config = {
'n_in': 2,
'hidden_dims': [30, 30, 30, 30, 30],
'n_out': 1,
'library_function': library_1D_in,
'library_args': {
'poly_order': 2,
'diff_order': 3
}
}
n_runs = 5
for run_idx in np.arange(n_runs):
X_train, y_train, rand_idx = dataset.create_dataset(x_grid.reshape(-1, 1),
t_grid.reshape(-1, 1),
n_samples=1000,
noise=0.1,
random=True,
return_idx=True)
theta = dataset.library(x_grid.reshape(-1, 1),
t_grid.reshape(-1, 1),
poly_order=2,
deriv_order=3)[rand_idx, :]
dt = dataset.time_deriv(x_grid.reshape(-1, 1),
t_grid.reshape(-1, 1))[rand_idx, :]
model = DeepMod(**config)
optimizer = torch.optim.Adam(model.parameters(),
betas=(0.99, 0.999),
amsgrad=True)
train(model,
# Settings for reproducibility
np.random.seed(42)
torch.manual_seed(0)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# Making data
v = 0.2
A = 1.0
x = np.linspace(-3, 4, 100)
t = np.linspace(0.5, 5.0, 50)
x_grid, t_grid = np.meshgrid(x, t, indexing='ij')
dataset = Dataset(BurgersDelta, v=v, A=A)
X_train, y_train = dataset.create_dataset(x_grid.reshape(-1, 1),
t_grid.reshape(-1, 1),
n_samples=1000,
noise=0.2)
# Configuring model
network = NN(2, [30, 30, 30, 30, 30], 1) # Function approximator
library = Library1D(poly_order=2, diff_order=2) # Library function
estimator = PINN([2, 4]) # active terms are 2 and 5
constraint = LeastSquares() # How to constrain
model = DeepMoD(network, library, estimator,
constraint) # Putting it all in the model
# Running model
sparsity_scheduler = Periodic(initial_epoch=0,
periodicity=1) # Defining when to apply sparsity
optimizer = torch.optim.Adam(model.parameters(),
betas=(0.99, 0.999),
from sklearn.linear_model import LassoLarsIC
if torch.cuda.is_available():
torch.set_default_tensor_type('torch.cuda.FloatTensor')
np.random.seed(42)
torch.manual_seed(42)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
v = 0.1
A = 1.0
# Making grid
x = np.linspace(-3, 4, 100)
t = np.linspace(0.5, 5.0, 50)
x_grid, t_grid = np.meshgrid(x, t, indexing='ij')
dataset = Dataset(BurgersDelta, v=v, A=A)
noise_range = np.arange(0.0, 1.01, 0.10)
n_runs = 5
for noise_level in noise_range:
for run in np.arange(n_runs):
X_train, y_train = dataset.create_dataset(x_grid.reshape(-1, 1), t_grid.reshape(-1, 1), n_samples=1000, noise=noise_level, random=True, return_idx=False, random_state=run)
estimator = Clustering(estimator=LassoLarsIC(fit_intercept=False))
config = {'n_in': 2, 'hidden_dims': [30, 30, 30, 30, 30], 'n_out': 1, 'library_function':library_1D_in, 'library_args':{'poly_order':2, 'diff_order': 3}, 'sparsity_estimator': estimator}
model = DeepModDynamic(**config)
optimizer = torch.optim.Adam(model.parameters(), betas=(0.99, 0.999), amsgrad=True)
train_dynamic(model, X_train, y_train, optimizer, 10000, log_dir=f'runs/cluster_{noise_level:.2f}_run_{run}/')
# Settings for reproducibility
np.random.seed(42)
torch.manual_seed(0)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# Making training set
dataset = Dataset(DoubleSoliton, c=(5, 2), x0=(-3, -1))
x_sample = np.linspace(-7, 5, 50)
t_sample = np.linspace(0.0, 1.0, 40)
x_grid_sample, t_grid_sample = np.meshgrid(x_sample, t_sample, indexing='ij')
X_train, y_train = dataset.create_dataset(x_grid_sample.reshape(-1, 1),
t_grid_sample.reshape(-1, 1),
n_samples=0,
noise=0.1,
normalize=True,
random=True)
# Configuring model
network = Siren(2, [30, 30, 30, 30, 30], 1) # Function approximator
