Skip to content

Commit d167139

Browse files
JZZ-NOTEJZZ-NOTE
committed
modify affine_channel convert test
1 parent b440689 commit d167139

File tree

2 files changed

+92
-54
lines changed

2 files changed

+92
-54
lines changed

‎paddle/fluid/inference/tensorrt/op_teller.cc‎

Lines changed: 8 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -383,6 +383,14 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
383383
auto data_layout = framework::StringToDataLayout(
384384
BOOST_GET_CONST(std::string, desc.GetAttr("data_layout")));
385385
if (data_layout != framework::DataLayout::kNCHW) return false;
386+
387+
auto* block = desc.Block();
388+
auto x_var_name = desc.Input("X")[0];
389+
auto* x_var_desc = block->FindVar(x_var_name);
390+
const auto x_shape = x_var_desc->GetShape();
391+
if (x_shape.size() == 2) {
392+
return false;
393+
}
386394
}
387395

388396
if (op_type == "multiclass_nms") {

‎python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_affine_channel.py‎

Lines changed: 84 additions & 54 deletions
Original file line numberDiff line numberDiff line change
@@ -25,59 +25,94 @@ def is_program_valid(self, program_config: ProgramConfig) -> bool:
2525
return True
2626

2727
def sample_program_configs(self):
28-
def generate_input1(batch, attrs: List[Dict[str, Any]]):
29-
if attrs[0]['data_layout'] == "NCHW":
30-
return np.ones([batch, 3, 64, 64]).astype(np.float32)
28+
def generate_input1(batch, dims, attrs: List[Dict[str, Any]]):
29+
if dims == 2:
30+
return np.ones([batch, 64]).astype(np.float32)
3131
else:
32-
return np.ones([batch, 64, 64, 3]).astype(np.float32)
33-
34-
def generate_weight1(attrs: List[Dict[str, Any]]):
35-
return np.random.random([3]).astype(np.float32)
36-
37-
for batch in [1, 2, 4]:
38-
for data_layout in ["NCHW", "NHWC"]:
39-
dics = [{"data_layout": data_layout}]
40-
41-
ops_config = [{
42-
"op_type": "affine_channel",
43-
"op_inputs": {
44-
"X": ["input_data"],
45-
"Scale": ["scale"],
46-
"Bias": ["bias"]
47-
},
48-
"op_outputs": {
49-
"Out": ["output_data"]
50-
},
51-
"op_attrs": dics[0]
52-
}]
53-
ops = self.generate_op_config(ops_config)
54-
55-
program_config = ProgramConfig(
56-
ops=ops,
57-
weights={
58-
"scale": TensorConfig(data_gen=partial(generate_weight1,
59-
dics)),
60-
"bias": TensorConfig(data_gen=partial(generate_weight1,
61-
dics))
62-
},
63-
inputs={
64-
"input_data": TensorConfig(data_gen=partial(
65-
generate_input1, batch, dics))
66-
},
67-
outputs=["output_data"])
68-
69-
yield program_config
32+
if attrs[0]['data_layout'] == "NCHW":
33+
return np.ones([batch, 3, 64, 64]).astype(np.float32)
34+
else:
35+
return np.ones([batch, 64, 64, 3]).astype(np.float32)
36+
37+
def generate_weight1(dims, attrs: List[Dict[str, Any]]):
38+
if dims == 2:
39+
return np.random.random([64]).astype(np.float32)
40+
else:
41+
return np.random.random([3]).astype(np.float32)
42+
43+
for dims in [2, 4]:
44+
for batch in [1, 2, 4]:
45+
for data_layout in ["NCHW", "NHWC"]:
46+
47+
self.dims = dims
48+
dics = [{"data_layout": data_layout}]
49+
50+
ops_config = [{
51+
"op_type": "affine_channel",
52+
"op_inputs": {
53+
"X": ["input_data"],
54+
"Scale": ["scale"],
55+
"Bias": ["bias"]
56+
},
57+
"op_outputs": {
58+
"Out": ["output_data"]
59+
},
60+
"op_attrs": dics[0]
61+
}]
62+
ops = self.generate_op_config(ops_config)
63+
64+
program_config = ProgramConfig(
65+
ops=ops,
66+
weights={
67+
"scale": TensorConfig(data_gen=partial(
68+
generate_weight1, dims, dics)),
69+
"bias": TensorConfig(data_gen=partial(
70+
generate_weight1, dims, dics))
71+
},
72+
inputs={
73+
"input_data": TensorConfig(data_gen=partial(
74+
generate_input1, batch, dims, dics))
75+
},
76+
outputs=["output_data"])
77+
78+
yield program_config
7079

