def depthxc_estimates(d1, d2, BASIS=BASIS, THETA=THETA, COV=COV):
n_params = THETA.shape[0]
bounds_x1 = (-8, +8)
bounds_x2 = (-2, +20)
shape = (100, 400)
test_X = grid(bounds_x1, bounds_x2, *shape)
xc_filt = ((test_X[:, 1] >= d1) * (test_X[:, 1] < d2))
D = np.sort(np.unique(test_X[:, 0]))
n_d = len(D)
DEPTH_IDX = 20
predmat = np.zeros((DEPTH_IDX, n_d))
maskmat = np.zeros((DEPTH_IDX, n_d))
for tidx, d in enumerate(D):
idxs = xc_filt * (test_X[:, 0] == d)
x_pts = test_X[idxs, 1] # 1 x 6
y_pts = BASIS[idxs, :] # 6 x 394
# trap integration for each column
int_basis = np.zeros(n_params)
var = np.zeros(DEPTH_IDX)
for pidx in range(n_params):
int_basis[pidx] = np.trapz(y_pts[:, pidx], x_pts)
for depth_idx in range(DEPTH_IDX):
var[depth_idx] = np.linalg.multi_dot(
(int_basis, COV[:, :, depth_idx], int_basis))
inprod = np.dot(int_basis, THETA)
predmat[:, tidx] = inprod / (max(x_pts) - min(x_pts))
sd = np.sqrt(var)
maskmat[:, tidx] = (inprod > 2 * sd) | (inprod < -2 * sd)
predmat_small = predmat
maskmat_small = maskmat
predmat = np.zeros((DEPTH_IDX, n_d * 4))
maskmat = np.zeros((DEPTH_IDX, n_d * 4))
xnew = np.linspace(-8, +8, n_d * 4)
for depth_idx in range(DEPTH_IDX):
predmat[depth_idx, :] = np.interp(xnew, D, predmat_small[depth_idx, :])
maskmat[depth_idx, :] = np.repeat(maskmat_small[depth_idx, :], 4)
mat = predmat.copy()
mat[~maskmat.astype(bool)] = np.nan
return mat
def depthtime_estimates(center, ws=0.5, BASIS=BASIS, THETA=THETA, COV=COV):
ws = 0.5
n_params = THETA.shape[0]
bounds_x1 = (-8, +8)
bounds_x2 = (-2, +20)
shape = (100, 400)
test_X = grid(bounds_x1, bounds_x2, *shape)
xc_filt = ((test_X[:, 0] > center - ws) * (test_X[:, 0] < center + ws))
TAU = np.sort(np.unique(test_X[:, 1]))
n_tau = len(TAU)
DEPTH_IDX = 20
predmat = np.zeros((DEPTH_IDX, n_tau))
maskmat = np.zeros((DEPTH_IDX, n_tau))
for tidx, tau in enumerate(TAU):
idxs = xc_filt * (test_X[:, 1] == tau)
x_pts = test_X[idxs, 0] # 1 x 6
y_pts = BASIS[idxs, :] # 6 x 394
# trap integration for each column
int_basis = np.zeros(n_params)
var = np.zeros(DEPTH_IDX)
for pidx in range(n_params):
int_basis[pidx] = np.trapz(y_pts[:, pidx], x_pts)
for depth_idx in range(DEPTH_IDX):
var[depth_idx] = np.linalg.multi_dot(
(int_basis, COV[:, :, depth_idx], int_basis))
inprod = np.dot(int_basis, THETA)
predmat[:, tidx] = inprod / (max(x_pts) - min(x_pts))
sd = np.sqrt(var)
maskmat[:, tidx] = (inprod > 2 * sd) | (inprod < -2 * sd)
mat = predmat.copy()
mat[~maskmat.astype(bool)] = np.nan
return mat
# lamb_choice = 'AdaptSignalInflate'
#lamb_choice = 'AdaptInflate' # This seems the selected one based on the description in the paper
lamb_choice = 'Custom'
# Loop through all cases
for sub in sub_list:
print(sub)
# Load output from B10 (Argo profile data)
df = pkl.load(
