def fetch_thread(i):
global outf
global grad_q
global b
s3 = boto3.resource('s3')
my_bucket = s3.Bucket('camus-pywren-489')
test = get_test()
num = 0
start_time = time.time()
while time.time() - start_time < total_time:
key = 'gradient_indiv_%d' % i
begin = time.time()
while time.time() - start_time < total_time and (not check_key(key)):
if time.time() - begin > 10:
print("Thread %d took too long" % i)
b.wait()
break
pass
obj = my_bucket.Object('gradient_indiv_%d' % i)
try:
grad = pickle.loads(obj.get()['Body'].read())
except:
continue
grad_q.put(grad)
#model = get_model()
#model = update_model(model, grad)
#store_model(model)
print("Thread %d waiting..." % i)
b.wait()
if num % 10 == 0:
print('[ERROR]', num,
time.time() - start_time, loglikelihood(test, model))
print("Thread %d moving..." % i)
obj.delete()
#if i == 0:
# b.reset()
num += 1
if time.time() - start_time > total_time:
return
def dump_thread(q, f):
global total_time
start = time.time()
print("DUMP THREAD STARTED")
outf = open(fname[:-4] + ".csv2", "w")
testdata = get_test()
while time.time() - start < total_time or not q.empty():
if time.time() - start > total_time and q.empty():
break
if not q.empty():
t, model = q.get()
print("dumping")
s = time.time()
#pickle.dump(time_model, f)
loss = loglikelihood(testdata, model)
print("wrote: %f %f" % (t, loss))
outf.write("%f, %f\n" % (t, loss))
print("dump done took", time.time() - s)
q.task_done()
outf.close()
print("DUMP THREAD STOPPED")
def get_local_test_data():
return get_test()
def error_thread(model, outf):
global grad_q
global log
global fname
global index
s3 = boto3.resource('s3')
my_bucket = s3.Bucket('camus-pywren-489')
num = 0
print("Starting error thread")
start_time = time.time()
# Clear existing gradients
test_data = get_test_data()
saves = 0
if True:
print(fname[:-4] + ".pkl")
f = open(fname[:-4] + ".pkl", 'wb')
time_model_lst = []
last_dump = -100
while time.time() - start_time < total_time:
if not grad_q.empty():
sz = grad_q.qsize()
print("Saw", sz)
grads = []
for _ in range(sz):
grad = grad_q.get()
model = update_model(model, grad)
store_model(model)
grad_q.task_done()
num += 1
#error = loglikelihood(test_data, model)
curr_time = time.time() - start_time
print("[ERROR_TASK]", curr_time, loglikelihood(test_data, model),
"this many grads:", num, "Sec / Grad:",
(time.time() - start_time) / num)
outf.write("[ERROR_TASK] " + str(curr_time) +
" this many grads: " + str(num) + " Sec / Grad: " +
str((time.time() - start_time) / num))
if True and curr_time - last_dump > 1:
print("dumping")
pickle.dump((curr_time, model), f)
print("dump done")
saves += 1
last_dump = curr_time
if time.time() - start_time > total_time:
break
print("Saves: ", saves, "Index:", index)
if True:
large_test = get_test()
f.close()
outf = open(fname[:-4] + ".csv", "w")
with open(fname[:-4] + ".pkl", 'rb') as f:
last = -1000
for i in range(saves):
t, model = pickle.load(f)
if t - last < 5:
continue
last = t
error = loglikelihood(large_test, model)
print("wrote: %f %f" % (t, error))
outf.write("%f, %f\n" % (t, error))
outf.close()
