@@ -108,24 +108,17 @@ def _run_test_case(self, output_names_with_port, feed_dict, **kwargs):
108108 kwargs ["constant_fold" ] = False
109109 return self .run_test_case (feed_dict , [], output_names_with_port , ** kwargs )
110110
111- def _test_expand_dims (self , idx ):
111+ def _test_expand_dims_known_rank (self , idx ):
112112 tf .reset_default_graph ()
113113 x_val = make_xval ([3 , 4 ])
114114 x = tf .placeholder (tf .float32 , shape = x_val .shape , name = _TFINPUT )
115115 op = tf .expand_dims (x , idx )
116116 _ = tf .identity (op , name = _TFOUTPUT )
117117 self ._run_test_case ([_OUTPUT ], {_INPUT : x_val })
118118
119- def test_expand_dims (self ):
119+ def test_expand_dims_known_rank (self ):
120120 for i in [- 1 , 0 , 1 , - 2 ]:
121- self ._test_expand_dims (i )
122-
123- def test_expand_dims_dynamic_inputs (self ):
124- x_val = make_xval ([3 , 4 ])
125- x = tf .placeholder (tf .float32 , shape = [None , None ], name = _TFINPUT )
126- op = tf .expand_dims (x , 0 )
127- _ = tf .identity (op , name = _TFOUTPUT )
128- self ._run_test_case ([_OUTPUT ], {_INPUT : x_val })
121+ self ._test_expand_dims_known_rank (i )
129122
130123 def test_expand_dims_one_unknown_rank (self ):
131124 tf .reset_default_graph ()
@@ -135,14 +128,18 @@ def test_expand_dims_one_unknown_rank(self):
135128 _ = tf .identity (op , name = _TFOUTPUT )
136129 self ._run_test_case ([_OUTPUT ], {_INPUT : x_val })
137130
138- def test_expand_dims_more_unknown_rank (self ):
131+ def _test_expand_dims_more_unknown_rank (self , idx ):
139132 tf .reset_default_graph ()
140133 x_val = make_xval ([3 , 4 ])
141134 x = tf .placeholder (tf .float32 , shape = [None , None ], name = _TFINPUT )
142- op = tf .expand_dims (x , 0 )
135+ op = tf .expand_dims (x , idx )
143136 _ = tf .identity (op , name = _TFOUTPUT )
144137 self ._run_test_case ([_OUTPUT ], {_INPUT : x_val })
145138
139+ def test_expand_dims_more_unknown_rank (self ):
140+ for i in [- 1 , 0 , 1 , - 2 ]:
141+ self ._test_expand_dims_more_unknown_rank (i )
142+
146143 @check_opset_min_version (9 , "ConstantOfShape" )
147144 def test_eye_non_const1 (self ):
148145 # tf.eye(num_rows), num_rows is not const here
@@ -444,7 +441,9 @@ def test_dropout(self):
444441 feed_dict = {"input_1:0" : x_val }
445442 input_names_with_port = ["input_1:0" ]
446443 output_names_with_port = ["output:0" ]
447- self .run_test_case (feed_dict , input_names_with_port , output_names_with_port )
444+ self .run_test_case (feed_dict , input_names_with_port , output_names_with_port ,
445+ graph_validator = lambda g : (check_op_count (g , "RandomUniform" , 0 ) and
446+ check_op_count (g , "RandomUniformLike" , 0 )))
448447
449448 def test_nn_dropout (self ):
450449 keep_prob = tf .placeholder_with_default (1. , (), "keep_prob" )
@@ -461,7 +460,10 @@ def test_nn_dropout(self):
461460 output_names_with_port = ["output:0" ]
462461 # when constant_fold is enabled, PlaceholderWithDefault will be folded into either a const or a placeholder.
