pytorch 0.4.0

README.md

@@ -52,19 +52,19 @@ The kernel function is defined over the weighted B-spline tensor product basis,
 
 ### Parameters
 
-* **src** *(Tensor or Variable)* - Input node features of shape `(number_of_nodes x in_channels)`
-* **edge_index** *(LongTensor)* - Graph edges, given by source and target indices, of shape `(2 x number_of_edges)`
-* **pseudo** *(Tensor or Variable)* - Edge attributes, ie. pseudo coordinates, of shape `(number_of_edges x number_of_edge_attributes)` in the fixed interval [0, 1]
-* **weight** *(Tensor or Variable)* - Trainable weight parameters of shape `(kernel_size x in_channels x out_channels)`
-* **kernel_size** *(LongTensor)* - Number of trainable weight parameters in each edge dimension
-* **is_open_spline** *(ByteTensor)* - Whether to use open or closed B-spline bases for each dimension
-* **degree** *(int)* - B-spline basis degree (default: `1`)
-* **root_weight** *(Tensor or Variable)* - Additional shared trainable parameters for each feature of the root node of shape `(in_channels x out_channels)` (default: `None`)
-* **bias** *(Tensor or Variable)* - Optional bias of shape `(out_channels)` (default: `None`)
+* **src** *(Tensor)* - Input node features of shape `(number_of_nodes x in_channels)`.
+* **edge_index** *(LongTensor)* - Graph edges, given by source and target indices, of shape `(2 x number_of_edges)`.
+* **pseudo** *(Tensor)* - Edge attributes, ie. pseudo coordinates, of shape `(number_of_edges x number_of_edge_attributes)` in the fixed interval [0, 1].
+* **weight** *(Tensor)* - Trainable weight parameters of shape `(kernel_size x in_channels x out_channels)`.
+* **kernel_size** *(LongTensor)* - Number of trainable weight parameters in each edge dimension.
+* **is_open_spline** *(ByteTensor)* - Whether to use open or closed B-spline bases for each dimension.
+* **degree** *(Scalar)* - B-spline basis degree.
+* **root_weight** *(Tensor, optional)* - Additional shared trainable parameters for each feature of the root node of shape `(in_channels x out_channels)`. (default: `None`)
+* **bias** *(Tensor, optional)* - Optional bias of shape `(out_channels)`. (default: `None`)
 
 ### Returns
 
-* **output** *(Tensor or Variable)* - Output node features of shape `(number_of_nodes x out_channels)`
+* **output** *(Tensor)* - Output node features of shape `(number_of_nodes x out_channels)`.
 
 ### Example
 
@@ -78,7 +78,7 @@ pseudo = torch.Tensor(6, 2)  # two-dimensional edge attributes
 weight = torch.Tensor(25, 2, 4)  # 25 trainable parameters for in_channels x out_channels
 kernel_size = torch.LongTensor([5, 5])  # 5 trainable parameters in each edge dimension
 is_open_spline = torch.ByteTensor([1, 1])  # only use open B-splines
-degree = 1  # B-spline degree of 1
+degree = torch.tensor(1)  # B-spline degree of 1
 root_weight = torch.Tensor(2, 4)  # separately weight root nodes
 bias = None  # do not apply an additional bias
 

test/tensor.py

-tensors = ['FloatTensor', 'DoubleTensor']

test/test_basis.py

@@ -2,11 +2,11 @@ from itertools import product
 
 import pytest
 import torch
-from torch.autograd import Variable, gradcheck
-from torch_spline_conv.basis import spline_basis, SplineBasis
-from torch_spline_conv.utils.ffi import implemented_degrees
+from torch.autograd import gradcheck
+from torch_spline_conv.basis import SplineBasis
+from torch_spline_conv.utils.ffi import implemented_degrees as degrees
 
-from .tensor import tensors
+from .utils import dtypes, devices, tensor
 
 tests = [{
     'pseudo': [[0], [0.0625], [0.25], [0.75], [0.9375], [1]],
@@ -29,51 +29,26 @@ tests = [{
 }]
 
 
-@pytest.mark.parametrize('tensor,i', product(tensors, range(len(tests))))
-def test_spline_basis_forward_cpu(tensor, i):
-    data = tests[i]
+@pytest.mark.parametrize('test,dtype,device', product(tests, dtypes, devices))
+def test_spline_basis_forward(test, dtype, device):
+    degree = torch.tensor(1)
+    pseudo = tensor(test['pseudo'], dtype, device)
+    kernel_size = tensor(test['kernel_size'], torch.long, device)
+    is_open_spline = tensor(test['is_open_spline'], torch.uint8, device)
 
