from torch.nn.functional import conv2d, pad
from escnn.nn import FieldType
from escnn.nn import GeometricTensor
import escnn.nn
from escnn.group import Representation, Group
from escnn.kernels import KernelBasis
from escnn.gspaces import GSpace2D
from .rd_convolution import _RdConv
from typing import Callable, Union, List
import torch
import numpy as np
import math
__all__ = ["R2Conv"]
[docs]class R2Conv(_RdConv):
def __init__(self,
in_type: FieldType,
out_type: FieldType,
kernel_size: int,
padding: int = 0,
stride: int = 1,
dilation: int = 1,
padding_mode: str = 'zeros',
groups: int = 1,
bias: bool = True,
sigma: Union[List[float], float] = None,
frequencies_cutoff: Union[float, Callable[[float], int]] = None,
rings: List[float] = None,
maximum_offset: int = None,
recompute: bool = False,
basis_filter: Callable[[dict], bool] = None,
initialize: bool = True,
):
r"""
G-steerable planar convolution mapping between the input and output :class:`~escnn.nn.FieldType` s specified by
the parameters ``in_type`` and ``out_type``.
This operation is equivariant under the action of :math:`\R^2\rtimes G` where :math:`G` is the
:attr:`escnn.nn.FieldType.fibergroup` of ``in_type`` and ``out_type``.
Specifically, let :math:`\rho_\text{in}: G \to \GL{\R^{c_\text{in}}}` and
:math:`\rho_\text{out}: G \to \GL{\R^{c_\text{out}}}` be the representations specified by the input and output
field types.
Then :class:`~escnn.nn.R2Conv` guarantees an equivariant mapping
.. math::
\kappa \star [\mathcal{T}^\text{in}_{g,u} . f] = \mathcal{T}^\text{out}_{g,u} . [\kappa \star f] \qquad\qquad \forall g \in G, u \in \R^2
where the transformation of the input and output fields are given by
.. math::
[\mathcal{T}^\text{in}_{g,u} . f](x) &= \rho_\text{in}(g)f(g^{-1} (x - u)) \\
[\mathcal{T}^\text{out}_{g,u} . f](x) &= \rho_\text{out}(g)f(g^{-1} (x - u)) \\
The equivariance of G-steerable convolutions is guaranteed by restricting the space of convolution kernels to an
equivariant subspace.
As proven in `3D Steerable CNNs <https://arxiv.org/abs/1807.02547>`_, this parametrizes the *most general
equivariant convolutional map* between the input and output fields.
For feature fields on :math:`\R^2` (e.g. images), the complete G-steerable kernel spaces for :math:`G \leq \O2`
is derived in `General E(2)-Equivariant Steerable CNNs <https://arxiv.org/abs/1911.08251>`_.
.. warning ::
This class implements a *discretized* convolution operator over a discrete grid.
This means that equivariance to continuous symmetries is *not* perfect.
In practice, by using sufficiently band-limited filters, the equivariance error introduced by the
discretization of the filters and the features is contained, but some design choices may have a negative
effect on the overall equivariance of the architecture.
We provide some :doc:`practical notes <conv_notes>` on using this discretized
convolution module.
During training, in each forward pass the module expands the basis of G-steerable kernels with learned weights
before calling :func:`torch.nn.functional.conv2d`.
When :meth:`~torch.nn.Module.eval()` is called, the filter is built with the current trained weights and stored
for future reuse such that no overhead of expanding the kernel remains.
.. warning ::
When :meth:`~torch.nn.Module.train()` is called, the attributes :attr:`~escnn.nn.R2Conv.filter` and
:attr:`~escnn.nn.R2Conv.expanded_bias` are discarded to avoid situations of mismatch with the
learnable expansion coefficients.
See also :meth:`escnn.nn.R2Conv.train`.
This behaviour can cause problems when storing the :meth:`~torch.nn.Module.state_dict` of a model while in
a mode and lately loading it in a model with a different mode, as the attributes of the class change.
To avoid this issue, we recommend converting the model to eval mode before storing or loading the state
dictionary.
The learnable expansion coefficients of the this module can be initialized with the methods in
:mod:`escnn.nn.init`.
By default, the weights are initialized in the constructors using :func:`~escnn.nn.init.generalized_he_init`.
.. warning ::
This initialization procedure can be extremely slow for wide layers.
In case initializing the model is not required (e.g. before loading the state dict of a pre-trained model)
or another initialization method is preferred (e.g. :func:`~escnn.nn.init.deltaorthonormal_init`), the
parameter ``initialize`` can be set to ``False`` to avoid unnecessary overhead.