463462 # here we set it False to test PlaceholderWithDefault bug: https://github.com/onnx/tensorflow-onnx/pull/446
464- self .run_test_case (feed_dict , input_names_with_port , output_names_with_port , constant_fold = False )
463+ # Dropout with ratio 1.0 will be optimized so that only one Identity is left
464+ self .run_test_case (feed_dict , input_names_with_port , output_names_with_port , constant_fold = False ,
465+ graph_validator = lambda g : (check_op_count (g , "RandomUniform" , 0 ) and
466+ check_op_count (g , "RandomUniformLike" , 0 )))
465467
466468 @check_tf_min_version ("1.13" )
467469 def test_nn_dropout_with_rate (self ):
@@ -477,7 +479,9 @@ def test_nn_dropout_with_rate(self):
477479 feed_dict = {"input_1:0" : x_val }
478480 input_names_with_port = ["input_1:0" ]
479481 output_names_with_port = ["output:0" ]
480- self .run_test_case (feed_dict , input_names_with_port , output_names_with_port , constant_fold = False )
482+ self .run_test_case (feed_dict , input_names_with_port , output_names_with_port , constant_fold = False ,
483+ graph_validator = lambda g : (check_op_count (g , "RandomUniform" , 0 ) and
484+ check_op_count (g , "RandomUniformLike" , 0 )))
481485
482486 def test_conv2d_with_input_transpose (self ):
483487 x_shape = [2 , 32 , 32 , 3 ]
@@ -925,6 +929,22 @@ def test_leaky_relu_int(self):
925929 self ._run_test_case ([_OUTPUT ], {_INPUT : x_val })
926930 tf .reset_default_graph ()
927931
932+ @skip_caffe2_backend ("fails on caffe2 with dim issue" )
933+ @check_onnxruntime_incompatibility ("Mul" )
934+ def test_leaky_relu_with_dependency (self ):
935+ x_val = 1000 * np .random .random_sample ([1000 , 100 ]).astype (np .float32 )
936+ x = tf .placeholder (x_val .dtype , [None ] * x_val .ndim , name = _TFINPUT )
937+ # simulate leaky_relu
938+ alpha = tf .constant (0.5 )
939+ y = alpha * x
940+ x_ = tf .maximum (y , x )
941+ dependency = y - 1
942+
943+ _ = tf .identity (x_ , name = _TFOUTPUT )
944+ _ = tf .identity (dependency , name = _TFOUTPUT1 )
945+ self ._run_test_case ([_OUTPUT , _OUTPUT1 ], {_INPUT : x_val })
946+ tf .reset_default_graph ()
947+
928948 @skip_caffe2_backend ("fails on caffe2 with dim issue" )
929949 @check_onnxruntime_incompatibility ("Mul" )
930950 def test_leaky_relu_float (self ):
@@ -1057,6 +1077,15 @@ def test_slice(self):
10571077 _ = tf .identity (x_ , name = _TFOUTPUT )
10581078 self ._run_test_case ([_OUTPUT ], {_INPUT : x_val })
10591079
1080+ def test_slice_neg_size (self ):
1081+ x_val = np .array ([[1 , 2 , 3 , 4 ], [5 , 6 , 7 , 8 ]], dtype = np .float32 )
1082+ t1 = tf .constant ([0 , 1 ], dtype = tf .int32 )
1083+ t2 = tf .constant ([- 1 , 2 ], dtype = tf .int32 )
1084+ x0 = tf .placeholder (tf .float32 , x_val .shape , name = _TFINPUT )
1085+ x_ = tf .slice (x0 , t1 , t2 )
1086+ _ = tf .identity (x_ , name = _TFOUTPUT )
1087+ self ._run_test_case ([_OUTPUT ], {_INPUT : x_val })
1088+
