Deep learning with C++ - an introduction to tiny-dnn

1. deep learning with c++ an introduction to tiny-dnn by Taiga Nomi embedded software engineer, Osaka, Japan

2. deep learning Icons made by Freepik from www.flaticon.com is licensed by CC 3.0 BY Facial recognition Image understanding Finance Game playing Translation Robotics Drug discovery Text recognition Video processing Text generation

3. Deep learning - Learning a complicated function from many data - Composed of trainable, simple mathematical functions Input OutputTrainable Building Blocks - text - audio - image - video - ... - text - audio - image - video - ...

4. deep learning framework

5. A modern deep learning framework for C++ programmers

6. 1400 stars 500 forks 35 contributors 100 clones/day

7. “A Modern Deep Learning module” by Edgar Riba “Deep Learning with Quantization for Semantic Saliency Detection” by Yida Wang https://summerofcode.withgoogle.com/archive/

8. 1.Easy to introduce 2.Simple syntax 3.Extensible backends

9. 1.Easy to introduce - Just put the following line into your cpp tiny-dnn is header only - No installation tiny-dnn is dependency-free - No prerequisites #include <tiny_dnn/tiny_dnn.h>

10. 1.Easy to introduce - You can bring Deep Learning to any target you have a C++ compiler - Officially supported (by CI builds): - Windows (msvc2013 32/64bit, msvc2015 32/64bit) - Linux (gcc4.9, clang3.5) - OSX(LLVM 7.3) - tiny-dnn might run on other compiler that support C++11

11. 1.Easy to introduce - Caffe model converter is also available - TensorFlow converter - coming soon! - Close the gap between researcher and engineer

13. 2.Simple syntax Example: Multi layer perceptron

14. Caffe prototxt input: "data" input_shape { dim: 1 dim: 1 dim: 1 dim: 20 } layer { name: "ip1" type: "InnerProduct" inner_product_param { num_output: 100 } bottom: "ip1" top: "ip2" } layer { name: "a1" type: "TanH" bottom: "ip1" top: "ip1" } layer { name: "ip2" type: "InnerProduct" inner_product_param { num_output: 10 } bottom: "ip1" top: "out" } layer { name: "a1" type: "TanH" bottom: "out" top: "out" }

15. Tensorflow w1 = tf.Variable(tf.random_normal([10, 100])) w2 = tf.Variable(tf.random_normal([100, 20])) b1 = tf.Variable(tf.random_normal([100])) b2 = tf.Variable(tf.random_normal([20])) layer1 = tf.add(tf.matmul(x, w1), b1) layer1 = tf.nn.relu(layer1) layer2 = tf.add(tf.matmul(x, w2), b2) layer2 = tf.nn.relu(layer2)

16. Keras model = Sequential([ Dense(100, input_dim=10), Activation('relu'), Dense(20), Activation('relu'), ])

17. tiny-dnn network<sequential> net; net << dense<relu>(10, 100) << dense<relu>(100, 20);

18. tiny-dnn, another solution auto net = make_mlp<relu>({10, 100, 20}); - modern C++ enable us to keep code simple - type inference, initializer list

19. 2.Simple syntax Example: Convolutional Neural Networks

20. Caffe prototxt name: "LeNet" layer { name: "data" type: "Input" top: "data" input_param { shape: { dim: 64 dim: 1 dim: 28 dim: 28 } } } layer { name: "conv1" type: "Convolution" bottom: "data" top: "conv1" param { lr_mult: 1 } param { lr_mult: 2 } convolution_param { num_output: 20 kernel_size: 5 stride: 1 weight_filler { type: "xavier" } bias_filler { type: "constant" } } } layer { name: "pool1" type: "Pooling" bottom: "conv1" top: "pool1" pooling_param { pool: MAX kernel_size: 2 stride: 2 } } } } layer { name: "conv2" type: "Convolution" bottom: "pool1" top: "conv2" param { lr_mult: 1 } param { lr_mult: 2 } convolution_param { num_output: 50 kernel_size: 5 stride: 1 weight_filler { type: "xavier" } bias_filler { type: "constant" } } } layer { name: "pool2" type: "Pooling" bottom: "conv2" top: "pool2" pooling_param { pool: MAX kernel_size: 2 stride: 2 } } layer { name: "ip1" type: "InnerProduct" bottom: "pool2" top: "ip1" param { param { lr_mult: 1 } param { lr_mult: 2 } inner_product_param { num_output: 500 weight_filler { type: "xavier" } bias_filler { type: "constant" } } } layer { name: "relu1" type: "ReLU" bottom: "ip1" top: "ip1" } layer { name: "ip2" type: "InnerProduct" bottom: "ip1" top: "ip2" param { lr_mult: 1 } param { lr_mult: 2 } inner_product_param { num_output: 10 weight_filler { type: "xavier" } bias_filler { type: "constant" } } } } layer { name: "relu1" type: "ReLU" bottom: "ip1" top: "ip1" } layer { name: "ip2" type: "InnerProduct" bottom: "ip1" top: "ip2" param { lr_mult: 1 } param { lr_mult: 2 } inner_product_param { num_output: 10 weight_filler { type: "xavier" } bias_filler { type: "constant" } } } layer { name: "prob" type: "Softmax" bottom: "ip2" top: "prob" }

