Skip to content

P
Paddle

PaddlePaddle / Paddle
1 年多 前同步成功

通知 2302
Star 20931
Fork 5422
P
Paddle
未验证 提交 f0422a28 编写于 作者: Z zhoutianzi666 提交者： GitHub
浏览文件
操作

[Paddle-TRT]fix slice, bilinear_interp_v2 in trt 7011 (#50187) 

* fix bug
* disable bilinear_interp_v2
* add verison check in  py UT
上级 fd679d31
隐藏空白更改
内联 并排
Showing with 173 addition and 42 deletion
paddle/fluid/inference/tensorrt/convert/slice_op.cc
浏览文件 @ f0422a28
...@@ -112,8 +112,8 @@ class SliceOpConverter : public OpConverter { ...@@ -112,8 +112,8 @@ class SliceOpConverter : public OpConverter {
} }
} }
// CI failed in trt 6015 but success in 7134, may be a trt bug // CI failed in trt 6015 but success in 7011/7134, may be a trt bug
#if IS_TRT_VERSION_GE(7134) #if IS_TRT_VERSION_GE(7000)
auto* size_tensor = auto* size_tensor =
Sub(Min(Concat(end_vec_tensor), shape_tensor), start_tensor); Sub(Min(Concat(end_vec_tensor), shape_tensor), start_tensor);
#else #else
......
paddle/fluid/inference/tensorrt/engine.cc
浏览文件 @ f0422a28
...@@ -199,8 +199,8 @@ void TensorRTEngine::FreezeNetwork() { ...@@ -199,8 +199,8 @@ void TensorRTEngine::FreezeNetwork() {
LOG(INFO) << "Run Paddle-TRT Dynamic Shape mode."; LOG(INFO) << "Run Paddle-TRT Dynamic Shape mode.";
for (int i = 0; i < max_profile_num_; i++) { for (int i = 0; i < max_profile_num_; i++) {
for (auto &input : min_input_shape_) { for (auto &input : min_input_shape_) {
#if IS_TRT_VERSION_LT(7000) #if IS_TRT_VERSION_LT(7100)
// trt6 will check all_of input > 0 // trt6/trt7011 will check all_of input > 0
if (!(std::all_of(input.second.begin(), if (!(std::all_of(input.second.begin(),
input.second.end(), input.second.end(),
[](int x) { return x > 0; }) && [](int x) { return x > 0; }) &&
......
paddle/fluid/inference/tensorrt/op_teller.cc
浏览文件 @ f0422a28
...@@ -792,6 +792,10 @@ struct SimpleOpTypeSetTeller : public Teller { ...@@ -792,6 +792,10 @@ struct SimpleOpTypeSetTeller : public Teller {
} }
if (op_type == "bilinear_interp_v2") { if (op_type == "bilinear_interp_v2") {
// trt 7011 result in test_solov2_trt_fp32.py TRT fp32 diff
#if IS_TRT_VERSION_LT(7100)
return false;
#endif
std::vector<std::string> attrs{"data_layout", std::vector<std::string> attrs{"data_layout",
"interp_method", "interp_method",
"align_corners", "align_corners",
......
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_bilinear_interp_v2.py
浏览文件 @ f0422a28
...@@ -22,7 +22,6 @@ import unittest ...@@ -22,7 +22,6 @@ import unittest
class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest): class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest):
def is_program_valid(self, program_config: ProgramConfig) -> bool: def is_program_valid(self, program_config: ProgramConfig) -> bool:
inputs = program_config.inputs inputs = program_config.inputs
weights = program_config.weights weights = program_config.weights
...@@ -33,13 +32,13 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest): ...@@ -33,13 +32,13 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest):
return True return True
def sample_program_configs(self): def sample_program_configs(self):
def generate_input1(attrs: List[Dict[str, Any]]): def generate_input1(attrs: List[Dict[str, Any]]):
return np.ones([1, 3, 64, 64]).astype(np.float32) return np.ones([1, 3, 64, 64]).astype(np.float32)
def generate_input2(attrs: List[Dict[str, Any]]): def generate_input2(attrs: List[Dict[str, Any]]):
return np.random.uniform(low=0.5, high=6.0, return np.random.uniform(low=0.5, high=6.0, size=(2)).astype(
size=(2)).astype("float32") "float32"
)
for data_layout in ["NCHW", "NHWC"]: for data_layout in ["NCHW", "NHWC"]:
for scale_y in [2.0, -1.0, 0.0]: for scale_y in [2.0, -1.0, 0.0]:
...@@ -47,48 +46,55 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest): ...@@ -47,48 +46,55 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest):
scale = [scale_y, scale_x] scale = [scale_y, scale_x]
for out_h in [32, 64, 128, 192]: for out_h in [32, 64, 128, 192]:
for out_w in [32, 64]: for out_w in [32, 64]:
dics = [{ dics = [
"data_layout": data_layout, {
"interp_method": "bilinear", "data_layout": data_layout,
"align_corners": False, "interp_method": "bilinear",
"align_mode": 0, "align_corners": False,
"scale": scale, "align_mode": 0,
"out_h": out_h, "scale": scale,
"out_w": out_w "out_h": out_h,
}] "out_w": out_w,
}
ops_config = [{ ]
"op_type": "bilinear_interp_v2",
"op_inputs": { ops_config = [
"X": ["input_data"], {
"Scale": ["input_scale"] "op_type": "bilinear_interp_v2",
}, "op_inputs": {
"op_outputs": { "X": ["input_data"],
"Out": ["bilinear_interp_v2_output_data"] "Scale": ["input_scale"],
}, },
"op_attrs": dics[0] "op_outputs": {
