def train_manifold_flow_sequential(args, dataset, model, simulator):
""" MFMF-A training """
assert not args.specified
trainer = ManifoldFlowTrainer(model) if simulator.parameter_dim(
) is None else ConditionalManifoldFlowTrainer(model)
common_kwargs = {
"dataset": dataset,
"batch_size": args.batchsize,
"initial_lr": args.lr,
"scheduler": optim.lr_scheduler.CosineAnnealingLR,
"clip_gradient": args.clip,
"validation_split": args.validationsplit,
}
if args.weightdecay is not None:
common_kwargs["optimizer_kwargs"] = {
"weight_decay": float(args.weightdecay)
}
logger.info("Starting training MF, phase 1: manifold training")
learning_curves = trainer.train(
loss_functions=[losses.mse],
loss_labels=["MSE"],
loss_weights=[args.msefactor],
epochs=args.epochs // 2,
parameters=list(model.outer_transform.parameters()) +
list(model.encoder.parameters())
if args.algorithm == "emf" else model.outer_transform.parameters(),
callbacks=[
callbacks.save_model_after_every_epoch(
create_filename("checkpoint", None, args)[:-3] +
"_epoch_A{}.pt")
],
forward_kwargs={"mode": "projection"},
**common_kwargs,
)
learning_curves = np.vstack(learning_curves).T
logger.info("Starting training MF, phase 2: density training")
learning_curves_ = trainer.train(
loss_functions=[losses.nll],
loss_labels=["NLL"],
loss_weights=[args.nllfactor],
epochs=args.epochs - (args.epochs // 2),
parameters=model.inner_transform.parameters(),
callbacks=[
callbacks.save_model_after_every_epoch(
create_filename("checkpoint", None, args)[:-3] +
"_epoch_B{}.pt")
],
forward_kwargs={"mode": "mf-fixed-manifold"},
**common_kwargs,
)
learning_curves_ = np.vstack(learning_curves_).T
learning_curves = learning_curves_ if learning_curves is None else np.vstack(
(learning_curves, learning_curves_))
return learning_curves
def train_pie(args, dataset, model, simulator):
""" PIE training """
trainer = ForwardTrainer(model) if simulator.parameter_dim(
) is None else ConditionalForwardTrainer(
model) if args.scandal is None else SCANDALForwardTrainer(model)
common_kwargs, scandal_loss, scandal_label, scandal_weight = make_training_kwargs(
args, dataset)
callbacks_ = [
callbacks.save_model_after_every_epoch(
create_filename("checkpoint", None, args))
]
if simulator.is_image():
callbacks_.append(
callbacks.plot_sample_images(
create_filename("training_plot", None, args),
context=None if simulator.parameter_dim() is None else
torch.zeros(30, simulator.parameter_dim())))
callbacks_.append(
callbacks.plot_reco_images(
create_filename("training_plot", "reco_epoch", args)))
logger.info("Starting training PIE on NLL")
learning_curves = trainer.train(
loss_functions=[losses.nll] + scandal_loss,
loss_labels=["NLL"] + scandal_label,
loss_weights=[args.nllfactor * nat_to_bit_per_dim(args.datadim)] +
scandal_weight,
epochs=args.epochs,
callbacks=callbacks_,
forward_kwargs={"mode": "pie"},
initial_epoch=args.startepoch,
**common_kwargs,
)
learning_curves = np.vstack(learning_curves).T
return learning_curves
def main():
# NOTE: All the following information will come from the UI once it's created
template_fname = (
r'C:/Users/erins/OneDrive - University of North Carolina at Chapel Hill/Protocols and SOPs/'
r'MARG-PTC-001 Primary Cell Spinoculation and Latency/MARG-PTC-001a-v1_1-Lewin Template - Copy.xlsx'
)
out_dir = r'C:/Users/erins/OneDrive - University of North Carolina at Chapel Hill/Protocols and SOPs/' \
r'MARG-PTC-001 Primary Cell Spinoculation and Latency/'
template_name: str = 'test'
template_desc: str = 'test'
########
outfile: str = create_filename(out_dir,
'formulas and names',
extension='xlsx')
template_data: dict = process_template_file(template_fname, outfile)
for name, items in template_data:
output_formulas_to_excel(outfile, name, items)
filename: str = create_filename(out_dir, template_desc)
