def test_Y_transformation_(self):
"""Test if the target Y whitens correctly
"""
# Training
train_data = XYData(True, self.Y_cols, 'FloatTensor', define_src_pos_wrt_lens=True, rescale_pixels=False, log_pixels=False, add_pixel_noise=False, eff_exposure_time={'TDLMC_F160W': self.original_exptime*2.0}, train_Y_mean=None, train_Y_std=None, train_baobab_cfg_path=self.train_baobab_cfg_path, val_baobab_cfg_path=self.val_baobab_cfg_path, for_cosmology=False)
_, actual_Y_0 = train_data[0]
_, actual_Y_1 = train_data[1]
actual_Y = np.stack([actual_Y_0, actual_Y_1], axis=0)
Y_df = self.train_metadata[self.Y_cols].copy()
Y_df['src_light_center_x'] -= Y_df['lens_mass_center_x']
Y_df['src_light_center_y'] -= Y_df['lens_mass_center_y']
expected_Y = Y_df.values
before_whitening_Y = Y_df.values
#expected_Y = (expected_Y - self.train_Y_mean.reshape([1, -1]))/self.train_Y_std.reshape([1, -1])
expected_Y[np.argmin(before_whitening_Y, axis=0), np.arange(len(self.Y_cols))] = -1
expected_Y[np.argmax(before_whitening_Y, axis=0), np.arange(len(self.Y_cols))] = 1
np.testing.assert_array_equal(actual_Y_0.shape, [len(self.Y_cols),], err_msg='shape of single example Y for training')
np.testing.assert_array_almost_equal(actual_Y, expected_Y, err_msg='transformed Y for training')
# Validation
val_data = XYData(False, self.Y_cols, 'FloatTensor', define_src_pos_wrt_lens=True, rescale_pixels=False, log_pixels=False, add_pixel_noise=False, eff_exposure_time={'TDLMC_F160W': self.original_exptime*2.0}, train_Y_mean=self.train_Y_mean, train_Y_std=self.train_Y_std, train_baobab_cfg_path=self.train_baobab_cfg_path, val_baobab_cfg_path=self.val_baobab_cfg_path, for_cosmology=False)
_, actual_Y_0 = val_data[0]
_, actual_Y_1 = val_data[1]
_, actual_Y_2 = val_data[2]
actual_Y = np.stack([actual_Y_0, actual_Y_1, actual_Y_2], axis=0)
expected_Y = self.val_metadata[self.Y_cols].copy()
expected_Y['src_light_center_x'] -= expected_Y['lens_mass_center_x']
expected_Y['src_light_center_y'] -= expected_Y['lens_mass_center_y']
expected_Y = expected_Y.values
expected_Y = (expected_Y - self.train_Y_mean.reshape([1, -1]))/self.train_Y_std.reshape([1, -1])
np.testing.assert_array_equal(actual_Y_0.shape, [len(self.Y_cols),], err_msg='shape of single example Y for validation for arbitrary train mean and std')
np.testing.assert_array_almost_equal(actual_Y, expected_Y, err_msg='transformed Y for validation for arbitrary train mean and std')
def test_train_vs_val(self):
"""Test if the images and metadata are loaded from the correct folder (train/val)
"""
train_data = XYData(True, self.Y_cols, 'FloatTensor', define_src_pos_wrt_lens=True, rescale_pixels=False, log_pixels=False, add_pixel_noise=False, eff_exposure_time={'TDLMC_F160W': self.original_exptime*2.0}, train_Y_mean=self.train_Y_mean, train_Y_std=self.train_Y_std, train_baobab_cfg_path=self.train_baobab_cfg_path, val_baobab_cfg_path=self.val_baobab_cfg_path, for_cosmology=False)
val_data = XYData(False, self.Y_cols, 'FloatTensor', define_src_pos_wrt_lens=True, rescale_pixels=False, log_pixels=False, add_pixel_noise=False, eff_exposure_time={'TDLMC_F160W': self.original_exptime*2.0}, train_Y_mean=self.train_Y_mean, train_Y_std=self.train_Y_std, train_baobab_cfg_path=self.train_baobab_cfg_path, val_baobab_cfg_path=self.val_baobab_cfg_path, for_cosmology=False)
np.testing.assert_equal(len(train_data), 2, err_msg='reading from correct folder (train/val)')
np.testing.assert_equal(len(val_data), 3, err_msg='reading from correct folder (train/val)')
def test_tensor_type(self):
"""Test if X, Y are of the configured data type
"""
# DoubleTensor
train_data = XYData(True, self.Y_cols, 'DoubleTensor', define_src_pos_wrt_lens=True, rescale_pixels=False, log_pixels=False, add_pixel_noise=False, eff_exposure_time={'TDLMC_F160W': self.original_exptime*2.0}, train_Y_mean=self.train_Y_mean, train_Y_std=self.train_Y_std, train_baobab_cfg_path=self.train_baobab_cfg_path, val_baobab_cfg_path=self.val_baobab_cfg_path, for_cosmology=False)
actual_X_0, actual_Y_0 = train_data[0]
assert actual_X_0.type() == 'torch.DoubleTensor'
assert actual_Y_0.type() == 'torch.DoubleTensor'
# FloatTensor
train_data = XYData(True, self.Y_cols, 'FloatTensor', define_src_pos_wrt_lens=True, rescale_pixels=False, log_pixels=False, add_pixel_noise=False, eff_exposure_time={'TDLMC_F160W': self.original_exptime*2.0}, train_Y_mean=self.train_Y_mean, train_Y_std=self.train_Y_std, train_baobab_cfg_path=self.train_baobab_cfg_path, val_baobab_cfg_path=self.val_baobab_cfg_path, for_cosmology=False)
actual_X_0, actual_Y_0 = train_data[0]
assert actual_X_0.type() == 'torch.FloatTensor'
assert actual_Y_0.type() == 'torch.FloatTensor'
def test_X_identity(self):
"""Test if the input iamge equals the dataset image, when nothing is done to the image at all
"""
train_data = XYData(True, self.Y_cols, 'FloatTensor', define_src_pos_wrt_lens=True, rescale_pixels=False, log_pixels=False, add_pixel_noise=False, eff_exposure_time={'TDLMC_F160W': self.original_exptime}, train_Y_mean=self.train_Y_mean, train_Y_std=self.train_Y_std, train_baobab_cfg_path=self.train_baobab_cfg_path, val_baobab_cfg_path=self.val_baobab_cfg_path, for_cosmology=False)
actual_img, _ = train_data[0]
expected_img = self.img_0
np.testing.assert_array_almost_equal(actual_img, expected_img, err_msg='test_X_identity')
def test_X_exposure_time_factor(self):
