Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add empty batch support for DeformConv2d #2782

Merged
merged 2 commits into from
Oct 11, 2020
Merged

Add empty batch support for DeformConv2d #2782

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
96ae800
Adding checks on forward and backward passes.
datumbox Oct 9, 2020
ca6df36
Adding unit-tests.
datumbox Oct 9, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 11 additions & 4 deletions test/test_ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -478,8 +478,7 @@ def expected_fn(self, x, weight, offset, bias, stride=1, padding=0, dilation=1):
out += bias.view(1, n_out_channels, 1, 1)
return out

def get_fn_args(self, device, contiguous):
batch_sz = 33
def get_fn_args(self, device, contiguous, batch_sz):
n_in_channels = 6
n_out_channels = 2
n_weight_grps = 2
Expand Down Expand Up @@ -516,7 +515,11 @@ def get_fn_args(self, device, contiguous):
return x, weight, offset, bias, stride, pad, dilation

def _test_forward(self, device, contiguous):
x, _, offset, _, stride, padding, dilation = self.get_fn_args(device, contiguous)
for batch_sz in [0, 33]:
self._test_forward_with_batchsize(device, contiguous, batch_sz)

def _test_forward_with_batchsize(self, device, contiguous, batch_sz):
x, _, offset, _, stride, padding, dilation = self.get_fn_args(device, contiguous, batch_sz)
in_channels = 6
out_channels = 2
kernel_size = (3, 2)
Expand All @@ -538,7 +541,11 @@ def _test_forward(self, device, contiguous):
res = layer(x, wrong_offset)

def _test_backward(self, device, contiguous):
x, weight, offset, bias, stride, padding, dilation = self.get_fn_args(device, contiguous)
for batch_sz in [0, 33]:
self._test_backward_with_batchsize(device, contiguous, batch_sz)

def _test_backward_with_batchsize(self, device, contiguous, batch_sz):
x, weight, offset, bias, stride, padding, dilation = self.get_fn_args(device, contiguous, batch_sz)

def func(x_, offset_, weight_, bias_):
return ops.deform_conv2d(x_, offset_, weight_, bias_, stride=stride, padding=padding, dilation=dilation)
Expand Down
9 changes: 9 additions & 0 deletions torchvision/csrc/cpu/DeformConv_cpu.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -326,6 +326,9 @@ at::Tensor DeformConv2d_forward_cpu(
out_w);

auto out = at::zeros({batch_sz, out_channels, out_h, out_w}, input.options());
if (batch_sz == 0) {
return out;
}

// Separate batches into blocks
out = out.view({batch_sz / n_parallel_imgs,
Expand Down Expand Up @@ -713,6 +716,9 @@ static std::tuple<at::Tensor, at::Tensor> deform_conv2d_backward_input_cpu(

auto grad_input = at::zeros_like(input);
auto grad_offset = at::zeros_like(offset);
if (batch_sz == 0) {
return std::make_tuple(grad_input, grad_offset);
}
auto columns = at::empty(
{n_in_channels * weight_w * weight_h, n_parallel_imgs * out_h * out_w},
input.options());
Expand Down Expand Up @@ -839,6 +845,9 @@ static at::Tensor deform_conv2d_backward_parameters_cpu(
long out_w = grad_out.size(3);

auto grad_weight = at::zeros_like(weight);
if (batch_sz == 0) {
return grad_weight;
}

at::Tensor grad_out_buf = grad_out
.reshape({batch_sz / n_parallel_imgs,
Expand Down
9 changes: 9 additions & 0 deletions torchvision/csrc/cuda/DeformConv_cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -343,6 +343,9 @@ at::Tensor DeformConv2d_forward_cuda(
out_w);

auto out = at::zeros({batch_sz, out_channels, out_h, out_w}, input.options());
if (batch_sz == 0) {
return out;
}

// Separate batches into blocks
out = out.view({batch_sz / n_parallel_imgs,
Expand Down Expand Up @@ -743,6 +746,9 @@ static std::tuple<at::Tensor, at::Tensor> deform_conv_backward_input_cuda(

auto grad_input = at::zeros_like(input);
auto grad_offset = at::zeros_like(offset);
if (batch_sz == 0) {
return std::make_tuple(grad_input, grad_offset);
}
auto columns = at::empty(
{n_in_channels * weight_w * weight_h, n_parallel_imgs * out_h * out_w},
input.options());
Expand Down Expand Up @@ -869,6 +875,9 @@ static at::Tensor deform_conv_backward_parameters_cuda(
long out_w = grad_out.size(3);

auto grad_weight = at::zeros_like(weight);
if (batch_sz == 0) {
return grad_weight;
}

at::Tensor grad_out_buf = grad_out.reshape(
{batch_sz / n_parallel_imgs,
Expand Down