Use the following steps to run vai_q_caffe.
- Prepare the Neural Network Model.
Table 1. vai_q_caffe Input Files No. Name Description 1 float.prototxt Floating-point model for ResNet-50. The data layer in the prototxt should be consistent with the path of the calibration dataset. 2 float.caffemodel Pre-trained weights file for ResNet-50. 3 calibration dataset A subset of the training set containing 100 to 1000 images.
Before running vai_q_caffe, prepare the Caffe model in floating-point format with the calibration data set, including the following:
Figure 1. Sample Caffe Layer for QuantizationNote: Three mean_value parameters for channels are recommended. If three mean_value parameters are specified, following the order BGR.
- Caffe floating-point network model prototxt file.
- Pre-trained Caffe floating-point network model caffemodel file.
- Calibration data set. The calibration set is usually a
subset of the training set or actual application images (at least 100
images). Make sure to set the source and root_folder in image_data_param
to the actual calibration image list and image folder path. For quantize
calibration, calibration data without a label is enough. But due to the
implementation, an image list file with two columns is required. Just
set the second column to a random value or zero. This is an example of
n01440764_985.JPEG 0 n01443537_9347.JPEG 0 n01484850_8799.JPEG 0 ...
- Run vai_q_caffe to generate a quantized
Because there are small differences between hardware platforms, the output formats of
vai_q_caffe quantize -model float.prototxt -weights float.caffemodel [options]
vai_q_caffeare also different. If the target hardware platform is DPUCAHX8H, the
-keep_fixed_neuronoption should be added to the command.
vai_q_caffe quantize -model float.prototxt -weights float.caffemodel -keep_fixed_neuron[options]
- After successful execution of the above command, four files are generated in
the output directory (default directory, /quantize_results/). The deploy.prototxt and deploy.caffemodel files are used as input files to the compiler.
The quantize_train_test.prototxt and
are used to test the accuracy on the GPU/CPU, and can be used as input files for
quantization aware training.
Table 2. vai_q_caffe Output Files No. Name Description 1 deploy.prototxt For Vitis AI compiler, quantized network description file. 2 deploy.caffemodel For Vitis AI compiler, quantized Caffe model parameter file (non-standard Caffe format). 3 quantize_train_test.prototxt For testing and finetuning, quantized network description file. 4 quantize_train_test.caffemodel For testing and finetuning, quantized Caffe model parameter file (non-standard Caffe format).To evaluate the accuracy of the quantized model, use a command similar to the following, or add an
-auto_testin step 2. Refer to the next section for vai_q_caffe argument details.
vai_q_caffe test -model ./quantize_results/quantize_train_test.prototxt -weights ./quantize_results/quantize_train_test.caffemodel -gpu 0 -test_iter 1000