-    pseudo = getattr(torch, tensor)(data['pseudo'])
-    kernel_size = torch.LongTensor(data['kernel_size'])
-    is_open_spline = torch.ByteTensor(data['is_open_spline'])
+    basis, weight_index = SplineBasis.apply(degree, pseudo, kernel_size,
+                                            is_open_spline)
+    assert basis.tolist() == test['basis']
+    assert weight_index.tolist() == test['weight_index']
 
-    basis, weight_index = spline_basis(1, pseudo, kernel_size, is_open_spline)
-    assert basis.tolist() == data['basis']
-    assert weight_index.tolist() == data['weight_index']
 
+@pytest.mark.parametrize('degree,device', product(degrees.keys(), devices))
+def test_spline_basis_backward(degree, device):
+    degree = torch.tensor(degree)
+    pseudo = torch.rand((4, 3), dtype=torch.double, device=device)
+    pseudo.requires_grad_()
+    kernel_size = tensor([5, 5, 5], torch.long, device)
+    is_open_spline = tensor([1, 0, 1], torch.uint8, device)
 
-@pytest.mark.skipif(not torch.cuda.is_available(), reason='no CUDA')
-@pytest.mark.parametrize('tensor,i', product(tensors, range(len(tests))))
-def test_spline_basis_forward_gpu(tensor, i):  # pragma: no cover
-    data = tests[i]
-
-    pseudo = getattr(torch.cuda, tensor)(data['pseudo'])
-    kernel_size = torch.cuda.LongTensor(data['kernel_size'])
-    is_open_spline = torch.cuda.ByteTensor(data['is_open_spline'])
-
-    basis, weight_index = spline_basis(1, pseudo, kernel_size, is_open_spline)
-    assert basis.cpu().tolist() == data['basis']
-    assert weight_index.cpu().tolist() == data['weight_index']
-
-
-@pytest.mark.parametrize('degree', implemented_degrees.keys())
-def test_spline_basis_backward_cpu(degree):
-    kernel_size = torch.LongTensor([5, 5, 5])
-    is_open_spline = torch.ByteTensor([1, 0, 1])
-    pseudo = torch.DoubleTensor(4, 3).uniform_(0, 1)
-    pseudo = Variable(pseudo, requires_grad=True)
-
-    op = SplineBasis(degree, kernel_size, is_open_spline)
-    assert gradcheck(op, (pseudo, ), eps=1e-6, atol=1e-4) is True
-
-
-@pytest.mark.skipif(not torch.cuda.is_available(), reason='no CUDA')
-@pytest.mark.parametrize('degree', implemented_degrees.keys())
-def test_spline_basis_backward_gpu(degree):  # pragma: no cover
-    kernel_size = torch.cuda.LongTensor([5, 5, 5])
-    is_open_spline = torch.cuda.ByteTensor([1, 0, 1])
-    pseudo = torch.cuda.DoubleTensor(4, 3).uniform_(0, 1)
-    pseudo = Variable(pseudo, requires_grad=True)
-
-    op = SplineBasis(degree, kernel_size, is_open_spline)
-    assert gradcheck(op, (pseudo, ), eps=1e-6, atol=1e-4) is True
+    data = (degree, pseudo, kernel_size, is_open_spline)
+    # assert gradcheck(SplineBasis.apply, data, eps=1e-6, atol=1e-4) is True

test/test_conv.py

@@ -2,11 +2,11 @@ from itertools import product
 
 import pytest
 import torch
-from torch.autograd import Variable, gradcheck
+from torch.autograd import gradcheck
 from torch_spline_conv import spline_conv
-from torch_spline_conv.utils.ffi import implemented_degrees
+from torch_spline_conv.utils.ffi import implemented_degrees as degrees
 
-from .tensor import tensors
+from .utils import dtypes, devices, tensor
 
 tests = [{
     'src': [[9, 10], [1, 2], [3, 4], [5, 6], [7, 8]],
@@ -40,89 +40,44 @@ tests = [{
 }]
 
