Edge AI and Vision Alliance, profile picture
Uploaded byEdge AI and Vision Alliance
PDF, PPTX1,016 views

"PyTorch Deep Learning Framework: Status and Directions," a Presentation from Facebook

The document provides an overview of PyTorch, focusing on mobile integration and the evolution of deploying models from research to production. It outlines key principles such as dynamic vs. static frameworks, the benefits of PyTorch's eager execution model, and the usability of TorchScript for optimizing Python models without Python dependencies. Additionally, it highlights efforts in model quantization and optimization for mobile devices, aiming for efficient inference in constrained environments.

Technology◦
Related topics:
Computer Vision Insights•
In this document
Powered by AI

Joseph Spisak, Product Manager at PyTorch for Mobile presentation introduction.

Overview of PyTorch mobile capabilities, dynamic vs static frameworks, and transition from research to production.

Essentials of PyTorch including developer efficiency, scale, simplicity over complexity, and dynamics of neural networks.

Key details of TorchScript as an optimizable subset of Python for model deployment without dependencies.

Highlights of PyTorch community growth, including ~1,230 contributors and 50% year-over-year growth.

Examining the differences between dynamic and static frameworks, including computation graph and implementation.

Focus on defining models through execution in imperative toolkits with Python native runtime.

How to run and export PyTorch models in ONNX format for mobile, focusing on support for mobile.

Explains the architecture supporting ONNX exporters, optimizations, and translations of PyTorch IR.

Introduction to PyTorch Mobile focusing on its capabilities, running TorchScript models, and avoiding full Python.

Capabilities of PyTorch Mobile including running TorchScript models, loops, quantization benefits, and performance.

Guide on how to implement and compile TorchScript models for mobile on Android and iOS.

Links to PyTorch mobile demo apps on Android and iOS, showcasing functionality and usage examples.

Upcoming features for PyTorch Mobile including optimization, new APIs, and potential GPU support.

