Skip to content

transposed convolution: add channels last tests #4208

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

Closed
wants to merge 1 commit into from
Closed

transposed convolution: add channels last tests #4208

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
31 changes: 15 additions & 16 deletions kernels/portable/cpu/op_convolution.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -136,17 +136,6 @@ void conv2d_impl(
}
}
} else { // transposed convolution
if (bias_ptr != nullptr) {
out_coord[2] = 0;
out_coord[3] = 0;
size_t out_c_start_idx =
calculate_linear_index(out_coord, out_strides.data(), 4);
size_t out_c_end_idx = out_c_start_idx + out_H * out_W;
for (size_t out_ix = out_c_start_idx; out_ix < out_c_end_idx; out_ix++) {
out_ptr[out_ix] = convert<CTYPE, CTYPE_BIAS>(bias_ptr[out_c]);
}
}

w_coord[1] = out_c - out_c_start;

for (size_t in_y = 0; in_y < in_H; ++in_y) {
Expand Down Expand Up @@ -295,12 +284,22 @@ void convolution_wrapper(
bias.has_value() ? bias.value().const_data_ptr<CTYPE_BIAS>() : nullptr;

size_t out_N = out.size(0);
size_t out_C_per_group = out.size(1) / groups;
size_t out_C = out.size(1);
size_t out_C_per_group = out_C / groups;

if (transposed && bias_ptr == nullptr) {
// If bias is not present, we need to initialize the output to 0
// before we can accumulate into it.
memset(out_ptr, 0, out.nbytes());
if (transposed) {
// For transposed convolution, we need to initialized the output before we
// can accumulate into it.
if (bias_ptr == nullptr) {
// If bias is not present, we need to initialize the output to 0
memset(out_ptr, 0, out.nbytes());
} else {
// If bias is present, we initialize the output to the bias value
for (size_t out_ix = 0; out_ix < out.numel(); ++out_ix) {
out_ptr[out_ix] = convert<CTYPE, CTYPE_BIAS>(
bias_ptr[(out_ix / out_strides[1]) % out_C]);
}
}
}

for (size_t batch = 0; batch < out_N; ++batch) {
Expand Down
114 changes: 113 additions & 1 deletion kernels/test/op_convolution_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -556,7 +556,7 @@ TEST_F(OpConvCorrectnessTest, TransposedNonDefaultParams) {
Tensor input = tf.full({2, 6, 4, 5}, 2.0);
Tensor weight = tf.full({6, 1, 2, 2}, 0.5);
Tensor bias = tf.make({3}, {1, 2, 3});
Tensor out = tf.zeros({2, 3, 3, 6});
Tensor out = tf.full({2, 3, 3, 6}, 0.7);
Tensor expected = tf.make(
{2, 3, 3, 6},
{1, 1, 1, 1, 1, 1, 1, 3, 3, 1, 3, 3, 1, 3, 3, 1, 3, 3, 2, 2, 2, 2,
Expand Down Expand Up @@ -587,6 +587,118 @@ TEST_F(OpConvCorrectnessTest, TransposedNonDefaultParams) {
EXPECT_TENSOR_CLOSE(out, expected);
}

template <typename T>
std::vector<T> get_channels_last_data(const Tensor& t) {
const std::vector<int32_t> sizes(t.sizes().begin(), t.sizes().end());
std::vector<T> contiguous_data(
t.const_data_ptr<T>(), t.const_data_ptr<T>() + t.numel());
std::vector<T> channels_last_data(t.numel());
int32_t N = sizes[0];
int32_t C = sizes[1];
int32_t H = sizes[2];
int32_t W = sizes[3];
for (int32_t n = 0; n < N; ++n) {
for (int32_t c = 0; c < C; ++c) {
for (int32_t h = 0; h < H; ++h) {
for (int32_t w = 0; w < W; ++w) {
// Calculate the index in the original blob
int32_t old_index = ((n * C + c) * H + h) * W + w;
// Calculate the index in the new blob
int32_t new_index = ((n * H + h) * W + w) * C + c;
// Copy the data
channels_last_data[new_index] = contiguous_data[old_index];
}
}
}
}
return channels_last_data;
}