"""Test if the images scale by the new effective exposure time correctly
"""
train_data = XYData(True, self.Y_cols, 'FloatTensor', define_src_pos_wrt_lens=True, rescale_pixels=False, log_pixels=False, add_pixel_noise=False, eff_exposure_time={'TDLMC_F160W': self.original_exptime*2.0}, train_Y_mean=self.train_Y_mean, train_Y_std=self.train_Y_std, train_baobab_cfg_path=self.train_baobab_cfg_path, val_baobab_cfg_path=self.val_baobab_cfg_path, for_cosmology=False)
actual_img, _ = train_data[0]
expected_img = self.img_0*2.0
np.testing.assert_array_almost_equal(actual_img, expected_img, err_msg='test_X_exposure_time_factor')
def test_X_transformation_rescale(self):
"""Test if the images transform as expected, with whitening
"""
train_data = XYData(True, self.Y_cols, 'FloatTensor', define_src_pos_wrt_lens=True, rescale_pixels=True, log_pixels=False, add_pixel_noise=False, eff_exposure_time={'TDLMC_F160W': self.original_exptime}, train_Y_mean=self.train_Y_mean, train_Y_std=self.train_Y_std, train_baobab_cfg_path=self.train_baobab_cfg_path, val_baobab_cfg_path=self.val_baobab_cfg_path, for_cosmology=False)
actual_img, _ = train_data[0]
expected_img = self.img_0
expected_img = (expected_img - np.mean(expected_img))/np.std(expected_img, ddof=1)
np.testing.assert_array_almost_equal(actual_img, expected_img, err_msg='test_X_transformation_rescale')
def test_X_transformation_log_rescale(self):
"""Test if the images transform as expected, with log(1+X) and whitening
"""
# Without exposure time factor
train_data = XYData(True, self.Y_cols, 'FloatTensor', define_src_pos_wrt_lens=True, rescale_pixels=True, log_pixels=True, add_pixel_noise=False, eff_exposure_time={'TDLMC_F160W': self.original_exptime}, train_Y_mean=self.train_Y_mean, train_Y_std=self.train_Y_std, train_baobab_cfg_path=self.train_baobab_cfg_path, val_baobab_cfg_path=self.val_baobab_cfg_path, for_cosmology=False)
actual_img, _ = train_data[0]
expected_img = self.img_0
expected_img = np.log1p(expected_img)
# Note torch std takes into account Bessel correction
expected_img = (expected_img - np.mean(expected_img))/np.std(expected_img, ddof=1)
np.testing.assert_array_almost_equal(actual_img, expected_img, err_msg='test_X_transformation_log_rescale, without exposure time factor')
# With exposure time factor
train_data = XYData(True, self.Y_cols, 'FloatTensor', define_src_pos_wrt_lens=True, rescale_pixels=True, log_pixels=True, add_pixel_noise=False, eff_exposure_time={'TDLMC_F160W': self.original_exptime*2.0}, train_Y_mean=self.train_Y_mean, train_Y_std=self.train_Y_std, train_baobab_cfg_path=self.train_baobab_cfg_path, val_baobab_cfg_path=self.val_baobab_cfg_path, for_cosmology=False)
actual_img, _ = train_data[0]
expected_img = self.img_0*2.0
expected_img = np.log1p(expected_img)
# Note torch std takes into account Bessel correction
expected_img = (expected_img - np.mean(expected_img))/np.std(expected_img, ddof=1)
np.testing.assert_array_almost_equal(actual_img, expected_img, err_msg='test_X_transformation_log_rescale, with exposure time factor')
def main():
args = script_utils.parse_inference_args()
test_cfg = TestConfig.from_file(args.test_config_file_path)
baobab_cfg = BaobabConfig.from_file(test_cfg.data.test_baobab_cfg_path)
cfg = TrainValConfig.from_file(test_cfg.train_val_config_file_path)
# Set device and default data type
device = torch.device(test_cfg.device_type)
if device.type == 'cuda':
torch.set_default_tensor_type('torch.cuda.FloatTensor')
else:
torch.set_default_tensor_type('torch.FloatTensor')
script_utils.seed_everything(test_cfg.global_seed)
############
# Data I/O #
############
train_data = XYData(is_train=True,
Y_cols=cfg.data.Y_cols,
float_type=cfg.data.float_type,
define_src_pos_wrt_lens=cfg.data.define_src_pos_wrt_lens,
rescale_pixels=cfg.data.rescale_pixels,
log_pixels=cfg.data.log_pixels,
add_pixel_noise=cfg.data.add_pixel_noise,
eff_exposure_time=cfg.data.eff_exposure_time,
train_Y_mean=None,
train_Y_std=None,
train_baobab_cfg_path=cfg.data.train_baobab_cfg_path,
val_baobab_cfg_path=test_cfg.data.test_baobab_cfg_path,
for_cosmology=False)
# Define val data and loader
test_data = XYData(is_train=False,
Y_cols=cfg.data.Y_cols,
float_type=cfg.data.float_type,
define_src_pos_wrt_lens=cfg.data.define_src_pos_wrt_lens,
rescale_pixels=cfg.data.rescale_pixels,
log_pixels=cfg.data.log_pixels,
add_pixel_noise=cfg.data.add_pixel_noise,
eff_exposure_time=cfg.data.eff_exposure_time,
train_Y_mean=train_data.train_Y_mean,
train_Y_std=train_data.train_Y_std,
train_baobab_cfg_path=cfg.data.train_baobab_cfg_path,
val_baobab_cfg_path=test_cfg.data.test_baobab_cfg_path,
for_cosmology=True)
cosmo_df = test_data.Y_df
if test_cfg.data.lens_indices is None:
n_test = test_cfg.data.n_test # number of lenses in the test set
lens_range = range(n_test)
else: # if specific lenses are specified
lens_range = test_cfg.data.lens_indices
n_test = len(lens_range)
print("Performing H0 inference on {:d} specified lenses...".format(n_test))
batch_size = max(lens_range) + 1
test_loader = DataLoader(test_data, batch_size=batch_size, shuffle=False, drop_last=True)
# Output directory into which the H0 histograms and H0 samples will be saved
out_dir = test_cfg.out_dir
if not os.path.exists(out_dir):
os.makedirs(out_dir)
print("Destination folder path: {:s}".format(out_dir))
else:
raise OSError("Destination folder already exists.")