Download as PDF, PPTX
J O S E P H S P I S A K P R O D U C T M A N A G E R P Y T O R C H | O N D E V I C E
OVERVIEW AND INTEGRATION PYTORCH MOBILE STATE OF THE STATE ON DEVICE DYNAMIC VS. STATIC FRAMEWORKS PRINCIPLES, MAKEUP, .. OVERVIEW
REIMPLEMENTATIONTAKESWEEKSORMONTHS Research to Production at Facebook – Early 2017
ENABLINGMODELORMODELFRAGMENTTRANSFER Research to Production at Facebook – Early 2018
Research to Production at Facebook
P Y T O R C H O V E R V I E W
SIMPLICITY OVER COMPLEXITY HARDWARE ACCELERATED INFERENCE DISTRIBUTED TRAINING DYNAMIC 
 NEURAL 
 NETWORKS EAGER & GRAPH-BASED EXECUTION WHAT IS PYTORCH?
P Y T O R C H R E S E A R C H P R O T O T Y P I N G P R O D U C T I O N D E P L O Y M E N T +
C O R E P R I N C I P L E S BUILDING FOR SCALE DEVELOPER EFFICIENCY
DEVELOPER EFFICIENCY ENABLING A HIGH VELOCITY OF MODEL ITERATION AND INNOVATION
` C L E A N A P I S
` T O R C H S C R I P T Models are Python TorchScript programs, an optimizable subset of Python + Same “models are programs” idea + Production deployment + No Python dependency + Compilation for performance optimization class RNN(nn.Module):   def __init__(self, W_h, U_h, W_y, b_h, b_y):     super(RNN, self).__init__()     self.W_h = nn.Parameter(W_h)     self.U_h = nn.Parameter(U_h)     self.W_y = nn.Parameter(W_y)     self.b_h = nn.Parameter(b_h)     self.b_y = nn.Parameter(b_y)   def forward(self, x, h):     y = []     for t in range(x.size(0)):       h = torch.tanh(x[t] @ self.W_h + h @ self.U_h + self.b_h)       y += [torch.tanh(h @ self.W_y + self.b_y)]       if t % 10 == 0:         print("stats: ", h.mean(), h.var())     return torch.stack(y), h 
 # one annotation! script_rnn = torch.jit.script(RNN(W_h, U_h, W_y, b_h, b_y))
` T O R C H S C R I P T Models are Python TorchScript programs, an optimizable subset of Python + Same “models are programs” idea + Prod deployment + No Python dependency + Optimizable (incl. codegen!)
` T E N S O R B O A R D
~1,230C O N T R I B U T O R S 50%+Y O Y G R O W T H 23KP Y T O R C H F O R U M U S E R S
GROW TH IN ARXIV MENTIONS IN RESEARCH PAPERS 0 100 200 300 400 500 Jan 17 Feb 17 M ar17 Apr17 M ay17 Jun 17 Jul17 Aug17 Sep 17 Jan 18 Feb 18 M ar18 Apr18 M ay18 Jun 18 Jul18 Aug18 Sep 18 Jan 19 Feb 19 M ar19 Apr19 M ay19 Jun 19 Jul19
F R A M E W O R K S D Y N A M I C V S . S T A T I C
Declarative Toolkits COMPUTATIONGRAPH
DECLARATIVETOOLKITS Declare and compile a model Repeatedly execute the model in a VM TOOLKIT VM PYTHON SCRIPT RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM
import tensorflow as tf
 import numpy as np trX = np.linspace(-1, 1, 101)
 trY = 2 * trX + np.random.randn(*trX.shape) * 0.33 X = tf.placeholder("float")
 Y = tf.placeholder("float") def model(X, w): return tf.multiply(X, w) w = tf.Variable(0.0, name="weights")
 y_model = model(X, w) cost = tf.square(Y - y_model) train_op = tf.train.GradientDescentOptimizer(0.01).minimize(cost) with tf.Session() as sess: tf.global_variables_initializer().run() for i in range(100): for (x, y) in zip(trX, trY): sess.run(train_op, feed_dict={X: x, Y: y}) print(sess.run(w)) DECLARATIVETOOLKITS Computation Graph • Declare a computation • Placeholder variables • Compile it • Run it in a Session
import tensorflow as tf
 import numpy as np trX = np.linspace(-1, 1, 101)
 trY = 2 * trX + np.random.randn(*trX.shape) * 0.33 X = tf.placeholder("float")
 Y = tf.placeholder("float") def model(X, w): return tf.multiply(X, w) w = tf.Variable(0.0, name="weights")
 y_model = model(X, w) cost = tf.square(Y - y_model) train_op = tf.train.GradientDescentOptimizer(0.01).minimize(cost) with tf.Session() as sess: tf.global_variables_initializer().run() for i in range(100): for (x, y) in zip(trX, trY): sess.run(train_op, feed_dict={X: x, Y: y}) print(sess.run(w)) X = tf.placeholder("float")
 Y = tf.placeholder("float") DECLARATIVETOOLKITS Computation Graph • Declare a computation • Placeholder variables • Compile it • Run it in a Session
import tensorflow as tf
 import numpy as np trX = np.linspace(-1, 1, 101)
 trY = 2 * trX + np.random.randn(*trX.shape) * 0.33 X = tf.placeholder("float")
 Y = tf.placeholder("float") def model(X, w): return tf.multiply(X, w) w = tf.Variable(0.0, name="weights")
 y_model = model(X, w) cost = tf.square(Y - y_model) train_op = tf.train.GradientDescentOptimizer(0.01).minimize(cost) with tf.Session() as sess: tf.global_variables_initializer().run() for i in range(100): for (x, y) in zip(trX, trY): sess.run(train_op, feed_dict={X: x, Y: y}) print(sess.run(w)) Model definition def model(X, w): return tf.multiply(X, w) w = tf.Variable(0.0, name="weights")
 y_model = model(X, w) cost = tf.square(Y - y_model) train_op = tf.train.GradientDescentOptimizer(0.01).minimize(cost) DECLARATIVETOOLKITS Computation Graph • Declare a computation • Placeholder variables • Compile it • Run it in a Session
import tensorflow as tf
 import numpy as np trX = np.linspace(-1, 1, 101)
 trY = 2 * trX + np.random.randn(*trX.shape) * 0.33 X = tf.placeholder("float")
 Y = tf.placeholder("float") def model(X, w): return tf.multiply(X, w) w = tf.Variable(0.0, name="weights")
 y_model = model(X, w) cost = tf.square(Y - y_model) train_op = tf.train.GradientDescentOptimizer(0.01).minimize(cost) with tf.Session() as sess: tf.global_variables_initializer().run() for i in range(100): for (x, y) in zip(trX, trY): sess.run(train_op, feed_dict={X: x, Y: y}) print(sess.run(w)) sess.run(train_op, feed_dict={X: x, Y: y}) print(sess.run(w)) A separate, Turing complete, virtual machine. for i in range(100): for (x, y) in zip(trX, trY): DECLARATIVETOOLKITS Computation Graph • Declare a computation • Placeholder variables • Compile it • Run it in a Session
Imperative Toolkits DEFINE-BY-RUN
IMPERATIVETOOLKITS Run a series of computation Implicitly defining the model as execution goes PYTHON NATIVE RUNTIME PYTHON INSTRUCTIONS RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM
Imperative Toolkits import torch
 from torch.autograd import Variable trX = torch.linspace(-1, 1, 101)
