id)):
_id = traf.id[i]
route = self.traffic_manager.routes[_id]
start_ids[i] = np.argwhere(
self.route_manager.idx_array == route[0])
end_ids[i] = np.argwhere(self.route_manager.idx_array == route[-1])
return start_ids, end_ids
# Normalise the state and context
def normalise_all(self, state, terminal_ac, g_dists, dist_matrix):
normal_states = self.normalise_state(state, terminal_ac, g_dists)
normal_context = []
start_ids, end_ids = self.get_all_nodes()
max_agents = 0
for _id in traf.id:
if terminal_ac[traf.id2idx(_id)] > 0 or len(
self.traffic_manager.active_sectors[traf.id2idx(
_id)]) <= 0:
continue
new_context = self.normalise_context(_id, terminal_ac, dist_matrix,
start_ids, end_ids)
max_agents = max(max_agents, len(new_context))
if len(normal_context) == 0:
normal_context = new_context
else:
normal_context = np.append(
keras.preprocessing.sequence.pad_sequences(
normal_context, max_agents, dtype='float32'),
keras.preprocessing.sequence.pad_sequences(
new_context, max_agents, dtype='float32'),
axis=0)
if len(normal_context) == 0:
normal_context = np.array([0, 0, 0, 0, 0, 0, 0]).reshape(1, 1, 7)
# print(normal_states.shape, normal_context.shape)
return normal_states, normal_context
# Normalise the agent state only
def normalise_state(self, state, terminal_ac, g_dists):
total_active = 0
for i in range(len(terminal_ac)):
if terminal_ac[i] == 0 and len(
self.traffic_manager.active_sectors[i]) > 0:
total_active += 1
normalised_state = np.zeros((total_active, STATE_SHAPE))
count = 0
for i in range(len(traf.id)):
if terminal_ac[i] > 0 or len(
self.traffic_manager.active_sectors[i]) <= 0:
continue
normalised_state[count, :] = self.normalise(state[i],
'state',
traf.id[i],
g_dist=g_dists[i])
count += 1
return normalised_state
# Get and normalise context
def normalise_context(self, _id, terminal_ac, dist_matrix, start_ids,
end_ids):
context = []
idx = traf.id2idx(_id)
distances = dist_matrix[:, idx]
this_sectors = self.traffic_manager.active_sectors[idx]
this_lat, this_lon = traf.lat[idx], traf.lon[idx]
for i in range(len(distances)):
# Ignore current aircraft
if i == idx:
continue
if terminal_ac[i] > 0 or len(
self.traffic_manager.active_sectors[i]) <= 0:
continue
sectors = self.traffic_manager.active_sectors[i]
# Only care if the ac in a matching sector
flag = False
for x in sectors:
if x in this_sectors:
flag = True
if not flag:
continue
dist = get_dist([this_lat, this_lon], [traf.lat[i], traf.lon[i]])
# Only care about visible distance aircraft
if dist > 40:
continue
spd = traf.tas[i]
alt = traf.alt[i]
trk = traf.trk[i]
vs = traf.vs[i]
start_id = start_ids[i]
end_id = end_ids[i]
self.dist[1] = max(self.dist[1], dist)
self.spd[1] = max(self.spd[1], spd)
self.vs[1] = max(self.vs[1], vs)
dist = dist / self.dist[1]
spd = spd / self.spd[1]
trk = trk / self.trk[1]
alt = ((alt/ft)-CONSTRAINTS["alt"]["min"]) / \
(CONSTRAINTS["alt"]["max"]-CONSTRAINTS["alt"]["min"])
vs = 0
if not vs == 0:
vs = vs / self.vs[1]
n_nodes, dist2next = get_n_nodes(traf.id[i], self.traffic_manager,
self.route_manager)
self.dist[1] = max(self.dist[1], dist2next)
dist2next = dist2next / self.dist[1]
if len(context) == 0:
context = np.array([
spd, alt, trk, vs, dist, dist2next, n_nodes[0], n_nodes[1],
n_nodes[2]
]).reshape(1, 1, 9)
else:
context = np.append(context,
np.array([
spd, alt, trk, vs, dist, dist2next,
n_nodes[0], n_nodes[1], n_nodes[2]
]).reshape(1, 1, 9),
axis=1)
if len(context) == 0:
context = np.zeros(9).reshape(1, 1, 9)
return context
# perform normalisation
def normalise(self, state, what, _id, g_dist=None):
# Normalise the entire state
if what == 'state':
if not g_dist:
raise Exception(
"For normalising a state please pass the distance to the goal."