# NOTE: Before the document is created, will need to update the formulas with:
# - Manual formula variable names
# - LaTeX for manual formula
# - How to account for potential multiple results?
# - These are formulas with Ifs, MIN, MAX
# - Dealing with Table formulas
document: Document = create_document(template_data, template_name,
template_desc)
document.save(filename)
def evaluate_model_samples(args, simulator, x_gen):
""" Evaluate model samples and save results """
logger.info("Calculating likelihood of generated samples")
try:
if simulator.parameter_dim() is None:
log_likelihood_gen = simulator.log_density(x_gen)
else:
params = simulator.default_parameters(true_param_id=args.trueparam)
params = np.asarray([params for _ in range(args.generate)])
log_likelihood_gen = simulator.log_density(x_gen,
parameters=params)
log_likelihood_gen[np.isnan(log_likelihood_gen)] = -1.0e-12
np.save(create_filename("results", "samples_likelihood", args),
log_likelihood_gen)
except IntractableLikelihoodError:
logger.info("True simulator likelihood is intractable for dataset %s",
args.dataset)
# Distance from manifold
try:
logger.info("Calculating distance from manifold of generated samples")
distances_gen = simulator.distance_from_manifold(x_gen)
np.save(create_filename("results", "samples_manifold_distance", args),
distances_gen)
except NotImplementedError:
logger.info("Cannot calculate distance from manifold for dataset %s",
args.dataset)
def evaluate_model_samples(args, simulator, x_gen):
""" Evaluate model samples and save results """
logger.info("Calculating likelihood of generated samples")
try:
if simulator.parameter_dim() is None:
log_likelihood_gen = simulator.log_density(x_gen)
else:
params = simulator.default_parameters(true_param_id=args.trueparam)
params = np.asarray([params for _ in range(args.generate)])
log_likelihood_gen = simulator.log_density(x_gen,
parameters=params)
log_likelihood_gen[np.isnan(log_likelihood_gen)] = -1.0e-12
np.save(create_filename("results", "samples_likelihood", args),
log_likelihood_gen)
except IntractableLikelihoodError:
logger.info("True simulator likelihood is intractable for dataset %s",
args.dataset)
logger.info("Calculating distance from manifold of generated samples")
try:
distances_gen = simulator.distance_from_manifold(x_gen)
np.save(create_filename("results", "samples_manifold_distance", args),
distances_gen)
except NotImplementedError:
logger.info("Cannot calculate distance from manifold for dataset %s",
args.dataset)
if simulator.is_image():
if calculate_fid_given_paths is None:
logger.warning(
"Cannot compute FID score, did not find FID implementation")
return
logger.info("Calculating FID score of generated samples")
# The FID script needs an image folder
with tempfile.TemporaryDirectory() as gen_dir:
logger.debug(
f"Storing generated images in temporary folder {gen_dir}")
array_to_image_folder(x_gen, gen_dir)
true_dir = create_filename("dataset", None, args) + "/test"
os.makedirs(os.path.dirname(true_dir), exist_ok=True)
if not os.path.exists(f"{true_dir}/0.jpg"):
array_to_image_folder(
simulator.load_dataset(train=False,
numpy=True,
dataset_dir=create_filename(
"dataset", None, args),
true_param_id=args.trueparam)[0],
true_dir)
logger.debug("Beginning FID calculation with batchsize 50")
fid = calculate_fid_given_paths([gen_dir, true_dir], 50, "", 2048)
logger.info(f"FID = {fid}")
np.save(create_filename("results", "samples_fid", args), [fid])
def train_specified_manifold_flow(args, dataset, model, simulator):
""" FOM training """
trainer = ForwardTrainer(model) if simulator.parameter_dim(
) is None else ConditionalForwardTrainer(
model) if args.scandal is None else SCANDALForwardTrainer(model)
common_kwargs, scandal_loss, scandal_label, scandal_weight = make_training_kwargs(
args, dataset)
logger.info("Starting training MF with specified manifold on NLL")
learning_curves = trainer.train(
loss_functions=[losses.mse, losses.nll] + scandal_loss,
loss_labels=["MSE", "NLL"] + scandal_label,
loss_weights=[
0.0, args.nllfactor * nat_to_bit_per_dim(args.modellatentdim)
] + scandal_weight,
epochs=args.epochs,
callbacks=[
callbacks.save_model_after_every_epoch(
create_filename("checkpoint", None, args))
],
forward_kwargs={"mode": "mf"},
initial_epoch=args.startepoch,
**common_kwargs,
)
learning_curves = np.vstack(learning_curves).T
return learning_curves
def train_generative_adversarial_manifold_flow(args, dataset, model,
simulator):
""" MFMF-OT training """
gen_trainer = AdversarialTrainer(model) if simulator.parameter_dim(
) is None else ConditionalAdversarialTrainer(model)
common_kwargs, scandal_loss, scandal_label, scandal_weight = make_training_kwargs(
args, dataset)
common_kwargs["batch_size"] = args.genbatchsize
logger.info("Starting training GAMF: Sinkhorn-GAN")
callbacks_ = [
callbacks.save_model_after_every_epoch(
create_filename("checkpoint", None, args))
]
if args.debug:
callbacks_.append(callbacks.print_mf_weight_statistics())
learning_curves_ = gen_trainer.train(
loss_functions=[losses.make_sinkhorn_divergence()],
loss_labels=["GED"],
loss_weights=[args.sinkhornfactor],
epochs=args.epochs,
callbacks=callbacks_,
compute_loss_variance=True,
initial_epoch=args.startepoch,
**common_kwargs,
)
learning_curves = np.vstack(learning_curves_).T
return learning_curves
def run_mcmc(args, simulator, model=None):
""" MCMC """
logger.info(
"Starting MCMC based on %s after %s observed samples, generating %s posterior samples with %s for parameter point number %s",
"true simulator likelihood" if model is None else "neural likelihood estimate",
args.observedsamples,
args.mcmcsamples,
"slice sampler" if args.slicesampler else "Metropolis-Hastings sampler (step = {})".format(args.mcmcstep),
args.trueparam,
)
# Data
true_parameters = simulator.default_parameters(true_param_id=args.trueparam)
x_obs, _ = simulator.load_dataset(
train=False, numpy=True, dataset_dir=create_filename("dataset", None, args), true_param_id=args.trueparam, joint_score=False, limit_samplesize=args.observedsamples
)
