Course UE D-Deep Learning at IMT Atlantique

Made by:

• David HURTADO @dr-hurtado
• Zijie NING @mm0806son

Thesis Reproduction and application: Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields

Main Activities:

• Read the paper to understand the mathematical principles and network structure.
• Read the code and understand the main functions.
• Development of new applications based on experimental conditions.

Our own applications:

• Processing of single images

  • Body joint confidence map
  

  • Part Affinity Fields (PAFs)

    

  • Generated body skeleton

    

• Processing video

• Use Google Colab to call a local webcam and process the images

• Layer visualization to see what happened in the network

  • For example, after the preprocessing network (here model0, first 10 layers of vgg19)

    

  • 3rd stage of PAFs branch (model3_2, first 20 output)

    

  • 3rd stage of confidence map branch (model3_2)

    

• Use Google Colab to call a local webcam and process the video (TODO)

• Control of a robot simulator using pybullet (TODO)

2021/10/01 & 08

Introduction to PyTorch

• Linear regression
• Logistique regression

2021/10/15

Multi-Layer Perception with MNIST dataset

• Build a 2-layer Neuro network
• Evolution of the loss function for both training and validation sets

2021/10/15

Convolutional Neural Networks (CNN) on MNIST

• Create data loaders
• Build and train a CNN network

2021/10/29 Graded session

Fashion-MNIST dataset

• Data loader

• MLP

• CNN

• Transfer learning with vgg-16

  Accuracy: 93.71% (1st in the class ✌️)

  Pretrained VGG16 with lr = 0.01, epoch = 7, Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))

2021/11/12 Graded session

Recurrent Neural Networks (RNN) on the the Lorenz-63 system (chaos system)