 trY = 2 * trX + torch.random(*trX.size()) * 0.33 w = Variable(trX.new([0.0]), requires_grad=True) for i in range(100): for (x, y) in zip(trX, trY): X = Variable(x) Y = Variable(y) print(X) print(Y) y_model = X * w.expand_as(X) cost = (Y - y_model) * 2 Cost.backward(torch.ones(*cost.size())) w.data = w.data + 0.01 * w.grad.data print(w) • Define a model by execution • No separate compilation stage • No separate execution engine
import torch
 from torch.autograd import Variable trX = torch.linspace(-1, 1, 101)
 trY = 2 * trX + torch.random(*trX.size()) * 0.33 w = Variable(trX.new([0.0]), requires_grad=True) for i in range(100): for (x, y) in zip(trX, trY): X = Variable(x) Y = Variable(y) print(X) print(Y) y_model = X * w.expand_as(X) cost = (Y - y_model) * 2 Cost.backward(torch.ones(*cost.size())) w.data = w.data + 0.01 * w.grad.data print(w) Imperative Toolkits Model constructed and values computed as we define it. • Define a model by execution • No separate compilation stage • No separate execution engine
P Y T O R C H F O R E M B E D D E D S T A T E O F T H E S T A T E
H O W D O I R U N P Y T O R C H M O D E L S O N D E V I C E ?
H O W D O I R U N P Y T O R C H M O D E L S O N D E V I C E ? E X P O R T O N N X F O R M A T T E D M O D E L S
H O W D O I R U N P Y T O R C H M O D E L S O N D E V I C E ? E X P O R T O N N X F O R M A T T E D M O D E L S P Y T O R C H M O B I L E
P Y T O R C H O N N X S U P P O R T
O N N X E c o s y s t e m
A R C H I T E C T U R E 
 A N D F L O W JIT Tracer and Torchscript ONNX
 Exporter Optimizer Torch IR to ONNX IR Translator Torch IR ONNX torch.onnx
 .export() PyTorch Model Sample Input ONNX Graph
Partners
P Y T O R C H M O B I L E P R E V I E W R E L E A S E
W H A T I S   P Y T O R C H   M O B I L E ?
W H A T I S   P Y T O R C H   M O B I L E ? I T ’ S P Y T O R C H
W H A T I S   P Y T O R C H   M O B I L E ? I T ’ S P Y T O R C H F O R M O B I L E 😃
W H A T I S   P Y T O R C H   M O B I L E ? I T ’ S P Y T O R C H F O R M O B I L E B U T N O P Y T H O N 😃
W H A T C A N I T R U N ? A N Y T O R C H S C R I P T M O D E L .
W H A T C A N I T R U N ? A N Y T O R C H S C R I P T M O D E L . L O O P S ? Y E S
W H A T C A N I T R U N ? A N Y T O R C H S C R I P T M O D E L . L O O P S ? F U N C T I O N S ? Y E S Y E S
W H A T C A N I T R U N ? A N Y T O R C H S C R I P T M O D E L . L O O P S ? F U N C T I O N S ? T U P L E S ? Y E S Y E S 
 Y E S
W H A T C A N I T R U N ? A N Y T O R C H S C R I P T M O D E L . L O O P S ? F U N C T I O N S ? T U P L E S ? N A M E D T U P L E ? Y E S Y E S 
 Y E S Y E S
ANDROID - MAVEN iOS - COCOAPODS MODEL OPTIMIZATION (OPTIONAL ) PY TORCH MOBILE • No separate runtime to export P Y T O R C H 1 . 3 AUTHOR A MODEL IN PYTORCH implementation 'org.pytorch:pytorch_ android:1.3.0' pod ‘LibTorch’ qmodel = quantization.convert(my_mobile_model) torch.jit.script(qmodel).save(“my_mobile_model.pt") C O M I N G S O O N • Build level optimization and selective compilation • Whole program optimization with link time optimization End-to-end workflows for mobile in iOS 
 and Android: EXPERIMENTAL
QUANTIZATION P Y T O R C H 1 . 3 model = ResNet50() model.load_state_dict(torch.load("model.pt")) qmodel = quantization.prepare( model, {"": quantization.default_qconfig}) qmodel.eval() for batch, target in data_loader: model(batch) qmodel = quantization.convert(qmodel) 4XL E S S M E M O R Y 
 U S A G E 2-4XS P E E D U P S I N 
 C O M P U T E EXPERIMENTAL • Neural networks inference is expensive • IoT and mobile devices have limited resources • Quantizing models enables efficient inference at scale
H O W D O I U S E I T ?
H O W D O I U S E I T ? TorchScript A static, high-performance subset of Python. 1. Prototype your model with PyTorch 2. Control flow is preserved 3. First-class support for lists, dicts, etc. import torch class MyModule(torch.nn.Module):     def __init__(self, N, M, state: List[Tensor]):         super(MyModule, self).__init__()         self.weight = torch.nn.Parameter(torch.rand(N, M))         self.state = state     def forward(self, input):         self.state.append(input)         if input.sum() > 0:             output = self.weight.mv(input)         else:             output = self.weight + input         return output # Compile the model code to a static representation my_module = MyModule(3, 4, [torch.rand(3, 4)]) my_script_module = torch.jit.script(my_module) # Save the compiled code and model data  # so it can be loaded elsewhere my_script_module.save("my_script_module.pt")
H O W D O I U S E I T ? TorchScript A static, high-performance subset of Python. 1. Prototype your model with PyTorch 2. Control flow is preserved 3. First-class support for lists, dicts, etc. import torch class MyModule(torch.nn.Module):     def __init__(self, N, M, state: List[Tensor]):         super(MyModule, self).__init__()         self.weight = torch.nn.Parameter(torch.rand(N, M))         self.state = state     def forward(self, input):         self.state.append(input)         if input.sum() > 0:             output = self.weight.mv(input)         else:             output = self.weight + input         return output # Compile the model code to a static representation my_module = MyModule(3, 4, [torch.rand(3, 4)]) my_script_module = torch.jit.script(my_module) # Save the compiled code and model data  # so it can be loaded elsewhere my_script_module.save("my_script_module.pt")
H O W D O I U S E I T ? # Compile the model code to a static representation my_module = MyModule(3, 4, [torch.rand(3, 4)]) my_script_module = torch.jit.script(my_module) # Save the compiled code and model data  # so it can be loaded elsewhere my_script_module.save("my_script_module.pt")
H O W D O I U S E I T ? ANDROID iOS implementation   'org.pytorch:pytorch_android:1.3.0' pod 'LibTorch'
H O W D O E S I T W O R K ? ANDROID iOS
H O W D O E S I T W O R K ? ANDROID iOS https://github.com/pytorch/android-demo-app https://github.com/pytorch/ios-demo-app
W H A T ' S H E R E T O D A Y ? Full TorchScript support. Pre-built binary releases in JCenter and CocoaPods. Java bindings. All forward CPU operators. Some optimized float operators (based on Caffe2Go). Some optimized quantized operators (based on QNNPACK w/ XNNPACK WIP).
W H A T ' S C O M I N G U P ? Faster. Smaller. Customized builds. Obj-C/Swift API? Kotlin wrapper? GPU support?? Accelerator support??
T H A N K Y O U

