lh solved (#55)

pina/span.py

@@ -3,6 +3,7 @@ import torch
 
 from .location import Location
 from .label_tensor import LabelTensor
+from .utils import torch_lhs
 
 
 class Span(Location):
@@ -41,10 +42,7 @@ class Span(Location):
         elif mode == 'grid':
             pts = torch.linspace(0, 1, n).reshape(-1, 1)
         elif mode == 'lh' or mode == 'latin':
-            from scipy.stats import qmc
-            sampler = qmc.LatinHypercube(d=dim)
-            pts = sampler.random(n)
-            pts = torch.from_numpy(pts)
+            pts = torch_lhs(n, dim)
 
         pts *= bounds[:, 1] - bounds[:, 0]
         pts += bounds[:, 0]
@@ -83,7 +81,7 @@ class Span(Location):
             return result
 
         def _Nd_sampler(n, mode, variables):
-            """ Sample ll the variables together """
+            """ Sample all the variables together """
             pairs = [(k, v) for k, v in self.range_.items() if k in variables]
             keys, values = map(list, zip(*pairs))
             bounds = torch.tensor(values)
@@ -107,7 +105,7 @@ class Span(Location):
 
         if mode in ['grid', 'chebyshev']:
             return _1d_sampler(n, mode, variables)
-        elif mode in ['random', 'lhs']:
+        elif mode in ['random', 'lh', 'latin']:
             return _Nd_sampler(n, mode, variables)
         else:
             raise ValueError(f'mode={mode} is not valid.')

pina/utils.py

@@ -5,6 +5,8 @@ from torch.utils.data import DataLoader, default_collate, ConcatDataset
 
 from .label_tensor import LabelTensor
 
+import torch
+
 
 def number_parameters(model, aggregate=True, only_trainable=True):  # TODO: check
     """
@@ -49,6 +51,40 @@ def merge_two_tensors(tensor1, tensor2):
     return tensor1.append(tensor2)
 
 
+def torch_lhs(n, dim):
+    """Latin Hypercube Sampling torch routine.
+    Sampling in range $[0, 1)^d$.
+
+    :param int n: number of samples
+    :param int dim: dimensions of latin hypercube
+    :return: samples
+    :rtype: torch.tensor
+    """
+
+    if not isinstance(n, int):
+        raise TypeError('number of point n must be int')
+
+    if not isinstance(dim, int):
+        raise TypeError('dim must be int')
+
+    if dim < 1:
+        raise ValueError('dim must be greater than one')
+
+    samples = torch.rand(size=(n, dim))
+
+    perms = torch.tile(torch.arange(1, n + 1), (dim, 1))
+
+    for row in range(dim):
+        idx_perm = torch.randperm(perms.shape[-1])
+        perms[row, :] = perms[row, idx_perm]
+
+    perms = perms.T
+
+    samples = (perms - samples) / n
+
+    return samples
+
+
 class PinaDataset():
 
     def __init__(self, pinn) -> None:

tests/test_pinn.py

@@ -63,6 +63,12 @@ def test_span_pts():
     pinn.span_pts(n, 'random', locations=['D'])
     assert pinn.input_pts['D'].shape[0] == n
 
+    pinn.span_pts(n, 'latin', locations=['D'])
+    assert pinn.input_pts['D'].shape[0] == n
+
+    pinn.span_pts(n, 'lh', locations=['D'])
+    assert pinn.input_pts['D'].shape[0] == n
+
 
 def test_train():
     pinn = PINN(problem, model)