nn-Meter is a novel and efficient system to accurately predict the inference latency of DNN models on diverse edge devices. The key idea is dividing a whole model inference into kernels, i.e., the execution units of fused operators on a device, and conduct kernel-level prediction. We currently evaluate four popular platforms on a large dataset of 26k models. It achieves 99.0% (mobile CPU), 99.1% (mobile Adreno 640 GPU), 99.0% (mobile Adreno 630 GPU), and 83.4% (Intel VPU) prediction accuracy.

The current supported hardware and inference frameworks:

nn-Meter has achieved the Mobisys 21 Best Paper Award! For more details, please check out paper:

nn-Meter: towards accurate latency prediction of deep-learning model inference on diverse edge devices

• Those who want to get the DNN inference latency on mobile and edge devices with no deployment efforts on real devices.
• Those who want to run hardware-aware NAS with NNI.
• Those who want to build latency predictors for their own devices (Documents of nn-Meter builder).
• Those who want to use the 26k latency benchmark dataset.

Currently nn-Meter has been tested on Linux and Windows system. Windows 10, Ubuntu 16.04 and 20.04 with python 3.6.10 are tested and supported. Please first install python3 before nn-Meter installation. Then nn-Meter could be installed by running:

pip install nn-meter

nn-meter==2.0 has been released now.

If you want to try latest code, please install nn-Meter from source code. First git clone nn-Meter package to local:

git clone [email protected]:microsoft/nn-Meter.git
cd nn-Meter

Then simply run the following pip install in an environment that has python >= 3.6. The command will complete the automatic installation of all necessary dependencies and nn-Meter.

pip install .

nn-Meter is a latency predictor of models with type of Tensorflow, PyTorch, Onnx, nn-meter IR graph and NNI IR graph. To use nn-Meter for specific model type, you also need to install corresponding required packages. The well tested versions are listed below:

[1] Please refer to nn-Meter Usage for more information.

Please also check the versions of numpy and scikit_learn. The different versions may change the prediction accuracy of kernel predictors.

The stable version of wheel binary package will be released soon.

To apply for hardware latency prediction, nn-Meter provides two types of interfaces：

• command line nn-meter after nn-meter installation.
• Python binding provided by the module nn_meter

Here is a summary of supported inputs of the two methods.

In both methods, users could appoint predictor name and version to target a specific hardware platform (device). Currently, nn-Meter supports prediction on the following four configs:

Users can get all predefined predictors and versions by running

# to list all predefined predictors
nn-meter --list-predictors 

After installation, a command named nn-meter is enabled. To predict the latency for a CNN model with a predefined predictor in command line, users can run the following commands (sample models can be downloaded here)

# for Tensorflow (*.pb) file
nn-meter predict --predictor <hardware> [--predictor-version <version>] --tensorflow <pb-file_or_folder> 
# Example Usage
nn-meter predict --predictor cortexA76cpu_tflite21 --predictor-version 1.0 --tensorflow mobilenetv3small_0.pb 

# for ONNX (*.onnx) file
nn-meter predict --predictor <hardware> [--predictor-version <version>] --onnx <onnx-file_or_folder>
#Example Usage
nn-meter predict --predictor cortexA76cpu_tflite21 --predictor-version 1.0 --onnx mobilenetv3small_0.onnx 

# for torch model from torchvision model zoo (str)
nn-meter predict --predictor <hardware> [--predictor-version <version>] --torchvision <model-name> <model-name>... 
#Example Usage
nn-meter predict --predictor cortexA76cpu_tflite21 --predictor-version 1.0 --torchvision mobilenet_v2

# for nn-Meter IR (*.json) file
nn-meter predict --predictor <hardware> [--predictor-version <version>] --nn-meter-ir <json-file_or_folder> 
#Example Usage
nn-meter predict --predictor cortexA76cpu_tflite21 --predictor-version 1.0 --nn-meter-ir mobilenetv3small_0.json 

--predictor-version <version> arguments is optional. When the predictor version is not specified by users, nn-meter will use the latest version of the predictor.

nn-Meter can support batch mode prediction. To predict latency for multiple models in the same model type once, user should collect all models in one folder and state the folder after --[model-type] liked argument.