x_obs_ = torch.tensor(x_obs, dtype=torch.float)
if model is None:
# MCMC based on ground truth likelihood
def log_posterior(params):
log_prob = np.sum(simulator.log_density(x_obs, parameters=params))
log_prob += simulator.evaluate_log_prior(params)
return float(log_prob)
else:
# MCMC based on neural likelihood estimator
def log_posterior(params):
params_ = np.broadcast_to(params.reshape((-1, params.shape[-1])), (x_obs.shape[0], params.shape[-1]))
params_ = torch.tensor(params_, dtype=torch.float)
if args.algorithm == "flow":
log_prob = np.sum(model.log_prob(x_obs_, context=params_).detach().numpy())
elif args.algorithm in ["pie", "slice"]:
log_prob = np.sum(model.log_prob(x_obs_, context=params_, mode=args.algorithm).detach().numpy())
elif not args.conditionalouter:
# Slow part of Jacobian drops out in LLR / MCMC acceptance ratio
log_prob = np.sum(model.log_prob(x_obs_, context=params_, mode="mf-fixed-manifold").detach().numpy())
else:
log_prob = np.sum(model.log_prob(x_obs_, context=params_, mode="mf").detach().numpy())
log_prob += simulator.evaluate_log_prior(params)
return float(log_prob)
if args.slicesampler:
logger.debug("Initializing slice sampler")
sampler = mcmc.SliceSampler(true_parameters, log_posterior, thin=args.thin)
else:
logger.debug("Initializing Gaussian Metropolis-Hastings sampler")
sampler = mcmc.GaussianMetropolis(true_parameters, log_posterior, step=args.mcmcstep, thin=args.thin)
if args.burnin > 0:
logger.info("Starting burn in")
sampler.gen(args.burnin)
logger.info("Burn in done, starting main chain")
posterior_samples = sampler.gen(args.mcmcsamples)
logger.info("MCMC done")
return posterior_samples
def train_pie(args, dataset, model, simulator):
""" PIE training """
trainer = ManifoldFlowTrainer(model) if simulator.parameter_dim(
) is None else ConditionalManifoldFlowTrainer(model)
logger.info("Starting training PIE on NLL")
common_kwargs = {
"dataset": dataset,
"batch_size": args.batchsize,
"initial_lr": args.lr,
"scheduler": optim.lr_scheduler.CosineAnnealingLR,
"clip_gradient": args.clip,
"validation_split": args.validationsplit,
}
if args.weightdecay is not None:
common_kwargs["optimizer_kwargs"] = {
"weight_decay": float(args.weightdecay)
}
learning_curves = trainer.train(
loss_functions=[losses.nll],
loss_labels=["NLL"],
loss_weights=[args.nllfactor],
epochs=args.epochs,
callbacks=[
callbacks.save_model_after_every_epoch(
create_filename("checkpoint", None, args)[:-3] +
"_epoch_{}.pt")
],
forward_kwargs={"mode": "pie"},
**common_kwargs,
)
learning_curves = np.vstack(learning_curves).T
return learning_curves
def sample_from_model(args, model, simulator, batchsize=200):
""" Generate samples from model and store """
logger.info("Sampling from model")
x_gen_all = []
while len(x_gen_all) < args.generate:
n = min(batchsize, args.generate - len(x_gen_all))
if simulator.parameter_dim() is None:
x_gen = model.sample(n=n).detach().numpy()
elif args.trueparam is None: # Sample from prior
params = simulator.sample_from_prior(n)
params = torch.tensor(params, dtype=torch.float)
x_gen = model.sample(n=n, context=params).detach().numpy()
else:
params = simulator.default_parameters(true_param_id=args.trueparam)
params = np.asarray([params for _ in range(n)])
params = torch.tensor(params, dtype=torch.float)
x_gen = model.sample(n=n, context=params).detach().numpy()
x_gen_all += list(x_gen)
x_gen_all = np.array(x_gen_all)
np.save(create_filename("results", "samples", args), x_gen_all)
return x_gen_all
def train_generative_adversarial_manifold_flow_alternating(
args, dataset, model, simulator):
""" MFMF-OTA training """
assert not args.specified
gen_trainer = GenerativeTrainer(model) if simulator.parameter_dim(
) is None else ConditionalGenerativeTrainer(model)
likelihood_trainer = ManifoldFlowTrainer(model) if simulator.parameter_dim(
) is None else ConditionalManifoldFlowTrainer(model)
metatrainer = AlternatingTrainer(model, gen_trainer, likelihood_trainer)
meta_kwargs = {
"dataset": dataset,
"initial_lr": args.lr,
"scheduler": optim.lr_scheduler.CosineAnnealingLR,
"validation_split": args.validationsplit
}
if args.weightdecay is not None:
meta_kwargs["optimizer_kwargs"] = {
"weight_decay": float(args.weightdecay)
}
phase1_kwargs = {"clip_gradient": args.clip}
phase2_kwargs = {
"forward_kwargs": {
"mode": "mf-fixed-manifold"
},
"clip_gradient": args.clip
}
phase1_parameters = model.parameters()
phase2_parameters = model.inner_transform.parameters()
logger.info(
"Starting training GAMF, alternating between Sinkhorn divergence and log likelihood"
)
learning_curves_ = metatrainer.train(
loss_functions=[losses.make_sinkhorn_divergence(), losses.nll],
loss_function_trainers=[0, 1],
loss_labels=["GED", "NLL"],
loss_weights=[args.sinkhornfactor, args.nllfactor],
batch_sizes=[args.genbatchsize, args.batchsize],
epochs=args.epochs // 2,
parameters=[phase1_parameters, phase2_parameters],
callbacks=[
callbacks.save_model_after_every_epoch(
create_filename("checkpoint", None, args)[:-3] +
"_epoch_{}.pt")
],
trainer_kwargs=[phase1_kwargs, phase2_kwargs],
subsets=args.subsets,
subset_callbacks=[callbacks.print_mf_weight_statistics()]
if args.debug else None,
**meta_kwargs,
)
learning_curves = np.vstack(learning_curves_).T
return learning_curves
def sample_from_model(args, model, simulator):
""" Generate samples from model and store """
logger.info("Sampling from model")
if simulator.parameter_dim() is None:
x_gen = model.sample(n=args.generate).detach().numpy()
else:
params = simulator.default_parameters(true_param_id=args.trueparam)
params = np.asarray([params for _ in range(args.generate)])
params = torch.tensor(params, dtype=torch.float)
x_gen = model.sample(n=args.generate, context=params).detach().numpy()
np.save(create_filename("results", "samples", args), x_gen)
return x_gen
def objective(trial):
global counter
counter += 1
# Hyperparameters
margs = pick_parameters(args, trial, counter)
logger.info("Starting training for the following hyperparameters:")
for k, v in margs.__dict__.items():
logger.info(" %s: %s", k, v)
