There is a bug in save_inference_model and prune when the input program is initialized by load_inference_model
Created by: Xreki
In some cases, users may need to use fluid.io.load_inference_model to get an inference_program, do some transformation to the inference_program, then use fluid.io.save_inference_model to save the transformed inference_program. However, the fluid.io.save_inference_model fails at the following code:
The reason is: t is a Variable and t.op is None, the use of t.block.idx will fail. We should handle the case when t.op is None.
In addition, because the input porgram is an inference program, there are operators which are marked as target, and there are feed_ops and fetch_ops. The following codes will insert feed_ops and fetch_ops to the program, which makes there are redundant feed_ops and fetch_ops in the resulted program. So we need to remove the original feed_ops and fetch_ops in the input program first.