######################
# Load trained state #
######################
# Instantiate loss function
loss_fn = getattr(h0rton.losses, cfg.model.likelihood_class)(Y_dim=train_data.Y_dim, device=device)
# Instantiate posterior (for logging)
bnn_post = getattr(h0rton.h0_inference.gaussian_bnn_posterior, loss_fn.posterior_name)(train_data.Y_dim, device, train_data.train_Y_mean, train_data.train_Y_std)
with torch.no_grad(): # TODO: skip this if lens_posterior_type == 'truth'
for X_, Y_ in test_loader:
X = X_.to(device)
Y = Y_.to(device)
break
# Export the input images X for later error analysis
if test_cfg.export.images:
for lens_i in range(n_test):
X_img_path = os.path.join(out_dir, 'X_{0:04d}.npy'.format(lens_i))
np.save(X_img_path, X[lens_i, 0, :, :].cpu().numpy())
################
# H0 Posterior #
################
h0_prior = getattr(stats, test_cfg.h0_prior.dist)(**test_cfg.h0_prior.kwargs)
kappa_ext_prior = getattr(stats, test_cfg.kappa_ext_prior.dist)(**test_cfg.kappa_ext_prior.kwargs)
aniso_param_prior = getattr(stats, test_cfg.aniso_param_prior.dist)(**test_cfg.aniso_param_prior.kwargs)
# FIXME: hardcoded
kwargs_model = dict(
lens_model_list=['PEMD', 'SHEAR'],
lens_light_model_list=['SERSIC_ELLIPSE'],
source_light_model_list=['SERSIC_ELLIPSE'],
point_source_model_list=['SOURCE_POSITION'],
cosmo=FlatLambdaCDM(H0=70.0, Om0=0.3)
#'point_source_model_list' : ['LENSED_POSITION']
)
h0_post = H0Posterior(
H0_prior=h0_prior,
kappa_ext_prior=kappa_ext_prior,
aniso_param_prior=aniso_param_prior,
exclude_vel_disp=test_cfg.h0_posterior.exclude_velocity_dispersion,
kwargs_model=kwargs_model,
baobab_time_delays=test_cfg.time_delay_likelihood.baobab_time_delays,
kinematics=baobab_cfg.bnn_omega.kinematics,
kappa_transformed=test_cfg.kappa_ext_prior.transformed,
Om0=baobab_cfg.bnn_omega.cosmology.Om0,
define_src_pos_wrt_lens=cfg.data.define_src_pos_wrt_lens,
kwargs_lens_eqn_solver={'min_distance': 0.05, 'search_window': baobab_cfg.instrument['pixel_scale']*baobab_cfg.image['num_pix'], 'num_iter_max': 100}
)
# Get H0 samples for each system
if not test_cfg.time_delay_likelihood.baobab_time_delays:
if 'abcd_ordering_i' not in cosmo_df:
raise ValueError("If the time delay measurements were not generated using Baobab, the user must specify the order of image positions in which the time delays are listed, in order of increasing dec.")
required_params = h0_post.required_params
########################
# Lens Model Posterior #
########################
n_samples = test_cfg.h0_posterior.n_samples # number of h0 samples per lens
sampling_buffer = test_cfg.h0_posterior.sampling_buffer # FIXME: dynamically sample more if we run out of samples
actual_n_samples = int(n_samples*sampling_buffer)
# Add artificial noise around the truth values
Y_orig = bnn_post.transform_back_mu(Y).cpu().numpy().reshape(batch_size, test_data.Y_dim)
Y_orig_df = pd.DataFrame(Y_orig, columns=cfg.data.Y_cols)
Y_orig_values = Y_orig_df[required_params].values[:, np.newaxis, :] # [n_test, 1, Y_dim]
artificial_noise = np.random.randn(batch_size, actual_n_samples, test_data.Y_dim)*Y_orig_values*test_cfg.fractional_error_added_to_truth # [n_test, buffer*n_samples, Y_dim]
lens_model_samples_values = Y_orig_values + artificial_noise # [n_test, buffer*n_samples, Y_dim]
# Placeholders for mean and std of H0 samples per system
mean_h0_set = np.zeros(n_test)
std_h0_set = np.zeros(n_test)
inference_time_set = np.zeros(n_test)
# For each lens system...
total_progress = tqdm(total=n_test)
sampling_progress = tqdm(total=n_samples)
prerealized_time_delays = test_cfg.error_model.prerealized_time_delays
if prerealized_time_delays:
realized_time_delays = pd.read_csv(test_cfg.error_model.realized_time_delays_path, index_col=None)
else:
realized_time_delays = pd.DataFrame()
realized_time_delays['measured_td_wrt0'] = [[]]*len(lens_range)
for i, lens_i in enumerate(lens_range):
lens_i_start_time = time.time()
# Each lens gets a unique random state for td and vd measurement error realizations.
rs_lens = np.random.RandomState(lens_i)
# BNN samples for lens_i
bnn_sample_df = pd.DataFrame(lens_model_samples_values[lens_i, :, :], columns=required_params)
# Cosmology observables for lens_i
cosmo = cosmo_df.iloc[lens_i]
true_td = np.array(literal_eval(cosmo['true_td']))
true_img_dec = np.array(literal_eval(cosmo['y_image']))
true_img_ra = np.array(literal_eval(cosmo['x_image']))
increasing_dec_i = np.argsort(true_img_dec)
true_img_dec = true_img_dec[increasing_dec_i]
true_img_ra = true_img_ra[increasing_dec_i]
measured_vd = cosmo['true_vd']*(1.0 + rs_lens.randn()*test_cfg.error_model.velocity_dispersion_frac_error)
if prerealized_time_delays:
measured_td_wrt0 = np.array(literal_eval(realized_time_delays.iloc[lens_i]['measured_td_wrt0']))
else:
true_td = true_td[increasing_dec_i]
true_td = true_td[1:] - true_td[0]
measured_td_wrt0 = true_td + rs_lens.randn(*true_td.shape)*test_cfg.error_model.time_delay_error # [n_img -1,]
realized_time_delays.at[lens_i, 'measured_td_wrt0'] = list(measured_td_wrt0)
#print("True: ", true_td)
#print("True img: ", true_img_dec)
#print("measured td: ", measured_td_wrt0)
h0_post.set_cosmology_observables(
z_lens=cosmo['z_lens'],
z_src=cosmo['z_src'],
measured_vd=measured_vd,
measured_vd_err=test_cfg.velocity_dispersion_likelihood.sigma,
measured_td_wrt0=measured_td_wrt0,
measured_td_err=test_cfg.time_delay_likelihood.sigma,
abcd_ordering_i=np.arange(len(true_td) + 1),
true_img_dec=true_img_dec,
true_img_ra=true_img_ra,
kappa_ext=cosmo['kappa_ext'], # not necessary
)
h0_post.set_truth_lens_model(sampled_lens_model_raw=bnn_sample_df.iloc[0])
# Initialize output array
h0_samples = np.full(n_samples, np.nan)
h0_weights = np.zeros(n_samples)
# For each sample from the lens model posterior of this lens system...
sampling_progress.reset()
valid_sample_i = 0
sample_i = 0
while valid_sample_i < n_samples:
if sample_i > actual_n_samples - 1:
break
#try:
# Each sample for a given lens gets a unique random state for H0, k_ext, and aniso_param realizations.
rs_sample = np.random.RandomState(int(str(lens_i) + str(sample_i).zfill(5)))
h0, weight = h0_post.get_h0_sample_truth(rs_sample)
h0_samples[valid_sample_i] = h0
h0_weights[valid_sample_i] = weight
sampling_progress.update(1)
time.sleep(0.001)
valid_sample_i += 1
sample_i += 1
#except:
# sample_i += 1
# continue
sampling_progress.refresh()
lens_i_end_time = time.time()
inference_time = (lens_i_end_time - lens_i_start_time)/60.0 # min
h0_dict = dict(
h0_samples=h0_samples,
h0_weights=h0_weights,
n_sampling_attempts=sample_i,
measured_td_wrt0=measured_td_wrt0,
inference_time=inference_time
)
h0_dict_save_path = os.path.join(out_dir, 'h0_dict_{0:04d}.npy'.format(lens_i))
np.save(h0_dict_save_path, h0_dict)
h0_stats = plot_weighted_h0_histogram(h0_samples, h0_weights, lens_i, cosmo['H0'], include_fit_gaussian=test_cfg.plotting.include_fit_gaussian, save_dir=out_dir)
mean_h0_set[i] = h0_stats['mean']
std_h0_set[i] = h0_stats['std']
inference_time_set[i] = inference_time
total_progress.update(1)
total_progress.close()
if not prerealized_time_delays:
realized_time_delays.to_csv(os.path.join(out_dir, 'realized_time_delays.csv'), index=None)
h0_stats = dict(
mean=mean_h0_set,
std=std_h0_set,
inference_time=inference_time_set,
)
h0_stats_save_path = os.path.join(out_dir, 'h0_stats')
np.save(h0_stats_save_path, h0_stats)
def main():
args = script_utils.parse_inference_args()
test_cfg = TestConfig.from_file(args.test_config_file_path)
baobab_cfg = BaobabConfig.from_file(test_cfg.data.test_baobab_cfg_path)
cfg = TrainValConfig.from_file(test_cfg.train_val_config_file_path)
# Set device and default data type
device = torch.device(test_cfg.device_type)
if device.type == 'cuda':
torch.set_default_tensor_type('torch.cuda.' + cfg.data.float_type)
else:
torch.set_default_tensor_type('torch.' + cfg.data.float_type)
script_utils.seed_everything(test_cfg.global_seed)
############
# Data I/O #
############
train_data = XYData(
is_train=True,
Y_cols=cfg.data.Y_cols,
float_type=cfg.data.float_type,
define_src_pos_wrt_lens=cfg.data.define_src_pos_wrt_lens,
rescale_pixels=cfg.data.rescale_pixels,
rescale_pixels_type=cfg.data.rescale_pixels_type,
log_pixels=cfg.data.log_pixels,
add_pixel_noise=cfg.data.add_pixel_noise,
eff_exposure_time=cfg.data.eff_exposure_time,
train_Y_mean=None,
train_Y_std=None,
train_baobab_cfg_path=cfg.data.train_baobab_cfg_path,
val_baobab_cfg_path=None,
for_cosmology=False)
# Define val data and loader
test_data = XYData(
is_train=False,
Y_cols=cfg.data.Y_cols,
float_type=cfg.data.float_type,
define_src_pos_wrt_lens=cfg.data.define_src_pos_wrt_lens,
rescale_pixels=cfg.data.rescale_pixels,
rescale_pixels_type=cfg.data.rescale_pixels_type,
log_pixels=cfg.data.log_pixels,
add_pixel_noise=cfg.data.add_pixel_noise,
eff_exposure_time=cfg.data.eff_exposure_time,
train_Y_mean=train_data.train_Y_mean,
train_Y_std=train_data.train_Y_std,
train_baobab_cfg_path=cfg.data.train_baobab_cfg_path,
val_baobab_cfg_path=test_cfg.data.test_baobab_cfg_path,
for_cosmology=True)
master_truth = test_data.Y_df
master_truth = metadata_utils.add_qphi_columns(master_truth)
master_truth = metadata_utils.add_gamma_psi_ext_columns(master_truth)
# Figure out how many lenses BNN will predict on (must be consecutive)
if test_cfg.data.lens_indices is None:
if args.lens_indices_path is None:
# Test on all n_test lenses in the test set
n_test = test_cfg.data.n_test
lens_range = range(n_test)
else:
# Test on the lens indices in a text file at the specified path
lens_range = []
with open(args.lens_indices_path, "r") as f:
for line in f:
lens_range.append(int(line.strip()))
n_test = len(lens_range)
print("Performing H0 inference on {:d} specified lenses...".format(
n_test))
else:
if args.lens_indices_path is None:
# Test on the lens indices specified in the test config file
lens_range = test_cfg.data.lens_indices
n_test = len(lens_range)
print("Performing H0 inference on {:d} specified lenses...".format(
n_test))
else:
raise ValueError(
"Specific lens indices were specified in both the test config file and the command-line argument."
)
batch_size = max(lens_range) + 1
test_loader = DataLoader(test_data,
batch_size=batch_size,
shuffle=False,
drop_last=True)
# Output directory into which the H0 histograms and H0 samples will be saved
out_dir = test_cfg.out_dir
if not os.path.exists(out_dir):
os.makedirs(out_dir)
print("Destination folder path: {:s}".format(out_dir))
else:
raise OSError("Destination folder already exists.")
#####################
# Parameter penalty #
#####################
# Instantiate original loss function with all BNN-predicted params
orig_Y_cols = cfg.data.Y_cols
loss_fn = getattr(h0rton.losses,
cfg.model.likelihood_class)(Y_dim=test_data.Y_dim,
device=device)
# Not all predicted params will be sampled via MCMC
params_to_remove = [] #'lens_light_R_sersic', 'src_light_R_sersic']
mcmc_Y_cols = [col for col in orig_Y_cols if col not in params_to_remove]
mcmc_Y_dim = len(mcmc_Y_cols)
# Instantiate loss function with just the MCMC params
mcmc_loss_fn = getattr(h0rton.losses, cfg.model.likelihood_class)(
Y_dim=test_data.Y_dim - len(params_to_remove), device=device)
remove_param_idx, remove_idx = mcmc_utils.get_idx_for_params(
mcmc_loss_fn.out_dim, orig_Y_cols, params_to_remove,
cfg.model.likelihood_class)
mcmc_train_Y_mean = np.delete(train_data.train_Y_mean, remove_param_idx)
mcmc_train_Y_std = np.delete(train_data.train_Y_std, remove_param_idx)
parameter_penalty = mcmc_utils.HybridBNNPenalty(
mcmc_Y_cols, cfg.model.likelihood_class, mcmc_train_Y_mean,
mcmc_train_Y_std, test_cfg.h0_posterior.exclude_velocity_dispersion,
device)
custom_logL_addition = parameter_penalty.evaluate
null_spread = False
###################
# BNN predictions #
###################
# Instantiate BNN model
net = getattr(h0rton.models,
cfg.model.architecture)(num_classes=loss_fn.out_dim,
dropout_rate=cfg.model.dropout_rate)
net.to(device)
# Load trained weights from saved state
net, epoch = train_utils.load_state_dict_test(test_cfg.state_dict_path,
net, cfg.optim.n_epochs,
device)
# When only generating BNN predictions (and not running MCMC), we can afford more n_dropout
# otherwise, we fix n_dropout = mcmc_Y_dim + 1
if test_cfg.export.pred:
n_dropout = 20
n_samples_per_dropout = test_cfg.numerics.mcmc.walkerRatio
else:
n_walkers = test_cfg.numerics.mcmc.walkerRatio * (
mcmc_Y_dim + 1) # (BNN params + D_dt) times walker ratio
n_dropout = n_walkers // test_cfg.numerics.mcmc.walkerRatio
n_samples_per_dropout = test_cfg.numerics.mcmc.walkerRatio
# Initialize arrays that will store samples and BNN predictions
init_pos = np.empty(
[batch_size, n_dropout, n_samples_per_dropout, mcmc_Y_dim])
mcmc_pred = np.empty([batch_size, n_dropout, mcmc_loss_fn.out_dim])
with torch.no_grad():
net.train()