)
self.dist[1] = max(self.dist[1], g_dist)
self.spd[1] = max(self.spd[1], state[4])
self.vs[1] = max(self.vs[1], state[5])
dist = g_dist / self.dist[1]
spd = state[4] / self.spd[1]
trk = state[2] / self.trk[1]
alt = ((state[3]/ft)-CONSTRAINTS["alt"]["min"]) / \
(CONSTRAINTS["alt"]["max"]-CONSTRAINTS["alt"]["min"])
vs = 0
if not state[5] == 0:
vs = state[5] / self.vs[1]
n_nodes, dist2next = get_n_nodes(_id, self.traffic_manager,
self.route_manager)
self.dist[1] = max(self.dist[1], dist2next)
dist2next = dist2next / self.dist[1]
return np.array([
spd, alt, trk, vs, dist, dist2next, n_nodes[0], n_nodes[1],
n_nodes[2]
])
# Get the terminal aircraft
def get_terminal(self, nearest_ac, g_dists):
terminal_ac = np.zeros(len(traf.id), dtype=int)
terminal_id = []
# Loop through all aircraft
for i in range(len(traf.id)):
# Terminal state 0 = not terminal, 1 = collision, 2 = success
T = 0
# Only care about aircraft in a sector
if len(self.traffic_manager.active_sectors[i]) > 0:
close_ac = nearest_ac[i]
n_ac_data = (close_ac[0], close_ac[1])
# Get the terminal state
T = self.agent.terminal(i, n_ac_data, g_dists[i])
# Only care about terminal aircraft
if not T == 0:
# Update collision aircraft
if T == 1:
terminal_ac[i] = 1
terminal_ac[traf.id2idx(close_ac[2])] = 1
elif not terminal_ac[i] == 1:
terminal_ac[i] = 2
_id = traf.id[i]
self.memory.store(_id, self.last_observation[_id],
self.previous_action[_id], nearest_ac[i],
T)
for i in range(len(terminal_ac)):
if terminal_ac[i] > 0:
terminal_id.append([traf.id[i], terminal_ac[i]])
return terminal_ac, terminal_id
# Handle terminal aircraft
def handle_terminal(self, terminal_id):
for ac in terminal_id:
stack.stack('DEL {}'.format(ac[0]))
self.traffic_manager.active -= 1
if ac[1] == 1:
self.results[1] += 1
elif ac[1] == 2:
self.results[0] += 1
# Generates a distance matrix of all aircraft in the system
def get_dist_martix(self):
size = traf.lat.shape[0]
return geo.latlondist_matrix(np.repeat(traf.lat, size),
np.repeat(traf.lon, size),
np.tile(traf.lat, size),
np.tile(traf.lon,
size)).reshape(size, size)
# Get the nearest aircraft to agents
def get_nearest_ac(self, dist_matrix):
nearest = []
# Loop through all aircraft
for i in range(len(traf.id)):
a_alt = traf.alt[i] / ft
ac_dists = dist_matrix[:, i]
close = 10e+25
alt_sep = 10e+25
nearest_id = None
# Loop through the row on the dist matrix
for x in range(len(ac_dists)):
# Ensure the aircraft is in controlled airspace and not the current aircraft
if not x == i and len(
self.traffic_manager.active_sectors[x]) > 0:
# See if it is closest and update
if ac_dists[x] < close:
close = float(ac_dists[x])
i_alt = traf.alt[x] / ft
alt_sep = abs(a_alt - i_alt)
nearest_id = traf.id[x]
nearest.append([close, alt_sep, nearest_id])
return np.array(nearest)
# returns a matrix of distances to a goal
def get_goal_distances(self):
goal_ds = np.zeros(len(traf.id), dtype=float)
for i in range(len(traf.id)):
