-
Notifications
You must be signed in to change notification settings - Fork 5.6k
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
* strided_slice * fix: compiler error because of size() * fix: warning * fix : warning * init input_shape * fix:forget punctuation
- Loading branch information
feng_shuai
authored
Apr 12, 2022
1 parent
43d5cca
commit b861022
Showing
5 changed files
with
255 additions
and
0 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
131 changes: 131 additions & 0 deletions
131
paddle/fluid/inference/tensorrt/convert/strided_slice_op.cc
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,131 @@ | ||
/* Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. | ||
Licensed under the Apache License, Version 2.0 (the "License"); | ||
you may not use this file except in compliance with the License. | ||
You may obtain a copy of the License at | ||
http://www.apache.org/licenses/LICENSE-2.0 | ||
Unless required by applicable law or agreed to in writing, software | ||
distributed under the License is distributed on an "AS IS" BASIS, | ||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
See the License for the specific language governing permissions and | ||
limitations under the License. */ | ||
|
||
#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" | ||
|
||
namespace paddle { | ||
namespace framework { | ||
class Scope; | ||
namespace proto { | ||
class OpDesc; | ||
} // namespace proto | ||
} // namespace framework | ||
} // namespace paddle | ||
|
||
namespace paddle { | ||
namespace inference { | ||
namespace tensorrt { | ||
|
||
/* | ||
* Stack converter from fluid to tensorRT. | ||
*/ | ||
class StridedSliceOpConverter : public OpConverter { | ||
public: | ||
void operator()(const framework::proto::OpDesc& op, | ||
const framework::Scope& scope, bool test_mode) override { | ||
VLOG(4) << "convert fluid StridedSlice op to tensorrt Slice layer"; | ||
|
||
framework::OpDesc op_desc(op, nullptr); | ||
auto* input = engine_->GetITensor(op_desc.Input("Input")[0]); | ||
nvinfer1::Dims input_dims = input->getDimensions(); | ||
|
||
std::vector<int> axes = | ||
BOOST_GET_CONST(std::vector<int>, op_desc.GetAttr("axes")); | ||
std::vector<int> starts = | ||
BOOST_GET_CONST(std::vector<int>, op_desc.GetAttr("starts")); | ||
std::vector<int> ends = | ||
BOOST_GET_CONST(std::vector<int>, op_desc.GetAttr("ends")); | ||
std::vector<int> strides = | ||
BOOST_GET_CONST(std::vector<int>, op_desc.GetAttr("strides")); | ||
|
||
nvinfer1::Dims start; | ||
start.nbDims = input_dims.nbDims; | ||
int axes_size = axes.size(); | ||
for (int i = 0; i < start.nbDims; i++) { | ||
start.d[i] = 0; | ||
} | ||
for (int i = 0; i < axes_size; i++) { | ||
start.d[axes[i]] = starts[i]; | ||
} | ||
|
||
nvinfer1::Dims stride; | ||
stride.nbDims = input_dims.nbDims; | ||
for (int i = 0; i < stride.nbDims; i++) { | ||
stride.d[i] = 1; | ||
} | ||
for (int i = 0; i < axes_size; i++) { | ||
stride.d[axes[i]] = strides[i]; | ||
} | ||
|
||
nvinfer1::Dims size; | ||
size.nbDims = input_dims.nbDims; | ||
for (int i = 0; i < size.nbDims; i++) { | ||
size.d[i] = 1; | ||
} | ||
|
||
auto output_name = op_desc.Output("Out")[0]; | ||
|
||
auto create_weights = [&](const std::vector<int>& data, | ||
const std::string& type) -> int* { | ||
std::unique_ptr<framework::Tensor> tmp_tensor(new framework::Tensor()); | ||
int data_size = data.size(); | ||
tmp_tensor->Resize({data_size}); | ||
auto* tmp_data = tmp_tensor->mutable_data<int>(platform::CPUPlace()); | ||
for (int i = 0; i < data_size; i++) { | ||
tmp_data[i] = data[i]; | ||
} | ||
|
||
engine_->SetWeights(output_name + "_add_slice_op_" + type, | ||
std::move(tmp_tensor)); | ||
return tmp_data; | ||
}; | ||
|
||
std::vector<int> const_weight(input_dims.nbDims, 1); | ||
for (int i = 0; i < axes_size; i++) { | ||
const_weight[axes[i]] = strides[i]; | ||
} | ||
|
||
int* weight_data = create_weights(const_weight, "size"); | ||
|
||
TensorRTEngine::Weight weight{nvinfer1::DataType::kINT32, | ||
static_cast<void*>(weight_data), | ||
static_cast<size_t>(input_dims.nbDims)}; | ||
|
||
int input_dim_size = input_dims.nbDims; | ||
nvinfer1::Dims input_shape; | ||
input_shape.nbDims = 1; | ||
input_shape.d[0] = input_dim_size; | ||
|
||
auto const_layer = | ||
TRT_ENGINE_ADD_LAYER(engine_, Constant, input_shape, weight.get()); | ||
|
||
auto shape_layer = TRT_ENGINE_ADD_LAYER(engine_, Shape, *input); | ||
|
||
auto size_layer = TRT_ENGINE_ADD_LAYER( | ||
