def distributed_strategy(args):
kappa_gen = NISGenerator( # only used to generate pixelated kappa fields
kappa_fov=args.kappa_fov,
src_fov=args.source_fov,
pixels=args.kappa_pixels,
z_source=args.z_source,
z_lens=args.z_lens
)
min_theta_e = 0.1 * args.image_fov if args.min_theta_e is None else args.min_theta_e
max_theta_e = 0.45 * args.image_fov if args.max_theta_e is None else args.max_theta_e
cosmos_files = glob.glob(os.path.join(args.cosmos_dir, "*.tfrecords"))
cosmos = tf.data.TFRecordDataset(cosmos_files, compression_type=args.compression_type)
cosmos = cosmos.map(decode_image).map(preprocess_image)
if args.shuffle_cosmos:
cosmos = cosmos.shuffle(buffer_size=args.buffer_size, reshuffle_each_iteration=True)
cosmos = cosmos.batch(args.batch_size)
window = tukey(args.src_pixels, alpha=args.tukey_alpha)
window = np.outer(window, window)
phys = PhysicalModel(
image_fov=args.image_fov,
kappa_fov=args.kappa_fov,
src_fov=args.source_fov,
pixels=args.lens_pixels,
kappa_pixels=args.kappa_pixels,
src_pixels=args.src_pixels,
method="conv2d"
)
noise_a = (args.noise_rms_min - args.noise_rms_mean) / args.noise_rms_std
noise_b = (args.noise_rms_max - args.noise_rms_mean) / args.noise_rms_std
psf_a = (args.psf_fwhm_min - args.psf_fwhm_mean) / args.psf_fwhm_std
psf_b = (args.psf_fwhm_max - args.psf_fwhm_mean) / args.psf_fwhm_std
options = tf.io.TFRecordOptions(compression_type=args.compression_type)
with tf.io.TFRecordWriter(os.path.join(args.output_dir, f"data_{THIS_WORKER}.tfrecords"), options) as writer:
print(f"Started worker {THIS_WORKER} at {datetime.now().strftime('%y-%m-%d_%H-%M-%S')}")
for i in range((THIS_WORKER - 1) * args.batch_size, args.len_dataset, N_WORKERS * args.batch_size):
for galaxies in cosmos:
break
galaxies = window[np.newaxis, ..., np.newaxis] * galaxies
noise_rms = truncnorm.rvs(noise_a, noise_b, loc=args.noise_rms_mean, scale=args.noise_rms_std, size=args.batch_size)
fwhm = truncnorm.rvs(psf_a, psf_b, loc=args.psf_fwhm_mean, scale=args.psf_fwhm_std, size=args.batch_size)
psf = phys.psf_models(fwhm, cutout_size=args.psf_cutout_size)
batch_size = galaxies.shape[0]
_r = tf.random.uniform(shape=[batch_size, 1, 1], minval=0, maxval=args.max_shift)
_theta = tf.random.uniform(shape=[batch_size, 1, 1], minval=-np.pi, maxval=np.pi)
x0 = _r * tf.math.cos(_theta)
y0 = _r * tf.math.sin(_theta)
ellipticity = tf.random.uniform(shape=[batch_size, 1, 1], minval=0., maxval=args.max_ellipticity)
phi = tf.random.uniform(shape=[batch_size, 1, 1], minval=-np.pi, maxval=np.pi)
einstein_radius = tf.random.uniform(shape=[batch_size, 1, 1], minval=min_theta_e, maxval=max_theta_e)
kappa = kappa_gen.kappa_field(x0, y0, ellipticity, phi, einstein_radius)
lensed_images = phys.noisy_forward(galaxies, kappa, noise_rms=noise_rms, psf=psf)
records = encode_examples(
kappa=kappa,
galaxies=galaxies,
lensed_images=lensed_images,
z_source=args.z_source,
z_lens=args.z_lens,
image_fov=phys.image_fov,
kappa_fov=phys.kappa_fov,
source_fov=args.source_fov,
noise_rms=noise_rms,
psf=psf,
fwhm=fwhm
)
for record in records:
writer.write(record)
print(f"Finished work at {datetime.now().strftime('%y-%m-%d_%H-%M-%S')}")
def distributed_strategy(args):
psf_pixels = 20
pixels = 128
model = os.path.join(os.getenv('CENSAI_PATH'), "models", args.model)