goal_ds[i] = get_goal_dist(traf.id[i], self.traffic_manager,
self.route_manager)
return goal_ds
# Reset the environment for the next epoch
def epoch_reset(self):
# Reset the traffic creation
self.traffic_manager.reset()
# Keep track of all success and failures
self.all_success.append(self.results[0])
self.all_fail.append(self.results[1])
# Calcuate total mean success
self.all_mean_success = np.mean(self.all_success)
# Calcuate rolling mean success
if (self.epoch_counter + 1) >= 50:
self.mean_success = np.mean(self.all_success[-50:])
if (self.epoch_counter + 1) % 5 == 0:
if self.mean_success > self.best:
if TRAIN:
print('::::::: Saving Best ::::::')
self.agent.save(path=NEW_FILE + "best.h5")
self.best = self.mean_success
if TRAIN:
print(':::::: Saving Model ::::::')
self.agent.save(path=NEW_FILE + ".h5")
print(":::::::: Training ::::::::")
self.agent.train(self.memory)
print(":::::::: Complete ::::::::")
temp = np.array([np.array(self.all_success), np.array(self.all_fail)])
np.savetxt("Files/" + NEW_FILE + "_numpy.csv", temp, delimiter=',')
# Stop the timer
self.stop = time.perf_counter()
# -------- Printing Outputs --------
string = "Epoch run in {:.2f} seconds".format(self.stop - self.start)
self.print_all(string)
string = "Success: {} | Fail: {} | Mean Success: {:.3f}% | (50) Mean Success Rolling {:.3f}% | Best {:.3f}%".format(
int(self.results[0]), int(self.results[1]),
(self.all_mean_success / MAX_AC) * 100,
(self.mean_success / MAX_AC) * 100, (self.best / MAX_AC) * 100)
self.print_all(string)
string = "Actions -> Descend: {}, Hold Current: {}, Climb: {}, Maintain Climb: {}".format(
self.epoch_actions[0], self.epoch_actions[1],
self.epoch_actions[2], self.epoch_actions[3])
# string = "Actions -> Descend: {}, Climb: {}".format(
# self.epoch_actions[1], self.epoch_actions[0])
self.print_all(string)
if self.epoch_counter + 1 >= EPOCHS:
super_stop = time.perf_counter()
stack.stack("STOP")
string = "::END:: Training {} episodes took {:.2f} hours".format(
EPOCHS, ((super_stop - self.super_start) / 60) / 60)
self.print_all(string)
return
self.epoch_counter += 1
string = "=================================\n UPDATE: RUNNING EPOCH {}\n=================================\n".format(
self.format_epoch())
self.print_all(string)
# Reset values
self.results = np.zeros(2)
self.stop = None
self.start = None
self.mean_rewards = []
self.epoch_actions = []
self.epoch_actions = np.zeros(ACTION_SHAPE)
self.previous_action = {}
self.last_observation = {}
self.observation = {}
# Scripts for printing values
def print_all(self, string):
stack.stack(f'ECHO {string}')
print(string)
def format_epoch(self):
epoch_string = ""
if self.epoch_counter + 1 < 10:
epoch_string += "0"
if self.epoch_counter + 1 < 100:
epoch_string += "0"
if self.epoch_counter + 1 < 1000:
epoch_string += "0"
if self.epoch_counter + 1 < 10000:
epoch_string += "0"
epoch_string += str(self.epoch_counter + 1)
return epoch_string