More Related Content

PDF
PyTorch Introduction
byYash Kawdiya
 
PDF
PyTorch Python Tutorial | Deep Learning Using PyTorch | Image Classifier Usin...
byEdureka!
 
PPTX
Introduction to PyTorch
byJun Young Park
 
PDF
Introduction to TensorFlow 2.0
byDatabricks
 
PPTX
Introduction to Keras
byJohn Ramey
 
PDF
Recurrent Neural Networks, LSTM and GRU
byananth
 
PDF
Recurrent Neural Networks (RNN) | RNN LSTM | Deep Learning Tutorial | Tensorf...
byEdureka!
 
PPTX
Pytorch
byehsan tr
 
PyTorch Introduction
byYash Kawdiya
 
PyTorch Python Tutorial | Deep Learning Using PyTorch | Image Classifier Usin...
byEdureka!
 
Introduction to PyTorch
byJun Young Park
 
Introduction to TensorFlow 2.0
byDatabricks
 
Introduction to Keras
byJohn Ramey
 
Recurrent Neural Networks, LSTM and GRU
byananth
 
Recurrent Neural Networks (RNN) | RNN LSTM | Deep Learning Tutorial | Tensorf...
byEdureka!
 
Pytorch
byehsan tr
 

What's hot

PPTX
Recurrent Neural Network (RNN) | RNN LSTM Tutorial | Deep Learning Course | S...
bySimplilearn
 
PDF
Xgboost
byVivian S. Zhang
 
PPTX
Deep learning with keras
byMOHITKUMAR1379
 
PDF
Introduction to Deep Learning, Keras, and TensorFlow
bySri Ambati
 
PPTX
Deep learning
byRajgupta258
 
PDF
Introduction To TensorFlow
bySpotle.ai
 
PDF
Feature Engineering - Getting most out of data for predictive models - TDC 2017
byGabriel Moreira
 
PPTX
Lstm
byMehrnaz Faraz
 
PPTX
Image classification using convolutional neural network
byKIRAN R
 
PPTX
What is TensorFlow? | Introduction to TensorFlow | TensorFlow Tutorial For Be...
bySimplilearn
 
PPTX
Image classification using CNN
byNoura Hussein
 
PDF
Introduction to Recurrent Neural Network
byYan Xu
 
PDF
Modern Convolutional Neural Network techniques for image segmentation
byGioele Ciaparrone
 
PDF
Deformable DETR Review [CDM]
byDongmin Choi
 
PPTX
Convolutional Neural Network - CNN | How CNN Works | Deep Learning Course | S...
bySimplilearn
 
PPTX
Attention Is All You Need
byIllia Polosukhin
 
PPTX
Rnn & Lstm
bySubash Chandra Pakhrin
 
PDF
Deep Feed Forward Neural Networks and Regularization
byYan Xu
 
PPTX
Autoencoders in Deep Learning
bymilad abbasi
 
PDF
論文紹介 "DARTS: Differentiable Architecture Search"
byYuta Koreeda
 
Recurrent Neural Network (RNN) | RNN LSTM Tutorial | Deep Learning Course | S...
bySimplilearn
 
Xgboost
byVivian S. Zhang
 
Deep learning with keras
byMOHITKUMAR1379
 
Introduction to Deep Learning, Keras, and TensorFlow
bySri Ambati
 
Deep learning
byRajgupta258
 
Introduction To TensorFlow
bySpotle.ai
 
Feature Engineering - Getting most out of data for predictive models - TDC 2017
byGabriel Moreira
 
Lstm
byMehrnaz Faraz
 
Image classification using convolutional neural network
byKIRAN R
 
What is TensorFlow? | Introduction to TensorFlow | TensorFlow Tutorial For Be...
bySimplilearn
 
Image classification using CNN
byNoura Hussein
 
Introduction to Recurrent Neural Network
byYan Xu
 
Modern Convolutional Neural Network techniques for image segmentation
byGioele Ciaparrone
 
Deformable DETR Review [CDM]
byDongmin Choi
 
Convolutional Neural Network - CNN | How CNN Works | Deep Learning Course | S...
bySimplilearn
 
Attention Is All You Need
byIllia Polosukhin
 
Rnn & Lstm
bySubash Chandra Pakhrin
 
Deep Feed Forward Neural Networks and Regularization
byYan Xu
 
Autoencoders in Deep Learning
bymilad abbasi
 
論文紹介 "DARTS: Differentiable Architecture Search"
byYuta Koreeda
 

Similar to "PyTorch Deep Learning Framework: Status and Directions," a Presentation from Facebook

PDF
Machine learning with py torch
byRiza Fahmi
 
PDF
Reproducible AI using MLflow and PyTorch
byDatabricks
 
PPTX
Soumith Chintala - Increasing the Impact of AI Through Better Software
byMLconf
 
PPTX
Transfer Leaning Using Pytorch synopsis Minor project pptx
byAnkit Gupta
 
PDF
Pytorch A Detailed Overview Agladze Mikhail
byilzobrzan47
 
PDF
pytdddddddddddddddddddddddddddddddddorch.pdf
bydrjigarsoni28
 
PDF
PyTorch Deep Learning Framework | USDSI®
byUSDSI
 
PPTX
Python - Deep Learning library-PyTorch.pptx
byakbarsayeedi32
 
PDF
pytorch-cheatsheet.pdf for ML study with pythroch
byJunZhao68
 
PPTX
python_libraries_for_artificial_intelligence.pptx
bysalehaalsaleh602
 
PPTX
Machine Learning Toolssssssssssssss.pptx
bysalehaalsaleh602
 
PPTX
Essentials of PyTorch Tutorial .pptx
bysurajkumartpoint
 
PDF
Scaling Up AI Research to Production with PyTorch and MLFlow
byDatabricks
 
PDF
Pytorch for tf_developers
byAbdul Muneer
 
PDF
Dive Into PyTorch
byIllarion Khlestov
 
PPTX
pytorch_tutorial_follow_this_to_start.pptx
bygyungmindenniskim
 
PPTX
PyTorch Tutorial for NTU Machine Learing Course 2017
byYu-Hsun (lymanblue) Lin
 
PDF
TensorFlow meetup: Keras - Pytorch - TensorFlow.js
byStijn Decubber
 
PDF
Icpp power ai-workshop 2018
byGanesan Narayanasamy
 
PDF
OpenPOWER Workshop in Silicon Valley
byGanesan Narayanasamy
 
Machine learning with py torch
byRiza Fahmi
 
Reproducible AI using MLflow and PyTorch
byDatabricks
 
Soumith Chintala - Increasing the Impact of AI Through Better Software
byMLconf
 
Transfer Leaning Using Pytorch synopsis Minor project pptx
byAnkit Gupta
 
Pytorch A Detailed Overview Agladze Mikhail
byilzobrzan47
 
pytdddddddddddddddddddddddddddddddddorch.pdf
bydrjigarsoni28
 
PyTorch Deep Learning Framework | USDSI®
byUSDSI
 
Python - Deep Learning library-PyTorch.pptx
byakbarsayeedi32
 
pytorch-cheatsheet.pdf for ML study with pythroch
byJunZhao68
 
python_libraries_for_artificial_intelligence.pptx
bysalehaalsaleh602
 
Machine Learning Toolssssssssssssss.pptx
bysalehaalsaleh602
 
Essentials of PyTorch Tutorial .pptx
bysurajkumartpoint
 
Scaling Up AI Research to Production with PyTorch and MLFlow
byDatabricks
 
Pytorch for tf_developers
byAbdul Muneer
 
Dive Into PyTorch
byIllarion Khlestov
 
pytorch_tutorial_follow_this_to_start.pptx
bygyungmindenniskim
 
PyTorch Tutorial for NTU Machine Learing Course 2017
byYu-Hsun (lymanblue) Lin
 
TensorFlow meetup: Keras - Pytorch - TensorFlow.js
byStijn Decubber
 
Icpp power ai-workshop 2018
byGanesan Narayanasamy
 
OpenPOWER Workshop in Silicon Valley
byGanesan Narayanasamy
 

More from Edge AI and Vision Alliance

PDF
“Transformer Networks: How They Work and Why They Matter,” a Presentation fro...
byEdge AI and Vision Alliance
 