ps_observation = PowerSpectrum(bins=args.observation_coherence_bins,
pixels=pixels)
ps_source = PowerSpectrum(bins=args.source_coherence_bins, pixels=pixels)
ps_kappa = PowerSpectrum(bins=args.kappa_coherence_bins, pixels=pixels)
phys = PhysicalModel(
pixels=pixels,
kappa_pixels=pixels,
src_pixels=pixels,
image_fov=7.68,
kappa_fov=7.68,
src_fov=3.,
method="fft",
)
with open(os.path.join(model, "unet_hparams.json")) as f:
unet_params = json.load(f)
unet_params["kernel_l2_amp"] = args.l2_amp
unet = Model(**unet_params)
ckpt = tf.train.Checkpoint(net=unet)
checkpoint_manager = tf.train.CheckpointManager(ckpt, model, 1)
checkpoint_manager.checkpoint.restore(
checkpoint_manager.latest_checkpoint).expect_partial()
with open(os.path.join(model, "rim_hparams.json")) as f:
rim_params = json.load(f)
rim_params["source_link"] = "relu"
rim = RIM(phys, unet, **rim_params)
kvae_path = os.path.join(os.getenv('CENSAI_PATH'), "models",
args.kappa_vae)
with open(os.path.join(kvae_path, "model_hparams.json"), "r") as f:
kappa_vae_hparams = json.load(f)
kappa_vae = VAE(**kappa_vae_hparams)
ckpt1 = tf.train.Checkpoint(step=tf.Variable(1), net=kappa_vae)
checkpoint_manager1 = tf.train.CheckpointManager(ckpt1, kvae_path, 1)
checkpoint_manager1.checkpoint.restore(
checkpoint_manager1.latest_checkpoint).expect_partial()
svae_path = os.path.join(os.getenv('CENSAI_PATH'), "models",
args.source_vae)
with open(os.path.join(svae_path, "model_hparams.json"), "r") as f:
source_vae_hparams = json.load(f)
source_vae = VAE(**source_vae_hparams)
ckpt2 = tf.train.Checkpoint(step=tf.Variable(1), net=source_vae)
checkpoint_manager2 = tf.train.CheckpointManager(ckpt2, svae_path, 1)
checkpoint_manager2.checkpoint.restore(
checkpoint_manager2.latest_checkpoint).expect_partial()
model_name = os.path.split(model)[-1]
wk = tf.keras.layers.Lambda(lambda k: tf.sqrt(k) / tf.reduce_sum(
tf.sqrt(k), axis=(1, 2, 3), keepdims=True))
with h5py.File(
os.path.join(
os.getenv("CENSAI_PATH"), "results", args.experiment_name +
"_" + model_name + f"_{THIS_WORKER:02d}.h5"), 'w') as hf:
data_len = args.size // N_WORKERS
hf.create_dataset(name="observation",
shape=[data_len, phys.pixels, phys.pixels, 1],
dtype=np.float32)
hf.create_dataset(name="psf",
shape=[data_len, psf_pixels, psf_pixels, 1],
dtype=np.float32)
hf.create_dataset(name="psf_fwhm", shape=[data_len], dtype=np.float32)
hf.create_dataset(name="noise_rms", shape=[data_len], dtype=np.float32)
hf.create_dataset(
name="source",
shape=[data_len, phys.src_pixels, phys.src_pixels, 1],
dtype=np.float32)
hf.create_dataset(
name="kappa",
shape=[data_len, phys.kappa_pixels, phys.kappa_pixels, 1],
dtype=np.float32)
hf.create_dataset(name="observation_pred",
shape=[data_len, phys.pixels, phys.pixels, 1],
dtype=np.float32)
hf.create_dataset(name="observation_pred_reoptimized",
shape=[data_len, phys.pixels, phys.pixels, 1],
dtype=np.float32)
hf.create_dataset(
name="source_pred",
shape=[data_len, rim.steps, phys.src_pixels, phys.src_pixels, 1],
dtype=np.float32)
hf.create_dataset(
name="source_pred_reoptimized",
shape=[data_len, phys.src_pixels, phys.src_pixels, 1],
dtype=np.float32)
hf.create_dataset(name="kappa_pred",
shape=[
data_len, rim.steps, phys.kappa_pixels,
phys.kappa_pixels, 1
