2222
2323class TrtConvertConv2dTest (TrtLayerAutoScanTest ):
2424 def is_program_valid (self , program_config : ProgramConfig ) -> bool :
25- # TODO: This is just the example to remove the wrong attrs.
2625 inputs = program_config .inputs
2726 weights = program_config .weights
2827 attrs = [
2928 program_config .ops [i ].attrs
3029 for i in range (len (program_config .ops ))
3130 ]
3231
33- # groups restriction.
3432 if inputs ['input_data' ].shape [1 ] != weights ['conv2d_weight' ].shape [
3533 1 ] * attrs [0 ]['groups' ]:
3634 return False
3735
38- # others restriction, todo.
39-
4036 return True
4137
4238 def sample_program_configs (self ):
43- def generate_input1 ( attrs : List [ Dict [ str , Any ]]):
44- # TODO: This is just the example to illustrate the releation between axis and input.
45- # for each attr, can generate different datas
39+ self . trt_param . workspace_size = 1073741824
40+
41+ def generate_input1 ( batch , attrs : List [ Dict [ str , Any ]]):
4642 if attrs [0 ]['groups' ] == 1 :
47- return np .ones ([2 , 3 , 64 , 64 ]).astype (np .float32 )
43+ return np .ones ([batch , 3 , 64 , 64 ]).astype (np .float32 )
44+ elif attrs [0 ]['groups' ] == 2 :
45+ return np .ones ([batch , 6 , 64 , 64 ]).astype (np .float32 )
4846 else :
49- return np .ones ([1 , 3 , 64 , 64 ]).astype (np .float32 )
47+ return np .ones ([batch , 9 , 64 , 64 ]).astype (np .float32 )
5048
5149 def generate_weight1 (attrs : List [Dict [str , Any ]]):
5250 return np .random .random ([24 , 3 , 3 , 3 ]).astype (np .float32 )
5351
54- # for strides in [[1, 1], [2, 2], [1, 2], [2, 3]]:
55- # for paddings in [[0, 3], [3, 1], [1, 1, 1, 1]]:
56- # for groups in [1, 2]:
57- # for padding_algotithm in ['EXPLICIT', 'SAME', 'VALID']:
58- # for dilations in [[1, 1], [1, 2]]:
59- # for data_format in ['NCHW']:
60- for strides in [[1 , 1 ], [2 , 2 ]]:
61- for paddings in [[0 , 3 ], [3 , 1 ]]:
62- for groups in [1 ]:
63- for padding_algotithm in ['EXPLICIT' ]:
64- for dilations in [[1 , 1 ]]:
65- for data_format in ['NCHW' ]:
66-
67- dics = [{
68- "data_fromat" : data_format ,
69- "dilations" : dilations ,
70- "padding_algorithm" : padding_algotithm ,
71- "groups" : groups ,
72- "paddings" : paddings ,
73- "strides" : strides ,
74- "data_format" : data_format
75- }, {}]
76-
77- ops_config = [{
78- "op_type" : "conv2d" ,
79- "op_inputs" : {
80- "Input" : ["input_data" ],
81- "Filter" : ["conv2d_weight" ]
82- },
83- "op_outputs" : {
84- "Output" : ["conv_output_data" ]
85- },
86- "op_attrs" : dics [0 ]
87- }, {
88- "op_type" : "relu" ,
89- "op_inputs" : {
90- "X" : ["conv_output_data" ]
91- },
92- "op_outputs" : {
93- "Out" : ["relu_output_data" ]
94- },
95- "op_attrs" : dics [1 ]
96- }]
97- ops = self .generate_op_config (ops_config )
98-
99- program_config = ProgramConfig (
100- ops = ops ,
101- weights = {
102- "conv2d_weight" : TensorConfig (
103- data_gen = partial (generate_weight1 ,
