class ChatMessageStore:
def __init__(self):
self.__id = 1
self.__storage = []
self.__rwlock = RWLock()
def add_message(self, message):
self.__rwlock.acquire_write()
self.__storage.append((self.__id, message))
self.__id += 1
self.__rwlock.release()
def get_messages(self, start_id=0):
self.__rwlock.acquire_read()
result = ([message for id_, message in self.__storage if id_ >= start_id], self.__id)
self.__rwlock.release()
return result
class TweetWorker(object):
def __init__(self, n, callback):
self.processing_users = []
self.lock = RWLock()
self.jobs = ThreadQueue.Queue()
self.results = ThreadQueue.Queue()
self.processes = []
for _ in range(0, n):
s = Thread(target=worker, args=(job_queue, self.results))
self.processes.append(s)
s.daemon = True
s.start()
print('Started {} worker processes'.format(len(self.processes)))
self.consumer = Thread(target=consumer_process,
args=(self.results, callback))
self.consumer.daemon = True
self.consumer.start()
print('Started consumer process')
def get(self, username):
if username is None:
return
self.lock.acquire_read()
if username in self.processing_users:
self.lock.release()
return
self.lock.release()
self.lock.acquire_write()
self.processing_users.append(username)
self.lock.release()
job_queue.put(username)
return
class TweetWorker(object):
def __init__(self, n, callback):
self.processing_users = []
self.lock = RWLock()
self.jobs = ThreadQueue.Queue()
self.results = ThreadQueue.Queue()
self.processes = []
for _ in range(0, n):
s = Thread(target=worker, args=(job_queue, self.results))
self.processes.append(s)
s.daemon = True
s.start()
print('Started {} worker processes'.format(len(self.processes)))
self.consumer = Thread(target=consumer_process, args=(self.results, callback))
self.consumer.daemon = True
self.consumer.start()
print('Started consumer process')
def get(self, username):
if username is None:
return
self.lock.acquire_read()
if username in self.processing_users:
self.lock.release()
return
self.lock.release()
self.lock.acquire_write()
self.processing_users.append(username)
self.lock.release()
job_queue.put(username)
return
class DeepDist:
def __init__(self, model, master='127.0.0.1:5000', min_updates=0, max_updates=4096):
"""DeepDist - Distributed deep learning.
:param model: provide a model that can be trained in parallel on the workers
"""
self.model = model
self.lock = RWLock()
self.descent = lambda model, gradient: model
self.master = master
self.state = 'serving'
self.served = 0
self.received = 0
#self.server = None
self.pmodel = None
self.min_updates = min_updates
self.max_updates = max_updates
def __enter__(self):
Thread(target=self.start).start()
# self.server = Process(target=self.start)
# self.server.start()
return self
def __exit__(self, type, value, traceback):
# self.server.terminate()
pass # need to shut down server here
def start(self):
from flask import Flask, request
app = Flask(__name__)
@app.route('/')
def index():
return 'DeepDist'
@app.route('/model', methods=['GET', 'POST', 'PUT'])
def model_flask():
i = 0
while (self.state != 'serving' or self.served >= self.max_updates) and (i < 1000):
time.sleep(1)
i += 1
# pickle on first read
pmodel = None
self.lock.acquire_read()
if not self.pmodel:
self.lock.release()
self.lock.acquire_write()
if not self.pmodel:
print(self.model)
self.pmodel = pickle.dumps(self.model, -1)
self.served += 1
pmodel = self.pmodel
self.lock.release()
else:
self.served += 1
pmodel = self.pmodel
self.lock.release()
return pmodel
@app.route('/update', methods=['GET', 'POST', 'PUT'])
def update_flask():
gradient = pickle.loads(request.data)
self.lock.acquire_write()
if self.min_updates <= self.served:
state = 'receiving'
self.received += 1
self.descent(self.model, gradient)
if self.received >= self.served and self.min_updates <= self.received:
self.received = 0
self.served = 0
self.state = 'serving'
self.pmodel = None
self.lock.release()
return 'OK'
print 'Listening to 0.0.0.0:5000...'
