halflings
commented
2 years ago For the specific use case of pre-processing, it would be great to have ops like the ones in tf transform: https://www.tensorflow.org/tfx/transform/get_started
Normalizing a feature would just require calling something like tft. scale_to_z_score(..) when building the model, and the rest is done more-or-less transparently (no need to manually serialize/deserialize metadata).
Here are some of the most common pre-processing steps I've used (either for MLConsole.com using custom JS code, or at work with TFX):
- Scaling numeric variables (metadata = mean, variance ; or min,max)
- Mapping categorical variables to an embedding ID (metadata = vocab)
- Tokenizing strings. (little to no metadata, but good to have as part of the graph ; usually combined with the mapping to embedding ID)
gaikwadrahul8
jasonmayes
pyu10055
Currently metadata is supported for conversion via the CLI converter on server side from models from Python however if you train a model in the browser there is no way to save custom key value pairs that you may need in future pre/post processing.
For example if you perform normalization on your training data you need to know the min/max values for a given attribute in order to normalize future data after the model is trained too at inference for any new prediction data that comes in.
Currently a user has to save this data in a different pipeline to the TensorFlow.js one which is inconvenient and does not make sense. It would be far more optimal to expose a simple interface to the end user to store their metadata as part of the model.json file and then able to read these values on model load such that they can be retrieved and used as needed for any custom pre/post processing when using the model.
All users training models in the browser either fully or via transfer learning could benefit from this feature.