├── BugDetails_IF.md ├── BugDetails_MCF.md └── BugDetails_DCF.md /BugDetails_IF.md: -------------------------------------------------------------------------------- 1 | We detail the IF bugs as follows. 2 | 3 | 4 | Inference Failure 5 | ================ 6 | 7 | ***IF-1 Reduceprod: The Reduceprod cannot output results. Inference aborted.*** 8 | ---------------- 9 | 10 | The Reduceprod cannot output results (on the left side of the structure ). This model contains 26 operators. 11 | 12 | ![image](https://user-images.githubusercontent.com/69624583/92420420-c1b53e80-f1a5-11ea-8c6c-37d64034fa32.png) 13 | ![image](https://user-images.githubusercontent.com/69624583/92420438-d98cc280-f1a5-11ea-8513-256922c35553.png) 14 | 15 | MNN Log: 16 | *** Error in `python': free(): invalid next size (fast): 0x0000000001e2cb90 *** 17 | ======= Backtrace: ========= 18 | /lib/x86_64-linux-gnu/libc.so.6(+0x777f5)[0x7f767300b7f5] 19 | /lib/x86_64-linux-gnu/libc.so.6(+0x8038a)[0x7f767301438a] 20 | /lib/x86_64-linux-gnu/libc.so.6(cfree+0x4c)[0x7f767301858c] 21 | /usr/local/lib/python2.7/dist-packages/_mnncengine.so(_ZN3MNN6TensorD1Ev+0x74)[0x7f7672028654] 22 | /usr/local/lib/python2.7/dist-packages/_mnncengine.so(+0x4f2a0)[0x7f7671fac2a0] 23 | 24 | #0 0x00007f0c6e11b438 in __GI_raise (sig=sig@entry=6) at ../sysdeps/unix/sysv/linux/raise.c:54 25 | 26 | #1 0x00007f0c6e11d03a in __GI_abort () at abort.c:89 27 | 28 | #2 0x00007f0c6e15d7fa in __libc_message (do_abort=do_abort@entry=2, fmt=fmt@entry=0x7f0c6e276f98 "*** Error in `%s': %s: 0x%s ***\n") at ../sysdeps/posix/libc_fatal.c:175 29 | 30 | #3 0x00007f0c6e16638a in malloc_printerr (ar_ptr=, ptr=, str=0x7f0c6e277010 "free(): invalid next size (fast)", action=3) at malloc.c:5020 31 | 32 | #4 _int_free (av=, p=, have_lock=0) at malloc.c:3874 33 | 34 | #5 0x00007f0c6e16a58c in __GI___libc_free (mem=) at malloc.c:2975 35 | 36 | #6 0x00007f0c6d17a654 in MNN::Tensor::~Tensor() () from /usr/local/lib/python2.7/dist-packages/_mnncengine.so 37 | 38 | #7 0x00007f0c6d0fe2a0 in PyMNNTensor_dealloc (self=0x7f0c6e86ab50) at /ruhuan/Projects/AliNNPrivate/pymnn/src/MNN.cc:1086 39 | 40 | 41 | 42 | 43 | ***IF-2 The shapes of the 4 output tensors of the model are 0 with a certain probability.*** 44 | ---------------- 45 | ![if_2_9_443247_0001](https://user-images.githubusercontent.com/69624583/92605466-56be5180-f2e4-11ea-8e29-f7ca1efb0b6e.png) 46 | 47 | The shapes of the 4 output tensors(mul_10100(left), mul_10800, reduceprod_10500,squeeze_10600) of the model are 0 with a certain probability. 48 | 49 | The shapes of these 4 output tensors in three inferences are listed as follows. And the results of three inferences are not correct. 50 | 51 | 52 | | Inference | mul_10100 | mul_10800 | reduceprod_10500 | squeeze_10600 | 53 | | ------------- |:-------------:|:----------:|:----------:| :----------: | 54 | | 1 | (0,0,0,0) | (0,0,0,0) |(0,0,0,0) | (0,0,0,0) | 55 | | 2 | (8,2,11,11) | (0,0,0,0) |(0,0,0,0) | (0,0,0,0) | 56 | | 3 | (8,2,11,11) | (0,0,0,0) | (11,) | (176, 11) | 57 | 58 | 59 | MNN log: 60 | Reshape error: 22 -> 2. 61 | due to the internal logic of MNN, if your MNN model doesn't have input shape, you may ignore this 'Resize error' information: 62 | ** Resize error for [Reshape], reshape_10800, code=3 ** 63 | it will work after you set the input tensor shape in MNN, and then resize the Session 64 | 65 | However, Slice's 'size' is (2,1,1,1). And Reshape's 'shape' is (2,1). That is to say, the number of input and output data of the reshape operator is 2. 66 | 67 | 68 | 69 | ***IF-3 The shape of the output tensor is 0.*** 70 | ---------------- 71 | 72 | ![if3_1_rn_23_133412_0001](https://user-images.githubusercontent.com/69624583/92605547-72295c80-f2e4-11ea-8c9d-fedf979d9e51.png) 73 | ![if3_2_rn_23_133412_0001](https://user-images.githubusercontent.com/69624583/92605554-75244d00-f2e4-11ea-98dd-f47dace3a03d.png) 74 | 75 | 76 | The shape of the output tensor is 0. 77 | 78 | 79 | 80 | 81 | 82 | 83 | ***IF-4 Split: When the num_or_size_splits parameter contains 0, the shape of the output tensor is 0.*** 84 | ---------------- 85 | 86 | 87 | When the num_or_size_splits parameter contains 0, the Split operator will yield a calculation error. 88 | 89 | The configuration of the Split operator is as follows. 90 | 91 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="placeholder_10000") 92 | 93 | out = tf.split(x,[0,3],axis = 2,name = "split" ) 94 | 95 | 96 | 97 | 98 | 99 | 100 | 101 | -------------------------------------------------------------------------------- /BugDetails_MCF.md: -------------------------------------------------------------------------------- 1 | We detail these bugs as follows. 2 | 3 | 4 | 5 | **Model Conversion Failure** 6 | ================ 7 | 8 | 9 | ***MCF-1 TopKV2: Segmentation fault in TopKV2 conversion.*** 10 | ---------------- 11 | 12 | MNN model cannot be generated. Almost all models that include TopKV2 operators cannot be converted and inferred successfully. 13 | 14 | MNN Log: Error for 18. Segmentation fault (core dumped). 15 | 16 | (gdb) bt 17 | 18 | #0 0x00007f2cc088c6e5 in MNN::TopKV2SizeComputer::onComputeSize(MNN::Op const*, std::vector > const&, std::vector > const&) const () from /usr/local/lib/python2.7/dist-packages/_tools.so 19 | 20 | #1 0x00007f2cc093e84b in MNN::Express::Executor::computeInfo(MNN::Express::Expr*) () from /usr/local/lib/python2.7/dist-packages/_tools.so 21 | 22 | 23 | ***MCF-2 Deconv： Conversion aborted.*** 24 | ---------------- 25 | ![mcf3fusion_fp16_randomnet_ws_2_832384_0001](https://user-images.githubusercontent.com/69624583/92605031-ce3fb100-f2e3-11ea-857e-44e21507ac68.png) 26 | 27 | The Deconv API of the pb model is tf.nn.conv2dtranspose. 28 | 29 | MNN log: /converter/source/common/writeFb.cpp:108 Check failed: (notSupportOps.size()) == (0). Not Support: Tensorflow::Conv2DBackpropInput, output\_shape is not consistent with inferred output shape in MNN. (height, width): (38,102) vs (4,21). Convert Tensorflow's Op deconv2d\_outputdata\_10100, type = Conv2DBackpropInput, failed. Aborted (core dumped). 30 | 31 | #0 0x00007f9ed925e438 in __GI_raise (sig=sig@entry=6) at ../sysdeps/unix/sysv/linux/raise.c:54 32 | 33 | #1 0x00007f9ed926003a in __GI_abort () at abort.c:89 34 | 35 | #2 0x00007f9ed7d2884d in __gnu_cxx::__verbose_terminate_handler() () from /usr/lib/x86_64-linux-gnu/libstdc++.so.6 36 | 37 | #3 0x00007f9ed7d266b6 in ?? () from /usr/lib/x86_64-linux-gnu/libstdc++.so.6 38 | 39 | #4 0x00007f9ed7d26701 in std::terminate() () from /usr/lib/x86_64-linux-gnu/libstdc++.so.6 40 | 41 | #5 0x00007f9ed7d26919 in __cxa_throw () from /usr/lib/x86_64-linux-gnu/libstdc++.so.6 42 | 43 | #6 0x00007f9e34fdcf9c in writeFb(std::unique_ptr >&, std::string const&, bool, bool) [clone .cold.981] () 44 | from /usr/local/lib/python2.7/dist-packages/_tools.so 45 | #7 0x00007f9e3500ff59 in PyTool_Converter(_object*, _object*) () at /ruhuan/Projects/AliNNPrivate/pymnn/src/MNNTools.cc:62 46 | 47 | 48 | 49 | 50 | ***MCF-3 A model with 10+ operators： Conversion aborted.