Visualizing joint angles of grasps In this post I am going to explain how to use the master motor map (MMM) in order to visualize grasping movements with joint angle data I downloaded from handcorpus.org. The MMM has a Whole-body...

Some impressions of LSTM architectures for simple math functions: seq2seq, seq2vec and then seq-seq-autoencoder. In particular the last part is an experiment of reconstructing sinoid waves with phase displacement from a single latent parameter. Sequence to sequence X-Y I generated...

The odyssey of writing a YOLO-like detection model for German traffic signs. Simply said: Give a CNN an image and let it draw boxes around the traffic signs. Getting started slowly Before detecting concrete boxes I thought it would be...

I have been using tensorflow for some examples in this blog. Building a sequential model and evaluating some metrics was easy. But when I wanted to do a little bit more (for example image data set augmentation) I got really...

Review of the video course When looking up resources on how to get a good overview of neural networks for computer vision you almost always get pointed to the youtube videos of the CS231n course held at Stanford. I watched...

Finding a computer science master in Germany This blog post describes my journey of finding a place for a master in computer science in Germany. It could be interesting to read because: You are also looking for computer science masters...

Transform images to increase the size of the GTSRB dataset My GTSRB Keras CNN is overfitting its training data. I know that because the accuracy on the training data ist above 98% but the accuracy on the test data was...

Using Quiver to visualize a Keras CNN Debugging a neural network is tricky. CNNs are a little bit more “understandable” because we can look which features of an image activates different neurons. The following video is from the CS231n class...

TL;DR You can try my Traffic Sign Recognition Model in your browser: https://pollithy.com/gtsrb/index.html. The accuracy won’t overwhelm you ;) This blog post explains how to build such a website. Using tensorflow.js to run a Keras model in browser Tensorflow.js is...

GTSRB with Keras In my last blog post I described how I built a model for traffic sign recognition with tensorflow. The data set contains 43 different German traffic signs: It is made out of 39.000 training images and the...

Classifying german traffic signs Having reached something like a checkpoint in the cs231n lecture plan I decided to train a CNN to detect German street signs before going on with recurrent neural networks. Tensorflow will be employed and the network...

VoG: Summarizing and understanding large graphs In summer 2017 I attended a data mining seminar. Every participant had to summarize and analyze a different algorithm. Mine was ”VoG - Summarizing and Understanding Large Graphs” by Danai Koutra (Carnegie Mellon University),...

Writing a CNN for cifar-10 Alex Krizhevsky, Vinod Nair and Geoffrey Hinton collected a data set called called CIFAR-10 where the Canadian Institute of Advanced Research was the name donor. It is a subset of 80 million tiny images with...

Pypcd Handling point cloud .pcd files with python. The point cloud library develops its own format called PCD (Point Cloud Data). Details can be found here. If I want to process the 3d data stored in pcd with python I...

Intro to Cython Python is not the fastest. Everything is an object and all of them are implemented as a dictionary. Numpy’s vectorized matrix operations can speed up a lot of things. Especially for-loops should not be written. But in...

Blogging with Jupyter I realized that most of my blog posts behave like a jupyter notebook (markdown, code and repeat). So why not just write a jupyter notebbok and post it as a blog entry. I am going evaluate if...

Experiment: Writing a CNN for tic tac toe Let’s write a CNN which detects the current state of a tic tac toe game. Although it might be super easy to write an opencv2 script which makes this I want to...

How I achieved 83% accuracy in the Kaggle Titanic challenge After trying different things (a custom linear model, different cvs, different random forests, randomized searches, pca) I found the GradientBoostingClassifier with {‘max_depth’: 4, ‘n_estimators’: 200} to perform the best with...

Training a CNN for MNIST on Google Cloud ML In this post I am going to describe how to setup a tensorflow graph to run on the Google Cloud. Sign up for the 12 months 300$ trial account Create a...

Example of a React SPA that logs into a Django backend with JWT A scenario for SPAs is that you might be building it to further use it in a mobile app (maybe react-native). To avoid switching from session_ids stored...

Processing data with scikit’s pipelines, pandas and numpy Fiddling around with my first kernels on Kaggle I found processing data time consuming and it easily looked messy in the jupyter notebook. Especially the fact that I splitted the data set...

Thoughts on how to design a registration figure As mentioned in my last post I was facing the challenge of registrating point clouds captured by multiple depth cameras. The general idea behind registration in this context is to rotate and...

How to create a point cloud from an arbitrary mesh in Blender If you want to registrate multiple tof cameras on one spot, chances are high that you are using Iterative Closest Point (ICP) to maximize the overlap between your...

Calculating backpropagation by hand My last post was about a feed forward net. I did not explain where the weights in the network came from. This entry explains back propagation and contains a calculation for the example from the last...

Let’s build a neural net from scratch without numpy and scikit In my last blog entries I have been discovering neural networks: Calculating a feed forward net by hand Backpropagation Update 2021: Thanks to Tejasvi S. Tomar who spotted that...