]
##########################################
########## OPTIMIZE SOLUTION ############
##########################################
model = Problem(objectives, constraints, components).to(device)
optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr)
visualizer = VisualizerOpen(dataset,
dynamics_model,
args.verbosity,
args.savedir,
training_visuals=args.train_visuals,
trace_movie=args.trace_movie)
# simulator = OpenLoopSimulator(model=model, dataset=dataset, eval_sim=not args.skip_eval_sim)
simulator = MHOpenLoopSimulator(model=model,
dataset=dataset,
eval_sim=not args.skip_eval_sim)
trainer = Trainer(model,
dataset,
optimizer,
logger=logger,
visualizer=visualizer,
simulator=simulator,
epochs=args.epochs,
eval_metric=args.eval_metric,
patience=args.patience,
warmup=args.warmup)
best_model = trainer.train()
trainer.evaluate(best_model)
logger.clean_up()
savedir=args.savedir,
verbosity=args.verbosity,
stdout=metrics)
logger.args.system = 'dpc_ref'
# device and optimizer
device = f"cuda:{args.gpu}" if args.gpu is not None else "cpu"
model = model.to(device)
optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr)
# trainer
trainer = Trainer(
model,
train_data,
dev_data,
test_data,
optimizer,
logger=logger,
epochs=args.epochs,
patience=args.patience,
eval_metric='nstep_dev_loss',
warmup=args.warmup,
)
# Train control policy
best_model = trainer.train()
best_outputs = trainer.test(best_model)
"""
# # # Plots and Analysis
"""
# plot closed loop trajectories from different initial conditions
cl_simulate(A,
B,
C,
inputs_lower_bound_penalty, inputs_upper_bound_penalty
]
##########################################
########## OPTIMIZE SOLUTION ############
##########################################
model = Problem(objectives, constraints, components).to(device)
freeze_weight(model, module_names=args.freeze)
unfreeze_weight(model, module_names=args.unfreeze)
optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr)
plot_keys = ['Y_pred', 'U_pred'] # variables to be plotted
visualizer = VisualizerClosedLoop(dataset,
dynamics_model,
plot_keys,
args.verbosity,
savedir=args.savedir)
emulator = dynamics_model
simulator = ClosedLoopSimulator(model=model,
dataset=dataset,
emulator=emulator)
trainer = Trainer(model,
dataset,
optimizer,
logger=logger,
visualizer=visualizer,
simulator=simulator,
epochs=args.epochs)
best_model = trainer.train()
trainer.evaluate(best_model)
logger.clean_up()
logger.args.system = 'integrator'
# device and optimizer
device = f"cuda:{args.gpu}" if args.gpu is not None else "cpu"
model = model.to(device)
optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr)
# simulator
simulator = ClosedLoopSimulator(model=model,
dataset=dataset,
emulator=dynamics_model,
policy=policy)
# trainer
trainer = Trainer(
model,
dataset,
optimizer,
logger=logger,
simulator=simulator,
epochs=args.epochs,
patience=args.patience,
warmup=args.warmup,
)
# Train control policy
best_model = trainer.train()
"""
# # # Plots and Analysis
"""
# plot closed loop trajectories
cl_simulate(A,
B,
policy.net,
nstep=40,
x0=1.5 * np.ones([2, 1]),