app.run(host='0.0.0.0', debug=True, threaded=True, use_reloader=False)
def train(self, rdd, gradient, descent):
master = self.master # will be pickled
if master == None:
master = rdd.ctx._conf.get('spark.master')
if master.startswith('local['):
master = 'localhost:5000'
else:
if master.startswith('spark://'):
master = '%s:5000' % urlparse.urlparse(master).netloc.split(':')[0]
else:
master = '%s:5000' % master.split(':')[0]
print '\n*** Master: %s\n' % master
self.descent = descent
def mapPartitions(data):
#a = fetch_model(master=master)
#print(a)
aa = gradient(fetch_model(master=master), data)
bb = pickle.dumps(aa)
#print aa
return [send_gradient(gradient(fetch_model(master=master), data), master=master)]
return rdd.mapPartitions(mapPartitions).collect()
class DeepDist:
def __init__(self,
model,
master='127.0.0.1:5000',
min_updates=0,
max_updates=4096):
"""DeepDist - Distributed deep learning.
:param model: provide a model that can be trained in parallel on the workers
"""
self.model = model
self.lock = RWLock()
self.descent = lambda model, gradient: model
self.master = master
self.state = 'serving'
self.served = 0
self.received = 0
#self.server = None
self.pmodel = None
self.min_updates = min_updates
self.max_updates = max_updates
print("THIS IS THE MASTER")
print(self.master)
print("\n")
print("THIS IS THE MODEL 1")
print(self.model)
print("\n")
def __enter__(self):
Thread(target=self.start).start()
# self.server = Process(target=self.start)
# self.server.start()
return self
def __exit__(self, type, value, traceback):
# self.server.terminate()
pass # need to shut down server here
def start(self):
from flask import Flask, request
app = Flask(__name__)
@app.route('/')
def index():
return 'DeepDist'
@app.route('/model', methods=['GET', 'POST', 'PUT'])
def model_flask():
i = 0
while (self.state != 'serving'
or self.served >= self.max_updates) and (i < 1000):
time.sleep(1)
i += 1
# pickle on first read
pmodel = None
self.lock.acquire_read()
if not self.pmodel:
self.lock.release()
self.lock.acquire_write()
if not self.pmodel:
self.pmodel = pickleDumper.dumps(self.model, -1)
self.served += 1
pmodel = self.pmodel
self.lock.release()
else:
self.served += 1
pmodel = self.pmodel
self.lock.release()
return pmodel
@app.route('/update', methods=['GET', 'POST', 'PUT'])
def update_flask():
gradient = pickle.loads(request.data)
self.lock.acquire_write()
if self.min_updates <= self.served:
state = 'receiving'
self.received += 1
self.descent(self.model, gradient)
if self.received >= self.served and self.min_updates <= self.received:
self.received = 0
self.served = 0
self.state = 'serving'
self.pmodel = None
self.lock.release()
return 'OK'
print 'Listening to 0.0.0.0:5000...'
app.run(host='0.0.0.0', debug=True, threaded=True, use_reloader=False)
def train(self, rdd, gradient, descent):
master = self.master # will be pickled
print("MASTER ROUND 2")
print(master)
if master == None:
master = rdd.ctx._conf.get('spark.master')
if master.startswith('local['):
master = 'localhost:5000'
else:
if master.startswith('spark://'):
master = '%s:5000' % urlparse.urlparse(master).netloc.split(
':')[0]
else:
master = '%s:5000' % master.split(':')[0]
print '\n*** Master: %s\n' % master
self.descent = descent
def mapPartitions(data):
return [
send_gradient(gradient(fetch_model(master=master), data),
master=master)
]
return rdd.mapPartitions(mapPartitions).collect()
class DeepDist:
def __init__(self,
model,
master='127.0.0.1:5000',
min_updates=0,
max_updates=4096):
"""DeepDist - Distributed deep learning.
:param model: provide a model that can be trained in parallel on the workers
"""
self.model = model
self.lock = RWLock()
self.descent = lambda model, gradient: model
self.master = master
self.state = 'serving'
self.served = 0
self.received = 0
self.server = '0.0.0.0'
self.pmodel = None
self.min_updates = min_updates
self.max_updates = max_updates
def __enter__(self):
Thread(target=self.start).start()
# self.server = Process(target=self.start)
# self.server.start()
return self
def __exit__(self, type, value, traceback):
url = "http://%s/shutdown" % self.master
response = urllib2.urlopen(url, '{}').read()
print("Exit requested...")
