Custom resource in Tensorflow

问题

For some reasons, I need to implement a custom resource for Tensorflow. I tried to get inspiration from lookup table implementations. If I understood well, I need to implement 3 TF operations:

creation of my resource
initialization of the resource (e.g. populate the hash table in case of the lookup table)
implementation of the find / lookup / query step.

To facilitate the implementation, I'm relying on tensorflow/core/framework/resource_op_kernel.h. I get the following error

[F tensorflow/core/lib/core/refcount.h:90] Check failed: ref_.load() == 0 (1 vs. 0)
1]    29701 abort      python test.py

Here is the full code to reproduce the issue:

using namespace tensorflow;

/** CUSTOM RESOURCE **/
class MyVector : public ResourceBase {
 public:
  string DebugString() override { return "MyVector"; };
 private:
  std::vector<int> vec_;
};

/** CREATE VECTOR **/
REGISTER_OP("CreateMyVector")
    .Attr("container: string = ''")
    .Attr("shared_name: string = ''")
    .Output("resource: resource")
    .SetIsStateful();

class MyVectorOp : public ResourceOpKernel<MyVector> {
 public:
  explicit MyVectorOp(OpKernelConstruction* ctx) : ResourceOpKernel(ctx) {}

 private:
  Status CreateResource(MyVector** resource) override {
    *resource = CHECK_NOTNULL(new MyVector);
    if(*resource == nullptr) {
      return errors::ResourceExhausted("Failed to allocate");
    }
    return Status::OK();
  }

  Status VerifyResource(MyVector* vec) override {
    return Status::OK();
  }
};

REGISTER_KERNEL_BUILDER(Name("CreateMyVector").Device(DEVICE_CPU), MyVectorOp)

and then, after compiling, the error can be reproduced with this Python snippet of code:

test_module = tf.load_op_library('./test.so')
my_vec = test_module.create_my_vector()
with tf.Session() as s:
  s.run(my_vec)

As a side question, I'd be interested in having tutorials / guidelines to implement custom resources. In particular, I'd like to have information about what needs to be implemented for checkpoints / graph export / serialization / etc.

Thanks a lot.

回答1:

Add -DNDEBUG to your build flags. This workaround is explained in TF issue 17316.

来源：https://stackoverflow.com/questions/55958530/custom-resource-in-tensorflow

标签

c++

tensorflow

tensorflow-serving