def prepare_param_exchanger(self):
from base.exchanger import EASGD_Exchanger
self.exchanger = EASGD_Exchanger(self.config, \
self.drv, \
self.model.params, \
etype='server')
def prepare_param_exchanger(self):
from base.exchanger import EASGD_Exchanger
self.exchanger = EASGD_Exchanger(self.config, \
self.drv, \
self.model.params, \
etype='server')
class EASGD_PTServer(PTServer):
'''
Server class based on a specific synchronization rule (EASGD)
Manage traininng ralted request from workers
'''
def __init__(self, port, config, device):
PTServer.__init__(self, port = port, \
config = config, \
device = device)
self.config['irank'] = 0
self.max_mb = self.config['n_epochs'] * len(self.data[0])
print 'max_mb = %d' % self.max_mb
self.uidx = {}
self.valid = {}
self.adj_lr = {}
self.start_time = None
self.last = None
self.last_uidx = 0
self.validFreq = len(self.data[0])
self.uepoch = 0
self.last_uepoch = 0
if self.config['resume_train']:
self.uepoch = self.config['load_epoch']
self.uidx['pretrained'] = self.uepoch * self.validFreq
def prepare_param_exchanger(self):
from base.exchanger import EASGD_Exchanger
self.exchanger = EASGD_Exchanger(self.config, \
self.drv, \
self.model.params, \
etype='server')
def process_request(self, worker_id, message):
# override PTServer class method, for training related request
reply = PTServer.process_request(self, worker_id, message)
if reply != None:
return reply
else:
pass
try:
valid = self.valid['%s' % worker_id]
amount = self.uidx['%s' % worker_id]
adj_lr = self.adj_lr['%s' % worker_id]
except KeyError:
self.valid['%s' % worker_id] = False
self.adj_lr['%s' % worker_id] = False
self.uidx['%s' % worker_id] = 0
self.adj_lr = self.adj_lr.fromkeys(self.adj_lr, True) # when a new worker joins
if message == 'next':
if self.start_time is None:
self.start_time = time.time()
if sum(self.uidx.values()) >= self.max_mb: # stop when finish all epochs
print "[Server] Total training time %.2fh" % ((time.time() - self.start_time)/3600.0)
reply = 'stop'
elif self.valid['%s' % worker_id]:
self.valid['%s' % worker_id] = False
reply = 'val'
elif self.adj_lr['%s' % worker_id]:
self.adj_lr['%s' % worker_id] = False
reply = 'adjust_lr'
else:
reply = 'train'
elif 'done' in message:
self.uidx['%s' % worker_id] += message['done']
#print '[Server] uidx %d' % sum(self.uidx.values())
elif message == 'uepoch':
reply = [self.uepoch, len(self.worker_comm)]
if message in ['next', 'uepoch'] or 'done' in message:
now_uidx = sum(self.uidx.values())
self.uepoch = int(now_uidx/self.validFreq)
if self.last_uepoch != self.uepoch:
#print "[Server] now global epoch %d" % self.uepoch
self.last_uepoch = self.uepoch
self.adj_lr = self.adj_lr.fromkeys(self.adj_lr, True) # when a epoch is finished
#self.valid = self.valid.fromkeys(self.valid, True)
self.valid["%s" % self.first_worker_id] = True # only the first worker validates
# tunning server alpha
a_step1, a_step2, a_step3 = self.config['alpha_step']
if self.uepoch>a_step1 and self.uepoch< a_step2:
step_idx = 1
elif self.uepoch>a_step2 and self.uepoch< a_step3:
step_idx = 2
elif self.uepoch>a_step3:
step_idx = 3
else:
step_idx = 0
self.exchanger.alpha=self.config['server_alpha'] - self.config['alpha_minus']*step_idx
print 'server alpha changed to %f' % self.exchanger.alpha
if self.last == None:
self.last = float(time.time())
if now_uidx - self.last_uidx >= 400:
now = float(time.time())
print '[Server] %d time per 5120 images: %.2f s' % \
(self.uepoch, (now - self.last)/(now_uidx - self.last_uidx)*40.0)
self.last_uidx = now_uidx
self.last = now
return reply
def action_after(self, worker_id, message):
# override PTServer class method, for training related action
PTServer.action_after(self, worker_id, message )
if message in ['connect', 'sync_register']:
self.prepare_param_exchanger()
elif message == 'exchange':
self.exchanger.comm = self.worker_comm[str(worker_id)]
self.exchanger.dest = self.worker_rank[str(worker_id)]
self.exchanger.exchange()