def start(self):
from flask import Flask, request
app = Flask(__name__)
@app.route('/')
def index():
return 'DeepDist'
@app.route('/model', methods=['GET', 'POST', 'PUT'])
def model_flask():
i = 0
while (self.state != 'serving'
or self.served >= self.max_updates) and (i < 1000):
time.sleep(1)
i += 1
# pickle on first read
pmodel = None
self.lock.acquire_read()
if not self.pmodel:
self.lock.release()
self.lock.acquire_write()
if not self.pmodel:
self.pmodel = pickleDumper.dumps(self.model, -1)
self.served += 1
pmodel = self.pmodel
self.lock.release()
else:
self.served += 1
pmodel = self.pmodel
self.lock.release()
return pmodel
@app.route('/update', methods=['GET', 'POST', 'PUT'])
def update_flask():
gradient = pickle.loads(request.data)
self.lock.acquire_write()
if self.min_updates <= self.served:
state = 'receiving'
self.received += 1
self.descent(self.model, gradient)
if self.received >= self.served and self.min_updates <= self.received:
self.received = 0
self.served = 0
self.state = 'serving'
self.pmodel = None
self.lock.release()
return 'OK'
@app.route('/shutdown', methods=['POST'])
def shutdown():
func = request.environ.get('werkzeug.server.shutdown')
if func is None:
raise RuntimeError('Not running with the Werkzeug Server')
func()
return 'Server shutting down...'
print 'Listening to 0.0.0.0:5000...'
app.run(host='0.0.0.0', debug=True, threaded=True, use_reloader=False)
def train(self, rdd, gradient, descent):
master = self.master # will be pickled
if master == None:
master = rdd.ctx._conf.get('spark.master')
if master.startswith('local['):
master = 'localhost:5000'
else:
if master.startswith('spark://'):
master = '%s:5000' % urlparse.urlparse(master).netloc.split(
':')[0]
else:
master = '%s:5000' % master.split(':')[0]
print '\n*** Master: %s\n' % master
self.descent = descent
def mapPartitions(data):
return [
send_gradient(gradient(fetch_model(master=master), data),
master=master)
]
return rdd.mapPartitions(mapPartitions).collect()
class DeepDist:
def __init__(self,
model,
master='127.0.0.1:5000',
min_updates=0,
max_updates=4096):
"""DeepDist - Distributed deep learning.
:param model: provide a model that can be trained in parallel on the workers
"""
self.model = model
self.lock = RWLock()
self.descent = lambda model, gradient: model
self.master = master
self.state = 'serving'
self.served = 0
self.received = 0
#self.server = None
self.pmodel = None
self.min_updates = min_updates
self.max_updates = max_updates
def start_server(self):
Thread(target=self.start).start()
def start(self):
from flask import Flask, request
app = Flask(__name__)
@app.route('/')
def index():
return 'DeepDist'
@app.route('/model', methods=['GET', 'POST', 'PUT'])
def model_flask():
i = 0
while (self.state != 'serving'
or self.served >= self.max_updates) and (i < 1000):
time.sleep(1)
i += 1
# pickle on first read
pmodel = None
self.lock.acquire_read()
if not self.pmodel:
self.lock.release()
self.lock.acquire_write()
if not self.pmodel:
self.pmodel = pickle.dumps(self.model, -1)
self.served += 1
pmodel = self.pmodel
self.lock.release()
else:
self.served += 1
pmodel = self.pmodel
self.lock.release()
print "model replica weights were updated via /model"
return pmodel
@app.route('/update', methods=['GET', 'POST', 'PUT'])
def update_flask():
gradient = pickle.loads(request.data)
self.lock.acquire_write()
if self.min_updates <= self.served:
state = 'receiving'
self.received += 1
old_syn0, old_syn1 = self.model.syn0.copy(), self.model.syn1.copy()
print "received gradient: " + str(gradient)
self.descent(self.model, gradient)
if self.received >= self.served and self.min_updates <= self.received:
self.received = 0
self.served = 0
self.state = 'serving'
self.pmodel = None
self.lock.release()
print "server weights were updated by model replica"
print "old weights: "
print old_syn0[0:3, 0:3], old_syn1[
0:3, 0:3] #printing just the first few weights
print "new weights: "
print self.model.syn0[0:3, 0:3], self.model.syn1[0:3, 0:3]
return 'OK'
print 'Listening to 0.0.0.0:5000...'