library = Library1D(poly_order=2, diff_order=3) # Library function
estimator = Threshold(0.1) #Clustering() # Sparse estimator
constraint = LeastSquares() # How to constrain
model = DeepMoD(network, library, estimator,
constraint) # Putting it all in the model
# Running model
sparsity_scheduler = Periodic(
initial_epoch=5000, periodicity=100) # Defining when to apply sparsity
optimizer = torch.optim.Adam(model.parameters(),
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
v = 0.1
A = 1.0
# Making grid
x = np.linspace(-3, 4, 100)
t = np.linspace(0.5, 5.0, 50)
x_grid, t_grid = np.meshgrid(x, t, indexing='ij')
# Making data
dataset = Dataset(BurgersDelta, v=v, A=A)
X_train, y_train = dataset.create_dataset(x_grid.reshape(-1, 1),
t_grid.reshape(-1, 1),
n_samples=2000,
noise=0.1,
random=True)
# Running deepmod
estimator = LassoLarsIC(fit_intercept=False)
config = {
'n_in': 2,
'hidden_dims': [30, 30, 30, 30, 30],
'n_out': 1,
'library_function': library_1D_in,
'library_args': {
'poly_order': 2,
'diff_order': 2
},
'sparsity_estimator': estimator
for run_idx in np.arange(n_runs):
network = NN(2, [30, 30, 30, 30, 30], 1)
library = Library1D(poly_order=2, diff_order=3) # Library function
estimator = Threshold(0.1) # Sparse estimator
constraint = LeastSquares() # How to constrain
model = DeepMoD(network, library, estimator,
constraint).to(device) # Putting it all in the model
sparsity_scheduler = Periodic(periodicity=25, initial_epoch=10000)
optimizer = torch.optim.Adam(model.parameters(),
betas=(0.9, 0.999),
amsgrad=True) # Defining optimizer
X, y = dataset.create_dataset(x_grid.reshape(-1, 1),
t_grid.reshape(-1, 1),
n_samples=1000,
noise=0.4,
random=True,
normalize=False)
X, y = X.to(device), y.to(device)
train(model,
X,
y,
optimizer,
sparsity_scheduler,
log_dir=f'data_high_noise/baseline_run_{run_idx}/',
write_iterations=25,
max_iterations=5000,
delta=0.00,
patience=100) # Running
# Making grid
x = np.linspace(-3, 4, 100)
t = np.linspace(0.5, 5.0, 50)
x_grid, t_grid = np.meshgrid(x, t, indexing='ij')
dataset = Dataset(BurgersDelta, v=v, A=A)
noise_range = np.arange(0.0, 1.61, 0.20) #np.arange(0.0, 0.51, 0.05)
n_runs = 5
for noise_level in noise_range:
for run in np.arange(n_runs):
X_train, y_train = dataset.create_dataset(
x_grid.reshape(-1, 1),
t_grid.reshape(-1, 1),
n_samples=1000,
noise=noise_level,
random=True,
return_idx=False,
random_state=run
) # use the same dataset for every run; only diff is in the network
estimator = Clustering(estimator=LassoLarsIC(fit_intercept=False))
config = {
'n_in': 2,
'hidden_dims': [30, 30, 30, 30, 30],
'n_out': 1,
'library_function': library_1D_in,
'library_args': {
'poly_order': 2,
'diff_order': 3
},
'sparsity_estimator': estimator
x_grid, t_grid = np.meshgrid(x, t)
dataset = Dataset(BurgersDelta, v=0.1, A=1.0)
dataset = Dataset(BurgersCos, v=0.1, a=0.1, b=0.1, k=2)
dataset = Dataset(BurgersSawtooth, v=0.1)
#dataset = Dataset(KdVSoliton, c=5.0, a = -1.0, b=1)
dataset.generate_solution(x_grid, t_grid).shape
dataset.parameters
dataset.time_deriv(x_grid, t_grid).shape
theta = dataset.library(x_grid.reshape(-1, 1),
t_grid.reshape(-1, 1),
poly_order=2,
deriv_order=2)
dt = dataset.time_deriv(x_grid.reshape(-1, 1), t_grid.reshape(-1, 1))
theta.shape
np.linalg.lstsq(theta, dt, rcond=None)[0]
X_train, y_train = dataset.create_dataset(x_grid,
t_grid,
n_samples=0,
noise=0.05)
y_train.shape
from phimal_utilities.analysis import load_tensorboard
