A smarter, faster and stronger solution for neural network optimization
While everyone says AI is the future, we’ve surged ahead and say optimized AI models are the future.
That’s why we developed AIminify, a pioneering solution that refines complex decision trees, ensuring that AI models are not only suitable for a wide range of applications, from edge devices to high-performance cloud infrastructures, but environmentally friendly too, as AIminfy reduces the carbon footprint associated with training and deploying large models.
At AIminify, we believe that small steps can have a big impact. We are a young, international team whose journey began as a spinoff of Tinify, the image compression industry leader.
When sustainability
meets innovation
Sustainable
We firmly believe that a win-win situation for mankind doesn’t have to come at the expense of our planet.
Accessible
We are committed to simplifying AI compression technology while driving a positive environmental change.
User-friendly
AIminify is user-friendly, blazingly fast, and highly deployable, all with minimal loss of quality and a reduced carbon footprint.
Making complex
AI models accessible
As AI technology evolves, larger and more complex models are developed. Without optimization, these models could become unwieldy and difficult to use in the real world. Our idea was to simplify neural network compression so that anyone can compress an AI model with just a few clicks.
With AIminify, AI professionals can save the time and money required for neural network compression, as this intelligent solution will handle all the work for them. Let AIminify take care of this repetitive and time-consuming task while you concentrate on growing your business.
AIminify uses three-model compression methods to refine complex decision trees and significantly reduce model size while the decrease of probability remains acceptable:
quantization, pruning and knowledge distillation.
Frequently Asked Questions
Why should I optimize my neural network?
Running and optimizing a neural network can be a costly operation. When a network has been optimized with AIminify’s compression, the operating cost should become lower due to a lower memory and CPU profile.
How much can I save by optimizing?
There are many factors that have influence on your cost savings. However, we see a general 30-50% reduction in average CPU and GPU utilization.
Will the optimization affect my neural network?
Yes, but in a good way! We ever so slightly change your network and apply various compression techniques which could change the way your network behaves. This does not mean your network will behave completely differently, but rather that there is a slightly increased chance for the network to produce more similar results.
What kind of neural networks are suitable?
We accept any PyTorch, Tensorflow, or Onnx models. However, Onnx models have a structure that doesn’t let us get the most out of the compression, so we recommend using our service before you create your Onnx models to get the most out of our optimization.
Can I adjust the level of optimization?
We are experts at our trade, so you can trust us with the optimization level (like with Tinify!). We don’t allow optimization adjustments because we know what we’re doing, and we will ensure the best (dynamic) optimization for your trained models.
What is quantization, and how will it affect my neural network?
To quantize means to restrict the number of possible values or states of a system. Quantization in neural networks or machine learning is a way to make your neural network use less memory by changing the precision of certain calculations just a tiny bit. For example, doing a multiplication by 1.2345 instead of 1.23456789.
Can I reverse this process?
No, once the network is optimized, we do not keep track of which things we threw away. We don’t keep any copies of your network, so we can’t restore how your model looked beforehand. We encourage good MLOps practices to make sure you always have a copy of a uncompressed network in case you need to compare strange outputs.
How long will you store my data?
Depending on the type of network, the compression techniques that are applied and customer specific requirements (if any) we will keep the files as long as necessary and until the date agreed with the customer. In the simplest form of network and compression technique applied, no files will be kept after the customer has acknowledged receipt of the compressed model.
Who has access to my information?
We operate our company the same way we operate Tinify. Our capable team (also behind Tinify) takes GDPR and privacy regulations seriously, and we promise never to send, sell, or otherwise use your data other than internally to further improve our services.
Will my data be safe?
Yes, your data will be totally safe! We only perform calculations on your model but never look into it or otherwise use your data. We do not store your neural network a second longer than necessary to perform compression and performance operations. Besides that, we never use your training or other data. Just your neural network.