Dilation for convolutional layers (#766)

* add dilation parameter to Conv1d layer

* space here too

* add conv1d dilation test

* add dilation parameter for Conv2d layer

* conv2d dilation test
This commit is contained in:
Piotr Rybiec 2024-03-04 15:43:00 +01:00 committed by GitHub
parent bc06cb9ff6
commit 6a665ea6ed
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 24 additions and 4 deletions

View File

@ -23,6 +23,7 @@ class Conv1d(Module):
Default: 1.
padding (int, optional): How many positions to 0-pad the input with.
Default: 0.
dilation (int, optional): The dilation of the convolution.
bias (bool, optional): If ``True`` add a learnable bias to the output.
Default: ``True``
"""
@ -34,6 +35,7 @@ class Conv1d(Module):
kernel_size: int,
stride: int = 1,
padding: int = 0,
dilation: int = 1,
bias: bool = True,
):
super().__init__()
@ -48,17 +50,19 @@ class Conv1d(Module):
self.bias = mx.zeros((out_channels,))
self.padding = padding
self.dilation = dilation
self.stride = stride
def _extra_repr(self):
return (
f"{self.weight.shape[-1]}, {self.weight.shape[0]}, "
f"kernel_size={self.weight.shape[1]}, stride={self.stride}, "
f"padding={self.padding}, bias={'bias' in self}"
f"padding={self.padding}, dilation={self.dilation}, "
f"bias={'bias' in self}"
)
def __call__(self, x):
y = mx.conv1d(x, self.weight, self.stride, self.padding)
y = mx.conv1d(x, self.weight, self.stride, self.padding, self.dilation)
if "bias" in self:
y = y + self.bias
return y
@ -81,6 +85,7 @@ class Conv2d(Module):
applying the filter. Default: 1.
padding (int or tuple, optional): How many positions to 0-pad
the input with. Default: 0.
dilation (int or tuple, optional): The dilation of the convolution.
bias (bool, optional): If ``True`` add a learnable bias to the
output. Default: ``True``
"""
@ -92,6 +97,7 @@ class Conv2d(Module):
kernel_size: Union[int, tuple],
stride: Union[int, tuple] = 1,
padding: Union[int, tuple] = 0,
dilation: Union[int, tuple] = 1,
bias: bool = True,
):
super().__init__()
@ -111,16 +117,18 @@ class Conv2d(Module):
self.padding = padding
self.stride = stride
self.dilation = dilation
def _extra_repr(self):
return (
f"{self.weight.shape[-1]}, {self.weight.shape[0]}, "
f"kernel_size={self.weight.shape[1:2]}, stride={self.stride}, "
f"padding={self.padding}, bias={'bias' in self}"
f"padding={self.padding}, dilation={self.dilation}, "
f"bias={'bias' in self}"
)
def __call__(self, x):
y = mx.conv2d(x, self.weight, self.stride, self.padding)
y = mx.conv2d(x, self.weight, self.stride, self.padding, self.dilation)
if "bias" in self:
y = y + self.bias
return y

View File

@ -586,6 +586,13 @@ class TestLayers(mlx_tests.MLXTestCase):
self.assertEqual(y.shape, (N, (L - ks + 1) // 2, C_out))
self.assertTrue("bias" in c.parameters())
dil = 2
c = nn.Conv1d(
in_channels=C_in, out_channels=C_out, kernel_size=ks, dilation=dil
)
y = c(x)
self.assertEqual(y.shape, (N, L - (ks - 1) * dil, C_out))
c = nn.Conv1d(in_channels=C_in, out_channels=C_out, kernel_size=ks, bias=False)
self.assertTrue("bias" not in c.parameters())
@ -632,6 +639,11 @@ class TestLayers(mlx_tests.MLXTestCase):
self.assertEqual(y.shape, (4, 3, 3, 8))
self.assertLess(mx.abs(y - c.weight.sum((1, 2, 3))).max(), 1e-4)
c = nn.Conv2d(3, 8, 3, dilation=2)
y = c(x)
self.assertEqual(y.shape, (4, 4, 4, 8))
self.assertLess(mx.abs(y - c.weight.sum((1, 2, 3))).max(), 1e-4)
def test_sequential(self):
x = mx.ones((10, 2))
m = nn.Sequential(nn.Linear(2, 10), nn.ReLU(), nn.Linear(10, 1))