Unable to install "curl" after installing Tensorflow

[John Nagle]

Reported this as a Tensorflow issue:

https://github.com/tensorflow/tensorflow/issues/36107

Linked that to here.

(Yeah, I know, "just use a Docker container on a virtual machine". Which a lot of people do with Tensorflow to avoid stuff like this. I may deploy that way, but I'm trying to develop on desktop right now.)

[John Nagle]

Let me know how that works out. A procedure for installing Rasa, CUDA, and Tensorflow locally without mucking up Ubuntu package management would be a win for a lot of people.

An NVidia 640 isn't really enough to do much with Tensorflow. I can't use it with Rasa. It has only "compute power 3.0", and Rasa needs at least "Compute performance 3.5". (Compute performance here means what features are in the hardware, not speed. Latest GPUs are at 7.5 or so.)

[John Nagle]

Filed as Tensorflow bug: https://github.com/tensorflow/tensorflow/issues/36107

Tensorflow, Ubuntu, and Rasa all have bugs filed now, so they can try to blame each other for breaking "curl" and the package manager.

The real problem is Tensorflow installing their own modded version of "curl" to support Google's remote file service.

[ubfan1]

Right, the GT 640 was to confirm my adapter (expresscard to PCIe) worked with an external GPU before I get a better one. (Worked perfectly, plug in and run). A compute Capability of 3.0 does get me to DNN, which is more than I had before (CC of 2.1, which only allowed CUDA 8.0). It'll probably awhile before I iron out the PCI memory issues on the new card anyway -- moving the buffers out of the lower 4G. The laptop vendor is refusing to acknowledge it's a 64 bit machine, and making nasty BIOS updates squeezing the top of the 4G limit.

[ubfan1]

Got a 4G GTX970, and had no issues with using the expresscard adapter. Its compute capability of 5.2 allowed the installation and running of Deep Neural Networking (DNN) and Tensorflow with CUDA10.1 (at the time). The .deb install stuck in an old Nvidia driver and make all the CUDA packages depend upon it, so an update of the Nvidia driver caused all the cuda files to be deleted. Restored the CUDA files from backup, cleaned up the CUDA package mess that was left, and it did work, but thought I'd try another way (which was easy for CUDA8, but has since gotten more complicated) I did another CUDA install of 10.2, avoiding the package manager by simply downloading and unpacking (repetetdly) the deb file in my chosen location. This worked, allowing the DNN samples to run and at least the Tensorflow "Is-It-Installed?" test to succeed. See my answer to https://askubuntu.com/questions/1256...256583#1256583 for details. The way to use old versions of the compiler or changed tools like curl would be to install the old/changed versions into the cuda/bin directory (or just links to the system-installed versions of old compilers if they are offered). Since the cuda/bin is put at the beginning of the PATH for doing CUDA work, that solves the problem of not wanting to clobber the standard system programs/configuration.