def optimize_sp_for_fixed_beta(params: pr.Parameters, groups, real_beta):
# print("!!! Beginning experiment !!!")
(x, y) = tst.generate_training_data(real_beta, params)
def f(log_sparsity_param):
params.sparsity_param = math.pow(2, log_sparsity_param)
(learned_beta, runtime, cycles,
convergence_type) = pr.learn(x, y, groups, params)
avg_error = tst.test(learned_beta, real_beta, params)
return avg_error
#Optimize sparsity parameter several times (stochastic optimization)
opt_log_params = []
for i in range(5):
best_log_param = minimize(f, 0.0, 12.0, 0.3)
opt_log_params.append(best_log_param)
print("optimizing sparsity param...", math.pow(2, best_log_param))
# print("sparsity params for fixed beta:", opt_log_params)
opt_sparsity_param = math.pow(2, np.mean(opt_log_params))
# print("TRAINING SPARSITY:", params.training_feature_sparsity)
print("Average optimum sparsity parameter: ", opt_sparsity_param, "= 2^",
np.mean(opt_log_params))
# Re-run experiment with optimal sparsity parameter
params.sparsity_param = opt_sparsity_param
(learned_beta, runtime, cycles,
convergence_type) = pr.learn(x, y, groups, params)
avg_error = tst.test(learned_beta, real_beta, params)
print("Performance on beta with optimum sparsity parameter",
params.sparsity_param, "runtime:", int(runtime), "cycles:", cycles,
"avg error:", round(avg_error, 3), "convergence:", convergence_type)
return runtime, cycles, avg_error, convergence_type, opt_sparsity_param
def run_experiment(params, generate_beta):
groups = tst.generate_groups(params)
real_beta = generate_beta(params, groups)
(x, y) = tst.generate_training_data(real_beta,params)
(learned_beta, runtime, cycles, convergence_type) = pr.learn(x, y, groups, params)
avg_error = tst.test(learned_beta, real_beta, params)
return runtime, cycles, avg_error, convergence_type
def experiment_with_fixed_params(params: pr.Parameters, gen_beta):
# print("!!! Beginning experiment !!!")
groups = tst.generate_groups(params)
real_beta = gen_beta(params, groups)
(x, y) = tst.generate_training_data(real_beta, params)
# Re-run experiment with optimal sparsity parameter
(learned_beta, runtime, cycles,
convergence_type) = pr.learn(x, y, groups, params)
avg_error = tst.test(learned_beta, real_beta, params)
print("Number of Groups:", params.num_groups,
"runtime:", int(runtime), "cycles:", cycles, "avg error:",
round(avg_error, 3), "convergence:", convergence_type)
return runtime, cycles, avg_error, convergence_type
def run_experiment(params: pr.Parameters, gen_beta):
#Generate learning algorithm
groups = tst.generate_groups(params)
real_beta = gen_beta(params, groups)
(x, y) = tst.generate_training_data(real_beta, params)
#Learning algorithm
(learned_beta, runtime, cycles,
convergence_type) = pr.learn(x, y, groups, params)
#Test accuracy
avg_error = tst.test(learned_beta, real_beta, params)
print("runtime:", int(runtime), "cycles:", cycles, "avg error:",
round(avg_error, 3), "convergence:", convergence_type)
return runtime, cycles, avg_error, convergence_type
def f(log_sparsity_param):
params.sparsity_param = math.pow(2, log_sparsity_param)
(learned_beta, runtime, cycles,
convergence_type) = pr.learn(x, y, groups, params)
avg_error = tst.test(learned_beta, real_beta, params)
return avg_error