# Bug fix related to some num_workers > 1 and CUDA. Bad things happen otherwise!
torch.multiprocessing.set_start_method("spawn", force=True)
# Load data
simulator = load_simulator(margs)
dataset = load_training_dataset(simulator, margs)
# Create model
model = create_model(margs, simulator)
# Train
_ = train.train_model(margs, dataset, model, simulator)
# Save
torch.save(model.state_dict(), create_filename("model", None, margs))
# Evaluate
model.eval()
# Evaluate test samples
log_likelihood_test, reconstruction_error_test, _ = evaluate.evaluate_test_samples(
margs, simulator, model, paramscan=True)
mean_log_likelihood_test = np.mean(log_likelihood_test)
mean_reco_error_test = np.mean(reconstruction_error_test)
# Generate samples
x_gen = evaluate.sample_from_model(margs, model, simulator)
distances_gen = simulator.distance_from_manifold(x_gen)
mean_gen_distance = np.mean(distances_gen)
# Report results
logger.info("Results:")
logger.info(" test log p: %s", mean_log_likelihood_test)
logger.info(" test reco err: %s", mean_reco_error_test)
logger.info(" gen distance: %s", mean_gen_distance)
return (-1.0 * margs.metricnllfactor * mean_log_likelihood_test +
margs.metricrecoerrorfactor * mean_reco_error_test +
margs.metricdistancefactor * mean_gen_distance)
def timing(args):
logger.info(
"Timing algorithm %s with %s outer layers with transformation %s and %s inner layers with transformation %s",
args.algorithm,
args.outerlayers,
args.outertransform,
args.innerlayers,
args.innertransform,
)
# Bug fix related to some num_workers > 1 and CUDA. Bad things happen otherwise!
torch.multiprocessing.set_start_method("spawn", force=True)
if torch.cuda.is_available():
torch.set_default_tensor_type("torch.cuda.DoubleTensor")
# Loop over data dims
all_times = []
for datadim in args.datadims:
logger.info("Starting timing for %s-dimensional data", datadim)
args.datadim = datadim
# Data
data = torch.randn(args.batchsize, datadim)
data.requires_grad = True
# Model
model = create_model(args, context_features=None)
if torch.cuda.is_available():
model = model.to(torch.device("cuda"))
# Time forward pass
times = []
for _ in range(args.repeats):
time_before = time.time()
_ = model(data)
times.append(time.time() - time_before)
logger.info("Mean time: %s s", np.mean(times))
all_times.append(times)
# Save results
logger.info("Saving results")
np.save(create_filename("timing", None, args), all_times)
def create_index_file(root_dir, width_tiles, height_tiles, xchunks, ychunks,
xtiles_per_chunk, ytiles_per_chunk):
filename = os.path.join(root_dir, "world.dat")
with open(filename, "wb") as f:
tile_size = np.array([width_tiles, height_tiles], dtype=np.uint32)
other_data = np.array(
[xchunks, ychunks, xtiles_per_chunk, ytiles_per_chunk],
dtype=np.uint16)
fname_length = np.array([utils.FILENAME_LENGTH], dtype=np.uint8)
f.write(tile_size.tobytes())
f.write(other_data.tobytes())
f.write(fname_length.tobytes())
f.write(bytes(pad_filename(utils.CHUNK_DIRECTORY), encoding="utf-8"))
for y in range(ychunks):
for x in range(xchunks):
f.write(
bytes(pad_filename(utils.create_filename(x, y, xchunks)),
encoding="utf-8"))
def create_empty(directory, width_tiles, height_tiles, xtiles_per_chunk,
ytiles_per_chunk):
xchunks = int(width_tiles / xtiles_per_chunk) + (
1 if width_tiles % xtiles_per_chunk != 0 else 0)
ychunks = int(height_tiles / ytiles_per_chunk) + (
1 if height_tiles % ytiles_per_chunk != 0 else 0)
ensure_directory_exists(directory)
clean_directory(directory)
ensure_directory_exists(os.path.join(directory, utils.CHUNK_DIRECTORY))
create_index_file(directory, width_tiles, height_tiles, xchunks, ychunks,
xtiles_per_chunk, ytiles_per_chunk)
data = np.ones(xtiles_per_chunk * ytiles_per_chunk,
dtype=utils.TILE_DTYPE) * tiles.TILE_NONE
for y in range(ychunks):
for x in range(xchunks):
filename = os.path.join(directory, utils.CHUNK_DIRECTORY,
utils.create_filename(x, y, xchunks))
with open(filename, "wb") as f:
write_chunk(data, f)
def train_generative_adversarial_manifold_flow(args, dataset, model,
simulator):
""" MFMF-OT training """
gen_trainer = GenerativeTrainer(model) if simulator.parameter_dim(
) is None else ConditionalGenerativeTrainer(model)
common_kwargs = {
"dataset": dataset,
"initial_lr": args.lr,
"scheduler": optim.lr_scheduler.CosineAnnealingLR,
"clip_gradient": args.clip,
"validation_split": args.validationsplit,
}
if args.weightdecay is not None:
common_kwargs["optimizer_kwargs"] = {
"weight_decay": float(args.weightdecay)
}
logger.info("Starting training GAMF: Sinkhorn-GAN")
callbacks_ = [
callbacks.save_model_after_every_epoch(
create_filename("checkpoint", None, args)[:-3] + "_epoch_{}.pt")
]
if args.debug:
callbacks_.append(callbacks.print_mf_weight_statistics())
learning_curves_ = gen_trainer.train(
loss_functions=[losses.make_sinkhorn_divergence()],
loss_labels=["GED"],
loss_weights=[args.sinkhornfactor],
epochs=args.epochs,
callbacks=callbacks_,
batch_size=args.genbatchsize,
compute_loss_variance=True,
**common_kwargs,
)
learning_curves = np.vstack(learning_curves_).T
return learning_curves
def download_file(item, path, course, output_dir):
filepath = create_filepath(course, path)
description = item["Description"]["Html"]
topic_type = item["TopicType"]
title = item["Title"]
if topic_type == 1:
filename = create_filename(item)
full_path = f"{output_dir}/{filepath}/{filename}"
# These documents are real files that we want to download
download_from_url(f"""{ufora}{item["Url"]}""", full_path)
if item["Url"].endswith(".html"):
# HTML files on Ufora need a little special treatment
# We'll prepend a title, <base> tag and convert them to pdf
with open(full_path, "r") as f:
content = f.read()
filename_without_extension = ".".join(filename.split(".")[:-1])
description_path = f"{output_dir}/{filepath}/{filename_without_extension}.pdf"
create_metadata(description_path, content,
filename_without_extension)
new_content = f"<base href={ufora}><h1>{title}</h1>{content}"
with open(full_path, "w") as f:
f.write(new_content)
elif description:
# Choosing this filename might cause an overlap...
filename_without_extension = ".".join(filename.split(".")[:-1])
description_path = f"{output_dir}/{filepath}/{filename_without_extension}.pdf"
create_metadata(description_path, description,
filename_without_extension)
elif topic_type == 3:
# These documents are just clickable links, we'll render them in a pdf
url = item["Url"]
filename = create_filename_without_extension(item)
full_path = f"{output_dir}/{filepath}/{filename}"
create_metadata(f"{full_path}.pdf",
f"<a href={url}>{url}</a>{description}", item["Title"])
else:
print(f"Don't know this topic type: {topic_type}")
exit()
def train_dough(args, dataset, model, simulator):
""" PIE with variable epsilons training """
trainer = VariableDimensionManifoldFlowTrainer(
model) if simulator.parameter_dim(
) is None else ConditionalVariableDimensionManifoldFlowTrainer(model)
common_kwargs = {
"dataset": dataset,
"batch_size": args.batchsize,
"initial_lr": args.lr,
"scheduler": optim.lr_scheduler.CosineAnnealingLR,
"clip_gradient": args.clip,
"validation_split": args.validationsplit,
}
if args.weightdecay is not None:
common_kwargs["optimizer_kwargs"] = {
"weight_decay": float(args.weightdecay)
}
logger.info(
"Starting training dough, phase 1: NLL without latent regularization")
learning_curves = trainer.train(
loss_functions=[losses.nll],
loss_labels=["NLL"],
loss_weights=[args.nllfactor],
epochs=args.epochs,
callbacks=[
callbacks.save_model_after_every_epoch(
create_filename("checkpoint", None, args)[:-3] +
"_epoch_{}.pt")
],
l1=args.doughl1reg,
**common_kwargs,
)
learning_curves = np.vstack(learning_curves).T
return learning_curves
def evaluate_test_samples(args, simulator, filename, model=None, ood=False, n_save_reco=100):
""" Likelihood evaluation """
logger.info(
"Evaluating %s samples according to %s, %s likelihood evaluation, saving in %s",
"the ground truth" if model is None else "a trained model",
"ood" if ood else "test",
"with" if not args.skiplikelihood else "without",
filename,
)
# Prepare
x, _ = simulator.load_dataset(
train=False, numpy=True, ood=ood, dataset_dir=create_filename("dataset", None, args), true_param_id=args.trueparam, joint_score=False, limit_samplesize=args.evaluate,
)
parameter_grid = [None] if simulator.parameter_dim() is None else simulator.eval_parameter_grid(resolution=args.gridresolution)
log_probs = []
x_recos = None
reco_error = None
# Evaluate
for i, params in enumerate(parameter_grid):
logger.debug("Evaluating grid point %s / %s", i + 1, len(parameter_grid))
if model is None:
params_ = None if params is None else np.asarray([params for _ in x])
log_prob = simulator.log_density(x, parameters=params_)
else:
log_prob = []
reco_error_ = []
x_recos_ = []
n_batches = (args.evaluate - 1) // args.evalbatchsize + 1
for j in range(n_batches):
x_ = torch.tensor(x[j * args.evalbatchsize : (j + 1) * args.evalbatchsize], dtype=torch.float)
if params is None:
params_ = None
else:
params_ = np.asarray([params for _ in x_])
params_ = torch.tensor(params_, dtype=torch.float)
if args.algorithm == "flow":
x_reco, log_prob_, _ = model(x_, context=params_)
elif args.algorithm in ["pie", "slice"]:
x_reco, log_prob_, _ = model(x_, context=params_, mode=args.algorithm if not args.skiplikelihood else "projection")
else:
x_reco, log_prob_, _ = model(x_, context=params_, mode="mf" if not args.skiplikelihood else "projection")
if not args.skiplikelihood:
log_prob.append(log_prob_.detach().numpy())
reco_error_.append((sum_except_batch((x_ - x_reco) ** 2) ** 0.5).detach().numpy())
x_recos_.append(x_reco.detach().numpy())
if not args.skiplikelihood:
log_prob = np.concatenate(log_prob, axis=0)
if reco_error is None:
reco_error = np.concatenate(reco_error_, axis=0)
if x_recos is None:
x_recos = np.concatenate(x_recos_, axis=0)
if not args.skiplikelihood:
log_probs.append(log_prob)
# Save results
if len(log_probs) > 0:
if simulator.parameter_dim() is None:
log_probs = log_probs[0]
np.save(create_filename("results", filename.format("log_likelihood"), args), log_probs)
if len(x_recos) > 0:
np.save(create_filename("results", filename.format("x_reco"), args), x_recos[:n_save_reco])
if reco_error is not None:
np.save(create_filename("results", filename.format("reco_error"), args), reco_error)
if parameter_grid is not None:
np.save(create_filename("results", "parameter_grid_test", args), parameter_grid)
create_modelname(args)
logger.info("Evaluating simulator truth")
else:
create_modelname(args)
logger.info("Evaluating model %s", args.modelname)
# Bug fix related to some num_workers > 1 and CUDA. Bad things happen otherwise!
torch.multiprocessing.set_start_method("spawn", force=True)
# Data set
simulator = load_simulator(args)
# Load model
if not args.truth:
model = create_model(args, simulator=simulator)
model.load_state_dict(torch.load(create_filename("model", None, args), map_location=torch.device("cpu")))
model.eval()
else:
model = None
# Evaluate generative performance
if args.skipgeneration:
logger.info("Skipping generative evaluation")
elif not args.truth:
x_gen = sample_from_model(args, model, simulator)
evaluate_model_samples(args, simulator, x_gen)
if args.skipinference:
logger.info("Skipping all inference tasks. Have a nice day!")
