Royce  Reinger

Royce Reinger


MNN: A Blazing Fast, Lightweight Deep Learning Framework


A blazing fast, lightweight deep learning framework, battle-tested by business-critical use cases in Alibaba


MNN is a highly efficient and lightweight deep learning framework. It supports inference and training of deep learning models, and has industry leading performance for inference and training on-device. At present, MNN has been integrated in more than 30 apps of Alibaba Inc, such as Taobao, Tmall, Youku, Dingtalk, Xianyu and etc., covering more than 70 usage scenarios such as live broadcast, short video capture, search recommendation, product searching by image, interactive marketing, equity distribution, security risk control. In addition, MNN is also used on embedded devices, such as IoT.


Inside Alibaba, MNN works as the basic module of the compute container in the Walle System, the first end-to-end, general-purpose, and large-scale production system for device-cloud collaborative machine learning, which has been published in the top system conference OSDI’22. The key design principles of MNN and the extensive benchmark testing results (vs. TensorFlow, TensorFlow Lite, PyTorch, PyTorch Mobile, TVM) can be found in the OSDI paper. The scripts and instructions for benchmark testing are put in the path “/benchmark”. If MNN or the design of Walle helps your research or production use, please cite our OSDI paper as follows:

@inproceedings {proc:osdi22:walle,
    author = {Chengfei Lv and Chaoyue Niu and Renjie Gu and Xiaotang Jiang and Zhaode Wang and Bin Liu and Ziqi Wu and Qiulin Yao and Congyu Huang and Panos Huang and Tao Huang and Hui Shu and Jinde Song and Bin Zou and Peng Lan and Guohuan Xu and Fei Wu and Shaojie Tang and Fan Wu and Guihai Chen},
    title = {Walle: An {End-to-End}, {General-Purpose}, and {Large-Scale} Production System for {Device-Cloud} Collaborative Machine Learning},
    booktitle = {16th USENIX Symposium on Operating Systems Design and Implementation (OSDI 22)},
    year = {2022},
    isbn = {978-1-939133-28-1},
    address = {Carlsbad, CA},
    pages = {249--265},
    url = {},
    publisher = {USENIX Association},
    month = jul,

Documentation and Workbench

MNN's docs are in placed in Yuque docs here and Read the docs.

MNN Workbench could be downloaded from MNN's homepage, which provides pretrained models, visualized training tools, and one-click deployment of models to devices.

Key Features


  • Optimized for devices, no dependencies, can be easily deployed to mobile devices and a variety of embedded devices.
  • iOS platform: static library size will full option for armv7+arm64 platforms is about 12MB, size increase of linked executables is about 2M.
  • Android platform: core so size is about 800KB (armv7a - c++_shared).
  • Use MNN_BUILD_MINI can reduce package size about 25% , with limit of fix model input size
  • Support FP16 / Int8 qunatize, can reduce model size 50%-70%


  • Supports Tensorflow, Caffe, ONNX,Torchscripts and supports common neural networks such as CNN, RNN, GAN, Transformork.
  • Supports AI model with multi-inputs or multi-outputs, every kind of dimenstion format, dynamic inputs, controlflow.
  • MNN supports approximate full OPs used for AI Model. The converter supports 178 Tensorflow OPs, 52 Caffe OPs, 163 Torchscripts OPs, 158 ONNX OPs.
  • Supports iOS 8.0+, Android 4.3+ and embedded devices with POSIX interface.
  • Supports hybrid computing on multiple devices. Currently supports CPU and GPU.

High performance

  • Implements core computing with lots of optimized assembly code to make full use of the ARM / x64 CPU.
  • Use Metal / OpenCL / Vulkan to support GPU inference on mobile.
  • Use CUDA and tensorcore to support NVIDIA GPU for better performance
  • Convolution and transposition convolution algorithms are efficient and stable. The Winograd convolution algorithm is widely used to better symmetric convolutions such as 3x3,4x4,5x5,6x6,7x7.
  • Twice speed increase for the new architecture ARM v8.2 with FP16 half-precision calculation support. 2.5 faster to use sdot for ARM v8.2 and VNNI.

Ease of use

  • Support use MNN's OP to do numerical calculating like numpy.
  • Support lightweight image process module like OpenCV, which is only 100k.
  • Support build model and train it on PC / mobile.
  • MNN Python API helps ML engineers to easily use MNN to inference, train, process image, without dipping their toes in C++ code.

The Architecture / Precision MNN supported is shown below:

  • S :Support and work well, deeply optimized, recommend to use
  • A :Support and work well, can use
  • B :Support but has bug or not optimized, no recommend to use
  • C :Not Support
Architecture / Precision NormalFP16BF16Int8
 ARMv7aSS (ARMv8.2)SS
 ARMv8SS (ARMv8.2)S(ARMv8.6)S


Base on MNN (Tensor compute engine), we provided a series of tools for inference, train and general computation.

