LuaPrimitiveObject append(const thpp::Tensor<T>& val, LuaRefList& refs,
thpp::SharingMode sharing) {
LuaRefObject ref;
ref.__isset.tensorVal = true;
val.serialize(ref.tensorVal, thpp::ThriftTensorEndianness::NATIVE, sharing);
return appendRef(std::move(ref), refs);
}
LuaObject make(thpp::Tensor<T>& val) {
LuaRefObject ref;
ref.__isset.tensorVal = true;
val.serialize(ref.tensorVal);
LuaObject obj;
obj.refs.push_back(std::move(ref));
obj.value.__isset.refVal = true;
obj.value.refVal = 0;
return obj;
}
PyObjectHandle LuaToPythonConverter::convertTensor(lua_State* L,
thpp::Tensor<T>& tensor,
int numpyType) {
npy_intp zero = 0;
int ndims;
std::unique_ptr<npy_intp[]> dims;
npy_intp* dimsPtr;
std::unique_ptr<npy_intp[]> strides;
// Numpy and Torch disagree on empty tensors. In Torch, an empty tensor
// is a tensor with zero dimensions. In Numpy, a tensor with zero dimensions
// is a scalar (with one element). So we'll convert an empty Torch tensor
// to a 1d Numpy tensor of shape [0]. Also see pushTensor in PythonToLua.cpp.
if (tensor.ndims() != 0) {
ndims = tensor.ndims();
auto tsizes = tensor.sizes();
DCHECK_EQ(tsizes.size(), ndims);
dims.reset(new npy_intp[ndims]);
dimsPtr = dims.get();
std::copy(tsizes.begin(), tsizes.end(), dims.get());
if (!tensor.isContiguous()) {
auto tstrides = tensor.strides();
DCHECK_EQ(tstrides.size(), ndims);
strides.reset(new npy_intp[ndims]);
// Numpy strides use bytes; Torch strides use element counts.
for (int i = 0; i < ndims; ++i) {
strides[i] = tstrides[i] * sizeof(T);
}
}
} else {
ndims = 1;
dimsPtr = &zero;
}
PyObjectHandle obj(PyArray_New(
&PyArray_Type, ndims, dimsPtr, numpyType,
strides.get(), tensor.data(), 0,
NPY_ARRAY_ALIGNED, nullptr));
checkPythonError(obj, L, "create numpy.ndarray of type {}", numpyType);
// Create a PythonStorage object to hold the reference count.
// PyArray_SetBaseObject steals the reference to the base object.
int r = PyArray_SetBaseObject(reinterpret_cast<PyArrayObject*>(obj.get()),
PythonStorage<T>::allocate(
L, tensor.storage()).release());
checkPythonError(r != -1, L, "SetBaseObject on numpy.ndarray");
return obj;
}