def connect(self, TIMEOUT):
if debug_level >= DEBUG_LEVEL.INFO:
TERM.write_info(
'Attempting to connect to {} (Timeout: {}s)'.format(
self.server, TIMEOUT))
# Retry to connect to the server
start = time.time()
while (time.time() - start) < TIMEOUT:
self.attempt_to_connect(TIMEOUT)
if self.connected:
break
# Run the client (if connection succesful)
if not self.connected:
if debug_level >= DEBUG_LEVEL.INFO:
TERM.write_failure(
'Time limit exceeded: {} not found'.format(SERVER))
else:
if debug_level >= DEBUG_LEVEL.INFO:
TERM.write_success('Successfully connected to {} as {}'.format(
SERVER, self.sock.getsockname()))
self.run()
def train(self):
while len(self.connected_clients_by_addr) > 0:
# select a subset of the clients and broadcast the model.
if debug_level >= DEBUG_LEVEL.INFO:
TERM.write_info("Selecting clients...")
self.select_clients(self.subset_size)
if debug_level >= DEBUG_LEVEL.INFO:
TERM.write_info("Broadcasting model...")
self.broadcast_model()
if debug_level >= DEBUG_LEVEL.INFO:
TERM.write_info("Waiting for updates...(Timeout: " +
str(self.TIMEOUT) + ")")
start = time.time()
# wait for each client's update and then aggregate.
while (self.aggregated_update is
None) and time.time() - start < self.TIMEOUT:
self.wait_for_updates()
self.attempt_to_aggregate_updates()
# if an aggregated update has been created...
if self.aggregated_update is not None:
# Stop communication (temporarily)
self.stop()
if debug_level >= DEBUG_LEVEL.INFO:
TERM.write_success("All updates received.")
TERM.write_info("Aggregating updates...")
# Update the model using aggregated update.
self.update_model(self.aggregated_update)
self.aggregated_update = None
# reset the selected client address list (to be re-selected)
self.selected_clients_by_addr = {}
self.selected_clients_updates = {}
if debug_level >= DEBUG_LEVEL.INFO:
TERM.write_info("Sending aggregated update to clients...")
# Restart communication.
self.start()
else:
if debug_level >= DEBUG_LEVEL.INFO:
TERM.write_warning(
"Time-limit exceeded: Some updates were not received.")
if debug_level >= DEBUG_LEVEL.INFO:
TERM.write_failure("All clients disconected.")
def listen_for_clients(self):
# TODO: make number bigger/variable
self.listener_sock.listen(5)
while self.listening:
# Accept connection.
client_sock, client_addr = self.listener_sock.accept()
if debug_level >= DEBUG_LEVEL.INFO:
TERM.write_success(
'Established connection to {}'.format(client_addr))
# cache the client socket.
self.connected_clients_by_addr[client_addr] = client_sock
self.connected_clients_by_sock[client_sock] = client_addr
def run(self):
while True:
# Wait for weights from the FLServer
if debug_level >= DEBUG_LEVEL.INFO:
TERM.write_info('Waiting for model from server...(Timeout: ' +
str(self.TIMEOUT) + ')')
start_time = time.time()
weights = None
while ((weights == None)
and (time.time() - start_time < self.TIMEOUT)):
# get the weights.
weights = Communication_Handler.recv_msg(self.sock)
if weights == None:
if debug_level >= DEBUG_LEVEL.INFO:
TERM.write_warning(
"Time Limit Exceeded: Weights not received.")
else:
if debug_level >= DEBUG_LEVEL.INFO:
TERM.write_success("Weights received.")
TERM.write_info("Training local model...")
# Load weights
self.trainer.load_weights(weights)
# Train model
self.trainer.train()
if debug_level >= DEBUG_LEVEL.INFO:
TERM.write_success("Training complete.")
TERM.write_info("Sending update to server...")
# Compute focused update
update = self.trainer.focused_update()
# Send update to the server
Communication_Handler.send_msg(self.sock, update)
if debug_level >= DEBUG_LEVEL.INFO:
TERM.write_success("Update sent.")