void GetTensor(const string& name,
                 std::unique_ptr<tensorflow::Tensor>* out_tensor,
                 TF_Status* out_status) const;

 private:
  // Uses "v2_reader_" to build "var name -> shape" and "var name -> data type"
  // maps; both owned by caller.
  // REQUIRES: "v2_reader_ != nullptr && v2_reader_.status().ok()".
  std::pair<std::unique_ptr<TensorSliceReader::VarToShapeMap>,

  TFE_TensorDebugInfo* debug_info = TFE_TensorHandleTensorDebugInfo(h, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  ASSERT_EQ(2, TFE_TensorDebugInfoOnDeviceNumDims(debug_info));
  // Shape is the same for CPU tensors.
  EXPECT_EQ(3, TFE_TensorDebugInfoOnDeviceDim(debug_info, 0));
  EXPECT_EQ(2, TFE_TensorDebugInfoOnDeviceDim(debug_info, 1));

  TFE_DeleteTensorDebugInfo(debug_info);
  TFE_DeleteTensorHandle(h);

// per device in `underlying_devices`. Implicit copies off of the device throw
// an error.
//
// All component tensors must have the same dtype. Currently they must also have
// the same shape, although this requirement may be relaxed in the future.
//
// `device_name` must not name an existing physical or custom device (see
// the documentation for TFE_RegisterCustomDevice for more information).
//

    python3 -m pip install "$TF_WHEEL"
}

@test "TensorFlow is importable" {
    source /tf/venv/bin/activate
    python3 -c 'import tensorflow as tf; t1=tf.constant([1,2,3,4]); t2=tf.constant([5,6,7,8]); print(tf.add(t1,t2).shape)'
}

# Is this still useful?
@test "TensorFlow has Keras" {
    source /tf/venv/bin/activate
    python3 -c 'import sys; import tensorflow as tf; sys.exit(0 if "keras" in tf.keras.__name__ else 1)'

inspected and debugged and it is intended to be used during the development
phase.

As part of the differences that make Eager mode easier to debug, the [shape
inference
functions](https://www.tensorflow.org/guide/create_op#define_the_op_interface)
are skipped, and any checks implemented inside the shape inference code are not
executed.

The security impact of skipping those checks should be low, since the attack

    *   Start enforcing input shape assumptions when calling Functional API
        Keras models. This may potentially break some users, in case there is a
        mismatch between the shape used when creating `Input` objects in a
        Functional model, and the shape of the data passed to that model. You
        can fix this mismatch by either calling the model with correctly-shaped

  // Create the variable handle.
  TFE_Op* op = TFE_NewOp(ctx, "VarHandleOp", status);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_OpSetAttrType(op, "dtype", TF_FLOAT);
  TFE_OpSetAttrShape(op, "shape", {}, 0, status);
  TFE_OpSetAttrString(op, "container", "localhost", 0);
  TFE_OpSetAttrString(op, "shared_name", "", 0);
  if (!device_name.empty()) {
    TFE_OpSetDevice(op, device_name.c_str(), status);
  }

 public:
  // Add a function parameter and return the corresponding tensor.
  virtual absl::Status AddParameter(DataType dtype,
                                    const PartialTensorShape& shape,
                                    TracingTensorHandle**) = 0;

  // Finalize this context and make a function out of it. The context is in a
  // invalid state after this call and must be destroyed.

  memcpy(theta_data.data(), TF_TensorData(theta_tensor),
         TF_TensorByteSize(theta_tensor));

  // Initialize space for the numerical gradient.
  vector<float> dtheta_approx(num_elems);

  // Get theta shape and store in theta_dims.
  int num_dims = TF_NumDims(theta_tensor);
  vector<int64_t> theta_dims(num_dims);
  GetDims(theta_tensor, theta_dims.data());

  // Initialize auxilary data structures.

TF_Operation* Neg(TF_Operation* n, TF_Graph* graph, TF_Status* s,
                  const char* name = "neg");

TF_Operation* LessThan(TF_Output l, TF_Output r, TF_Graph* graph, TF_Status* s);

TF_Operation* RandomUniform(TF_Operation* shape, TF_DataType dtype,
                            TF_Graph* graph, TF_Status* s);

// Split `input` along the first dimension into 3 tensors
TF_Operation* Split3(TF_Operation* input, TF_Graph* graph, TF_Status* s,