See also `this issue <https://github.com/QUVA-Lab/escnn/issues/54>`_
The parameters ``basisexpansion``, ``sigma``, ``frequencies_cutoff``, ``rings`` and ``maximum_offset`` are
optional parameters used to control how the basis for the filters is built, how it is sampled on the filter
grid and how it is expanded to build the filter. We suggest to keep these default values.
.. warning ::
Even if the input tensor has a `coords` attribute, the output of this module will not have one.
Args:
in_type (FieldType): the type of the input field, specifying its transformation law
out_type (FieldType): the type of the output field, specifying its transformation law
kernel_size (int): the size of the (square) filter
padding(int, optional): implicit zero paddings on both sides of the input. Default: ``0``
stride(int, optional): the stride of the kernel. Default: ``1``
dilation(int, optional): the spacing between kernel elements. Default: ``1``
padding_mode(str, optional): ``zeros``, ``reflect``, ``replicate`` or ``circular``. Default: ``zeros``
groups (int, optional): number of blocked connections from input channels to output channels.
It allows depthwise convolution. When used, the input and output types need to be
divisible in ``groups`` groups, all equal to each other.
Default: ``1``.
bias (bool, optional): Whether to add a bias to the output (only to fields which contain a
trivial irrep) or not. Default ``True``
sigma (list or float, optional): width of each ring where the bases are sampled. If only one scalar
is passed, it is used for all rings.
frequencies_cutoff (callable or float, optional): function mapping the radii of the basis elements to the
maximum frequency accepted. If a float values is passed, the maximum frequency is equal to the
radius times this factor. By default (``None``), a more complex policy is used.
rings (list, optional): radii of the rings where to sample the bases
maximum_offset (int, optional): number of additional (aliased) frequencies in the intertwiners for finite
groups. By default (``None``), all additional frequencies allowed by the frequencies cut-off
are used.
recompute (bool, optional): if ``True``, recomputes a new basis for the equivariant kernels.
By Default (``False``), it caches the basis built or reuse a cached one, if it is found.
basis_filter (callable, optional): function which takes as input a descriptor of a basis element
(as a dictionary) and returns a boolean value: whether to preserve (``True``) or discard (``False``)
the basis element. By default (``None``), no filtering is applied.
initialize (bool, optional): initialize the weights of the model. Default: ``True``
Attributes:
~.weights (torch.Tensor): the learnable parameters which are used to expand the kernel
~.filter (torch.Tensor): the convolutional kernel obtained by expanding the parameters
in :attr:`~escnn.nn.R2Conv.weights`
~.bias (torch.Tensor): the learnable parameters which are used to expand the bias, if ``bias=True``
~.expanded_bias (torch.Tensor): the equivariant bias which is summed to the output, obtained by expanding
the parameters in :attr:`~escnn.nn.R2Conv.bias`
"""
assert isinstance(in_type.gspace, GSpace2D)
assert isinstance(out_type.gspace, GSpace2D)
basis_filter, self._rings, self._sigma, self._maximum_frequency = compute_basis_params(
kernel_size, frequencies_cutoff, rings, sigma, dilation, basis_filter
)
super(R2Conv, self).__init__(
in_type,
out_type,
2,
kernel_size,
padding,
stride,
dilation,
padding_mode,
groups,
bias,
basis_filter,
recompute,
)
if initialize:
# by default, the weights are initialized with a generalized form of He's weight initialization
escnn.nn.init.generalized_he_init(self.weights.data, self.basisexpansion)
def _build_kernel_basis(self, in_repr: Representation, out_repr: Representation) -> KernelBasis:
return self.space.build_kernel_basis(in_repr, out_repr, self._sigma, self._rings,
maximum_frequency=self._maximum_frequency
)
[docs] def forward(self, input: GeometricTensor):
r"""
Convolve the input with the expanded filter and bias.
Args:
input (GeometricTensor): input feature field transforming according to ``in_type``
Returns:
output feature field transforming according to ``out_type``
"""
assert input.type == self.in_type
if not self.training:
_filter = self.filter
_bias = self.expanded_bias
else:
# retrieve the filter and the bias
_filter, _bias = self.expand_parameters()
# use it for convolution and return the result
if self.padding_mode == 'zeros':
output = conv2d(input.tensor, _filter,
stride=self.stride,
padding=self.padding,
dilation=self.dilation,
groups=self.groups,
bias=_bias)
else:
output = conv2d(pad(input.tensor, self._reversed_padding_repeated_twice, self.padding_mode),
_filter,
stride=self.stride,
dilation=self.dilation,
groups=self.groups,
bias=_bias)
return GeometricTensor(output, self.out_type, coords=None)
def check_equivariance(self, atol: float = 0.1, rtol: float = 0.1, assertion: bool = True, verbose: bool = True, device: str = 'cpu'):
# np.set_printoptions(precision=5, threshold=30 *self.in_type.size**2, suppress=False, linewidth=30 *self.in_type.size**2)
feature_map_size = 33
last_downsampling = 5
first_downsampling = 5
initial_size = (feature_map_size * last_downsampling - 1 + self.kernel_size - self.padding) * first_downsampling
c = self.in_type.size
import matplotlib.image as mpimg
from skimage.measure import block_reduce
from skimage.transform import resize
# x = mpimg.imread('../group/testimage.jpeg').transpose((2, 0, 1))[np.newaxis, 0:c, :, :]
import scipy
x = scipy.datasets.face().transpose((2, 0, 1))[np.newaxis, 0:c, :, :]
x = resize(
x,
(x.shape[0], x.shape[1], initial_size, initial_size),
anti_aliasing=True
)
assert x.shape[0] == 1, x.shape
x = x / 255.0 - 0.5
if x.shape[1] < c:
to_stack = [x for i in range(c // x.shape[1])]
if c % x.shape[1] > 0:
to_stack += [x[:, :(c % x.shape[1]), ...]]
