Compressing a Model to FP16¶
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
Working with devices.
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)