open(f'{data_dir}/HurricaneVariableCovDF_Subset{sub}.pkl', 'rb'))
X = np.array(df[['standard_signed_angle', 'hurricane_dtd']]).copy()
bounds_x1 = (-8, +8)
bounds_x2 = (-2, +20)
train_knots = grid(bounds_x1, bounds_x2, 33, 45)
n_param = train_knots.shape[0] + 3
shape = (100, 400)
test_X = grid(bounds_x1, bounds_x2, *shape)
n_test = test_X.shape[0]
y = variable_diff(df, stage, 0)
ymat = np.zeros((1, DEPTH_IDX, len(y)))
for depth_idx in range(DEPTH_IDX):
ymat[0, depth_idx, :] = variable_diff(df, stage, depth_idx)
# Load results from B32 (lambda values and LOOCV scores)
#LAMB_SMALL, LOOCV_SMALL = pkl.load(open(f'{data_dir}/B32_LOOCV_Data_{sub}.pkl', 'rb'))
# Load results from B35 (lambda values and LOOCV scores)
LAMB_LARGE, LOOCV_LARGE = pkl.load(
BASIS = pkl.load(
open(f'{tps_est_dir}/TPS_Basis_Block_0.5_{filt}.pkl', 'rb'))
THETA = pkl.load(
open(f'{tps_est_dir}/TPS_Theta_Block_0.5_{filt}.pkl', 'rb'))
COV = pkl.load(
open(f'{tps_est_dir}/TPS_CovTheta_Block_0.5_{filt}.pkl', 'rb'))
center = 0 #int(sys.argv[1]) da cambiare a mano in base al cross track angle che si vuole valutare
ws = 0.5
n_params = THETA.shape[0]
bounds_x1 = (-8, +8)
bounds_x2 = (-2, +20)
shape = (100, 400)
test_X = grid(bounds_x1, bounds_x2, *shape)
xc_filt = ((test_X[:, 0] > center - ws) * (test_X[:, 0] < center + ws))
TAU = np.sort(np.unique(test_X[:, 1]))
n_tau = len(TAU)
grid_start = int(grid_lower)
grid_end = int(grid_upper)
grid_stride = int(grid_stride)
DEPTH_IDX = int((grid_end - grid_start) / grid_stride + 1)
predmat = np.zeros((DEPTH_IDX, n_tau))
maskmat = np.zeros((DEPTH_IDX, n_tau))
for tidx, tau in enumerate(TAU):
idxs = xc_filt * (test_X[:, 1] == tau)
def __init__(self, lamb, knots=grid((-8, 8), (-2, 20), 17, 23)):
self.lamb = lamb
self.knots = knots
self.train_X = NULL
self.train_y = NULL
sub_list = (
'all_combined',
'all_hurricanes',
'all_tstd',
)
for sub in sub_list:
print(sub)
df = pkl.load(
open(f'{data_dir}/HurricaneVariableCovDF_Subset{sub}.pkl', 'rb'))
X = np.array(df[['standard_signed_angle', 'hurricane_dtd']]).copy()
bounds_x1 = (-8, +8)
bounds_x2 = (-2, +20)
train_knots = grid(bounds_x1, bounds_x2, 17, 23)
n_param = train_knots.shape[0] + 3
shape = (100, 400)
test_X = grid(bounds_x1, bounds_x2, *shape)
n_test = test_X.shape[0]
PREDS_NOREW = np.zeros((n_test, DEPTH_IDX))
PREDS_DIAG = np.zeros((n_test, DEPTH_IDX))
PREDS_BLOCK = np.zeros((n_test, DEPTH_IDX))
STDEV_DIAG = np.zeros((n_test, DEPTH_IDX))
STDEV_BLOCK = np.zeros((n_test, DEPTH_IDX))
MASK_DIAG = np.zeros((n_test, DEPTH_IDX))
MASK_BLOCK = np.zeros((n_test, DEPTH_IDX))
THETA_BLOCK = np.zeros((n_param, DEPTH_IDX))
BASIS_BLOCK = np.zeros((n_test, n_param)) # Invariant to depth
COV_THETA_BLOCK = np.zeros((n_param, n_param, DEPTH_IDX))
