jyh2986 / active-shift-tf Goto Github PK

how can i use it in pytorch

Since the environment is different as follows I changed some options.
[environment]
CentOS 17.4
tensorflow-gpu 2.0.0
CUDA 10.0
CUDNN7.4.2
gcc8.3.2

By using tf_upgrade_v2, test_forward_ASL.py run and I can the test_forward_ASL.py
However, when I run test_gradient_ASL.py
I firstly met following error. Warning]WARNING:tensorflow:From test_gradient_ASL-TF20.py:45: compute_gradient_error (from tensorflow.python.ops.gradient_checker) is deprecated and will be removed in a future version. Instructions for updating: Use tf.test.compute_gradient in 2.0, which has better support for functions. Note that the two versions have different usage, so code change is needed.

[compile Errror] for code line 45, 73,
Attribute Error err = gradient_checker.compute_gradient_error(a, arr.shape, result, result.get_shape().as_list(), x_init_value=arr)
err = gradient_checker.compute_gradient_error(c, shift.shape,result, result.get_shape().as_list(), x_init_value=shift)
AttributeError: Tensor.graph is meaningless when eager execution is enabled. Thus I changed gradient_checker.compute_gradient_error function like below.

    strides = [1, 1, stride_h, stride_w]
    paddings = [0, 0, pad_h, pad_w]
    strides_tf = tf.constant(strides, dtype=tf.int64)
    paddings_tf = tf.constant(paddings, dtype=tf.int64)

    theoretical, numerical = tf.test.compute_gradient(active_shift2d_op.active_shift2d_op, [a, c, strides_tf, paddings_tf])

However, I met another error
Traceback (most recent call last): File "", line 54, in active_shift2d_op tensorflow.python.eager.core._FallbackException: Expecting int64_t value for attr strides, got tensorflow.python.framework.ops.EagerTensor
TypeError: Expected list for 'strides' argument to 'active_shift2d_op' Op, not .
library file is made by c++, thus it has int64_t data type, however tensorflow doesn't have int64_t type.

How can I resolve the error by removing type difference?

Thank you in advance~

Hi, I have followed your instructions to build the Op successfully. Then I tried to run the test code. And I got the error message:
InvalidArgumentError (see above for traceback): No OpKernel was registered to support Op 'ActiveShift2DOp' with these attrs. Registered devices: [CPU], Registered kernels:
device='GPU'; T in [DT_DOUBLE]
device='GPU'; T in [DT_FLOAT]

Could you help to figure out what is wrong? Thank you!

I tried to run build.sh and I ran into this trouble:
tensorflow/include/tensorflow/core/util/cuda_kernel_helper.h:24:31: fatal error: cuda/include/cuda.h: No such file or directory
#include "cuda/include/cuda.h"

I use tensorflow 1.4.1, cuda 8.0 cudnn 6.0 and g++4.9
Any help?