],
dtype=np.float32)
hf.create_dataset(
name="kappa_pred_reoptimized",
shape=[data_len, phys.kappa_pixels, phys.kappa_pixels, 1],
dtype=np.float32)
hf.create_dataset(name="chi_squared",
shape=[data_len, rim.steps],
dtype=np.float32)
hf.create_dataset(name="chi_squared_reoptimized",
shape=[data_len, rim.steps],
dtype=np.float32)
hf.create_dataset(name="chi_squared_reoptimized_series",
shape=[data_len, rim.steps, args.re_optimize_steps],
dtype=np.float32)
hf.create_dataset(name="sampled_chi_squared_reoptimized_series",
shape=[data_len, args.re_optimize_steps],
dtype=np.float32)
hf.create_dataset(name="source_optim_mse",
shape=[data_len],
dtype=np.float32)
hf.create_dataset(name="source_optim_mse_series",
shape=[data_len, args.re_optimize_steps],
dtype=np.float32)
hf.create_dataset(name="sampled_source_optim_mse_series",
shape=[data_len, args.re_optimize_steps],
dtype=np.float32)
hf.create_dataset(name="kappa_optim_mse",
shape=[data_len],
dtype=np.float32)
hf.create_dataset(name="kappa_optim_mse_series",
shape=[data_len, args.re_optimize_steps],
dtype=np.float32)
hf.create_dataset(name="sampled_kappa_optim_mse_series",
shape=[data_len, args.re_optimize_steps],
dtype=np.float32)
hf.create_dataset(name="latent_kappa_gt_distance_init",
shape=[data_len, kappa_vae.latent_size],
dtype=np.float32)
hf.create_dataset(name="latent_source_gt_distance_init",
shape=[data_len, source_vae.latent_size],
dtype=np.float32)
hf.create_dataset(name="latent_kappa_gt_distance_end",
shape=[data_len, kappa_vae.latent_size],
dtype=np.float32)
hf.create_dataset(name="latent_source_gt_distance_end",
shape=[data_len, source_vae.latent_size],
dtype=np.float32)
hf.create_dataset(name="source_coherence_spectrum",
shape=[data_len, args.source_coherence_bins],
dtype=np.float32)
hf.create_dataset(name="source_coherence_spectrum_reoptimized",
shape=[data_len, args.source_coherence_bins],
dtype=np.float32)
hf.create_dataset(name="observation_coherence_spectrum",
shape=[data_len, args.observation_coherence_bins],
dtype=np.float32)
hf.create_dataset(name="observation_coherence_spectrum_reoptimized",
shape=[data_len, args.observation_coherence_bins],
dtype=np.float32)
hf.create_dataset(name="kappa_coherence_spectrum",
shape=[data_len, args.kappa_coherence_bins],
dtype=np.float32)
hf.create_dataset(name="kappa_coherence_spectrum_reoptimized",
shape=[data_len, args.kappa_coherence_bins],
dtype=np.float32)
hf.create_dataset(name="observation_frequencies",
shape=[args.observation_coherence_bins],
dtype=np.float32)
hf.create_dataset(name="source_frequencies",
shape=[args.source_coherence_bins],
dtype=np.float32)
hf.create_dataset(name="kappa_frequencies",
shape=[args.kappa_coherence_bins],
dtype=np.float32)
hf.create_dataset(name="kappa_fov", shape=[1], dtype=np.float32)
hf.create_dataset(name="source_fov", shape=[1], dtype=np.float32)
hf.create_dataset(name="observation_fov", shape=[1], dtype=np.float32)
for i in range(data_len):
checkpoint_manager.checkpoint.restore(
checkpoint_manager.latest_checkpoint).expect_partial(
) # reset model weights
# Produce an observation
kappa = 10**kappa_vae.sample(1)
source = tf.nn.relu(source_vae.sample(1))
source /= tf.reduce_max(source, axis=(1, 2, 3), keepdims=True)
noise_rms = 10**tf.random.uniform(shape=[1],
