m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Refine error message of leaky_relu, tensorrt_engine, split, prelu op converter (#23661)

Browse Source
revert-22778-infer_var_type
Pei Yang 5 years ago committed by GitHub
parent 1b5122ba07
commit 015acdbfb1
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
5 changed files with 98 additions and 28 deletions
Show Stats Download Patch File Download Diff File

45
paddle/fluid/inference/tensorrt/convert/leaky_relu_op.cc
Unescape View File

@ -27,12 +27,20 @@ class LeakyReluOpConverter : public OpConverter {
framework::OpDesc op_desc(op, nullptr);
// Declare inputs
int input_num = op_desc.Input("X").size();
PADDLE_ENFORCE(input_num == 1);
size_t input_num = op_desc.Input("X").size();
PADDLE_ENFORCE_EQ(input_num, 1UL,
platform::errors::InvalidArgument(
"Invalid number of TRT leaky_relu op converter "
"inputs. Expected 1, but received %d",
input_num));
auto* input = engine_->GetITensor(op_desc.Input("X")[0]);
// Get output
size_t output_num = op_desc.Output("Out").size();
PADDLE_ENFORCE(output_num == 1);
PADDLE_ENFORCE_EQ(output_num, 1UL,
platform::errors::InvalidArgument(
"Invalid number of TRT leaky_relu op converter "
"outputs. Expected 1, but received %d",
output_num));
// Get attrs
float alpha = boost::get<float>(op_desc.GetAttr("alpha"));
nvinfer1::ILayer* output_layer = nullptr;
@ -66,11 +74,17 @@ class LeakyReluOpConverter : public OpConverter {
auto* scale_layer = TRT_ENGINE_ADD_LAYER(
engine_, Scale, *input, nvinfer1::ScaleMode::kUNIFORM, shift.get(),
scale.get(), power.get());
PADDLE_ENFORCE(nullptr != scale_layer);
PADDLE_ENFORCE_NOT_NULL(
scale_layer, platform::errors::InvalidArgument(
"Invalid scale layer in leaky_relu TRT op converter. "
"The scale layer should not be null."));
// y_relu = (x > 0) : x : 0
auto* relu_layer = TRT_ENGINE_ADD_LAYER(engine_, Activation, *input,
nvinfer1::ActivationType::kRELU);
PADDLE_ENFORCE(nullptr != relu_layer);
PADDLE_ENFORCE_NOT_NULL(
relu_layer, platform::errors::InvalidArgument(
"Invalid relu layer in leaky_relu TRT op converter. "
"The relu layer should not be null."));
//
TensorRTEngine::Weight sub_scale{nvinfer1::DataType::kFLOAT, &alpha_data[1],
1};
@ -78,16 +92,29 @@ class LeakyReluOpConverter : public OpConverter {
TRT_ENGINE_ADD_LAYER(engine_, Scale, *(relu_layer->getOutput(0)),
nvinfer1::ScaleMode::kUNIFORM, shift.get(),
sub_scale.get(), power.get());
PADDLE_ENFORCE(nullptr != scale_relu_layer);
PADDLE_ENFORCE_NOT_NULL(
scale_relu_layer,
platform::errors::InvalidArgument(
"Invalid scale_relu layer in leaky_relu TRT op converter. The "
"scale_relu layer should not be null."));
output_layer =
TRT_ENGINE_ADD_LAYER(engine_, ElementWise, *(scale_layer->getOutput(0)),
*(scale_relu_layer->getOutput(0)),
nvinfer1::ElementWiseOperation::kSUM);
PADDLE_ENFORCE(nullptr != output_layer);
PADDLE_ENFORCE_NOT_NULL(
output_layer, platform::errors::InvalidArgument(
"Invalid output layer in leaky_relu TRT op "
"converter. The output layer should not be null."));
// keep alpha tensor to avoid release it's memory
std::string alpha_name = op_desc.Output("Out")[0] + "_alpha";
PADDLE_ENFORCE(engine_->weight_map.find(alpha_name) ==
engine_->weight_map.end());
bool alpha_not_in_weight_map =
(engine_->weight_map.find(alpha_name) == engine_->weight_map.end());
PADDLE_ENFORCE_EQ(alpha_not_in_weight_map, true,
platform::errors::InvalidArgument(
"The name of parameter alpha in leaky_relu TRT op "
"converter is already "
"found in the weight map. The same weight cannot be "
"set twice. Please check if it is already set."));
engine_->SetWeights(alpha_name, std::move(alpha_tensor));
#endif
auto output_name = op_desc.Output("Out")[0];

