when dealing with untrusted data. Based on the security history of these
libraries we consider that it is safe to work with untrusted inputs for PNG,
BMP, GIF, WAV, RAW, RAW\_PADDED, CSV and PROTO formats. All other input formats,
including tensorflow-io should be sandboxed if used to process untrusted data.

For example, if an attacker were to upload a malicious video file, they could

                               tensorflow::DataType dtype, void* data,
                               TFE_CustomDeviceTensorHandleMethods methods)
      : tensorflow::CustomDeviceTensorHandle(context, device, dtype),
        data_(data),
        methods_(methods) {}

  ~CAPICustomDeviceTensorHandle() override { methods_.deallocator(data_); }
  void* DevicePointer() const override { return data_; }
  Status NumDims(int* num_dims) const override {

    }

    auto s = TF_NewStatus();
    auto func = TF_GraphToFunction(graph_.get(), name_.data(), 0, -1, nullptr,
                                   inputs_.size(), inputs_.data(),
                                   graph_outputs.size(), graph_outputs.data(),
                                   nullptr, nullptr, name_.data(), s);
    *f = new GraphFunction(std::move(func->record->fdef()));
    TF_DeleteFunction(func);

*   `tf.data`:
    *   Graduated experimental APIs:
        * [`tf.data.Dataset.ragged_batch`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset/#ragged_batch), which batches elements of `tf.data.Dataset`s into `tf.RaggedTensor`s.
        * [`tf.data.Dataset.sparse_batch`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset/#sparse_batch), which batches elements of `tf.data.Dataset`s into `tf.sparse.SparseTensor`s.

///   1. Opaque plugin private data structures and wrappers for type safety;
///   2. Function tables for plugin functionality;
///   3. Versioning metadata;
///   4. Plugin registration API and the DSO entry point.

#ifdef __cplusplus
extern "C" {
#endif  // __cplusplus

/// SECTION 1. Opaque data structures to hold plugin specific data

      : region_(std::move(region)), ops_(ops) {}

  ~ModularReadOnlyMemoryRegion() override { ops_->cleanup(region_.get()); };

  const void* data() override { return ops_->data(region_.get()); }
  uint64 length() override { return ops_->length(region_.get()); }

 private:
  std::unique_ptr<TF_ReadOnlyMemoryRegion> region_;
  const TF_ReadOnlyMemoryRegionOps* ops_;  // not owned

  }

  void CompareInt32Tensor(const std::vector<int32_t>& expected, TF_Tensor* t) {
    int32_t* data = static_cast<int32_t*>(TF_TensorData(t));
    size_t size = TF_TensorByteSize(t);
    ASSERT_EQ(expected.size() * sizeof(int32_t), size);
    for (int i = 0; i < expected.size(); ++i) {
      ASSERT_EQ(expected[i], data[i]) << "Different data at index " << i;
    }
  }

    ctx.reset(ctx_raw);
  }

  AbstractTensorHandlePtr x;
  {
    AbstractTensorHandle* x_raw = nullptr;
    float data[] = {0., 0., 0., 0., 0., 0., 0., 0};
    int64_t dim_sizes[] = {2, 4};
    Status s = TestTensorHandleWithDims<float, TF_FLOAT>(ctx.get(), data,
                                                         dim_sizes, 2, &x_raw);
    ASSERT_EQ(errors::OK, s.code()) << s.message();
    x.reset(x_raw);

  typedef void (*MemoryReleaser)(void* data, size_t len, void* arg);

  // Create a tensor instance from the given data buffer and description.
  // `memory_releaser` will be called on destruction, and it's responsible for
  // cleaning up the underlying buffer.
  virtual AbstractTensorInterface* CreateTensor(
      DataType dtype, const int64_t* dims, int num_dims, void* data, size_t len,

// op executed under the tape. A separate map (`tensorflow::eager::OpTape`)
// maintains the map from `op_id` to a `OpTapeEntry` which stores the `op_type`,
// inputs and outputs and the gradient function These data structures combined
// allow us to trace the data dependencies between operations and hence compute
// gradients.
//
// `ZerosLike` is not expected to be called and returns a nullptr. The creation