Missing pytorch package

[kizziah4]

I am attempting to re-install topaz on a local workstation. Following the recommended installation instructions using Anaconda, I have created a new environment for topaz, but when I run the installation from the tbepler and pytorch channels, I get a PackageNotFoundError for pytorch 0.2.0. I have tried installing this pytorch package alone according to the instructions here (https://pytorch.org/get-started/previous-versions/) with the same result. Seems it is truly missing from the official channel. Has anyone else encountered this?

[tbepler]

What version of Topaz are you using? The newer versions support pytorch 1.0+.

[tbepler]

How are you installing topaz? When I run conda install topaz -c tbepler -c pytorch everything installs normally and it does not try to install pytorch 0.2 (which is too old for the current version of topaz anyway).

[kizziah4]

Using conda 4.3.21, I ran conda create -n topaz python=3.6 (accepting installation of new packages), then conda install topaz -c tbepler -c pytorch, which returned:

	PackageNotFoundError: Dependencies missing in current linux-64 channels:
	  - topaz -> pytorch >=1.0.0 -> libuv >=1.40.0,<2.0a0
	  - topaz -> pytorch >=1.0.0 -> mkl >=2018
	  - topaz -> pytorch >=1.0.0 -> typing_extensions
	  - topaz -> pytorch >=1.0.0 -> dataclasses
	  - topaz -> pytorch >=1.0.0 -> cudatoolkit >=10.0,<10.1
	  - topaz -> pytorch >=1.0.0 -> python_abi 3.9.* *_cp39

	Close matches found; did you mean one of these?

		python_abi: python, ipython

	(and similarly for the other packages)

On a whim, I tried conda install topaz -c tbepler -c pytorch again, and the error changed to:

	PackageNotFoundError: Dependencies missing in current linux-64 channels:
	  - topaz -> numpy >=1.13.3,<2.0a0 -> libgcc-ng >=7.3.0
	  - topaz -> numpy >=1.13.3,<2.0a0 -> libgfortran-ng >=7,<8.0a0
	  - topaz -> numpy >=1.13.3,<2.0a0 -> libopenblas >=0.3.3,<1.0a0
	  - topaz -> numpy >=1.13.3,<2.0a0 -> numpy-base 1.17.0 py37h2f8d375_0 -> python >=3.7,<3.8.0a0
	  - topaz -> numpy >=1.13.3,<2.0a0 -> mkl >=2019.4,<2020.0a0
	  - topaz -> numpy >=1.13.3,<2.0a0 -> mkl-service >=2,<3.0a0
	  - topaz -> numpy >=1.13.3,<2.0a0 -> mkl_fft >=1.0.14,<2.0a0 -> libstdcxx-ng >=7.3.0
	  - topaz -> numpy >=1.13.3,<2.0a0 -> mkl_random >=1.0.4,<2.0a0
	  - topaz -> pytorch 0.2.0.*

	Close matches found; did you mean one of these?

		libgcc-ng: libgcc
		libgfortran-ng: libgfortran
		libopenblas: openblas

	(and similarly for the other packages)

There you can see the first request for pytorch 0.2.0, specifically. A third conda install topaz -c tbepler -c pytorch yielded only:

	PackageNotFoundError: Dependency missing in current linux-64 channels:
	  - topaz -> pytorch 0.2.0.*

I did this a few more times with the missing numpy dependency intermittently returning, but specifying a few different numpy versions. The missing pytorch 0.2.0* was persistent, however. Hence, I thought it must be required, but I could not get it through the pytorch channel even with conda install pytorch=0.2.0 -c pytorch yesterday.

Eventually, I individually installed the missing dependencies from the initial error from the conda-forge or anaconda channels:

conda install -c conda-forge python_abi
conda install -c anaconda cudatoolkit
conda install -c conda-forge dataclasses
conda install -c conda-forge typing_extensions
conda install -c conda-forge libuv
conda install -c anaconda mkl

After that, conda install topaz -c tbepler -c pytorch worked with no issue. I have not confirmed that topaz actually runs yet, but it is installed at least.

[taputechnic]

I have perhaps a related problem. When I ran topaz denoise, I got the following warning:

CudaWarning: module 'torch._C' has no attribute '_cuda_setDevice'
Falling back to CPU.
# using device=1 with cuda=False

I had installed Topaz using conda, and when I do a conda list, I get:

pytorch                   1.10.0              py3.6_cpu_0    pytorch
pytorch-mutex             1.0                         cpu    pytorch

When I googled that warning, I saw suggestions to install pytorch-gpu, which I couldn't get to work in a new environment. (I started with conda install -c conda-forge pytorch-gpu, and then tried to install Topaz.)

I'm not running a big denoising job, and it runs reasonably fast on the CPU, but I'm wasting the GPUs.

[tbepler]

@taputechnic you have the CPU-only version of pytorch installed. You probably need to install pytorch with cudatoolkit=X.X (your CUDA version). See here.

[taputechnic]

Thanks! That was the problem.