18
paddle/fluid/inference/tensorrt/convert/prelu_op.cc
Unescape View File

@ -30,17 +30,27 @@ class PReluOpConverter : public OpConverter {
framework::OpDesc op_desc(op, nullptr);
// Declare inputs
int input_num = op_desc.Input("X").size();
PADDLE_ENFORCE(input_num == 1);
size_t input_num = op_desc.Input("X").size();
PADDLE_ENFORCE_EQ(input_num, 1UL,
platform::errors::InvalidArgument(
"Invalid input X's size of prelu TRT converter. "
"Expected 1, received %d.",
input_num));
auto* input = engine_->GetITensor(op_desc.Input("X")[0]);
// Get output
size_t output_num = op_desc.Output("Out").size();
PADDLE_ENFORCE(output_num == 1);
PADDLE_ENFORCE_EQ(output_num, 1UL,
platform::errors::InvalidArgument(
"Invalid output Out's size of prelu TRT converter. "
"Expected 1, received %d.",
output_num));
// Get attrs
std::string mode = boost::get<std::string>(op_desc.GetAttr("mode"));
//
auto* alpha_var = scope.FindVar(op_desc.Input("Alpha")[0]);
PADDLE_ENFORCE_NOT_NULL(alpha_var);
PADDLE_ENFORCE_NOT_NULL(
alpha_var, platform::errors::NotFound(
"Variable Alpha of prelu TRT converter is not found."));
auto* alpha_tensor = alpha_var->GetMutable<framework::LoDTensor>();
platform::CPUPlace cpu_place;

23
paddle/fluid/inference/tensorrt/convert/split_op.cc
Unescape View File

@ -29,14 +29,21 @@ class SplitOpConverter : public OpConverter {
// Declare inputs
auto* input = engine_->GetITensor(op_desc.Input("X")[0]);
auto input_dims = input->getDimensions();
int input_num = op_desc.Input("X").size();
size_t input_num = op_desc.Input("X").size();
size_t output_num = op_desc.Output("Out").size();
// Get Attrs
PADDLE_ENFORCE(input_num == 1);
PADDLE_ENFORCE_EQ(input_num, 1UL,
platform::errors::InvalidArgument(
"Invalid input X's size of split TRT converter. "
"Expected 1, received %d.",
input_num));
int axis = boost::get<int>(op_desc.GetAttr("axis"));
// split on batch is not supported in TensorRT
PADDLE_ENFORCE(axis != 0);
PADDLE_ENFORCE_NE(
axis, 0,
platform::errors::InvalidArgument(
"Invalid split axis. Split on batch is not supported in TensorRT"));
std::vector<int> output_lengths =
boost::get<std::vector<int>>(op_desc.GetAttr("sections"));
@ -58,9 +65,13 @@ class SplitOpConverter : public OpConverter {
"The (%d) dim of input should not be -1", axis));
if (num > 0) {
int64_t in_axis_dim = input_dims.d[axis];
PADDLE_ENFORCE_EQ(in_axis_dim % num, 0,
"Tensor split does not result"
" in an equal division");
PADDLE_ENFORCE_EQ(
in_axis_dim % num, 0,
platform::errors::InvalidArgument(
"Invalid number to split. Tensor split does not result"
" in an equal division of dimensions. Axis dim = %d %% num = %d "
"!= 0",
in_axis_dim, num));
size_t out_axis_dim = in_axis_dim / num;
for (int i = 0; i < num; ++i) {
output_lengths.push_back(out_axis_dim);