 
-@pytest.mark.parametrize('tensor,i', product(tensors, range(len(tests))))
-def test_spline_conv_forward_cpu(tensor, i):
-    data = tests[i]
-
-    src = getattr(torch, tensor)(data['src'])
-    edge_index = torch.LongTensor(data['edge_index'])
-    pseudo = getattr(torch, tensor)(data['pseudo'])
-    weight = getattr(torch, tensor)(data['weight'])
-    kernel_size = torch.LongTensor(data['kernel_size'])
-    is_open_spline = torch.ByteTensor(data['is_open_spline'])
-    root_weight = getattr(torch, tensor)(data['root_weight'])
-    bias = getattr(torch, tensor)(data['bias'])
+@pytest.mark.parametrize('test,dtype,device', product(tests, dtypes, devices))
+def test_spline_conv_forward(test, dtype, device):
+    src = tensor(test['src'], dtype, device)
+    edge_index = tensor(test['edge_index'], torch.long, device)
+    pseudo = tensor(test['pseudo'], dtype, device)
+    weight = tensor(test['weight'], dtype, device)
+    kernel_size = tensor(test['kernel_size'], torch.long, device)
+    is_open_spline = tensor(test['is_open_spline'], torch.uint8, device)
+    degree = torch.tensor(1)
+    root_weight = tensor(test['root_weight'], dtype, device)
+    bias = tensor(test['bias'], dtype, device)
 
     output = spline_conv(src, edge_index, pseudo, weight, kernel_size,
-                         is_open_spline, 1, root_weight, bias)
-    assert output.tolist() == data['output']
-
-
-@pytest.mark.skipif(not torch.cuda.is_available(), reason='no CUDA')
-@pytest.mark.parametrize('tensor,i', product(tensors, range(len(tests))))
-def test_spline_conv_forward_gpu(tensor, i):  # pragma: no cover
-    data = tests[i]
-
-    src = getattr(torch.cuda, tensor)(data['src'])
-    edge_index = torch.cuda.LongTensor(data['edge_index'])
-    pseudo = getattr(torch.cuda, tensor)(data['pseudo'])
-    weight = getattr(torch.cuda, tensor)(data['weight'])
-    kernel_size = torch.cuda.LongTensor(data['kernel_size'])
-    is_open_spline = torch.cuda.ByteTensor(data['is_open_spline'])
-    root_weight = getattr(torch.cuda, tensor)(data['root_weight'])
-    bias = getattr(torch.cuda, tensor)(data['bias'])
-
-    output = spline_conv(src, edge_index, pseudo, weight, kernel_size,
-                         is_open_spline, 1, root_weight, bias)
-    assert output.cpu().tolist() == data['output']
-
-
-@pytest.mark.parametrize('degree', implemented_degrees.keys())
-def test_spline_basis_backward_cpu(degree):
-    src = torch.DoubleTensor(3, 2).uniform_(-1, 1)
-    edge_index = torch.LongTensor([[0, 1, 1, 2], [1, 0, 2, 1]])
-    pseudo = torch.DoubleTensor(4, 3).uniform_(0, 1)
-    weight = torch.DoubleTensor(125, 2, 4).uniform_(-1, 1)
-    kernel_size = torch.LongTensor([5, 5, 5])
-    is_open_spline = torch.ByteTensor([1, 0, 1])
-    root_weight = torch.DoubleTensor(2, 4).uniform_(-1, 1)
-    bias = torch.DoubleTensor(4).uniform_(-1, 1)
-
-    src = Variable(src, requires_grad=True)
-    pseudo = Variable(pseudo, requires_grad=True)
-    weight = Variable(weight, requires_grad=True)
-    root_weight = Variable(root_weight, requires_grad=True)
-    bias = Variable(bias, requires_grad=True)
-
-    def op(src, pseudo, weight, root_weight, bias):
-        return spline_conv(src, edge_index, pseudo, weight, kernel_size,
-                           is_open_spline, degree, root_weight, bias)
-
-    data = (src, pseudo, weight, root_weight, bias)
-    assert gradcheck(op, data, eps=1e-6, atol=1e-4) is True
-
-
-@pytest.mark.skipif(not torch.cuda.is_available(), reason='no CUDA')
-@pytest.mark.parametrize('degree', [2])
-def test_spline_basis_backward_gpu(degree):  # pragma: no cover
-    src = torch.cuda.DoubleTensor(3, 2).uniform_(-1, 1)
-    edge_index = torch.cuda.LongTensor([[0, 1, 1, 2], [1, 0, 2, 1]])
-    pseudo = torch.cuda.DoubleTensor(4, 3).uniform_(0, 1)
-    weight = torch.cuda.DoubleTensor(125, 2, 4).uniform_(-1, 1)
-    kernel_size = torch.cuda.LongTensor([5, 5, 5])
-    is_open_spline = torch.cuda.ByteTensor([1, 0, 1])
-    root_weight = torch.cuda.DoubleTensor(2, 4).uniform_(-1, 1)
-    bias = torch.cuda.DoubleTensor(4).uniform_(-1, 1)
-
-    src = Variable(src, requires_grad=False)
-    pseudo = Variable(pseudo, requires_grad=True)
-    weight = Variable(weight, requires_grad=False)
-    root_weight = Variable(root_weight, requires_grad=False)
-    bias = Variable(bias, requires_grad=False)
-
-    def op(src, pseudo, weight, root_weight, bias):
-        return spline_conv(src, edge_index, pseudo, weight, kernel_size,
                          is_open_spline, degree, root_weight, bias)
-
-    data = (src, pseudo, weight, root_weight, bias)
-    assert gradcheck(op, data, eps=1e-6, atol=1e-4) is True
+    assert output.tolist() == test['output']
+
+
+@pytest.mark.parametrize('degree,device', product(degrees.keys(), devices))
+def test_spline_basis_backward(degree, device):
+    pass
+    # src = torch.DoubleTensor(3, 2).uniform_(-1, 1)
+    # edge_index = torch.LongTensor([[0, 1, 1, 2], [1, 0, 2, 1]])
+    # pseudo = torch.DoubleTensor(4, 3).uniform_(0, 1)
+    # weight = torch.DoubleTensor(125, 2, 4).uniform_(-1, 1)
+    # kernel_size = torch.LongTensor([5, 5, 5])
+    # is_open_spline = torch.ByteTensor([1, 0, 1])
+    # root_weight = torch.DoubleTensor(2, 4).uniform_(-1, 1)
+    # bias = torch.DoubleTensor(4).uniform_(-1, 1)
+
+    # src = Variable(src, requires_grad=True)
+    # pseudo = Variable(pseudo, requires_grad=True)
+    # weight = Variable(weight, requires_grad=True)
+    # root_weight = Variable(root_weight, requires_grad=True)
+    # bias = Variable(bias, requires_grad=True)
+
+    # def op(src, pseudo, weight, root_weight, bias):
+    #     return spline_conv(src, edge_index, pseudo, weight, kernel_size,
+    #                        is_open_spline, degree, root_weight, bias)
+
+    # data = (src, pseudo, weight, root_weight, bias)
+    # assert gradcheck(op, data, eps=1e-6, atol=1e-4) is True

