[Relay][Pass]Do we want to allow LetNode in FoldScaleAxis pass?

#1

I’m compiling a tensorflow model and FoldScaleAxis fails because it doesn’t allow LetNode: https://github.com/apache/incubator-tvm/blob/master/src/relay/transforms/fold_scale_axis.cc#L249

If I comment out that visit function, FoldScaleAxis pass succeeds. Do we really need this check?

@tqchen @zhiics

1 Like
[Pytorch] [Quantization] Error during quantization
#2

That simply means the pass was not designed with Let node in mind, we probably need a careful revisit, and enforce everything to basic block normal form as in Basic Block Normal Form then we can formally support the IR with let

#3

Thanks. Do you think it make sense to change FATAL to WARN for now?

#4

I think we will at least need to investigate into the impl before doing such a switch since it might affect the correctness. Disabling would be safer

#5

I also hit this error after I started using prelude in the torch frontend. If I do infer type with prelude, there is always warning about the presence of free variables, which I think is related to the introduction of LetNode. I wondered if this is a bug in the Torch frontend.

#6

Hit the same error when compiling ssd-mobilenet. Any formal solution now? BTW: I am using TVM 0.7.0.

#7

You can skip this pass when compiling. You can refer to test_forward_ssd in tensorflow tests.

#8

Great! Thanks, @kevinthesun Is there any regular test to track this unit test status?

I delete the other test and keep test_forward_ssd() only. Then I run the unit test in this way

python test_forward.py

It fails with the error message in the frontend:

TypeError: batch_norm() got an unexpected keyword argument 'exponential_avg_factor'
#9

I think the above issue can be fixed by optimize the BN into BiasAdd for inference. But I still got the another error as:

tvm._ffi.base.TVMError: Traceback (most recent call last):
  [bt] (8) /home/tvm/tvm/build/libtvm.so(tvm::relay::StorageAllocaInit::VisitExpr_(tvm::relay::CallNode const*)+0x94) [0x7f0557b778a4]
  [bt] (7) /home/tvm/tvm/build/libtvm.so(tvm::relay::ExprVisitor::VisitExpr(tvm::RelayExpr const&)+0x7b) [0x7f0557bfaa3b]
  [bt] (6) /home/tvm/tvm/build/libtvm.so(tvm::relay::ExprFunctor<void (tvm::RelayExpr const&)>::VisitExpr(tvm::RelayExpr const&)+0x5b) [0x7f0557bbc65b]
  [bt] (5) /home/tvm/tvm/build/libtvm.so(tvm::relay::StorageAllocaInit::VisitExpr_(tvm::relay::CallNode const*)+0x94) [0x7f0557b778a4]
  [bt] (4) /home/tvm/tvm/build/libtvm.so(tvm::relay::ExprVisitor::VisitExpr(tvm::RelayExpr const&)+0x7b) [0x7f0557bfaa3b]
  [bt] (3) /home/tvm/tvm/build/libtvm.so(tvm::relay::ExprFunctor<void (tvm::RelayExpr const&)>::VisitExpr(tvm::RelayExpr const&)+0x5b) [0x7f0557bbc65b]
  [bt] (2) /home/tvm/tvm/build/libtvm.so(tvm::relay::StorageAllocaInit::VisitExpr_(tvm::relay::CallNode const*)+0x1e) [0x7f0557b7782e]
  [bt] (1) /home/tvm/tvm/build/libtvm.so(tvm::relay::StorageAllocaInit::CreateToken(tvm::RelayExprNode const*, bool)+0x61e) [0x7f0557b7716e]
  [bt] (0) /home/tvm/tvm/build/libtvm.so(+0x268a567) [0x7f0557b75567]
  File "/home/tvm/tvm/src/relay/backend/graph_plan_memory.cc", line 160

How can I know which visited op throws this error?