HI I’m new to Rasa, and I’m taking a course for beginners… I’m training my first model but I have notice that Rasa seems don’t use GPU accellerator…?? is right? probably Rasa run shallow models for NLU training or core training? thanks
Hi! This depends on what kind of configuration you have. Some components (
DIET, for example) benefit from GPU acceleration (provided you have a GPU, and it can be used by
tensorflow), whereas other components (
RegexEntityExtractor) do not. Some models are shallow, and others can be configured to be deeper, but for many applications you don’t require a deep model. If you want to check whether the GPU is available to
tensorflow you can run:
import tensorflow as tf print("Num GPUs Available: ", len(tf.config.experimental.list_physical_devices('GPU')))
HI thanks and sorry for cross posting (in other section)…
"the problem is that for a older TF versions,like 2.3 that Rasa use, we need a specific cuda, specific cudnn e visual studio…you can’t install cuda on virtual env, then You need to handle different versions on your system…this isn’t very efficient way to go…strange that Rasa works with this version…also because, probably, when for ex, You fine tuning an external transformer model, go with transfer learning, or any other external models, will be "probably"compatibily issue when it will be used in Rasa…(probably…I don’t know still Rasa framework, but I work daily with transformers, and RNNs…)
thanks for your support
Hi, not sure I understand the above. TF 2.3 is compatible with CUDA 10. You should be able to manage multiple CUDA versions on one system. You can also compile a tensorflow 2.3 that is compatible with other versions to CUDA.
strange that Rasa works with this version
We are currently working on updating our version, more on this here.
"probably"compatibily issue when it will be used in Rasa
Not sure what this part means, what models are you trying to use with Rasa?
Blockquote TF 2.3 is compatible with CUDA 10. You should be able to manage multiple CUDA versions on one system. You can also compile a tensorflow 2.3 that is compatible with other versions to CUDA.
oh, yes…it is simply a matter of to ease the workflow
Blockquote Not sure what this part means, what models are you trying to use with Rasa?
None at the moment…I’m very noob of Rasa…but i can’t wait to try , but if I have, for example, a roberta model trained on TF 2.6…and I try to load a checkpoint on enviroment that works with TF 2.3 I could have (or not) some issue of compatibly…“potentially backwards incompatible changes”…
thanks for your useful clarification! and sorry my poor english…
Sorry for my slow response!
I see. You’re right that you’d probably run into compatibility issues, sadly. We are working on upgrading our TF version and hope to be able to do so soon.
The good news is that for most beginner and small to medium projects you shouldn’t need a GPU, we aim to make our training lightweight, and try to design systems so that you don’t need a huge amount of training data.
As for pre-trained models, you can load various pre-trained Hugging Face Transformers models, including RoBERTa (see here). I hope that helps!
HI! Thank You very much for your infos and support! that helps me surely !
Sure! Feel free to post again if you run into issues
hi. i have same problem . please help me! tanks