How to install pytorch 1.9 or below in jetson orin

[Raymond_456]

I have a model that runs successfully on jetson AGX (pytorch 1.6), but it gets an error on jetson orin(pytorch 1.11):

    736 restore_location = _get_restore_location(map_location)
--> 738 class UnpicklerWrapper(pickle_module.Unpickler):  # type: ignore[name-defined]
    740     def find_class(self, mod_name, name):
    741         if type(name) is str and 'Storage' in name:

TypeError: the first argument must be callable

I read the internet saying the solution is to downgrade pytorch

[AastaLLL]

Hi,

To get the previous PyTorch package, you can build it from the source.
Please find the detailed command in the below topic:

PyTorch for Jetson

Below are pre-built PyTorch pip wheel installers for Python on Jetson Nano, Jetson TX1/TX2, Jetson Xavier NX/AGX, and Jetson AGX Orin with JetPack 4.2 and newer. Download one of the PyTorch binaries from below for your version of JetPack, and see the installation instructions to run on your Jetson. These pip wheels are built for ARM aarch64 architecture, so run these commands on your Jetson (not on a host PC). PyTorch pip wheels

PyTorch v1.12.0

JetPack 5.0 (L4T R34.1.0) / JetPack 5.0.1 (L4T …

Thanks.

[Raymond_456]

I followed the tutorial to compile pytorch and it gives the following error

[ 81%] Building NVCC (Device) object caffe2/CMakeFiles/torch_cuda.dir/operators/torch_cuda_generated_depthwise_3x3_conv_op_cudnn.cu.o /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(149): warning: the “visibility” attribute can only appear on functions and variables with external linkage /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(200): warning: the “visibility” attribute can only appear on functions and variables with external linkage /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(236): warning: the “visibility” attribute can only appear on functions and variables with external linkage /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(908): error: namespace “thrust” has no member “host_vector” /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(908): error: expected an expression /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(909): error: namespace “thrust” has no member “host_vector” /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(909): error: expected an expression /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(910): error: namespace “thrust” has no member “host_vector” /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(910): error: type name is not allowed /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(910): error: expected an expression /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(912): error: identifier “A_array” is undefined /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(913): error: identifier “B_array” is undefined /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(914): error: identifier “C_array” is undefined /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(917): error: identifier “A_array” is undefined /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(919): error: identifier “B_array” is undefined /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(920): error: identifier “C_array” is undefined /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(1763): warning: the “visibility” attribute can only appear on functions and variables with external linkage /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(2234): warning: the “visibility” attribute can only appear on functions and variables with external linkage /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(2282): warning: the “visibility” attribute can only appear on functions and variables with external linkage /home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu(2846): warning: the “visibility” attribute can only appear on functions and variables with external linkage 13 errors detected in the compilation of “/home/py/Downloads/pytorch/caffe2/utils/math_gpu.cu”. CMake Error at torch_cuda_generated_math_gpu.cu.o.Release.cmake:281 (message): Error generating file /home/py/Downloads/pytorch/build/caffe2/CMakeFiles/torch_cuda.dir/utils/./torch_cuda_generated_math_gpu.cu.o make[2]: *** [caffe2/CMakeFiles/torch_cuda.dir/build.make:1388: caffe2/CMakeFiles/torch_cuda.dir/utils/torch_cuda_generated_math_gpu.cu.o] Error 1 make[2]: *** Waiting for unfinished jobs… make[1]: *** [CMakeFiles/Makefile2:6966: caffe2/CMakeFiles/torch_cuda.dir/all] Error 2 make: *** [Makefile:141: all] Error 2

[dusty_nv]

Hi @Raymond_456, unfortunately it would seem from your attempt to build this older version of PyTorch for JetPack 5.0, there were some updates required in PyTorch needed to support CUDA 11.4:

PyTorch for Jetson

Hi @Raymond_456, I haven’t tried to build this older version of PyTorch for JetPack 5.0, so JetPack 5.0’s newer CUDA version may have required some updates in PyTorch. I think the first version I built for it was PyTorch 1.11.

So this pickling/unpickling error you get seems to be related to loading/saving the model that came from an older version of PyTorch. It may be worth trying to install a newer version on your Xavier device (or PC) and seeing you can re-save or export it so that the model gets updated.

[Raymond_456]

Means that I need to retrain once on the new pytorch version? Because I try to load the model weights with the new pytorch version, I get an error: “TypeError: the first argument must be callable”

[dusty_nv]

If you aren’t already, can you try using torch’s load_state_dict() function for restoring the model weights?

Aside from the fact of not being able to build older PyTorch versions on JetPack 5.0 due to newer CUDA requirements, this issue you are having of not being able to load older PyTorch checkpoints on newer versions of PyTorch doesn’t seem specific to Jetson - so you might want to try asking on the PyTorch discourse forums on PyTorch github about it.