Feature Drafts

OpenVINO is an Intel optimization software that allows models to get highly increased performance (we have seen up to 55x speed increases) during inference. It would become an export option in the tool.

See the weight and activation distributions over time in each component, similar to how tensorboard does it:

See the loss landscape and how it changes over time. Reference: https://proceedings.neurips.cc/paper/2018/file/a41b3bb3e6b050b6c9067c67f663b915-Paper.pdf https://github.com/tomgoldstein/loss-landscape

By providing a pre-processing alternative that allows you to convert any datatype into a timeseries, and has the option to enter how long the timeseries should be through an entrybox.

This feature would let you choose a model, then choose which layer you wanted features from and finally let you run through some data to output the those features in a CSV file. Requires batch inference: https://perceptilabs.canny.io/feature-requests/p/use-a-model-inside-perceptilabs-batch-inference

https://github.com/SysCV/bdd100k-models

So that any saved component can be shared to other people for re-use in their models. This requires "Saving components" ( https://perceptilabs.canny.io/feature-requests/p/saving-components ) first though.

Add a "Analyze data" button in the Data Wizard which generates a report using pandas profiling ( https://github.com/pandas-profiling/pandas-profiling ). The report will likely open in a new tab to start with, but might be embedded in a model as the feature progress.