Tensorflow: How does tensorflow connect python and C++ in details? -

-

what i'm trying do

i'm trying make modification conv_ops in tensorflow filters used convolution in consist of binary numbers.

what have done

following suggestion in this tensorflow issue, didn't create custom op. instead, copy code in conv_ops.cc , change names from

register_kernel_builder( name("conv2d").device(device_gpu).typeconstraint<eigen::half>("t"), conv2dop<gpudevice, eigen::half>); register_kernel_builder( name("conv2d").device(device_gpu).typeconstraint<float>("t"), conv2dop<gpudevice, float>);

to

register_kernel_builder( name("binaryconv2d").device(device_gpu).typeconstraint<eigen::half>("t"), binaryconv2dop<gpudevice, eigen::half>); register_kernel_builder( name("binaryconv2d").device(device_gpu).typeconstraint<float>("t"), binaryconv2dop<gpudevice, float>);

after modification, have make new op same conv_ops. in order bypass problem functor:sign in //tensorflow/core/kernels/cwise_ops.h not easy use simple functors functor:relu in //tensorflow/core/kernels/relu_op_functor.h. decide add input(binarized_filter) new op binaryconv2d. in way can use op tf.sign safely make binary filter.

what problem is

although have changed registration in nn_ops.cc as

register_op("binaryconv2d")     .input("input: t")     .input("filter: t")     .input("binarized_filter: t")     .output("output: t")     .attr("t: {half, float, double}")     .attr("strides: list(int)")     .attr("use_cudnn_on_gpu: bool = true")     .attr(getpaddingattrstring())     .attr(getconvnetdataformatattrstring())

and generated python code becomes:

result = _op_def_lib.apply_op("binaryconv2d", input=input, filter=filter,                                 binarized_filter=binarized_filter,                                 strides=strides, padding=padding,                                 use_cudnn_on_gpu=use_cudnn_on_gpu,                                 data_format=data_format, name=name)

i use sign op inside binarized_filter:

  import tensorflow tf   binarized_filter = tf.sign(filter)

i got error as:

w tensorflow/core/framework/op_kernel.cc:926] invalid argument: signature mismatch, have: float, float, float->float expected: float, float->float e tensorflow/core/common_runtime/executor.cc:334] executor failed create kernel. invalid argument: signature mismatch, have: float, float, float->float expected: float, float->float      [[node: conv1/binaryconv2d = binaryconv2d[t=dt_float, data_format="nhwc", padding="same", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true, _device="/job:localhost/replica:0/task:0/cpu:0"](shuffle_batch, conv1/weights/read, conv1/sign)]] e tensorflow/core/client/tensor_c_api.cc:485] signature mismatch, have: float, float, float->float expected: float, float->float      [[node: conv1/binaryconv2d = binaryconv2d[t=dt_float, data_format="nhwc", padding="same", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true, _device="/job:localhost/replica:0/task:0/cpu:0"](shuffle_batch, conv1/weights/read, conv1/sign)]] traceback (most recent call last):

i have checked value of input types in graph in //tensorflow/python/framework/op_def_library.py. right. assume input number defined in c++ still not correct, can find no example or document except source code can provide information how c++ code , python code of tensorflow associated. want know if have missed part of building op in procedure of modification.


Comments

Post a Comment

Popular posts from this blog

Combining PHP Registration and Login into one class with multiple functions in one PHP file -

-
i'm trying combine registration, activation , login scripts php website backend script front end developer can pass variables different forms to. my question whether appropriate approach this. don't want have lot of php files different pieces of application developing. far have written following 2 functions login, register , activate user front end developer can call: <?php /** * created phpstorm. * user: karl * date: 26/07/2016 * time: 02:25 */ class users { function register_user($email, $password, $user_name) { $server_name = "localhost"; $u_name = "root"; $db_password = "root"; $db_name = "betamath_graspe"; //email notification variable $from_address="info@slack.com"; //registration form $msg_reg_user='username taken. please choose different username'; $msg_reg_email='email registered'; $msg_reg_active=...
Read more

magento2 - Magento 2 admin grid add filter to collection -

-
i created grid in admin xml ui components. need filter collection via url parameter , dont know how achieve that. tried inject requestinterface collection, filter didnt work. di.xml <virtualtype name="slidelistingdataprovider" type="magento\framework\view\element\uicomponent\dataprovider\dataprovider"> <arguments> <argument name="collection" xsi:type="object" shared="false">xxx\xxx\model\resourcemodel\grid\slide\collection</argument> <argument name="filterpool" xsi:type="object" shared="false">slidelistingfilterpool</argument> <!-- define new object filters --> </arguments> </virtualtype> <virtualtype name="slidelistingfilterpool" type="magento\framework\view\element\uicomponent\dataprovider\filterpool"> <arguments> <argument name="applier...
Read more

Android volley - avoid multiple requests of the same kind to the server? -

-
i using volley upload list of images server. happens within service. once image uploaded remove list. problem arises when internet connection breaks. when internet connection breaks , comes again thinking of adding images yet uploaded request queue. this can result in more 1 copy of same image saved on server request same image might have gone through before getting rseponse. how can address scenario? i'm not quite sure why you're using service schedule requests volley automatically runs it's request in seperate thread. nevertheless can listen both request results, successful , unsuccessful adding listeners. you can remove image list schedule request using volley. if succeeds can carry on desired, if fails can add again list. public static request getimageuploadrequest(final string image) { response.listener<t> responselistener = new response.listener<t>() { @override public void onresponse(t response...
Read more