exit()
if containing_file == path_item:
is_file_found = True
break
if is_file_found:
utils.chdir(path_item)
is_file_found = False
else:
print(f'NOT FOUND FILE: {path_item}')
is_allow_continue = False
return is_allow_continue
file_locs = utils.get_all_file_locs(TARGETS)
file_name = utils.create_filename(FILE_NAME_PREFIX, FILE_EXT)
path_array = DESTINATION.split('\\')
path_root = path_array[0]
utils.chdir(path_root)
utils.log_folders_destination_targets(DESTINATION, TARGETS, file_name)
if validate_target_loc():
with utils.zipfile.ZipFile(file_name, 'x', utils.zipfile.ZIP_LZMA, True) as backup_file:
backup_file.close()
# Starting pool with x workers.
with Pool(processes=MAX_PROCESSES) as pool:
# Launching multiple evaluations asynchronously *may* use more processes
multiple_results = [pool.apply_async(utils.f, (i, file_name,)) for i in file_locs]
help='save the agent and the results')
parser.add_argument('--flipped_terminals',
action='store_true',
default=False,
help='flip the rewards associated '
'with terminal 1 and terminal 2')
parser.add_argument('--flipped_actions',
action='store_true',
default=False,
help='Shuffle the actions to cancel '
'the effect of model learning')
args = parser.parse_args()
experiment_settings = get_experiment_setting(args)
domain_settings = get_domain_setting(args)
filename = create_filename(args)
print("file: ", filename)
experiment_settings['filename'] = filename
if experiment_settings['method'] == 'sarsa_lambda':
agent_config = []
from sarsa_lambda.sarsa_lambda import build_agent, load_agent
elif experiment_settings['method'] == 'MuZero':
from muzero.MuZeroAgent import MuZeroAgent, build_agent
from muzero.env import muzero_config
agent_config = muzero_config
agent_config.flippedTask = args.flipped_terminals
agent_config.flippedActions = args.flipped_actions
from muzero.MuZeroAgent import build_agent, load_agent
else:
assert False, 'HvS: Invalid method id.'
def train_manifold_flow_sequential(args, dataset, model, simulator):
""" Sequential MFMF-M/D training """
assert not args.specified
if simulator.parameter_dim() is None:
trainer1 = ForwardTrainer(model)
trainer2 = ForwardTrainer(model)
else:
trainer1 = ConditionalForwardTrainer(model)
if args.scandal is None:
trainer2 = ConditionalForwardTrainer(model)
else:
trainer2 = SCANDALForwardTrainer(model)
common_kwargs, scandal_loss, scandal_label, scandal_weight = make_training_kwargs(
args, dataset)
callbacks1 = [
callbacks.save_model_after_every_epoch(
create_filename("checkpoint", "A", args)),
callbacks.print_mf_latent_statistics(),
callbacks.print_mf_weight_statistics()
]
callbacks2 = [
callbacks.save_model_after_every_epoch(
create_filename("checkpoint", "B", args)),
callbacks.print_mf_latent_statistics(),
callbacks.print_mf_weight_statistics()
]
if simulator.is_image():
callbacks1.append(
callbacks.plot_sample_images(
create_filename("training_plot", "sample_epoch_A", args),
context=None if simulator.parameter_dim() is None else
torch.zeros(30, simulator.parameter_dim()),
))
callbacks2.append(
callbacks.plot_sample_images(
create_filename("training_plot", "sample_epoch_B", args),
context=None if simulator.parameter_dim() is None else
torch.zeros(30, simulator.parameter_dim()),
))
callbacks1.append(
callbacks.plot_reco_images(
create_filename("training_plot", "reco_epoch_A", args)))
callbacks2.append(
callbacks.plot_reco_images(
create_filename("training_plot", "reco_epoch_B", args)))
logger.info("Starting training MF, phase 1: manifold training")
learning_curves = trainer1.train(
loss_functions=[losses.smooth_l1_loss if args.l1 else losses.mse] +
([] if args.uvl2reg is None else [losses.hiddenl2reg]),
loss_labels=["L1" if args.l1 else "MSE"] +
([] if args.uvl2reg is None else ["L2_lat"]),
loss_weights=[args.msefactor] +
([] if args.uvl2reg is None else [args.uvl2reg]),
epochs=args.epochs // 2,
parameters=list(model.outer_transform.parameters()) +
list(model.encoder.parameters()) if args.algorithm == "emf" else list(
model.outer_transform.parameters()),
callbacks=callbacks1,
forward_kwargs={
"mode": "projection",
"return_hidden": args.uvl2reg is not None
},
initial_epoch=args.startepoch,
**common_kwargs,
)
learning_curves = np.vstack(learning_curves).T
logger.info("Starting training MF, phase 2: density training")
learning_curves_ = trainer2.train(
loss_functions=[losses.nll] + scandal_loss,
loss_labels=["NLL"] + scandal_label,
loss_weights=[
args.nllfactor * nat_to_bit_per_dim(args.modellatentdim)
] + scandal_weight,
epochs=args.epochs - (args.epochs // 2),
parameters=list(model.inner_transform.parameters()),
callbacks=callbacks2,
forward_kwargs={"mode": "mf-fixed-manifold"},
initial_epoch=args.startepoch - args.epochs // 2,
**common_kwargs,
)
learning_curves = np.vstack(
