Getting started with ONNX optimization
In this section, we will learn about the 4 main steps needed to optimize your ONNX models:
- Input your model and data
- Run the optimization
- Save your optimized model
- Load and run your optimized model in production
1) Input model and data
In order to optimize a model with
Speedster, first you should input the model you want to optimize and load some sample data that will be needed to test the optimization performances (latency, throughput, accuracy loss, etc).
import numpy as np # Load a resnet as example # Model was downloaded from here: # https://github.com/onnx/models/tree/main/vision/classification/resnet model = "resnet50-v1-12.onnx" # Provide input data for the model input_data = [((np.random.randn(1, 3, 224, 224).astype(np.float32), ), np.array()) for _ in range(100)]
Now your input model and data are ready, you can move on to Run the optimization section 🚀.
2) Run the optimization
input_data have been defined, everything is ready to use Speedster's
optimize_model function to optimize your model.
The function takes the following arguments as inputs:
model: model to be optimized in your preferred framework (ONNX in this case)
input_data: sample data needed to test the optimization performances (latency, throughput, accuracy loss, etc)
optimization_time: if "constrained" mode,
Speedstertakes advantage only of compilers and precision reduction techniques, such as quantization. "unconstrained" optimization_time allows it to exploit more time-consuming techniques, such as pruning and distillation
metric_drop_ths: maximum drop in your preferred accuracy metric that you are willing to trade to gain in acceleration
and returns the accelerated version of your model 🚀.
from speedster import optimize_model # Run Speedster optimization optimized_model = optimize_model( model, input_data=input_data, optimization_time="constrained", metric_drop_ths=0.05 )
Speedster tries to use all the compilers and optimization techniques at its disposal along the software to hardware stack to optimize the model. From these, it will choose the ones with the lowest latency on the specific hardware.
At the end of the optimization, you are going to see the results in a summary table like the following:
If the speedup you obtained is good enough for your application, you can move to the Save your optimized model section to save your model and use it in production.
If you want to squeeze out even more acceleration out of the model, please see the
optimize_model API section. Consider if in your application you can trade off a little accuracy for much higher performance and use the
optimization_time arguments accordingly.
3) Save your optimized model
After accelerating the model, it can be saved using the
Now you are all set to use your optimized model in production. To explore how to do it, see the Load and run your optimized model in production section.
4) Load and run your optimized model in production
Once the optimized model has been saved, it can be loaded with the
The optimized model can be used for accelerated inference in the same way as the original model:
# Use the accelerated version of your ONNX model in production output = optimized_model(input_sample)
The first 1-2 inferences could be a bit slower than expected because some compilers still perform some optimizations during the first iterations. After this warm-up time, the next ones will be faster than ever.
If you want to know more about how to squeeze out more performances from your models, please visit the Advanced options section.