TEST_F(OpConvCorrectnessTest, TransposedDefaultParamsChannelsLast) {
TensorFactory<ScalarType::Float> tf;

Tensor input = tf.full_channels_last({2, 4, 3, 2}, 2.0);
Tensor weight = tf.full_channels_last({4, 1, 2, 2}, 0.5);
optional<Tensor> bias;
Tensor out = tf.full_channels_last({2, 2, 4, 3}, 0.7);
Tensor expected =
tf.make({2, 2, 4, 3}, {2, 4, 2, 4, 8, 4, 4, 8, 4, 2, 4, 2, 2, 4, 2, 4,
8, 4, 4, 8, 4, 2, 4, 2, 2, 4, 2, 4, 8, 4, 4, 8,
4, 2, 4, 2, 2, 4, 2, 4, 8, 4, 4, 8, 4, 2, 4, 2});

const std::vector<int32_t> sizes(
expected.sizes().begin(), expected.sizes().end());
std::vector<float> channels_last_data =
get_channels_last_data<float>(expected);
Tensor expected_channels_last =
tf.make_channels_last(sizes, channels_last_data);

int64_t stride[1] = {1};
int64_t padding[1] = {0};
int64_t dilation[1] = {1};
bool transposed = true;
int64_t output_padding[1] = {0};
int64_t groups = 2;

op_convolution_out(
input,
weight,
exec_aten::optional<Tensor>(bias),
exec_aten::ArrayRef<int64_t>{stride, 1},
exec_aten::ArrayRef<int64_t>{padding, 1},
exec_aten::ArrayRef<int64_t>{dilation, 1},
transposed,
exec_aten::ArrayRef<int64_t>{output_padding, 1},
groups,
out);

EXPECT_TENSOR_CLOSE(out, expected_channels_last);
}

TEST_F(OpConvCorrectnessTest, TransposedNonDefaultParamsChannelsLast) {
TensorFactory<ScalarType::Float> tf;

Tensor input = tf.full_channels_last({2, 6, 4, 5}, 2.0);
Tensor weight = tf.full_channels_last({6, 1, 2, 2}, 0.5);
Tensor bias = tf.make({3}, {1, 2, 3});
Tensor out = tf.full_channels_last({2, 3, 3, 6}, 0.7);
Tensor expected = tf.make(
{2, 3, 3, 6},
{1, 1, 1, 1, 1, 1, 1, 3, 3, 1, 3, 3, 1, 3, 3, 1, 3, 3, 2, 2, 2, 2,
2, 2, 2, 4, 4, 2, 4, 4, 2, 4, 4, 2, 4, 4, 3, 3, 3, 3, 3, 3, 3, 5,
5, 3, 5, 5, 3, 5, 5, 3, 5, 5, 1, 1, 1, 1, 1, 1, 1, 3, 3, 1, 3, 3,
1, 3, 3, 1, 3, 3, 2, 2, 2, 2, 2, 2, 2, 4, 4, 2, 4, 4, 2, 4, 4, 2,
4, 4, 3, 3, 3, 3, 3, 3, 3, 5, 5, 3, 5, 5, 3, 5, 5, 3, 5, 5});

const std::vector<int32_t> sizes(
expected.sizes().begin(), expected.sizes().end());
std::vector<float> channels_last_data =
get_channels_last_data<float>(expected);
Tensor expected_channels_last =
tf.make_channels_last(sizes, channels_last_data);

int64_t stride[1] = {3};
int64_t padding[1] = {7};
int64_t dilation[1] = {5};
bool transposed = true;
int64_t output_padding[1] = {2};
int64_t groups = 3;

op_convolution_out(
input,
weight,
exec_aten::optional<Tensor>(bias),
exec_aten::ArrayRef<int64_t>{stride, 1},
exec_aten::ArrayRef<int64_t>{padding, 1},
exec_aten::ArrayRef<int64_t>{dilation, 1},
transposed,
exec_aten::ArrayRef<int64_t>{output_padding, 1},
groups,
out);

EXPECT_TENSOR_CLOSE(out, expected_channels_last);
}

TEST_F(OpConvCorrectnessTest, InvalidOutputPadding) {
TensorFactory<ScalarType::Float> tf;

Expand Down
Loading