Hi, I’m using two ways to create executor for a PyTorch resnet18 model.
tvm_model = relay.create_executor(“graph”, mod, dev, target).evaluate()
tvm_model = graph_executor.GraphModule(lib[ " default " ] ( dev ) )
The result of graph_executor is correct, but if I use relay.create_executor(“graph”), the accuracy drops to almost 0. However, if I use relay.create_executor(“graph”) to create executor for a quantized resnet18 model, the result is still correct. I’m confused about this. What’s the difference between relay.create_executor(“graph”) and graph_executor.GraphModule?
Thanks so much for your time!