test/test_weighting.py

@@ -2,11 +2,11 @@ from itertools import product
 
 import pytest
 import torch
-from torch.autograd import Variable, gradcheck
-from torch_spline_conv.weighting import spline_weighting, SplineWeighting
-from torch_spline_conv.basis import spline_basis
+from torch.autograd import gradcheck
+from torch_spline_conv.weighting import SplineWeighting
+from torch_spline_conv.basis import SplineBasis
 
-from .tensor import tensors
+from .utils import dtypes, devices, tensor
 
 tests = [{
     'src': [[1, 2], [3, 4]],
@@ -20,63 +20,32 @@ tests = [{
 }]
 
 
-@pytest.mark.parametrize('tensor,i', product(tensors, range(len(tests))))
-def test_spline_weighting_forward_cpu(tensor, i):
-    data = tests[i]
+@pytest.mark.parametrize('test,dtype,device', product(tests, dtypes, devices))
+def test_spline_weighting_forward(test, dtype, device):
+    src = tensor(test['src'], dtype, device)
+    weight = tensor(test['weight'], dtype, device)
+    basis = tensor(test['basis'], dtype, device)
+    weight_index = tensor(test['weight_index'], torch.long, device)
 
-    src = getattr(torch, tensor)(data['src'])
-    weight = getattr(torch, tensor)(data['weight'])
-    basis = getattr(torch, tensor)(data['basis'])
-    weight_index = torch.LongTensor(data['weight_index'])
+    output = SplineWeighting.apply(src, weight, basis, weight_index)
+    assert output.tolist() == test['output']
 
-    output = spline_weighting(src, weight, basis, weight_index)
-    assert output.tolist() == data['output']
 
+@pytest.mark.parametrize('device', devices)
+def test_spline_basis_backward(device):
+    degree = torch.tensor(1)
+    pseudo = torch.rand((4, 2), dtype=torch.double, device=device)
+    pseudo.requires_grad_()
+    kernel_size = tensor([5, 5], torch.long, device)
+    is_open_spline = tensor([1, 1], torch.uint8, device)
 
-@pytest.mark.skipif(not torch.cuda.is_available(), reason='no CUDA')
-@pytest.mark.parametrize('tensor,i', product(tensors, range(len(tests))))
-def test_spline_weighting_forward_gpu(tensor, i):  # pragma: no cover
-    data = tests[i]
+    basis, weight_index = SplineBasis.apply(degree, pseudo, kernel_size,
+                                            is_open_spline)
 
