MACE supports two kinds of quantization mechanisms, i.e.,
- Quantization-aware training (Recommend)
After pre-training model using float point, insert simulated quantization operations into the model. Fine tune the new model. Refer to Tensorflow quantization-aware training.
- Post-training quantization
After pre-training model using float point, estimate output range of each activation layer using sample inputs.
quantize_weights and quantize_nodes should not be specified when using TransformGraph tool if using MACE quantization.
It is recommended that developers fine tune the fixed-point model, as experiments show that by this way accuracy could be improved, especially for lightweight models, e.g., MobileNet. The only thing you need to make it run using MACE is to add the following config to model yaml file:
- input_ranges: the ranges of model's inputs, e.g., -1.0,1.0.
- quantize: set quantize to be 1.
MACE supports post-training quantization if you want to take a chance to quantize model directly without fine tuning. This method requires developer to calculate tensor range of each activation layer statistically using sample inputs. MACE provides tools to do statistics with following steps (using inception-v3 from MACE Model Zoo as an example):
1. Convert original model to run on CPU host without obfuscation (by setting target_abis to host, runtime to cpu, and obfuscate to 0, appending :0 to output_tensors if missing in yaml config).# For CMake users: python tools/python/convert.py --config ../mace-models/inception-v3/inception-v3.yml --quantize_stat # For Bazel users: python tools/converter.py convert --config ../mace-models/inception-v3/inception-v3.yml --quantize_stat
2. Log tensor range of each activation layer by inferring several samples on CPU host. Sample inputs should be representative to calculate the ranges of each layer properly.# Convert images to input tensors for MACE, see image_to_tensor.py for more arguments. python tools/image/image_to_tensor.py --input /path/to/directory/of/input/images --output_dir /path/to/directory/of/input/tensors --image_shape=299,299,3 # Rename input tensors to start with input tensor name(to differentiate multiple # inputs of a model), input tensor name is what you specified as "input_tensors" # in yaml config. For example, "input" is the input tensor name of InceptionV3 as below. rename 's/^/input/' * # Run with input tensors # For CMake users: python tools/python/run_model.py --config ../mace-models/inception-v3/inception-v3.yml --quantize_stat --input_dir /path/to/directory/of/input/tensors > range_log # For Bazel users: python tools/converter.py run --config ../mace-models/inception-v3/inception-v3.yml --quantize_stat --input_dir /path/to/directory/of/input/tensors > range_log
3. Calculate overall range of each activation layer. You may specify --percentile or --enhance and --enhance_ratio to try different ranges and see which is better. Experimentation shows that the default percentile and enhance_ratio works fine for several common models.python tools/python/quantize/quantize_stat.py --log_file range_log > overall_range
4. Convert quantized model (by setting target_abis to the final target abis, e.g., armeabi-v7a, quantize to 1 and quantize_range_file to the overall_range file path in yaml config).
As quantization-aware training is still evolving, there are some operations that are not supported, which leaves some activation layers without tensor range. In this case, post-training quantization can be used to calculate these missing ranges. To mix the usage, just get a quantization-aware training model and then go through all the steps of post-training quantization. MACE will use the tensor ranges from the overall_range file of post-training quantization if the ranges are missing from the quantization-aware training model.
MACE supports running quantized models on ARM CPU and other acceleration devices, e.g., Qualcomm Hexagon DSP, MediaTek APU. ARM CPU is ubiquitous, which can speed up most of edge devices. However, AI specialized devices may run much faster than ARM CPU, and in the meantime consume much lower power. Headers and libraries of these devices can be found in third_party directory.
To run models on ARM CPU, users should
- Set runtime in yaml config to cpu (Armv8.2+dotproduct instructions will be used automatically if detected by getauxval, which can greatly improve convolution/gemm performance).
To run models on Hexagon DSP, users should
- Set runtime in yaml config to dsp.
- Make sure SOCs of the phone is manufactured by Qualcomm and has HVX supported.
- Make sure the phone disables secure boot (once enabled, cannot be reversed, so you probably can only get that type phones from manufacturers). This can be checked by executing the following command.
adb shell getprop ro.boot.secureboot
The return value should be 0.
- Root the phone.
- Sign the phone by using testsig provided by Qualcomm. (Download Qualcomm Hexagon SDK first, plugin the phone to PC, run scripts/testsig.py)
- Push third_party/nnlib/v6x/libhexagon_nn_skel.so to /system/vendor/lib/rfsa/adsp/. You can check docs/feature_matrix.html in Hexagon SDK to make sure which version to use.
Then, there you go, you can run Mace on Hexagon DSP. This indeed seems like a whole lot of work to do. Well, the good news is that starting in the SM8150 family(some devices with old firmware may still not work), signature-free dynamic module offload is enabled on cDSP. So, steps 2-4 can be skipped. This can be achieved by calling SetHexagonToUnsignedPD() before creating MACE engine.
- To run models on MediaTek APU, users should
- Set runtime in yaml config to apu.
- Make sure SOCs of the phone is manufactured by MediaTek and has APU supported.
- Push third_party/apu/mtxxxx/libapu-platform.so to /vendor/lib64/.