torch/nn/_functions/vision.py

import torch
from torch.autograd import Function
from torch.autograd.function import once_differentiable
from torch._thnn import type2backend
from .thnn.auto import function_by_name
import torch.backends.cudnn as cudnn

MODE_ZEROS = 0
MODE_BORDER = 1


def grid_sampler(input, grid, padding_mode):
    if cudnn.is_acceptable(input.data) and padding_mode == 'zeros':
        return torch._C._VariableBase.cudnn_grid_sampler(input, grid)
    else:
        return GridSampler.apply(input, grid, padding_mode)


def affine_grid_generator(theta, size):
    if theta.data.is_cuda:
        if not cudnn.enabled:
            raise RuntimeError("AffineGridGenerator needs CuDNN for "
                               "processing CUDA inputs, but CuDNN is not enabled")
        if not cudnn.is_acceptable(theta.data):
            raise RuntimeError("AffineGridGenerator generator theta not acceptable for CuDNN")
        N, C, H, W = size
        return torch._C._VariableBase.cudnn_affine_grid_generator(theta, N, C, H, W)
    else:
        return AffineGridGenerator.apply(theta, size)


# TODO: Port these completely into C++


class GridSampler(Function):

    @staticmethod
    def forward(ctx, input, grid, padding_mode='zeros'):
        ctx.save_for_backward(input, grid)

        if padding_mode == 'zeros':
            ctx.padding_mode = MODE_ZEROS
        elif padding_mode == 'border':
            ctx.padding_mode = MODE_BORDER
        else:
            raise ValueError("padding_mode needs to be 'zeros' or 'border', but got {}"
                             .format(padding_mode))

        grid_sz = grid.size()
        backend = type2backend[type(input)]
        output = input.new(grid_sz[0], input.size(1), grid_sz[1], grid_sz[2])
        backend.SpatialGridSamplerBilinear_updateOutput(backend.library_state, input, grid, output, ctx.padding_mode)
        return output

    @staticmethod
    @once_differentiable
    def backward(ctx, grad_output):
        input, grid = ctx.saved_tensors
        padding_mode = ctx.padding_mode

        backend = type2backend[type(input)]
        grad_input = input.new(input.size())
        grad_grid = grid.new(grid.size())
        backend.SpatialGridSamplerBilinear_updateGradInput(
            backend.library_state, input, grad_input,
            grid, grad_grid, grad_output, padding_mode)
        return grad_input, grad_grid, None


class AffineGridGenerator(Function):

    @staticmethod
    def _enforce_cudnn(input):
        if not cudnn.enabled:
            raise RuntimeError("AffineGridGenerator needs CuDNN for "
                               "processing CUDA inputs, but CuDNN is not enabled")
        assert cudnn.is_acceptable(input)

    @staticmethod
    def forward(ctx, theta, size):
        assert type(size) == torch.Size
        N, C, H, W = size
        ctx.size = size
        if theta.is_cuda:
            AffineGridGenerator._enforce_cudnn(theta)
            assert False
        ctx.is_cuda = False
        base_grid = theta.new(N, H, W, 3)
        linear_points = torch.linspace(-1, 1, W) if W > 1 else torch.Tensor([-1])
        base_grid[:, :, :, 0] = torch.ger(torch.ones(H), linear_points).expand_as(base_grid[:, :, :, 0])
        linear_points = torch.linspace(-1, 1, H) if H > 1 else torch.Tensor([-1])
        base_grid[:, :, :, 1] = torch.ger(linear_points, torch.ones(W)).expand_as(base_grid[:, :, :, 1])
        base_grid[:, :, :, 2] = 1
        ctx.base_grid = base_grid
        grid = torch.bmm(base_grid.view(N, H * W, 3), theta.transpose(1, 2))
        grid = grid.view(N, H, W, 2)
        return grid

    @staticmethod
    @once_differentiable
    def backward(ctx, grad_grid):
        N, C, H, W = ctx.size
        assert grad_grid.size() == torch.Size([N, H, W, 2])
        assert ctx.is_cuda == grad_grid.is_cuda
        if grad_grid.is_cuda:
            AffineGridGenerator._enforce_cudnn(grad_grid)
            assert False
        base_grid = ctx.base_grid
        grad_theta = torch.bmm(
            base_grid.view(N, H * W, 3).transpose(1, 2),
            grad_grid.view(N, H * W, 2))
        grad_theta = grad_theta.transpose(1, 2)
        return grad_theta, None