def _init_dask(self):
if self._cluster is not None:
print('WARNING: reinitiailizing dask')
self._cluster.close()
if self._client is not None:
self._client.close()
del self._dask_futures
if self.pub_out is not None:
del self.pub_out
if self.sub_in is not None:
del self.sub_in
self._cluster = LocalCluster(
n_workers=self.num_processes,
# silence_logs=0,
memory_limit=None)
self._client = Client(self._cluster)
# always define publishers first then subscribers
pub_out = [
Pub('env{}_input'.format(env_idx))
for env_idx in range(self.num_processes)
]
self._dask_futures = self._client_map()
sub_in = Sub('observations')
self.pub_out = pub_out
self.sub_in = sub_in
# wait until all the peers are created
time.sleep(5)
def init_workers(env_name, num_processes):
cluster = LocalCluster(n_workers=num_processes)
client = Client(cluster)
pubs_reset = [
Pub('env{}_reset'.format(seed)) for seed in range(num_processes)
]
client.map(run_env, [env_name] * num_processes, range(num_processes))
sub_obs = Sub('observations')
# sleep while sub/pub is initialized
time.sleep(5)
return client, pubs_reset, sub_obs
def worker2():
pub_try = Pub('lets')
sub_try = Sub('receive')
pub_try.put("I can send...")
try:
s = sub_try.get()
print("Let see what i get,", s)
except TimeoutError:
print("I CANT")
return
def __init__(self, args, obs_running_stats=None):
self.parse_args(args)
torch.set_num_threads(1)
# create variables
self.env = self.create_env(args['seed'])
self.pub_out = Pub('observations')
self.sub_in = Sub('env{}_input'.format(int(self.env_idx)))
self.step_counter = 0
self.step_counter_after_new_action = 0
self.reset_env(reset_mime=False)
# start the environment loop
self.env_loop()
def run_env(env_name, seed):
try:
import gym
import mime
env = gym.make(env_name)
env.seed(seed)
pub_obs = Pub('observations')
sub_reset = Sub('env{}_reset'.format(int(seed)))
print('seed {} ready'.format(seed))
for counter in sub_reset:
frames, scalars = get_frames_scalars(env, counter)
pub_obs.put((frames, scalars, seed))
except Exception as e:
print('Exeception: {}'.format(e))
def coordinator(clf,e,n_minibatch,total_workers):
pub_results = Pub('results')
pub_init = Pub('Initialize')
pub_th = Pub('Theta')
pub_endr = Pub('EndRound')
pub_endsub = Pub('EndSubRound')
#pub_ask_state = Pub('AskState')
sub_incr = Sub('Increment')
sub_f = Sub('Fs')
sub_x = Sub('Xs')
# sub_try=Sub('lets')
# pub_try=Pub('receive')
# get increments from workers
def get_incr():
try:
incr=sub_incr.get(timeout=5)
print("Coo Received increments...",incr)
if incr<0: # works as a flag to let coordinator know that chunks are out
print("Coo received notice of chunks ended...")
return incr
except TimeoutError:
return 0
# get fi's from all workers
def get_fi(n_workers):
fis=[]
print("try to get fis workers:",n_workers)
for i in range(n_workers):
try:
fi=sub_f.get(timeout=5)
print("Coo received",i+1,"fi")
fis.append(fi)
except TimeoutError:
print('Fis Lost worker/workers num=',len(fis))
break
return fis
# get xi's from all workers
def get_xi(n_workers):
drifts=[]
print("try to get xi workers:",n_workers)
for i in range(n_workers):
try:
xi=sub_x.get(timeout=6)
print("Coo received",i+1,"xi")
drifts.append(xi)
except TimeoutError:
print('Lost worker/workers')
break
print("Num of workers",len(drifts))
return drifts
def check_subcribers(pub,n_workers):
print("Check...")
if n_workers==0:
print("No workers left")
return "end"
while len(pub.subscribers)<n_workers: #if not all workers subscribe sleep
time.sleep(0.01)
print("OK Check")
return "ok"
#____________________________Start coordinator_________________________________
E=None
th=0
fis=0
drifts=0
sum_xi=0
incr=0
e_y=0.01
workers=[]
time_stamb=0
n_rounds=0
print("Coo started ...")
client= get_client()
for i in range(len(total_workers)-1):
workers.append(client.submit(worker_f,i,clf,n_minibatch,e,workers=total_workers[i+1]))
time.sleep(1)
flag=True #use this flag to finish future if chunks are out
start_time=time.time()
while flag==True:
n_subs=0
workers_status=[w.status for w in workers]
k=workers_status.count('pending')
print("NUMBER OF WORKERS...",k)
if E is None: #if E=0 we need to update E
pub_init.put(None)
print("Warmup...Coo Sended E=0...")