*** 51 | ---------------- 52 | ![8](https://user-images.githubusercontent.com/69624583/92605134-ef080680-f2e3-11ea-8987-a2da9e408f3a.jpg) 53 | 54 | MNN log: 55 | 56 | *** Error in `/usr/bin/python': free(): invalid pointer: 0x0000000001db1940 *** 57 | ======= Backtrace: ========= 58 | /lib/x86_64-linux-gnu/libc.so.6(+0x777f5)[0x7f0ed0ec27f5] 59 | /lib/x86_64-linux-gnu/libc.so.6(+0x8038a)[0x7f0ed0ecb38a] 60 | /lib/x86_64-linux-gnu/libc.so.6(cfree+0x4c)[0x7f0ed0ecf58c] 61 | /usr/local/lib/python2.7/dist-packages/_tools.so(_ZNSt15_Sp_counted_ptrIPN3MNN7Express4Expr6InsideELN9__gnu_cxx12_Lock_policyE2EE10_M_disposeEv+0x7e)[0x7 f0e2cd7cdce] 62 | /usr/local/lib/python2.7/dist-packages/_tools.so(_ZNSt16_Sp_counted_baseILN9__gnu_cxx12_Lock_policyE2EE10_M_releaseEv+0x47)[0x7f0e2cc46347] 63 | /usr/local/lib/python2.7/dist-packages/_tools.so(_ZN3MNN7Express4ExprD1Ev+0x25)[0x7f0e2cd7ca95] 64 | /usr/local/lib/python2.7/dist-packages/_tools.so(_ZNSt15_Sp_counted_ptrIPN3MNN7Express4ExprELN9__gnu_cxx12_Lock_policyE2EE10_M_disposeEv+0x12)[0x7f0e2cd7 cd32] 65 | 66 | #0 0x00007fd945390438 in __GI_raise (sig=sig@entry=6) at ../sysdeps/unix/sysv/linux/raise.c:54 67 | 68 | #1 0x00007fd94539203a in __GI_abort () at abort.c:89 69 | 70 | #2 0x00007fd9453d27fa in __libc_message (do_abort=do_abort@entry=2, fmt=fmt@entry=0x7fd9454ebf98 "*** Error in `%s': %s: 0x%s ***\n") at ../sysdeps/posix/libc_fatal.c:175 71 | 72 | #3 0x00007fd9453db38a in malloc_printerr (ar_ptr=, ptr=, str=0x7fd9454e8d6f "free(): invalid pointer", action=3) at malloc.c:5020 73 | 74 | #4 _int_free (av=, p=, have_lock=0) at malloc.c:3874 75 | 76 | #5 0x00007fd9453df58c in __GI___libc_free (mem=) at malloc.c:2975 77 | 78 | #6 0x00007fd8a11454a6 in std::string::_Rep::_M_dispose (__a=..., this=) at /opt/rh/devtoolset-8/root/usr/include/c++/8/ext/atomicity.h:82 79 | 80 | #7 std::string::_Rep::_M_dispose (__a=..., this=) at /opt/rh/devtoolset-8/root/usr/include/c++/8/bits/basic_string.h:3250 81 | 82 | #8 std::basic_string, std::allocator >::~basic_string (this=0x12e9320, __in_chrg=) 83 | at /opt/rh/devtoolset-8/root/usr/include/c++/8/bits/basic_string.h:3640 84 | 85 | #9 MNN::AttributeT::~AttributeT (this=0x12e9310, __in_chrg=) at /ruhuan/Projects/AliNNPrivate/schema/current/Tensor_generated.h:378 86 | 87 | #10 std::default_delete::operator() (this=, __ptr=0x12e9310) at /opt/rh/devtoolset-8/root/usr/include/c++/8/bits/unique_ptr.h:81 88 | 89 | #11 std::default_delete::operator() (this=0x12e9648, __ptr=0x12e9310) at /opt/rh/devtoolset-8/root/usr/include/c++/8/bits/unique_ptr.h:75 90 | 91 | #12 std::unique_ptr >::~unique_ptr (this=0x12e9648, __in_chrg=) at /opt/rh/devtoolset-8/root/usr/include/c++/8/bits/unique_ptr.h:274 92 | 93 | #13 std::_Destroy > > (__pointer=0x12e9648) at /opt/rh/devtoolset-8/root/usr/include/c++/8/bits/stl_construct.h:98 94 | 95 | #14 std::_Destroy_aux::__destroy >*> (__last=, __first=0x12e9648) 96 | 97 | 98 | ***MCF-4 A model with 5 operators： Conversion aborted.*** 99 | ---------------- 100 | 101 | ![10_mcf](https://user-images.githubusercontent.com/69624583/92605209-08a94e00-f2e4-11ea-9042-4e018fbef622.png) 102 | 103 | MNN log: 104 | Start to Optimize the MNN Net... 105 | [22:08:40] :20: Inputs: placeholder_10000 106 | [22:08:40] :37: Outputs: reducesum_outputdata_10400, Type = Reduction 107 | 108 | #0 _int_malloc (av=av@entry=0x7f1376b8ab20 , bytes=bytes@entry=4097) at malloc.c:3619 109 | 110 | #1 0x00007f137684a1d4 in __GI___libc_malloc (bytes=4097) at malloc.c:2920 111 | 112 | #2 0x00000000004a41bf in PyImport_ImportModuleLevel () 113 | 114 | #3 0x00000000004a5ce4 in ?? () 115 | 116 | #4 0x00000000004a59e5 in PyObject_CallFunction () 117 | 118 | #5 0x00000000004a3d27 in PyImport_Import () 119 | 120 | #6 0x00007f13757dec8e in importName (symbol=0x7f1375a1777f "Tensor", name=0x7f1375a16a3b "MNN") at /ruhuan/Projects/AliNNPrivate/pymnn/src/MNN.cc:50 121 | 122 | #7 0x00007f13757e2912 in PyMNNInterpreter_getSessionOutputAll(PyMNNInterpreter*, _object*) () at /ruhuan/Projects/AliNNPrivate/pymnn/src/MNN.cc:904 123 | 124 | #8 0x00000000004bc9ba in PyEval_EvalFrameEx () 125 | 126 | 127 | ***MCF-5： StopGradient： Model Generation Failure In MNN.*** 128 | ---------------- 129 | No error was reported when converting models that include unsupported StopGradient operators. The conversion process did not intercept an error and was aborted, and an empty model with a size of 0 was generated. 130 | MNN log: Model has no oplist. 131 | 132 | 133 | ***Concat： Model Generation Failure In TensorFlow.*** 134 | ---------------- 135 | It is worthy of mentioning that some exceptions of TensorFlow are found in model generation. When generating a model containing two Concats whose two inputs come from the same two constants, TensorFlow will get stuck. 136 | 137 | The configuration of the Concats is as follows. 138 | 139 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="input_1") 140 | 141 | y=tf.constant(y_data,dtype=tf.float32, shape=(n,h,w,c),name="constant_1") 142 | 143 | m=tf.constant(m_data,dtype=tf.float32, shape=(n,h,w,c),name="constant_3") 144 | 145 | c_1= tf.concat([x,y,m], axis=2, name="concat") 146 | 147 | out = tf.concat([c_1,y,m], axis=2, name="concat_1") 148 | 149 | -------------------------------------------------------------------------------- /BugDetails_DCF.md: -------------------------------------------------------------------------------- 1 | We detail the DCF bugs as follows. 2 | 3 | The threshold for the comparison of MNN and TensorFlow is that numbers with a relative error greater than 0.1% and less than 0.1% of the total data. 4 | Formally, let _RE(mnn)_ be the ratio of the numbers with the relative error between MNN and TensorFlow over the total output data of an operator. 5 | If _RE_ of a result is greater than 99.9%, when _RE(mnn) >= 99.9%_, the result is considered as a success. 6 | 7 | 8 | Output Comparison Failure (Inconsistencies) 9 | ============== 10 | 11 | 12 | ***DCF-1 Gather. core dumped and data comparison failure on MNN X86CPU and TensorFlow .*** 13 | --------------- 14 | ![dcfgather1fusion_fp16_randomnet_rn_5_540172_0001](https://user-images.githubusercontent.com/69624583/92605943-f24fc200-f2e4-11ea-86f1-10842e7e15b3.png) 15 | 16 | _RE(mnn_X86CPU)_ of Sub1 is 52.08%, _RE(mnn_X86CPU)_ of Sub2 is 76.04\%. 17 | MNN log: Error in `python': double free or corruption (!prev): 0x0000000002 4ae8b0. Backtrace: /dist-packages/\_mnncengine.so (\_ZN3MNN15BufferAllocator4NodeD1Ev+0x88) [0x7ff0e81 7a098]. Aborted (core dumped). 18 | 19 | #0 0x00007ff418033438 in __GI_raise (sig=sig@entry=6) at ../sysdeps/unix/sysv/linux/raise.c:54 20 | 21 | #1 0x00007ff41803503a in __GI_abort () at abort.c:89 22 | 23 | #2 0x00007ff4180757fa in __libc_message (do_abort=do_abort@entry=2, fmt=fmt@entry=0x7ff41818ef98 "*** Error in `%s': %s: 0x%s ***\n") at ../sysdeps/posix/libc_fatal.c:175 24 | 25 | #3 0x00007ff41807e38a in malloc_printerr (ar_ptr=, ptr=, str=0x7ff41818f0c8 "double free or corruption (!prev)", action=3) at malloc.c:5020 26 | 27 | #4 _int_free (av=, p=, have_lock=0) at malloc.c:3874 28 | 29 | #5 0x00007ff41808258c in __GI___libc_free (mem=) at malloc.c:2975 30 | 31 | #6 0x00007ff417070098 in MNN::BufferAllocator::Node::~Node() () from /usr/local/lib/python2.7/dist-packages/_mnncengine.so 32 | 33 | #7 0x00007ff4170700c2 in std::_Sp_counted_ptr::_M_dispose() () 34 | from /usr/local/lib/python2.7/dist-packages/_mnncengine.so 35 | 36 | #8 0x00007ff417070b02 in std::_Rb_tree >, std::_Select1st > >, std::less, std::allocator > > >::_M_erase(std::_Rb_tree_node > >*) () from /usr/local/lib/python2.7/dist-packages/_mnncengine.so 37 | 38 | 39 | 40 | ***DCF-2 Sigmoid. NaN: data comparison failure on MNN X86CPU and TensorFlow.