search

udacity / CarND-Term1-Starter-Kit

MIT License
494 stars 602 forks source link

Environment file for new GPU compatibility #113

Closed StavrosD closed 5 years ago

StavrosD commented 5 years ago

The environment-gpu.yml creates an environment that is not compatible with the latest GPUs.

I created a new environment file (environment-gpu-2019) that is compatible with the latest GPUs.

Read the environment-gpu-2019-notes-warning.txt for more information and compatibility warnings.

Warning: This can be used for the projects but not for the final trained model that will be uploaded on Carla.

mvirgo commented 5 years ago

Unfortunately, we are not able to support versions of TensorFlow higher than 1.3.0 at this time, which matches the version we use in the classroom. We plan to upgrade this in the coming months.

StavrosD commented 5 years ago

I know that Carla runs TensorFlow 1.3.0 and the final project should be implemented with that version. On the notes I have a warning about that.

There are many students that have a new PC or a new GPU and they could use it to save time with GPU acceleration. The users with new GPUs are also Udacity's potential customers. Someone who have a $1000 GPU is more likely to enroll a $1600 nanodegree than someone who does not have a GPU.

Last year I lost many hours waiting for the training. I had a slow GPU on my laptop and I did not have internet all the time, I could not use AWS. Recently I upgraded to a new PC with 2080 Ti. I planed to use it for the capstone project. I lost three days experimenting with the environment and the system setup just to realize that tensorflow 1.3.0 is not compatible with the new GPUs.

I think it will be good to have this environment, even with a notice that it is unsupported. I just run my submitted projects (except the capstone project) with this environment and everything run fine.

mvirgo commented 5 years ago

All classroom code is implemented with TensorFlow 1.3.0. The only GPU's that can't use this are the RTX GPUs, which a substantially small portion of our student base uses. We can re-visit this at a later time, but for now, since the vast majority of students are able to use this version and have lower end GPU's than late 2018/2019, we will not make this change at this time.