Skip to content

Adjust bound_shape_inferencer to take 4 inputs for FCs #41452

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

Adjust bound_shape_inferencer to take 4 inputs for FCs #41452

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
28 changes: 28 additions & 0 deletions caffe2/opt/bound_shape_inference_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -632,6 +632,8 @@ TEST(BoundShapeInference, FC) {
CreateOperatorDef("FC", "", {"X0", "W0", "B0"}, {"Out0"}, {}));
net.add_op()->CopyFrom(
CreateOperatorDef("FCTransposed", "", {"X1", "W1", "B1"}, {"Out1"}, {}));
net.add_op()->CopyFrom(CreateOperatorDef(
"Int8FC", "", {"X2", "W2", "B2", "quant_param"}, {"Out2"}, {}));
ShapeInfoMap shape_map;
shape_map.emplace(
"W0",
Expand All @@ -649,6 +651,18 @@ TEST(BoundShapeInference, FC) {
{16, 1024}));
shape_map.emplace(
"B1", makeTensorInfo({TensorBoundShape_DimType_CONSTANT}, {1024}));

shape_map.emplace(
"W2",
makeTensorInfo(
{TensorBoundShape_DimType_CONSTANT,
TensorBoundShape_DimType_CONSTANT},
{16, 1024}));
shape_map.emplace(
"B2", makeTensorInfo({TensorBoundShape_DimType_CONSTANT}, {16}));
shape_map.emplace(
"quant_param", makeTensorInfo({TensorBoundShape_DimType_CONSTANT}, {1}));

BoundShapeSpec spec(20, 1000);
BoundShapeInferencer eng(spec);
eng.InferBoundShapeAndType(net, shape_map, nullptr);
Expand All @@ -673,6 +687,20 @@ TEST(BoundShapeInference, FC) {
"Out1",
{TensorBoundShape_DimType_BATCH, TensorBoundShape_DimType_CONSTANT},
{spec.max_batch_size, 1024});
verifyShapeInfo(
out_shape,
"X2",
{TensorBoundShape_DimType_BATCH, TensorBoundShape_DimType_CONSTANT},
{spec.max_batch_size, 1024},
TensorProto_DataType_UINT8,
true);
verifyShapeInfo(
out_shape,
"Out2",
{TensorBoundShape_DimType_BATCH, TensorBoundShape_DimType_CONSTANT},
{spec.max_batch_size, 16},
TensorProto_DataType_UINT8,
true);
}

TEST(BoundShapeInference, FC3D) {
Expand Down
12 changes: 11 additions & 1 deletion caffe2/opt/bound_shape_inferencer.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -603,7 +603,7 @@ void BoundShapeInferencer::InferConcat(const OperatorDef& op) {
}

void BoundShapeInferencer::InferFC(const OperatorDef& op) {
CAFFE_ENFORCE_EQ(op.input_size(), 3, "FC has to have 3 inputs");
CAFFE_ENFORCE(op.input_size() == 3 || op.input_size() == 4, "FC has to have 3 or 4 inputs");
const auto w_it = shape_info_.find(op.input(1));
CAFFE_ENFORCE(
w_it != shape_info_.end(),
Expand Down Expand Up @@ -668,6 +668,16 @@ void BoundShapeInferencer::InferFC(const OperatorDef& op) {
// Standard shape inference for outputs
std::vector<TensorShape> input_shapes{
shape_info_[op.input(0)].shape, w_shape_info.shape, b_shape_info.shape};
if(op.input_size() == 4) {
const auto quant_param_it = shape_info_.find(op.input(3));
CAFFE_ENFORCE(
quant_param_it != shape_info_.end(),
"Shape of quant_param input of FC ",
op.input(3),
" needs to be presented");
const ShapeInfo& quant_param_shape_info = quant_param_it->second;
input_shapes.emplace_back(quant_param_shape_info.shape);
}
std::vector<TensorShape> output_shapes = InferOutput(op, input_shapes);
CAFFE_ENFORCE_EQ(output_shapes.size(), 1);
TensorProto::DataType output_data_type;
Expand Down
10 changes: 10 additions & 0 deletions caffe2/quantization/server/int8_quant_scheme_blob_fill.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,6 +12,16 @@ REGISTER_CPU_OPERATOR(
OPERATOR_SCHEMA(Int8QuantSchemeBlobFill)
.NumInputs(0)
.NumOutputs(1)
.TensorInferenceFunction([](const OperatorDef& /* def */,
const vector<TensorShape>& in) {
vector<TensorShape> out;
TensorShape X = in[0];
X.clear_dims();
X.add_dims(1);
out.emplace_back(std::move(X));
out[0].set_data_type(TensorProto_DataType_STRING);
return out;
})
.Arg(
"quantization_kind",
"The kind of quant scheme that would be used to generate quant param")
Expand Down