*** 41 | --------------- 42 | When the input of a sigmoid operator is nan (coming from Rsqrt), the results are nan in TensorFlow and 1 in MNN X86 CPU respectively. 43 | 44 | 45 | 46 | ***DCF-3 Sigmoid. -inf: data comparison failure on MNN X86CPU and TensorFlow.*** 47 | --------------- 48 | When the input of a sigmoid operator is -inf (coming from Rsqrt), the results are 0 in TensorFlow and 1.64581e-38 in MNN X86 CPU respectively. 49 | 50 | 51 | 52 | 53 | 54 | ***DCF-4 Cast. Data comparison failure on MNN X86CPU and TensorFlow.*** 55 | --------------- 56 | 57 | The model contains two operators connected in sequence. The first Cast operator is configured as fp32 to int8, and the second is configured as int8 to uint8. The cast operator will yield a calculation error. 58 | 59 | The input tensor: 0.40432 0.117169 0.962151 0.183967 0.49692 0.0275081 0.909009 0.604157 0.493754 0.897677 0.535287 0.429582 0.266089 0.458383 0.997908 0.868963 60 | 0.232218 0.0356807 0.056058 0.476565 0.593033 0.700972 0.781927 0.485121 61 | 62 | The output tensor of TF: 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 . The output tensor of mnn X86 CPU: 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1. 63 | 64 | _RE(mnn_X86CPU)_ of Cast is 0%. 65 | 66 | 67 | 68 | 69 | ***DCF-5 LRN. Data comparison failure on MNN X86CPU and TensorFlow.*** 70 | --------------- 71 | The configuration of the LRN operator is as follows. 72 | 73 | alpha=1 beta=0.5 bias = 1 depth_raduis=5 74 | 75 | x_data =(np.ones((1,1,1,1))*2).astype(np.float32) 76 | 77 | x_data.tofile(name + "_inputdata_10000_0_fp32.bin") 78 | 79 | out = tf.nn.local_response_normalization(x,name="lrn") 80 | 81 | The output tensor of TF is 0.894427. The output tensor of mnn X86 CPU is 1.712698. 82 | 83 | 84 | 85 | 86 | 87 | ***DCF-6 Realdiv. Data comparison failure on MNN X86CPU and TensorFlow.*** 88 | --------------- 89 | ![dcffusion_fp16_randomnet_rn_3_832028_0001](https://user-images.githubusercontent.com/69624583/92606037-13b0ae00-f2e5-11ea-9567-db98bcdba823.png) 90 | 91 | The results of dividing by zero (coming from Pad ) in Realdiv are inf in TensorFlow and nan in MNN respectively. 92 | 93 | 94 | The x of Realdiv: array([ 1.2786534 , 1.0426922 , -0.11910683, 1.5758498 , 1.4206867 , 1.6706531 , 1.0834675 , 0.25468162, 0.09891693, 1.4739399 , 0.32884791, 0.11667726, 1.8873491 , 0.5553351 , 1.5132152 , 1.5450279 , 0.5409553 , 1.2378874 , 1.78609 , 0.61828995, 1.5263603 , 0.7822404 , 0.9433871 , 1.5502377 , 0.0717475 , 1.421492 , 1.7305765 , 1.0178283 , 1.8812672 , 1.0458821 , 0.52104974, 1.1909834 , 0.50230515, 1.6570765 , 1.3553871 , 0.9626574 , 0.25978124, 1.4960558 , 0.4294046 , 1.5899268 , -0.08192477, -0.22359428, 0.73022735, -0.20688827, 1.5729395 , 1.7708656 , 0.5167163 , 0.17279305, -0.18313001, 0.69207156, 1.8150818 , 1.3760335 , 0.63115156, 0.198356 , 0.80806834, 0.8461137 , 0.396619 , 0.95837283, 1.126148 , -0.00151169, 1.55765 , 0.43355164, 1.1452124 , 1.7827791 , 0.06954476, 0.6177971 , 0.6698139 , 0.5043217 , 1.2868638 , -0.21163152, 1.8892969 , 0.80474234], dtype=float32) 95 | 96 | The y of Realdiv: array([1.5803587, 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 1.5677226, 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 1.574414 , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. ], dtype=float32) 97 | 98 | The output tensor of TF: 99 | array([0.8090906 , inf, -inf, inf, inf, inf, inf,inf, inf, inf, inf, inf, 1.2038796 , inf, inf, inf, inf, inf, inf,inf, 100 | inf, inf, inf, inf,0.04557092, inf, inf, inf, inf, inf, inf, inf, inf, inf, inf, inf, inf, inf, inf, inf, 101 | -inf, -inf, inf, inf, inf, inf, inf, inf, inf, inf, inf, inf, inf, inf, inf, inf, inf, inf, inf,-inf, 102 | inf, inf, inf, inf, inf, inf, inf, inf, inf, -inf, inf, inf], dtype=float32) 103 | 104 | The output tensor of mnn X86 CPU: 105 | array([0.8347643 , nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, 1.201485 , nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,0.04603206, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan], dtype=float32) 106 | 107 | 108 | ***DCF-7 Concat. Data comparison failure on MNN X86CPU and TensorFlow.*** 109 | --------------- 110 | 111 | When the number of inputs is more than 2, the Concat operator will yield a calculation error. 112 | 113 | The values of the input and output tensors of the Concat operator are as follows. 114 | 115 | The input tensor1: 3.60285 9.53354 116 | 117 | The input tensor2: 8.67775 2.24457 118 | 119 | The input tensor3: 1.38196 2.91937 120 | 121 | The output tensor of TF: 3.602852 9.533535 8.677751 2.244566 1.381964 2.919371. The output tensor of mnn X86 CPU: 1.381964 2.919371 8.677751 2.244566 1.381964 2.919371. 122 | 123 | No errors are reported in the log of the Concat operator. The reference guide of MNN also does not indicate that the number of inputs of the Concat operator does not support more than 2. 124 | 125 | 126 | 127 | 128 | ***DCF-8 Deconv. Data comparison failure on MNN X86CPU and TensorFlow.*** 129 | --------------- 130 | The shape of the output tensor of Tf and mnn is inconsistent. 131 | The configuration of the Deconv operator is as follows. 132 | n = 10, h = 1, w =3, c = 3 133 | 134 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="placeholder") 135 | 136 | filter_data = np.random.uniform(-1, 1, size=[15,11,3,3]).astype(np.float32) * 0.3 # HWCK 137 | 138 | c = tf.nn.conv2d(x, filter_data,strides=[1,2,1,1],padding='SAME',dilations=[1,1,1,1],name="conv2d") 139 | 140 | filter_data = np.random.uniform(-1, 1, size=[1,3,3,3]).astype(np.float32) * 0.3 # HWCK 141 | 142 | out = tf.nn.conv2d_transpose(c,filter_data, strides=[1,2,2,1],output_shape=[10,2,7,3], padding="VALID",name="deconv") 143 | 144 | The shape of the output tensor of Tf and mnn is inconsistent. The output tensor shape of TF: [10, 2, 7,3 ] (NHWC). The output tensor shape of MNN X86 CPU: [10,3,1,7]. 145 | 146 | 147 | 148 | 149 | ***DCF-9 Argmax. Data comparison failure on MNN X86CPU and TensorFlow.*** 150 | --------------- 151 | The configuration of the Argmax operator is as follows. 152 | 153 | n = 3, h = 1, w = 1, c = 1, argmax/dimension = 0 154 | 155 | The values of the input and output tensors of the Argmax operator are as follows. 156 | 157 | The input tensor: [[[[3.5494123]]] [[[1.2762802]]] [[[5.062324 ]]]] 158 | 159 | The output tensor of TF: 2. The output tensor of mnn X86 CPU: 1. 160 | 161 | 162 | 163 | 164 | ***DCF-10 Depthwith. Data comparison failure on MNN X86CPU and TensorFlow.*** 165 | --------------- 166 | Data comparison failure on MNN X86CPU and TensorFlow. The configuration of the Depthwith operator is as follows. 