app.run(host='0.0.0.0', debug=True, threaded=True, use_reloader=False)
def train(self, rdd, gradient, descent):
master = self.master
print '\n*** Master: %s\n' % master
self.descent = descent
def mapPartitions(data):
return [
send_gradient(gradient(fetch_model(master=master), data),
master=master)
]
return rdd.mapPartitions(mapPartitions).collect()
class Volumn(object):
_rpc_methods = [
'assign_volumn', 'store', 'replica', 'download', 'status', 'balance',
'migrate_volumn_to', 'migrate_volumn_from', 'delete_file',
'delete_volumn'
]
def __init__(self, logger, host, port):
self.logger = logger
self.host = host
self.port = port
self.lock = RWLock()
self.vdb = dict()
self.fdb = dict()
if os.path.isfile('vdb'):
self.vdb = pickle.load(open('vdb', 'rb'))
if os.path.isfile('fdb'):
self.fdb = pickle.load(open('fdb', 'rb'))
self.act_mst_serv = list()
self.serv = ThreadXMLRPCServer((self.host, self.port),
logRequests=True)
for name in self._rpc_methods:
self.serv.register_function(getattr(self, name))
def _update_vdb(self):
pickle.dump(self.vdb, open('vdb', 'wb'))
def _update_fdb(self):
pickle.dump(self.fdb, open('fdb', 'wb'))
def update_master(self, masters):
self.act_mst_serv = masters
def get_master(self):
return ServerProxy(random.choice(self.act_mst_serv))
def assign_volumn(self, vid, size):
path = 'data/%s' % vid
if not os.path.isdir('data'):
os.mkdir('data')
try:
self.lock.acquire_read()
with open(path, 'wb') as f:
f.seek(size - 1)
f.write(b'\0')
vdoc = dict()
vdoc['vid'] = vid
vdoc['path'] = path
vdoc['size'] = size
vdoc['counter'] = 0
self.vdb[vid] = vdoc
self._update_vdb()
return True
except:
return False
finally:
self.lock.release()
def migrate_volumn_to(self, vid, to_addr):
try:
vdoc = self.vdb[vid]
path = vdoc['path']
s = ServerProxy(to_addr)
with open(path, 'rb') as f:
while True:
data = f.read(64 * 1024 * 1024)
if data:
s.migrate_volumn_from(vid, data, vdoc)
self.logger.info('Send data...')
else:
fdocs = {
k: v
for k, v in self.fdb.items()
if k.startswith('%d,' % vid)
}
self.logger.info('Send metadata')
s.migrate_volumn_from(vid, data, vdoc, fdocs, True)
break
return True
except:
return False
def migrate_volumn_from(self, vid, data, vdoc, fdocs=None, done=False):
path = vdoc['path']
if not os.path.isdir('data'):
os.mkdir('data')
if done:
self.vdb[vid] = vdoc
self._update_vdb()
self.fdb = {**self.fdb, **fdocs}
self._update_fdb()
self.logger.info('Build replica %d success' % vid)
else:
with open(path, 'ab') as f:
f.write(data.data)
return True
def store(self, fid, data):
vid, _ = fid.split(',')
vid = int(vid)
try:
self.replica(fid, data)
master = self.get_master()
volumns = master.find_writable_volumn(vid)
if not volumns:
return False
for volumn in volumns:
if volumn != 'http://%s:%d' % (self.host, self.port):
s = ServerProxy(volumn)
s.replica(fid, data)
return True
except Exception as e:
self.logger.exception('Got an exception')
return False
def replica(self, fid, data):
data = data.data
vid, _ = fid.split(',')
vid = int(vid)
self.lock.acquire_write()
vdoc = self.vdb[vid]
path = vdoc['path']
offset = vdoc['counter']
size = len(data)
vdoc['counter'] += size
self.lock.release()
with open(path, 'r+b') as f:
f.seek(offset)
f.write(data)
fdoc = dict()
fdoc['fid'] = fid
fdoc['offset'] = offset
fdoc['size'] = size
fdoc['delete'] = False
self.vdb[vid] = vdoc
self._update_vdb()
self.fdb[fid] = fdoc
self._update_fdb()
return True
def update_file(self, fid, data):
pass
def delete_file(self, fid, sync=True):
vid, _ = fid.split(',')
vid = int(vid)
try:
if sync:
master = self.get_master()
volumns = master.find_writable_volumn(vid)
if not volumns:
return False
