Hi all, I came across a confusing problem, the compiled module has a different inference with onnxruntime.
TVM predict result:
ONNX predict result:
Script for reproduce this bug:
import onnx
from tvm import relay
from tvm.contrib import graph_runtime
import tvm
import numpy as np
onnx_model_path = "model_seq.onnx"
model = onnx.load(onnx_model_path)
irmod, params = relay.frontend.from_onnx(model, {'X': [2, 3, 4], 'Y': [2, 3, 4], 'Z': [2, 3, 4]}, freeze_params=True)
graph, lib, params = relay.build(irmod, target='llvm', params=params)
input_data_x = np.full([2, 3, 4], 1, dtype='float32')
input_data_y = np.full([2, 3, 4], 0, dtype='float32')
input_data_z = np.full([2, 3, 4], 2, dtype='float32')
module = graph_runtime.create(graph, lib, tvm.cpu(0))
module.set_input('X', input_data_x)
module.set_input('Y', input_data_y)
module.set_input('Z', input_data_z)
module.set_input(**params)
module.run()
outputs_dict = {'out': [2, 3, 4]}
res_tvm = module.get_output(0, tvm.nd.empty([2, 3, 4])).asnumpy()
print("tvm predict result: ", res_tvm)
# ---------------------------------------------------------------------------
import onnxruntime as rt
sess = rt.InferenceSession(onnx_model_path)
res_frame = sess.run(['out'], {'X': input_data_x, 'Y': input_data_y, 'Z': input_data_z})
print("onnx predict result:", res_frame)
related model used in the script:
you can receive the simplest onnx model by this link: model_seq.onnx