[Relay][TFLite] Fix in qnn.conv2d when parameter groups not equal to 1

[padreofthegame]

Bug observed in TFLite frontend when working with quantized conv2d operator with parameter groups != 1

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• cc @leandron, @shingjan _{See #10317 for details}

_{Generated by tvm-bot}

[padreofthegame]

@tvm-bot rerun

[ekalda]

Thanks @padreofthegame for fixing this! Would you mind providing more detail in the commit message body as per our commit message guideline, as in, what is the expected behaviour and what was happening before that bugfix?

[padreofthegame]

Hello @ekalda, thank you for the comment.

I will try to explain my observations on this problem.

Basically, I was working with simple quantized tflite model with conv2D and bias_add layer, which was working fine with parameter groups == 1, but failing for example with groups == 2 with error message:

TVMError: Traceback (most recent call last):
  12: TVMFuncCall
  11: tvm::runtime::PackedFuncObj::Extractor<tvm::runtime::PackedFuncSubObj<tvm::runtime::TypedPackedFunc<tvm::IRModule (tvm::transform::Pass, tvm::IRModule)>::AssignTypedLambda<tvm::transform::{lambda(tvm::transform::Pass, tvm::IRModule)#7}>(tvm::transform::{lambda(tvm::transform::Pass, tvm::IRModule)#7}, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >)::{lambda(tvm::runtime::TVMArgs const&, tvm::runtime::TVMRetValue*)#1}> >::Call(tvm::runtime::PackedFuncObj const*, tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*)
  10: tvm::transform::Pass::operator()(tvm::IRModule) const
  9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
  8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
  7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
  6: tvm::transform::ModulePassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
  5: tvm::runtime::PackedFuncObj::Extractor<tvm::runtime::PackedFuncSubObj<tvm::runtime::TypedPackedFunc<tvm::IRModule (tvm::IRModule, tvm::transform::PassContext)>::AssignTypedLambda<tvm::relay::transform::InferType()::{lambda(tvm::IRModule, tvm::transform::PassContext const&)#1}>(tvm::relay::transform::InferType()::{lambda(tvm::IRModule, tvm::transform::PassContext const&)#1})::{lambda(tvm::runtime::TVMArgs const&, tvm::runtime::TVMRetValue*)#1}> >::Call(tvm::runtime::PackedFuncObj const*, tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*)
  4: tvm::relay::transform::InferType()::{lambda(tvm::IRModule, tvm::transform::PassContext const&)#1}::operator()(tvm::IRModule, tvm::transform::PassContext const&) const [clone .isra.0]
  3: tvm::DiagnosticContext::Render()
  2: tvm::DiagnosticRenderer::Render(tvm::DiagnosticContext const&)
  1: tvm::runtime::PackedFuncObj::Extractor<tvm::runtime::PackedFuncSubObj<tvm::runtime::TypedPackedFunc<void (tvm::DiagnosticContext)>::AssignTypedLambda<tvm::TerminalRenderer(std::ostream&)::{lambda(tvm::DiagnosticContext const&)#1}>(tvm::TerminalRenderer(std::ostream&)::{lambda(tvm::DiagnosticContext const&)#1})::{lambda(tvm::runtime::TVMArgs const&, tvm::runtime::TVMRetValue*)#1}> >::Call(tvm::runtime::PackedFuncObj const*, tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*)
  0: tvm::ReportAt(tvm::DiagnosticContext const&, std::ostream&, tvm::Span const&, tvm::Diagnostic const&)
  File "/home/padre/TVM_full_repo/tvm/src/ir/diagnostic.cc", line 264
TVMError: The source maps are not populated for this module.` 
Please use `tvm.relay.transform.AnnotateSpans` `to attach source maps for error reporting.

Error: The Relay type checker is unable to show the following types match:
  Tensor[(16), float32]
  Tensor[(2), float32]
In particular:
  dimension 0 conflicts: 16 does not match 2.

While further testing I realized that problem occures every time when the parameter groups differ from 1 or depth of input tensor (should be equivalent to depth convolution). When looking deeper into the code I track that problem occurs in qnn/op/convolution.cc, specifically in the part

	AssignType(types[5], DataType::Float(32), weight->shape[i_axis] * weight->shape[o_axis],
               reporter);  // weight_scale

Since the types[5] parameter corresponds to kernel_scale, and is a tensor of length num_kernels, weight->shape[i_axis] * weight->shape[o_axis] will generally differ from num_kernels, thus the error of tensor type matching will occur.

In solution, I just added an additional if statement that will check if the convolution is depthwise (similar to the code in Conv2dRel), and kept the current workflow, while in other case the shape of types[5] will be set according to conv2d requirements.

[ekalda]

Thanks for the explanation, makes sense to me :) What I was asking was to amend the commit message body as I don't think "There was a bug" is particularly informative in the commit history and doesn't really comply with the commit message guidelines. Some version of what you wrote in the explaining comment (without the stack trace) would do :)

[ekalda]

LGTM!

[ekalda]

Thanks @padreofthegame!