21. Tensorflow x = tf.Variable(tf.random_normal([-1, 28, 28, 1])) wc1 = tf.Variable(tf.random_normal([5, 5, 1, 32])) wc2 = tf.Variable(tf.random_normal([5, 5, 32, 64])) wd1 = tf.Variable(tf.random_normal([7*7*64, 1024])) wout = tf.Variable(tf.random_normal([1024, n_classes])) bc1 = tf.Variable(tf.random_normal([32])) bc2 = tf.Variable(tf.random_normal([64])) bd1 = tf.Variable(tf.random_normal([1024])) bout = tf.Variable(tf.random_normal([n_classes])) conv1 = conv2d(x, wc1, bc1) conv1 = maxpool2d(conv1, k=2) conv1 = tf.nn.relu(conv1) conv2 = conv2d(conv1, wc2, bc2) conv2 = maxpool2d(conv2, k=2) conv2 = tf.nn.relu(conv2) fc1 = tf.reshape(conv2, [-1, wd1.get_shape().as_list()[0]]) fc1 = tf.add(tf.matmul(fc1, wd1), bd1) fc1 = tf.nn.relu(fc1) fc1 = tf.nn.dropout(fc1, dropout) out = tf.add(tf.matmul(fc1, wout), bout)

22. Keras model = Sequential([ Convolution2D(32, 5, 5, input_shape=[28,28,5]), MaxPooling2D(pool_size=2), Activation('relu'), Convolution2D(64, 5, 5), MaxPooling2D(pool_size=2), Activation('relu'), Dense(1024), Dropout(0.5), Dense(10), ])

23. tiny-dnn network<sequential> net; net << conv<>(28, 28, 5, 1, 32) << max_pool<relu>(24, 24, 2) << conv<>(12, 12, 5, 32, 64) << max_pool<relu>(8, 8, 64, 2) << fc<relu>(4*4*64, 1024) << dropout(1024, 0.5f) << fc<>(1024, 10);

25. 3.Extensible backends Common scenario1: “We have a good GPU machine to train networks, but we need to deploy trained model into mobile device” Common scenario2: “We need to write platform-specific code to get production-level performance... but it’s painful to understand whole framework”

26. 3.Extensible backends Some performance critical layers have backend engine Layer API backend::internal pure-c++ code backend::avx avx-optimized code … backend::nnpack x86/ARM backend::opencl GPU Optional

27. 3.Extensible backends // select an engine explicitly net << conv<>(28, 28, 5, 1, 32, backend::avx) << ...; // switch them seamlessly net[0]->set_backend_type(backend::opencl);

28. Model serialization (binary/json) Regression training Basic image processing Layer freezing Graph visualization Multi-thread execution Double precision support Basic functionality

29. Caffe importer (requires protobuf) OpenMP support Intel TBB support NNPACK backend (same to caffe2) libdnn backend (same to caffe-opencl)Extra modules (requires 3rd-party libraries)

30. Future plans

31. - GPU integration - GPU backend is still experimental - cudnn backend - More mobile-oriented - iOS/Android examples - Quantized operation for less RAM - TensorFlow Importer - Performance profiling tools - OpenVX support We need your help!

32. User chat for QA: https://gitter.im/tiny-dnn Official documents: http://tiny-dnn.readthedocs.io/en/latest/ For users

33. Join our developer chat: https://gitter.im/tiny-dnn/developers or Check out docs, and our issues marked as “contributions welcome”: https://github.com/tiny-dnn/tiny-dnn/blob/master/docs/developer_gui des/How-to-contribute.md https://github.com/tiny-dnn/tiny-dnn/labels/contributions%20welcome For developers

34. code: github.com/tiny-dnn/tiny-dnn slide: https://goo.gl/Se2rzu

Deep learning with C++ - an introduction to tiny-dnn

大河 能見

Deep learning with C++ - an introduction to tiny-dnn

大河能見