104- dics ))
105- },
106- inputs = {
107- "input_data" : TensorConfig (
108- data_gen = partial (generate_input1 ,
109- dics ))
110- },
111- outputs = ["relu_output_data" ])
112-
113- yield program_config
52+ for batch in [1 , 2 , 4 ]:
53+ for strides in [[1 , 1 ], [2 , 2 ], [1 , 2 ]]:
54+ for paddings in [[0 , 3 ], [1 , 2 , 3 , 4 ]]:
55+ for groups in [1 , 2 , 3 ]:
56+ for padding_algorithm in ['EXPLICIT' , 'SAME' , 'VALID' ]:
57+ for dilations in [[1 , 1 ], [2 , 2 ], [1 , 2 ]]:
58+ for data_format in ['NCHW' ]:
59+
60+ dics = [{
61+ "data_fromat" : data_format ,
62+ "dilations" : dilations ,
63+ "padding_algorithm" : padding_algorithm ,
64+ "groups" : groups ,
65+ "paddings" : paddings ,
66+ "strides" : strides ,
67+ "data_format" : data_format
68+ }, {}]
69+
70+ if padding_algorithm == 'EXPLICIT' :
71+ ops_config = [{
72+ "op_type" : "conv2d" ,
73+ "op_inputs" : {
74+ "Input" : ["input_data" ],
75+ "Filter" : ["conv2d_weight" ]
76+ },
77+ "op_outputs" : {
78+ "Output" : ["conv_output_data" ]
79+ },
80+ "op_attrs" : dics [0 ]
81+ }, {
82+ "op_type" : "relu" ,
83+ "op_inputs" : {
84+ "X" : ["conv_output_data" ]
85+ },
86+ "op_outputs" : {
87+ "Out" : ["output_data" ]
88+ },
89+ "op_attrs" : dics [1 ]
90+ }]
91+ else :
92+ ops_config = [{
93+ "op_type" : "conv2d" ,
94+ "op_inputs" : {
95+ "Input" : ["input_data" ],
96+ "Filter" : ["conv2d_weight" ]
97+ },
98+ "op_outputs" : {
99+ "Output" : ["output_data" ]
100+ },
101+ "op_attrs" : dics [0 ]
102+ }]
103+ ops = self .generate_op_config (ops_config )
104+
105+ program_config = ProgramConfig (
106+ ops = ops ,
107+ weights = {
108+ "conv2d_weight" :
109+ TensorConfig (data_gen = partial (
110+ generate_weight1 , dics ))
111+ },
112+ inputs = {
113+ "input_data" :
114+ TensorConfig (data_gen = partial (
115+ generate_input1 , batch , dics ))
116+ },
117+ outputs = ["output_data" ])
118+
119+ yield program_config
114120
115121 def sample_predictor_configs (
116122 self , program_config ) -> (paddle_infer .Config , List [int ], float ):
117123 def generate_dynamic_shape (attrs ):
118- if len ( attrs [0 ]['paddings' ]) == 4 :
124+ if attrs [0 ]['groups' ] == 1 :
119125 self .dynamic_shape .min_input_shape = {
120126 "input_data" : [1 , 3 , 32 , 32 ],
121- '' : []
127+ "output_data" : [1 , 24 , 32 , 32 ]
122128 }
123129 self .dynamic_shape .max_input_shape = {
124130 "input_data" : [4 , 3 , 64 , 64 ],
125- '' : []
131+ "output_data" : [4 , 24 , 64 , 64 ]
126132 }
127133 self .dynamic_shape .opt_input_shape = {
128134 "input_data" : [1 , 3 , 64 , 64 ],
129- '' : []
135+ "output_data" : [1 , 24 , 64 , 64 ]
136+ }
137+ elif attrs [0 ]['groups' ] == 2 :
138+ self .dynamic_shape .min_input_shape = {
139+ "input_data" : [1 , 6 , 32 , 32 ],
140+ "output_data" : [1 , 24 , 32 , 32 ]
141+ }
142+ self .dynamic_shape .max_input_shape = {
143+ "input_data" : [4 , 6 , 64 , 64 ],
