def worker_process(rank):
from learning_objective.hidden_function import evaluate, get_settings
lim_domain = get_settings(lim_domain_only=True)
while True:
comm.send("WORKER is ready", dest=0, tag=WORKER_READY) # tell Master node that I need a new query
query = comm.recv(source=0, tag=MPI.ANY_TAG, status=status)
tag = status.Get_tag()
if tag == SEND_WORKER:
# string = "WORKER %3d: The query is: " % rank
# print string + str(query)
result = evaluate(query, lim_domain)
comm.send(result, dest=0, tag=WORKER_DONE)
elif tag == EXIT_WORKER:
# Worker dies!
print "WORKER: Worker %2d exiting" % rank
break
comm.send(None, dest=0, tag=EXIT_WORKER) # Suicide complete
def worker_process(rank):
from learning_objective.hidden_function import evaluate, get_settings
lim_domain = get_settings(lim_domain_only=True)
while True:
comm.send("WORKER is ready", dest=0,
tag=WORKER_READY) # tell Master node that I need a new query
query = comm.recv(source=0, tag=MPI.ANY_TAG, status=status)
tag = status.Get_tag()
if tag == SEND_WORKER:
# string = "WORKER %3d: The query is: " % rank
# print string + str(query)
result = evaluate(query, lim_domain)
comm.send(result, dest=0, tag=WORKER_DONE)
elif tag == EXIT_WORKER:
# Worker dies!
print "WORKER: Worker %2d exiting" % rank
break
comm.send(None, dest=0, tag=EXIT_WORKER) # Suicide complete
"""
import time
from learning_objective.hidden_function import evaluate, true_evaluate, get_settings
import matplotlib.pyplot as plt
import utilities.optimizer as op
import numpy as np
# Open file to write times for comparison
file_record = open("data/seq_time_data.csv", "a")
# Freeze plotting
plot_it = False
print_statements = False
# Get settings relevant to the hidden function being used
lim_domain, init_size, additional_query_size, init_query, domain, selection_size = get_settings()
# Construct the dataset
dataset = evaluate(init_query[0,:], lim_domain)
print "Randomly query a set of initial points... ",
for query in init_query[1:,:]:
dataset = np.concatenate((dataset, evaluate(query, lim_domain)), axis=0)
print dataset.shape
print "Complete initial dataset acquired"
# Begin sequential optimization using NN-LR based query system
optimizer = op.Optimizer(dataset, domain)
optimizer.train()
def master_process(print_statements):
file_record = open("data/mpi_time_data.csv", "a") # record times for experiment
from learning_objective.hidden_function import evaluate, true_evaluate, get_settings
import random
import utilities.optimizer as op
print "MASTER: starting with %d workers" % (size - 1)
# Setup
t1 = time.time() # Get amount of time taken
num_workers = size - 1 # Get number of workers
closed_workers = 0 # Get number of workers EXIT'ed
# Get settings relevant to the hidden function being used
lim_domain, init_size, additional_query_size, init_query, domain, selection_size = get_settings()
# Acquire an initial data set
dataset = None
init_assigned = 0 # init query counters
init_done = 0
print "Randomly query a set of initial points... ",
while init_done < init_size:
# Get a worker (trainer does not initiate conversation with master)
data = comm.recv(source=MPI.ANY_SOURCE, tag=MPI.ANY_TAG, status=status)
source = status.Get_source()
tag = status.Get_tag()
if tag == WORKER_READY:
if init_assigned < init_size:
# Send a (m,) array query to worker
comm.send(init_query[init_assigned, :], dest=source, tag=SEND_WORKER)
init_assigned += 1
else:
# No more intial work available. Give random work
comm.send(domain[random.choice(range(domain.shape[0])), :],
dest=source, tag=SEND_WORKER)
if tag == WORKER_DONE:
# data is a (1, m) array
if dataset == None:
dataset = data
else:
dataset = np.concatenate((dataset, data), axis=0)
if contains_row(data[0, :-1], init_query):
init_done += 1
if print_statements:
string1 = "MASTER: Init queries done: %3d. " % init_done
string2 = "Submission from WORKER %2d is: " % source
print string1 + string2 + str(data)
print "Complete initial dataset acquired"
# NN-LR based query system
optimizer = op.Optimizer(dataset, domain)
optimizer.train()
# Select a series of points to query
selected_points = optimizer.select_multiple(selection_size) # (#points, m) array
# Set counters
listen_to_trainer = True
trainer_is_ready = True # Determines if trainer will be used
trainer_index = 0 # Keeps track of data that trainer doesn't have
selection_index = 0 # Keeps track of unqueried selected_points
queries_done = 0 # Keeps track of total queries done
queries_total = additional_query_size
t0 = time.time()
print "Performing optimization..."