}] "Out": [
"bilinear_interp_v2_output_data"
]
},
"op_attrs": dics[0],
}
]
ops = self.generate_op_config(ops_config) ops = self.generate_op_config(ops_config)
program_config = ProgramConfig( program_config = ProgramConfig(
ops=ops, ops=ops,
weights={ weights={
"input_scale": "input_scale": TensorConfig(
TensorConfig( data_gen=partial(generate_input2, dics)
data_gen=partial(generate_input2, dics)) )
}, },
inputs={ inputs={
"input_data": "input_data": TensorConfig(
TensorConfig( data_gen=partial(generate_input1, dics)
data_gen=partial(generate_input1, dics)) )
}, },
outputs=["bilinear_interp_v2_output_data"]) outputs=["bilinear_interp_v2_output_data"],
)
yield program_config yield program_config
def sample_predictor_configs( def sample_predictor_configs(
self, program_config) -> (paddle_infer.Config, List[int], float): self, program_config
) -> (paddle_infer.Config, List[int], float):
def generate_dynamic_shape(attrs): def generate_dynamic_shape(attrs):
self.dynamic_shape.min_input_shape = {"input_data": [1, 3, 64, 64]} self.dynamic_shape.min_input_shape = {"input_data": [1, 3, 64, 64]}
self.dynamic_shape.max_input_shape = {"input_data": [4, 3, 64, 64]} self.dynamic_shape.max_input_shape = {"input_data": [4, 3, 64, 64]}
...@@ -100,6 +106,123 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest): ...@@ -100,6 +106,123 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest):
self.dynamic_shape.opt_input_shape = {} self.dynamic_shape.opt_input_shape = {}
def generate_trt_nodes_num(attrs, dynamic_shape): def generate_trt_nodes_num(attrs, dynamic_shape):
# Here is consistent with op_teller.cc
ver = paddle_infer.get_trt_compile_version()
if ver[0] * 1000 + ver[1] * 100 + ver[2] * 10 < 7100:
return 0, 3
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
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
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:
return True
def sample_program_configs(self):
def generate_input1(attrs: List[Dict[str, Any]]):
return np.random.random(attrs[0]['input_size']).astype(np.float32)
for data_layout in ["NCHW", "NHWC"]:
for input_size, scale, out_h, out_w in [
([1, 258, 40, 40], [], 24, 24),
([1, 258, 40, 40], [2.0, 2.0], -1, -1),
([1, 256, 160, 160], [0.5, 0.5], -1, -1),
([1, 258, 20, 20], [], 16, 16),
]:
dics = [
{
"data_layout": data_layout,
"interp_method": "bilinear",
"align_corners": False,
"align_mode": 0,
"scale": scale,
"out_h": out_h,
"out_w": out_w,
# below attrs are used just for unitest, not used by inference
"input_size": input_size,
}
]
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": attrs[0]['input_size']
}
self.dynamic_shape.max_input_shape = {
"input_data": [attrs[0]['input_size'][0] + 2]
+ attrs[0]['input_size'][1:4]
}
self.dynamic_shape.opt_input_shape = {
"input_data": attrs[0]['input_size']
}
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):
# Here is consistent with op_teller.cc
ver = paddle_infer.get_trt_compile_version()
if ver[0] * 1000 + ver[1] * 100 + ver[2] * 10 < 7100:
return 0, 3
return 1, 2 return 1, 2
attrs = [ attrs = [
...@@ -110,19 +233,23 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest): ...@@ -110,19 +233,23 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest):
clear_dynamic_shape() clear_dynamic_shape()
self.trt_param.precision = paddle_infer.PrecisionType.Float32 self.trt_param.precision = paddle_infer.PrecisionType.Float32
yield self.create_inference_config(), generate_trt_nodes_num( yield self.create_inference_config(), generate_trt_nodes_num(
attrs, False), 1e-5 attrs, False
), 1e-5
self.trt_param.precision = paddle_infer.PrecisionType.Half self.trt_param.precision = paddle_infer.PrecisionType.Half
yield self.create_inference_config(), generate_trt_nodes_num( yield self.create_inference_config(), generate_trt_nodes_num(
attrs, False), 1e-2 attrs, False
), 1e-2
# for dynamic_shape # for dynamic_shape
generate_dynamic_shape(attrs) generate_dynamic_shape(attrs)
self.trt_param.precision = paddle_infer.PrecisionType.Float32 self.trt_param.precision = paddle_infer.PrecisionType.Float32
yield self.create_inference_config(), generate_trt_nodes_num( yield self.create_inference_config(), generate_trt_nodes_num(
attrs, True), 1e-5 attrs, True
), 1e-5
self.trt_param.precision = paddle_infer.PrecisionType.Half self.trt_param.precision = paddle_infer.PrecisionType.Half
yield self.create_inference_config(), generate_trt_nodes_num( yield self.create_inference_config(), generate_trt_nodes_num(
attrs, True), 1e-2 attrs, True
), 1e-2
def test(self): def test(self):
self.run_test() self.run_test()
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或