10601089 @check_opset_min_version (10 , "Slice in opset 10 can accept dymaic 'start' and 'ends'" )
10611090 def test_slice_with_non_const (self ):
10621091 x_val = np .array ([[1 , 2 , 3 , 4 ], [5 , 6 , 7 , 8 ]], dtype = np .float32 )
@@ -2237,7 +2266,6 @@ def test_sparse_softmax_cross_entropy_with_logits_large_class(self):
22372266
22382267 self ._run_test_case ([_OUTPUT ], {_INPUT : label_val , _INPUT1 : logits_val }, rtol = 1e-6 )
22392268
2240- @skip_onnxruntime_backend ("onnxruntime Slice did not supported BOOL" )
22412269 def test_matrix_band_part (self ):
22422270 input_val = np .random .randint (0 , 666 , (10 , 15 )).astype (np .int32 )
22432271 input_x = tf .placeholder (dtype = tf .int32 , shape = [None , None ], name = _TFINPUT )
@@ -2247,7 +2275,6 @@ def test_matrix_band_part(self):
22472275 _ = tf .identity (res1 , name = _TFOUTPUT1 )
22482276 self ._run_test_case ([_OUTPUT , _OUTPUT1 ], {_INPUT : input_val })
22492277
2250- @skip_onnxruntime_backend ("onnxruntime Slice did not supported BOOL." )
22512278 def test_matrix_band_part_2 (self ):
22522279 input_val = np .random .randint (0 , 666 , (1 , 1 )).astype (np .int32 )
22532280 input_x = tf .placeholder (dtype = tf .int32 , shape = [None , None ], name = _TFINPUT )
@@ -2416,7 +2443,7 @@ def test_softsign(self):
24162443
24172444 def test_batch_to_spacend (self ):
24182445 block_size = [2 , 2 ]
2419- crop = [[0 , 1 ], [2 , 1 ]]
2446+ crop = [[1 , 0 ], [2 , 1 ]]
24202447
24212448 input_val = np .random .random_sample ([40 , 3 , 5 , 100 ]).astype (np .float32 )
24222449 input_x = tf .placeholder (dtype = tf .float32 , shape = input_val .shape , name = _TFINPUT ) # NHWC
@@ -2551,6 +2578,80 @@ def test_selu(self):
25512578 _ = tf .identity (y , name = _TFOUTPUT )
25522579 self ._run_test_case ([_OUTPUT ], {_INPUT : x_val })
25532580
2581+ # test for gemm pattern0: alpha*A*B + beta*C
2582+ def test_gemm_pattern0 (self ):
2583+ max_number = 10
2584+ m = np .random .randint (max_number )
2585+ n = np .random .randint (max_number )
2586+ k = np .random .randint (max_number )
2587+ x_val1 = np .random .rand (m , n ).astype ("float32" )
2588+ x_val2 = np .random .rand (n , k ).astype ("float32" )
2589+ x_val3 = np .random .rand (m , k ).astype ("float32" )
2590+ a = tf .placeholder (tf .float32 , x_val1 .shape , name = _TFINPUT )
2591+ b = tf .placeholder (tf .float32 , x_val2 .shape , name = _TFINPUT1 )
2592+ c = tf .placeholder (tf .float32 , x_val3 .shape , name = _TFINPUT2 )
2593+ alpha = tf .constant (1.0 , dtype = tf .float32 )
2594+ beta = tf .constant (2.0 , dtype = tf .float32 )
2595+ mul1 = tf .multiply (alpha , tf .matmul (a , b ))
2596+ mul2 = tf .multiply (beta , c )
2597+ x_ = mul1 + mul2
2598+ _ = tf .identity (x_ , name = _TFOUTPUT )
2599+ self ._run_test_case ([_OUTPUT ], {_INPUT : x_val1 , _INPUT1 : x_val2 , _INPUT2 : x_val3 },
2600+ graph_validator = lambda g : check_op_count (g , "Gemm" , 1 ))
2601+
2602+ # test for gemm pattern1: alpha*A*B + C
2603+ def test_gemm_pattern1 (self ):
