def new_neuron():
# 1. Init Config item.
config = {
'session': {
'datapath': 'data/',
'learning_rate': 0.01,
'momentum': 0.9,
'batch_size_train': 64,
'batch_size_test': 64,
'log_interval': 10,
'sync_interval': 100,
'priority_interval': 100,
'name': 'mnist',
'trial_id': '1608070667'
},
'synapse': {
'target_dim': 10
},
'dendrite': {
'key_dim': 100,
'topk': 10,
'stale_emit_filter': 10000,
'pass_gradients': True,
'timeout': 0.5,
'do_backoff': True,
'max_backoff': 100
},
'axon': {
'local_port': 8091,
'external_ip': '191.97.53.53',
'max_workers': 5,
'max_gradients': 1000
},
'nucleus': {
'max_workers': 5,
'queue_timeout': 5,
'queue_maxsize': 1000
},
'metagraph': {
'chain_endpoint': '206.189.254.5:12345',
'stale_emit_filter': 10000
},
'meta_logger': {
'log_dir': 'data/'
},
'neuron': {
'keyfile': None,
'keypair': None
}
}
config = Munch.fromDict(config)
logger.info(Config.toString(config))
mnemonic = Keypair.generate_mnemonic()
keypair = Keypair.create_from_mnemonic(mnemonic)
neuron = Neuron(config)
neuron.keypair = keypair
return neuron
self.tensorboard.add_scalar('Neuron/Rloss',
output.remote_target_loss.item(),
self.global_step)
self.tensorboard.add_scalar('Neuron/Lloss',
output.local_target_loss.item(),
self.global_step)
self.tensorboard.add_scalar('Neuron/Dloss',
output.distillation_loss.item(),
self.global_step)
# ---- Step increments ----
self.global_step += 1
self.training_loss += output.local_target_loss.item()
# --- Memory clean up ----
torch.cuda.empty_cache()
del output
if __name__ == "__main__":
# ---- Config ----
parser = argparse.ArgumentParser()
Session.add_args(parser)
config = Config.to_config(parser)
Session.check_config(config)
logger.info(Config.toString(config))
# ---- Build + Run ----
session = Session(config)
session.run()
def default_config() -> Munch:
parser = argparse.ArgumentParser()
Session.add_args(parser)
config = Config.to_config(parser)
Session.check_config(config)
return config
self.tensorboard.add_scalar('Dloss',
output.distillation_loss.item(),
self.global_step)
# --- Test epoch ----
def test(self):
with torch.no_grad(
): # Turns off gradient computation for inference speed up.
self.model.eval() # Turns off Dropoutlayers, BatchNorm etc.
loss = 0.0
accuracy = 0.0
for _, (images, labels) in enumerate(self.testloader):
# ---- Local Forward pass ----
outputs = self.model.local_forward(
images=images.to(self.device),
targets=torch.LongTensor(labels).to(self.device),
)
loss += outputs.local_target_loss.item()
accuracy += outputs.local_accuracy.item()
return loss / len(self.testloader), accuracy / len(self.testloader)
if __name__ == "__main__":
# ---- Build and Run ----
config = Miner.build_config()
logger.info(Config.toString(config))
miner = Miner(config)
miner.run()
self.neuron.dendrite)
self.tensorboard.add_scalar('Rloss', output.remote_target_loss.item(), self.global_step)
self.tensorboard.add_scalar('Lloss', output.local_target_loss.item(), self.global_step)
self.tensorboard.add_scalar('Dloss', output.distillation_loss.item(), self.global_step)
# --- Test epoch ----
def test (self):
with torch.no_grad(): # Turns off gradient computation for inference speed up.
self.model.eval() # Turns off Dropoutlayers, BatchNorm etc.
loss = 0.0; accuracy = 0.0
for _, (images, labels) in enumerate(self.testloader):
# ---- Local Forward pass ----
outputs = self.model.local_forward(
images = images.to(self.device),
targets = torch.LongTensor(labels).to(self.device),
)
loss += outputs.local_target_loss.item()
accuracy += outputs.local_accuracy.item()
return loss / len(self.testloader), accuracy / len(self.testloader)
if __name__ == "__main__":
# ---- Build and Run ----
config = Session.config(); logger.info(Config.toString(config))
session = Session(config)
session.run()