[LEGACY] Compressing a Model to FP16#

Danger

The code described here has been deprecated! Do not use it to avoid working with a legacy solution. It will be kept for some time to ensure backwards compatibility, but you should not use it in contemporary applications.

This guide describes a deprecated conversion method. The guide on the new and recommended method can be found in the Conversion Parameters article.

By default, when IR is saved all relevant floating-point weights are compressed to FP16 data type during model conversion. It results in creating a “compressed FP16 model”, which occupies about half of the original space in the file system. The compression may introduce a minor drop in accuracy, but it is negligible for most models. In case if accuracy drop is significant user can disable compression explicitly.

To disable compression, use the compress_to_fp16=False option:

from openvino.runtime import save_model
ov_model = save_model(INPUT_MODEL, compress_to_fp16=False)
mo --input_model INPUT_MODEL --compress_to_fp16=False

For details on how plugins handle compressed FP16 models, see Inference Devices and Modes.

Note

FP16 compression is sometimes used as the initial step for INT8 quantization. Refer to the Post-training optimization guide for more information about that.

Note

Some large models (larger than a few GB) when compressed to FP16 may consume an overly large amount of RAM on the loading phase of the inference. If that is the case for your model, try to convert it without compression: convert_model(INPUT_MODEL, compress_to_fp16=False) or convert_model(INPUT_MODEL)