2604+ max_number = 10
2605+ m = np .random .randint (max_number )
2606+ n = np .random .randint (max_number )
2607+ k = np .random .randint (max_number )
2608+ x_val1 = np .random .rand (m , n ).astype ("float32" )
2609+ x_val2 = np .random .rand (n , k ).astype ("float32" )
2610+ x_val3 = np .random .rand (m , k ).astype ("float32" )
2611+ a = tf .placeholder (tf .float32 , x_val1 .shape , name = _TFINPUT )
2612+ b = tf .placeholder (tf .float32 , x_val2 .shape , name = _TFINPUT1 )
2613+ c = tf .placeholder (tf .float32 , x_val3 .shape , name = _TFINPUT2 )
2614+ alpha = tf .constant (1.0 , dtype = tf .float32 )
2615+ x_ = tf .multiply (alpha , tf .matmul (a , b )) + c
2616+ _ = tf .identity (x_ , name = _TFOUTPUT )
2617+ self ._run_test_case ([_OUTPUT ], {_INPUT : x_val1 , _INPUT1 : x_val2 , _INPUT2 : x_val3 },
2618+ graph_validator = lambda g : check_op_count (g , "Gemm" , 1 ))
2619+
2620+ # test for gemm pattern2: A*B + beta*C
2621+ def test_gemm_pattern2 (self ):
2622+ max_number = 10
2623+ m = np .random .randint (max_number )
2624+ n = np .random .randint (max_number )
2625+ k = np .random .randint (max_number )
2626+ x_val1 = np .random .rand (m , n ).astype ("float32" )
2627+ x_val2 = np .random .rand (n , k ).astype ("float32" )
2628+ x_val3 = np .random .rand (m , k ).astype ("float32" )
2629+ a = tf .placeholder (tf .float32 , x_val1 .shape , name = _TFINPUT )
2630+ b = tf .placeholder (tf .float32 , x_val2 .shape , name = _TFINPUT1 )
2631+ c = tf .placeholder (tf .float32 , x_val3 .shape , name = _TFINPUT2 )
2632+ beta = tf .constant (2.0 , dtype = tf .float32 )
2633+ x_ = tf .matmul (a , b ) + tf .multiply (beta , c )
2634+ _ = tf .identity (x_ , name = _TFOUTPUT )
2635+ self ._run_test_case ([_OUTPUT ], {_INPUT : x_val1 , _INPUT1 : x_val2 , _INPUT2 : x_val3 },
2636+ graph_validator = lambda g : check_op_count (g , "Gemm" , 1 ))
2637+
2638+ # test for gemm pattern3: A*B + C
2639+ def test_gemm_pattern3 (self ):
2640+ max_number = 10
2641+ m = np .random .randint (max_number )
2642+ n = np .random .randint (max_number )
2643+ k = np .random .randint (max_number )
2644+ x_val1 = np .random .rand (m , n ).astype ("float32" )
2645+ x_val2 = np .random .rand (n , k ).astype ("float32" )
2646+ x_val3 = np .random .rand (m , k ).astype ("float32" )
2647+ a = tf .placeholder (tf .float32 , x_val1 .shape , name = _TFINPUT )
2648+ b = tf .placeholder (tf .float32 , x_val2 .shape , name = _TFINPUT1 )
2649+ c = tf .placeholder (tf .float32 , x_val3 .shape , name = _TFINPUT2 )
2650+ x_ = tf .matmul (a , b ) + c
2651+ _ = tf .identity (x_ , name = _TFOUTPUT )
2652+ self ._run_test_case ([_OUTPUT ], {_INPUT : x_val1 , _INPUT1 : x_val2 , _INPUT2 : x_val3 },
2653+ graph_validator = lambda g : check_op_count (g , "Gemm" , 1 ))
2654+
25542655 def test_graph_matcher (self ):
25552656 shape = [2 , 6 ]
25562657 x_val = np .random .random (shape ).astype (np .float32 )
0 commit comments