7180
def sample_predictor_configs(
7281
self, program_config) -> (paddle_infer.Config, List[int], float):
7382
def generate_dynamic_shape(attrs):
74-
self.dynamic_shape.min_input_shape = {"input_data": [1, 3, 32, 32]}
75-
self.dynamic_shape.max_input_shape = {"input_data": [4, 3, 64, 64]}
76-
self.dynamic_shape.opt_input_shape = {"input_data": [1, 3, 64, 64]}
83+
if self.dims == 2:
84+
self.dynamic_shape.min_input_shape = {"input_data": [1, 32]}
85+
self.dynamic_shape.max_input_shape = {"input_data": [4, 64]}
86+
self.dynamic_shape.opt_input_shape = {"input_data": [2, 64]}
87+
else:
88+
if attrs[0]['data_layout'] == "NCHW":
89+
self.dynamic_shape.min_input_shape = {
90+
"input_data": [1, 3, 32, 32]
91+
}
92+
self.dynamic_shape.max_input_shape = {
93+
"input_data": [4, 3, 64, 64]
94+
}
95+
self.dynamic_shape.opt_input_shape = {
96+
"input_data": [1, 3, 64, 64]
97+
}
98+
else:
99+
self.dynamic_shape.min_input_shape = {
100+
"input_data": [1, 32, 32, 3]
101+
}
102+
self.dynamic_shape.max_input_shape = {
103+
"input_data": [4, 64, 64, 3]
104+
}
105+
self.dynamic_shape.opt_input_shape = {
106+
"input_data": [1, 64, 64, 3]
107+
}
108+
109+
def clear_dynamic_shape():
110+
self.dynamic_shape.min_input_shape = {}
111+
self.dynamic_shape.max_input_shape = {}
112+
self.dynamic_shape.opt_input_shape = {}
77113

78114
def generate_trt_nodes_num(attrs, dynamic_shape):
79-
# TODO: This is just the example, need to be fixed.
80-
if attrs[0]['data_layout'] == "NCHW":
115+
if self.dims == 4 and attrs[0]['data_layout'] == "NCHW":
81116
return 1, 2
82117
else:
83118
return 0, 3
@@ -88,15 +123,13 @@ def generate_trt_nodes_num(attrs, dynamic_shape):
88123
]
89124

90125
# for static_shape
126+
clear_dynamic_shape()
91127
self.trt_param.precision = paddle_infer.PrecisionType.Float32
92128
yield self.create_inference_config(), generate_trt_nodes_num(
93129
attrs, False), 1e-5
94130
self.trt_param.precision = paddle_infer.PrecisionType.Half
95131
yield self.create_inference_config(), generate_trt_nodes_num(
96-
attrs, False), 1e-2
97-
self.trt_param.precision = paddle_infer.PrecisionType.Int8
98-
yield self.create_inference_config(), generate_trt_nodes_num(
99-
attrs, False), 1e-1
132+
attrs, False), 1e-5
100133

101134
# for dynamic_shape
102135
generate_dynamic_shape(attrs)
@@ -105,10 +138,7 @@ def generate_trt_nodes_num(attrs, dynamic_shape):
105138
True), 1e-5
106139
self.trt_param.precision = paddle_infer.PrecisionType.Half
107140
yield self.create_inference_config(), generate_trt_nodes_num(attrs,
108-
True), 1e-2
109-
self.trt_param.precision = paddle_infer.PrecisionType.Int8
110-
yield self.create_inference_config(), generate_trt_nodes_num(attrs,
111-
True), 1e-1
141+
True), 1e-5
112142

113143
def test(self):
114144
self.run_test()

0 commit comments

Comments
 (0)