167 | 168 | n = 2, h = 2, w =2, *c =1* ,padding = 'VALID' 169 | 170 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="placeholder") 171 | 172 | filter = np.random.uniform(-1, 1, size=[1,1,1,1]).astype(np.float32) #hwcn 173 | 174 | out = tf.nn.depthwise_conv2d(x,filter,strides = [1,1,1,1],padding = *'VALID'*,rate = [2,2],name = "depthwise") 175 | 176 | The values of the input and output tensors of the Depthwith operator are as follows. 177 | 178 | The input tensor: 1.89075 1.01571 1.24118 1.15506 1.58546 1.69911 1.12812 1.8123 179 | 180 | The output tensor of TF: 181 | -0.426098 182 | -0.228900 183 | -0.279712 184 | -0.260302 185 | -0.357297 186 | -0.382909 187 | -0.254232 188 | -0.408417 189 | . 190 | 191 | The output tensor of mnn X86 CPU: 192 | -0.426098 193 | -0.357297 194 | -0.228900 195 | -0.382909 196 | -0.279712 197 | -0.254232 198 | -0.260302 199 | -0.408417 200 | . 201 | _RE(mnn_X86CPU)_ of Depthwith is 37.50%. 202 | 203 | 204 | 205 | 206 | ***DCF-11 Depthwith. Data comparison failure on MNN X86CPU and TensorFlow.*** 207 | --------------- 208 | The shape of the output tensor of Tf and mnn is inconsistent. 209 | The configuration of the Depthwith operator is as follows. 210 | 211 | n = 2，h = 2，w =2，c =1 212 | 213 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="placeholder") 214 | 215 | filter = np.random.uniform(-1, 1, size=[1,1,1,2]).astype(np.float32)#hwcn 216 | 217 | out = tf.nn.depthwise_conv2d(x,filter,strides = [1,1,1,1],padding = 'VALID',rate = [2,2],name = "depthwise") 218 | 219 | The input tensor: 1.45974 1.36636 1.07671 1.97315 1.79221 1.52585 1.2357 1.98181. 220 | The input tensor shape is [2,2,2,1]. 221 | 222 | The output tensor of TF: 0.0435238 0.915834 0.0407397 0.85725 0.0321035 0.675526 0.0588319 1.23795 0.0534368 1.12442 0.0454951 0.957314 0.0368439 0.775274 0.0590901 1.24338 223 | The output tensor shape of TF is [2,2,2,2]. 224 | 225 | The output tensor of mnn X86 CPU: 0.0435238 0.0534368 0.0407397 0.0454951 0.0321035 0.0368439 0.0588319 0.0590901 226 | The output tensor shape of MNN X86 CPU is [8,1,1,1]. 227 | 228 | The shape of the output tensor of Tf and mnn is inconsistent. 229 | 230 | 231 | 232 | ***DCF-12 Deconv. Data comparison failure on MNN X86CPU and TensorFlow.*** 233 | --------------- 234 | The configuration of the Deconv operator is as follows. 235 | 236 | n = 1，h = 1，w = 3，c = 3 237 | 238 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="placeholder_10000") 239 | x_data =(np.ones((n,h,w,c))).astype(np.float32) 240 | 241 | filter_data = (np.ones((1,3,3,3))).astype(np.float32) # HWCK 242 | out_2 = tf.nn.conv2d_transpose(x,filter_data, strides=[1,1,1,1],output_shape=[1,1,3,3], padding="SAME",name="deconv") 243 | 244 | 245 | The input tensor:1 1 1 1 1 1 1 1 1 246 | 247 | The output tensor of TF: 6 6 6 9 9 9 6 6 6 248 | 249 | The output tensor of mnn X86 CPU: 3 3 3 6 6 6 9 9 9 250 | 251 | The shape of the output tensor of Tf and MNN X86 CPU is inconsistent. The output tensor shape of TF: [1,1,3,3](NHWC). The output tensor shape of MNN X86 CPU: [1, 3, 1, 3]. 252 | 253 | _RE(mnn_X86CPU)_ of Deconv is 0%. 254 | 255 | 256 | 257 | ***DCF-13 Reducemin. Data comparison failure on MNN X86CPU and TensorFlow.*** 258 | --------------- 259 | The cause of this inconsistency should be the calculation error of TensorFlow. 260 | 261 | The configuration of the Reducemin operator is as follows. 262 | 263 | n = 1，h = 2，w =3，c = 4 264 | 265 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="input_1") 266 | 267 | x_data = (np.ones((n,h,w,c))*16).astype(np.float32) 268 | 269 | a = tf.math.sqrt(x,name="sqrt_1") 270 | 271 | res_reducemin_0= tf.reduce_min(a,axis = None,keep_dims = True,name = "redecemin_0") 272 | 273 | 274 | The output tensor of TF: 16.0 275 | 276 | The output tensor of mnn X86 CPU: 4.0 277 | 278 | If an Add operator is added after the Reducemin operator to add 0 to the result of the reducemin operator, then the inference result of tf is correct(4.0). 279 | 280 | And thus, the cause of this inconsistency should be the calculation error of TensorFlow. 281 | 282 | 283 | ***DCF-14 Batchnorm. Data comparison failure on MNN ARMCPU and TensorFlow.*** 284 | --------------- 285 | The configuration of the Batchnorm operator is as follows. 286 | 287 | n = 1，h = 2，w = 3，c = 4 288 | 289 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="placeholder_10000") 290 | 291 | scale = [0,0.6856415271759033,0.6856415271759033,0.6856415271759033] 292 | 293 | offset = [0.6856415271759033,0.6856415271759033,0.6856415271759033,0.6856415271759033] 294 | 295 | out = tf.nn.fused_batch_norm(x,scale = scale,offset = offset,mean=scale,variance=scale,epsilon=0.001,data_format='NHWC',is_training=False,name="bn")[0] 296 | 297 | 298 | The input tensor:1.80969 1.92234 1.10676 1.50458 1.37146 1.7904 1.26449 1.17676 1.98465 1.04982 1.87448 1.15665 1.45925 1.50845 1.80422 1.95404 1.46966 1.75505 1.13797 1.45105 1.68181 1.70959 1.38402 1.23861 299 | 300 | The output tensor of TF: 0.685642 1.70892 1.03409 1.36326 0.685642 1.59976 1.1646 1.09201 0.685642 0.986972 1.66932 1.07537 0.685642 1.36646 1.61119 1.73515 0.685642 1.5705 1.05991 1.31896 0.685642 1.53289 1.2635 1.14318 301 | 302 | The output tensor of mnn ARM CPU: 0.685642 0.685642 0.685642 0.685642 0.685642 0.685642 1.1646 1.09201 1.76048 0.986972 1.66932 1.07537 1.32575 1.36646 1.61119 1.73515 303 | 1.33437 1.5705 1.05991 1.31896 1.5099 1.53289 1.2635 1.14318 304 | 305 | The output tensor of mnnX86 CPU: 0.685642 1.70892 1.03409 1.36326 0.685642 1.59976 1.1646 1.09201 0.685642 0.986972 1.66932 1.07537 0.685642 1.36646 1.61119 1.73515 306 | 0.685642 1.5705 1.05991 1.31896 0.685642 1.53289 1.2635 1.14318 307 | 308 | 309 | _RE(mnn_ARMCPU)_ of Batchnorm is 66.67%. 310 | 311 | 312 | 313 | ***DCF-15 Maxpooling. Data comparison failure on MNN ARMCPU and TensorFlow.*** 314 | --------------- 315 | 316 | 317 | The configuration of the Maxpooling operator is as follows. 318 | 319 | n = 4，h = 2，w = 2，c = 2 320 | 321 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="placeholder_10000") 322 | 323 | out_2= tf.nn.max_pool(x, ksize =[1,1,1,1] ,strides=[1,2,1,1],padding="VALID",name="maxpooling") 324 | 325 | 326 | 327 | The input tensor:1.85204 1.14694 1.97292 1.33441 1.53921 1.43205 1.9677 1.68036 1.01731 1.15119 1.28549 1.4035 1.2546 1.54129 1.75193 1.77484 328 | 1.58525 1.38116 1.10232 1.92552 1.87033 1.57231 1.842 1.51482 1.09474 1.88493 1.51141 1.08825 1.92578 1.16217 1.98962 1.87622 329 | 330 | The output tensor of TF: 1.85204 1.14694 1.97292 1.33441 1.01731 1.15119 1.28549 1.4035 1.58525 1.38116 1.10232 1.92552 1.09474 1.88493 1.51141 1.08825 331 | 332 | The output tensor of mnn ARM CPU: 1.85204 1.14694 1.53921 1.43205 1.01731 1.15119 1.2546 1.54129 1.58525 1.38116 1.87033 1.57231 1.09474 1.88493 1.92578 1.16217 333 | 334 | The output tensor of mnnX86 CPU: 1.85204 1.14694 1.97292 1.33441 1.01731 1.15119 1.28549 1.4035 1.58525 1.38116 1.10232 1.92552 1.09474 1.88493 1.51141 1.08825 335 | 336 | _RE(mnn_ARMCPU)_ of Maxpooling is 50.00%. 337 | 338 | 339 | 340 | 341 | ***DCF-16 Avgpooling. Data comparison failure on MNN ARMCPU and TensorFlow.*** 342 | --------------- 343 | The configuration of the Avgpooling operator is as follows. 