Merge pull request #150 from NeuroDiffGym/v0.5.0

fix bug in H1 and H1-semi loss

Author: shuheng-liu

neurodiffeq/losses.py

@@ -16,13 +16,13 @@ def _infinity_norm(residual, funcs, coords):
 
 def _h1_norm(residual, funcs, coords):
     g = grad(residual, *coords)
-    rg = torch.cat([residual, *g])
+    rg = torch.cat([residual, *g], dim=1)
     return (rg ** 2).mean()
 
 
 def _h1_semi_norm(residual, funcs, coords):
     g = grad(residual, *coords)
-    g = torch.cat(g)
+    g = torch.cat(g, dim=1)
     return (g ** 2).mean()
 
 

neurodiffeq/solvers.py

@@ -57,8 +57,19 @@ class BaseSolver(ABC, PretrainedSolver):
         The optimizer to be used for training.
     :type optimizer: `torch.nn.optim.Optimizer`, optional
     :param criterion:
-        A function that maps a PDE residual vector (torch tensor with shape (-1, 1)) to a scalar loss.
-    :type criterion: callable, optional
+        The loss function used for training.
+
+        - If a str, must be present in the keys of `neurodiffeq.losses._losses`.
+        - If a `torch.nn.modules.loss._Loss` instance, just pass the instance.
+        - If any other callable, it must map
+          A) a residual tensor (shape `(n_points, n_equations)`),
+          B) a function values tuple (length `n_funcs`, each element a tensor of shape `(n_points, 1)`), and
+          C) a coordinate values tuple (length `n_coords`, each element a tensor of shape `(n_coords, 1)`
+          to a tensor of empty shape (i.e. a scalar). The returned tensor must be connected to the computational graph,
+          so that backpropagation can be performed.
+
+    :type criterion:
+        str or `torch.nn.moduesl.loss._Loss` or callable
     :param n_batches_train:
         Number of batches to train in every epoch, where batch-size equals ``train_generator.size``.
         Defaults to 1.
@@ -687,8 +698,19 @@ class SolverSpherical(BaseSolver):
         Defaults to a ``torch.optim.Adam`` instance that trains on all parameters of ``nets``.
     :type optimizer: ``torch.nn.optim.Optimizer``, optional
     :param criterion:
-        Function that maps a PDE residual tensor (of shape (-1, 1)) to a scalar loss.
-    :type criterion: callable, optional
+        The loss function used for training.
+
+        - If a str, must be present in the keys of `neurodiffeq.losses._losses`.
+        - If a `torch.nn.modules.loss._Loss` instance, just pass the instance.
+        - If any other callable, it must map
+          A) a residual tensor (shape `(n_points, n_equations)`),
+          B) a function values tuple (length `n_funcs`, each element a tensor of shape `(n_points, 1)`), and
+          C) a coordinate values tuple (length `n_coords`, each element a tensor of shape `(n_coords, 1)`
+          to a tensor of empty shape (i.e. a scalar). The returned tensor must be connected to the computational graph,
+          so that backpropagation can be performed.
+
+    :type criterion:
+        str or `torch.nn.moduesl.loss._Loss` or callable
     :param n_batches_train:
         Number of batches to train in every epoch, where batch-size equals ``train_generator.size``.
         Defaults to 1.
@@ -935,8 +957,19 @@ class Solver1D(BaseSolver):
         Defaults to a ``torch.optim.Adam`` instance that trains on all parameters of ``nets``.
     :type optimizer: ``torch.nn.optim.Optimizer``, optional
     :param criterion:
-        Function that maps a ODE residual tensor (of shape (-1, 1)) to a scalar loss.
-    :type criterion: callable, optional
+        The loss function used for training.
+
+        - If a str, must be present in the keys of `neurodiffeq.losses._losses`.
+        - If a `torch.nn.modules.loss._Loss` instance, just pass the instance.
+        - If any other callable, it must map
+          A) a residual tensor (shape `(n_points, n_equations)`),
+          B) a function values tuple (length `n_funcs`, each element a tensor of shape `(n_points, 1)`), and
+          C) a coordinate values tuple (length `n_coords`, each element a tensor of shape `(n_coords, 1)`
+          to a tensor of empty shape (i.e. a scalar). The returned tensor must be connected to the computational graph,
+          so that backpropagation can be performed.
+
+    :type criterion:
+        str or `torch.nn.moduesl.loss._Loss` or callable
     :param n_batches_train:
         Number of batches to train in every epoch, where batch-size equals ``train_generator.size``.
         Defaults to 1.
@@ -1108,8 +1141,19 @@ class BundleSolver1D(BaseSolver):
         Defaults to a ``torch.optim.Adam`` instance that trains on all parameters of ``nets``.
     :type optimizer: ``torch.nn.optim.Optimizer``, optional
     :param criterion:
-        Function that maps a ODE residual tensor (of shape (-1, 1)) to a scalar loss.
-    :type criterion: callable, optional
+        The loss function used for training.
+
+        - If a str, must be present in the keys of `neurodiffeq.losses._losses`.
+        - If a `torch.nn.modules.loss._Loss` instance, just pass the instance.
+        - If any other callable, it must map
+          A) a residual tensor (shape `(n_points, n_equations)`),
+          B) a function values tuple (length `n_funcs`, each element a tensor of shape `(n_points, 1)`), and
+          C) a coordinate values tuple (length `n_coords`, each element a tensor of shape `(n_coords, 1)`
+          to a tensor of empty shape (i.e. a scalar). The returned tensor must be connected to the computational graph,
+          so that backpropagation can be performed.
+
+    :type criterion:
+        str or `torch.nn.moduesl.loss._Loss` or callable
     :param n_batches_train:
         Number of batches to train in every epoch, where batch-size equals ``train_generator.size``.
         Defaults to 1.
@@ -1308,8 +1352,19 @@ class Solver2D(BaseSolver):
         Defaults to a ``torch.optim.Adam`` instance that trains on all parameters of ``nets``.
     :type optimizer: ``torch.nn.optim.Optimizer``, optional
     :param criterion:
-        Function that maps a PDE residual tensor (of shape (-1, 1)) to a scalar loss.
-    :type criterion: callable, optional
+        The loss function used for training.
+
+        - If a str, must be present in the keys of `neurodiffeq.losses._losses`.
+        - If a `torch.nn.modules.loss._Loss` instance, just pass the instance.
+        - If any other callable, it must map
+          A) a residual tensor (shape `(n_points, n_equations)`),
+          B) a function values tuple (length `n_funcs`, each element a tensor of shape `(n_points, 1)`), and
+          C) a coordinate values tuple (length `n_coords`, each element a tensor of shape `(n_coords, 1)`
+          to a tensor of empty shape (i.e. a scalar). The returned tensor must be connected to the computational graph,
+          so that backpropagation can be performed.
+
+    :type criterion:
+        str or `torch.nn.moduesl.loss._Loss` or callable
     :param n_batches_train:
         Number of batches to train in every epoch, where batch-size equals ``train_generator.size``.
         Defaults to 1.

