End2End Captcha Recognition (OCR) [Segmentation fault]

[ZHmao]

When I run the End2End Captcha Recognition (OCR) example, get a segmentation fault. This is the original blog url I changed the devs, from gpu to cpu, for lack of hardware[gpu] Below is the main code:

class OCRIter(mx.io.DataIter):
    def __init__(self, count, batch_size, num_label, height, width):
        super(OCRIter, self).__init__()
        self.captcha = ImageCaptcha(fonts=['./data/OpenSans-Regular.ttf'])
        self.batch_size = batch_size
        self.count = count
        self.height = height
        self.width = width
        self.provide_data = [('data', (batch_size, 3, height, width))]
        self.provide_label = [('softmax_label', (self.batch_size, num_label))]

    def __iter__(self):
        for k in range(self.count / self.batch_size):
            data = []
            label = []
            for i in range(self.batch_size):
                num = gen_rand()
                img = self.captcha.generate(num)
                img = np.fromstring(img.getvalue(), dtype='uint8')
                img = cv2.imdecode(img, cv2.IMREAD_COLOR)
                img = cv2.resize(img, (self.width, self.height))
                cv2.imwrite("./tmp" + str(i % 10) + ".png", img)
                img = np.multiply(img, 1/255.0)
                img = img.transpose(2, 0, 1)
                data.append(img)
                label.append(get_label(num))

            data_all = [mx.nd.array(data)]
            label_all = [mx.nd.array(label)]
            data_names = ['data']
            label_names = ['softmax_label']

            data_batch = OCRBatch(data_names, data_all, label_names, label_all)
            yield data_batch

    def reset(self):
        pass

def get_ocrnet():
    data = mx.symbol.Variable('data')
    label = mx.symbol.Variable('softmax_label')
    conv1 = mx.symbol.Convolution(data=data, kernel=(5,5), num_filter=32)
    pool1 = mx.symbol.Pooling(data=conv1, pool_type="max", kernel=(2,2), stride=(1, 1))
    relu1 = mx.symbol.Activation(data=pool1, act_type="relu")

    conv2 = mx.symbol.Convolution(data=relu1, kernel=(5,5), num_filter=32)
    pool2 = mx.symbol.Pooling(data=conv2, pool_type="avg", kernel=(2,2), stride=(1, 1))
    relu2 = mx.symbol.Activation(data=pool2, act_type="relu")

    conv3 = mx.symbol.Convolution(data=relu2, kernel=(3,3), num_filter=32)
    pool3 = mx.symbol.Pooling(data=conv3, pool_type="avg", kernel=(2,2), stride=(1, 1))
    relu3 = mx.symbol.Activation(data=pool3, act_type="relu")

    flatten = mx.symbol.Flatten(data = relu3)
    fc1 = mx.symbol.FullyConnected(data=flatten, num_hidden=512)
    fc21 = mx.symbol.FullyConnected(data=fc1, num_hidden=10)
    fc22 = mx.symbol.FullyConnected(data=fc1, num_hidden=10)
    fc23 = mx.symbol.FullyConnected(data=fc1, num_hidden=10)
    fc24 = mx.symbol.FullyConnected(data=fc1, num_hidden=10)
    fc2 = mx.symbol.Concat(*[fc21, fc22, fc23, fc24], dim=0)
    label = mx.symbol.transpose(data=label)
    label = mx.symbol.Reshape(data=label, target_shape=(0, ))
    return mx.symbol.SoftmaxOutput(data=fc2, label=label, name="softmax")

def accuracy(label, pred):
    label = label.T.reshape((-1, ))
    hit = 0
    total = 0
    for i in range(pred.shape[0] / 4):
        ok = True
        for j in range(4):
            k = i * 4 + j
            if np.argmax(pred[k]) != int(label[k]):
                ok = False
                break
        if ok:
            hit += 1
        total += 1
    return 1.0 * hit / total

if __name__ == '__main__':
    network = get_ocrnet()
    devs = [mx.cpu()]
    model = mx.model.FeedForward(ctx=devs,
                                 symbol=network,
                                 num_epoch=15,
                                 learning_rate=0.001,
                                 wd=0.00001,
                                 initializer=mx.init.Xavier(factor_type="in", magnitude=2.34),
                                 momentum=0.9)

    data_train = OCRIter(100000, 50, 4, 30, 80)
    data_test = OCRIter(1000, 50, 4, 30, 80)

    import logging
    head = '%(asctime)-15s %(message)s'
    logging.basicConfig(level=logging.DEBUG, format=head)

    model.fit(X=data_train, eval_data=data_test, eval_metric=accuracy, 
              batch_end_callback=mx.callback.Speedometer(32, 50),)