30
paddle/fluid/inference/tensorrt/plugin/split_op_plugin.cu
Unescape View File

@ -45,8 +45,17 @@ __device__ int upper_bound(T const* vals, int n, T const& key) {
nvinfer1::Dims SplitPlugin::getOutputDimensions(
int index, const nvinfer1::Dims* input_dims, int num_inputs) {
PADDLE_ENFORCE_EQ(num_inputs, 1);
PADDLE_ENFORCE_LT(index, this->getNbOutputs());
PADDLE_ENFORCE_EQ(num_inputs, 1,
platform::errors::InvalidArgument(
"Invalid number of inputs of split TRT plugin. "
"Expected 1, received %d.",
num_inputs));
PADDLE_ENFORCE_LT(
index, this->getNbOutputs(),
platform::errors::InvalidArgument(
"Index of output should be less than the total number of outputs in "
"split TensorRT plugin. Received index = %d >= total outputs = %d",
index, this->getNbOutputs()));
nvinfer1::Dims output_dims = input_dims[0];
output_dims.d[axis_] = output_length_.at(index);
@ -54,7 +63,11 @@ nvinfer1::Dims SplitPlugin::getOutputDimensions(
}
int SplitPlugin::initialize() {
PADDLE_ENFORCE_LE(axis_, nvinfer1::Dims::MAX_DIMS);
PADDLE_ENFORCE_LE(axis_, nvinfer1::Dims::MAX_DIMS,
platform::errors::InvalidArgument(
"Axis dimension exceeds max dimension in TensorRT. "
"Received axis = %d > MAX_DIMS = %d",
axis_, nvinfer1::Dims::MAX_DIMS));
// notice input dims is [C, H, W]
nvinfer1::Dims dims = this->getInputDims(0);
outer_rows_ = 1;
@ -111,9 +124,12 @@ int SplitPlugin::enqueue(int batchSize, const void* const* inputs,
float const* input_ptr = reinterpret_cast<float const*>(inputs[0]);
float* const* h_odatas = reinterpret_cast<float* const*>(outputs);
float** output_ptrs = thrust::raw_pointer_cast(&d_output_ptrs_[0]);
PADDLE_ENFORCE_CUDA_SUCCESS(cudaMemcpyAsync(
output_ptrs, h_odatas, d_output_ptrs_.size() * sizeof(float*),
cudaMemcpyHostToDevice, stream));
PADDLE_ENFORCE_CUDA_SUCCESS(
cudaMemcpyAsync(output_ptrs, h_odatas,
d_output_ptrs_.size() * sizeof(float*),
cudaMemcpyHostToDevice, stream),
platform::errors::External(
"CUDA Memcpy failed during split plugin run."));
int outer_rows = outer_rows_ * batchSize;
@ -159,7 +175,7 @@ bool SplitPluginDynamic::supportsFormatCombination(
int nb_outputs) {
PADDLE_ENFORCE_NOT_NULL(
in_out, platform::errors::InvalidArgument(
"The input of swish plugin shoule not be nullptr."));
"The input of split plugin should not be nullptr."));
PADDLE_ENFORCE_LT(
pos, nb_inputs + nb_outputs,

10
paddle/fluid/operators/tensorrt/tensorrt_engine_op.h
Unescape View File

@ -232,8 +232,14 @@ class TensorRTEngineOp : public framework::OperatorBase {
auto t_shape = framework::vectorize<int64_t>(t.dims());
runtime_batch = t_shape[0];
const int bind_index = engine->engine()->getBindingIndex(x.c_str());
PADDLE_ENFORCE(bind_index < num_bindings,
"The bind index should be less than num_bindings");
PADDLE_ENFORCE_LT(
bind_index, num_bindings,
platform::errors::InvalidArgument(
"Wrong TRT engine input binding index. Expected The "
"binding index of TRT engine input to be less than "
"the number of inputs and outputs. Received binding "
"index=%d >= total inputs and outputs=%d",
bind_index, num_bindings));
if (!engine->with_dynamic_shape()) {
// check if the input shapes are consistent with model.
if (HasAttr(x + "_shape")) {

Write Preview
Loading…
Cancel
Save