81114 Since I last posted, I’ve been revising and rewriting my convolutional neural network incorporating transfer learning from one of the five algorithms mentioned below, to make dog breed predictions from random photos selected from the imagenet database as part of the Udacity Deep Learning Nanodegree ‘dog-breed-project’ in a jupyter notebook using the python programming language.
I’ve really enjoyed this project a great deal as it involves computer vision algorithms used to make image content predictions. Although I am not quite finished with the project at this moment, I suddenly realized this morning that there is a possibility that I might be able to join the five transfer learning pre-trained algorithms into a single classifier. As I leave this post, I will experiment with this ensemble concept of creating an algorithm with higher predictive accuracy today and then move on to the next Udacity Deep Learning Nanodegree Project called ‘Generate TV Scripts’ using a recurrent neural network. Wish me good luck as I continue to study this interesting technology.
81110 Continued working on the second project in the Udacity Deep Learning Nanodegree this afternoon. I’ve written and tested the following image processing models for sequential convolutional neural network classification of human face and dog images attempting to successfully classify 133 different dog breeds using one of the following pre-trained models; ResNet50, VGG16, VGG19, InceptionV3 and Xception. So fare have been making relatively good progress.
81109 Today I am approaching the end of this project and will be attempting to close it out and submit it to Udacity for grading by next Wednesday at the latest. Then on to the third project in the Udacity Deep Learning Nanodegree called “Generate TV scripts”. The goal of this project is to train a recurrent neural network on a set of existing TV scripts and then generate an original TV script using the recurrent neural network that was created for this project.
81106 Today more progress was made on this project in fact I was able to hit a predictions accuracy of about 76% when the minimum target goal was 60%. This was accomplished using a VGG19 Sequential model using a ‘relu’, ‘sigmoid, and finally followed by a ‘softmax’ activation function.
81105-2 Later today I was able to move beyond the sticking point that I bumped into yesterday and the day before so I should be able to move on to parts 5 & 6 shortly.
81105 For the past two days I’d been making great progress on the second project in this ND. Unfortunately, yesterday I came to a virtual standstill on the 4th and 5th sections of the jupyter notebook. One thing I found that is quite frustrating, is the linearity involved in the different sections of the project. Once you get into the fourth and fifth sections running these code blocks are dependent on information in the previous three sections and the third section involves a block that runs very slowly so your progress moving forward in code blocks 4 & 5 is constrained by running the third slow block of code repeatedly every time you want to try out something new in sections 4 and 5. Of course I could be doing something wrong here and haven’t realized how to deal with it properly yet.
81103 Just a quick update on my DLND progress. Yesterday I was working on the second project in the course. This project is about building and training a convolutional neural network on data in the form of human face images and dog images in order to get the neural network model to predict dog breeds. The good news was that on part three of the project I am building a CNN from scratch to detect dog breeds and yesterday I had made some significant progress along those lines. So the rest of today I’ll be picking up where I left off yesterday and attempting to complete this part of the project.
81101 Today is a national holiday here in Hungary, it is ‘All Saints Day’. Its a day when most people visit a cemetery where one or more family member’s have been laid to rest for the purpose of placing candles at the graves sites to burn throughout the night and possible longer. In my case, I’ll be working on the next project in the Deep Learning Nanodegree by Udacity called the Dog Breed Classifier project using a convolutional neural network model architecture. Holidays for me, are typically when I get a lot of work done on a project.
81031 Successfully completed the first project in the Udacity Deep Learning Nanodegree yesterday. Starting on the second project using a convolutional neural network called dog breed classifier.
81030 Submitted the first project: Predicting Bike Ridership with the first neural network project in the Udacity Nanodegree today.
81028 This recurring post is on the topic of Deep Learning in Artificial Intelligence.
After some weeks away from this course in order that I might put more emphasis on another course, I’ve returned to it now so that I might continue on and possibly complete this Deep Learning Nanodegree.
The course begins with an introduction module followed by five additional modules or sections which are in order; neural networks, convolutional networks, recurrent networks, generative adversarial networks and finally deep reinforcement learning.
Each section has lecture materials in the form of videos with supporting slides and exercises that promote better understanding of the concepts as well as project fully covering the topics in each section so that you will not only have a good mental understanding but also gain experience in the practical application of the concepts.
My experience has been that it is really in the project phase of each section that the majority of true learning occurs. At the moment I am on the Neural Networks section. Yesterday I spent a good amount of time on the project in this module called first neural network and expect to be completing it shortly.