-
InfluxDB
Power Real-Time Data Analytics at Scale. Get real-time insights from all types of time series data with InfluxDB. Ingest, query, and analyze billions of data points in real-time with unbounded cardinality.
In dfdx backprop, there's no fixed set of operations, and operation is just a Box> https://github.com/coreylowman/dfdx/blob/main/src/gradients.rs#L37. dfdx backprop used to be similar with the enum for operation types, but supporting batching with that is really hard. Now every tensor operation in dfdx defines the backward op as a closure in the function itself. That also means every operation's backward op can be optimized specifically for that function.
There's other differences in how nn layers are implemented if you compare the source of linear layers: https://github.com/coreylowman/dfdx/blob/main/src/nn/linear.rs vs https://github.com/c0dearm/mushin/blob/main/src/nn/layers/linear.rs
Incredible work! I’ve messed around with making DL libraries in rust for my startup. https://github.com/Sidekick-AI/condor is my most recent attempt, which basically wraps tch for speed and provides const generic support. It’s really early and not super robust, so I’m excited to see other efforts in this area.
I’ve made a library for just that: https://github.com/Sidekick-AI/dataflow