It should also be noted that for PyTorch model, nn-meter can only support existing models in torchvision model zoo. The string followed by --torchvision should be exactly one or more string indicating name(s) of some existing torchvision models. To apply latency prediction for torchvision model in command line, onnx and onnx-simplifier packages are required.

Furthermore, users may be interested to convert tensorflow pb-file or onnx file to nn-Meter IR graph. Users could convert nn-Meter IR graph and save to .json file be running

# for Tensorflow (*.pb) file
nn-meter get_ir --tensorflow <pb-file> [--output <output-name>]

# for ONNX (*.onnx) file
nn-meter get_ir --onnx <onnx-file> [--output <output-name>]

Output name is default to be /path/to/input/file/<input_file_name>_<model-type>_ir.json if not specified by users.

After installation, users can import nn-Meter in python code

from nn_meter import load_latency_predictor

predictor = load_latency_predictor(hardware_name, hardware_predictor_version) # case insensitive in backend

# build your model (e.g., model instance of torch.nn.Module)
model = ... 

lat = predictor.predict(model, model_type) # the resulting latency is in unit of ms

By calling load_latency_predictor, user selects the target hardware and loads the corresponding predictor. nn-Meter will try to find the right predictor file in ~/.nn_meter/data. If the predictor file doesn't exist, it will download from the Github release.

In predictor.predict(), the allowed items of the parameter model_type include ["pb", "torch", "onnx", "nnmeter-ir", "nni-ir"], representing model types of tensorflow, torch, onnx, nn-meter IR graph and NNI IR graph, respectively.

<span id="torch-model-converters"> For Torch models, the shape of feature maps is unknown merely based on the given network structure, which is, however, significant parameters in latency prediction. Therefore, torch model requires a shape of input tensor for inference as a input of predictor.predict(). Based on the given input shape, a random tensor according to the shape will be generated and used. Another thing for Torch model prediction is that users can install the onnx and onnx-simplifier packages for latency prediction (referred to as Onnx-based latency prediction for torch model), or alternatively install the nni package (referred to as NNI-based latency prediction for torch model). Note that the nni option does not support command line calls. In addition, if users use nni for latency prediction, the PyTorch modules should be defined by the nn interface from NNI import nni.retiarii.nn.pytorch as nn (view NNI doc for more information), and the parameter apply_nni should be set as True in the function predictor.predict(). Here is an example of NNI-based latency prediction for Torch model:

import nni.retiarii.nn.pytorch as nn
from nn_meter import load_latency_predictor

predictor = load_latency_predictor(...)

# build your model using nni.retiarii.nn.pytorch as nn
model = nn.Module ...

input_shape = (1, 3, 224, 224)
lat = predictor.predict(model, model_type='torch', input_shape=input_shape, apply_nni=True) 

The Onnx-based latency prediction for torch model is stable but slower, while the NNI-based latency prediction for torch model is unstable as it could fail in some case but much faster compared to the Onnx-based model. The Onnx-based model is set as the default one for Torch model latency prediction in nn-Meter. Users could choose which one they preferred to use according to their needs.

Users could view the information all built-in predictors by list_latency_predictors or view the config file in nn_meter/configs/predictors.yaml.

Users could get a nn-Meter IR graph by applying model_file_to_graph and model_to_graph by calling the model name or model object and specify the model type. The supporting model types of model_file_to_graph include "onnx", "pb", "torch", "nnmeter-ir" and "nni-ir", while the supporting model types of model_to_graph include "onnx", "torch" and "nni-ir".

nn-Meter builder is an open source tool for users to build latency predictor on their own devices. There are three main parts in nn-Meter builder:

backend: the module of connecting backends;

backend_meta: the meta tools related to backend. Here we provide the fusion rule tester to detect fusion rules for users' backend;

kernel_predictor_builder: the tool to build different kernel latency predictors.

