kevinLu1114 opened a new issue #8495: URL: https://github.com/apache/tvm/issues/8495
> Description When I run the model(ssd_mobilenetv1), I found that the answer was wrong, and finally found that the problem was in [Expand ](https://github.com/apache/tvm/blob/d3db5d65c9aefacda128756c15c7ec3f0a7b91ea/python/tvm/relay/frontend/onnx.py#L1982) operator, so I wrote a small script to reproduce the problem. In my case, my comment out code line was resolved. https://github.com/apache/tvm/blob/d3db5d65c9aefacda128756c15c7ec3f0a7b91ea/python/tvm/relay/frontend/onnx.py#L2038 > Environment - tvm version: d3db5d65c9aefacda128756c15c7ec3f0a7b91ea - onnxruntime version: 1.8.1 > Script ```python import tvm import onnx import numpy as np from tvm import relay from tvm.runtime.vm import VirtualMachine from tvm.testing import assert_allclose from onnx import TensorProto, helper from onnxruntime.backend import prepare # Create the outer network graph_proto = helper.make_graph( [ helper.make_node("Expand", ["input", "shape"], ["output"]), ], "graph", [ helper.make_tensor_value_info('shape', TensorProto.INT64, [1]), ], [ helper.make_tensor_value_info("output", TensorProto.INT32, ['?']), ], initializer= [ helper.make_tensor( 'input', TensorProto.INT32, dims=[], vals=[0], ) ], ) model = helper.make_model(graph_proto) onnx.checker.check_model(model) i_shape = np.array([0], dtype=np.int64) shape_list = {'shape' : i_shape.shape} mod, params = relay.frontend.from_onnx(model, shape_list, freeze_params=True) target = "llvm" with tvm.transform.PassContext(opt_level=3): vm_exec = relay.vm.compile(mod, target=target, params=params) dev = tvm.cpu() vm = VirtualMachine(vm_exec, dev) tvm_out = vm.run(tvm.nd.array(i_shape)) model_rep = prepare(model) ort_out = model_rep.run(i_shape)[0] assert_allclose(ort_out, tvm_out) ``` Expected output : > `[]` Error output: > `[0]` -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