elif message == 'copy_to_local':
self.exchanger.comm = self.worker_comm[str(worker_id)]
self.exchanger.dest = self.worker_rank[str(worker_id)]
self.exchanger.copy_to_local()
class EASGD_PTServer(PTServer):
'''
Server class based on a specific synchronization rule (EASGD)
Manage traininng ralted request from workers
'''
def __init__(self, port, config, device):
PTServer.__init__(self, port = port, \
config = config, \
device = device)
self.config['irank'] = 0
self.max_mb = self.config['n_epochs'] * len(self.data[0])
print 'max_mb = %d' % self.max_mb
self.uidx = {}
self.valid = {}
self.adj_lr = {}
self.start_time = None
self.last = None
self.last_uidx = 0
self.validFreq = len(self.data[0])
self.uepoch = 0
self.last_uepoch = 0
if self.config['resume_train']:
self.uepoch = self.config['load_epoch']
self.uidx['pretrained'] = self.uepoch * self.validFreq
def prepare_param_exchanger(self):
from base.exchanger import EASGD_Exchanger
self.exchanger = EASGD_Exchanger(self.config, \
self.drv, \
self.model.params, \
etype='server')
def process_request(self, worker_id, message):
# override PTServer class method, for training related request
reply = PTServer.process_request(self, worker_id, message)
if reply != None:
return reply
else:
pass
try:
valid = self.valid['%s' % worker_id]
amount = self.uidx['%s' % worker_id]
adj_lr = self.adj_lr['%s' % worker_id]
except KeyError:
self.valid['%s' % worker_id] = False
self.adj_lr['%s' % worker_id] = False
self.uidx['%s' % worker_id] = 0
self.adj_lr = self.adj_lr.fromkeys(self.adj_lr,
True) # when a new worker joins
if message == 'next':
if self.start_time is None:
self.start_time = time.time()
if sum(self.uidx.values()
) >= self.max_mb: # stop when finish all epochs
print "[Server] Total training time %.2fh" % (
(time.time() - self.start_time) / 3600.0)
reply = 'stop'
elif self.valid['%s' % worker_id]:
self.valid['%s' % worker_id] = False
reply = 'val'
elif self.adj_lr['%s' % worker_id]:
self.adj_lr['%s' % worker_id] = False
reply = 'adjust_lr'
else:
reply = 'train'
elif 'done' in message:
self.uidx['%s' % worker_id] += message['done']
#print '[Server] uidx %d' % sum(self.uidx.values())
elif message == 'uepoch':
reply = [self.uepoch, len(self.worker_comm)]
if message in ['next', 'uepoch'] or 'done' in message:
now_uidx = sum(self.uidx.values())
self.uepoch = int(now_uidx / self.validFreq)
if self.last_uepoch != self.uepoch:
#print "[Server] now global epoch %d" % self.uepoch
self.last_uepoch = self.uepoch
self.adj_lr = self.adj_lr.fromkeys(
self.adj_lr, True) # when a epoch is finished
#self.valid = self.valid.fromkeys(self.valid, True)
self.valid[
"%s" % self.
first_worker_id] = True # only the first worker validates
# tunning server alpha
a_step1, a_step2, a_step3 = self.config['alpha_step']
if self.uepoch > a_step1 and self.uepoch < a_step2:
step_idx = 1
elif self.uepoch > a_step2 and self.uepoch < a_step3:
step_idx = 2
elif self.uepoch > a_step3:
step_idx = 3
else:
step_idx = 0
self.exchanger.alpha = self.config[
'server_alpha'] - self.config['alpha_minus'] * step_idx
print 'server alpha changed to %f' % self.exchanger.alpha
if self.last == None:
self.last = float(time.time())
if now_uidx - self.last_uidx >= 400:
now = float(time.time())
print '[Server] %d time per 5120 images: %.2f s' % \
(self.uepoch, (now - self.last)/(now_uidx - self.last_uidx)*40.0)
self.last_uidx = now_uidx
self.last = now
return reply
def action_after(self, worker_id, message):
# override PTServer class method, for training related action
PTServer.action_after(self, worker_id, message)
if message in ['connect', 'sync_register']:
self.prepare_param_exchanger()
elif message == 'exchange':
self.exchanger.comm = self.worker_comm[str(worker_id)]
self.exchanger.dest = self.worker_rank[str(worker_id)]
self.exchanger.exchange()
elif message == 'copy_to_local':
self.exchanger.comm = self.worker_comm[str(worker_id)]
self.exchanger.dest = self.worker_rank[str(worker_id)]
self.exchanger.copy_to_local()
