def test_gp_dfixed():
'''
Sets up uniform generation with all discrete parameters fixed to set values.
'''
n_samples = 100
seed = 2
np.random.seed(seed)
domain = Domain()
sampling_strategy = UniformSamplingStrategy()
domain.fix_param(domain.params[1], 'tungsten')
domain.fix_param(domain.params[2], 'SiC')
domain.fix_param(domain.params[3], 'H2O')
domain.fix_param(domain.params[5], 'SiC')
domain.fix_param(domain.params[6], 'Li4SiO4')
domain.fix_param(domain.params[7], 'Be')
domain.fix_param(domain.params[8], 'H20')
df = domain.gen_data_frame(sampling_strategy, n_samples)
df.to_csv('params/100params0000000.csv', index=False)
def main():
'''
Perform quality-adaptive sampling algorithm
'''
# Parse inputs and store in relevant variables.
args = input_parse()
init_samples = args.init_samples
step_samples = args.step_samples
step_candidates = args.step_candidates
d_params = disctrans(args.disc_fix)
# Collect surrogate model type and theory under study.
thismodel = get_model_factory()[args.model](cli_args=sys.argv[7:])
thistheory = globals()["theory_" + args.theory]
domain = Domain()
if args.saved_init:
# load data as initial evaluated samples
df = load_batches(args.saved_init, (0, 1 + int(init_samples/1000)))
X_init, d, y_multiple = c_d_y_split(df.iloc[0:init_samples])
d_params = d.values[0]
print(d.values[0][0])
y_init = y_multiple['tbr']
domain.fix_param(domain.params[1], d_params[0])
domain.fix_param(domain.params[2], d_params[1])
domain.fix_param(domain.params[3], d_params[2])
domain.fix_param(domain.params[5], d_params[3])
domain.fix_param(domain.params[6], d_params[4])
domain.fix_param(domain.params[7], d_params[5])
domain.fix_param(domain.params[8], d_params[6])
if not args.saved_init:
# generate initial parameters
sampling_strategy = UniformSamplingStrategy()
c = domain.gen_data_frame(sampling_strategy, init_samples)
print(c.columns)
# evaluate initial parameters in given theory
print("Evaluating initial " + str(init_samples) + " samples in " + args.theory + " theory.")
output = thistheory(params = c, domain = domain, n_samples = init_samples)
X_init, d, y_multiple = c_d_y_split(output)
y_init = y_multiple['tbr']
current_samples, current_tbr = X_init, y_init
# MAIN QASS LOOP
complete_condition = False
iter_count = 0
err_target = 0.0001
max_iter_count = 70
all_metrics = pd.DataFrame()
current_samples = current_samples.sort_index(axis=1)
print(f'Features in order are: {list(current_samples.columns)}')
X_train, X_test, y_train, y_test = train_test_split(current_samples, current_tbr,
test_size=0.5, random_state=1)
thismodel.enable_renormalization(100)
while not complete_condition:
iter_count += 1
samp_size = X_train.shape[0] * 2
print("Iteration " + str(iter_count) + " -- Total samples: " + str(samp_size))
# Train surrogate for theory, and plot results
if iter_count == 1:
new_samples, new_tbr = X_train, y_train
train(thismodel, new_samples, new_tbr)
test(thismodel, X_test, y_test)
plot("qassplot", thismodel, X_test, y_test)
this_metrics = get_metrics(thismodel, X_test, y_test)
this_metrics['numdata'] = samp_size
print(this_metrics)
# Calculate error data for this training iteration
y_train_pred = thismodel.predict(X_train.to_numpy())
y_test_pred = thismodel.predict(X_test.to_numpy())
train_err = abs(y_train - y_train_pred)
test_err = abs(y_test - y_test_pred)
# Train neural network surrogate for error function (Failed)
X_test = X_test.sort_index(axis=1)
X_test1, X_test2, test_err1, test_err2 = train_test_split(X_test, test_err,
test_size=0.5, random_state=1)
#errmodel = get_model_factory()["nn"](cli_args=["--epochs=50", "--batch-size=200"
# ,"--arch-type=4F_512"])
#errmodel = get_model_factory()["rbf"](cli_args=["--d0=20"])
#scaled_X_test1, scaled_test_err1 = errmodel.scale_training_set(X_test1, test_err1)
#scaled_X_test2, scaled_test_err2 = errmodel.scale_testing_set(X_test2, test_err2)
#dtest1 = pd.DataFrame(scaled_X_test1, columns = X_test1.columns,