PDF
“Understanding Human Activity from Visual Data,” a Presentation from Sportlogiq
byEdge AI and Vision Alliance
 
PDF
“Depth Estimation from Monocular Images Using Geometric Foundation Models,” a...
byEdge AI and Vision Alliance
 
PDF
“Introduction to Depth Sensing: Technologies, Trade-offs and Applications,” a...
byEdge AI and Vision Alliance
 
PDF
“Scaling Computer Vision at the Edge,” a Presentation from Invisible AI
byEdge AI and Vision Alliance
 
PDF
“Multimodal Enterprise-scale Applications in the Generative AI Era,” a Presen...
byEdge AI and Vision Alliance
 
PDF
“Deploying Accelerated ML and AI: The Role of Khronos Open Standards,” a Pres...
byEdge AI and Vision Alliance
 
PDF
“Toward Hardware-agnostic ADAS Implementations for Software-defined Vehicles,...
byEdge AI and Vision Alliance
 
PDF
“Object Detection Models: Balancing Speed, Accuracy and Efficiency,” a Presen...
byEdge AI and Vision Alliance
 
PDF
“Developing a GStreamer-based Custom Camera System for Long-range Biometric D...
byEdge AI and Vision Alliance
 
PDF
“A Lightweight Camera Stack for Edge AI,” a Presentation from Meta
byEdge AI and Vision Alliance
 
PDF
“Introduction to DNN Training: Fundamentals, Process and Best Practices,” a P...
byEdge AI and Vision Alliance
 
PDF
“Unlocking Visual Intelligence: Advanced Prompt Engineering for Vision-langua...
byEdge AI and Vision Alliance
 
PDF
“Lessons Learned Building and Deploying a Weed-killing Robot,” a Presentation...
byEdge AI and Vision Alliance
 
PDF
“Scaling Machine Learning with Containers: Lessons Learned,” a Presentation f...
byEdge AI and Vision Alliance
 
PDF
“Vision LLMs in Multi-agent Collaborative Systems: Architecture and Integrati...
byEdge AI and Vision Alliance
 
PDF
“Vision-language Models on the Edge,” a Presentation from Hugging Face
byEdge AI and Vision Alliance
 
PDF
“Real-world Deployment of Mobile Material Handling Robotics in the Supply Cha...
byEdge AI and Vision Alliance
 
PDF
“SKAIVISION: Transforming Automotive Dealerships with Computer Vision,” a Pre...
byEdge AI and Vision Alliance
 
PDF
“Scaling Artificial Intelligence and Computer Vision for Conservation,” a Pre...
byEdge AI and Vision Alliance
 
“Transformer Networks: How They Work and Why They Matter,” a Presentation fro...
byEdge AI and Vision Alliance
 
“Understanding Human Activity from Visual Data,” a Presentation from Sportlogiq
byEdge AI and Vision Alliance
 
“Depth Estimation from Monocular Images Using Geometric Foundation Models,” a...
byEdge AI and Vision Alliance
 
“Introduction to Depth Sensing: Technologies, Trade-offs and Applications,” a...
byEdge AI and Vision Alliance
 
“Scaling Computer Vision at the Edge,” a Presentation from Invisible AI
byEdge AI and Vision Alliance
 
“Multimodal Enterprise-scale Applications in the Generative AI Era,” a Presen...
byEdge AI and Vision Alliance
 
“Deploying Accelerated ML and AI: The Role of Khronos Open Standards,” a Pres...
byEdge AI and Vision Alliance
 
“Toward Hardware-agnostic ADAS Implementations for Software-defined Vehicles,...
byEdge AI and Vision Alliance
 
“Object Detection Models: Balancing Speed, Accuracy and Efficiency,” a Presen...
byEdge AI and Vision Alliance
 
“Developing a GStreamer-based Custom Camera System for Long-range Biometric D...
byEdge AI and Vision Alliance
 
“A Lightweight Camera Stack for Edge AI,” a Presentation from Meta
byEdge AI and Vision Alliance
 
“Introduction to DNN Training: Fundamentals, Process and Best Practices,” a P...
byEdge AI and Vision Alliance
 
“Unlocking Visual Intelligence: Advanced Prompt Engineering for Vision-langua...
byEdge AI and Vision Alliance
 
“Lessons Learned Building and Deploying a Weed-killing Robot,” a Presentation...
byEdge AI and Vision Alliance
 
“Scaling Machine Learning with Containers: Lessons Learned,” a Presentation f...
byEdge AI and Vision Alliance
 
“Vision LLMs in Multi-agent Collaborative Systems: Architecture and Integrati...
byEdge AI and Vision Alliance
 
“Vision-language Models on the Edge,” a Presentation from Hugging Face
byEdge AI and Vision Alliance
 
“Real-world Deployment of Mobile Material Handling Robotics in the Supply Cha...
byEdge AI and Vision Alliance
 
“SKAIVISION: Transforming Automotive Dealerships with Computer Vision,” a Pre...
byEdge AI and Vision Alliance
 
“Scaling Artificial Intelligence and Computer Vision for Conservation,” a Pre...
byEdge AI and Vision Alliance
 

Recently uploaded

PDF
Integrare AI e Automazione con UiPath Agent Builder
byUiPathCommunity
 
PDF
FME Realize in the Real World - The Power of Spatial Computing in Action
bySafe Software
 
PDF
Immer vorne mit dabei sein - Roaming leicht gemacht
bypanagenda
 
PDF
Staying Ahead of the Curve - Roaming Just Got Easier
bypanagenda
 
PDF
The Invisible Threads: How 'Connection' Builds the Smartest Organizations. By...
byTheta Prime Institute, Cheyenne, Wyoming, USA.
 
PDF
AI64 - Mapping the Next Generation of Enterprise Software | Redpoint Ventures
byRazin Mustafiz
 
PDF
Top 11 Sites to Buy Verified GitHub Accounts in 2025
byBuy GitHub Accounts Looking for a verified GitHub account?
 
