Installing theseus with pip install theseus-ai with a specific version of PyTorch

[DanielTakeshi]

From reading the README, I am trying to install theseus using pip install but with a specific version of PyTorch (for example PyTorch 1.13.0). For that, I followed these commands:

conda create --name theseus python=3.8 -y && conda activate theseus
conda install ipython jupyter -y
conda install pytorch==1.13.0 torchvision==0.14.0 torchaudio==0.13.0 pytorch-cuda=11.6 -c pytorch -c nvidia -y
pip install theseus-ai

You can find the raw pastebin with the output here:

https://pastebin.com/raw/jFEXvawV

However I notice that pip install theseus-ai will override my older torch version to use version 2.0.1, which is the latest version as of today. Is there a reason for this, out of curiosity? is it related to this deprecation of older torch? https://github.com/facebookresearch/theseus/pull/525/files

[luisenp]

HI @DanielTakeshi. It's sort of related, but not a consequence of that PR. A few months ago we made this the behavior of our wheels uploaded to pypi, after torch 1.13 deprecated some older CUDA versions, and in anticipation of eventually moving to torch 2.0.

If you want a more custom install, you could try installing from source, or also try modifying this script, which is the one we use to build our wheels. In line 67 you can see the version is set to ">=1.13", unless you are using CUDA 10.2 or 11.3. Changing that line to equality and running the script should generate what you need. You can then run as

./build_scripts/build_wheel.sh ROOT_DIR 0.1.4 11.6

And then pip install from the generated wheel that's saved under ROOT_DIR.

Let me know if you have any questions.

[DanielTakeshi]

Thanks! I think that answers my question.

[DanielTakeshi]

Update: just following up that I was able to get this working with PyTorch 1.13. The change that I had to do to the build_wheel.sh script was in 2 places (python 3.8 and torch version):

(theseus) seita@starship:~/theseus (main) $ git diff
diff --git a/build_scripts/build_wheel.sh b/build_scripts/build_wheel.sh
index c969b31..140013b 100755
--- a/build_scripts/build_wheel.sh
+++ b/build_scripts/build_wheel.sh
@@ -37,7 +37,7 @@ CUDA_VERSION=$3
 TH_VERSION=${4:-${COMMIT}}
 INCLUDE_LABS=${5:-0}
 NIGHTLY=${6:-0}
-PY_VER=${7:-3.10}
+PY_VER=${7:-3.8}
 
 if [[ ${INCLUDE_LABS} != 0 ]]
 then
@@ -66,7 +66,7 @@ CUDA_VERSION_IS_SUPPORTED=$(echo "cpu ${SUPPORTED_CUDA_VERSIONS}" | grep -w ${CU
 
 CUDA_SUFFIX=$(echo ${CUDA_VERSION} | sed 's/[.]//g')
 
-TORCH_VERSION='"torch>=1.13"'
+TORCH_VERSION='"torch==1.13"'
 if [[ ${CUDA_VERSION} == "cpu" ]]
 then 
     DEVICE_TAG=cpu
(theseus) seita@starship:~/theseus (main) $ 

The command to build the script:

./build_scripts/build_wheel.sh ROOT_DIR 0.2.1 11.6

Then create and install stuff:

conda create --name theseus python=3.8 -y && conda activate theseus
conda install pytorch==1.13.0 torchvision==0.14.0 torchaudio==0.13.0 pytorch-cuda=11.6 -c pytorch -c nvidia -y
pip install ROOT_DIR/theseus_docker_3.8/theseus_ai-0.2.1+cu116-cp38-cp38-manylinux_2_17_x86_64.whl

After that, I can do the following and it works:

(theseus) seita@starship:~/theseus (main) $ python
Python 3.8.17 (default, Jul  5 2023, 21:04:15) 
[GCC 11.2.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch, theseus
/home/seita/theseus/theseus/_version.py:30: FutureWarning: Using torch < 2.0 for theseus is deprecated and compatibility will be discontinued in future releases.
  warnings.warn(
>>> torch.__version__
'1.13.0'
>>> theseus.__version__
'0.2.1'
>>> exit()
(theseus) seita@starship:~/theseus (main) $ 

[luisenp]

Thanks for sharing! Would you be interested in submiting a PR adding a CLI flag to specify the torch version to install (e.g. "==1.13", ">=1.13")? The python version can already be specified via CLI, so no need to add one for that.