def plot_val(name,
width=64,
lstm_width=49,
nu=1,
ny=2,
layers=2,
train_batches=50,
nl=None,
T=1000,
plot_str='k',
x0=None,
LSTM=False):
path = "./experimental_results/pendulum/"
non_lin = torch.relu if nl is None else nl
data = io.loadmat(path + name + ".mat")
if LSTM:
model = lstm.lstm(nu, lstm_width, ny, layers)
model.load_state_dict(torch.load(path + "p_" + name))
model.output_layer = model.output
else:
model = diRNN.diRNN(nu,
width,
ny,
layers,
nBatches=train_batches,
nl=non_lin,
learn_init_state=False)
model.load_state_dict(torch.load(path + "p_" + name))
u, yest = get_ic_response(model, x0, T=T, batches=1000)
plt.plot(yest[:, 0].detach().numpy().T, yest[:, 1].detach().numpy().T,
plot_str)
plt.pause(0.01)
# Optimize over models with weights confined to spectral norm ball (I.e. what Miller and Hardt did)
name = "spectral_norm_sub{:d}_val{:d}".format(subject, val_set)
model = dnb.dnbRNN(nu, width, ny, layers, nBatches=9, init_var=1.3)
model.project_norm_ball(eps)
log, best_model = train.train_model(model,
train_loader=train_loader,
val_loader=val_loader,
test_loader=test_loader,
options=solver_options,
LMIs=model.norm_ball_lmi(eps))
test_and_save_model(name, best_model, train_loader, val_loader,
test_loader, log)
# Train Contracting model
name = "contracting_sub{:d}_val{:d}".format(subject, val_set)
model = diRNN.diRNN(nu, width, ny, layers, nBatches=9)
# model.project_l2(mu=mu, epsilon=eps)
model.init_l2(mu=mu, epsilon=eps, init_var=1.3)
log, best_model = train.train_model(model,
train_loader=train_loader,
val_loader=val_loader,
test_loader=test_loader,
options=solver_options,
LMIs=model.contraction_lmi(
mu=mu, epsilon=eps))
test_and_save_model(name, best_model, train_loader, val_loader,
test_loader, log)
max_epochs=max_epochs,
lr=lr,
mu0=10,
lr_decay=lr_decay,
patience=patience)
# for gamma in [0.5, 1, 1.5, 2.5, 5]:
for gamma in [0.5, 1, 1.5, 2.5, 5]:
# Train l2 gain bounded implicit model ------------------------------------------------------------------------
name = "dl2_gamma{:1.2f}_sub{:d}_val{:d}".format(
gamma, subject, val_set)
print("training model: " + name)
model = diRNN.diRNN(nu,
width,
ny,
layers,
nBatches=9,
nl=torch.tanh)
# Add 0.1 to ensure we are strictly feasible after initialization
model.init_dl2(epsilon=eps + init_offset,
gamma=gamma,
init_var=init_var,
custom_seed=val_set + this_seed)
log, best_model = train.train_model_ipm(model,
train_loader=train_loader,
val_loader=val_loader,
test_loader=test_loader,
options=solver_options,
LMIs=model.dl2_lmi(
gamma=gamma,
nu = train_loader.nu
ny = train_loader.ny
# Options for the solver
solver_options = nlsdp.make_stochastic_nlsdp_options(
max_epochs=max_epochs,
lr=0.1E-4,
mu0=2000,
lr_decay=0.96,
patience=10)
# Train Contracting implicit model ------------------------------------------------------------------------
name = "contracting_sub{:d}_val{:d}".format(subject, val_set)
model = diRNN.diRNN(nu,
width,
ny,
layers,
nBatches=dataset_options["train_batch_size"] -
1)
model.init_l2(mu=mu, epsilon=eps, init_var=1.2)
log, best_model = train.train_model(model,
train_loader=train_loader,
val_loader=val_loader,
test_loader=test_loader,
options=solver_options,
LMIs=model.contraction_lmi(
mu=mu, epsilon=eps))
# eModel = best_model.make_explicit()
# compare_exp_imp(eModel, best_model, train_loader)
test_and_save_model(name, best_model, train_loader, val_loader,
ss = 0.5
plot_sim_step_response(step_size=ss, T=T, Ts=0.1, meas_sd=0.01)
gammas = [0.2, 0.5, 1.0, 5.0, 10.0, 50.0]
colors = cm.hot(np.linspace(0.25, 0.75, gammas.__len__()))
for val_set in range(1, 9):
for (cid, gamma) in enumerate(gammas):
color = colors[cid]
plot_args = {"color": color, "linestyle": '-'}
model_c = diRNN.diRNN(nu,
width,
ny,
layers,
nBatches=train_batches,
nl=torch.relu,
learn_init_state=False)
model_c.load_state_dict(
torch.load(
"./experimental_results/pendulum/p_dl2_gamma{:1.1f}_val{:d}"
.format(gamma, val_set)))
plot_step_response(model_c, step_size=ss, T=T, plot_args=plot_args)
plot_args = {"color": 'g', "linestyle": '-'}
model_c = diRNN.diRNN(nu,
width,
ny,
layers,
nBatches=train_batches,