minval=-2.5,
maxval=-1)
fwhm = tf.random.uniform(shape=[1], minval=0.06, maxval=0.3)
psf = phys.psf_models(fwhm, cutout_size=psf_pixels)
observation = phys.noisy_forward(source, kappa, noise_rms, psf)
# RIM predictions for kappa and source
source_pred, kappa_pred, chi_squared = rim.predict(
observation, noise_rms, psf)
observation_pred = phys.forward(source_pred[-1], kappa_pred[-1],
psf)
source_o = source_pred[-1]
kappa_o = kappa_pred[-1]
# Latent code of model predictions
z_source, _ = source_vae.encoder(source_o)
z_kappa, _ = kappa_vae.encoder(log_10(kappa_o))
# Ground truth latent code for oracle metrics
z_source_gt, _ = source_vae.encoder(source)
z_kappa_gt, _ = kappa_vae.encoder(log_10(kappa))
# Re-optimize weights of the model
STEPS = args.re_optimize_steps
learning_rate_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
initial_learning_rate=args.learning_rate,
decay_rate=args.decay_rate,
decay_steps=args.decay_steps,
staircase=args.staircase)
optim = tf.keras.optimizers.RMSprop(
learning_rate=learning_rate_schedule)
chi_squared_series = tf.TensorArray(DTYPE, size=STEPS)
source_mse = tf.TensorArray(DTYPE, size=STEPS)
kappa_mse = tf.TensorArray(DTYPE, size=STEPS)
sampled_chi_squared_series = tf.TensorArray(DTYPE, size=STEPS)
sampled_source_mse = tf.TensorArray(DTYPE, size=STEPS)
sampled_kappa_mse = tf.TensorArray(DTYPE, size=STEPS)
best = chi_squared
source_best = source_pred[-1]
kappa_best = kappa_pred[-1]
source_mse_best = tf.reduce_mean((source_best - source)**2)
kappa_mse_best = tf.reduce_mean((kappa_best - log_10(kappa))**2)
# ===================== Optimization ==============================
for current_step in tqdm(range(STEPS)):
# ===================== VAE SAMPLING ==============================
# L1 distance with ground truth in latent space -- this is changed by an user defined value when using real data
# z_source_std = tf.abs(z_source - z_source_gt)
# z_kappa_std = tf.abs(z_kappa - z_kappa_gt)
z_source_std = args.source_vae_ball_size
z_kappa_std = args.kappa_vae_ball_size
# Sample latent code, then decode and forward
z_s = tf.random.normal(
shape=[args.sample_size, source_vae.latent_size],
mean=z_source,
stddev=z_source_std)
z_k = tf.random.normal(
shape=[args.sample_size, kappa_vae.latent_size],
mean=z_kappa,
stddev=z_kappa_std)
sampled_source = tf.nn.relu(source_vae.decode(z_s))
sampled_source /= tf.reduce_max(sampled_source,
axis=(1, 2, 3),
keepdims=True)
sampled_kappa = kappa_vae.decode(z_k) # output in log_10 space
sampled_observation = phys.noisy_forward(
sampled_source, 10**sampled_kappa, noise_rms,
tf.tile(psf, [args.sample_size, 1, 1, 1]))
with tf.GradientTape() as tape:
tape.watch(unet.trainable_variables)
s, k, chi_sq = rim.call(
sampled_observation,
noise_rms,
tf.tile(psf, [args.sample_size, 1, 1, 1]),
outer_tape=tape)
_kappa_mse = tf.reduce_sum(wk(10**sampled_kappa) *
(k - sampled_kappa)**2,
axis=(2, 3, 4))
cost = tf.reduce_mean(_kappa_mse)
cost += tf.reduce_mean((s - sampled_source)**2)
cost += tf.reduce_sum(rim.unet.losses) # weight decay
grads = tape.gradient(cost, unet.trainable_variables)
optim.apply_gradients(zip(grads, unet.trainable_variables))
# Record performance on sampled dataset
sampled_chi_squared_series = sampled_chi_squared_series.write(