Users could get access to nn-Meter builder by calling nn_meter.builder. For more details to use nn-Meter builder, please check the document of nn-Meter builder.

To empower affordable DNN on the edge and mobile devices, hardware-aware NAS searches both high accuracy and low latency models. In particular, the search algorithm only considers the models within the target latency constraints during the search process.

Currently we provide two examples of hardware-aware NAS, including end-to-end multi-trial NAS which is a random search algorithm on SPOS NAS search space, and the popular ProxylessNAS, which is a one-shot NAS algorithm with hardware-efficient loss function. More examples of other widely-used hardware-aware NAS and model compression algorithms are coming soon.

To run multi-trail SPOS demo, NNI should be installed through source code by following NNI Doc

python setup.py develop

Then run multi-trail SPOS demo:

python ${NNI_ROOT}/examples/nas/oneshot/spos/multi_trial.py

Refer to NNI Doc for how to perform NAS by NNI.

To support hardware-aware NAS, you first need a Strategy that supports filtering the models by latency. We provide such a filter named LatencyFilter in NNI and initialize a Random strategy with the filter:

simple_strategy = strategy.Random(model_filter=LatencyFilter(threshold=100, predictor=base_predictor))

LatencyFilter will predict the models' latency by using nn-Meter and filter out the models whose latency with the given predictor are larger than the threshold (i.e., 100 in this example). You can also build your own strategies and filters to support more flexible NAS such as sorting the models according to latency.

Then, pass this strategy to RetiariiExperiment:

exp = RetiariiExperiment(base_model, trainer, strategy=simple_strategy)

exp_config = RetiariiExeConfig('local')
...
exp_config.dummy_input = [1, 3, 32, 32]

exp.run(exp_config, port)

In exp_config, dummy_input is required for tracing shape info.

To run the one-shot ProxylessNAS demo, users can run the NNI ProxylessNAS training demo:

python ${NNI_ROOT}/examples/nas/oneshot/proxylessnas/main.py --applied_hardware <hardware> --reference_latency <reference latency (ms)>

Refer to NNI Doc for how to perform NAS by NNI.

ProxylessNAS currently builds a lookup table, that stores the measured latency of each candidate building block in the search space. The latency sum of all building blocks in a candidate model will be treated as the model inference latency. With leveraging nn-Meter in NNI, users can apply ProxylessNAS to search efficient DNN models on more types of edge devices. In NNI implementation, a HardwareLatencyEstimator predicts expected latency for the mixed operation based on the path weight of ProxylessLayerChoice. To call nn-Meter in NNI ProxylessNAS, users can add the arguments of "--applied_hardware <hardware> --reference_latency <reference latency (ms)>" in the example.

To evaluate the effectiveness of a prediction model on an arbitrary DNN model, we need a representative dataset that covers a large prediction scope. nn-Meter collects and generates 26k CNN models. (Please refer the paper for the dataset generation method.)

We release the dataset, and provide an interface of nn_meter.dataset for users to get access to the dataset. Users can also download the data from the Download Link on their own.

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.

The entire codebase is under MIT license

The dataset is under Open Use of Data Agreement

If you find that nn-Meter helps your research, please consider citing it:

@inproceedings{nnmeter,
    author = {Zhang, Li Lyna and Han, Shihao and Wei, Jianyu and Zheng, Ningxin and Cao, Ting and Yang, Yuqing and Liu, Yunxin},
    title = {nn-Meter: Towards Accurate Latency Prediction of Deep-Learning Model Inference on Diverse Edge Devices},
    year = {2021},
    publisher = {ACM},
    address = {New York, NY, USA},
    url = {https://doi.org/10.1145/3458864.3467882},
    doi = {10.1145/3458864.3467882},
    booktitle = {Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services},
    pages = {81–93},
}

@misc{nnmetercode,
    author = {Microsoft Research nn-Meter Team},
    title = {nn-Meter: Towards Accurate Latency Prediction of Deep-Learning Model Inference on Diverse Edge Devices},
    year = {2021},
    url = {https://github.com/microsoft/nn-Meter},
}