while closed_workers < num_workers:
if selection_index == selection_size:
# Update optimizer's dataset and retrain LR
optimizer.retrain_LR()
selected_points = optimizer.select_multiple(selection_size) # Select new points
selection_size = selected_points.shape[0] # Get number of selected points
selection_index = 0 # Restart index
if queries_done < queries_total and trainer_is_ready and (dataset.shape[0] - trainer_index - 1) >= 100:
# Trainer ready and enough new data for trainer to train a new NN.
if print_statements:
print "MASTER: Trainer has been summoned"
comm.send(dataset[trainer_index: -1, :], dest=TRAINER, tag=SEND_TRAINER)
trainer_index = dataset.shape[0] - 1
trainer_is_ready = not trainer_is_ready # Trainer is not longer available.
if queries_done >= queries_total and trainer_is_ready:
comm.send("MASTER has fired Trainer", dest=TRAINER, tag=EXIT_TRAINER)
# Check for data from either worker or trainer
data = comm.recv(source=MPI.ANY_SOURCE, tag=MPI.ANY_TAG, status=status)
source = status.Get_source()
tag = status.Get_tag()
if tag == WORKER_READY:
if queries_done < queries_total:
comm.send(selected_points[selection_index, :],
dest=source, tag=SEND_WORKER)
selection_index += 1
else:
comm.send(None, dest=source, tag=EXIT_WORKER)
elif tag == WORKER_DONE:
dataset = np.concatenate((dataset, data), axis=0) # data is (m, 1) array
optimizer.update_data(data) # add data to optimizer
queries_done += 1
if print_statements:
string1 = "MASTER: Queries done: %3d. " % queries_done
string2 = "Submission from Worker %2d: " % source
print string1 + string2 + str(data)
else:
# Print some sort of progress:
if queries_done % (queries_total/10) == 0:
print "%.3f completion..." % ((1.*queries_done)/queries_total)
if queries_done <= queries_total:
info = "%.3f," % (time.time()-t0)
file_record.write(info)
elif tag == TRAINER_DONE:
if listen_to_trainer:
if print_statements:
print "MASTER: Updating neural network"
W, B, architecture = data
optimizer.update_params(W, B, architecture)
trainer_is_ready = not trainer_is_ready
elif tag == EXIT_WORKER or tag == EXIT_TRAINER:
closed_workers += 1
file_record.write("NA\n")
file_record.close()
t2 = time.time()
print "MASTER: Total update time is: %3.3f" % (t2-t1)
print "Best evaluated point is:"
print dataset[np.argmax(dataset[:, -1]), :]
print "MASTER: Predicted best point is:"
optimizer.retrain_LR()
domain, pred, hi_ci, lo_ci, nn_pred, ei, gamma = optimizer.get_prediction()
index = np.argmax(pred[:, 0])
print np.concatenate((np.atleast_2d(domain[index, :]), np.atleast_2d(pred[index, 0])), axis=1)[0, :]
if np.max(ei) <= 0:
select_index = np.argmax(sig)
else:
select_index = np.argmax(ei)
return np.atleast_2d(domain[select_index, :])
if __name__ == "__main__":
print_statements = False
# Open file to write times for comparison
file_record = open("data/gp_time_data.csv", "a")
# Get settings relevant to the hidden function being used
lim_domain, init_size, additional_query_size, init_query, domain, selection_size = get_settings(
)
# Construct the dataset
dataset = evaluate(init_query[0, :], lim_domain)
print "Randomly query a set of initial points... ",
for query in init_query[1:, :]:
dataset = np.concatenate((dataset, evaluate(query, lim_domain)),
axis=0)
print "Complete initial dataset acquired"
# Begin sequential optimization using Gaussian process based query system
model = pyGPs.GPR()
model.getPosterior(dataset[:, :-1], dataset[:, -1:])
import os
import re
from learning_objective.hidden_function import true_evaluate, get_settings
lim_domain = get_settings(lim_domain_only=True)
scribe = open("./data/regret_analysis/gp_hm.csv", 'w')
for f in os.listdir("./data/regret_analysis"):
if f.startswith("gp_hm"):
print f
f = "data/regret_analysis/" + f
for line in open(f, 'r'):
r = re.compile('Tasks done:(.*?). New')
m = r.search(line)
if m:
print line,
r = re.compile('\[ (.*?)\]')
n = r.search(line)
print n.group(1).split()
val = n.group(1).split()
val[0] = val[0].replace("[", "")
print val
query = [float(elem) for elem in val[0:4]]
print query
tasknum = int(m.group(1))
y_val = true_evaluate(query, lim_domain)[0, -1]
scribe.write(str(tasknum) + "," + str(y_val) + "\n")