  • MNN-Converter: Convert other model to MNN model for inference, such as Tensorflow(lite), Caffe, ONNX, Torchscripts. And do graph optimization to reduce computation.
  • MNN-Compress: Compress model to reduce size and increase performance / speed
  • MNN-Express: Support model with controlflow, use MNN's OP to do general-purpose compute.
  • MNN-CV: An OpenCV liked library, but based on MNN and then much more lightweight.
  • MNN-Train: Support train MNN model.

How to Discuss and Get Help From MNN Community

The group discussions are predominantly Chinese. But we welcome and will help English speakers.

Dingtalk discussion groups:

Group 1 (Full): 23329087

Group 2 (Full): 23350225

Group 3:

Historical Paper

The preliminary version of MNN, as mobile inference engine and with the focus on manual optimization, has also been published in MLSys 2020. Please cite the paper, if MNN previously helped your research:

  author = {Jiang, Xiaotang and Wang, Huan and Chen, Yiliu and Wu, Ziqi and Wang, Lichuan and Zou, Bin and Yang, Yafeng and Cui, Zongyang and Cai, Yu and Yu, Tianhang and Lv, Chengfei and Wu, Zhihua},
  title = {MNN: A Universal and Efficient Inference Engine},
  booktitle = {MLSys},
  year = {2020}


MNN participants: Taobao Technology Department, Search Engineering Team, DAMO Team, Youku and other Alibaba Group employees.

MNN refers to the following projects:


MNN Homepage

Download Details:

Author: Alibaba
Source Code: 
License: Apache 2.0

#machinelearning #neuralnetworks #deeplearning 

What is GEEK

Buddha Community

MNN: A Blazing Fast, Lightweight Deep Learning Framework
Marget D

Marget D


Top Deep Learning Development Services | Hire Deep Learning Developer

View more:

We at Inexture, strategically work on every project we are associated with. We propose a robust set of AI, ML, and DL consulting services. Our virtuoso team of data scientists and developers meticulously work on every project and add a personalized touch to it. Because we keep our clientele aware of everything being done associated with their project so there’s a sense of transparency being maintained. Leverage our services for your next AI project for end-to-end optimum services.

#deep learning development #deep learning framework #deep learning expert #deep learning ai #deep learning services

Mikel  Okuneva

Mikel Okuneva


Top 10 Deep Learning Sessions To Look Forward To At DVDC 2020

The Deep Learning DevCon 2020, DLDC 2020, has exciting talks and sessions around the latest developments in the field of deep learning, that will not only be interesting for professionals of this field but also for the enthusiasts who are willing to make a career in the field of deep learning. The two-day conference scheduled for 29th and 30th October will host paper presentations, tech talks, workshops that will uncover some interesting developments as well as the latest research and advancement of this area. Further to this, with deep learning gaining massive traction, this conference will highlight some fascinating use cases across the world.

Here are ten interesting talks and sessions of DLDC 2020 that one should definitely attend:

Also Read: Why Deep Learning DevCon Comes At The Right Time

Adversarial Robustness in Deep Learning

By Dipanjan Sarkar

**About: **Adversarial Robustness in Deep Learning is a session presented by Dipanjan Sarkar, a Data Science Lead at Applied Materials, as well as a Google Developer Expert in Machine Learning. In this session, he will focus on the adversarial robustness in the field of deep learning, where he talks about its importance, different types of adversarial attacks, and will showcase some ways to train the neural networks with adversarial realisation. Considering abstract deep learning has brought us tremendous achievements in the fields of computer vision and natural language processing, this talk will be really interesting for people working in this area. With this session, the attendees will have a comprehensive understanding of adversarial perturbations in the field of deep learning and ways to deal with them with common recipes.

Read an interview with Dipanjan Sarkar.

Imbalance Handling with Combination of Deep Variational Autoencoder and NEATER

By Divye Singh

**About: **Imbalance Handling with Combination of Deep Variational Autoencoder and NEATER is a paper presentation by Divye Singh, who has a masters in technology degree in Mathematical Modeling and Simulation and has the interest to research in the field of artificial intelligence, learning-based systems, machine learning, etc. In this paper presentation, he will talk about the common problem of class imbalance in medical diagnosis and anomaly detection, and how the problem can be solved with a deep learning framework. The talk focuses on the paper, where he has proposed a synergistic over-sampling method generating informative synthetic minority class data by filtering the noise from the over-sampled examples. Further, he will also showcase the experimental results on several real-life imbalanced datasets to prove the effectiveness of the proposed method for binary classification problems.