drifts=get_xi(k) #get local drifts (Xi's)
print("Coo received xi's...workers=",k)
sum_xi=add_x(drifts)
e1=sum_xi[0]/len(drifts)
e2=sum_xi[1]/len(drifts)
E=[e1,e2]
pub_init.put(E)
print("Coo Sended E")
else:
pub_init.put(E)
print("Coo Sended E")
n_rounds+=1
y=k*f([[0],0],E,e)
barrier=e_y*k*f([[0],0],E,e)
#start of the round...
print("START ROUND:",n_rounds," workers ",k)
while y<=barrier:
th=-y/(2*k)
pub_th.put(th) #send theta
print("Coo Sended theta")
n_subs+=1
print("START SUBROUND:",n_subs," workers ",k)
c=0
fis=[]
#start of the subround...
while c<k:
incr=get_incr() #Get increments
if incr<0: # works as a flag to let coordinator know that chunks are out
incr=0
workers_status=[w.status for w in workers]
k=workers_status.count('pending')
if k==0:
flag=False
c=c+incr
#subrounds ended...
pub_endsub.put(0) #let workers know that subrounds ended
print("Coo Sended endofSub... num_workers",k)
workers_status=[w.status for w in workers]
k=workers_status.count('pending')
fis=get_fi(k) #get F(Xi)'s from workers
if len(fis)==0:
pub_endr.put(0)
break
print("Coo Received fi's workers=",k)
y=add_f(fis)
print("y",y)
if flag==False: #if false chunks are out end future
print("Coo Sended endofSub..")
break
#rounds ended...
pub_endr.put(0) #let workers know that rounds ended
print("Coo Sended endofround... num_workers",k)
drifts=get_xi(len(fis)) #get local drifts (Xi's)
print("len of drifts",len(drifts))
print("Coo Received xi's workers=",k)
if len(drifts)==0: break
sum_xi=add_x(drifts)
e1=E[0]+(sum_xi[0]/len(drifts)) #len(drifts)
e2=E[1]+(sum_xi[1]/len(drifts)) #len(drifts)
E=[e1,e2]
time_stamb=time.time()-start_time
pub_results.put([E,n_subs,k,time_stamb])
if flag==False:
break
print("Coo ended...")
return E,n_rounds,n_subs,k,time_stamb
def main(client, w, new, dataset_params, e, chunks, n_minibatch):
print("-------------------------------------------------------------\n")
print("Start with num of chunks:", chunks, ",e:", e)
E = None
Acc = []
worker = []
n_rounds = 0
time_stamps = []
total_time = []
total_acc = []
sub_results = Sub('results')
#make a dataset and save training X and y ,give sample number and future number
if new == 'yes':
create_dataset(dataset_params, chunks, w)
X_test = np.load("np_arrays/X_test.npy")
y_test = np.load("np_arrays/y_test.npy")
size = dataset_params["n_samples"] - len(X_test)
print("Minibatch size:", size / (chunks * n_minibatch))
clf = linear_model.SGDClassifier(shuffle=False)
clf_results = [clf] * (len(w) - 1)
start_time = time.time()
start_rounds = 0
for p in range(1):
s_run_time = time.time()
random_assign(len(w) - 1, chunks)
# for i in range(2):
# worker.append(client.submit(worker_f,i,clf_results[i],n_minibatch,e,workers=w[i+1]))
#TAG Changed the call of random assign + added a for loop + return the clf and feed it again + print after finished
coo = client.submit(coordinator, clf, e, n_minibatch, w, workers=w[0])
print("In progress...")
# acc=pred(E,clf,X_test,y_test)
# Acc.append(acc)
while coo.status == 'pending':
try:
results = sub_results.get(timeout=0.01)
E = results[0]
time_stamps.append(results[3])
print(results[1:])
acc = pred(E, clf, X_test, y_test)
Acc.append(acc)
except TimeoutError:
continue
#return
# t_run_time=time.time()
# start_rounds=n_rounds
# print("End of chunks...")
# status_l=[w.status for w in worker]
# clf_results=[x.result() for x in worker]
# print("Coordinator:",coo.status,"...\nWorkers:",status_l)
# if check_coo(coo)=="ok":
# print("Finished ",p," pass with no error...\n\n")
# del coo
# for f in worker: del f
# worker=[]
# total_time.append(t_run_time-s_run_time)
# total_acc.append(Acc[-1])
# name1="np_arrays/total_time"+str(len(w)-1)
# name2="np_arrays/total_acc"+str(len(w)-1)
# time.sleep(5)
# np.save(name1,total_time)
# np.save(name2,total_acc)
return Acc, len(Acc), time_stamps
def worker_f(name, clf, parts, e):
sub_init = Sub('Initialize')
sub_th = Sub('Theta')
sub_endr = Sub('EndRound')
sub_endsub = Sub('EndSubRound')
pub_incr = Pub('Increment')
pub_f = Pub('Fs')
pub_x = Pub('Xs')
# get initial E value from coordinator
def get_init():
w_id = get_worker().name
try:
print(w_id, "waits to receive E...")