# Send some empty forward passes through the test data without backprop to adjust batchnorm weights
# (This is often not necessary. Beware if using for just 1 lens.)
for nograd_pass in range(5):
for X_, Y_ in test_loader:
X = X_.to(device)
_ = net(X)
# Obtain MC dropout samples
for d in range(n_dropout):
net.eval()
for X_, Y_ in test_loader:
X = X_.to(device)
Y = Y_.to(device)
pred = net(X)
break
mcmc_pred_d = pred.cpu().numpy()
# Replace BNN posterior's primary gaussian mean with truth values
if test_cfg.lens_posterior_type == 'default_with_truth_mean':
mcmc_pred_d[:, :len(mcmc_Y_cols)] = Y[:, :len(mcmc_Y_cols
)].cpu().numpy()
# Leave only the MCMC parameters in pred
mcmc_pred_d = mcmc_utils.remove_parameters_from_pred(
mcmc_pred_d, remove_idx, return_as_tensor=False)
# Populate pred that will define the MCMC penalty function
mcmc_pred[:, d, :] = mcmc_pred_d
# Instantiate posterior to generate BNN samples, which will serve as initial positions for walkers
bnn_post = getattr(h0rton.h0_inference.gaussian_bnn_posterior_cpu,
loss_fn.posterior_name + 'CPU')(
mcmc_Y_dim, mcmc_train_Y_mean,
mcmc_train_Y_std)
bnn_post.set_sliced_pred(mcmc_pred_d)
init_pos[:, d, :, :] = bnn_post.sample(
n_samples_per_dropout, sample_seed=test_cfg.global_seed +
d) # contains just the lens model params, no D_dt
gc.collect()
# Terminate right after generating BNN predictions (no MCMC)
if test_cfg.export.pred:
import sys
samples_path = os.path.join(out_dir, 'samples.npy')
np.save(samples_path, init_pos)
sys.exit()
#############
# MCMC loop #
#############
# Convolve MC dropout iterates with aleatoric samples
init_pos = init_pos.transpose(0, 3, 1, 2).reshape(
[batch_size, mcmc_Y_dim,
-1]).transpose(0, 2, 1) # [batch_size, n_samples, mcmc_Y_dim]
init_D_dt = np.random.uniform(0.0,
15000.0,
size=(batch_size, n_walkers, 1))
pred_mean = np.mean(init_pos, axis=1) # [batch_size, mcmc_Y_dim]
# Define assumed model profiles
kwargs_model = dict(lens_model_list=['PEMD', 'SHEAR'],
point_source_model_list=['SOURCE_POSITION'],
source_light_model_list=['SERSIC_ELLIPSE'])
astro_sig = test_cfg.image_position_likelihood.sigma # astrometric uncertainty
# Get H0 samples for each system
if not test_cfg.time_delay_likelihood.baobab_time_delays:
if 'abcd_ordering_i' not in master_truth:
raise ValueError(
"If the time delay measurements were not generated using Baobab, the user must specify the order of image positions in which the time delays are listed, in order of increasing dec."
)
kwargs_lens_eqn_solver = {
'min_distance':
0.05,
'search_window':
baobab_cfg.instrument['pixel_scale'] * baobab_cfg.image['num_pix'],
'num_iter_max':
200
}
total_progress = tqdm(total=n_test)
realized_time_delays = pd.read_csv(
test_cfg.error_model.realized_time_delays, index_col=None)
# For each lens system...
for i, lens_i in enumerate(lens_range):
# Each lens gets a unique random state for time delay measurement error realizations.
#rs_lens = np.random.RandomState(lens_i) # replaced with externally rendered time delays
###########################
# Relevant data and prior #
###########################
data_i = master_truth.iloc[lens_i].copy()
# Set BNN pred defining parameter penalty for this lens, batch processes across n_dropout
parameter_penalty.set_bnn_post_params(mcmc_pred[lens_i, :, :])
# Initialize lens model params walkers at the predictive mean
init_info = dict(
zip(mcmc_Y_cols,
pred_mean[lens_i, :] * mcmc_train_Y_std + mcmc_train_Y_mean))
lcdm = LCDM(z_lens=data_i['z_lens'],
z_source=data_i['z_src'],
flat=True)
true_img_dec = literal_eval(data_i['y_image'])
n_img = len(true_img_dec)
measured_td_sig = test_cfg.time_delay_likelihood.sigma
measured_td_wrt0 = np.array(
literal_eval(
realized_time_delays.iloc[lens_i]['measured_td_wrt0']))
kwargs_data_joint = dict(
time_delays_measured=measured_td_wrt0,
time_delays_uncertainties=measured_td_sig,
)
#############################
# Parameter init and bounds #
#############################
lens_kwargs = mcmc_utils.get_lens_kwargs(init_info,
null_spread=null_spread)
ps_kwargs = mcmc_utils.get_ps_kwargs_src_plane(init_info, astro_sig)
src_light_kwargs = mcmc_utils.get_light_kwargs(
init_info['src_light_R_sersic'], null_spread=null_spread)
special_kwargs = mcmc_utils.get_special_kwargs(
n_img, astro_sig
) # image position offset and time delay distance, aka the "special" parameters
kwargs_params = {
'lens_model': lens_kwargs,
'point_source_model': ps_kwargs,
'source_model': src_light_kwargs,
'special': special_kwargs,
}
if test_cfg.numerics.solver_type == 'NONE':
solver_type = 'NONE'
else:
solver_type = 'PROFILE_SHEAR' if n_img == 4 else 'CENTER'
#solver_type = 'NONE'
kwargs_constraints = {
'num_point_source_list': [n_img],
'Ddt_sampling': True,
'solver_type': solver_type,
}
kwargs_likelihood = {
'time_delay_likelihood': True,
'sort_images_by_dec': True,
'prior_lens': [],
'prior_special': [],
'check_bounds': True,
'check_matched_source_position': False,
'source_position_tolerance': 0.01,
'source_position_sigma': 0.01,
'source_position_likelihood': False,
'custom_logL_addition': custom_logL_addition,
'kwargs_lens_eqn_solver': kwargs_lens_eqn_solver
}
###########################
# MCMC posterior sampling #
###########################
fitting_seq = FittingSequence(kwargs_data_joint,
kwargs_model,
kwargs_constraints,
kwargs_likelihood,
kwargs_params,
verbose=False,
mpi=False)
if i == 0:
param_class = fitting_seq._updateManager.param_class
n_params, param_class_Y_cols = param_class.num_param()
init_pos = mcmc_utils.reorder_to_param_class(
mcmc_Y_cols, param_class_Y_cols, init_pos, init_D_dt)
# MCMC sample from the post-processed BNN posterior jointly with cosmology
lens_i_start_time = time.time()
if test_cfg.lens_posterior_type == 'default':
test_cfg.numerics.mcmc.update(init_samples=init_pos[lens_i, :, :])
fitting_kwargs_list_mcmc = [['MCMC', test_cfg.numerics.mcmc]]
#try:
with script_utils.HiddenPrints():
