H ANDS - ON T UTORIAL DEEP LEARNING (F OR R OBOTIC V ISION ) Juxi Leitner @Juxi http://Juxi.net R ESEARCHER R OBOTICS /AI h7p:/ /Juxi.net/workshop/SoAIR/
roboticvision.org
http://Juxi.net/aboutme create agents that see & interact with the real-world Pixar (2008) http://roboticvision.org/
deep learning for vision-based manipulation
H ANDS - ON T UTORIAL introduction to deep learning first contact with pytorch create a neural network train a classifier $$$ @Juxi
what is deep learning @Juxi
frameworks for deep learning @Juxi Read more at Wikipedia or check this blog for an overview https://jameskle.com/writes/deep-learning-frameworks
pytorch (2016) FAIR open source library tensor computation on GPU + autodif @Juxi
let’s play > jupyter notebook Install via conda conda create --name rvss2019 python=3 conda activate rvss2019 conda install jupyter conda install scipy numpy conda install pytorch=0.4.1 torchvision -c pytorch conda install opencv @Juxi https://realpython.com/jupyter-notebook-introduction/
what is deep learning @Juxi
what is machine learning supervised learning … @Juxi
what is machine learning unsupervised learning @Juxi
what is machine learning reinforcement learning run see agent act hug reward agent ? age @Juxi
machine learning classification regression @Juxi
@Juxi
@Juxi http://www.asimovinstitute.org/neural-network-zoo/
artificial neuron @Juxi
activation functions @Juxi
back propagation @Juxi
training step @Juxi
gradient descent @Juxi
train a simple neural network @Juxi
convolutional neural networks Convolution of 5 x 5 pixel image with 3 x 3 pixel filter (stride = 1 x 1 pixel) max_pooling @Juxi
convolutional neural networks https://towardsdatascience.com/image-classification-in-10-minutes-with-mnist-dataset-54c35b77a38d @Juxi
train a convolutional neural network Want to try more: https://pytorch.org/tutorials/beginner/blitz/cifar10_tutorial.html @Juxi
break a neural network @Juxi
debugging networks http://cs231n.github.io/understanding-cnn/ use network architectures https://modelzoo.co/framework/pytorch @Juxi
deep learning visual control fully conn. + ReLU fully conn. + ReLU 7 × 7 conv + ReLU 4 × 4 conv + ReLU 3 × 3 conv + ReLU 84 × 84 fully conn. fully conn. stride 2 stride 2 stride 1 Q-values 400 units 300 units 5 units 9 units Or 64 � lters 64 � lters 64 � lters θ Conv1 Conv2 Conv3 I BN FC_c1 FC_c2 FC_c3 [Zhang et al, arxiv ] Bottleneck Control Module Perception Module Occlusion Occlusion Occlusion Occlusion A B C D E
deep learning visual control [Zhuang et al, RA:L 2018 (submitted) ]
generative grasp network [Morrison et al, RSS 2018]
thank you Juxi Leitner arc centre of excellence for robotic vision queensland university of technology j.leitner@qut.edu.au http://Juxi.net Juxi
Recommend
More recommend