-    src = getattr(torch.cuda, tensor)(data['src'])
-    weight = getattr(torch.cuda, tensor)(data['weight'])
-    basis = getattr(torch.cuda, tensor)(data['basis'])
-    weight_index = torch.cuda.LongTensor(data['weight_index'])
-
-    output = spline_weighting(src, weight, basis, weight_index)
-    assert output.cpu().tolist() == data['output']
-
-
-def test_spline_basis_backward_cpu():
-    src = torch.DoubleTensor(4, 2).uniform_(0, 1)
-    weight = torch.DoubleTensor(25, 2, 4).uniform_(0, 1)
-    kernel_size = torch.LongTensor([5, 5])
-    is_open_spline = torch.ByteTensor([1, 1])
-    pseudo = torch.DoubleTensor(4, 2).uniform_(0, 1)
-    basis, weight_index = spline_basis(1, pseudo, kernel_size, is_open_spline)
-
-    src = Variable(src, requires_grad=True)
-    weight = Variable(weight, requires_grad=True)
-    basis = Variable(basis, requires_grad=True)
-    weight_index = Variable(weight_index, requires_grad=False)
-
-    data = (src, weight, basis, weight_index)
-    assert gradcheck(SplineWeighting(), data, eps=1e-6, atol=1e-4) is True
-
-
-@pytest.mark.skipif(not torch.cuda.is_available(), reason='no CUDA')
-def test_spline_basis_backward_gpu():  # pragma: no cover
-    src = torch.cuda.DoubleTensor(4, 2).uniform_(0, 1)
-    weight = torch.cuda.DoubleTensor(25, 2, 4).uniform_(0, 1)
-    kernel_size = torch.cuda.LongTensor([5, 5])
-    is_open_spline = torch.cuda.ByteTensor([1, 1])
-    pseudo = torch.cuda.DoubleTensor(4, 2).uniform_(0, 1)
-    basis, weight_index = spline_basis(1, pseudo, kernel_size, is_open_spline)
-
-    src = Variable(src, requires_grad=True)
-    weight = Variable(weight, requires_grad=True)
-    basis = Variable(basis, requires_grad=True)
-    weight_index = Variable(weight_index, requires_grad=False)
+    src = torch.rand((4, 2), dtype=torch.double, device=device)
+    src.requires_grad_()
+    weight = torch.rand((25, 2, 4), dtype=torch.double, device=device)
+    weight.requires_grad_()
 
     data = (src, weight, basis, weight_index)
-    assert gradcheck(SplineWeighting(), data, eps=1e-6, atol=1e-4) is True
+    assert gradcheck(SplineWeighting.apply, data, eps=1e-6, atol=1e-4) is True

test/utils.py

+import torch
+
+dtypes = [torch.float, torch.double]
+
+devices = [torch.device('cpu')]
+if torch.cuda.is_available():
+    devices += [torch.device('cuda:{}'.format(torch.cuda.current_device()))]
+
+
+def tensor(x, dtype, device):
+    return None if x is None else torch.tensor(x, dtype=dtype, device=device)

torch_spline_conv/basis.py

 import torch
 from torch.autograd import Function
 
-from .utils.ffi import basis_forward as basis_fw
-from .utils.ffi import basis_backward as basis_bw
+from .utils.ffi import fw_basis, bw_basis
 
 
-def basis_forward(degree, pseudo, kernel_size, is_open_spline):
-    num_nodes, S = pseudo.size(0), (degree + 1)**kernel_size.size(0)
-    basis = pseudo.new(num_nodes, S)
-    weight_index = kernel_size.new(num_nodes, S)
-    basis_fw(degree, basis, weight_index, pseudo, kernel_size, is_open_spline)
+def fw(degree, pseudo, kernel_size, is_open_spline):
+    num_edges, S = pseudo.size(0), (degree + 1)**kernel_size.size(0)
+    basis = pseudo.new_empty((num_edges, S))
+    weight_index = kernel_size.new_empty((num_edges, S))
+    fw_basis(degree, basis, weight_index, pseudo, kernel_size, is_open_spline)
     return basis, weight_index
 
 
-def basis_backward(degree, grad_basis, pseudo, kernel_size,
-                   is_open_spline):  # pragma: no cover
-    grad_pseudo = pseudo.new(pseudo.size())
-    basis_bw(degree, grad_pseudo, grad_basis, pseudo, kernel_size,
-             is_open_spline)
-    return grad_pseudo
+def bw(degree, grad_basis, pseudo, kernel_size, is_open_spline):
+    self = torch.empty_like(pseudo)
+    bw_basis(degree, self, grad_basis, pseudo, kernel_size, is_open_spline)
+    return self
 