chain_list_mcmc = fitting_seq.fit_sequence(
fitting_kwargs_list_mcmc)
kwargs_result_mcmc = fitting_seq.best_fit()
lens_i_end_time = time.time()
inference_time = (lens_i_end_time - lens_i_start_time) / 60.0 # min
#############################
# Plotting the MCMC samples #
#############################
# sampler_type : 'EMCEE'
# samples_mcmc : np.array of shape `[n_mcmc_eval, n_params]`
# param_mcmc : list of str of length n_params, the parameter names
sampler_type, samples_mcmc, param_mcmc, _ = chain_list_mcmc[0]
new_samples_mcmc = mcmc_utils.postprocess_mcmc_chain(
kwargs_result_mcmc, samples_mcmc, kwargs_model, lens_kwargs[2],
ps_kwargs[2], src_light_kwargs[2], special_kwargs[2],
kwargs_constraints)
# Plot D_dt histogram
D_dt_samples = new_samples_mcmc['D_dt'].values
true_D_dt = lcdm.D_dt(H_0=data_i['H0'], Om0=0.3)
data_i['D_dt'] = true_D_dt
# Export D_dt samples for this lens
lens_inference_dict = dict(
D_dt_samples=D_dt_samples, # kappa_ext=0 for these samples
inference_time=inference_time,
true_D_dt=true_D_dt,
)
lens_inference_dict_save_path = os.path.join(
out_dir, 'D_dt_dict_{0:04d}.npy'.format(lens_i))
np.save(lens_inference_dict_save_path, lens_inference_dict)
# Optionally export the MCMC samples
if test_cfg.export.mcmc_samples:
mcmc_samples_path = os.path.join(
out_dir, 'mcmc_samples_{0:04d}.csv'.format(lens_i))
new_samples_mcmc.to_csv(mcmc_samples_path, index=None)
# Optionally export the D_dt histogram
if test_cfg.export.D_dt_histogram:
cleaned_D_dt_samples = h0_utils.remove_outliers_from_lognormal(
D_dt_samples, 3)
_ = plotting_utils.plot_D_dt_histogram(cleaned_D_dt_samples,
lens_i,
true_D_dt,
save_dir=out_dir)
# Optionally export the plot of MCMC chain
if test_cfg.export.mcmc_chain:
mcmc_chain_path = os.path.join(
out_dir, 'mcmc_chain_{0:04d}.png'.format(lens_i))
plotting_utils.plot_mcmc_chain(chain_list_mcmc, mcmc_chain_path)
# Optionally export posterior cornerplot of select lens model parameters with D_dt
if test_cfg.export.mcmc_corner:
mcmc_corner_path = os.path.join(
out_dir, 'mcmc_corner_{0:04d}.png'.format(lens_i))
plotting_utils.plot_mcmc_corner(
new_samples_mcmc[test_cfg.export.mcmc_cols],
data_i[test_cfg.export.mcmc_cols],
test_cfg.export.mcmc_col_labels, mcmc_corner_path)
total_progress.update(1)
gc.collect()
realized_time_delays.to_csv(os.path.join(out_dir,
'realized_time_delays.csv'),
index=None)
total_progress.close()
def main():
args = script_utils.parse_inference_args()
test_cfg = TestConfig.from_file(args.test_config_file_path)
baobab_cfg = BaobabConfig.from_file(test_cfg.data.test_baobab_cfg_path)
cfg = TrainValConfig.from_file(test_cfg.train_val_config_file_path)
# Set device and default data type
device = torch.device(test_cfg.device_type)
if device.type == 'cuda':
torch.set_default_tensor_type('torch.cuda.' + cfg.data.float_type)
else:
torch.set_default_tensor_type('torch.' + cfg.data.float_type)
script_utils.seed_everything(test_cfg.global_seed)
############
# Data I/O #
############
# Define val data and loader
test_data = XYData(
is_train=False,
Y_cols=cfg.data.Y_cols,
float_type=cfg.data.float_type,
define_src_pos_wrt_lens=cfg.data.define_src_pos_wrt_lens,
rescale_pixels=False,
rescale_pixels_type=None,
log_pixels=False,
add_pixel_noise=cfg.data.add_pixel_noise,
eff_exposure_time={"TDLMC_F160W": test_cfg.data.eff_exposure_time},
train_Y_mean=np.zeros((1, len(cfg.data.Y_cols))),
train_Y_std=np.ones((1, len(cfg.data.Y_cols))),
train_baobab_cfg_path=cfg.data.train_baobab_cfg_path,
val_baobab_cfg_path=test_cfg.data.test_baobab_cfg_path,
for_cosmology=True)
master_truth = test_data.Y_df
master_truth = metadata_utils.add_qphi_columns(master_truth)
master_truth = metadata_utils.add_gamma_psi_ext_columns(master_truth)
# Figure out how many lenses BNN will predict on (must be consecutive)
if test_cfg.data.lens_indices is None:
if args.lens_indices_path is None:
# Test on all n_test lenses in the test set
n_test = test_cfg.data.n_test
lens_range = range(n_test)
else:
# Test on the lens indices in a text file at the specified path
lens_range = []
with open(args.lens_indices_path, "r") as f:
for line in f:
lens_range.append(int(line.strip()))
n_test = len(lens_range)
print("Performing H0 inference on {:d} specified lenses...".format(
n_test))
else:
if args.lens_indices_path is None:
# Test on the lens indices specified in the test config file
lens_range = test_cfg.data.lens_indices
n_test = len(lens_range)
print("Performing H0 inference on {:d} specified lenses...".format(
n_test))
else:
raise ValueError(
"Specific lens indices were specified in both the test config file and the command-line argument."
)
batch_size = max(lens_range) + 1
test_loader = DataLoader(test_data,
batch_size=batch_size,
shuffle=False,
drop_last=True)
# Output directory into which the H0 histograms and H0 samples will be saved
out_dir = test_cfg.out_dir
if not os.path.exists(out_dir):
os.makedirs(out_dir)
print("Destination folder path: {:s}".format(out_dir))
else:
warnings.warn("Destination folder already exists.")
################
# Compile data #
################
# Image data
with torch.no_grad():
for X_, Y_ in test_loader:
X = X_.to(device)
break
X = X.detach().cpu().numpy()
#############
# MCMC loop #
#############
kwargs_lens_eqn_solver = dict(
min_distance=0.05,
search_window=baobab_cfg.instrument['pixel_scale'] *
baobab_cfg.image['num_pix'],
num_iter_max=200)
fm_posterior = ForwardModelingPosterior(
kwargs_lens_eqn_solver=kwargs_lens_eqn_solver,
astrometric_sigma=test_cfg.image_position_likelihood.sigma,
supersampling_factor=baobab_cfg.numerics.supersampling_factor)
# Get H0 samples for each system
if not test_cfg.time_delay_likelihood.baobab_time_delays:
if 'abcd_ordering_i' not in master_truth:
raise ValueError(
"If the time delay measurements were not generated using Baobab, the user must specify the order of image positions in which the time delays are listed, in order of increasing dec."
