feat: 支持用 numpy.ndarray 构造张量

Signed-off-by: YdrMaster <ydrml@hotmail.com>

pyinfinitensor/src/pyinfinitensor/onnx.py

@@ -1,10 +1,15 @@
-import backend
+from numpy import ndarray
+import backend
 from onnx import ModelProto, NodeProto, TensorProto, AttributeProto, numpy_helper
 from onnx.helper import make_model, make_node, make_graph, make_tensor_value_info
 from backend import DimExpr, refactor_tensor, refactor_operator, refactor_graph
 from typing import Any
 
 
+def build_tensor(array: ndarray) -> backend.Tensor:
+    return _parse_tensor(numpy_helper.from_array(array))
+
+
 def build_graph(model: ModelProto) -> backend.Graph:
     edges: dict[str, backend.Tensor] = dict()
 

src/ffi/ffi_infinitensor.cc

@@ -35,7 +35,7 @@ class Handler {
     std::unordered_set<Name> fillEdgeInfo() { return _g.fillEdgeInfo(); }
     void setInput(size_t index, std::shared_ptr<Tensor> tensor) {
         ASSERT(_g.setInput(index, std::move(tensor)),
-               fmt::format("set input {} failed", index));
+               fmt::format("set input {} failed with wrong shape", index));
     }
     void substitute(const char *name, int64_t value) {
         ASSERT(_g.substitute(name, value),
@@ -53,7 +53,7 @@ class Handler {
 #endif
     }
     template <class T> std::vector<T> copyout(size_t i) {
-        return _outputs[i]->copyout<T>();
+        return _outputs.at(i)->copyout<T>();
     }
 };