344 | 345 | n = 1，h =3，w =3，c =2 346 | 347 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="placeholder_10000") 348 | 349 | out= tf.nn.avg_pool(x, [1,1,1,1],strides=[1, 1, 2, 1],padding='VALID',name="avgpooling") 350 | 351 | The input tensor:1.19919 1.64421 1.52825 1.80458 1.21612 1.13802 1.33801 1.79613 1.22788 1.59343 1.22308 1.82913 1.18881 1.68926 1.76186 1.06241 352 | 1.3741 1.17902 353 | 354 | The output tensor of TF: 1.19919 1.64421 1.21612 1.13802 1.33801 1.79613 1.22308 1.82913 1.18881 1.68926 1.3741 1.17902 355 | 356 | The output tensor of mnn ARM CPU: 1.19919 1.52825 1.80458 1.13802 1.33801 1.22788 1.59343 1.82913 1.18881 1.76186 1.06241 1.17902 357 | 358 | The output tensor of mnnX86 CPU: 1.19919 1.64421 1.21612 1.13802 1.33801 1.79613 1.22308 1.82913 1.18881 1.68926 1.3741 1.17902 359 | 360 | _RE(mnn_ARMCPU)_ of Avgpooling is 50.00%. 361 | 362 | 363 | 364 | ***DCF-17 Deconv. Data comparison failure on MNN ARMCPU and TensorFlow.*** 365 | --------------- 366 | The configuration of the Deconv operator is as follows. 367 | 368 | n = 3，h = 2，w = 2，c = 3 369 | 370 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="placeholder_10000") 371 | 372 | filter_data = np.random.uniform(-1, 1, size=[1,1,1,3]).astype(np.float32) * 0.3 # HWNC 373 | 374 | out_2 = tf.nn.conv2d_transpose(x,filter_data, strides=[1,1,1,1],output_shape=[3,2,2,1], padding="SAME",name="deconv") 375 | 376 | The input tensor:1.1954 1.46606 1.21522 1.40054 1.24535 1.60786 1.99609 1.4368 1.92193 1.71225 1.15864 1.6657 1.46595 1.74937 1.86695 1.5089 377 | 1.37309 1.5522 1.91413 1.46281 1.57776 1.94598 1.99442 1.76259 1.20255 1.82843 1.52027 1.57966 1.60674 1.7726 1.86414 1.26261 378 | 1.58428 1.17788 1.14936 1.46023 379 | 380 | The output tensor of TF: 0.301035 0.41749 0.562585 0.490874 0.432062 0.419105 0.491333 0.485923 0.322798 0.449112 0.498304 0.360348 381 | 382 | The output tensor of mnn ARM CPU: 0.428883 0.418508 0.255569 0.412048 0.415169 0.511926 0.513831 0.445365 0.349692 0.389087 0.320603 0.426135 383 | 384 | The output tensor of mnn X86 CPU: 0.301035 0.41749 0.562585 0.490874 0.432062 0.419105 0.491333 0.485923 0.322798 0.449112 0.498304 0.360348 385 | 386 | _RE(mnn_ARMCPU)_ of Deconv is 8.33%. 387 | 388 | 389 | 390 | 391 | ***DCF-18 Resize_bilinear. Data comparison failure on MNN ARMCPU and TensorFlow.*** 392 | --------------- 393 | The configuration of the Resize_bilinear operator is as follows. 394 | 395 | n = 2，h = 2，w = 2，c = 2 396 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="placeholder_10000") 397 | re = tf.image.resize_bilinear(x,size = [7,2],align_corners=False,name="resize_bilinear") 398 | 399 | The input tensor:2.52656 0.141995 5.55546 5.83816 6.57609 1.52823 4.67499 7.81567 6.39217 8.54632 7.56845 4.99505 9.92131 8.55646 8.7727 7.98398 400 | 401 | The output tensor of TF: 2.52656 0.141995 5.55546 5.83816 3.68357 0.538063 5.30389 6.40316 4.84057 0.93413 5.05233 6.96816 5.99758 1.3302 4.80077 7.53317 402 | 6.57609 1.52823 4.67499 7.81567 6.57609 1.52823 4.67499 7.81567 6.57609 1.52823 4.67499 7.81567 6.39217 8.54632 7.56845 4.99505 403 | 7.4005 8.54922 7.91252 5.84903 8.40882 8.55212 8.25659 6.70301 9.41715 8.55501 8.60067 7.55699 9.92131 8.55646 8.7727 7.98398 404 | 9.92131 8.55646 8.7727 7.98398 9.92131 8.55646 8.7727 7.98398 405 | 406 | The output tensor of mnn ARM CPU: 2.52656 0.141995 3.39196 1.76947 4.25736 3.39694 5.12276 5.02442 5.55546 5.83816 5.55546 5.83816 5.55546 5.83816 6.57609 1.52823 407 | 6.03292 3.32464 5.48975 5.12105 4.94658 6.91746 4.67499 7.81567 4.67499 7.81567 4.67499 7.81567 6.39217 8.54632 6.72825 7.53167 408 | 7.06433 6.51703 7.40041 5.50238 7.56845 4.99505 7.56845 4.99505 7.56845 4.99505 9.92131 8.55646 9.59314 8.39289 9.26496 8.22933 409 | 8.93679 8.06576 8.7727 7.98398 8.7727 7.98398 8.7727 7.98398 410 | 411 | The output tensor of mnnX86 CPU: 2.52656 0.141995 5.55546 5.83816 3.68357 0.538063 5.30389 6.40316 4.84057 0.93413 5.05233 6.96816 5.99758 1.3302 4.80077 7.53317 412 | 6.57609 1.52823 4.67499 7.81567 6.57609 1.52823 4.67499 7.81567 6.57609 1.52823 4.67499 7.81567 6.39217 8.54632 7.56845 4.99505 413 | 7.4005 8.54922 7.91252 5.84903 8.40882 8.55212 8.25659 6.70301 9.41715 8.55501 8.60067 7.55699 9.92131 8.55646 8.7727 7.98398 414 | 9.92131 8.55646 8.7727 7.98398 9.92131 8.55646 8.7727 7.98398 415 | 416 | _RE(mnn_ARMCPU)_ of Resize_bilinear is 21.428571%. 417 | 418 | 419 | 420 | ***DCF-19 Conv2d. Data comparison failure on MNN ARMCPU and TensorFlow.*** 421 | --------------- 422 | The configuration of the Conv2d operator is as follows. 423 | 424 | n = 1，h = 2，w = 1，c = 2 425 | 426 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="placeholder_10000") 427 | 428 | filter_data = (np.random.rand(1,1,2,1)*10).astype(np.float32)#HWCN 429 | 430 | out = tf.nn.conv2d(x, filter_data, strides=[1,1,1,1],padding="SAME",dilations=[1,1,1,1],name="conv") 431 | 432 | 433 | The input tensor: 3.32932 7.1221 8.35369 3.85031 434 | 435 | The output tensor of TF: 76.9141 73.1483 436 | 437 | The output tensor of mnn ARM CPU: 87.4414 67.2161 438 | 439 | The output tensor of mnnX86 CPU: 76.9141 73.1483 440 | 441 | _RE(mnn_ARMCPU)_ of Conv2d is 0%. 442 | 443 | 444 | 445 | ***DCF-20 Deconv2d. Data comparison failure on MNN ARMCPU and TensorFlow.*** 446 | --------------- 447 | 448 | 449 | The configuration of the Deconv2d operator is as follows. 450 | 451 | n = 3，h = 2，w = 2，c = 2 452 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="placeholder_10000") 453 | filter_data = np.random.uniform(-1, 1, size=[1,1,1,2]).astype(np.float32) * 0.3 # HWNC 454 | 455 | o_h = int((h + (h - 1) * (strides[1] - 1) + (kernel[0] - 1) * 2 - kernel[0]) / 1 + 1) 456 | 457 | o_w = int((w + (w - 1) * (strides[2] - 1) + (kernel[2] - 1) * 2 - kernel[2]) / 1 + 1) 458 | 459 | out_2 = tf.nn.conv2d_transpose(x,filter_data, strides=[1,1,1,1],output_shape=[3,2,2,1], padding="VALID",name="deconv") 460 | 461 | 462 | The input tensor: 1.02216 1.56782 1.13286 1.1632 1.2089 1.97479 1.98287 1.29203 1.82363 1.04148 1.4673 1.13671 1.88557 1.50993 1.14716 1.86642 463 | 1.86152 1.25844 1.81675 1.32558 1.19287 1.18547 1.63139 1.80021 464 | 465 | The output tensor of TF: -0.409532 -0.376534 -0.500582 -0.558878 -0.49423 -0.437887 -0.569335 -0.474004 -0.5308 -0.53115 -0.391181 -0.559082 466 | 467 | The output tensor of mnn ARM CPU: -0.3612 -0.570252 -0.486908 -0.399772 -0.607893 -0.405487 -0.439293 -0.471977 -0.52197 -0.403909 -0.572329 -0.499721 468 | 469 | The output tensor of mnnX86 CPU: -0.409532 -0.376534 -0.500582 -0.558878 -0.49423 -0.437887 -0.569335 -0.474004 -0.5308 -0.53115 -0.391181 -0.559082 470 | 471 | _RE(mnn_ARMCPU)_ of Deconv2d is 0%. 472 | 473 | 474 | 475 | 476 | ***DCF-21 SpaceToBatchND. Data comparison failure on MNN ARMCPU and TensorFlow.*** 477 | --------------- 478 | 479 | After analyzing the data of each layer of the model, it is found that the calculation of the first SpaceToBatchND operator is incorrect. The configuration of the SpaceToBatchND operator is as follows. 