def run_experiment(datasets, alg_engine, epsilons, seed, num_bins):
total_runs = len(epsilons)*len(datasets)
print("total runs: ", total_runs)
num_done = 0
experiment_results = []
for i in in range(len(datasets)):
dataset = datasets[i]
if sum(dataset) == 0 or len(dataset) <= 2:
print("bad dataset")
continue # for some reason there are '0' data vectors
# also, for branching, length should be at least 3
dataset = np.array(dataset)
scale = sum(dataset)
domain_size = len(dataset)
data_range = max(dataset) - min(dataset)
std_dev = math.sqrt(np.var(dataset))
uniform_distance = algs.uniform_distance(dataset)
for epsilon in epsilons:
w = workload.Prefix1D(domain_shape_int=len(dataset))
dataset_hat = alg_engine.Run(w, dataset, epsilon, seed)
histogram, bin_size = algs.get_histogram(dataset, num_bins)
private_hist, bin_size = algs.get_histogram(dataset_hat, num_bins)
error = algs.get_scaled_error(histogram, private_hist)
experiment_results.append((scale, domain_size, error, data_range, std_dev, uniform_distance, epsilon, data_set_index, i))
num_done +=1
if num_done % 50 ==0 :
print("num done: ", num_done)
return experiment_results
def get_error_data():
for i in range(reps):
for data_vector in dataset_vectors_ext:
for epsilon in epsilons:
#num_iterations = 5*math.log(sum(data_vector))
num_iterations = 2
queries = algs.get_queries(num_bins, len(data_vector))
private_dataset = algs.mwem(data_vector, queries,
num_iterations, epsilon)
histogram, bin_size = algs.get_histogram(data_vector, num_bins)
private_hist, bin_size = algs.get_histogram(
private_dataset, num_bins)
# collect statistics
scale = sum(data_vector)
domain_size = len(data_vector)
std_dev = math.sqrt(np.var(data_vector))
uniform_distance = algs.uniform_distance(data_vector)
private_histograms_data.append(
(private_hist, histogram, sum(histogram), domain_size,
data_range, std_dev, uniform_distance, epsilon))
done += 1
if done % 500 == 0:
print "num done = ", done
error_errors = []
all_results = []
num_correct = 0
for i in range(len(data_files)):
data_file = data_files[i]
dataset = np.load(data_file)
epsilon = .01
w = workload.Prefix1D(domain_shape_int=len(dataset))
results = {}
predicted_error = predictions[i]['dataset_stat'][2]
for alg_engine in alg_engines:
predicted_epsilon = predictions[i][alg_engine.short_name][0]
dataset_hat = alg_engine.Run(w, dataset, predicted_epsilon, seed)
histogram, bin_size = algs.get_histogram(dataset, num_bins)
private_hist, bin_size = algs.get_histogram(dataset_hat, num_bins)
error = algs.get_scaled_error(histogram, private_hist)
error_errors.append(abs(predicted_error - error))
results[alg_engine.short_name] = error
actual_best = min(results, key=results.get)
predictions_algs = {}
for key in predictions[i].keys():
if key == 'dataset_stat':
continue
predictions_algs[key] = predictions[i][key][0] #isolate epsilons
predicted_best = min(predictions_algs, key=predictions_algs.get)
if actual_best == predicted_best:
num_correct += 1
num_total = reps * len(dataset_vectors_ext) * len(epsilons)
print "num total: ", num_total
# for data_vector in dataset_vectors_ext:
# print "my scale: ", sum(data_vector)
# print "my domain size: ", len(data_vector)
for i in range(reps):
for data_vector in dataset_vectors_ext:
for epsilon in epsilons:
#num_iterations = 5*math.log(sum(data_vector))
num_iterations = 2
queries = algs.get_queries(num_bins, len(data_vector))
private_dataset = algs.mwem(data_vector, queries, num_iterations,
epsilon)
histogram, bin_size = algs.get_histogram(data_vector, num_bins)
private_hist, bin_size = algs.get_histogram(
private_dataset, num_bins)
# collect statistics
scale = sum(data_vector)
domain_size = len(data_vector)
data_range = max(data_vector) - min(data_vector)
std_dev = math.sqrt(np.var(data_vector))
uniform_distance = algs.uniform_distance(data_vector)
private_histograms_data.append(
(private_hist, histogram, sum(histogram), domain_size,
data_range, std_dev, uniform_distance, epsilon))
done += 1
if done % 500 == 0: