graph.initializers[weights_value_name] KeyError, because ONNX without graph.initializer (inc. Sollution)

[kdirkjan]

The current ONNX2TORCH assumes that all weights are available through the model.graph.initializer, but this is not true for ONNX files generated by PaddleOCR .
I solved this by adding the following code in model.py:

if len(onnx_model.graph.initializer) == 0:
    new_initializers = []
    for node in onnx_model.graph.node:
        if node.op_type == "Constant":
            for attr in node.attribute:
                if attr.name == "value":
                    new_initializers.append(attr.t)
    onnx_model.graph.initializer.extend(new_initializers)

Then the conversion worked flawlessly with onnx2torch (while onnx2pytorch still failed), good job!