Default Rate Prediction Models for Self-Employment in Korea using Ridge, Random Forest & Deep Neural Network

By Dongsuk Hong

About: This is a paper presentation given by Dongsuk Hong, who is a PhD in Computer Science, and works in the big data centre of Korea Credit Information Services. This talk will introduce the attendees with machine learning and deep learning models for predicting self-employment default rates using credit information. He will talk about the study, where the DNN model is implemented for two purposes — a sub-model for the selection of credit information variables; and works for cascading to the final model that predicts default rates. Hong’s main research area is data analysis of credit information, where she is particularly interested in evaluating the performance of prediction models based on machine learning and deep learning. This talk will be interesting for the deep learning practitioners who are willing to make a career in this field.

#opinions #attend dldc 2020 #deep learning #deep learning sessions #deep learning talks #dldc 2020 #top deep learning sessions at dldc 2020 #top deep learning talks at dldc 2020

Few Shot Learning — A Case Study (2)

In the previous blog, we looked into the fact why Few Shot Learning is essential and what are the applications of it. In this article, I will be explaining the Relation Network for Few-Shot Classification (especially for image classification) in the simplest way possible. Moreover, I will be analyzing the Relation Network in terms of:

  1. Effectiveness of different architectures such as Residual and Inception Networks
  2. Effects of transfer learning via using pre-trained classifier on ImageNet dataset

Moreover, effectiveness will be evaluated on the accuracy, time required for training, and the number of required training parameters.

Please watch the GitHub repository to check out the implementations and keep updated with further experiments.

Introduction to Few-Shot Classification

In few shot classification, our objective is to design a method which can identify any object images by analyzing few sample images of the same class. Let’s the take one example to understand this. Suppose Bob has a client project to design a 5 class classifier, where 5 classes can be anything and these 5 classes can even change with time. As discussed in previous blog, collecting the huge amount of data is very tedious task. Hence, in such cases, Bob will rely upon few shot classification methods where his client can give few set of example images for each classes and after that his system can perform classification young these examples with or without the need of additional training.

In general, in few shot classification four terminologies (N way, K shot, support set, and query set) are used.

  1. N way: It means that there will be total N classes which we will be using for training/testing, like 5 classes in above example.
  2. K shot: Here, K means we have only K example images available for each classes during training/testing.
  3. Support set: It represents a collection of all available K examples images from each classes. Therefore, in support set we have total N*K images.
  4. Query set: This set will have all the images for which we want to predict the respective classes.

At this point, someone new to this concept will have doubt regarding the need of support and query set. So, let’s understand it intuitively. Whenever humans sees any object for the first time, we get the rough idea about that object. Now, in future if we see the same object second time then we will compare it with the image stored in memory from the when we see it for the first time. This applied to all of our surroundings things whether we see, read, or hear. Similarly, to recognise new images from query set, we will provide our model a set of examples i.e., support set to compare.

And this is the basic concept behind Relation Network as well. In next sections, I will be giving the rough idea behind Relation Network and I will be performing different experiments on 102-flower dataset.

About Relation Network

The Core idea behind Relation Network is to learn the generalized image representations for each classes using support set such that we can compare lower dimensional representation of query images with each of the class representations. And based on this comparison decide the class of each query images. Relation Network has two modules which allows us to perform above two tasks:

  1. Embedding module: This module will extract the required underlying representations from each input images irrespective of the their classes.
  2. Relation Module: This module will score the relation of embedding of query image with each class embedding.

Training/Testing procedure:

We can define the whole procedure in just 5 steps.

  1. Use the support set and get underlying representations of each images using embedding module.
  2. Take the average of between each class images and get the single underlying representation for each class.
  3. Then get the embedding for each query images and concatenate them with each class’ embedding.
  4. Use the relation module to get the scores. And class with highest score will be the label of respective query image.
  5. [Only during training] Use MSE loss functions to train both (embedding + relation) modules.

Few things to know during the training is that we will use only images from the set of selective class, and during the testing, we will be using images from unseen classes. For example, from the 102-flower dataset, we will use 50% classes for training, and rest will be used for validation and testing. Moreover, in each episode, we will randomly select 5 classes to create the support and query set and follow the above 5 steps.

That is all need to know about the implementation point of view. Although the whole process is simple and easy to understand, I’ll recommend reading the published research paper, Learning to Compare: Relation Network for Few-Shot Learning, for better understanding.

