Tensorflow has not supported operator [Softmax] with name [cls_fc/Softmax]

[PapaMadeleine2022]

hello, my tf is tensorflow-gpu-1.4.1. when I run mmconvert -sf tensorflow -in RNet-20.meta -iw RNet-20 --dstNodeName cls_fc/Softmax -df caffe -om tf_RNet on my RNet-20, it still shows:

Parse file [RNet-20.meta] with binary format successfully.
Tensorflow model file [RNet-20.meta] loaded successfully.
Tensorflow checkpoint file [RNet-20] loaded successfully. [29] variables loaded.
Tensorflow has not supported operator [Softmax] with name [cls_fc/Softmax].

but my RNet-20.meta does has cls_fc/Softmax, how to fix it? my RNet-20.meta is here

[PapaMadeleine2022]

need help~

[namizzz]

Hi @IvyGongoogle , thanks for your issue, we have fixed the problems!

[Archernarkiu]

@IvyGongoogle I am tring to convert tensorflow model of mtcnn now, but another error occurred :"

F0926 17:53:15.171332 1120 reshape_layer.cpp:87] Check failed: top[0]->count() == bottom[0]->count() (768 vs. 2) output count must match input count ** Check failure stack trace: ./run.sh: line 3: 1120 Aborted (core dumped) mmconvert -sf tensorflow -in model-30.meta -iw model-30.ckpt --dstNodeName cls_prob -df caffe -om tf_pnet**

". Could you offer me some advice, thanks.

[namizzz]

Hi @Archernarkiu , maybe it's better to new a issue~

[liangheng]

@IvyGongoogle I am tring to convert tensorflow model of mtcnn now, but another error occurred :"

F0926 17:53:15.171332 1120 reshape_layer.cpp:87] Check failed: top[0]->count() == bottom[0]->count() (768 vs. 2) output count must match input count ** Check failure stack trace: ./run.sh: line 3: 1120 Aborted (core dumped) mmconvert -sf tensorflow -in model-30.meta -iw model-30.ckpt --dstNodeName cls_prob -df caffe -om tf_pnet**

". Could you offer me some advice, thanks.

@Archernarkiu how about convert tensorflow model of mtcnn? I have the same doubt, please inform me. Thank you

[Archernarkiu]

@liangheng Emmmm, It has been half a year since I proposed this question. I failed to convert my model then, however I almost forgot the details. But I conclude that, maybe this tool would fix the graph's input while converting your input graph. I didn't feel strange now, cause many tools (tensortRT, tensorflow Lite), as far as I know, have this trait.

[liangheng]

@Archernarkiu thank you very much @namizzz @IvyGongoogle I use the following command: mmconvert -sf tensorflow -in E:/tmp/pnet-30.meta -iw E:/tmp/pnet-30.index --dstNode conv4_1/Softmax conv4_1/BiasAdd conv4_3/BiasAdd -df caffe -om E:/tmp/outmodel

it happen: Traceback (most recent call last): File "d:\softinstall\anaconda3\lib\runpy.py", line 184, in _run_module_as_main "main", mod_spec) File "d:\softinstall\anaconda3\lib\runpy.py", line 85, in _run_code exec(code, run_globals) File "D:\SoftInstall\Anaconda3\Scripts\mmconvert.exe__main__.py", line 9, in File "d:\softinstall\anaconda3\lib\site-packages\mmdnn\conversion_script\convert.py", line 102, in _main ret = convertToIR._convert(ir_args) File "d:\softinstall\anaconda3\lib\site-packages\mmdnn\conversion_script\convertToIR.py", line 115, in _convert parser.run(args.dstPath) File "d:\softinstall\anaconda3\lib\site-packages\mmdnn\conversion\common\DataStructure\parser.py", line 22, in run self.gen_IR() File "d:\softinstall\anaconda3\lib\site-packages\mmdnn\conversion\tensorflow\tensorflow_parser.py", line 424, in gen_IR func(current_node) File "d:\softinstall\anaconda3\lib\site-packages\mmdnn\conversion\tensorflow\tensorflow_parser.py", line 552, in rename_Conv2D self.set_weight(node_name, 'weights', self.ckpt_data[W.name]) KeyError: 'conv1/weights'

need help

[liangheng]

I used my own script to implement the transformation, but the precision was lost, I don't know why