index=current_step,
value=tf.squeeze(tf.reduce_mean(chi_sq[-1])))
sampled_source_mse = sampled_source_mse.write(
index=current_step,
value=tf.reduce_mean((s[-1] - sampled_source)**2))
sampled_kappa_mse = sampled_kappa_mse.write(
index=current_step,
value=tf.reduce_mean((k[-1] - sampled_kappa)**2))
# Record model prediction on data
s, k, chi_sq = rim.call(observation, noise_rms, psf)
chi_squared_series = chi_squared_series.write(
index=current_step, value=tf.squeeze(chi_sq))
source_o = s[-1]
kappa_o = k[-1]
# oracle metrics, remove when using real data
source_mse = source_mse.write(index=current_step,
value=tf.reduce_mean(
(source_o - source)**2))
kappa_mse = kappa_mse.write(index=current_step,
value=tf.reduce_mean(
(kappa_o - log_10(kappa))**2))
if abs(chi_sq[-1, 0] - 1) < abs(best[-1, 0] - 1):
source_best = tf.nn.relu(source_o)
kappa_best = 10**kappa_o
best = chi_sq
source_mse_best = tf.reduce_mean((source_best - source)**2)
kappa_mse_best = tf.reduce_mean(
(kappa_best - log_10(kappa))**2)
source_o = source_best
kappa_o = kappa_best
y_pred = phys.forward(source_o, kappa_o, psf)
chi_sq_series = tf.transpose(chi_squared_series.stack())
source_mse = source_mse.stack()
kappa_mse = kappa_mse.stack()
sampled_chi_squared_series = sampled_chi_squared_series.stack()
sampled_source_mse = sampled_source_mse.stack()
sampled_kappa_mse = sampled_kappa_mse.stack()
# Latent code of optimized model predictions
z_source_opt, _ = source_vae.encoder(tf.nn.relu(source_o))
z_kappa_opt, _ = kappa_vae.encoder(log_10(kappa_o))
# Compute Power spectrum of converged predictions
_ps_observation = ps_observation.cross_correlation_coefficient(
observation[..., 0], observation_pred[..., 0])
_ps_observation2 = ps_observation.cross_correlation_coefficient(
observation[..., 0], y_pred[..., 0])
_ps_kappa = ps_kappa.cross_correlation_coefficient(
log_10(kappa)[..., 0],
log_10(kappa_pred[-1])[..., 0])
_ps_kappa2 = ps_kappa.cross_correlation_coefficient(
log_10(kappa)[..., 0], log_10(kappa_o[..., 0]))
_ps_source = ps_source.cross_correlation_coefficient(
source[..., 0], source_pred[-1][..., 0])
_ps_source2 = ps_source.cross_correlation_coefficient(
source[..., 0], source_o[..., 0])
# save results
hf["observation"][i] = observation.numpy().astype(np.float32)
hf["psf"][i] = psf.numpy().astype(np.float32)
hf["psf_fwhm"][i] = fwhm.numpy().astype(np.float32)
hf["noise_rms"][i] = noise_rms.numpy().astype(np.float32)
hf["source"][i] = source.numpy().astype(np.float32)
hf["kappa"][i] = kappa.numpy().astype(np.float32)
hf["observation_pred"][i] = observation_pred.numpy().astype(
np.float32)
hf["observation_pred_reoptimized"][i] = y_pred.numpy().astype(
np.float32)
hf["source_pred"][i] = tf.transpose(
source_pred, perm=(1, 0, 2, 3, 4)).numpy().astype(np.float32)
hf["source_pred_reoptimized"][i] = source_o.numpy().astype(
np.float32)
hf["kappa_pred"][i] = tf.transpose(
kappa_pred, perm=(1, 0, 2, 3, 4)).numpy().astype(np.float32)
hf["kappa_pred_reoptimized"][i] = kappa_o.numpy().astype(
np.float32)
hf["chi_squared"][i] = tf.squeeze(chi_squared).numpy().astype(
np.float32)
hf["chi_squared_reoptimized"][i] = tf.squeeze(best).numpy().astype(
np.float32)
hf["chi_squared_reoptimized_series"][i] = chi_sq_series.numpy(
).astype(np.float32)