(learning_curves, np.vstack(learning_curves_).T))
return learning_curves
def train_manifold_flow_alternating(args, dataset, model, simulator):
""" MFMF-A training """
assert not args.specified
trainer1 = ForwardTrainer(model) if simulator.parameter_dim(
) is None else ConditionalForwardTrainer(model)
trainer2 = ForwardTrainer(model) if simulator.parameter_dim(
) is None else ConditionalForwardTrainer(
model) if args.scandal is None else SCANDALForwardTrainer(model)
metatrainer = AlternatingTrainer(model, trainer1, trainer2)
meta_kwargs = {
"dataset": dataset,
"initial_lr": args.lr,
"scheduler": optim.lr_scheduler.CosineAnnealingLR,
"validation_split": args.validationsplit
}
if args.weightdecay is not None:
meta_kwargs["optimizer_kwargs"] = {
"weight_decay": float(args.weightdecay)
}
_, scandal_loss, scandal_label, scandal_weight = make_training_kwargs(
args, dataset)
phase1_kwargs = {
"forward_kwargs": {
"mode": "projection"
},
"clip_gradient": args.clip
}
phase2_kwargs = {
"forward_kwargs": {
"mode": "mf-fixed-manifold"
},
"clip_gradient": args.clip
}
phase1_parameters = list(model.outer_transform.parameters()) + list(
model.encoder.parameters(
)) if args.algorithm == "emf" else model.outer_transform.parameters()
phase2_parameters = list(model.inner_transform.parameters())
logger.info(
"Starting training MF, alternating between reconstruction error and log likelihood"
)
learning_curves_ = metatrainer.train(
loss_functions=[
losses.smooth_l1_loss if args.l1 else losses.mse, losses.nll
] + scandal_loss,
loss_function_trainers=[0, 1] +
[1] if args.scandal is not None else [],
loss_labels=["L1" if args.l1 else "MSE", "NLL"] + scandal_label,
loss_weights=[
args.msefactor,
args.nllfactor * nat_to_bit_per_dim(args.modellatentdim)
] + scandal_weight,
epochs=args.epochs // 2,
subsets=args.subsets,
batch_sizes=[args.batchsize, args.batchsize],
parameters=[phase1_parameters, phase2_parameters],
callbacks=[
callbacks.save_model_after_every_epoch(
create_filename("checkpoint", None, args))
],
trainer_kwargs=[phase1_kwargs, phase2_kwargs],
**meta_kwargs,
)
learning_curves = np.vstack(learning_curves_).T
return learning_curves
def train_manifold_flow(args, dataset, model, simulator):
""" MFMF-S training """
assert not args.specified
trainer = ForwardTrainer(model) if simulator.parameter_dim(
) is None else ConditionalForwardTrainer(
model) if args.scandal is None else SCANDALForwardTrainer(model)
common_kwargs, scandal_loss, scandal_label, scandal_weight = make_training_kwargs(
args, dataset)
logger.info(
"Starting training MF, phase 1: pretraining on reconstruction error")
learning_curves = trainer.train(
loss_functions=[losses.mse],
loss_labels=["MSE"],
loss_weights=[args.msefactor],
epochs=args.epochs // 3,
callbacks=[
callbacks.save_model_after_every_epoch(
create_filename("checkpoint", "A", args))
],
forward_kwargs={"mode": "projection"},
initial_epoch=args.startepoch,
**common_kwargs,
)
learning_curves = np.vstack(learning_curves).T
logger.info("Starting training MF, phase 2: mixed training")
learning_curves_ = trainer.train(
loss_functions=[losses.mse, losses.nll] + scandal_loss,
loss_labels=["MSE", "NLL"] + scandal_label,
loss_weights=[
args.msefactor,
args.addnllfactor * nat_to_bit_per_dim(args.modellatentdim)
] + scandal_weight,
epochs=args.epochs - 2 * (args.epochs // 3),
parameters=list(model.parameters()),
callbacks=[
callbacks.save_model_after_every_epoch(
create_filename("checkpoint", "B", args))
],
forward_kwargs={"mode": "mf"},
initial_epoch=args.startepoch - (args.epochs // 3),
**common_kwargs,
)
learning_curves_ = np.vstack(learning_curves_).T
learning_curves = learning_curves_ if learning_curves is None else np.vstack(
(learning_curves, learning_curves_))
logger.info(
"Starting training MF, phase 3: training only inner flow on NLL")
learning_curves_ = trainer.train(
loss_functions=[losses.mse, losses.nll] + scandal_loss,
loss_labels=["MSE", "NLL"] + scandal_label,
loss_weights=[
0.0, args.nllfactor * nat_to_bit_per_dim(args.modellatentdim)
] + scandal_weight,
epochs=args.epochs // 3,
parameters=list(model.inner_transform.parameters()),
callbacks=[
callbacks.save_model_after_every_epoch(
create_filename("checkpoint", "C", args))
],
forward_kwargs={"mode": "mf-fixed-manifold"},
initial_epoch=args.startepoch - (args.epochs - (args.epochs // 3)),
**common_kwargs,
)
learning_curves_ = np.vstack(learning_curves_).T
learning_curves = np.vstack(
(learning_curves, np.vstack(learning_curves_).T))
return learning_curves
if __name__ == "__main__":
# Logger
args = parse_args()
logging.basicConfig(
format="%(asctime)-5.5s %(name)-20.20s %(levelname)-7.7s %(message)s",
datefmt="%H:%M",
level=logging.DEBUG if args.debug else logging.INFO)
logger.info("Hi!")
