PyTorch YoloV5 Model is not getting compiled with Neo.

[amitmukh]

IHAC who has some use cases where they want to build an object detection model using Pytorch YoloV5. They were able to build the model but not able to compile through Neo. here is the error we are getting :

"ClientError: InputConfiguration: TVM cannot convert the PyTorch model. Invalid model or input-shape mismatch. Make sure that inputs are lexically ordered and of the correct dimensionality. Traceback (most recent call last):\n 2: TVMFuncCall\n 1: std::_Function_handler<void (tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue), tvm::runtime::TypedPackedFunc<tvm::relay::Tuple (tvm::runtime::Array<tvm::RelayExpr, void>, tvm::Span)>::AssignTypedLambda<tvm::relay::{lambda(tvm::runtime::Array<tvm::RelayExpr, void>, tvm::Span)#5}>(tvm::relay::{lambda(tvm::runtime::Array<tvm::RelayExpr, void>, tvm::Span)#5}, std::__cxx11::basic_string<char, std::char_traits, std::allocator >)::{lambda(tvm::runtime::TVMArgs const&, tvm::runtime::TVMRetValue)#1}>::_M_invoke(std::_Anydata const&, tvm::runtime::TVMArgs&&, tvm::runtime::TVMRetValue*&&)\n 0: tvm::runtime::TVMMovableArgValueWithContext::operator tvm::runtime::Array<tvm::RelayExpr, void><tvm::runtime::Array<tvm::RelayExpr, void> >() const\n 3: TVMFuncCall\n 2: std::_Function_handler<void (tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue), tvm::runtime::TypedPackedFunc<tvm::relay::Tuple (tvm::runtime::Array<tvm::RelayExpr, void>, tvm::Span)>::AssignTypedLambda<tvm::relay::{lambda(tvm::runtime::Array<tvm::RelayExpr, void>, tvm::Span)#5}>(tvm::relay::{lambda(tvm::runtime::Array<tvm::RelayExpr, void>, tvm::Span)#5}, std::__cxx11::basic_string<char, std::char_traits, std::allocator >)::{lambda(tvm::runtime::TVMArgs const&, tvm::runtime::TVMRetValue)#1}>::_M_invoke(std::_Anydata const&, tvm::runtime::TVMArgs&&, tvm::runtime::TVMRetValue*&&)\n 1: tvm::runtime::TVMMovableArgValueWithContext::operator tvm::runtime::Array<tvm::RelayExpr, void><tvm::runtime::Array<tvm::RelayExpr, void> >() const\n 0: tvm::runtime::Array<tvm::RelayExpr, void> tvm::runtime::TVMPODValue_::AsObjectRef<tvm::runtime::Array<tvm::RelayExpr, void> >() const\n File "/tvm/include/tvm/runtime/packed_func.h", line 714\nTVMError: In function relay.ir.Tuple: error while converting argument 0: [09:36:18] /tvm/include/tvm/runtime/packed_func.h:1591: \n---------------------------------------------------------------\nAn error occurred during the execution of TVM.\nFor more information, please see: https://tvm.apache.org/docs/errors.html\n---------------------------------------------------------------\n Check failed: (!checked_type.defined()) is false: Expected Array[RelayExpr], but got Array[index 1: Array]\n”

[jk1333]

I'm not sure this guide also applies to YoloV5, but I am able to run bytetrack/yolox by following this patch https://github.com/mrtj/yolox-panorama-tutorial/blob/main/yolox-torchscript.ipynb

[jk1333]

Here is my approach for neo compile and tested through simulator

1. Use compatible version of torch and vision pip install torch==1.8.1 torchvision==0.9.1

2. Use yolov5 latest release to export and bring pretrained model using yolov5 v4.0 (higher not supported)

3. modify export.py of yolov5 latest version

   import models import torch.nn as nn from utils.activations import Hardswish, SiLU

   In 'def export_torchscript' add following code before 'ts = torch.jit.trace(model, im, strict=False)'

   Update model

   for k, m in model.named_modules():
       m._non_persistent_buffers_set = set()  # pytorch 1.6.0 compatibility
       if isinstance(m, models.common.Conv):  # assign export-friendly activations
           if isinstance(m.act, nn.Hardswish):
               m.act = Hardswish()
           elif isinstance(m.act, nn.SiLU):
               m.act = SiLU()
       # elif isinstance(m, models.yolo.Detect):
       #     m.forward = m.forward_export  # assign forward (optional)
   model.model[-1].export = True  # set Detect() layer export=True

4. Run 'python3 yolov5/export.py --weights ./yolov5m.pt' to create torchscript and change ext to .pth (yolov5m.pt is yolov5 v4.0 pretrained model asset)

5. create tar.gz and do neo compile

6. call yolov5 model and post_process like this (non_max_suppression is in utils/general.py) inference_results = self.call({"input0":image_data}, self.MODEL_NODE)[0] inference_results = non_max_suppression(torch.tensor(inference_results))[0]