hf["sampled_chi_squared_reoptimized_series"][
i] = 2 * sampled_chi_squared_series.numpy().astype(np.float32)
hf["source_optim_mse"][i] = source_mse_best.numpy().astype(
np.float32)
hf["source_optim_mse_series"][i] = source_mse.numpy().astype(
np.float32)
hf["sampled_source_optim_mse_series"][
i] = sampled_source_mse.numpy().astype(np.float32)
hf["kappa_optim_mse"][i] = kappa_mse_best.numpy().astype(
np.float32)
hf["kappa_optim_mse_series"][i] = kappa_mse.numpy().astype(
np.float32)
hf["sampled_kappa_optim_mse_series"][i] = sampled_kappa_mse.numpy(
).astype(np.float32)
hf["latent_source_gt_distance_init"][i] = tf.abs(
z_source - z_source_gt).numpy().squeeze().astype(np.float32)
hf["latent_kappa_gt_distance_init"][i] = tf.abs(
z_kappa - z_kappa_gt).numpy().squeeze().astype(np.float32)
hf["latent_source_gt_distance_end"][i] = tf.abs(
z_source_opt - z_source_gt).numpy().squeeze().astype(
np.float32)
hf["latent_kappa_gt_distance_end"][i] = tf.abs(
z_kappa_opt - z_kappa_gt).numpy().squeeze().astype(np.float32)
hf["observation_coherence_spectrum"][i] = _ps_observation
hf["observation_coherence_spectrum_reoptimized"][
i] = _ps_observation2
hf["source_coherence_spectrum"][i] = _ps_source
hf["source_coherence_spectrum_reoptimized"][i] = _ps_source2
hf["kappa_coherence_spectrum"][i] = _ps_kappa
hf["kappa_coherence_spectrum_reoptimized"][i] = _ps_kappa2
if i == 0:
_, f = np.histogram(np.fft.fftfreq(phys.pixels)[:phys.pixels //
2],
bins=ps_observation.bins)
f = (f[:-1] + f[1:]) / 2
hf["observation_frequencies"][:] = f
_, f = np.histogram(np.fft.fftfreq(
phys.src_pixels)[:phys.src_pixels // 2],
bins=ps_source.bins)
f = (f[:-1] + f[1:]) / 2
hf["source_frequencies"][:] = f
_, f = np.histogram(np.fft.fftfreq(
phys.kappa_pixels)[:phys.kappa_pixels // 2],
bins=ps_kappa.bins)
f = (f[:-1] + f[1:]) / 2
hf["kappa_frequencies"][:] = f
hf["kappa_fov"][0] = phys.kappa_fov
hf["source_fov"][0] = phys.src_fov
def main(args):
files = glob.glob(os.path.join(args.dataset, "*.tfrecords"))
files = tf.data.Dataset.from_tensor_slices(files)
dataset = files.interleave(lambda x: tf.data.TFRecordDataset(
x, compression_type=args.compression_type),
block_length=1,
num_parallel_calls=tf.data.AUTOTUNE)
for physical_params in dataset.map(decode_physical_model_info):
break
dataset = dataset.map(decode_train)
# files = glob.glob(os.path.join(args.source_dataset, "*.tfrecords"))
# files = tf.data.Dataset.from_tensor_slices(files)
# source_dataset = files.interleave(lambda x: tf.data.TFRecordDataset(x, compression_type=args.compression_type),
# block_length=1, num_parallel_calls=tf.data.AUTOTUNE)
# source_dataset = source_dataset.map(decode_image).map(preprocess_image).shuffle(10000).batch(args.sample_size)
with open(os.path.join(args.kappa_vae, "model_hparams.json"), "r") as f:
kappa_vae_hparams = json.load(f)
kappa_vae = VAE(**kappa_vae_hparams)
ckpt1 = tf.train.Checkpoint(step=tf.Variable(1), net=kappa_vae)
checkpoint_manager1 = tf.train.CheckpointManager(ckpt1, args.kappa_vae, 1)
checkpoint_manager1.checkpoint.restore(
checkpoint_manager1.latest_checkpoint).expect_partial()
with open(os.path.join(args.source_vae, "model_hparams.json"), "r") as f:
source_vae_hparams = json.load(f)
source_vae = VAE(**source_vae_hparams)