480 | 481 | n = 1，h = 2，w = 1，c = 2 482 | 483 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="placeholder_10000") 484 | 485 | filter_data = (np.random.rand(1,1,2,1)*10).astype(np.float32)#HWCN 486 | 487 | out = tf.nn.conv2d(x, filter_data, strides=[1,1,1,1],padding="SAME",dilations=[1,1,1,1],name="conv") 488 | 489 | 490 | The input tensor: 1.81285 1.80182 1.09175 1.34579 1.72089 1.51118 1.35476 1.83301 1.34466 1.81689 1.74305 1.17221 1.89568 1.43885 1.61227 1.67711 1.3753 1.71589 1.17715 1.28493 1.53955 1.91123 1.71185 1.50172 1.90004 1.10641 1.4819 1.40583 1.13994 1.18919 1.84038 1.36894 1.40226 1.32921 1.10263 1.00787 1.30912 1.74669 1.19233 1.14202 1.94152 1.43464 1.74921 1.40845 1.01566 1.70362 1.00647 1.0829 1.6929 1.10756 1.9646 1.42497 1.67611 1.22671 1.94018 1.8546 1.82265 1.48128 1.72428 1.46487 1.31575 1.15607 1.49664 1.84624 1.23121 1.16041 1.41739 1.44333 1.33098 1.24006 1.04439 1.71925 1.14984 1.81575 1.90262 1.10342 1.24691 1.10301 1.77748 1.67415 1.79952 1.42616 1.45274 1.04644 1.48503 1.07376 1.09438 1.4688 1.32712 1.69093 1.92985 1.95699 1.64589 1.42173 1.54272 1.14737 1.86412 1.31165 1.62311 1.93979 1.84567 1.88417 1.16312 1.34032 1.82499 1.23435 1.17927 1.06113 1.98209 1.59234 1.0961 1.74577 1.9278 1.91586 1.75939 1.35022 1.13791 1.80659 1.98983 1.80961 1.22276 1.63678 1.72411 1.87997 1.34435 1.84229 1.10311 1.39226 1.84723 1.82624 1.92291 1.80067 1.15709 1.43695 1.6523 1.35589 1.17243 1.6111 1.90181 1.26177 1.31176 1.64643 1.54058 1.74862 1.14093 1.55403 1.25646 1.2847 1.2333 1.50363 1.78761 1.51293 1.95894 1.50829 1.2121 1.01706 1.61223 1.78511 1.52368 1.57139 1.48224 1.92002 1.43546 1.99533 1.86006 1.40947 1.62721 1.14133 491 | 492 | The output tensor of TF: 1.81285 1.80182 1.13994 1.18919 1.82265 1.48128 1.48503 1.07376 1.9278 1.91586 1.31176 1.64643 1.09175 1.34579 1.84038 1.36894 1.72428 1.46487 1.09438 1.4688 1.75939 1.35022 1.54058 1.74862 1.72089 1.51118 1.40226 1.32921 1.31575 1.15607 1.32712 1.69093 1.13791 1.80659 1.14093 1.55403 1.35476 1.83301 1.10263 1.00787 1.49664 1.84624 1.92985 1.95699 1.98983 1.80961 1.25646 1.2847 1.34466 1.81689 1.30912 1.74669 1.23121 1.16041 1.64589 1.42173 1.22276 1.63678 1.2333 1.50363 1.74305 1.17221 1.19233 1.14202 1.41739 1.44333 1.54272 1.14737 1.72411 1.87997 1.78761 1.51293 1.89568 1.43885 1.94152 1.43464 1.33098 1.24006 1.86412 1.31165 1.34435 1.84229 1.95894 1.50829 1.61227 1.67711 1.74921 1.40845 1.04439 1.71925 1.62311 1.93979 1.10311 1.39226 1.2121 1.01706 1.3753 1.71589 1.01566 1.70362 1.14984 1.81575 1.84567 1.88417 1.84723 1.82624 1.61223 1.78511 1.17715 1.28493 1.00647 1.0829 1.90262 1.10342 1.16312 1.34032 1.92291 1.80067 1.52368 1.57139 1.53955 1.91123 1.6929 1.10756 1.24691 1.10301 1.82499 1.23435 1.15709 1.43695 1.48224 1.92002 1.71185 1.50172 1.9646 1.42497 1.77748 1.67415 1.17927 1.06113 1.6523 1.35589 1.43546 1.99533 1.90004 1.10641 1.67611 1.22671 1.79952 1.42616 1.98209 1.59234 1.17243 1.6111 1.86006 1.40947 1.4819 1.40583 1.94018 1.8546 1.45274 1.04644 1.0961 1.74577 1.90181 1.26177 1.62721 1.14133 493 | 494 | The output tensor of mnn ARM CPU: 1.81285 1.61227 1.13994 1.74921 1.82265 1.04439 1.48503 1.62311 1.9278 1.10311 1.31176 1.2121 1.80182 1.67711 1.18919 1.40845 1.48128 1.71925 1.07376 1.93979 1.91586 1.39226 1.64643 1.01706 1.09175 1.3753 1.84038 1.01566 1.72428 1.14984 1.09438 1.84567 1.75939 1.84723 1.54058 1.61223 1.34579 1.71589 1.36894 1.70362 1.46487 1.81575 1.4688 1.88417 1.35022 1.82624 1.74862 1.78511 1.72089 1.17715 1.40226 1.00647 1.31575 1.90262 1.32712 1.16312 1.13791 1.92291 1.14093 1.52368 1.51118 1.28493 1.32921 1.0829 1.15607 1.10342 1.69093 1.34032 1.80659 1.80067 1.55403 1.57139 1.35476 1.53955 1.10263 1.6929 1.49664 1.24691 1.92985 1.82499 1.98983 1.15709 1.25646 1.48224 1.83301 1.91123 1.00787 1.10756 1.84624 1.10301 1.95699 1.23435 1.80961 1.43695 1.2847 1.92002 1.34466 1.71185 1.30912 1.9646 1.23121 1.77748 1.64589 1.17927 1.22276 1.6523 1.2333 1.43546 1.81689 1.50172 1.74669 1.42497 1.16041 1.67415 1.42173 1.06113 1.63678 1.35589 1.50363 1.99533 1.74305 1.90004 1.19233 1.67611 1.41739 1.79952 1.54272 1.98209 495 | 1.72411 1.17243 1.78761 1.86006 1.17221 1.10641 1.14202 1.22671 1.44333 1.42616 1.14737 1.59234 1.87997 1.6111 1.51293 1.40947 1.89568 1.4819 1.94152 1.94018 1.33098 1.45274 1.86412 1.0961 1.34435 1.90181 1.95894 1.62721 1.43885 1.40583 1.43464 1.8546 1.24006 1.04644 1.31165 1.74577 1.84229 1.26177 1.50829 1.14133 496 | 497 | The output tensor of mnnX86 CPU: 1.81285 1.80182 1.13994 1.18919 1.82265 1.48128 1.48503 1.07376 1.9278 1.91586 1.31176 1.64643 1.09175 1.34579 1.84038 1.36894 1.72428 1.46487 1.09438 1.4688 1.75939 1.35022 1.54058 1.74862 1.72089 1.51118 1.40226 1.32921 1.31575 1.15607 1.32712 1.69093 1.13791 1.80659 1.14093 1.55403 1.35476 1.83301 1.10263 1.00787 1.49664 1.84624 1.92985 1.95699 1.98983 1.80961 1.25646 1.2847 1.34466 1.81689 1.30912 1.74669 1.23121 1.16041 1.64589 1.42173 1.22276 1.63678 1.2333 1.50363 1.74305 1.17221 1.19233 1.14202 1.41739 1.44333 1.54272 1.14737 1.72411 1.87997 1.78761 1.51293 1.89568 1.43885 1.94152 1.43464 1.33098 1.24006 1.86412 1.31165 1.34435 1.84229 1.95894 1.50829 1.61227 1.67711 1.74921 1.40845 1.04439 1.71925 1.62311 1.93979 1.10311 1.39226 1.2121 1.01706 1.3753 1.71589 1.01566 1.70362 1.14984 1.81575 1.84567 1.88417 1.84723 1.82624 1.61223 1.78511 1.17715 1.28493 1.00647 1.0829 1.90262 1.10342 1.16312 1.34032 1.92291 1.80067 1.52368 1.57139 1.53955 1.91123 1.6929 1.10756 1.24691 1.10301 1.82499 1.23435 498 | 1.15709 1.43695 1.48224 1.92002 1.71185 1.50172 1.9646 1.42497 1.77748 1.67415 1.17927 1.06113 1.6523 1.35589 1.43546 1.99533 1.90004 1.10641 1.67611 1.22671 1.79952 1.42616 1.98209 1.59234 1.17243 1.6111 1.86006 1.40947 1.4819 1.40583 1.94018 1.8546 1.45274 1.04644 1.0961 1.74577 1.90181 1.26177 1.62721 1.14133 499 | 500 | _RE(mnn_ARMCPU)_ of SpaceToBatchND is 7.14% . 501 | 502 | 503 | 504 | 505 | ***DCF-22 Reducemax. Data comparison failure on MNN X86CPU and TensorFlow.*** 506 | --------------- 507 | ![dcf22_reducemax_rn_3_832029_0001](https://user-images.githubusercontent.com/69624583/92606825-f6301400-f2e5-11ea-9d74-6ac27b614389.png) 508 | 509 | After analyzing the data of each layer of the model, it is found that the calculation of the Reducemax operator is incorrect. When comparing nan and -inf(output by sqrt), TensorFlow returns -inf, MNN returns nan. The result of calculation with nan should be nan. And thus, the cause of this inconsistency should be the calculation error of TensorFlow. 510 | 511 | The configuration of the Reducemax operator is as follows. 512 | 513 | n = 1，h = 2，w = 7，c = 1 514 | 515 | nx =np.array(x,dtype=np.float32).reshape(n,h,w,c) 516 | 517 | placeholder = tf.placeholder(shape=(n,h,w,c),dtype=tf.float32,name="placeholder_10000") 518 | 519 | res_reducemax = tf.reduce_max(placeholder, axis=1, keep_dims=True, name = "reducemax") 520 | 521 | 522 | The input tensor： [[[[ 1.] [ 2.] [ 3.] [ 4.] [ nan] [ 5.] [ 6.]] [[ 7.] [ 8.] [ 9.] [ inf] [-inf] [ nan] [ 10.]]]] 523 | 524 | The output tensor of TF: [[[[7.] [8.] [9.] [inf] [-inf] [5.] [10.]]]] 525 | 526 | The output tensor of mnn X86 CPU: [[[[7.] [8.] [9.] [inf] [nan] [5] [10.]]]] 527 | 528 | 529 | 530 | ***DCF-23 Reducemax. Data comparison failure on MNN X86CPU and TensorFlow.