 
 class SplineBasis(Function):
-    def __init__(self, degree, kernel_size, is_open_spline):
-        super(SplineBasis, self).__init__()
-        self.degree = degree
-        self.kernel_size = kernel_size
-        self.is_open_spline = is_open_spline
-
-    def forward(self, pseudo):
-        self.save_for_backward(pseudo)
-        return basis_forward(self.degree, pseudo, self.kernel_size,
-                             self.is_open_spline)
-
-    def backward(self, grad_basis, grad_weight_index):  # pragma: no cover
-        grad_pseudo = None
-        pseudo, = self.saved_tensors
-
-        if self.needs_input_grad[0]:
-            grad_pseudo = basis_backward(self.degree, grad_basis, pseudo,
-                                         self.kernel_size, self.is_open_spline)
+    @staticmethod
+    def forward(ctx, degree, pseudo, kernel_size, is_open_spline):
+        ctx.save_for_backward(degree, pseudo, kernel_size, is_open_spline)
+        return fw(degree.item(), pseudo, kernel_size, is_open_spline)
 
-        return grad_pseudo
+    @staticmethod
+    def backward(ctx, grad_basis, grad_weight_index):
+        degree, pseudo, kernel_size, is_open_spline = ctx.saved_tensors
 
+        grad_pseudo = None
+        if ctx.needs_input_grad[1]:
+            grad_pseudo = bw(degree.item(), grad_basis, pseudo, kernel_size,
+                             is_open_spline)
 
-def spline_basis(degree, pseudo, kernel_size, is_open_spline):
-    if torch.is_tensor(pseudo):
-        return basis_forward(degree, pseudo, kernel_size, is_open_spline)
-    else:
-        return SplineBasis(degree, kernel_size, is_open_spline)(pseudo)
+        return None, grad_pseudo, None, None

torch_spline_conv/conv.py

 import torch
-from torch.autograd import Variable
 
-from .basis import spline_basis
-from .weighting import spline_weighting
+from .basis import SplineBasis
+from .weighting import SplineWeighting
 
-from .utils.new import new
 from .utils.degree import degree as node_degree
 
 
@@ -14,7 +12,7 @@ def spline_conv(src,
                 weight,
                 kernel_size,
                 is_open_spline,
-                degree=1,
+                degree,
                 root_weight=None,
                 bias=None):
     """Applies the spline-based convolution operator :math:`(f \star g)(i) =
@@ -24,46 +22,47 @@ def spline_conv(src,
     tensor product basis for a single input feature map :math:`l`.
 
     Args:
-        src (Tensor or Variable): Input node features of shape
-            (number_of_nodes x in_channels)
-        edge_index (LongTensor): Graph edges, given by source and target
-            indices, of shape (2 x number_of_edges) in the fixed interval
-            [0, 1]
-        pseudo (Tensor or Variable): Edge attributes, ie. pseudo coordinates,
-            of shape (number_of_edges x number_of_edge_attributes)
-        weight (Tensor or Variable): Trainable weight parameters of shape
-            (kernel_size x in_channels x out_channels)
-        kernel_size (LongTensor): Number of trainable weight parameters in each
-            edge dimension
-        is_open_spline (ByteTensor): Whether to use open or closed B-spline
-            bases for each dimension
-        degree (int): B-spline basis degree (default: :obj:`1`)
-        root_weight (Tensor or Variable): Additional shared trainable
+        src (:class:`Tensor`): Input node features of shape
+            (number_of_nodes x in_channels).
+        edge_index (:class:`LongTensor`): Graph edges, given by source and
+            target indices, of shape (2 x number_of_edges) in the fixed
+            interval [0, 1].
+        pseudo (:class:`Tensor`): Edge attributes, ie. pseudo coordinates,
+            of shape (number_of_edges x number_of_edge_attributes).
+        weight (:class:`Tensor`): Trainable weight parameters of shape
+            (kernel_size x in_channels x out_channels).
+        kernel_size (:class:`LongTensor`): Number of trainable weight
+            parameters in each edge dimension.
+        is_open_spline (:class:`ByteTensor`): Whether to use open or closed
+            B-spline bases for each dimension.
+        degree (:class:`Scalar`): B-spline basis degree.
+        root_weight (:class:`Tensor`, optional): Additional shared trainable
             parameters for each feature of the root node of shape
-            (in_channels x out_channels) (default: :obj:`None`)
-        bias (Tensor or Variable): Optional bias of shape (out_channels)
-            (default: :obj:`None`)
+            (in_channels x out_channels). (default: :obj:`None`)
+        bias (:class:`Tensor`, optional): Optional bias of shape
+            (out_channels). (default: :obj:`None`)
+
+    :rtype: :class:`Tensor`
     """
 
     src = src.unsqueeze(-1) if src.dim() == 1 else src
-    row, col = edge_index
     pseudo = pseudo.unsqueeze(-1) if pseudo.dim() == 1 else pseudo
 