for volumn in volumns:
if volumn != 'http://%s:%d' % (self.host, self.port):
s = ServerProxy(volumn)
s.delete_file(fid, False)
fdoc = self.fdb[fid]
fdoc['delete'] = True
self._update_fdb()
return True
except Exception as e:
self.logger.exception('Got an exception')
return False
def delete_volumn(self, vid):
pass
def download(self, fid):
vid, _ = fid.split(',')
vid = int(vid)
if vid not in self.vdb or fid not in self.fdb:
return None
try:
self.lock.acquire_read()
vdoc = self.vdb[vid]
fdoc = self.fdb[fid]
if fdoc['delete'] == True:
return None
path = vdoc['path']
offset = fdoc['offset']
size = fdoc['size']
with open(path, 'rb') as f:
f.seek(offset)
data = f.read(size)
return data
except:
return None
finally:
self.lock.release()
def balance(self, vid):
try:
self.lock.acquire_write()
vdoc = self.vdb[vid]
fdocs = self.fdb
tfdocs = fdocs.copy()
tvdoc = vdoc.copy()
tvdoc['counter'] = 0
path = vdoc['path']
size = vdoc['size']
with open(path + '.tmp', 'wb') as f:
f.seek(size - 1)
f.write(b'\0')
with open(path, 'r+b') as from_file, open(path + '.tmp',
'r+b') as to_file:
to_file.seek(0)
for fdoc in fdocs.values():
if fdoc['fid'].startswith(
'%d,' % vid) and fdoc['delete'] == False:
from_file.seek(fdoc['size'])
data = from_file.read(fdoc['size'])
to_file.write(data)
tfdoc = fdoc.copy()
tfdoc['offset'] = tvdoc['counter']
tvdoc['counter'] += fdoc['size']
tfdocs[fdoc['fid']] = tfdoc
os.remove(path)
os.rename(path + '.tmp', path)
self.vdb[vid] = tvdoc
self.fdocs = tfdocs
return True
except:
self.logger.exception('Got an exception')
return False
finally:
self.lock.release()
def status(self):
status = dict()
total, used, free = shutil.disk_usage(__file__)
status['total'] = str(total)
status['used'] = str(used)
status['free'] = str(free)
status['vdb'] = {str(vid): vdoc for vid, vdoc in self.vdb.items()}
return status
def start(self):
self.logger.info('Start serving at %s:%d' % (self.host, self.port))
self.serv.serve_forever()
class MSGControler:
#def __init__(self, listen_ip, listen_port, log_file, debug):
def __init__(self, configDic):
"""
Init Controler class
"""
self.running = True
self.config = configDic
self.log_file = self.config["log"]
self.debug = self.config["debug"]
self.security = self.config["security"]
self.clients = {}
self.lock = RWLock()
## Start TCP comm server ##
listen_ip = self.config["listen_ip"]
listen_port = self.config["listen_port"]
try:
self.server = tcpServer((listen_ip,listen_port), handleConnection, self.clients, self.debug, self.security )
except:
self.log_error("Unable to bind TCP socket %s:%s !" % (listen_ip,listen_port))
proc = subprocess.Popen(["ss", "-pant"], stdout=subprocess.PIPE)
code = proc.wait()
for aline in proc.stdout:
if (str(listen_ip)+':'+str(listen_port)) in aline and "LISTEN" in aline:
tmpstr1 = re.sub(').*', '', re.sub('.*(', '', aline))
pid = re.sub(',.*', '', re.sub('.*pid=', tmpstr1))
prog = re.sub('.*"', '', re.sub('",.*', '', aline))
self.log_warning("Process %s, PID %s, is binding port %s. It will be killed." % (prog, pid, listen_port))
os.system("kill -9 %s" % pid)
time.sleep(10)
self.log_info("Trying again to bind %s on %s." % (listen_port, listen_ip))
self.server = tcpServer((listen_ip,listen_port), handleConnection, self.clients, self.debug, self.security )
self.comm_thread = threading.Thread(target=self.server.serve_forever)
self.comm_thread.daemon = True
self.comm_thread.start()
##### Send a keepalive message every minutes (60 sec) ##
self.keepalive = KeepAliveTimer(60, self.send_keepalive, ["KeepAliveTimer"])
self.keepalive.start()
def log_error(self, newline):
self.log(newline, "ERROR")
def log_warning(self, newline):
