Failed installed and "Needs linking cache DATA/EL/legacy_linker"

[clement1234567891011]

Hi,

When checking if correctly installed, it says failed installing although it's not very clear to me why. The hint being: "maybe LD_LIBRARY_PATH unconfigured?" (see attached). I've also downloaded manually spacy 2.0.16 thinking from this post that this could help but it did not https://github.com/IBM/transition-amr-parser/issues/13

When running the test file, I also encounter another error: "Needs linking cache DATA/EL/legacy_linker_amr3.0/". Is there a file missing?

Many thanks in advance for any help! Best wishes

[ramon-astudillo]

When checking if correctly installed, it says failed installing although it's not very clear to me why. The hint being: "maybe LD_LIBRARY_PATH unconfigured?" (see attached)

the message is misleading, this just means that pytorch-scatter is not installed or failed to install. Try installing manually as indicated in the README

When running the test file, I also encounter another error: "Needs linking cache DATA/EL/legacy_linker_amr3.0/". Is there a file missing?

See https://github.com/IBM/transition-amr-parser/issues/17 for this

[clement1234567891011]

Thanks. With the file, minimal_test now runs up to a point and then crashes with:

fairseq 0.10.2 is installed. Anything I can do?

Re correctly_installed: it did not help that I manually installed pytorch_scatter and reinstalled pytorch==1.4.0. Still exactly the same message.

Many thanks!

[ramon-astudillo]

Did you follow the README install instructions? I find odd that fairseq_ext is installed inside of you pyenv, should be locally on the folder. Can you post the output of

bash tests/correctly_installed.sh

?

[clement1234567891011]

I did: again, pytorch 1.4, and fairseq 0.10.2. Here is the output, same as before...

pytorch 1.4.0 No CUDA available pytorch-scatter installed maybe LD_LIBRARY_PATH unconfigured?, import torch_scatter.scatter_cuda dies fairseq works [FAILED] installing

[clement1234567891011]

I thought that it might have come from me using python 3.8. So, I installed python 3.6, reinstalled the 2 packages pytorch and fairseq, and I still get exactly the same error messages both on the correctly_installed.sh and minimal_test.sh

Also, here is what I get when trying to install pytorch-scatter

Any other ideas? Many thanks in advance!

[clement1234567891011]

Did you follow the README install instructions? I find odd that fairseq_ext is installed inside of you pyenv, should be locally on the folder. Can you post the output of

bash tests/correctly_installed.sh

?

So, you think that I should install it differently? Or copy it into another folder? Where to?

[ramon-astudillo]

maybe LD_LIBRARY_PATH unconfigured?, import torch_scatter.scatter_cuda dies

you can safely ignore this, since you are installing for GPU. This means that the install was successful. We will change this.

I think I found the problem, please try to install again (new virtualenv) but this time install as editable.

pip install --editable .

[clement1234567891011]

Thanks, it now goes much further but still shoots up up yet another error message about torch not being compiled with CUDA. I'm on OS X 11.6.2, so not sure whether I can install CUDA at all. Is there a way to run it without CUDA?

.

[logicReasoner]

@clement1234567891011 have you managed to run it without CUDA, using the CPU only?

[clement1234567891011]

No, I didn't so far (but may try again later on another non-mac computer where this very issue at least shouldn't crop up)

[ramon-astudillo]

Latest version should run on CPU on OSX