ckpt2 = tf.train.Checkpoint(step=tf.Variable(1), net=source_vae)
checkpoint_manager2 = tf.train.CheckpointManager(ckpt2, args.source_vae, 1)
checkpoint_manager2.checkpoint.restore(
checkpoint_manager2.latest_checkpoint).expect_partial()
phys = PhysicalModel(pixels=physical_params["pixels"].numpy(),
kappa_pixels=physical_params["kappa pixels"].numpy(),
src_pixels=physical_params["src pixels"].numpy(),
image_fov=physical_params["image fov"].numpy(),
kappa_fov=physical_params["kappa fov"].numpy(),
src_fov=physical_params["source fov"].numpy(),
method="fft")
# simulate observations
kappa = 10**kappa_vae.sample(args.sample_size)
source = preprocess_image(source_vae.sample(args.sample_size))
# for source in source_dataset:
# break
fwhm = tf.random.normal(shape=[args.sample_size],
mean=1.5 * phys.image_fov / phys.pixels,
stddev=0.5 * phys.image_fov / phys.pixels)
# noise_rms = tf.random.normal(shape=[args.sample_size], mean=args.noise_mean, stddev=args.noise_std)
psf = phys.psf_models(fwhm, cutout_size=20)
y_vae = phys.forward(source, kappa, psf)
with h5py.File(
os.path.join(os.getenv("CENSAI_PATH"), "results",
args.output_name + ".h5"), 'w') as hf:
# rank these observations against the dataset with L2 norm
for i in tqdm(range(args.sample_size)):
distances = []
for y_d, _, _, _, _ in dataset:
distances.append(
tf.sqrt(tf.reduce_sum(
(y_d - y_vae[i][None, ...])**2)).numpy().astype(
np.float32))
k_indices = np.argsort(distances)[:args.k]
# save results
g = hf.create_group(f"sample_{i:02d}")
g.create_dataset(name="matched_source",
shape=[args.k, phys.src_pixels, phys.src_pixels],
dtype=np.float32)
g.create_dataset(
name="matched_kappa",
shape=[args.k, phys.kappa_pixels, phys.kappa_pixels],
dtype=np.float32)
g.create_dataset(name="matched_obs",
shape=[args.k, phys.pixels, phys.pixels],
dtype=np.float32)
g.create_dataset(name="matched_psf",
shape=[args.k, 20, 20],
dtype=np.float32)
g.create_dataset(name="matched_noise_rms",
shape=[args.k],
dtype=np.float32)
g.create_dataset(name="obs_L2_distance",
shape=[args.k],
dtype=np.float32)
g["vae_source"] = source[i, ..., 0].numpy().astype(np.float32)
g["vae_kappa"] = kappa[i, ..., 0].numpy().astype(np.float32)
g["vae_obs"] = y_vae[i, ..., 0].numpy().astype(np.float32)
g["vae_psf"] = psf[i, ..., 0].numpy().astype(np.float32)
for rank, j in enumerate(k_indices):
# fetch back the matched observation
for y_d, source_d, kappa_d, noise_rms_d, psf_d in dataset.skip(
j):
break
# g["vae_noise_rms"] = noise_rms[i].numpy().astype(np.float32)
g["matched_source"][rank] = source_d[..., 0].numpy().astype(
np.float32)
g["matched_kappa"][rank] = kappa_d[..., 0].numpy().astype(
np.float32)
g["matched_obs"][rank] = y_d[..., 0].numpy().astype(np.float32)
g["matched_noise_rms"][rank] = noise_rms_d.numpy().astype(
np.float32)
g["matched_psf"][rank] = psf_d[...,
0].numpy().astype(np.float32)
g["obs_L2_distance"][rank] = distances[j]
def distributed_strategy(args):
kappa_datasets = []
for path in args.kappa_datasets:
files = glob.glob(os.path.join(path, "*.tfrecords"))
files = tf.data.Dataset.from_tensor_slices(files).shuffle(
len(files), reshuffle_each_iteration=True)
dataset = files.interleave(lambda x: tf.data.TFRecordDataset(
x, compression_type=args.compression_type),
block_length=args.block_length,