self.log(newline, "WARNING")
def log_info(self, newline):
self.log(newline, "INFO")
def log_event(self, newline):
self.log(newline, "EVENT")
def log_debug(self, newline):
if self.debug == True :
self.log(newline, "DEBUG")
def log(self, newline, level="INFO"):
LOG_SIZE = os.path.getsize(self.log_file)
# if > 1M create a new file
if LOG_SIZE > 1000000:
if os.path.exists(self.log_file+".4"):
os.remove(self.log_file+".4")
os.rename(self.log_file+".3", self.log_file+".4")
if os.path.exists(self.log_file+".3"):
os.rename(self.log_file+".3", self.log_file+".4")
if os.path.exists(self.log_file+".2"):
os.rename(self.log_file+".2", self.log_file+".3")
if os.path.exists(self.log_file+".1"):
os.rename(self.log_file+".1", self.log_file+".2")
os.rename(self.log_file, self.log_file+".1")
if os.path.exists('/opt/virtualisation/openkvi/debug'):
os.remove('/opt/virtualisation/openkvi/debug')
logs = open(self.log_file,'w')
else:
logs = open(self.log_file,'a')
timestamp = time.strftime("%Y-%m-%d %H:%M:%S")
logs.write(timestamp+"::["+level+"]::"+newline+"\n")
logs.close()
def print_debug(self, msg):
if self.debug == True :
self.log_debug(msg)
print msg
def tell_all(self, event, data):
self.keepalive.stop()
self.print_debug("telling all %s %s"% (event, data))
line = event+";"+json.dumps(data)
## Acquire lock so that no messages are sent
## simultanously
self.lock.acquire_write()
res = self.server.writeToAll(line)
## Wait 500 ms between two message to prevent
## clients being overwhelmed
time.sleep(0.5)
self.lock.release()
self.keepalive.start()
def stop(self):
self.print_debug("stop tcp server")
self.keepalive.stop()
self.server.socket.close()
def send_keepalive(self):
res = self.server.writeToAll("keep alive")
class tcpServer(SocketServer.ThreadingTCPServer):
def __init__(self, server_address, RequestHandlerClass, clients, debug, security):
self.allow_reuse_address = True
SocketServer.ThreadingTCPServer.__init__(self,server_address,RequestHandlerClass)
self.clients = clients
#self.arg2 = arg2
self.rwlock = RWLock()
self.debug = debug
self.security = security
def print_debug(self, msg):
if self.debug == True :
print msg
def writeToAll(self, line):
lines = []
lines.append(line)
msg = {}
msg['messages'] = lines
try:
self.rwlock.acquire_read()
keys = self.clients.keys()
self.rwlock.release()
except:
err = str(sys.exc_info()[1]).strip("'")
return 0
for an_id in keys:
self.clients[an_id]['lock'].acquire_write()
conn = self.clients[an_id]['connection']
try :
#self.print_debug("trying to say %s to %s" % (line.strip(), an_id))
conn.wfile.write(json.dumps(msg)+"\n")
self.clients[an_id]['lock'].release()
except:
self.print_debug("Not able to speak to %s" % an_id)
self.clients[an_id]['lock'].release()
timestamp = self.clients[an_id]["timestamp"]
if timestamp != 0:
timeout = datetime.datetime.now() - timestamp
if timeout > datetime.timedelta(seconds = 20):
self.print_debug("connection to %s timed out !" % an_id)
try:
conn.finish()
except:
self.print_debug("connection to %s is already finished" % an_id)
del self.clients[an_id]
if self.security != "low":
#curl -k https://localhost/_auth/unset/?id=12345"
url = 'https://localhost/_auth/unset/?id='+an_id
proc = subprocess.Popen(['curl','-k',url], stdout=subprocess.PIPE)
code = proc.wait()
else:
self.clients[an_id]['lock'].acquire_write()
self.clients[an_id]['timestamp'] = datetime.datetime.now()
self.clients[an_id]['messages'].append(line)
self.clients[an_id]['lock'].release()
return 0
class DeepDist:
def __init__(self, model, master='127.0.0.1:5000', min_updates=0, max_updates=4096):
"""DeepDist - Distributed deep learning.