PDF
Top-10-Cloud-Service-Companies-in-2025-Market-Leaders-and-Innovators.pdf
byArtjoker Software Development Company
 
PPTX
Taming Time in PHP: Best Practices and Gotchas at Longhorn PHP 2025
byScott Keck-Warren
 
PPTX
Artificial Intelligence lecture slides.pptx
bymaimelabernard6
 
PDF
Igalia and WebKit: Status update and plans (2025)
byIgalia
 
PDF
Energia Nowa and byteLAKE Partner to Develop Intelligent Services for Energy ...
bybyteLAKE
 
PDF
AI Demands More: Help Ensure Network Readiness With ThousandEyes
byThousandEyes
 
PPTX
Introduction to WordPress Workshop - WCEH 2025
byShanta Nathwani
 
PDF
Transcript: Embedding sustainability: Tips for ebook and print production - T...
byBookNet Canada
 
DOCX
Formula 3 - MACO using general limit.docx
byMarkus Janssen
 
DOCX
Formula 2 - APIC HBEL general calculation.docx
byMarkus Janssen
 
PDF
The State of Adventure Capital - Grant Gregory | Cantos
byRazin Mustafiz
 
PDF
Embedding sustainability: Tips for ebook and print production - Tech Forum 2025
byBookNet Canada
 
PDF
UiPath Community Day UNI - Nuevos productos de UiPath.pdf
byfelixluen
 
Integrare AI e Automazione con UiPath Agent Builder
byUiPathCommunity
 
FME Realize in the Real World - The Power of Spatial Computing in Action
bySafe Software
 
Immer vorne mit dabei sein - Roaming leicht gemacht
bypanagenda
 
Staying Ahead of the Curve - Roaming Just Got Easier
bypanagenda
 
The Invisible Threads: How 'Connection' Builds the Smartest Organizations. By...
byTheta Prime Institute, Cheyenne, Wyoming, USA.
 
AI64 - Mapping the Next Generation of Enterprise Software | Redpoint Ventures
byRazin Mustafiz
 
Top 11 Sites to Buy Verified GitHub Accounts in 2025
byBuy GitHub Accounts Looking for a verified GitHub account?
 
Top-10-Cloud-Service-Companies-in-2025-Market-Leaders-and-Innovators.pdf
byArtjoker Software Development Company
 
Taming Time in PHP: Best Practices and Gotchas at Longhorn PHP 2025
byScott Keck-Warren
 
Artificial Intelligence lecture slides.pptx
bymaimelabernard6
 
Igalia and WebKit: Status update and plans (2025)
byIgalia
 
Energia Nowa and byteLAKE Partner to Develop Intelligent Services for Energy ...
bybyteLAKE
 
AI Demands More: Help Ensure Network Readiness With ThousandEyes
byThousandEyes
 
Introduction to WordPress Workshop - WCEH 2025
byShanta Nathwani
 
Transcript: Embedding sustainability: Tips for ebook and print production - T...
byBookNet Canada
 
Formula 3 - MACO using general limit.docx
byMarkus Janssen
 
Formula 2 - APIC HBEL general calculation.docx
byMarkus Janssen
 
The State of Adventure Capital - Grant Gregory | Cantos
byRazin Mustafiz
 
Embedding sustainability: Tips for ebook and print production - Tech Forum 2025
byBookNet Canada
 
UiPath Community Day UNI - Nuevos productos de UiPath.pdf
byfelixluen
 

"PyTorch Deep Learning Framework: Status and Directions," a Presentation from Facebook

