diff --git a/paddle/fluid/inference/tensorrt/convert/bilinear_interp_v2_op.cc b/paddle/fluid/inference/tensorrt/convert/bilinear_interp_v2_op.cc index bd9e53bd7fd1f2725413bee1d51d54ba8174298d..77153d8ade56db9b180189d0d3fee8b1ec330b1d 100644 --- a/paddle/fluid/inference/tensorrt/convert/bilinear_interp_v2_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/bilinear_interp_v2_op.cc @@ -48,10 +48,21 @@ class BilinearInterpolateV2OpConverter : public OpConverter { auto out_w = PADDLE_GET_CONST(int, op_desc.GetAttr("out_w")); auto layer = TRT_ENGINE_ADD_LAYER(engine_, Resize, *input); - if (align_mode == 0 && !align_corners) { + if (align_mode == 0) { layer->setResizeMode(nvinfer1::ResizeMode::kLINEAR); } - +#if IS_TRT_VERSION_GE(8000) + if (align_corners == true) { + layer->setCoordinateTransformation( + nvinfer1::ResizeCoordinateTransformation::kALIGN_CORNERS); + } else { + layer->setCoordinateTransformation( + nvinfer1::ResizeCoordinateTransformation::kHALF_PIXEL); + } +#endif +#if !IS_TRT_VERSION_GE(8000) + layer->setAlignCorners(align_corners); +#endif auto in_dim = input->getDimensions(); float scale_h = -1.f; float scale_w = -1.f; @@ -95,7 +106,7 @@ class BilinearInterpolateV2OpConverter : public OpConverter { } } - if (out_h > 0 && out_w > 0) { + if (out_h > 0 && out_w > 0 && !(scale_w > 0. && scale_h > 0.)) { scale_h = static_cast(out_h) / static_cast(in_dim.d[h_axis]); scale_w = diff --git a/paddle/fluid/inference/tensorrt/op_teller.cc b/paddle/fluid/inference/tensorrt/op_teller.cc index 40b9590c519d83a53d3a87c6ec8cf2caee00b837..f6c09a7df9582c171f53f547c4a4f7b898740cc8 100644 --- a/paddle/fluid/inference/tensorrt/op_teller.cc +++ b/paddle/fluid/inference/tensorrt/op_teller.cc @@ -919,7 +919,6 @@ struct SimpleOpTypeSetTeller : public Teller { return false; } } - if (resize_inputs.find("OutSize") != resize_inputs.end()) { if (!with_dynamic_shape) { VLOG(3) << "Static shape don't support the OutSize for op_type " @@ -944,18 +943,11 @@ struct SimpleOpTypeSetTeller : public Teller { return false; } - auto align_corners = - PADDLE_GET_CONST(bool, desc.GetAttr("align_corners")); - if (align_corners != false) { - VLOG(3) - << "The bilinear_interp_v2 only supports align_corners with false."; - return false; - } - bool has_scale_input_size = (resize_inputs.find("Scale") != resize_inputs.end()); - if (has_scale_input_size && desc.Input("Scale").size() != 1) { + if (!has_scale_input_size || + (has_scale_input_size && desc.Input("Scale").size() != 1)) { const std::vector scale = PADDLE_GET_CONST(std::vector, desc.GetAttr("scale")); if (scale.size() <= 1) { diff --git a/test/ir/inference/test_trt_convert_bilinear_interp_v2.py b/test/ir/inference/test_trt_convert_bilinear_interp_v2.py index 5fb36deec1ab2dd6fac607fe785d10e41da08647..005b5da41a2fefbd9a9cc7ca0ee6a60644ee141b 100644 --- a/test/ir/inference/test_trt_convert_bilinear_interp_v2.py +++ b/test/ir/inference/test_trt_convert_bilinear_interp_v2.py @@ -38,7 +38,9 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest): def sample_program_configs(self): def generate_input1(attrs: List[Dict[str, Any]]): - return np.ones([1, 3, 64, 64]).astype(np.float32) + return np.random.uniform( + low=0.0, high=1.0, size=[1, 3, 64, 64] + ).astype(np.float32) def generate_input2(attrs: List[Dict[str, Any]]): return np.random.uniform(low=0.5, high=6.0, size=(2)).astype( @@ -46,56 +48,53 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest): ) for data_layout in ["NCHW", "NHWC"]: - for scale_y in [2.0, 1.0]: - for scale_x in [2.0]: - scale = [scale_y, scale_x] - for out_h in [32, 128]: - for out_w in [64]: - dics = [ - { - "data_layout": data_layout, - "interp_method": "bilinear", - "align_corners": False, - "align_mode": 