:param model: provide a model that can be trained in parallel on the workers
"""
self.model = model
self.lock = RWLock()
self.descent = lambda model, gradient: model
self.master = master
self.state = 'serving'
self.served = 0
self.received = 0
#self.server = None
self.pmodel = None
self.min_updates = min_updates
self.max_updates = max_updates
def start_server(self):
Thread(target=self.start).start()
def start(self):
from flask import Flask, request
app = Flask(__name__)
@app.route('/')
def index():
return 'DeepDist'
@app.route('/model', methods=['GET', 'POST', 'PUT'])
def model_flask():
i = 0
while (self.state != 'serving' or self.served >= self.max_updates) and (i < 1000):
time.sleep(1)
i += 1
# pickle on first read
pmodel = None
self.lock.acquire_read()
if not self.pmodel:
self.lock.release()
self.lock.acquire_write()
if not self.pmodel:
self.pmodel = pickle.dumps(self.model, -1)
self.served += 1
pmodel = self.pmodel
self.lock.release()
else:
self.served += 1
pmodel = self.pmodel
self.lock.release()
print "model replica weights were updated via /model"
return pmodel
@app.route('/update', methods=['GET', 'POST', 'PUT'])
def update_flask():
gradient = pickle.loads(request.data)
self.lock.acquire_write()
if self.min_updates <= self.served:
state = 'receiving'
self.received += 1
old_syn0, old_syn1 = self.model.syn0.copy(), self.model.syn1.copy()
print "received gradient: " + str(gradient)
self.descent(self.model, gradient)
if self.received >= self.served and self.min_updates <= self.received:
self.received = 0
self.served = 0
self.state = 'serving'
self.pmodel = None
self.lock.release()
print "server weights were updated by model replica"
print "old weights: "
print old_syn0[0:3, 0:3], old_syn1[0:3, 0:3] #printing just the first few weights
print "new weights: "
print self.model.syn0[0:3, 0:3], self.model.syn1[0:3, 0:3]
return 'OK'
print 'Listening to 0.0.0.0:5000...'
app.run(host='0.0.0.0', debug=True, threaded=True, use_reloader=False)
def train(self, rdd, gradient, descent):
master = self.master
print '\n*** Master: %s\n' % master
self.descent = descent
def mapPartitions(data):
return [send_gradient(gradient(fetch_model(master=master), data), master=master)]
return rdd.mapPartitions(mapPartitions).collect()
class DeepDist:
def __init__(self, model, master="127.0.0.1:5000", min_updates=0, max_updates=4096):
"""DeepDist - Distributed deep learning.
:param model: provide a model that can be trained in parallel on the workers
"""
self.model = model
self.lock = RWLock()
self.descent = lambda model, gradient: model
self.master = master
self.state = "serving"
self.served = 0
self.received = 0
# self.server = None
self.pmodel = None
self.min_updates = min_updates
self.max_updates = max_updates
def __enter__(self):
Thread(target=self.start).start()
# self.server = Process(target=self.start)
# self.server.start()
return self
def __exit__(self, type, value, traceback):
# self.server.terminate()
pass # need to shut down server here
def start(self):
from flask import Flask, request
app = Flask(__name__)
@app.route("/")
def index():
return "DeepDist"
@app.route("/model", methods=["GET", "POST", "PUT"])
def model_flask():
i = 0
while (self.state != "serving" or self.served >= self.max_updates) and (i < 1000):
time.sleep(1)
i += 1
# pickle on first read
pmodel = None
self.lock.acquire_read()
if not self.pmodel:
self.lock.release()
self.lock.acquire_write()
if not self.pmodel:
self.pmodel = pickle.dumps(self.model, -1)
self.served += 1
pmodel = self.pmodel
self.lock.release()
else:
self.served += 1
pmodel = self.pmodel
self.lock.release()
return pmodel
@app.route("/update", methods=["GET", "POST", "PUT"])
def update_flask():
gradient = pickle.loads(request.data)
self.lock.acquire_write()
if self.min_updates <= self.served:
state = "receiving"
self.received += 1
self.descent(self.model, gradient)
if self.received >= self.served and self.min_updates <= self.received:
self.received = 0
self.served = 0
self.state = "serving"
self.pmodel = None
self.lock.release()
return "OK"
print "Listening to 0.0.0.0:5000..."
app.run(host="0.0.0.0", debug=True, threaded=True, use_reloader=False)
def train(self, rdd, gradient, descent):
master = self.master # will be pickled
if master == None:
master = rdd.ctx._conf.get("spark.master")
if master.startswith("local["):
master = "localhost:5000"
else:
if master.startswith("spark://"):
master = "%s:5000" % urlparse.urlparse(master).netloc.split(":")[0]
else:
master = "%s:5000" % master.split(":")[0]
print "\n*** Master: %s\n" % master
self.descent = descent
def mapPartitions(data):
return [send_gradient(gradient(fetch_model(master=master), data), master=master)]
return rdd.mapPartitions(mapPartitions).collect()
class DeepDist:
def __init__(self, model, batch=None, master='127.0.0.1:5000'):
"""DeepDist - Distributed deep learning.