# index = X_test1.index)
#dtest2 = pd.DataFrame(scaled_X_test2, columns = X_test2.columns,
# index = X_test2.index)
#derr1 = pd.Series(scaled_test_err1, index = test_err1.index)
#derr2 = pd.Series(scaled_test_err2, index = test_err2.index)
#print(type(test_err1))
#print(type(scaled_test_err1))
#train(errmodel, dtest1, derr1)
#test(errmodel, dtest2, derr2)
#print(X_test1)
#print(scaled_X_test1)
#print(dtest1)
#plot("qassplot3", errmodel, dtest2, derr2)
#tri = Delaunay(X_test1.values, qhull_options="Qc QbB Qx Qz")
#f = interpolate.LinearNDInterpolator(tri, test_err1.values)
# Test surrogate (nearest neighbor interpolator) on split error data
errordist_test = interpolate.NearestNDInterpolator(X_test1.values, test_err1.values)
pred_err1 = errordist_test(X_test1.values)
pred_err2 = errordist_test(X_test2.values)
# Train surrogate (nearest neighbor interpolator) for error function
errordist = interpolate.NearestNDInterpolator(X_test.values, test_err.values)
pred_err = errordist(X_test.values)
max_err = max(test_err.values)
print('Max error: ' + str(max_err))
this_metrics['maxerr'] = max_err
plot_results("qassplot2", pred_err1, test_err1)
plt.figure()
plot_results("qassplot3", pred_err2, test_err2)
plt.figure()
plt.hist(test_err.values, bins=100)
plt.savefig("qassplot4.pdf", format="pdf")
# Perform MCMC on error function
saveinterval = 1
nburn = 1000
nrun = 10000
initial_sample = X_train.iloc[0]
#print(initial_sample.values)
#print(errordist(initial_sample.values))
burnin_sample, burnin_dist, burnin_acceptance = burnin_MH(errordist, initial_sample.values, nburn)
saved_samples, saved_dists, run_acceptance = run_MH(errordist, burnin_sample, nrun, saveinterval)
plt.figure()
plt.hist(saved_dists, bins=100)
plt.savefig("qassplot5.pdf", format="pdf")
print('MCMC run finished.')
print('Burn-In Acceptance: ' + str(burnin_acceptance))
print('Run Acceptance: ' + str(run_acceptance))
this_metrics['burn_acc'] = burnin_acceptance
this_metrics['run_acc'] = run_acceptance
# Extract candidate samples from MCMC output and calculate mutual crowding distance
cand_cdms = []
print(saved_samples.shape)
samplestep = int(saved_samples.shape[0] / step_candidates)
print(samplestep)
candidates = saved_samples[::samplestep]
for candidate in candidates:
cand_cdms.append( cdm(candidate,candidates) )
# Rank candidate samples by error value, and filter out crowded samples
new_samples = pd.DataFrame(candidates, columns = current_samples.columns)
new_samples['error'] = saved_dists[::samplestep]
new_samples['cdm'] = cand_cdms
#print(new_samples)
#print(new_samples.shape)
new_samples = new_samples.sort_values(by='error', ascending=False)
indexNames = new_samples[ new_samples['cdm'] <= new_samples['cdm'].median() ].index
new_samples.drop(indexNames , inplace=True)
new_samples.drop(columns=['error', 'cdm'])
new_samples = new_samples.head(step_samples).reset_index()
# Add new samples and corresponding TBR evaluations to current sample set
new_output = thistheory(params = new_samples.join(pd.concat([d.head(1)]*step_samples, ignore_index=True)), domain = domain, n_samples = step_samples)
new_samples, new_d, new_y_multiple = c_d_y_split(new_output)
new_tbr = new_y_multiple['tbr']
#print(new_samples)
new_samples = new_samples.sort_index(axis=1)
#new_X_train, new_X_test, new_y_train, new_y_test = train_test_split(new_samples, new_tbr,test_size=0.5, random_state=1)
X_train = pd.concat([X_train, new_samples], ignore_index=True)
#X_test = pd.concat([X_test, new_X_test], ignore_index=True)
y_train = pd.concat([y_train, new_tbr], ignore_index=True)
#y_test = pd.concat([y_test, new_y_test], ignore_index=True)
# Check completion conditions and close loop
if max_err < err_target or iter_count > max_iter_count:
complete_condition = True
all_metrics = pd.concat([all_metrics,this_metrics], ignore_index=True)
print(all_metrics)
all_metrics.to_csv('qassmetrics.csv')
print('QASS finished.')
