Hyperparameter tuning of custom components #10447
-
Hello!I am currently working on a project for building a custom entity based sentiment analyzer. I have built and trained my custom components on a set of data. I would now like to tune the hyperparameters to get the best set of parameters for the mode. I understand that we can use Weights and Biases for building a sweep to tune hyperparamters. However, I am a bit confused about which are the hyperparameters to be tuned for both NER component and a textcat model. I have gone through spacy docs but would appreciate any more inputs. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 4 replies
-
Hello, |
Beta Was this translation helpful? Give feedback.
Hello,
Yes, you can combine the train and dev dataset to train custom components. In terms of what the most important parameters to change are: It really depends on the data you're working with. We're also currently working on a FAQ for this particular topic to help users get a better feeling/understanding about hyperparameters/sections in general. We don't have a detailed explanation about every hyperparameter for every pipeline in our docs but if you're interested you can look into the NER and Textcat component code to get a more detailed insight into how the parameters influence the model/training and decide which suits best for your use-case based on that.