)
total_progress = tqdm(total=n_test)
realized_time_delays = pd.read_csv(
test_cfg.error_model.realized_time_delays, index_col=None)
# For each lens system...
for i, lens_i in enumerate(lens_range):
###########################
# Relevant data and prior #
###########################
data_i = master_truth.iloc[lens_i].copy()
lcdm = LCDM(z_lens=data_i['z_lens'],
z_source=data_i['z_src'],
flat=True)
measured_td_wrt0 = np.array(
literal_eval(
realized_time_delays.iloc[lens_i]['measured_td_wrt0']))
n_img = len(measured_td_wrt0) + 1
#print(baobab_cfg.survey_object_dict)
fm_posterior.set_kwargs_data_joint(
image=X[lens_i, 0, :, :],
measured_td=measured_td_wrt0,
measured_td_sigma=test_cfg.time_delay_likelihood.sigma,
survey_object_dict=baobab_cfg.survey_object_dict,
eff_exposure_time=test_cfg.data.eff_exposure_time,
)
# Update solver according to number of lensed images
if test_cfg.numerics.solver_type == 'NONE':
fm_posterior.kwargs_constraints['solver_type'] = 'NONE'
else:
fm_posterior.kwargs_constraints[
'solver_type'] = 'PROFILE_SHEAR' if n_img == 4 else 'ELLIPSE'
fm_posterior.kwargs_constraints['num_point_source_list'] = [n_img]
#print(fm_posterior.kwargs_params['point_source_model'][0][0])
true_D_dt = lcdm.D_dt(H_0=data_i['H0'], Om0=0.3)
# Pull truth param values and initialize walkers there
if test_cfg.numerics.initialize_walkers_to_truth:
fm_posterior.kwargs_lens_init = metadata_utils.get_kwargs_lens_mass(
data_i)
fm_posterior.kwargs_lens_light_init = metadata_utils.get_kwargs_lens_light(
data_i)
fm_posterior.kwargs_source_init = metadata_utils.get_kwargs_src_light(
data_i)
fm_posterior.kwargs_ps_init = metadata_utils.get_kwargs_ps_lensed(
data_i)
fm_posterior.kwargs_special_init = dict(D_dt=true_D_dt)
###########################
# MCMC posterior sampling #
###########################
lens_i_start_time = time.time()
#with script_utils.HiddenPrints():
chain_list_mcmc, kwargs_result_mcmc = fm_posterior.run_mcmc(
test_cfg.numerics.mcmc)
lens_i_end_time = time.time()
inference_time = (lens_i_end_time - lens_i_start_time) / 60.0 # min
#############################
# Plotting the MCMC samples #
#############################
# sampler_type : 'EMCEE'
# samples_mcmc : np.array of shape `[n_mcmc_eval, n_params]`
# param_mcmc : list of str of length n_params, the parameter names
sampler_type, samples_mcmc, param_mcmc, _ = chain_list_mcmc[0]
new_samples_mcmc = mcmc_utils.postprocess_mcmc_chain(
kwargs_result_mcmc,
samples_mcmc,
fm_posterior.kwargs_model,
fm_posterior.kwargs_params['lens_model'][2],
fm_posterior.kwargs_params['point_source_model'][2],
fm_posterior.kwargs_params['source_model'][2],
fm_posterior.kwargs_params['special'][2],
fm_posterior.kwargs_constraints,
kwargs_fixed_lens_light=fm_posterior.
kwargs_params['lens_light_model'][2],
verbose=False)
#from lenstronomy.Plots import chain_plot
model_plot = ModelPlot(fm_posterior.multi_band_list,
fm_posterior.kwargs_model,
kwargs_result_mcmc,
arrow_size=0.02,
cmap_string="gist_heat")
plotting_utils.plot_forward_modeling_comparisons(model_plot, out_dir)
# Plot D_dt histogram
D_dt_samples = new_samples_mcmc[
'D_dt'].values # may contain negative values
data_i['D_dt'] = true_D_dt
# Export D_dt samples for this lens
lens_inference_dict = dict(
D_dt_samples=D_dt_samples, # kappa_ext=0 for these samples
inference_time=inference_time,
true_D_dt=true_D_dt,
)
lens_inference_dict_save_path = os.path.join(
out_dir, 'D_dt_dict_{0:04d}.npy'.format(lens_i))
np.save(lens_inference_dict_save_path, lens_inference_dict)
# Optionally export the MCMC samples
if test_cfg.export.mcmc_samples:
mcmc_samples_path = os.path.join(
out_dir, 'mcmc_samples_{0:04d}.csv'.format(lens_i))
new_samples_mcmc.to_csv(mcmc_samples_path, index=None)
# Optionally export the D_dt histogram
if test_cfg.export.D_dt_histogram:
cleaned_D_dt_samples = h0_utils.remove_outliers_from_lognormal(
D_dt_samples, 3)
_ = plotting_utils.plot_D_dt_histogram(cleaned_D_dt_samples,
lens_i,
true_D_dt,
save_dir=out_dir)
# Optionally export the plot of MCMC chain
if test_cfg.export.mcmc_chain:
mcmc_chain_path = os.path.join(
out_dir, 'mcmc_chain_{0:04d}.png'.format(lens_i))
plotting_utils.plot_mcmc_chain(chain_list_mcmc, mcmc_chain_path)
# Optionally export posterior cornerplot of select lens model parameters with D_dt
if test_cfg.export.mcmc_corner:
mcmc_corner_path = os.path.join(
out_dir, 'mcmc_corner_{0:04d}.png'.format(lens_i))
plotting_utils.plot_mcmc_corner(
new_samples_mcmc[test_cfg.export.mcmc_cols],
data_i[test_cfg.export.mcmc_cols],
test_cfg.export.mcmc_col_labels, mcmc_corner_path)
total_progress.update(1)
gc.collect()
realized_time_delays.to_csv(os.path.join(out_dir,
'realized_time_delays.csv'),
index=None)
total_progress.close()
def main():
args = parse_args()
cfg = TrainValConfig.from_file(args.user_cfg_path)
# Set device and default data type
device = torch.device(cfg.device_type)
if device.type == 'cuda':
torch.set_default_tensor_type('torch.cuda.' + cfg.data.float_type)
else:
torch.set_default_tensor_type('torch.' + cfg.data.float_type)
script_utils.seed_everything(cfg.global_seed)
############
# Data I/O #
############
# Define training data and loader
#torch.multiprocessing.set_start_method('spawn', force=True)
train_data = XYData(
is_train=True,
Y_cols=cfg.data.Y_cols,
float_type=cfg.data.float_type,
define_src_pos_wrt_lens=cfg.data.define_src_pos_wrt_lens,
rescale_pixels=cfg.data.rescale_pixels,
log_pixels=cfg.data.log_pixels,
add_pixel_noise=cfg.data.add_pixel_noise,
eff_exposure_time=cfg.data.eff_exposure_time,
train_Y_mean=None,
train_Y_std=None,
train_baobab_cfg_path=cfg.data.train_baobab_cfg_path,
val_baobab_cfg_path=cfg.data.val_baobab_cfg_path,
for_cosmology=False)
train_loader = DataLoader(train_data,
batch_size=cfg.optim.batch_size,
shuffle=True,
drop_last=True)
n_train = len(train_data) - (len(train_data) % cfg.optim.batch_size)
# Define val data and loader
val_data = XYData(is_train=False,
