Post-training Optimization Tool Frequently Asked Questions¶
If your question is not covered below, use the OpenVINO™ Community Forum page, where you can participate freely.
I have successfully quantized my model with a low accuracy drop and improved performance but the output video generated from the low precision model is much worse than from the full precision model. What could be the root cause?
What do the messages “Output name: <result_operation_name> not found” or “Output node with <result_operation_name> is not found in graph” mean?
Is the Post-training Optimization Tool (POT) opensourced?¶
Yes, POT is developed on GitHub as a part of https://github.com/openvinotoolkit/openvino under Apache-2.0 License.
Can I quantize my model without a dataset?¶
In general, you should have a dataset. The dataset should be annotated if you want to validate the accuracy. If your dataset is not annotated, you can use Default Quantization to quantize the model or command-line interface with Simplified mode.
Can a model in any framework be quantized by the POT?¶
The POT accepts models in the OpenVINO Intermediate Representation (IR) format only. For that you need to convert your model to the IR format using Model Optimizer.
I’d like to quantize a model and I’ve converted it to IR but I don’t have the Accuracy Checker config. What can I do?¶
Try quantization using Python* API of the Post-training Optimization Tool. For more details see [Default Quantization](Quantizing Model.
If you consider command-line usage only refer to Accuracy Checker documentation to create the Accuracy Checker configuration file, and try to find the configuration file for your model among the ones available in the Accuracy Checker examples.
An alternative way is to quantize the model in the Simplified mode but you will not be able to measure the accuracy.
What is a tradeoff when you go to low precision?¶
The tradeoff is between the accuracy drop and performance. When a model is in low precision, it is usually performed compared to the same model in full precision but the accuracy might be worse. You can find some benchmarking results in INT8 vs FP32 Comparison on Select Networks and Platforms. The other benefit of having a model in low precision is its smaller size.
I tried all recommendations from “Post-Training Optimization Best Practices” but either have a high accuracy drop or bad performance after quantization. What else can I do?¶
First of all, you should validate the POT compression pipeline you are running, which can be done with the following steps:
Make sure the accuracy of the original uncompressed model has the value you expect. Run your POT pipeline with an empty compression config and evaluate the resulting model metric. Compare this uncompressed model accuracy metric value with your reference.
Run your compression pipeline with a single compression algorithm (Default Quantization or Accuracy-aware Quantization) without any parameter values specified in the config (except for
stat_subset_size). Make sure you get the desirable accuracy drop/performance gain in this case.
Finally, if you have done the steps above and the problem persists, you could try to compress your model using the Neural Network Compression Framework (NNCF). Note that NNCF usage requires you to have a PyTorch or TensorFlow 2 based training pipeline of your model to perform Quantization-aware Training. See Model Optimization Guide for more details.
I get “RuntimeError: Cannot get memory” and “RuntimeError: Output data was not allocated” when I quantize my model by the POT.¶
These issues happen due to insufficient available amount of memory for statistics collection during the quantization process of a huge model or due to a very high resolution of input images in the quantization dataset. If you do not have a possibility to increase your RAM size, one of the following options can help:
inplace_statisticparameters to “True”. In that case the POT will change method collect statistics and use less memory. Note that such change might increase time required for quantization.
stat_requests_numberparameters to 1. In that case the POT will limit the number of infer requests by 1 and use less memory. Note that such change might increase time required for quantization.
false. In that case the POT will switch from the FastBiasCorrection algorithm to the full BiasCorrection algorithm which is usually more accurate and takes more time but requires less memory. See Post-Training Optimization Best Practices for more details.
Reshape your model to a lower resolution and resize the size of images in the dataset. Note that such change might impact the accuracy.
I have successfully quantized my model with a low accuracy drop and improved performance but the output video generated from the low precision model is much worse than from the full precision model. What could be the root cause?¶
It can happen due to the following reasons:
A wrong or not representative dataset was used during the quantization and accuracy validation. Please make sure that your data and labels are correct and they sufficiently reflect the use case.
If the command-line interface was used for quantization, a wrong Accuracy Checker configuration file could lead to this problem. Refer to Accuracy Checker documentation for more information.
The quantization process of my model takes a lot of time. Can it be decreased somehow?¶
Quantization time depends on multiple factors such as the size of the model and the dataset. It also depends on the algorithm: the Default Quantization algorithm takes less time than the [Accuracy-aware Quantization algorithm. The following configuration parameters also impact the quantization time duration (see details in Post-Training Optimization Best Practices):
use_fast_bias: when set to
false, it increases the quantization time
stat_subset_size: the higher the value of this parameter, the more time will be required for the quantization
tune_hyperparams: if set to
truewhen the AccuracyAwareQuantization algorithm is used, it increases the quantization time
stat_requests_number: the lower number, the more time might be required for the quantization
eval_requests_number: the lower number, the more time might be required for the quantization Note that higher values of
eval_requests_numberincrease memory consumption by POT.
When I execute POT CLI, I get “File “/workspace/venv/lib/python3.6/site-packages/nevergrad/optimization/base.py”, line 35… SyntaxError: invalid syntax”. What is wrong?¶
This error is reported when you have a Python version older than 3.6 in your environment. Upgrade your Python version.
What does a message “ModuleNotFoundError: No module named ‘some_module_name’” mean?¶
It means that some required python module is not installed in your environment. To install it, run
pip install some_module_name.
Is there a way to collect an intermidiate IR when the AccuracyAware mechanism fails?¶
You can add
"dump_intermediate_model": true to the POT configuration file and it will drop an intermidiate IR to
What do the messages “Output name: <result_operation_name> not found” or “Output node with <result_operation_name> is not found in graph” mean?¶
Errors are caused by missing output nodes names in a graph when using the POT tool for model quantization. It might appear for some models only for IRs converted from ONNX models using new frontend (which is the default conversion path starting from 2022.1 release). To avoid such errors, use legacy MO frontend to convert a model to IR by passing the use_legacy_frontend option. Then, use the produced IR for quantization.