init = sub_init.get(timeout=20)
print(w_id, "Received E")
return init
except TimeoutError:
print(w_id, 'Error E not received')
return False
#get theta from cordinator
def get_th():
w_id = get_worker().name
try:
print(w_id, "waits to receive th...")
th = sub_th.get(timeout=1)
print(w_id, "Received theta")
return th
except TimeoutError:
print(w_id, 'Theta aknowlegment not received')
return None
#get aknowlegment for continue or stop the rounds
def get_endr():
try:
endr = sub_endr.get(timeout=1)
print(w_id, 'End of round received')
return endr
except TimeoutError:
return None
#get aknowlegment for continue or stop the subrounds
def get_endsub():
try:
endsub = sub_endsub.get(timeout=1)
print(w_id, 'End of subround received')
return endsub
except TimeoutError:
return None
# ____Start of worker____
th = 0
w_id = get_worker().name #get worker id
print("worker", w_id, "started...")
flag = True
E = [[0], 0]
Si = [0, 0]
S_prev = [0, 0]
Xi = [[0], 0]
count_chunks = 0
minibatches = 0
#TAG chunks assigned and load first one
X_chunk_array, y_chunk_array = load_chunks(
name) #get the array with the chunk names assigned to this worker
X_chunk, y_chunk = load_np(X_chunk_array, y_chunk_array, count_chunks)
count_chunks += 1
while flag == True: #while this flag stays true there are chunks
E = get_init() # get E from coordinator
if E is False:
pub_incr.put(-1)
return clf
if E is None: #if E=0 compute Xi and return Xi to update E
#TODO make it prettier
print(w_id, "Warmup....")
temp = get_minibatch(X_chunk, y_chunk, minibatches,
parts) #get_newSi(count_chunks,f_name)
if temp is None:
minibatches = 0
load = load_np(X_chunk_array, y_chunk_array, count_chunks)
if load is None:
print(w_id, "End of chunks")
flag = False
pub_incr.put(-1)
break
X_chunk, y_chunk = load
count_chunks += 1
temp = get_minibatch(X_chunk, y_chunk, minibatches, parts)
minibatches += 1
X, y = temp
clf.partial_fit(X, y, np.unique(([0, 1])))
Si = [clf.coef_[0], clf.intercept_[0]]
Xi = [clf.coef_[0], clf.intercept_[0]]
while len(pub_x.subscribers) != 1:
time.sleep(0.01)
pub_x.put(Xi)
print(w_id, "Sended Xi")
E = get_init() # get E from coordinator
if E is False:
pub_incr.put(-1)
break
print(w_id, "Start of round")
clf.coef_[0] = E[0]
clf.intercept_[0] = E[1]
S_prev[0] = np.array(list(E[0]))
S_prev[1] = E[1]
Xi = [[0], 0]
#begin of round...
#FIXME do not send message every time & check rounds and subrounds
while get_endr() == None:
ci = 0
# Xi=[[0],0]
th = get_th() #get theta
if th == None:
print("nonreceive")
continue
print(w_id, "Received start of subround")
#begin of subround...
while get_endsub() == None:
zi = f(Xi, E, e)
temp = get_minibatch(X_chunk, y_chunk, minibatches, parts)
while temp is None:
load = load_np(X_chunk_array, y_chunk_array, count_chunks)
if load is None:
print(w_id, "End of chunks")
flag = False
break
X_chunk, y_chunk = load
count_chunks += 1
minibatches = 0
temp = get_minibatch(X_chunk, y_chunk, minibatches, parts)
if flag == False:
break
else:
minibatches += 1
X, y = temp
clf.partial_fit(X, y, np.unique([0, 1]))
Si[0] = clf.coef_[0]
Si[1] = clf.intercept_[0]
Xi = [Si[0] - S_prev[0], Si[1] - S_prev[1]]
c_th = 0
if th != 0: #avoid division with 0 if th=0 c_th=0
c_th = (f(Xi, E, e) - zi) / th
ci_new = max(ci, math.floor(c_th))
if ci != ci_new: #if we detect a difference send it to the coordinator
incr = ci_new - ci
pub_incr.put(incr)
ci = ci_new
print(w_id, "Sended...", incr)
while len(pub_f.subscribers) != 1:
time.sleep(0.01)
pub_f.put(f(Xi, E, e))
print(w_id, "Sended Fi")
print(w_id, "End of subround")
if flag == False:
break
#end of subround...
if all([v == 0 for v in Xi[0]]):
print(w_id, "ZERO XI")
else:
pub_x.put(Xi) # send Xi
print(w_id, "Sended Xi")
Xi = [[0], 0]
if flag == False:
break
# pub_incr.put(-1)
print(w_id, "Ended...")
return clf