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
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" linlayers: {margs.linlayers}")
logger.info(f" linchannelfactor: {margs.linchannelfactor}")
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" actnorm: {margs.actnorm}")
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, margs),
limit_samplesize=margs.samplesize)
# Create model
model = create_model(margs, simulator)
# Train
try:
trainer = ForwardTrainer(model) if simulator.parameter_dim(
) is None else ConditionalForwardTrainer(model)
common_kwargs, _, _, _ = train.make_training_kwargs(margs, dataset)
logger.info("Starting training MF: manifold training")
np.random.seed(123)
_, val_losses = trainer.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,
)
# Save
torch.save(model.state_dict(),
create_filename("model", None, margs))
# Evaluate reco error
logger.info("Evaluating reco error")
model.eval()
torch.cuda.empty_cache()
np.random.seed(123)
dataloader = trainer.make_dataloader(
simulator.load_dataset(train=True,
dataset_dir=create_filename(
"dataset", None, margs),
limit_samplesize=margs.samplesize),
args.validationsplit, 20, 4)[1]
reco_errors = []
x_plot, x_reco_plot = None, None
for x, params in dataloader:
x = x.to(device=trainer.device, dtype=trainer.dtype)
params = None if simulator.parameter_dim(
) is None else params.to(device=trainer.device,
dtype=trainer.dtype)
x_reco, _, _ = model(x, context=params, mode="projection")
reco_errors.append((torch.sum(
(x - x_reco)**2, dim=1)**0.5).detach().cpu().numpy())
if x_plot is None:
x_plot = x.detach().cpu().numpy()
x_reco_plot = x_reco.detach().cpu().numpy()
reco_error = np.mean(reco_errors)
if not np.isfinite(reco_error):
raise RuntimeError()
# Report results
logger.info("Results:")
logger.info(" reco err: %s", reco_error)
# Plot reco error
x = np.clip(np.transpose(x_plot, [0, 2, 3, 1]) / 256.0, 0.0, 1.0)
x_reco = np.clip(
np.transpose(x_reco_plot, [0, 2, 3, 1]) / 256.0, 0.0, 1.0)
plt.figure(figsize=(6 * 3.0, 5 * 3.0))
for i in range(15):
plt.subplot(5, 6, 2 * i + 1)
plt.imshow(x[i])
plt.gca().get_xaxis().set_visible(False)
plt.gca().get_yaxis().set_visible(False)
plt.subplot(5, 6, 2 * i + 2)
plt.imshow(x_reco[i])
plt.gca().get_xaxis().set_visible(False)
plt.gca().get_yaxis().set_visible(False)
plt.tight_layout()
filename = create_filename("training_plot", "reco", margs)
plt.savefig(filename.format(""))
except RuntimeError as e:
logger.info("Error during training, returning 1e9\n %s", e)
return 1e9
return reco_error