:param model: provide a model that can be trained in parallel on the workers
"""
self.model = model
self.lock = RWLock()
self.descent = lambda model, gradient: model
self.master = master
self.state = 'serving'
self.served = 0
self.received = 0
self.batch = batch
self.server = None
def __enter__(self):
Thread(target=self.start).start()
# self.server = Process(target=self.start)
# self.server.start()
return self
def __exit__(self, type, value, traceback):
# self.server.terminate()
pass # need to shut down server here
def start(self):
from flask import Flask, request
app = Flask(__name__)
@app.route('/')
def index():
return 'DeepDist'
@app.route('/model', methods=['GET', 'POST', 'PUT'])
def model_flask():
i = 0
while (self.state != 'serving') and (i < 1000):
time.sleep(1)
i += 1
self.lock.acquire_read()
self.served += 1
model = copy.deepcopy(self.model)
self.lock.release()
return pickle.dumps(model, -1)
@app.route('/update', methods=['GET', 'POST', 'PUT'])
def update_flask():
gradient = pickle.loads(request.data)
self.lock.acquire_write()
state = 'receiving'
self.received += 1
self.descent(self.model, gradient)
if self.received >= self.served:
self.received = 0
self.served = 0
self.state = 'serving'
self.lock.release()
return 'OK'
print 'Listening to 0.0.0.0:5000...'
app.run(host='0.0.0.0', debug=True, threaded=True, use_reloader=False)
def train(self, rdd, gradient, descent):
master = self.master # will be pickled
print 'master0: ', master
if master == None:
master = rdd.ctx._conf.get('spark.master')
print 'master1: ', master
if master.startswith('local['):
master = 'localhost:5000'
else:
if master.startswith('spark://'):
master = '%s:5000' % urlparse.urlparse(master).netloc.split(':')[0]
else:
master = '%s:5000' % master.split(':')[0]
print '\n*** master: %s\n' % master
self.descent = descent
batch = self.batch
def mapPartitions(data):
last = 'dummy'
class Iter:
def __iter__(self):
self.i = 0
return self
def next(self):
if (batch == None) or (self.i < batch):
self.i += 1
last = data.next()
return last
else:
return None
res = []
while last != None:
res.append(send_gradient(gradient(fetch_model(master=master), Iter()), master=master))
return res
return rdd.mapPartitions(mapPartitions).collect()
class DeepDist:
def __init__(self, model, host='127.0.0.1:5000'):
"""DeepDist - Distributed deep learning.
:param model: provide a model that can be trained in parallel on the workers
"""
self.model = model
self.lock = RWLock()
self.descent = lambda model, gradient: model
self.host = host
self.state = 'serving'
self.served = 0
self.received = 0
def __enter__(self):
Thread(target=self.start).start()
return self
def __exit__(self, type, value, traceback):
pass # need to shut down server here
def start(self):
from flask import Flask, request
app = Flask(__name__)
@app.route('/')
def index():
return 'DeepDist'
@app.route('/model', methods=['GET', 'POST', 'PUT'])
def model_flask():
i = 0
while (self.state != 'serving') and (i < 20):
time.sleep(1)
i += 1
self.lock.acquire_read()
self.served += 1
model = copy.deepcopy(self.model)
self.lock.release()
return pickle.dumps(model, -1)
@app.route('/update', methods=['GET', 'POST', 'PUT'])
def update_flask():
gradient = pickle.loads(request.data)
self.lock.acquire_write()
state = 'receiving'
self.received += 1
self.descent(self.model, gradient)
if self.received >= self.served:
self.received = 0
self.served = 0
self.state = 'serving'
self.lock.release()
return 'OK'
print 'Listening to 0.0.0.0:5000...'
app.run(host='0.0.0.0', debug=True, threaded=True, use_reloader=False)
def train(self, rdd, gradient, descent):
self.descent = descent
host = self.host # will be pickled by rdd.mapPartitions
def mapPartitions(data):
return (send_gradient(gradient(fetch_model(host=host), data), host=host))
return rdd.mapPartitions(mapPartitions).collect()
