quantize¶
Quantize a trained model to reduce its memory footprint.
Additional Documentation¶
Usage¶
Usage: mltk quantize [OPTIONS] <model>
Quantize a model into a .tflite file
The model is automatically quantized after training completes.
This is useful if the mltk_model.tflite_converter parameters
are modified after the model is trained.
For more details see:
https://siliconlabs.github.io/mltk/docs/guides/model_quantization
----------
Examples
----------
# Quantize the previously trained model
# and update its associated model archive audio_example1.mltk.zip
# with the generated .tflite model file
mltk quantize audio_example1
# Generate a .tflite in the current directory from the given model archive
mltk quantize audio_example1.mltk.zip --output .
# Generate a .tflite from the given model python script
# The .tflite is generated in the same directory as the Python script
mltk quantize my_model.py --build
Arguments
* model <model> One of the following:
- Name of MLTK model
- Path to trained model's archive (.mltk.zip)
- Path to MLTK model's python script
[default: None]
[required]
Options
--verbose -v Enable verbose
console logs
--output -o <path> One of the following:
- Path to generated
output .tflite file
- Directory where
output .tflite is
generated
- If omitted, .tflite
is generated in the
MLTK model's log
directory and the
model archive is
updated
[default: None]
--build -b Build the Keras model
rather than loading
from a pre-trained
.h5 file in the MLTK
model's archive.
This is useful if a
.tflite needs to be
generated to view its
structure
--weights -w <value> Optional, load
weights from previous
training session.
May be one of the
following:
- If option omitted
then quantize using
output .h5 from
training
- Absolute path to a
generated weights .h5
file generated by
Keras during training
- The keyword `best`;
find the best weights
in <model log
dir>/train/weights
- Filename of .h5 in
<model log
dir>/train/weights
[default: None]
--update-archive --no-update-archive Update the model
archive with the
quantized model
[default:
no-update-archive]
--help Show this message and
exit.