novioleo / crnn.mxnet Goto Github PK

Hello @novioleo

How to get the confidence score of output result ?

Hello, @novioleo ! Your work is very nice, I just use the demo，the result is good.But when I use my dataset(a picture contains two lines of characters), the result is not satisfactory. So, can you give me some help for training(include preparing dataset, and parameter);or can you give me some example, I can't start this project!Thank you very much!

i use the algorithm from opencv_contrib to location the text(temporary)，to detect the character area.this algorithm is suit for non-chinese text.i'll replace it with other simpler algorithm ,so that the mobile device can detect the text area much more efficient.

Recently I've changed the structure of the web to make it more generic, adapting to more than just generating data sets. The current model can very good fitting we generated data, but the reality data and background too miscellaneous and too much, we can not collect enough data, therefore, in order to improve the ability of generic OCR,I put forward several solutions:
1. Background filtering: the operation of max-pooling minus avg-pooling is performed on the first or second layer.
2. Reduce the number of pooling: not compress the number of rows to 1 row, and increase the input steps to LSTM.
3. Reduce the number of kernels of convolution: it can improve generalization, but it will slow down the convergence rate and reduce the accuracy slightly. BUT It also reduces the size of the model.

I am trying to test the cpp version using main.cpp file but i am unable to understand how to read file
in bytes as const std::vector so that i can pass it to predict function in predictor class.

Please help this is the main function i am using.

int main(int argc, char** argv) {
    predictor obj = predictor(argv[2],argv[3]);
   // read image here in bytes format
    try {
        std::string result = obj.predict(image);
    }catch(std::exception& e){
        std::cout<<e.what()<<std::endl;
    }
 return 0;
}

and this is my prediction function prototype

std::string predict(const std::vector<uchar> &byte_data);

if above things are unclear i can try to explain again. thank you very much.

楼主你好，请问这个只能识别中文吗？

@novioleo Hi, I'm trying to train crnn on synth90k with your code.But it behaves badly.While the origin version get 93% on that dataset ,my model can only get 60% accuracy. A little different from the original version , I use the adam optimizer as you used with learning rate set to be 1e-4.Could you give me some advice?
Thank you very much.

IN THIS ISSUE ,I WILL POST THE SCRIPT WHICH I BUILD FOR CPP INVOKER.

i run the python predictor.py ,there are some errors:
[11:38:15] /home/xuting/ocr/mxnet/dmlc-core/include/dmlc/./logging.h:308: [11:38:15] src/core/op.cc:55: Check failed: op != nullptr Operator WarpCTC is not registered

Stack trace returned 10 entries:
[bt] (0) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(_ZN4nnvm2Op3GetERKSs+0x329) [0x7f1833825449]
[bt] (1) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(+0x28aecf8) [0x7f183380ccf8]
[bt] (2) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(_ZN4dmlc20JSONObjectReadHelper13ReadAllFieldsEPNS_10JSONReaderE+0x100) [0x7f18338130a0]
[bt] (3) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(+0x28adb86) [0x7f183380bb86]
[bt] (4) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(ZNSt17_Function_handlerIFN4nnvm5GraphES1_EPS2_E9_M_invokeERKSt9_Any_dataS1+0x11f) [0x7f1832487d4f]
[bt] (5) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(_ZN4nnvm11ApplyPassesENS_5GraphERKSt6vectorISsSaISsEE+0x501) [0x7f1833818a21]
[bt] (6) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(_ZN5mxnet18LoadLegacyJSONPassEN4nnvm5GraphE+0x180) [0x7f1832480e60]
[bt] (7) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(ZNSt17_Function_handlerIFN4nnvm5GraphES1_EPS2_E9_M_invokeERKSt9_Any_dataS1+0x11f) [0x7f1832487d4f]
[bt] (8) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(_ZN4nnvm11ApplyPassesENS_5GraphERKSt6vectorISsSaISsEE+0x501) [0x7f1833818a21]
[bt] (9) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(_ZN4nnvm9ApplyPassENS_5GraphERKSs+0x8e) [0x7f18327e7f7e]

Traceback (most recent call last):
File "predictor.py", line 133, in
my_predictor = predict(images,(256,32),'model/digit',0,'./digit.txt',32,24,128,False)
File "predictor.py", line 77, in init
label_names=('label',)
File "/usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/module/module.py", line 143, in load
sym, args, auxs = load_checkpoint(prefix, epoch)
File "/usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/model.py", line 394, in load_checkpoint
symbol = sym.load('%s-symbol.json' % prefix)
File "/usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/symbol/symbol.py", line 1913, in load
check_call(_LIB.MXSymbolCreateFromFile(c_str(fname), ctypes.byref(handle)))
File "/usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/base.py", line 143, in check_call
raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: Failed loading Op warpctc0 of type WarpCTC: [11:38:15] src/core/op.cc:55: Check failed: op != nullptr Operator WarpCTC is not registered

Stack trace returned 10 entries:
[bt] (0) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(_ZN4nnvm2Op3GetERKSs+0x329) [0x7f1833825449]
[bt] (1) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(+0x28aecf8) [0x7f183380ccf8]
[bt] (2) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(_ZN4dmlc20JSONObjectReadHelper13ReadAllFieldsEPNS_10JSONReaderE+0x100) [0x7f18338130a0]
[bt] (3) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(+0x28adb86) [0x7f183380bb86]
[bt] (4) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(ZNSt17_Function_handlerIFN4nnvm5GraphES1_EPS2_E9_M_invokeERKSt9_Any_dataS1+0x11f) [0x7f1832487d4f]
[bt] (5) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(_ZN4nnvm11ApplyPassesENS_5GraphERKSt6vectorISsSaISsEE+0x501) [0x7f1833818a21]
[bt] (6) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(_ZN5mxnet18LoadLegacyJSONPassEN4nnvm5GraphE+0x180) [0x7f1832480e60]
[bt] (7) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(ZNSt17_Function_handlerIFN4nnvm5GraphES1_EPS2_E9_M_invokeERKSt9_Any_dataS1+0x11f) [0x7f1832487d4f]
[bt] (8) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(_ZN4nnvm11ApplyPassesENS_5GraphERKSt6vectorISsSaISsEE+0x501) [0x7f1833818a21]
[bt] (9) /usr/lib/python2.7/site-packages/mxnet-0.11.1-py2.7.egg/mxnet/libmxnet.so(_ZN4nnvm9ApplyPassENS_5GraphERKSs+0x8e) [0x7f18327e7f7e]

I see the language used in this project is C++ and the code is coding in python, so is there any problem I use python to run this code

hi,there
did you test the code on windows?
thx!

hi,novioleo,thanks for your job.can you plz support the multi gpus training version?thanks.

hi, there
did you find some new about repeated letter of result ?

请问一下，模型批量预测有说明文档吗？