  • 1.
    J O SE P H S P I S A K P R O D U C T M A N A G E R P Y T O R C H | O N D E V I C E
  • 2.
    OVERVIEW AND INTEGRATION PYTORCHMOBILE STATE OF THE STATE ON DEVICE DYNAMIC VS. STATIC FRAMEWORKS PRINCIPLES, MAKEUP, .. OVERVIEW
  • 3.
  • 4.
  • 5.
  • 6.
    P Y TO R C H O V E R V I E W
  • 7.
  • 8.
    P Y TO R C H R E S E A R C H P R O T O T Y P I N G P R O D U C T I O N D E P L O Y M E N T +
  • 9.
    C O RE P R I N C I P L E S BUILDING FOR SCALE DEVELOPER EFFICIENCY
  • 10.
    DEVELOPER EFFICIENCY ENABLING AHIGH VELOCITY OF MODEL ITERATION AND INNOVATION
  • 11.
    ` C L EA N A P I S
  • 12.
    ` T O RC H S C R I P T Models are Python TorchScript programs, an optimizable subset of Python + Same “models are programs” idea + Production deployment + No Python dependency + Compilation for performance optimization class RNN(nn.Module):   def __init__(self, W_h, U_h, W_y, b_h, b_y):     super(RNN, self).__init__()     self.W_h = nn.Parameter(W_h)     self.U_h = nn.Parameter(U_h)     self.W_y = nn.Parameter(W_y)     self.b_h = nn.Parameter(b_h)     self.b_y = nn.Parameter(b_y)   def forward(self, x, h):     y = []     for t in range(x.size(0)):       h = torch.tanh(x[t] @ self.W_h + h @ self.U_h + self.b_h)       y += [torch.tanh(h @ self.W_y + self.b_y)]       if t % 10 == 0:         print("stats: ", h.mean(), h.var())     return torch.stack(y), h 
 # one annotation! script_rnn = torch.jit.script(RNN(W_h, U_h, W_y, b_h, b_y))
  • 13.
    ` T O RC H S C R I P T Models are Python TorchScript programs, an optimizable subset of Python + Same “models are programs” idea + Prod deployment + No Python dependency + Optimizable (incl. codegen!)
  • 14.
    ` T E NS O R B O A R D
  • 15.
    ~1,230C O NT R I B U T O R S 50%+Y O Y G R O W T H 23KP Y T O R C H F O R U M U S E R S
  • 16.
    GROW TH INARXIV MENTIONS IN RESEARCH PAPERS 0 100 200 300 400 500 Jan 17 Feb 17 M ar17 Apr17 M ay17 Jun 17 Jul17 Aug17 Sep 17 Jan 18 Feb 18 M ar18 Apr18 M ay18 Jun 18 Jul18 Aug18 Sep 18 Jan 19 Feb 19 M ar19 Apr19 M ay19 Jun 19 Jul19
  • 17.
    F R AM E W O R K S D Y N A M I C V S . S T A T I C
  • 18.
  • 19.
    DECLARATIVETOOLKITS Declare and compilea model Repeatedly execute the model in a VM TOOLKIT VM PYTHON SCRIPT RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM
  • 20.
    import tensorflow astf
 import numpy as np trX = np.linspace(-1, 1, 101)
 trY = 2 * trX + np.random.randn(*trX.shape) * 0.33 X = tf.placeholder("float")
 Y = tf.placeholder("float") def model(X, w): return tf.multiply(X, w) w = tf.Variable(0.0, name="weights")
 y_model = model(X, w) cost = tf.square(Y - y_model) train_op = tf.train.GradientDescentOptimizer(0.01).minimize(cost) with tf.Session() as sess: tf.global_variables_initializer().run() for i in range(100): for (x, y) in zip(trX, trY): sess.run(train_op, feed_dict={X: x, Y: y}) print(sess.run(w)) DECLARATIVETOOLKITS Computation Graph • Declare a computation • Placeholder variables • Compile it • Run it in a Session
  • 21.
    import tensorflow astf
 import numpy as np trX = np.linspace(-1, 1, 101)
 trY = 2 * trX + np.random.randn(*trX.shape) * 0.33 X = tf.placeholder("float")
 Y = tf.placeholder("float") def model(X, w): return tf.multiply(X, w) w = tf.Variable(0.0, name="weights")
 y_model = model(X, w) cost = tf.square(Y - y_model) train_op = tf.train.GradientDescentOptimizer(0.01).minimize(cost) with tf.Session() as sess: tf.global_variables_initializer().run() for i in range(100): for (x, y) in zip(trX, trY): sess.run(train_op, feed_dict={X: x, Y: y}) print(sess.run(w)) X = tf.placeholder("float")
 Y = tf.placeholder("float") DECLARATIVETOOLKITS Computation Graph • Declare a computation • Placeholder variables • Compile it • Run it in a Session
  • 22.
    import tensorflow astf
 import numpy as np trX = np.linspace(-1, 1, 101)
 trY = 2 * trX + np.random.randn(*trX.shape) * 0.33 X = tf.placeholder("float")
 Y = tf.placeholder("float") def model(X, w): return tf.multiply(X, w) w = tf.Variable(0.0, name="weights")
 y_model = model(X, w) cost = tf.square(Y - y_model) train_op = tf.train.GradientDescentOptimizer(0.01).minimize(cost) with tf.Session() as sess: tf.global_variables_initializer().run() for i in range(100): for (x, y) in zip(trX, trY): sess.run(train_op, feed_dict={X: x, Y: y}) print(sess.run(w)) Model definition def model(X, w): return tf.multiply(X, w) w = tf.Variable(0.0, name="weights")
 y_model = model(X, w) cost = tf.square(Y - y_model) train_op = tf.train.GradientDescentOptimizer(0.01).minimize(cost) DECLARATIVETOOLKITS Computation Graph • Declare a computation • Placeholder variables • Compile it • Run it in a Session
  • 23.
    import tensorflow astf
 import numpy as np trX = np.linspace(-1, 1, 101)
 trY = 2 * trX + np.random.randn(*trX.shape) * 0.33 X = tf.placeholder("float")
 Y = tf.placeholder("float") def model(X, w): return tf.multiply(X, w) w = tf.Variable(0.0, name="weights")
 y_model = model(X, w) cost = tf.square(Y - y_model) train_op = tf.train.GradientDescentOptimizer(0.01).minimize(cost) with tf.Session() as sess: tf.global_variables_initializer().run() for i in range(100): for (x, y) in zip(trX, trY): sess.run(train_op, feed_dict={X: x, Y: y}) print(sess.run(w)) sess.run(train_op, feed_dict={X: x, Y: y}) print(sess.run(w)) A separate, Turing complete, virtual machine. for i in range(100): for (x, y) in zip(trX, trY): DECLARATIVETOOLKITS Computation Graph • Declare a computation • Placeholder variables • Compile it • Run it in a Session
  • 24.
  • 25.
    IMPERATIVETOOLKITS Run a seriesof computation Implicitly defining the model as execution goes PYTHON NATIVE RUNTIME PYTHON INSTRUCTIONS RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM RELU CONNV2D BATCHNORM
  • 26.
    Imperative Toolkits import torch
 fromtorch.autograd import Variable trX = torch.linspace(-1, 1, 101)
 trY = 2 * trX + torch.random(*trX.size()) * 0.33 w = Variable(trX.new([0.0]), requires_grad=True) for i in range(100): for (x, y) in zip(trX, trY): X = Variable(x) Y = Variable(y) print(X) print(Y) y_model = X * w.expand_as(X) cost = (Y - y_model) * 2 Cost.backward(torch.ones(*cost.size())) w.data = w.data + 0.01 * w.grad.data print(w) • Define a model by execution • No separate compilation stage • No separate execution engine