Environment info

Operating System: virtual machine Linux Mint 18[base on Ubuntu 16.04], 64bit Compiler: gcc 5.4.0 Package used (Python/R/Scala/Julia): Python 2.7.12 MXNet version: get from git master[2017-2-23] If you are using python package, please provide opencv 3.2

Error Message:

run in PyCharm

Process finished with exit code 139 (interrupted by signal 11: SIGSEGV)

gdb py-bt

Traceback (most recent call first): File "test-g.py", line 56, in iter img = cv2.imdecode(img, cv2.IMREAD_COLOR) File "/home/mao/mxnet/python/mxnet/model.py", line 236, in _train_multi_device for data_batch in train_data: File "/home/mao/mxnet/python/mxnet/model.py", line 816, in fit sym_gen=self.sym_gen) File "test-g.py", line 135, in model.fit(X = data_train, eval_data = data_test, eval_metric = Accuracy, batch_end_callback=mx.callback.Speedometer(32, 50),)

gdb bt

0 0x0000000000000000 in ?? ()

1 0x00007ffff18149ee in cv::imdecode(cv::_InputArray const&, int) ()

from /usr/lib/x86_64-linux-gnu/libopencv_highgui.so.2.4

2 0x00007fffd9728f92 in pyopencv_cv_imdecode(_object, _object, _object*) ()

from /usr/lib/python2.7/dist-packages/cv2.so

3 0x00000000004c468a in call_function (oparg=,

pp_stack=0x7fffffffd440) at ../Python/ceval.c:4350

4 PyEval_EvalFrameEx () at ../Python/ceval.c:2987

5 0x00000000004dddca in gen_send_ex.isra.0.lto_priv ()

at ../Objects/genobject.c:85

6 0x00000000004c4c6f in PyEval_EvalFrameEx () at ../Python/ceval.c:2806

7 0x00000000004c2765 in PyEval_EvalCodeEx () at ../Python/ceval.c:3582

8 0x00000000004ca099 in fast_function (nk=17, na=,

n=, pp_stack=0x7fffffffd7e0, func=<function at remote 0x7fffd4b0c488>) at ../Python/ceval.c:4445

9 call_function (oparg=, pp_stack=0x7fffffffd7e0)

at ../Python/ceval.c:4370

10 PyEval_EvalFrameEx () at ../Python/ceval.c:2987

11 0x00000000004c2765 in PyEval_EvalCodeEx () at ../Python/ceval.c:3582

12 0x00000000004ca099 in fast_function (nk=4, na=,

n=, pp_stack=0x7fffffffd9f0, func=<function at remote 0x7fffd4b0d0c8>) at ../Python/ceval.c:4445

13 call_function (oparg=, pp_stack=0x7fffffffd9f0)

---Type to continue, or q to quit--- at ../Python/ceval.c:4370

14 PyEval_EvalFrameEx () at ../Python/ceval.c:2987

15 0x00000000004c2765 in PyEval_EvalCodeEx () at ../Python/ceval.c:3582

16 0x00000000004c2509 in PyEval_EvalCode (co=,

globals=, locals=) at ../Python/ceval.c:669

17 0x00000000004f1def in run_mod.lto_priv () at ../Python/pythonrun.c:1376

18 0x00000000004ec652 in PyRun_FileExFlags () at ../Python/pythonrun.c:1362

19 0x00000000004eae31 in PyRun_SimpleFileExFlags ()

at ../Python/pythonrun.c:948

20 0x000000000049e14a in Py_Main () at ../Modules/main.c:640

21 0x00007ffff7811830 in __libc_start_main (main=0x49dab0

, argc=2,

argv=0x7fffffffde38, init=, fini=, rtld_fini=, stack_end=0x7fffffffde28) at ../csu/libc-start.c:291

22 0x000000000049d9d9 in _start ()

What have you tried to solve it?

I find the point of interrupte is img = cv2.imdecode(img, cv2.IMREAD_COLOR). So I split this part to another file, only generate img from captcha module, and decode it by opencv. It can run well. But when I run the example, with mxnet, decode by opencv will be segmentation fault.

What's wrong?

[zht3344]

Do you solve this problem?

[ZHmao]

No. I try my best, but there it is. So I change my platform to Windows.

[baiyancheng20]

I have also met with this problem. Anyone has fixed it?

[szha]

This issue is closed due to lack of activity in the last 90 days. Feel free to ping me to reopen if this is still an active issue. Thanks!

[roccolocko]

The problem happens when using the latest version of opencv, try with a previous one like 3.1.x That solved it for me.

[ZHmao]

Thank you.