*** 531 | --------------- 532 | ![dcf23_reducemax_ws_2_832024_0001](https://user-images.githubusercontent.com/69624583/92606774-e7496180-f2e5-11ea-81f7-14b0591cff10.png) 533 | 534 | After analyzing the data of each layer of the model, it is found that the calculation of the Reducemax operator is incorrect.The result of calculation should be 9.0 and 10.0. And thus, the cause of this inconsistency should be the calculation error of TensorFlow. 535 | 536 | The configuration of the Reducemax operator is as follows. 537 | 538 | n = 1，h = 2，w = 2，c = 2 539 | 540 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="placeholder_10000") 541 | 542 | x_data = (np.random.rand(n,h,w,c)*10).astype(np.float32) 543 | 544 | ceil =tf.math.ceil(x,name="ceil") 545 | 546 | out = tf.reduce_max(ceil,axis =[0,1,3], keep_dims=True,name="reducemax") 547 | 548 | 549 | The input tensor： [[[[8.370346 7.894142 ] [4.6929383 0.80955863]] [[4.7769775 8.743156 ] [2.3229198 9.342096 ]]]] 550 | 551 | The output tensor of Ceil operator:[[[[ 9. 8.] [ 5. 1.]] [[ 5. 9.] [ 3. 10.]]]] 552 | 553 | The output tensor of TF: 8.74316 9.3421 554 | 555 | The output tensor of mnn X86 CPU: 9 10 556 | 557 | 558 | 559 | ***DCF-24 operators unsupported in MNN reference guide.*** 560 | --------------- 561 | 562 | We also tried to infer some models that contain operators unsupported in MNN reference guide, such as Addn, Clip and Asin. And these models are successfully inferred by MNN on X86 CPU and ARM CPU. 563 | 564 | 565 | ***DCF-25 Data types unsupported in MNN reference guide.*** 566 | --------------- 567 | 568 | We also tried to infer some models that Data types unsupported in MNN reference guide, such as uint8 and bool of Cast. And these models are successfully inferred by MNN on X86 CPU and ARM CPU. 569 | 570 | 571 | 572 | 573 | 574 | ***DCF-26 BatchToSpaceND. Data comparison failure on MNN ARMCPU and TensorFlow.*** 575 | --------------- 576 | ![dcf26_BatchToSpaceND_rn_2_268332_0001](https://user-images.githubusercontent.com/69624583/92606701-cf71dd80-f2e5-11ea-9cd0-e297966183f4.png) 577 | 578 | After analyzing the data of each layer of the model, it is found that the calculation of the first SpaceToBatchND operator is incorrect. The configuration of the SpaceToBatchND operator is as follows. 579 | n = 2,h = 2,w = 3,c = 4 580 | x = tf.placeholder(shape=(n,h,w,c), dtype=tf.float32,name="placeholder_10000") 581 | out = tf.batch_to_space_nd(x,crops=[[0,0],[0,0]], block_shape=[1,2], name="batch") 582 | 583 | 584 | The input tensor: 1.03884 1.16265 1.76784 1.78022 1.1338 1.21764 1.25461 1.51277 1.48314 1.70554 1.19038 1.47641 1.07206 1.07769 1.47213 1.23621 1.68363 1.49882 1.0184 1.8426 1.07029 1.28709 1.45593 1.01174 1.47935 1.06786 1.02764 1.64876 1.058 1.7861 1.34905 1.58041 1.00863 1.9733 1.76953 1.06869 1.36675 1.32908 1.38327 1.39296 1.1866 1.20684 1.53111 1.92347 1.58694 1.74022 1.87438 1.38115 585 | 586 | The output tensor of TF: 1.03884 1.16265 1.76784 1.78022 1.47935 1.06786 1.02764 1.64876 1.1338 1.21764 1.25461 1.51277 1.058 1.7861 1.34905 1.58041 587 | 1.48314 1.70554 1.19038 1.47641 1.00863 1.9733 1.76953 1.06869 1.07206 1.07769 1.47213 1.23621 1.36675 1.32908 1.38327 1.39296 588 | 1.68363 1.49882 1.0184 1.8426 1.1866 1.20684 1.53111 1.92347 1.07029 1.28709 1.45593 1.01174 1.58694 1.74022 1.87438 1.38115 589 | 590 | The output tensor of mnn ARM CPU: 1.03884 1.47935 1.16265 1.06786 1.76784 1.02764 1.78022 1.64876 1.1338 1.058 1.21764 1.7861 1.25461 1.34905 1.51277 1.58041 1.48314 1.00863 1.70554 1.9733 1.19038 1.76953 1.47641 1.06869 1.07206 1.36675 1.07769 1.32908 1.47213 1.38327 1.23621 1.39296 1.68363 1.1866 1.49882 1.20684 1.0184 1.53111 1.8426 1.92347 1.07029 1.58694 1.28709 1.74022 1.45593 1.87438 1.01174 1.38115 591 | 592 | The output tensor of mnnX86 CPU: 1.03884 1.16265 1.76784 1.78022 1.47935 1.06786 1.02764 1.64876 1.1338 1.21764 1.25461 1.51277 1.058 1.7861 1.34905 1.58041 1.48314 1.70554 1.19038 1.47641 1.00863 1.9733 1.76953 1.06869 1.07206 1.07769 1.47213 1.23621 1.36675 1.32908 1.38327 1.39296 1.68363 1.49882 1.0184 1.8426 1.1866 1.20684 1.53111 1.92347 1.07029 1.28709 1.45593 1.01174 1.58694 1.74022 1.87438 1.38115 593 | 594 | _RE(mnn_ARMCPU)_ of SpaceToBatchND is 25.00% . 595 | 596 | 597 | 598 | 599 | 600 | 601 | ***DCF-27 Maximum. Data comparison failure on MNN X86CPU and TensorFlow.*** 602 | --------------- 603 | 604 | After analyzing the data of each layer of the model, it is found that the calculation of the Maximum operator is incorrect. When comparing nan and -inf(output by sqrt), TensorFlow returns nan , MNN returns -inf. The result of calculation with nan should be nan. And thus, the cause of this inconsistency should be the calculation error of MNN. 605 | 606 | The configuration of the Maximum operator is as follows. 607 | 608 | placeholder1 = tf.placeholder(shape=shape1,dtype=tf.float32,name="placeholder_10000") 609 | 610 | placeholder2 = tf.placeholder(shape=shape2,dtype=tf.float32,name="placeholder_10200") 611 | 612 | res_maximum = tf.math.minimum(placeholder1, placeholder2,name = "minimum_outputdata_10100") 613 | 614 | 615 | The input tensor1： nan nan inf inf nan nan inf nan inf inf nan inf inf nan nan nan inf inf inf inf nan inf nan inf 616 | 617 | The input tensor2：nan nan -inf nan -inf -inf -inf nan -inf -inf nan nan nan nan nan nan -inf -inf -inf nan -inf -inf nan -inf 618 | 619 | The output tensor of TF: nan, nan, inf, inf, nan, nan, inf, nan, inf, inf, nan, inf, inf, nan, nan, nan, inf, inf, inf, inf, nan, inf, nan, inf 620 | 621 | The output tensor of mnn X86 CPU: nan, nan, inf, inf, -inf, -inf, inf, nan, inf, inf, nan, inf, nan, nan, nan, nan, inf, inf, inf, inf, -inf, inf, nan, inf 622 | 623 | 624 | 625 | ***DCF-28 Minimum. Data comparison failure on MNN X86CPU and TensorFlow.*** 626 | --------------- 627 | 628 | After analyzing the data of each layer of the model, it is found that the calculation of the Minimum operator is incorrect. When comparing nan and -inf(output by sqrt), TensorFlow returns nan , MNN returns -inf. The result of calculation with nan should be nan. And thus, the cause of this inconsistency should be the calculation error of MNN. 629 | 630 | The configuration of the Minimum operator is as follows. 631 | 632 | placeholder1 = tf.placeholder(shape=shape1,dtype=tf.float32,name="placeholder_10000") 633 | 634 | placeholder2 = tf.placeholder(shape=shape2,dtype=tf.float32,name="placeholder_10200") 635 | 636 | res_maximum = tf.math.minimum(placeholder1, placeholder2,name = "minimum_outputdata_10100") 637 | 638 | 639 | The input tensor1： nan inf nan nan nan nan nan inf inf inf nan nan nan nan inf inf inf inf inf nan nan nan nan nan 640 | 641 | The input tensor2：-inf -inf nan -inf -inf nan -inf nan nan -inf nan nan -inf -inf nan nan nan nan -inf -inf nan -inf nan -inf 642 | 643 | The output tensor of TF: nan, -inf, nan, nan, nan, nan, nan, inf, inf, -inf, nan, nan, nan, nan, inf, inf, inf, inf, -inf, nan, nan, nan, nan, nan 644 | 645 | The output tensor of mnn X86 CPU: -inf, -inf, nan, -inf, -inf, nan, -inf, nan, nan, -inf, nan, nan, -inf, -inf, nan, nan, nan, nan, -inf, -inf, nan, -inf, nan, -inf 646 | 647 | 648 | 649 | 650 | 651 | 652 | ***DCF-29 Reducemin. Data comparison failure on MNN X86CPU and TensorFlow.