def run_ps(trainers):
utils.timed_log("Start training")
ps_rref = rpc.RRef(bups.BatchUpdateParameterServer())
futs = []
for trainer in trainers:
futs.append(rpc.rpc_async(trainer, run_trainer, args=(ps_rref, )))
torch.futures.wait_all(futs)
utils.timed_log("Finish training")
def core(config):
for attr in initialize_dag(config):
attr = dict(config, **attr)
attr = get_raw_table(attr)
if ((attr.get('raw_table') is None) or # skip if table was not selected in config file
(attr.get('pass') == True)): # skip if pass == True in dependency graph
with timed_log(name=attr['name'], config=config, time_chunk='seconds', force='skiped'):
continue
with timed_log(name=attr['name'], config=config, time_chunk='seconds', force=attr['force']):
run_process(attr)
def update_and_fetch_model(ps_rref, grads, rewards):
self = ps_rref.local_value()
utils.timed_log(
f"PS got {self.curr_update_size}/{self.batch_update_size} updates")
for p, g in zip(self.model.parameters(), grads):
p.grad += g
for reward in rewards:
self.current_rewards.append(reward)
self.curr_update_size += 1
with self.lock:
fut = self.future_model
if self.curr_update_size >= self.batch_update_size:
for p in self.model.parameters():
p.grad /= self.batch_update_size
self.curr_update_size = 0
self.optimizer.step()
self.optimizer.zero_grad()
fut.set_result(self.model)
utils.timed_log(
f"rewards length is {len(self.current_rewards)}")
utils.timed_log(
f"average reward is {np.mean(self.current_rewards)}")
utils.timed_log("PS updated model")
self.future_model = torch.futures.Future()
#save the model
cwd = os.getcwd()
parameter_file = C.trained_model_name
cwd = os.path.join(cwd, parameter_file)
torch.save(self.model.state_dict(), cwd)
#record optimization history
self.optimization_history.append(np.mean(self.current_rewards))
optimization_history = {}
optimization_history['history'] = self.optimization_history
#store the history
cwd = os.getcwd()
#cwd = os.path.join(cwd, 'data_folder')
parameter_file = 'optimization_history.json'
cwd = os.path.join(cwd, parameter_file)
with open(cwd, 'w') as statusFile:
statusFile.write(jsonpickle.encode(optimization_history))
self.current_rewards = []
return fut
def single(self, config_path='configs/config-template.yaml'):
"""Run the pipeline for a single region
Parameters
----------
config_path : str, optional
config.yaml path, by default 'configs/config.yaml'
"""
config = get_config(config_path)
config.update(self.args)
if self.verbose:
print(config)
with timed_log(name='Start process', config=config, time_chunk='minutes',
force=config['force']):
pass
with timed_log(name='Full process', config=config, time_chunk='minutes',
force=config['force']):
core(config)
def single(self, config_path="configs/config-template.yaml"):
"""Run the pipeline for a single region
Parameters
----------
config_path : str, optional
config.yaml path, by default 'configs/config.yaml'
"""
config = get_config(config_path)
config.update(self.args)
if self.verbose:
print(config)
with timed_log(
name="Start process",
config=config,
time_chunk="minutes",
force=config["force"],
):
pass
with timed_log(
name="Full process",
config=config,
time_chunk="minutes",
force=config["force"],
):
if config["slug"] == "prod":
try:
core(config)
except:
if config["post_log"]:
logger.post(traceback.format_exc())
print(traceback.format_exc())
else:
core(config)
def core(config):
for attr in initialize_dag(config):
attr = dict(config, **attr)
attr = get_raw_table(attr)
if (attr.get("raw_table") is
None) or ( # skip if table was not selected in config file
attr.get("pass")
== True): # skip if pass == True in dependency graph
with timed_log(name=attr["name"],
config=config,
time_chunk="seconds",
force="skiped"):
continue
with timed_log(name=attr["name"],
config=config,
time_chunk="seconds",
force=attr["force"]):
run_process(attr)
def update_and_fetch_model(ps_rref, grads, rewards):
self = ps_rref.local_value()
utils.timed_log(
f"PS got {self.curr_update_size}/{self.batch_update_size} updates")
for p, g in zip(self.model.parameters(), grads):
p.grad += g
for reward in rewards:
self.current_rewards.append(reward)
self.curr_update_size += 1
with self.lock:
fut = self.future_model
if self.curr_update_size >= self.batch_update_size:
for p in self.model.parameters():
p.grad /= self.batch_update_size
self.curr_update_size = 0
self.optimizer.step()
self.optimizer.zero_grad()
fut.set_result(self.model)
utils.timed_log(
f"rewards length is {len(self.current_rewards)}")
utils.timed_log(
f"average reward is {np.mean(self.current_rewards)}")
utils.timed_log("PS updated model")
self.future_model = torch.futures.Future()
self.current_rewards = []
#save the model
cwd = os.getcwd()
parameter_file = 'cartpole_rpc_trained_model.pt'
cwd = os.path.join(cwd, parameter_file)
torch.save(self.model.state_dict(), cwd)
return fut