logger.debug("Starting train.py with arguments %s", args)
create_modelname(args)
if args.resume is not None:
resume_filename = create_filename("resume", None, args)
args.startepoch = args.resume
logger.info(
"Resuming training. Loading file %s and continuing with epoch %s.",
resume_filename, args.resume + 1)
elif args.load is None:
logger.info("Training model %s with algorithm %s on data set %s",
args.modelname, args.algorithm, args.dataset)
else:
logger.info(
"Loading model %s and training it as %s with algorithm %s on data set %s",
args.load, args.modelname, args.algorithm, args.dataset)
# Bug fix related to some num_workers > 1 and CUDA. Bad things happen otherwise!
torch.multiprocessing.set_start_method("spawn", force=True)
# Generate samples
logger.info("Evaluating sample closure")
x_gen = evaluate.sample_from_model(margs, model, simulator)
distances_gen = simulator.distance_from_manifold(x_gen)
mean_gen_distance = np.mean(distances_gen)
# Report results
logger.info("Results:")
logger.info(" reco err: %s", reco_error)
logger.info(" gen distance: %s", mean_gen_distance)
return margs.metricrecoerrorfactor * reco_error + margs.metricdistancefactor * mean_gen_distance
# Load saved study object
if args.resumestudy:
filename = create_filename("paramscan", None, args)
logger.info("Loading parameter scan from %s", filename)
with open(filename, "rb") as file:
study = pickle.load(file)
else:
study = optuna.create_study(study_name=args.paramscanstudyname,
direction="minimize")
# Optimize!
try:
study.optimize(objective, n_trials=args.trials)
except (KeyboardInterrupt, SystemExit):
logger.warning("Optimization interrupted!")
def objective(trial):
global counter
counter += 1
# Hyperparameters
margs = pick_parameters(args, trial, counter)
logger.info(f"Starting run {counter} / {args.trials}")
logger.info(f"Hyperparams:")
logger.info(f" outer layers: {margs.outerlayers}")
logger.info(f" inner layers: {margs.innerlayers}")
logger.info(f" linear transform: {margs.lineartransform}")
logger.info(f" spline range: {margs.splinerange}")
logger.info(f" spline bins: {margs.splinebins}")
logger.info(f" batchnorm: {margs.batchnorm}")
logger.info(f" dropout: {margs.dropout}")
logger.info(f" batch size: {margs.batchsize}")
logger.info(f" MSE factor: {margs.msefactor}")
logger.info(f" latent L2 reg: {margs.uvl2reg}")
logger.info(f" weight decay: {margs.weightdecay}")
logger.info(f" gradient clipping: {margs.clip}")
# Bug fix related to some num_workers > 1 and CUDA. Bad things happen otherwise!
torch.multiprocessing.set_start_method("spawn", force=True)
# Load data
simulator = load_simulator(margs)
dataset = simulator.load_dataset(train=True,
dataset_dir=create_filename(
"dataset", None, args),
limit_samplesize=margs.samplesize)
# Create model
model = create_model(margs, simulator)
# Train
trainer1 = ForwardTrainer(model) if simulator.parameter_dim(
) is None else ConditionalForwardTrainer(model)
trainer2 = ForwardTrainer(model) if simulator.parameter_dim(
) is None else ConditionalForwardTrainer(model)
common_kwargs, _, _, _ = train.make_training_kwargs(margs, dataset)
logger.info("Starting training MF, phase 1: manifold training")
np.random.seed(123)
_, val_losses = trainer1.train(
loss_functions=[losses.mse, losses.hiddenl2reg],
loss_labels=["MSE", "L2_lat"],
loss_weights=[
margs.msefactor,
0.0 if margs.uvl2reg is None else margs.uvl2reg
],
epochs=margs.epochs,
parameters=(list(model.outer_transform.parameters()) +
list(model.encoder.parameters()) if args.algorithm
== "emf" else model.outer_transform.parameters()),
forward_kwargs={
"mode": "projection",
"return_hidden": True
},
**common_kwargs,
)
logger.info("Starting training MF, phase 2: density training")
np.random.seed(123)
_ = trainer2.train(
loss_functions=[losses.nll],
loss_labels=["NLL"],
loss_weights=[args.nllfactor],
epochs=args.densityepochs,
parameters=model.inner_transform.parameters(),
forward_kwargs={"mode": "mf-fixed-manifold"},
**common_kwargs,
)
# Save
torch.save(model.state_dict(), create_filename("model", None, margs))
# Evaluate reco error
logger.info("Evaluating reco error")
model.eval()
np.random.seed(123)
x, params = next(
iter(
trainer1.make_dataloader(
simulator.load_dataset(train=True,
dataset_dir=create_filename(
"dataset", None, args),
limit_samplesize=args.samplesize),
args.validationsplit, 1000, 0)[1]))
x = x.to(device=trainer1.device, dtype=trainer1.dtype)
params = None if simulator.parameter_dim() is None else params.to(
device=trainer1.device, dtype=trainer1.dtype)
x_reco, _, _ = model(x, context=params, mode="projection")
reco_error = torch.mean(torch.sum((x - x_reco)**2,
dim=1)**0.5).detach().cpu().numpy()
# Generate samples
logger.info("Evaluating sample closure")
x_gen = evaluate.sample_from_model(margs, model, simulator)
distances_gen = simulator.distance_from_manifold(x_gen)
mean_gen_distance = np.mean(distances_gen)
# Report results
logger.info("Results:")
logger.info(" reco err: %s", reco_error)
logger.info(" gen distance: %s", mean_gen_distance)
return margs.metricrecoerrorfactor * reco_error + margs.metricdistancefactor * mean_gen_distance
logger.info("Evaluating simulator truth")
else:
create_modelname(args)
logger.info("Evaluating model %s", args.modelname)
# Bug fix related to some num_workers > 1 and CUDA. Bad things happen otherwise!
torch.multiprocessing.set_start_method("spawn", force=True)
# Data set
simulator = load_simulator(args)
# Load model
if not args.truth:
model = create_model(args, simulator=simulator)
model.load_state_dict(
torch.load(create_filename("model", None, args),
map_location=torch.device("cpu")))
model.eval()
else:
model = None
# Evaluate generative performance
if args.skipgeneration:
logger.info("Skipping generative evaluation as per request.")
elif not args.truth:
x_gen = sample_from_model(args, model, simulator)
evaluate_model_samples(args, simulator, x_gen)
if args.skipinference:
logger.info(
"Skipping all inference tasks as per request. Have a nice day!")