x = np.concatenate(to_stack, axis=1)
x = torch.FloatTensor(x)
x = self.in_type(x)
def shrink(t: GeometricTensor, s) -> GeometricTensor:
# return GeometricTensor(torch.FloatTensor(block_reduce(t.tensor.detach().numpy(), s, func=np.mean)), t.type)
return t.type(torch.nn.functional.avg_pool2d(t.tensor, kernel_size=(s, s), stride=s, padding=0))
with torch.no_grad():
self.to(device)
gx = self.in_type(torch.cat([x.transform(el).tensor for el in self.space.testing_elements], dim=0))
gx = gx.to(device)
gx = shrink(gx, first_downsampling)
assert gx.shape[-2:] == (initial_size // first_downsampling, initial_size // first_downsampling), (gx.shape, initial_size//first_downsampling)
outs_2 = self(gx)
outs_2 = shrink(outs_2, last_downsampling)
outs_2 = outs_2.tensor.detach().cpu().numpy()
assert outs_2.shape[-2:] == (feature_map_size, feature_map_size), (gx.shape, feature_map_size)
out_1 = self(shrink(x.to(device), first_downsampling)).to('cpu')
errors = []
for i, el in enumerate(self.space.testing_elements):
# out1 = self(shrink(x, (1, 1, 5, 5))).transform(el).tensor.detach().numpy()
# out2 = self(shrink(x.transform(el), (1, 1, 5, 5))).tensor.detach().numpy()
out1 = shrink(out_1.transform(el), last_downsampling).tensor.detach().numpy()
out2 = outs_2[i:i+1]
b, c, h, w = out2.shape
center_mask = np.stack(np.meshgrid(*[np.arange(0, _w) - _w//2 for _w in [h, w]]), axis=0)
assert center_mask.shape == (2, h, w), (center_mask.shape, h, w)
center_mask = center_mask[0, :, :] ** 2 + center_mask[1, :, :] ** 2 < (h / 4) ** 2
out1 = out1[..., center_mask]
out2 = out2[..., center_mask]
out1 = out1.reshape(-1)
out2 = out2.reshape(-1)
errs = np.abs(out1 - out2)
esum = np.maximum(np.abs(out1), np.abs(out2))
esum[esum == 0.0] = 1
relerr = errs / esum
if verbose:
print(el, relerr.max(), relerr.mean(), relerr.var(), errs.max(), errs.mean(), errs.var())
tol = rtol * esum + atol
if np.any(errs > tol) and verbose:
print(out1[errs > tol])
print(out2[errs > tol])
print(tol[errs > tol])
if assertion:
assert np.all(
errs < tol), 'The error found during equivariance check with element "{}" is too high: max = {}, mean = {} var ={}'.format(
el, errs.max(), errs.mean(), errs.var())
errors.append((el, errs.mean()))
return errors
# init.deltaorthonormal_init(self.weights.data, self.basisexpansion)
# filter = self.basisexpansion()
# center = self.s // 2
# filter = filter[..., center, center]
# assert torch.allclose(torch.eye(filter.shape[1]), filter.t() @ filter, atol=3e-7)
[docs] def export(self):
r"""
Export this module to a normal PyTorch :class:`torch.nn.Conv2d` module and set to "eval" mode.