Y_cols=cfg.data.Y_cols,
float_type=cfg.data.float_type,
define_src_pos_wrt_lens=cfg.data.define_src_pos_wrt_lens,
rescale_pixels=cfg.data.rescale_pixels,
log_pixels=cfg.data.log_pixels,
add_pixel_noise=cfg.data.add_pixel_noise,
eff_exposure_time=cfg.data.eff_exposure_time,
train_Y_mean=train_data.train_Y_mean,
train_Y_std=train_data.train_Y_std,
train_baobab_cfg_path=cfg.data.train_baobab_cfg_path,
val_baobab_cfg_path=cfg.data.val_baobab_cfg_path,
for_cosmology=False)
val_loader = DataLoader(
val_data,
batch_size=min(len(val_data), cfg.optim.batch_size),
shuffle=False,
drop_last=True,
)
n_val = len(val_data) - (len(val_data) %
min(len(val_data), cfg.optim.batch_size))
#########
# Model #
#########
Y_dim = val_data.Y_dim
# Instantiate loss function
loss_fn = getattr(h0rton.losses, cfg.model.likelihood_class)(Y_dim=Y_dim,
device=device)
# Instantiate posterior (for logging)
bnn_post = getattr(h0rton.h0_inference.gaussian_bnn_posterior,
loss_fn.posterior_name)(val_data.Y_dim, device,
val_data.train_Y_mean,
val_data.train_Y_std)
# Instantiate model
net = getattr(h0rton.models,
cfg.model.architecture)(num_classes=loss_fn.out_dim,
dropout_rate=cfg.model.dropout_rate)
net.to(device)
################
# Optimization #
################
# Instantiate optimizer
optimizer = optim.Adam(net.parameters(),
lr=cfg.optim.learning_rate,
amsgrad=False,
weight_decay=cfg.optim.weight_decay)
#optimizer = optim.SGD(net.parameters(), lr=cfg.optim.learning_rate, weight_decay=cfg.optim.weight_decay)
lr_scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
mode='min',
factor=0.75,
patience=50,
cooldown=50,
min_lr=1e-5,
verbose=True)
#lr_scheduler = optim.lr_scheduler.CyclicLR(optimizer, base_lr=cfg.optim.learning_rate*0.2, max_lr=cfg.optim.learning_rate, step_size_up=cfg.optim.lr_scheduler.step_size_up, step_size_down=None, mode='triangular2', gamma=1.0, scale_fn=None, scale_mode='cycle', cycle_momentum=True, base_momentum=0.8, max_momentum=0.9, last_epoch=-1)
# Saving/loading state dicts
checkpoint_dir = cfg.checkpoint.save_dir
if not os.path.exists(checkpoint_dir):
os.mkdir(checkpoint_dir)
if cfg.model.load_state:
epoch, net, optimizer, train_loss, val_loss = train_utils.load_state_dict(
cfg.model.state_path, net, optimizer, cfg.optim.n_epochs, device)
epoch += 1 # resume with next epoch
last_saved_val_loss = val_loss
print(lr_scheduler.state_dict())
print(optimizer.state_dict())
else:
epoch = 0
last_saved_val_loss = np.inf
logger = SummaryWriter()
model_path = ''
print("Training set size: {:d}".format(n_train))
print("Validation set size: {:d}".format(n_val))
progress = tqdm(range(epoch, cfg.optim.n_epochs))
n_iter = 0
for epoch in progress:
#net.apply(h0rton.models.deactivate_batchnorm)
train_loss = 0.0
for batch_idx, (X_tr, Y_tr) in enumerate(train_loader):
n_iter += 1
net.train()
X_tr = X_tr.to(device)
Y_tr = Y_tr.to(device)
# Update weights
optimizer.zero_grad()
pred_tr = net.forward(X_tr)
loss = loss_fn(pred_tr, Y_tr)
loss.backward()
optimizer.step()
# For logging
train_loss += (loss.detach().item() - train_loss) / (1 + batch_idx)
# Step lr_scheduler every batch
lr_scheduler.step(train_loss)
tqdm.write("Iter [{}/{}/{}]: TRAIN Loss: {:.4f}".format(
n_iter, epoch + 1, cfg.optim.n_epochs, train_loss))
if (n_iter) % (cfg.monitoring.interval) == 0:
net.eval()
with torch.no_grad():
#net.apply(h0rton.models.deactivate_batchnorm)
val_loss = 0.0
for batch_idx, (X_v, Y_v) in enumerate(val_loader):
X_v = X_v.to(device)
Y_v = Y_v.to(device)
pred_v = net.forward(X_v)
nograd_loss_v = loss_fn(pred_v, Y_v)
val_loss += (nograd_loss_v.detach().item() -
val_loss) / (1 + batch_idx)
tqdm.write("Epoch [{}/{}]: VALID Loss: {:.4f}".format(
epoch + 1, cfg.optim.n_epochs, val_loss))
# Subset of validation for plotting
n_plotting = cfg.monitoring.n_plotting
#X_plt = X_v[:n_plotting].cpu().numpy()
#Y_plt = Y[:n_plotting].cpu().numpy()
Y_plt_orig = bnn_post.transform_back_mu(
Y_v[:n_plotting]).cpu().numpy()
pred_plt = pred_v[:n_plotting]
# Slice pred_plt into meaningful Gaussian parameters for this batch
bnn_post.set_sliced_pred(pred_plt)
mu_orig = bnn_post.transform_back_mu(
bnn_post.mu).cpu().numpy()
# Log train and val metrics
loss_dict = {'train': train_loss, 'val': val_loss}
logger.add_scalars('metrics/loss', loss_dict, n_iter)
#mae = train_utils.get_mae(mu, Y_plt)
mae_dict = train_utils.get_mae(mu_orig, Y_plt_orig,
cfg.data.Y_cols)
logger.add_scalars('metrics/mae', mae_dict, n_iter)
# Log log determinant of the covariance matrix
if cfg.model.likelihood_class in [
'DoubleGaussianNLL', 'FullRankGaussianNLL'
]:
logdet = train_utils.get_logdet(
bnn_post.tril_elements.cpu().numpy(), Y_dim)
logger.add_histogram('logdet_cov_mat', logdet, n_iter)
# Log second Gaussian stats
if cfg.model.likelihood_class in [
'DoubleGaussianNLL', 'DoubleLowRankGaussianNLL'
]:
# Log histogram of w2
logger.add_histogram('val_pred/weight_gaussian2',
bnn_post.w2.cpu().numpy(), n_iter)
# Log RMSE of second Gaussian
mu2_orig = bnn_post.transform_back_mu(
bnn_post.mu2).cpu().numpy()
mae2_dict = train_utils.get_mae(
mu2_orig, Y_plt_orig, cfg.data.Y_cols)
logger.add_scalars('metrics/mae2', mae2_dict, n_iter)
# Log logdet of second Gaussian
logdet2 = train_utils.get_logdet(
bnn_post.tril_elements2.cpu().numpy(), Y_dim)
logger.add_histogram('logdet_cov_mat2', logdet2,
n_iter)
if val_loss < last_saved_val_loss:
os.remove(model_path) if os.path.exists(
model_path) else None
model_path = train_utils.save_state_dict(
net, optimizer, lr_scheduler, train_loss, val_loss,
checkpoint_dir, cfg.model.architecture, epoch)
last_saved_val_loss = val_loss
logger.close()
# Save final state dict
if val_loss < last_saved_val_loss:
os.remove(model_path) if os.path.exists(model_path) else None
model_path = train_utils.save_state_dict(net, optimizer, lr_scheduler,
train_loss, val_loss,
checkpoint_dir,
cfg.model.architecture, epoch)
print("Saved model at {:s}".format(os.path.abspath(model_path)))