*** 653 | --------------- 654 | 655 | After analyzing the data of each layer of the model, it is found that the calculation of the Reducemax operator is incorrect. 656 | The decimal part of the calculation result should be 0. And thus, the cause of this inconsistency should be the calculation error of TensorFlow. 657 | 658 | The configuration of the Maximum operator is as follows. 659 | 660 | n = 8, h = 5, w = 3, = 4 661 | 662 | placeholder = tf.placeholder(shape=(n,h,w,c),dtype=tf.float32,name="placeholder_10000") 663 | 664 | res_ceil = tf.math.ceil(placeholder,name="celi_outputdata_10000") 665 | 666 | res_reducemin = tf.reduce_min(res_ceil,axis=[1,2,3], keep_dims=True,name="reducemin") 667 | 668 | The input tensor of Ceil ： 0.692225 0.93858 0.712277 0.225081 0.949312 0.161038 0.0869423 0.529012 0.213447 0.505565 0.104097 0.12932 0.91826 0.41102 0.639792 0.651933 0.138182 0.338399 0.354983 0.5047 0.821886 0.719497 0.888915 0.260079 0.979791 0.530006 0.636251 0.172793 0.6888 0.996615 0.277687 0.397159 0.0931961 0.414661 0.307734 0.296113 0.944499 0.509749 0.126067 0.833146 0.669131 0.952248 0.987454 0.362709 0.010182 0.693371 0.710924 0.304081 0.741464 0.494248 0.880531 0.284276 0.477406 0.0302742 0.939258 0.712152 0.648007 0.867603 0.577887 0.460249 0.00215121 0.304308 0.775361 0.999424 0.354603 0.221053 0.895589 0.140296 0.760546 0.766981 0.81381 0.815747 0.190551 0.0709158 0.33612 0.161643 0.187871 0.962751 0.504239 0.425821 0.286352 0.56067 0.556322 0.96662 0.509775 0.951699 0.987685 0.352054 0.955333 0.0554917 0.236716 0.967292 0.95307 0.841556 0.0750637 0.80934 0.217561 0.672645 0.946269 0.596971 0.42113 0.500691 0.11898 0.504732 0.169392 0.138503 0.19804 0.000392505 0.388237 0.878493 0.949584 0.122674 0.991812 0.380613 0.599924 0.580828 0.898459 0.698719 0.395472 0.198427 0.62872 0.0436476 0.575488 0.271558 0.709563 0.294651 0.0722937 0.996233 0.405947 0.865826 0.753831 0.443388 0.456224 0.440658 0.374573 0.309357 0.371972 0.5165 0.45658 0.824428 0.3738 0.54454 0.117894 0.671438 0.405704 0.151209 0.77726 0.236028 0.0347466 0.797211 0.676711 0.941772 0.668484 0.68935 0.814786 0.017036 0.213829 0.306827 0.000708339 0.561073 0.885485 0.38042 0.601363 0.910492 0.817526 0.433484 0.998961 0.0579396 0.0354243 0.208774 0.572147 0.303477 0.795528 0.470845 0.0611874 0.248836 0.756074 0.741556 0.37829 0.836266 0.171869 0.648324 0.985769 0.479648 0.0410855 0.178417 0.966587 0.818564 0.779305 0.123689 0.454796 0.0167353 0.31844 0.636537 0.533819 0.157864 0.0354836 0.678528 0.916046 0.957519 0.922705 0.234027 0.883145 0.261043 0.859908 0.665565 0.329577 0.849441 0.504726 0.274414 0.129874 0.635124 0.905432 0.932581 0.259519 0.0896101 0.0422479 0.544967 0.630855 0.799092 0.930978 0.709916 0.651128 0.128197 0.622245 0.802919 0.952257 0.269869 0.101824 0.994266 0.874646 0.416595 0.527015 0.391797 0.169745 0.161902 0.817495 0.549307 0.241133 0.504831 0.454197 0.555321 0.62291 0.31103 0.261787 0.214955 0.97376 0.183044 0.0249169 0.0371966 0.0570485 0.925517 0.870879 0.698749 0.830457 0.938765 0.389112 0.00295612 0.485772 0.313742 0.392849 0.0771843 0.0539229 0.838047 0.464059 0.880103 0.545655 0.0732461 0.72186 0.413933 0.804667 0.514696 0.954016 0.937535 0.216915 0.398839 0.0536812 0.733149 0.184013 0.901414 0.35778 0.886324 0.219489 0.530526 0.108065 0.31779 0.147274 0.739766 0.933226 0.791286 0.0877398 0.438235 0.508087 0.0988351 0.827244 0.670466 0.941197 0.253886 0.444018 0.447521 0.132748 0.626819 0.470801 0.624326 0.200641 0.272204 0.149866 0.0652196 0.961756 0.675887 0.762648 0.35655 0.262811 0.202773 0.790241 0.134048 0.55609 0.832261 0.830504 0.768885 0.760313 0.704658 0.667812 0.0166153 0.266377 0.281874 0.269549 0.472509 0.0273559 0.436977 0.848994 0.380641 0.279575 0.29261 0.449598 0.660364 0.923326 0.401439 0.256106 0.626845 0.973464 0.107286 0.99078 0.162073 0.496947 0.69799 0.647916 0.552273 0.234355 0.261813 0.122905 0.246059 0.0180538 0.560145 0.19938 0.73054 0.454653 0.127464 0.997552 0.711823 0.151022 0.245862 0.685221 0.12892 0.00692515 0.717766 0.215804 0.215127 0.220378 0.470857 0.860159 0.866804 0.00436794 0.128258 0.331338 0.177153 0.400856 0.586165 0.238451 0.814553 0.962278 0.151696 0.821186 0.922021 0.272277 0.763492 0.332733 0.131988 0.572224 0.455728 0.809886 0.827012 0.559146 0.499985 0.75889 0.061708 0.659621 0.0512265 0.802936 0.578957 669 | 0.870488 0.874845 0.340076 0.435762 0.542651 0.443902 0.726688 0.857814 0.919516 0.583788 0.711802 0.439743 0.911983 0.234367 0.178278 0.851924 0.231062 0.592798 0.0952106 0.356535 0.140992 0.0791117 0.958825 0.577415 0.578921 0.212631 0.739432 0.555858 0.716206 0.288654 0.256814 0.0622695 0.206076 0.741558 0.100626 0.685045 0.0792263 0.826257 0.561469 0.102868 0.848553 0.259438 0.518111 0.33108 0.890541 0.0106224 0.474491 0.29854 0.453791 0.139653 0.251248 0.657253 0.20492 0.529513 0.776427 0.28525 0.559803 0.531715 0.510169 0.544994 0.741698 0.5395 0.94634 0.907282 0.204054 0.921818 0.072024 0.866193 0.553041 0.296966 0.656794 0.810907 0.0374486 0.378581 0.6274 0.35115 0.612189 0.944294 0.135113 0.974301 670 | 671 | The input tensor of Reduce_min： 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 672 | 673 | The output tensor of TF: 0.010182 0.000392505 0.000708339 0.0167353 0.00295612 0.0166153 0.00436794 0.0106224 674 | 675 | The output tensor of mnn X86 CPU: 1 1 1 1 1 1 1 1 676 | 677 | 678 | 679 | 680 | 681 | ***DCF-30-35 Reducemax, Reducemin, Reduceany, Reducemean, Reducesum and Reduceall. Data comparison failure on MNN X86CPU and TensorFlow.*** 682 | --------------- 683 | When the axis parameter is [], the shape of the output tensor is 0 in MNN. The output result of tensorflow is consistent with the data input to the operator. 684 | 685 | MNN log: 686 | The reduce_XXXX's input is not ready. 687 | 688 | 689 | 690 | 691 | 692 | ***DCF-36 Data comparison failure on MNN X86CPU and TensorFlow.*** 693 | --------------- 694 | 695 | ![dcf36_1_29_646065_0001](https://user-images.githubusercontent.com/69624583/92746508-eda70e80-f3b5-11ea-91e4-8f773449bade.png) 696 | ![dcf36_2_29_646065_0001](https://user-images.githubusercontent.com/69624583/92746539-f39cef80-f3b5-11ea-9718-f0a80d4efc42.png) 697 | 698 | This model has 4 output results, one of which is inconsistent between MNN X86CPU and TensorFlow. The main difference in the results is that the output is nan(-nan) or inf(-inf). 699 | 700 | 701 | _RE(mnn_ARMCPU)_ of the addn_12800 is 86.11%. 702 | 703 | _RE(mnn_ARMCPU)_ of the addn_12000 is 100%. 704 | 705 | _RE(mnn_ARMCPU)_ of the rsqrt_10100 is 100%. 706 | 707 | _RE(mnn_ARMCPU)_ of the rsqrt_12300 is 100%. 708 | 709 | The output tensor of TF: inf inf inf 24.0558 21.6177 18.5425 20.3782 17.1967 19.8565 -nan -nan -nan 19.2809 19.2534 25.4682 21.8413 26.5787 19.5057 -nan -nan inf 4.89995 6.50453 5.52601 4.82719 5.15199 7.9205 inf inf inf 2.4 2.4 2.4 2.4 2.4 2.4 710 | 711 | 712 | The output tensor of mnn X86 CPU: inf inf inf 24.0558 21.6177 18.5425 20.3782 17.1967 19.8565 inf inf inf 19.2809 19.2534 25.4682 21.8413 26.5787 19.5057 inf inf inf 4.89995 6.50453 5.52601 4.82719 5.15199 7.9205 inf inf inf 2.4 2.4 2.4 2.4 2.4 2.4 713 | 714 | 715 | 716 | 717 | 718 | 719 | --------------------------------------------------------------------------------