num_parallel_calls=tf.data.AUTOTUNE)
kappa_datasets.append(
dataset.shuffle(args.buffer_size, reshuffle_each_iteration=True))
kappa_dataset = tf.data.experimental.sample_from_datasets(
kappa_datasets, weights=args.kappa_datasets_weights)
# Read off global parameters from first example in dataset
for example in kappa_dataset.map(decode_kappa_info):
kappa_fov = example["kappa fov"].numpy()
kappa_pixels = example["kappa pixels"].numpy()
break
kappa_dataset = kappa_dataset.map(decode_kappa).batch(args.batch_size)
cosmos_datasets = []
for path in args.cosmos_datasets:
files = glob.glob(os.path.join(path, "*.tfrecords"))
files = tf.data.Dataset.from_tensor_slices(files).shuffle(
len(files), reshuffle_each_iteration=True)
dataset = files.interleave(lambda x: tf.data.TFRecordDataset(
x, compression_type=args.compression_type),
block_length=args.block_length,
num_parallel_calls=tf.data.AUTOTUNE)
cosmos_datasets.append(
dataset.shuffle(args.buffer_size, reshuffle_each_iteration=True))
cosmos_dataset = tf.data.experimental.sample_from_datasets(
cosmos_datasets, weights=args.cosmos_datasets_weights)
# Read off global parameters from first example in dataset
for src_pixels in cosmos_dataset.map(decode_cosmos_info):
src_pixels = src_pixels.numpy()
break
cosmos_dataset = cosmos_dataset.map(decode_cosmos).map(
preprocess_cosmos).batch(args.batch_size)
window = tukey(src_pixels, alpha=args.tukey_alpha)
window = np.outer(window, window)[np.newaxis, ..., np.newaxis]
window = tf.constant(window, dtype=DTYPE)
phys = PhysicalModel(image_fov=kappa_fov,
src_fov=args.source_fov,
pixels=args.lens_pixels,
kappa_pixels=kappa_pixels,
src_pixels=src_pixels,
kappa_fov=kappa_fov,
method="conv2d")
noise_a = (args.noise_rms_min - args.noise_rms_mean) / args.noise_rms_std
noise_b = (args.noise_rms_max - args.noise_rms_mean) / args.noise_rms_std
psf_a = (args.psf_fwhm_min - args.psf_fwhm_mean) / args.psf_fwhm_std
psf_b = (args.psf_fwhm_max - args.psf_fwhm_mean) / args.psf_fwhm_std
options = tf.io.TFRecordOptions(compression_type=args.compression_type)
with tf.io.TFRecordWriter(
os.path.join(args.output_dir, f"data_{THIS_WORKER}.tfrecords"),
options) as writer:
print(
f"Started worker {THIS_WORKER} at {datetime.now().strftime('%y-%m-%d_%H-%M-%S')}"
)
for i in range((THIS_WORKER - 1) * args.batch_size, args.len_dataset,
N_WORKERS * args.batch_size):
for galaxies in cosmos_dataset: # select a random batch from our dataset that is reshuffled each iterations
break
for kappa in kappa_dataset:
break
galaxies = window * galaxies
noise_rms = truncnorm.rvs(noise_a,
noise_b,
loc=args.noise_rms_mean,
scale=args.noise_rms_std,
size=args.batch_size)
fwhm = truncnorm.rvs(psf_a,
psf_b,
loc=args.psf_fwhm_mean,
scale=args.psf_fwhm_std,
size=args.batch_size)
psf = phys.psf_models(fwhm, cutout_size=args.psf_cutout_size)
lensed_images = phys.noisy_forward(galaxies,
kappa,
noise_rms=noise_rms,
psf=psf)
records = encode_examples(kappa=kappa,
galaxies=galaxies,
lensed_images=lensed_images,
z_source=args.z_source,
z_lens=args.z_lens,
image_fov=phys.image_fov,
kappa_fov=phys.kappa_fov,
source_fov=args.source_fov,
noise_rms=noise_rms,
psf=psf,
fwhm=fwhm)
for record in records:
writer.write(record)
print(f"Finished work at {datetime.now().strftime('%y-%m-%d_%H-%M-%S')}")