diff --git a/paddle/fluid/inference/tensorrt/op_teller.cc b/paddle/fluid/inference/tensorrt/op_teller.cc index d11b4c68cd245f2ba3d2f9ac5d3f6abfce92dcdf..3e6a275e14c388f263827d35de7ac0ca22bf1f70 100644 --- a/paddle/fluid/inference/tensorrt/op_teller.cc +++ b/paddle/fluid/inference/tensorrt/op_teller.cc @@ -1265,6 +1265,24 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8, } } + if (op_type == "hard_sigmoid") { + if (!with_dynamic_shape) { + auto* block = desc.Block(); + if (block == nullptr) { + VLOG(3) << "The block is null."; + return false; + } + auto x_var_name = desc.Input("X")[0]; + auto* x_var_desc = block->FindVar(x_var_name); + const auto x_shape = x_var_desc->GetShape(); + if (x_shape.size() <= 2) { + VLOG(3) << "hard_sigmoid op does not support input's dim less than 3 " + "in tensorrt."; + return false; + } + } + } + if ((*teller)(op_type, desc, use_no_calib_int8)) return true; } diff --git a/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_hard_sigmoid.py b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_hard_sigmoid.py new file mode 100644 index 0000000000000000000000000000000000000000..d803d9e4616139d832d58f3a02a2e8aa4d7a32f1 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_hard_sigmoid.py @@ -0,0 +1,126 @@ +# 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 + + +class TrtConvertHardSigmoidTest_dim_2(TrtLayerAutoScanTest): + def is_program_valid(self, program_config: ProgramConfig) -> bool: + return True + + def sample_program_configs(self): + def generate_input(shape): + return np.random.random(shape).astype(np.float32) + + for batch in [1, 2, 4]: + for shape in [[batch, 64], [batch, 32, 64], [batch, 64, 32, 128]]: + self.input_dim = len(shape) + for slope in [0.1, 0.5]: + for offset in [0.2, 0.7]: + dics = [{"slope": slope, "offset": offset}] + ops_config = [{ + "op_type": "hard_sigmoid", + "op_inputs": { + "X": ["input_data"], + }, + "op_outputs": { + "Out": ["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_input, shape)) + }, + outputs=["output_data"]) + + yield program_config + + def sample_predictor_configs( + self, program_config) -> (paddle_infer.Config, List[int], float): + def generate_dynamic_shape(attrs): + if self.input_dim == 2: + self.dynamic_shape.min_input_shape = {"input_data": [1, 8]} + self.dynamic_shape.max_input_shape = {"input_data": [64, 128]} + self.dynamic_shape.opt_input_shape = {"input_data": [2, 16]} + elif self.input_dim == 3: + self.dynamic_shape.min_input_shape = {"input_data": [1, 8, 8]} + self.dynamic_shape.max_input_shape = { + "input_data": [64, 128, 256] + } + self.dynamic_shape.opt_input_shape = {"input_data": [2, 16, 64]} + elif self.input_dim == 4: + self.dynamic_shape.min_input_shape = { + "input_data": [1, 8, 8, 4] + } + self.dynamic_shape.max_input_shape = { + "input_data": [64, 128, 256, 512] + } + self.dynamic_shape.opt_input_shape = { + "input_data": [2, 16, 64, 128] + } + + def clear_dynamic_shape(): + self.dynamic_shape.max_input_shape = {} + self.dynamic_shape.min_input_shape = {} + self.dynamic_shape.opt_input_shape = {} + + attrs = [ + program_config.ops[i].attrs + for i in range(len(program_config.ops)) + ] + + # for static_shape + clear_dynamic_shape() + self.trt_param.precision = paddle_infer.PrecisionType.Float32 + yield self.create_inference_config(), (1, 2), 1e-5 + self.trt_param.precision = paddle_infer.PrecisionType.Half + yield self.create_inference_config(), (1, 2), 1e-5 + + # for dynamic_shape + generate_dynamic_shape(attrs) + self.trt_param.precision = paddle_infer.PrecisionType.Float32 + yield self.create_inference_config(), (1, 2), 1e-5 + self.trt_param.precision = paddle_infer.PrecisionType.Half + yield self.create_inference_config(), (1, 2), 1e-5 + + def add_skip_trt_case(self): + def teller(program_config, predictor_config): + if len(self.dynamic_shape. + min_input_shape) == 0 and self.input_dim == 2: + return True + return False + + self.add_skip_case( + teller, SkipReasons.TRT_NOT_SUPPORT, + "Need to repair the case: the output of trt and GPU has diff when inputs' dims is 2 in static shape mode." + ) + + def test(self): + self.add_skip_trt_case() + self.run_test() + + +if __name__ == "__main__": + unittest.main()