setup.py

@@ -17,14 +17,14 @@ def func(m):
 
 setuptools.setup(
     name="neurodiffeq",
-    version="0.4.0",
+    version="0.5.0",
     author="neurodiffgym",
     author_email="shuheng_liu@g.harvard.edu",
     description="A light-weight & flexible library for solving differential equations using neural networks based on PyTorch. ",
     long_description=long_description,
     long_description_content_type="text/markdown",
     url="https://github.com/NeuroDiffGym/neurodiffeq",
-    download_url="https://github.com/NeuroDiffGym/neurodiffeq/archive/v0.4.0.tar.gz",
+    download_url="https://github.com/NeuroDiffGym/neurodiffeq/archive/v0.5.0.tar.gz",
     keywords=[
         "neural network",
         "deep learning",

tests/test_losses.py

@@ -0,0 +1,37 @@
+import pytest
+import torch
+from neurodiffeq import diff
+from neurodiffeq.losses import _losses as losses
+
+N = 100
+
+
+def pde_system(u, v, w, x, y):
+    return [
+        diff(u, x, order=2) + diff(u, y, order=2),
+        diff(v, x, order=2) + diff(v, y, order=2),
+        diff(w, x, order=2) + diff(w, y, order=2),
+    ]
+
+
+def get_rfx(n_input, n_output, n_equation):
+    coords = [torch.rand((N, 1), requires_grad=True) for _ in range(n_input)]
+    coords_tensor = torch.cat(coords, dim=1)
+    funcs = [torch.sigmoid(torch.sum(coords_tensor, dim=1, keepdim=True)) for _ in range(n_output)]
+    residual = [diff(funcs[0], coords[0]) + funcs[0] for _ in range(n_equation)]
+    residual = torch.cat(residual, dim=1)
+    return residual, funcs, coords
+
+
+@pytest.mark.parametrize(argnames='n_input', argvalues=[1, 3])
+@pytest.mark.parametrize(argnames='n_output', argvalues=[1, 3])
+@pytest.mark.parametrize(argnames='n_equation', argvalues=[1, 3])
+@pytest.mark.parametrize(
+    argnames=('loss_name', 'loss_fn'),
+    argvalues=losses.items(),
+)
+def test_losses(n_input, n_output, n_equation, loss_name, loss_fn):
+    r, f, x = get_rfx(n_input, n_output, n_equation)
+    loss = loss_fn(r, f, x)
+    assert loss.shape == (), f"{loss_name} doesn't output scalar"
+    assert loss.requires_grad, f"{loss_name} doesn't require gradient"