def main():
'''
Perform FAKE quality-adaptive sampling algorithm
'''
# Parse inputs and store in relevant variables.
args = input_parse()
init_samples = args.init_samples
step_samples = args.step_samples
step_candidates = args.step_candidates
eval_samples = args.eval_samples
retrain = args.retrain
d_params = disctrans(args.disc_fix)
# Collect surrogate model type and theory under study.
thismodel = get_model_factory()[args.model](cli_args=sys.argv[9:])
thistheory = globals()["theory_" + args.theory]
domain = Domain()
if args.saved_init:
# load data as initial evaluated samples
df = load_batches(args.saved_init, (0, 1 + int(init_samples / 1000)))
X_init, d, y_multiple = c_d_y_split(df.iloc[0:init_samples])
d_params = d.values[0]
print(d.values[0][0])
y_init = y_multiple['tbr']
domain.fix_param(domain.params[1], d_params[0])
domain.fix_param(domain.params[2], d_params[1])
domain.fix_param(domain.params[3], d_params[2])
domain.fix_param(domain.params[5], d_params[3])
domain.fix_param(domain.params[6], d_params[4])
domain.fix_param(domain.params[7], d_params[5])
domain.fix_param(domain.params[8], d_params[6])
if not args.saved_init:
# generate initial parameters
sampling_strategy = UniformSamplingStrategy()
c = domain.gen_data_frame(sampling_strategy, init_samples)
print(c.columns)
# evaluate initial parameters in given theory
print("Evaluating initial " + str(init_samples) + " samples in " +
args.theory + " theory.")
output = thistheory(params=c, domain=domain, n_samples=init_samples)
X_init, d, y_multiple = c_d_y_split(output)
y_init = y_multiple['tbr']
current_samples, current_tbr = X_init, y_init
# MAIN QASS LOOP
complete_condition = False
iter_count = 0
trigger_retrain = False
similarity = 0
err_target = 0.0001
max_iter_count = 10000
all_metrics = pd.DataFrame()
while not complete_condition:
iter_count += 1
samp_size = current_samples.shape[0]
print("Iteration " + str(iter_count) + " -- Total samples: " +
str(samp_size))
# Train surrogate for theory, and plot results
X_train, X_test, y_train, y_test = train_test_split(
current_samples, current_tbr, test_size=0.5,
random_state=1) #play with this
# Goldilocks retraining scheme
if iter_count > 1:
alt_scaler = thismodel.create_scaler()
Xy_in = thismodel.join_sets(X_train, y_train)
alt_scaler.fit(Xy_in)
similarity = thismodel.scaler_similarity(thismodel.scaler,
alt_scaler)
if iter_count % 10000 == 0: #restart with new random weights
#thismodel = get_model_factory()[args.model](cli_args=sys.argv[8:])
thismodel.scaler = alt_scaler
train(thismodel, X_train, y_train)
test(thismodel, X_test, y_test)
plot("qassplot", thismodel, X_test, y_test)
this_metrics = get_metrics(thismodel, X_test, y_test)
this_metrics['numdata'] = samp_size
this_metrics['similarity'] = similarity
print(this_metrics)
# True evaluation test on uniform random data
evaltest_samples = domain.gen_data_frame(sampling_strategy,
eval_samples)
eval_output = thistheory(params=evaltest_samples,
domain=domain,
n_samples=eval_samples)
evaltest_samples, evaltest_d, evaltest_y_multiple = c_d_y_split(
eval_output)
evaltest_tbr = evaltest_y_multiple['tbr']
test(thismodel, evaltest_samples, evaltest_tbr)
plot("qassplot2", thismodel, evaltest_samples, evaltest_tbr)
eval_metrics = get_metrics(thismodel, evaltest_samples, evaltest_tbr)
print(eval_samples)
this_metrics['E_MAE'] = eval_metrics['MAE']
this_metrics['E_S'] = eval_metrics['S']
this_metrics['E_R2'] = eval_metrics['R2']
this_metrics['E_R2(adj)'] = eval_metrics['R2(adj)']
# Generate uniform random new samples
new_samples = domain.gen_data_frame(sampling_strategy, step_samples)
new_output = thistheory(params=new_samples,
domain=domain,
n_samples=step_samples)
new_samples, new_d, new_y_multiple = c_d_y_split(new_output)
new_tbr = new_y_multiple['tbr']
current_samples = pd.concat([current_samples, new_samples],
ignore_index=True)
current_tbr = pd.concat([current_tbr, new_tbr], ignore_index=True)
# Check completion conditions and close loop
if iter_count > max_iter_count:
complete_condition = True
all_metrics = pd.concat([all_metrics, this_metrics], ignore_index=True)
print(all_metrics)
all_metrics.to_csv('qassfakemetrics.csv')
print('FAKE QASS finished.')