#deep-learning #few-shot-learning #computer-vision #machine-learning #deep learning #deep learning

Tia  Gottlieb

Tia Gottlieb


Deep Reinforcement Learning for Video Games Made Easy

In this post, we will investigate how easily we can train a Deep Q-Network (DQN) agent (Mnih et al., 2015) for Atari 2600 games using the Google reinforcement learning library Dopamine. While many RL libraries exist, this library is specifically designed with four essential features in mind:

  • Easy experimentation
  • Flexible development
  • Compact and reliable
  • Reproducible

_We believe these principles makes __Dopamine _one of the best RL learning environment available today. Additionally, we even got the library to work on Windows, which we think is quite a feat!

In my view, the visualization of any trained RL agent is an absolute must in reinforcement learning! Therefore, we will (of course) include this for our own trained agent at the very end!

We will go through all the pieces of code required (which is** minimal compared to other libraries**), but you can also find all scripts needed in the following Github repo.

1. Brief Introduction to Reinforcement Learning and Deep Q-Learning

The general premise of deep reinforcement learning is to

“derive efficient representations of the environment from high-dimensional sensory inputs, and use these to generalize past experience to new situations.”

  • Mnih et al. (2015)

As stated earlier, we will implement the DQN model by Deepmind, which only uses raw pixels and game score as input. The raw pixels are processed using convolutional neural networks similar to image classification. The primary difference lies in the objective function, which for the DQN agent is called the optimal action-value function

Image for post

where_ rₜ is the maximum sum of rewards at time t discounted by γ, obtained using a behavior policy π = P(a_∣_s)_ for each observation-action pair.

There are relatively many details to Deep Q-Learning, such as Experience Replay (Lin, 1993) and an _iterative update rule. _Thus, we refer the reader to the original paper for an excellent walk-through of the mathematical details.

One key benefit of DQN compared to previous approaches at the time (2015) was the ability to outperform existing methods for Atari 2600 games using the same set of hyperparameters and only pixel values and game score as input, clearly a tremendous achievement.

2. Installation

This post does not include instructions for installing Tensorflow, but we do want to stress that you can use both the CPU and GPU versions.

Nevertheless, assuming you are using Python 3.7.x, these are the libraries you need to install (which can all be installed via pip):

tensorflow-gpu=1.15   (or tensorflow==1.15  for CPU version)

#reinforcement-learning #q-learning #games #machine-learning #deep-learning #deep learning

Learn Transfer Learning for Deep Learning by implementing the project.

Project walkthrough on Convolution neural networks using transfer learning

From 2 years of my master’s degree, I found that the best way to learn concepts is by doing the projects. Let’s start implementing or in other words learning.

Problem Statement

Take an image as input and return a corresponding dog breed from 133 dog breed categories. If a dog is detected in the image, it will provide an estimate of the dog’s breed. If a human is detected, it will give an estimate of the dog breed that is most resembling the human face. If there’s no human or dog present in the image, we simply print an error.

Let’s break this problem into steps

  1. Detect Humans
  2. Detect Dogs
  3. Classify Dog breeds

For all these steps, we use pre-trained models.

Pre-trained models are saved models that were trained on a huge image-classification task such as Imagenet. If these datasets are huge and generalized enough, the saved weights can be used for multiple image detection task to get a high accuracy quickly.

Detect Humans

For detecting humans, OpenCV provides many pre-trained face detectors. We use OpenCV’s implementation of Haar feature-based cascade classifiers to detect human faces in images.

### returns "True" if face is detected in image stored at img_path
def face_detector(img_path):
    img = cv2.imread(img_path)
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    faces = face_cascade.detectMultiScale(gray)
    return len(faces) > 0

Image for post

Detect Dogs

For detecting dogs, we use a pre-trained ResNet-50 model to detect dogs in images, along with weights that have been trained on ImageNet, a very large, very popular dataset used for image classification and other vision tasks.

from keras.applications.resnet50 import ResNet50

### define ResNet50 model
ResNet50_model_detector = ResNet50(weights='imagenet')
### returns "True" if a dog is detected
def dog_detector(img_path):
    prediction = ResNet50_predict_labels(img_path)
    return ((prediction <= 268) & (prediction >= 151))

Classify Dog Breeds

For classifying Dog breeds, we use transfer learning

Transfer learning involves taking a pre-trained neural network and adapting the neural network to a new, different data set.

To illustrate the power of transfer learning. Initially, we will train a simple CNN with the following architecture:

Image for post

Train it for 20 epochs, and it gives a test accuracy of just 3% which is better than a random guess from 133 categories. But with more epochs, we can increase accuracy, but it takes up a lot of training time.

To reduce training time without sacrificing accuracy, we will train the CNN model using transfer learning.

#data-science #transfer-learning #project-based-learning #cnn #deep-learning #deep learning