blog page - 1/2¶
x / y != x * (1 / y)¶
Custom C++ TopK¶
It started with the fact the python runtime for the AdaBoostRegressor was quite slow. I noticed three operators were quite slow even though their implementation was based on numpy: TopK, ArrayFeatureExtractor and GatherElement. I made a custom implementation of the first two.
RandomForestClassifier - prediction for one observation¶
I was meeting with Olivier Grisel this morning and we were wondering why scikit-learn was slow to compute the prediction of a random forest for one observation compare to what onnxruntime does, and more specically some optimized C++ code inspired from onnxruntime. We used py-spy and wrote the following script:
The bug which makes you waste time¶
It is not a bug but it is something which makes
you waste some significant time just to understand
what’s going on. asv would refuse to detect
the benchmark I was trying to set up just because
filenames did contain dots. So, for asv
don’t add a file
name.option.py but use
A couple of benchmark for tries:
Notebooks Pairwise distances with ONNX (pdist) shows how much slower
an ONNX implementation of function
cdist, from 3 to 10 times slower.
One way to optimize the converted model is to
create dedicated operator such as one for function
cdist. Tutorial Converters with options
explains how to tell function
to_onnx to use
the custom operator CDist.
Float, double with ONNX¶
Replicating what a library does, scikit-learn for example, is different from implementing a function defined in a paper. Every trick needs to be replicated. scikit-learn trees implement a prediction function which takes float features and compares them to double thresholds. Knowning the ONNX assumes that comparison only happens numbers of the same type, you end up with discrepencies.
The python runtime is now almost complete for all the supported numerical operator implemented in sklearn-onnx. A couple of notebooks introduces a couple of way to investigates issues, to benchmark ONNX models with onnxruntime or python runtime, to check the differences between the same model. It also extend ONNX with operators not in the specification to experiment some assumptions and check it is more efficient. Notebook Precision loss due to float32 conversion with ONNX introduces a way to guess the margins introduced by the conversion from double to single. There also exists a function to convert numpy function into ONNX (see From numpy to ONNX). Its coverage is probably low but it will improve.
ONNX, runtime, converters¶
I have been recently working on sklearn-onnx to write converter from scikit-learn operators to ONNX serialization format. I was talking about that a month ago and somebody asked me if there was a runtime implemented in RUST. Not that I know of but I said it would not be too complex to implement one.