Learning Slip Perception

This repository contains python scripts to train a slip detection Temporal Convolution Network (TCN) using the data collected from the Takktile sensor kit. Details about this work can be found in this paper: Under Pressure: Learning to Detect Slip with Barometric Tactile Sensors.

Table of Content

• Learning Slip Perception
• Setup
  • Pre-requisites
  • Installation
• Usage
  • Training a New Model
  • Monitoring Training with Tensorboard
• Software Architecture

Setup

Pre-requisites

The training scripts are all written in Python2, in order to ensure compatibility with ROS Kinetic. The authors use Keras for building and training the TCN models.

• Install TensorFlow 2.1.0 (CPU or GPU) by following the instructions here. Keras is now packaged as part of TensorFlow framework. This is last TensorFlow version that supports Python2.

• Email the author to request access to the Takktile training dataset.

Installation

• Clone this repository

• cd <path to learning-slip-perception>

• git submodule init && git submodule update

Usage

The main script takes a yaml configuration file as input, which provides values for all the data, network, and training parameters. The base config files are present in the configs directory.

Training a New Model

• Modify the base config files to create a training experiment. Modify the data_home parameter and replace it with the path to your Takktile data folder. You can use the epochs parameter to controls the number of epochs.

• python2.7 main.py ./configs/<file name>.yaml

  replace the name with that of the config file you edited.
• A new folder will be created with in the models directory.

Continue Training of an Exisiting Model

• Locate the model you would like to continue training. (A copy of the config file is saved with each model in the models directory)
• Modify the epochs parameter to indicate the number of additional epochs you would like to train.

• python2.7 main.py ./logs/models/<model_name>/config.yaml
  

  replace the name with that of the config file you edited. No new directory will be created.
• You can use the bash script, ./train.sh, to automate training experiments.

Monitoring Training with Tensorboard

The tensorboard scalars are saves in the scalars directory.

tensorboard --logdir ./logs/scalars/<scaler_dir>

Software Architecture

TODO