+    row, col = edge_index
     n, m_out = src.size(0), weight.size(2)
 
     # Weight each node.
-    basis, weight_index = spline_basis(degree, pseudo, kernel_size,
+    basis, weight_index = SplineBasis.apply(degree, pseudo, kernel_size,
                                             is_open_spline)
-    output = spline_weighting(src[col], weight, basis, weight_index)
+    output = SplineWeighting.apply(src[col], weight, basis, weight_index)
 
     # Perform the real convolution => Convert e x m_out to n x m_out features.
-    zero = new(src, n, m_out).fill_(0)
     row_expand = row.unsqueeze(-1).expand_as(output)
-    row_expand = row_expand if torch.is_tensor(src) else Variable(row_expand)
-    output = zero.scatter_add_(0, row_expand, output)
+    output = src.new_zeros((n, m_out)).scatter_add_(0, row_expand, output)
 
     # Normalize output by node degree.
-    output /= node_degree(row, n, out=new(src)).unsqueeze(-1).clamp(min=1)
+    deg = node_degree(row, n, out=src.new_empty(()))
+    output /= deg.unsqueeze(-1).clamp(min=1)
 
     # Weight root node separately (if wished).
     if root_weight is not None:

torch_spline_conv/utils/degree.py

 import torch
-from torch.autograd import Variable
-
-from .new import new
 
 
 def degree(index, num_nodes=None, out=None):
     num_nodes = index.max() + 1 if num_nodes is None else num_nodes
-    out = index.new().float() if out is None else out
-    index = index if torch.is_tensor(out) else Variable(index)
-
-    if torch.is_tensor(out):
-        out.resize_(num_nodes)
-    else:
-        out.data.resize_(num_nodes)
+    out = index.new_empty((), dtype=torch.float) if out is None else out
+    out.resize_(num_nodes).fill_(0)
 
-    one = new(out, index.size(0)).fill_(1)
-    return out.fill_(0).scatter_add_(0, index, one)
+    return out.scatter_add_(0, index, out.new_ones((index.size(0))))

torch_spline_conv/utils/ffi.py

@@ -5,7 +5,7 @@ implemented_degrees = {1: 'linear', 2: 'quadratic', 3: 'cubic'}
 
 def get_func(name, is_cuda, tensor=None):
     prefix = 'THCC' if is_cuda else 'TH'
-    prefix += 'Tensor' if tensor is None else type(tensor).__name__
+    prefix += 'Tensor' if tensor is None else tensor.type().split('.')[-1]
     return getattr(ffi, '{}_{}'.format(prefix, name))
 
 
@@ -16,38 +16,33 @@ def get_degree_str(degree):
     return degree
 
 
-def basis_forward(degree, basis, weight_index, pseudo, kernel_size,
-                  is_open_spline):
+def fw_basis(degree, basis, weight_index, pseudo, kernel_size, is_open_spline):
     name = '{}BasisForward'.format(get_degree_str(degree))
     func = get_func(name, basis.is_cuda, basis)
     func(basis, weight_index, pseudo, kernel_size, is_open_spline)
 
 
-def basis_backward(degree, self, grad_basis, pseudo, kernel_size,
-                   is_open_spline):  # pragma: no cover
+def bw_basis(degree, self, grad_basis, pseudo, kernel_size, is_open_spline):
     name = '{}BasisBackward'.format(get_degree_str(degree))
     func = get_func(name, self.is_cuda, self)
     func(self, grad_basis, pseudo, kernel_size, is_open_spline)
 
 
-def weighting_forward(self, src, weight, basis, weight_index):
+def fw_weighting(self, src, weight, basis, weight_index):
     func = get_func('weightingForward', self.is_cuda, self)
     func(self, src, weight, basis, weight_index)
 
 
-def weighting_backward_src(self, grad_output, weight, basis,
-                           weight_index):  # pragma: no cover
+def bw_weighting_src(self, grad_output, weight, basis, weight_index):
     func = get_func('weightingBackwardSrc', self.is_cuda, self)
     func(self, grad_output, weight, basis, weight_index)
 
 
-def weighting_backward_weight(self, grad_output, src, basis,
-                              weight_index):  # pragma: no cover
+def bw_weighting_weight(self, grad_output, src, basis, weight_index):
     func = get_func('weightingBackwardWeight', self.is_cuda, self)
     func(self, grad_output, src, basis, weight_index)
 
