Hi @akelad we are training nlu and core on the cpu right now, but now nlu data has grown up in size and we are using embedding intent classifier and it takes more than 2 hours which is unbearable. And in future we are also planning to add sentiment analyzer that will further explode the training time. So my question is that, there was a bottleneck that Rasa didn’t optimised models for GPUs. Does this bottleneck has gone now. And can we now harness the power of GPUs for training our nlu and core models? Thanks.
We have a couple of machines here with GPU and the speed increase has been significant.
We use RASA 1.4.6. In relation to training times. Here’s a rough comparison:
- 8 x i7 cores - around 6 hours
- 8 x i7 cores + Rtx 1050(4Gb) - estimates show around 45 minutes but it runs out of memory at around 50% of the EmbeddingIntentClassifier training
- 8 x i7 cores + V100 (cloud based) - around 25 minutes
Setting up the GPU was a bit of a pain, because I couldn’t get the latest nvidia drivers to work with tensorflow. We use Tensorflow 1.15.0, nvidia drivers 418.87.01 and CUDA toolkit 10.1.
@akelad what do you suggest? Thanks
@samscudder wow, how much data do you have? When it comes to that size of data a GPU does help. Generally if you don’t have that much data though, since Rasas models are quite shallow, a GPU won’t help much. As for running out of memory, as of version 1.6.0 some of our featurizers now use sparse features (see more info here). So you should consider upgrading.
@noman you can definitely try a GPU and see if that speeds up your training time. Which version of Rasa are you on?
As for your SentimentAnalyzer, that will depend on how it’s implemented
Actually, I don’t think we have that much data…We have two datasets. The smaller one has 273 intents, 356 actions, 551 stories, and 2694 phrases in the nlu. The largest one (with the numbers I mentioned above) has 663 intents, 665 actions, 660 stories, and 5542 phrases in the nlu.
The chatbots are in portuguese (Brazilian).
We are unable to upgrade at the moment, as accuracy drops from around .90 to .40 in versions > 1.6.0. Can’t wait to get it sorted out though… the tests I ran, the training time was incredibly quick.
@samscudder that is quite a bit of data though especially with the amount of intents. It seems like you have on average less than 10 examples per intent though? That’s generally not advisable.
Hm, the accuracy drop happens for version > 1.6.0? That’s worrying - could you tell me what pipeline you’re using for this?
@akelad There is a bug open for this (Loss of confidence in Rasa > 1.6.0 nlu (compared to 1.4.6) · Issue #5004 · RasaHQ/rasa · GitHub)
Here’s my pipeline:
# Configuration for Rasa NLU. # https://rasa.com/docs/rasa/nlu/components/ language: pt pipeline: - name: "SpacyNLP" - name: "SpacyTokenizer" - name: "RegexFeaturizer" - name: "CRFEntityExtractor" - name: "CountVectorsFeaturizer" strip_accents: "unicode" - name: "CountVectorsFeaturizer" strip_accents: "unicode" analyzer: "char_wb" min_ngram: 1 max_ngram: 4 - name: "components.lemmatization.lemma.CustomLemmatization" use_cls_token: false - name: "EmbeddingIntentClassifier" epochs: 200 random_seed: 2614 policies: - name: FormPolicy - name: AugmentedMemoizationPolicy max_history: 1 - name: KerasPolicy random_seed: 2614 batch_size: 32 epochs: 350 - name: MappingPolicy - name: FallbackPolicy nlu_threshold: 0.6 core_threshold: 0.6 fallback_action_name: action_padrao
Our threshold in 1.4.6 was 0.7, but we reduced it to 0.6.
In > 1.6.0, though, I’m getting 0.4-0.5 per intent
@akelad right now we are on 1.3.3 soon we’ll upgrade it to the recent stable version. So you think we should try gpu for intent classification having Embedding intent classifier in the pipeline and evaluate the performance, right? But my question how can i reduce the training time from hours to minutes?
the GPU would potentially speed up performance, not improve performance. As for getting training time from hours to minutes, I think upgrading to the newer Rasa version should help, and potentially using the GPU as well
hi @akelad i have one concern regarding GPU like is it really required to have GPU or not .
it is definitely not required.
hi @akelad , I has spend a hour in training the project rasa-demo(called sara), i feel too slow, do you think the speed is normal? I will be thankful you if you give me suggestion. my rasa version is the newlest.
asking another query related to GPU here.
I have a GTX 1660 ti GPU, rasa model trainings are noticeably faster in GPU compared to CPU. However when I was looking at the GPU usage using the command nvidia-smi, i saw only 24% of GPU is utilised at max.
I was just wondering if this is normal or is there any hyperparameters that I can tune to maximise the usage of available GPU.
Hi @samscudder, how did you utilize GPU with Rasa? I can’t configure the conda environment.
Thanks in advance