144+ "output_data" : [4 , 24 , 64 , 64 ]
145+ }
146+ self .dynamic_shape .opt_input_shape = {
147+ "input_data" : [1 , 6 , 64 , 64 ],
148+ "output_data" : [1 , 24 , 64 , 64 ]
130149 }
131150 else :
132151 self .dynamic_shape .min_input_shape = {
133- "input_data" : [1 , 3 , 32 , 32 ]
152+ "input_data" : [1 , 9 , 32 , 32 ],
153+ "output_data" : [1 , 24 , 32 , 32 ]
134154 }
135155 self .dynamic_shape .max_input_shape = {
136- "input_data" : [4 , 3 , 64 , 64 ]
156+ "input_data" : [4 , 9 , 64 , 64 ],
157+ "output_data" : [4 , 24 , 64 , 64 ]
137158 }
138159 self .dynamic_shape .opt_input_shape = {
139- "input_data" : [1 , 3 , 64 , 64 ]
160+ "input_data" : [1 , 9 , 64 , 64 ],
161+ "output_data" : [1 , 24 , 64 , 64 ]
140162 }
141163
142164 def clear_dynamic_shape ():
@@ -145,11 +167,7 @@ def clear_dynamic_shape():
145167 self .dynamic_shape .opt_input_shape = {}
146168
147169 def generate_trt_nodes_num (attrs , dynamic_shape ):
148- # TODO: This is just the example, need to be fixed.
149- if len (attrs [0 ]['paddings' ]) == 4 :
150- return 1 , 2
151- else :
152- return 1 , 2
170+ return 1 , 2
153171
154172 attrs = [
155173 program_config .ops [i ].attrs
@@ -169,6 +187,7 @@ def generate_trt_nodes_num(attrs, dynamic_shape):
169187 attrs , False ), (1e-5 , 1e-5 )
170188
171189 # for dynamic_shape
190+
172191 generate_dynamic_shape (attrs )
173192 self .trt_param .precision = paddle_infer .PrecisionType .Float32
174193 yield self .create_inference_config (), generate_trt_nodes_num (attrs ,
@@ -181,29 +200,18 @@ def generate_trt_nodes_num(attrs, dynamic_shape):
181200 attrs , True ), (1e-5 , 1e-5 )
182201
183202 def add_skip_trt_case (self ):
184- # TODO(wilber): This is just the example to illustrate the skip usage.
185203 def teller1 (program_config , predictor_config ):
186- if len (program_config .ops [0 ].attrs ['paddings' ]) == 4 :
204+ if program_config .ops [0 ].attrs [
205+ 'padding_algorithm' ] == "SAME" or program_config .ops [
206+ 0 ].attrs ['padding_algorithm' ] == "VALID" :
187207 return True
188208 return False
189209
190210 self .add_skip_case (
191211 teller1 , SkipReasons .TRT_NOT_IMPLEMENTED ,
192- "NOT Implemented: we need to add support in the future ....TODO, just for the example "
212+ "When padding_algorithm is 'SAME' or 'VALID', Trt dose not support. In this case, trt build error is caused by scale op. "
193213 )
194214
195- def teller2 (program_config , predictor_config ):
196- if (
197- program_config .ops [0 ].attrs ['dilations' ][0 ] == 1 and
198- program_config .ops [0 ].attrs ['dilations' ][0 ] == 2
199- ) or program_config .ops [0 ].attrs ['padding_algorithm' ] != 'EXPLICIT' :
200- return True
201- return False
202-
203- self .add_skip_case (teller2 , SkipReasons .TRT_NOT_SUPPORT ,
204- "TODO, just for the example" )
205- pass
206-
207215 def test (self ):
208216 self .add_skip_trt_case ()
209217 self .run_test ()
0 commit comments