  • 27.
    import torch
 from torch.autogradimport Variable trX = torch.linspace(-1, 1, 101)
 trY = 2 * trX + torch.random(*trX.size()) * 0.33 w = Variable(trX.new([0.0]), requires_grad=True) for i in range(100): for (x, y) in zip(trX, trY): X = Variable(x) Y = Variable(y) print(X) print(Y) y_model = X * w.expand_as(X) cost = (Y - y_model) * 2 Cost.backward(torch.ones(*cost.size())) w.data = w.data + 0.01 * w.grad.data print(w) Imperative Toolkits Model constructed and values computed as we define it. • Define a model by execution • No separate compilation stage • No separate execution engine
  • 28.
    P Y TO R C H F O R E M B E D D E D S T A T E O F T H E S T A T E
  • 29.
    H O WD O I R U N P Y T O R C H M O D E L S O N D E V I C E ?
  • 30.
    H O WD O I R U N P Y T O R C H M O D E L S O N D E V I C E ? E X P O R T O N N X F O R M A T T E D M O D E L S
  • 31.
    H O WD O I R U N P Y T O R C H M O D E L S O N D E V I C E ? E X P O R T O N N X F O R M A T T E D M O D E L S P Y T O R C H M O B I L E
  • 32.
    P Y TO R C H O N N X S U P P O R T
  • 33.
    O N NX E c o s y s t e m
  • 34.
    A R CH I T E C T U R E 
 A N D F L O W JIT Tracer and Torchscript ONNX
 Exporter Optimizer Torch IR to ONNX IR Translator Torch IR ONNX torch.onnx
 .export() PyTorch Model Sample Input ONNX Graph
  • 35.
  • 36.
    P Y TO R C H M O B I L E P R E V I E W R E L E A S E
  • 37.
    W H AT I S   P Y T O R C H   M O B I L E ?
  • 38.
    W H AT I S   P Y T O R C H   M O B I L E ? I T ’ S P Y T O R C H
  • 39.
    W H AT I S   P Y T O R C H   M O B I L E ? I T ’ S P Y T O R C H F O R M O B I L E 😃
  • 40.
    W H AT I S   P Y T O R C H   M O B I L E ? I T ’ S P Y T O R C H F O R M O B I L E B U T N O P Y T H O N 😃
  • 41.
    W H AT C A N I T R U N ? A N Y T O R C H S C R I P T M O D E L .
  • 42.
    W H AT C A N I T R U N ? A N Y T O R C H S C R I P T M O D E L . L O O P S ? Y E S
  • 43.
    W H AT C A N I T R U N ? A N Y T O R C H S C R I P T M O D E L . L O O P S ? F U N C T I O N S ? Y E S Y E S
  • 44.
    W H AT C A N I T R U N ? A N Y T O R C H S C R I P T M O D E L . L O O P S ? F U N C T I O N S ? T U P L E S ? Y E S Y E S 
 Y E S
  • 45.
    W H AT C A N I T R U N ? A N Y T O R C H S C R I P T M O D E L . L O O P S ? F U N C T I O N S ? T U P L E S ? N A M E D T U P L E ? Y E S Y E S 
 Y E S Y E S
  • 46.
    ANDROID - MAVENiOS - COCOAPODS MODEL OPTIMIZATION (OPTIONAL ) PY TORCH MOBILE • No separate runtime to export P Y T O R C H 1 . 3 AUTHOR A MODEL IN PYTORCH implementation 'org.pytorch:pytorch_ android:1.3.0' pod ‘LibTorch’ qmodel = quantization.convert(my_mobile_model) torch.jit.script(qmodel).save(“my_mobile_model.pt") C O M I N G S O O N • Build level optimization and selective compilation • Whole program optimization with link time optimization End-to-end workflows for mobile in iOS 
 and Android: EXPERIMENTAL
  • 47.
    QUANTIZATION P Y TO R C H 1 . 3 model = ResNet50() model.load_state_dict(torch.load("model.pt")) qmodel = quantization.prepare( model, {"": quantization.default_qconfig}) qmodel.eval() for batch, target in data_loader: model(batch) qmodel = quantization.convert(qmodel) 4XL E S S M E M O R Y 
 U S A G E 2-4XS P E E D U P S I N 
 C O M P U T E EXPERIMENTAL • Neural networks inference is expensive • IoT and mobile devices have limited resources • Quantizing models enables efficient inference at scale
  • 48.
    H O WD O I U S E I T ?
  • 49.
    H O WD O I U S E I T ? TorchScript A static, high-performance subset of Python. 1. Prototype your model with PyTorch 2. Control flow is preserved 3. First-class support for lists, dicts, etc. import torch class MyModule(torch.nn.Module):     def __init__(self, N, M, state: List[Tensor]):         super(MyModule, self).__init__()         self.weight = torch.nn.Parameter(torch.rand(N, M))         self.state = state     def forward(self, input):         self.state.append(input)         if input.sum() > 0:             output = self.weight.mv(input)         else:             output = self.weight + input         return output # Compile the model code to a static representation my_module = MyModule(3, 4, [torch.rand(3, 4)]) my_script_module = torch.jit.script(my_module) # Save the compiled code and model data  # so it can be loaded elsewhere my_script_module.save("my_script_module.pt")
  • 50.
    H O WD O I U S E I T ? TorchScript A static, high-performance subset of Python. 1. Prototype your model with PyTorch 2. Control flow is preserved 3. First-class support for lists, dicts, etc. import torch class MyModule(torch.nn.Module):     def __init__(self, N, M, state: List[Tensor]):         super(MyModule, self).__init__()         self.weight = torch.nn.Parameter(torch.rand(N, M))         self.state = state     def forward(self, input):         self.state.append(input)         if input.sum() > 0:             output = self.weight.mv(input)         else:             output = self.weight + input         return output # Compile the model code to a static representation my_module = MyModule(3, 4, [torch.rand(3, 4)]) my_script_module = torch.jit.script(my_module) # Save the compiled code and model data  # so it can be loaded elsewhere my_script_module.save("my_script_module.pt")
  • 51.
    H O WD O I U S E I T ? # Compile the model code to a static representation my_module = MyModule(3, 4, [torch.rand(3, 4)]) my_script_module = torch.jit.script(my_module) # Save the compiled code and model data  # so it can be loaded elsewhere my_script_module.save("my_script_module.pt")
  • 52.
    H O WD O I U S E I T ? ANDROID iOS implementation   'org.pytorch:pytorch_android:1.3.0' pod 'LibTorch'
  • 53.
    H O WD O E S I T W O R K ? ANDROID iOS
  • 54.
    H O WD O E S I T W O R K ? ANDROID iOS https://github.com/pytorch/android-demo-app https://github.com/pytorch/ios-demo-app
  • 55.
    W H AT ' S H E R E T O D A Y ? Full TorchScript support. Pre-built binary releases in JCenter and CocoaPods. Java bindings. All forward CPU operators. Some optimized float operators (based on Caffe2Go). Some optimized quantized operators (based on QNNPACK w/ XNNPACK WIP).
  • 56.
    W H AT ' S C O M I N G U P ? Faster. Smaller. Customized builds. Obj-C/Swift API? Kotlin wrapper? GPU support?? Accelerator support??
  • 57.
    T H AN K Y O U