engine_, ElementWise, *shape_layer->getOutput(0), | ||
*const_layer->getOutput(0), nvinfer1::ElementWiseOperation::kDIV); | ||
|
||
auto* layer = | ||
TRT_ENGINE_ADD_LAYER(engine_, Slice, *input, start, size, stride); | ||
layer->setInput(2, *size_layer->getOutput(0)); | ||
|
||
RreplenishLayerAndOutput(layer, "strided_slice", {output_name}, test_mode); | ||
} | ||
}; | ||
|
||
} // namespace tensorrt | ||
} // namespace inference | ||
} // namespace paddle | ||
|
||
REGISTER_TRT_OP_CONVERTER(strided_slice, StridedSliceOpConverter); |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
120 changes: 120 additions & 0 deletions
120
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_strided_slice.py
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,120 @@ | ||
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved. | ||
# | ||
# Licensed under the Apache License, Version 2.0 (the "License"); | ||
# you may not use this file except in compliance with the License. | ||
# You may obtain a copy of the License at | ||
# | ||
# http://www.apache.org/licenses/LICENSE-2.0 | ||
# | ||
# Unless required by applicable law or agreed to in writing, software | ||
# distributed under the License is distributed on an "AS IS" BASIS, | ||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
# See the License for the specific language governing permissions and | ||
# limitations under the License. | ||
|
||
from trt_layer_auto_scan_test import TrtLayerAutoScanTest, SkipReasons | ||
from program_config import TensorConfig, ProgramConfig | ||
import numpy as np | ||
import paddle.inference as paddle_infer | ||
from functools import partial | ||
from typing import Optional, List, Callable, Dict, Any, Set | ||
import unittest | ||
|
||
|
||
class TrtConvertStridedSliceTest(TrtLayerAutoScanTest): | ||
def is_program_valid(self, program_config: ProgramConfig) -> bool: | ||
inputs = program_config.inputs | ||
weights = program_config.weights | ||
attrs = [ | ||
program_config.ops[i].attrs | ||
for i in range(len(program_config.ops)) | ||
] | ||
return True | ||
|
||
def sample_program_configs(self): | ||
def generate_input1(attrs: List[Dict[str, Any]]): | ||
return np.ones([1, 56, 56, 192]).astype(np.float32) | ||
|
||
for axes in [[1, 2]]: | ||
for starts in [[1, 1]]: | ||
for ends in [[10000000, 10000000]]: | ||
for decrease_axis in [[]]: | ||
for infer_flags in [[1, 1]]: | ||
for strides in [[2, 2]]: | ||
dics = [{ | ||
"axes": axes, | ||
"starts": starts, | ||
"ends": ends, | ||
"decrease_axis": decrease_axis, | ||
"infer_flags": infer_flags, | ||
"strides": strides | ||
}] | ||
|
||
ops_config = [{ | ||
"op_type": "strided_slice", | ||
"op_inputs": { | ||
"Input": ["input_data"] | ||
}, | ||
"op_outputs": { | ||
"Out": ["slice_output_data"] | ||
}, | ||
"op_attrs": dics[0] | ||
}] | ||
ops = self.generate_op_config(ops_config) | ||
|
||
program_config = ProgramConfig( | ||
ops=ops, | ||
weights={}, | ||
inputs={ | ||
"input_data": TensorConfig( | ||
data_gen=partial(generate_input1, | ||
dics)) | ||
}, | ||
outputs=["slice_output_data"]) | ||
|
||
yield program_config | ||
|
||
def sample_predictor_configs( | ||
self, program_config) -> (paddle_infer.Config, List[int], float): | ||
def generate_dynamic_shape(attrs): | ||
self.dynamic_shape.min_input_shape = { | ||
"input_data": [1, 56, 56, 192] | ||
} | ||
self.dynamic_shape.max_input_shape = { | ||
"input_data": [8, 56, 56, 192] | ||
} | ||
self.dynamic_shape.opt_input_shape = { | ||
"input_data": [4, 56, 56, 192] | ||
} | ||
|
||
def clear_dynamic_shape(): | ||
self.dynamic_shape.min_input_shape = {} | ||
self.dynamic_shape.max_input_shape = {} | ||
self.dynamic_shape.opt_input_shape = {} | ||
|
||
def generate_trt_nodes_num(attrs, dynamic_shape): | ||
inputs = program_config.inputs | ||
|
||
if dynamic_shape: | ||
for i in range(len(attrs[0]["starts"])): | ||
if attrs[0]["starts"][i] < 0 or attrs[0]["ends"][i] < 0: | ||
return 0, 3 | ||
if not dynamic_shape: | ||
for x in attrs[0]["axes"]: | ||
if x == 0: | ||
return 0, 3 | ||
return 1, 2 | ||
|
||
attrs = [ | ||
program_config.ops[i].attrs | ||
for i in range(len(program_config.ops)) | ||
] | ||
|
||
# for dynamic_shape | ||
generate_dynamic_shape(attrs) | ||
self.trt_param.precision = paddle_infer.PrecisionType.Float32 | ||
yield self.create_inference_config(), generate_trt_nodes_num(attrs, | ||
True), 1e-5 | ||
|
||
def test(self): | ||
self.run_test() |