class EASGD_PTServer(PTServer):
"""
Server class based on a specific synchronization rule (EASGD)
Manage traininng related request from workers
This class reimplements the "controller" in platoon. See the original implementation here:
https://github.com/mila-udem/platoon/blob/master/platoon/channel.py
"""
def __init__(self, port, config, device):
PTServer.__init__(self, port=port, config=config, device=device)
self.config["irank"] = 0
self.max_mb = self.config["n_epochs"] * len(self.data[0])
print "max_mb = %d" % self.max_mb
self.uidx = {}
self.valid = {}
self.adj_lr = {}
self.start_time = None
self.last = None
self.last_uidx = 0
self.validFreq = len(self.data[0])
self.uepoch = 0
self.last_uepoch = 0
if self.config["resume_train"]:
self.uepoch = self.config["load_epoch"]
self.uidx["pretrained"] = self.uepoch * self.validFreq
def prepare_param_exchanger(self):
from base.exchanger import EASGD_Exchanger
self.exchanger = EASGD_Exchanger(self.config, self.drv, self.model.params, etype="server")
def process_request(self, worker_id, message):
# override PTServer class method, for training related request
reply = PTServer.process_request(self, worker_id, message)
if reply != None:
return reply
else:
pass
try:
valid = self.valid["%s" % worker_id]
amount = self.uidx["%s" % worker_id]
adj_lr = self.adj_lr["%s" % worker_id]
except KeyError:
self.valid["%s" % worker_id] = False
self.adj_lr["%s" % worker_id] = False
self.uidx["%s" % worker_id] = 0
self.adj_lr = self.adj_lr.fromkeys(self.adj_lr, True) # when a new worker joins
if message == "next":
if self.start_time is None:
self.start_time = time.time()
if sum(self.uidx.values()) >= self.max_mb: # stop when finish all epochs
print "[Server] Total training time %.2fh" % ((time.time() - self.start_time) / 3600.0)
reply = "stop"
elif self.valid["%s" % worker_id]:
self.valid["%s" % worker_id] = False
reply = "val"
elif self.adj_lr["%s" % worker_id]:
self.adj_lr["%s" % worker_id] = False
reply = "adjust_lr"
else:
reply = "train"
elif "done" in message:
self.uidx["%s" % worker_id] += message["done"]
# print '[Server] uidx %d' % sum(self.uidx.values())
elif message == "uepoch":
reply = [self.uepoch, len(self.worker_comm)]
if message in ["next", "uepoch"] or "done" in message:
now_uidx = sum(self.uidx.values())
self.uepoch = int(now_uidx / self.validFreq)
if self.last_uepoch != self.uepoch:
# print "[Server] now global epoch %d" % self.uepoch
self.last_uepoch = self.uepoch
self.adj_lr = self.adj_lr.fromkeys(self.adj_lr, True) # when a epoch is finished
# self.valid = self.valid.fromkeys(self.valid, True)
self.valid["%s" % self.first_worker_id] = True # only the first worker validates
# tunning server alpha
a_step1, a_step2, a_step3 = self.config["alpha_step"]
if self.uepoch > a_step1 and self.uepoch < a_step2:
step_idx = 1
elif self.uepoch > a_step2 and self.uepoch < a_step3:
step_idx = 2
elif self.uepoch > a_step3:
step_idx = 3
else:
step_idx = 0
self.exchanger.alpha = self.config["server_alpha"] - self.config["alpha_minus"] * step_idx
print "server alpha changed to %f" % self.exchanger.alpha
if self.last == None:
self.last = float(time.time())
if now_uidx - self.last_uidx >= 400:
now = float(time.time())
print "[Server] %d time per 5120 images: %.2f s" % (
self.uepoch,
(now - self.last) / (now_uidx - self.last_uidx) * 40.0,
)
self.last_uidx = now_uidx
self.last = now
return reply
def action_after(self, worker_id, message):
# override PTServer class method, for training related action
PTServer.action_after(self, worker_id, message)
if message in ["connect", "sync_register"]:
self.prepare_param_exchanger()
elif message == "exchange":
self.exchanger.comm = self.worker_comm[str(worker_id)]
self.exchanger.dest = self.worker_rank[str(worker_id)]
self.exchanger.exchange()
elif message == "copy_to_local":
self.exchanger.comm = self.worker_comm[str(worker_id)]
self.exchanger.dest = self.worker_rank[str(worker_id)]
self.exchanger.copy_to_local()