Block Floating Point (BFP) Example

Block Floating Point (BFP) Example#

Note

For information on accessing Quark PyTorch examples, refer to Accessing PyTorch Examples. This example and the relevant files are available at /onnx/accuracy_improvement/BFP.

This is an example of quantizing a mobilenetv2_050.lamb_in1k model using the ONNX quantizer of Quark with BFP16. Int8 quantization performs poorly on the model, but BFP16 and ADAQUANT can significantly mitigate the quantization loss.

Block Floating Point (BFP) quantization computational complexity by grouping numbers to share a common exponent, preserving accuracy efficiently. BFP has both reduced storage requirements and high quantization precision.

Pip requirements#

Install the necessary python packages:

python -m pip install -r ../utils/requirements.txt

Prepare model#

Export onnx model from mobilenetv2_050.lamb_in1k torch model. The corresponding model link is https://huggingface.co/timm/mobilenetv2_050.lamb_in1k:

mkdir models && python ../utils/export_onnx.py mobilenetv2_050.lamb_in1k

Prepare data#

ILSVRC 2012, commonly known as ‘ImageNet’. This dataset provides access to ImageNet (ILSVRC) 2012 which is the most commonly used subset of ImageNet. This dataset spans 1000 object classes and contains 50,000 validation images.

If you already have an ImageNet datasets, you can directly use your dataset path.

To prepare the test data, please check the download section of the main website: https://huggingface.co/datasets/imagenet-1k/tree/main/data. You need to register and download val_images.tar.gz.

Then, create the validation dataset and calibration dataset:

mkdir val_data && tar -xzf val_images.tar.gz -C val_data
python ../utils/prepare_data.py val_data calib_data

The storage format of the val_data of the ImageNet dataset organized as follows:

The storage format of the calib_data of the ImageNet dataset organized as follows:

BFP16 Quantization#

The quantizer takes the float model and produce a BFP16 quantized model.

python quantize_model.py --model_name mobilenetv2_050.lamb_in1k \
                         --input_model_path models/mobilenetv2_050.lamb_in1k.onnx \
                         --output_model_path models/mobilenetv2_050.lamb_in1k_quantized.onnx \
                         --calibration_dataset_path calib_data \
                         --config BFP16

This command will generate a BFP16 quantized model under the models folder, which was quantized by BFP16 configuration.

BFP16 Quantization with ADAQUANT#

The quantizer takes the float model and produce a BFP16 quantized model with ADAQUANT.

Note: If the model has dynamic shapes, you need to convert the model to fixed shapes before performing ADAQUANT.

python -m  quark.onnx.tools.convert_dynamic_to_fixed  --fix_shapes 'input:[1,3,224,224]' models/mobilenetv2_050.lamb_in1k.onnx  models/mobilenetv2_050.lamb_in1k_fix.onnx

python quantize_model.py --model_name mobilenetv2_050.lamb_in1k \
                         --input_model_path models/mobilenetv2_050.lamb_in1k_fix.onnx \
                         --output_model_path models/mobilenetv2_050.lamb_in1k_adaquant_quantized.onnx \
                         --calibration_dataset_path calib_data \
                         --config  BFP16_ADAQUANT

This command will generate a BFP16 quantized model under the models folder, which was quantized by BFP16 configuration with ADAQUANT.

Evaluation#

Test the accuracy of the float model on ImageNet val dataset:

python ../utils/onnx_validate.py val_data --model-name mobilenetv2_050.lamb_in1k --batch-size 1 --onnx-input models/mobilenetv2_050.lamb_in1k.onnx

Test the accuracy of the BFP16 quantized model on ImageNet val dataset:

python ../utils/onnx_validate_with_custom_op.py val_data --model-name mobilenetv2_050.lamb_in1k --batch-size 1 --onnx-input models/mobilenetv2_050.lamb_in1k_quantized.onnx

If want to run faster with GPU support, you can also execute the following command:

python ../utils/onnx_validate_with_custom_op.py val_data --model-name mobilenetv2_050.lamb_in1k --batch-size 1 --onnx-input models/mobilenetv2_050.lamb_in1k_quantized.onnx --gpu

Test the accuracy of the BFP16 quantized model with ADAQUANT on ImageNet val dataset:

python ../utils/onnx_validate_with_custom_op.py val_data --model-name mobilenetv2_050.lamb_in1k --batch-size 1 --onnx-input models/mobilenetv2_050.lamb_in1k_adaquant_quantized.onnx

If want to run faster with GPU support, you can also execute the following command:

python ../utils/onnx_validate_with_custom_op.py val_data --model-name mobilenetv2_050.lamb_in1k --batch-size 1 --onnx-input models/mobilenetv2_050.lamb_in1k_adaquant_quantized.onnx --gpu

Quantization Results#

Note

Different machine models can lead to minor variations in the accuracy of quantized model with adaquant.