Got error when GraphExecutor instance executes relay backward function


#1

Hi, i am a green hand of using TVM and curious about how to build up a training process with Relay.

With reference to the file: test_op_grad_level1.py, I found ‘Interpreter’ is used to execute relay backward function, so i tried to execute a relay backward function with GraphExecutor, but errors occurred.

So currently I was wondering whether GraphExecutor can not be applied to execute a relay backward function, or something wrong happened in the process.

Looking forward to your reply. Thanks : )


#2

It is possible to execute the backwards function on the graph runtime, but unsupported language features must be evaluated away using the partial evaluation pass. The new VM which will be the preferred execution mechanism: https://github.com/dmlc/tvm/pull/2889 and will execute all programs without optimization.

We will be writing more tutorials on executing Relay models in a training setting in the next few weeks.


#3

It sounds great, thanks for your reply :grinning:.