"""
# set to eval mode so the filter and the bias are updated with the current
# values of the weights
self.eval()
_filter = self.filter
_bias = self.expanded_bias
# build the PyTorch Conv2d module
has_bias = self.bias is not None
conv = torch.nn.Conv2d(self.in_type.size,
self.out_type.size,
self.kernel_size,
padding=self.padding,
stride=self.stride,
dilation=self.dilation,
groups=self.groups,
bias=has_bias)
# set the filter and the bias
conv.weight.data = _filter.data
if has_bias:
conv.bias.data = _bias.data
return conv
def bandlimiting_filter(frequency_cutoff: Union[float, Callable[[float], float]]) -> Callable[[dict], bool]:
r"""
Returns a method which takes as input the attributes (as a dictionary) of a basis element and returns a boolean
value: whether to preserve that element (true) or not (false)
If the parameter ``frequency_cutoff`` is a scalar value, the maximum frequency allowed at a certain radius is
proportional to the radius itself. in thi case, the parameter ``frequency_cutoff`` is the factor controlling this
proportionality relation.
If the parameter ``frequency_cutoff`` is a callable, it needs to take as input a radius (a scalar value) and return
the maximum frequency which can be sampled at that radius.
args:
frequency_cutoff (float): factor controlling the bandlimiting
returns:
a function which checks the attributes of individual basis elements and chooses whether to discard them or not
"""
if isinstance(frequency_cutoff, float):
frequency_cutoff = lambda r, fco=frequency_cutoff: r * frequency_cutoff
def bl_filter(attributes: dict) -> bool:
return math.fabs(attributes["irrep:frequency"]) <= frequency_cutoff(attributes["radius"])
return bl_filter
def compute_basis_params(
kernel_size: int,
frequencies_cutoff: Union[float, Callable[[float], float]] = None,
rings: List[float] = None,
sigma: List[float] = None,
dilation: int = 1,
custom_basis_filter: Callable[[dict], bool] = None,
):
width = dilation * (kernel_size - 1) / 2
max_radius = width * np.sqrt(2)
# by default, the number of rings equals half of the filter size
if rings is None:
n_rings = math.ceil(kernel_size / 2)
rings = torch.linspace(0, (kernel_size - 1) // 2, n_rings) * dilation
rings = rings.tolist()
assert all([max_radius >= r >= 0 for r in rings])
if sigma is None:
sigma = [0.6] * (len(rings) - 1) + [0.4]
for i, r in enumerate(rings):
if r == 0.:
sigma[i] = 0.005
elif isinstance(sigma, float):
sigma = [sigma] * len(rings)
if frequencies_cutoff is None:
frequencies_cutoff = -1.
if isinstance(frequencies_cutoff, float):
if frequencies_cutoff == -3:
frequencies_cutoff = _manual_fco3(kernel_size // 2)
elif frequencies_cutoff == -2:
frequencies_cutoff = _manual_fco2(kernel_size // 2)
elif frequencies_cutoff == -1:
frequencies_cutoff = _manual_fco1(kernel_size // 2)
else:
frequencies_cutoff = lambda r, fco=frequencies_cutoff: fco * r
# check if the object is a callable function
assert callable(frequencies_cutoff)
maximum_frequency = int(max(frequencies_cutoff(r) for r in rings))
fco_filter = bandlimiting_filter(frequencies_cutoff)
if custom_basis_filter is not None:
basis_filter = lambda d, custom_basis_filter=custom_basis_filter, fco_filter=fco_filter: (
custom_basis_filter(d) and fco_filter(d))
else:
basis_filter = fco_filter
return basis_filter, rings, sigma, maximum_frequency
def _manual_fco3(max_radius: float) -> Callable[[float], float]:
r"""
Returns a method which takes as input the radius of a ring and returns the maximum frequency which can be sampled
on that ring.
Args:
max_radius (float): radius of the last ring touching the border of the grid
Returns:
a function which checks the attributes of individual basis elements and chooses whether to discard them or not
"""
def bl_filter(r: float) -> float:
max_freq = 0 if r == 0. else 1 if r == max_radius else 2
return max_freq
return bl_filter
def _manual_fco2(max_radius: float) -> Callable[[float], float]:
r"""
Returns a method which takes as input the radius of a ring and returns the maximum frequency which can be sampled
on that ring.
Args:
max_radius (float): radius of the last ring touching the border of the grid
Returns:
a function which checks the attributes of individual basis elements and chooses whether to discard them or not
"""
def bl_filter(r: float) -> float:
max_freq = 0 if r == 0. else min(2 * r, 1 if r == max_radius else 2 * r - (r + 1) % 2)
return max_freq
return bl_filter
def _manual_fco1(max_radius: float) -> Callable[[float], float]:
r"""
Returns a method which takes as input the radius of a ring and returns the maximum frequency which can be sampled
on that ring.
Args:
max_radius (float): radius of the last ring touching the border of the grid
Returns:
a function which checks the attributes of individual basis elements and chooses whether to discard them or not
"""
def bl_filter(r: float) -> float:
max_freq = 0 if r == 0. else min(2 * r, 2 if r == max_radius else 2 * r - (r + 1) % 2)
return max_freq
return bl_filter