CMeiyi
commented
5 months ago Have you found a solution? I had the same problem.
Open Po-Ting-lin opened 5 months ago
CMeiyi
commented
5 months ago Have you found a solution? I had the same problem.
jiri-kula
commented
5 months ago @Po-Ting-lin I read your log thoroughly and it seems to me that we have the exact same setup, and also the same type of error. The only difference was that I got stuck with the trt_pose_hand project, which is a derivative of the trt_pose.
What fixe the issue for me was that I went through initial dependencies setup again while putting more attention to download correct and not mix up the version of packages.
This is the procedure that I have followed, after which I was able to run the notebook.
wget https://nvidia.box.com/shared/static/fjtbno0vpo676a25cgvuqc1wty0fkkg6.whl -O torch-1.10.0-cp36-cp36m-linux_aarch64.whlsudo apt-get install \
python3-pip \
libopenblas-base \
libopenmpi-dev \
libomp-devpip3 install 'Cython<3'
pip3 install numpy torch-1.10.0-cp36-cp36m-linux_aarch64.whlsudo apt-get install \
libjpeg-dev \
zlib1g-dev \
libpython3-dev \
libopenblas-dev \
libavcodec-dev \
libavformat-dev \
libswscale-devgit clone --branch v0.11.1 https://github.com/pytorch/vision torchvision
cd torchvision
export BUILD_VERSION=v0.11.1 # where 0.x.0 is the torchvision version
pip install 'pillow<7'
python3 setup.py install --userWhich fixed the issue for me. At least I can tell that the code is working and maybe give you a little clue that the source of the problem might be in the environment variables or packages that are installed, not the code of the script itself.
Chang558
commented
2 months ago Thanks, @jiri-kula! I encountered the same issue, but I managed to resolve it after seeing your explanation.
The issue occurs because torch2trt does not support PyTorch 2.x versions. Since I’m using PyTorch 2.x, that’s why the problem arose.
Some suggest downgrading to PyTorch 1.x to resolve the issue, but I’m unable to do that since I’m working on another project that requires PyTorch 2.x. So, instead of using torch2trt, I converted the model directly into a TensorRT engine.
I hope this solution can help others who are facing similar issues.
This is the code I used:
print('-----------------------Model Load Part------------------')
WIDTH=224 HEIGHT=224
data=torch.zeros((1,3, HEIGHT, WIDTH)).cuda()
torch.onnx.export( model, data, "trt-pose.onnx", export_params=True, opset_version=11, do_constant_folding=True, input_names=['input'], output_names=['output'], dynamic_axes={'input': {0: 'batch_size'}, 'output':{0:'batch_size'}} )
TRT_LOGGER = trt.Logger(trt.Logger.WARNING) builder = trt.Builder(TRT_LOGGER) network = builder.create_network(1<<int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH))
parser = trt.OnnxParser(network, TRT_LOGGER)
with open("trt-pose.onnx",'rb') as mf: if not parser.parse(mf.read()): print("ERROR: Failed to parse the ONNX file.") for error in range(parser.num_errors): print(parser.get_error(error)) exit()
config = builder.create_builder_config() config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, 1 << 35)
if builder.platform_has_fast_fp16: config.set_flag(trt.BuilderFlag.FP16)
profile = builder.create_optimization_profile() input_tensor_name ='input' profile.set_shape(input_tensor_name, min=(1, 3, 224, 224), opt=(8, 3, 224, 224), max=(32, 3, 224, 224)) config.add_optimization_profile(profile) engine = builder.build_engine(network, config)
Hi Guys,
I tried to run this repo on Jetson nano, and I did the same steps in the instruction. Somehow it failed to convert model to TRT. It seems the dimension is not correct?
Is anyone know why, Thanks!