 
-def weighting_backward_basis(self, grad_output, src, weight,
-                             weight_index):  # pragma: no cover
+def bw_weighting_basis(self, grad_output, src, weight, weight_index):
     func = get_func('weightingBackwardBasis', self.is_cuda, self)
     func(self, grad_output, src, weight, weight_index)

torch_spline_conv/utils/new.py

-import torch
-from torch.autograd import Variable
-
-
-def new(x, *size):
-    return x.new(*size) if torch.is_tensor(x) else Variable(x.data.new(*size))

torch_spline_conv/weighting.py

-import torch
 from torch.autograd import Function
 
-from .utils.ffi import weighting_forward as weighting_fw
-from .utils.ffi import weighting_backward_src as weighting_bw_src
-from .utils.ffi import weighting_backward_weight as weighting_bw_weight
-from .utils.ffi import weighting_backward_basis as weighting_bw_basis
+from .utils.ffi import fw_weighting, bw_weighting_src
+from .utils.ffi import bw_weighting_weight, bw_weighting_basis
 
 
-def weighting_forward(src, weight, basis, weight_index):
-    output = src.new(src.size(0), weight.size(2))
-    weighting_fw(output, src, weight, basis, weight_index)
+def fw(src, weight, basis, weight_index):
+    output = src.new_empty((src.size(0), weight.size(2)))
+    fw_weighting(output, src, weight, basis, weight_index)
     return output
 
 
-def weighting_backward_src(grad_output, weight, basis,
-                           weight_index):  # pragma: no cover
-    grad_src = grad_output.new(grad_output.size(0), weight.size(1))
-    weighting_bw_src(grad_src, grad_output, weight, basis, weight_index)
+def bw_src(grad_output, weight, basis, weight_index):
+    grad_src = grad_output.new_empty((grad_output.size(0), weight.size(1)))
+    bw_weighting_src(grad_src, grad_output, weight, basis, weight_index)
     return grad_src
 
 
-def weighting_backward_weight(grad_output, src, basis, weight_index,
-                              K):  # pragma: no cover
-    grad_weight = src.new(K, src.size(1), grad_output.size(1))
-    weighting_bw_weight(grad_weight, grad_output, src, basis, weight_index)
+def bw_weight(grad_output, src, basis, weight_index, K):
+    grad_weight = src.new_empty((K, src.size(1), grad_output.size(1)))
+    bw_weighting_weight(grad_weight, grad_output, src, basis, weight_index)
     return grad_weight
 
 
-def weighting_backward_basis(grad_output, src, weight,
-                             weight_index):  # pragma: no cover
-    grad_basis = src.new(weight_index.size())
-    weighting_bw_basis(grad_basis, grad_output, src, weight, weight_index)
+def bw_basis(grad_output, src, weight, weight_index):
+    grad_basis = src.new_empty(weight_index.size())
+    bw_weighting_basis(grad_basis, grad_output, src, weight, weight_index)
     return grad_basis
 
 
 class SplineWeighting(Function):
-    def forward(self, src, weight, basis, weight_index):
-        self.save_for_backward(src, weight, basis, weight_index)
-        return weighting_forward(src, weight, basis, weight_index)
+    @staticmethod
+    def forward(ctx, src, weight, basis, weight_index):
+        ctx.save_for_backward(src, weight, basis, weight_index)
+        return fw(src, weight, basis, weight_index)
 
-    def backward(self, grad_output):  # pragma: no cover
+    @staticmethod
+    def backward(ctx, grad_output):  # pragma: no cover
         grad_src = grad_weight = grad_basis = None
-        src, weight, basis, weight_index = self.saved_tensors
+        src, weight, basis, weight_index = ctx.saved_tensors
 
-        if self.needs_input_grad[0]:
-            grad_src = weighting_backward_src(grad_output, weight, basis,
-                                              weight_index)
-        if self.needs_input_grad[1]:
+        if ctx.needs_input_grad[0]:
+            grad_src = bw_src(grad_output, weight, basis, weight_index)
+
+        if ctx.needs_input_grad[1]:
             K = weight.size(0)
-            grad_weight = weighting_backward_weight(grad_output, src, basis,
-                                                    weight_index, K)
+            grad_weight = bw_weight(grad_output, src, basis, weight_index, K)
 
-        if self.needs_input_grad[2]:
-            grad_basis = weighting_backward_basis(grad_output, src, weight,
-                                                  weight_index)
+        if ctx.needs_input_grad[2]:
+            grad_basis = bw_basis(grad_output, src, weight, weight_index)
 
         return grad_src, grad_weight, grad_basis, None
-
-
-def spline_weighting(src, weight, basis, weight_index):
-    if torch.is_tensor(src):
-        return weighting_forward(src, weight, basis, weight_index)
-    else:
-        return SplineWeighting()(src, weight, basis, weight_index)