0, - "scale": scale, - "out_h": out_h, - "out_w": out_w, - } - ] - - ops_config = [ - { - "op_type": "bilinear_interp_v2", - "op_inputs": { - "X": ["input_data"], - "Scale": ["input_scale"], - }, - "op_outputs": { - "Out": [ - "bilinear_interp_v2_output_data" - ] - }, - "op_attrs": dics[0], - } - ] - ops = self.generate_op_config(ops_config) - - program_config = ProgramConfig( - ops=ops, - weights={ - "input_scale": TensorConfig( - data_gen=partial(generate_input2, dics) - ) + for align_corners in [False, True]: + for scale_y in [2.0, 1.0]: + for scale_x in [2.0]: + scale = [scale_y, scale_x] + dics = [ + { + "data_layout": data_layout, + "interp_method": "bilinear", + "align_corners": align_corners, + "align_mode": 0, + "scale": scale, + "out_h": -1, + "out_w": -1, + } + ] + + ops_config = [ + { + "op_type": "bilinear_interp_v2", + "op_inputs": { + "X": ["input_data"], + "Scale": ["input_scale"], }, - inputs={ - "input_data": TensorConfig( - data_gen=partial(generate_input1, dics) - ) + "op_outputs": { + "Out": ["bilinear_interp_v2_output_data"] }, - outputs=["bilinear_interp_v2_output_data"], - ) + "op_attrs": dics[0], + } + ] + ops = self.generate_op_config(ops_config) + + program_config = ProgramConfig( + ops=ops, + weights={ + "input_scale": TensorConfig( + data_gen=partial(generate_input2, dics) + ) + }, + inputs={ + "input_data": TensorConfig( + data_gen=partial(generate_input1, dics) + ) + }, + outputs=["bilinear_interp_v2_output_data"], + ) - yield program_config + yield program_config def sample_predictor_configs( self, program_config @@ -136,9 +135,125 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest): program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True + ), (1e-5, 1e-5) + self.trt_param.precision = paddle_infer.PrecisionType.Half + yield self.create_inference_config(), generate_trt_nodes_num( + attrs, True + ), 1e-2 + + def test(self): + self.run_test() + + +class TrtConvertBilinearInterpV2Test1(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)) + ] + ver = paddle_infer.get_trt_compile_version() + # here is consistent with op_teller.cc + if ver[0] * 1000 + ver[1] * 100 + ver[2] * 10 < 7100: + return False + return True + + def sample_program_configs(self): + self.workspace_size = 1 << 32 + + def generate_input1(attrs: List[Dict[str, Any]]): + return np.random.uniform( + low=0.0, high=1.0, size=[1, 18, 144, 144] + ).astype(np.float32) + + for data_layout in ["NCHW", "NHWC"]: + for align_corners in [False, True]: + for out_h in [128, 288]: + for out_w in [288]: + dics = [ + { + "data_layout": data_layout, + "interp_method": "bilinear", + "align_corners": align_corners, + "align_mode": 0, + "scale": [], + "out_h": out_h, + "out_w": out_w, + } + ] + + ops_config = [ + { + "op_type": "bilinear_interp_v2", + "op_inputs": { + "X": ["input_data"], + }, + "op_outputs": { + "Out": ["bilinear_interp_v2_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=["bilinear_interp_v2_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, 18, 144, 144] + } + self.dynamic_shape.max_input_shape = { + "input_data": [8, 18, 144, 144] + } + self.dynamic_shape.opt_input_shape = { + "input_data": [4, 18, 144, 144] + } + + 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): + return 1, 2 + + 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(), generate_trt_nodes_num( + attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half program_config.set_input_type(np.float16) + yield self.create_inference_config(), generate_trt_nodes_num( + attrs, False + ), 1e-2 + + # 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, 1e-5) + self.trt_param.precision = paddle_infer.PrecisionType.Half yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2