This week we focused on finalizing our design review slides and document. It allowed us to finalize important details in our architecture. We plan to have the basic framework for our networks by the time we come back from Spring Break in order to start training. Then we can start working on the application portion of the project. We’ve received our Raspberry Pi 4 and Camera have been delivered. As for the mic, we’re in Read more…
This week I focused on finishing the design review presentation and report. I worked on the text emotion recognition portion of the design review and helped design our slide deck. I also incorporated the feedback from the presentations and solidified my architecture for the text sentiment analysis. No matter which model and word vectorize I use, the general layout of the text analysis module is the same. Our team still needs to hash out how we Read more…
I started to work on the text-based emotional analysis. For the first iteration, I’m using Term Frequency-Inverse Document Frequency. This parameter gives the relative importance of a term in the data and is a measure of how frequently and rarely it appears in the text. It’s how we turn words into numbers. I used sklearn’sTfidfVectorizer to do this. I tried 3 models: a Naive Bayes’ Classified, Linear SVM and logistic regression. All 3 produced accuracies around 50% Read more…
Updates: Received AWS credits, ordered parts (RPi and camera). There are no changes to the design or schedule. The team is working on individual components (writing algorithms, writing app, etc.) There are no significant risks we can think of at this time; however we did discuss the possibility of storing the neural nets on the rPi instead of the cloud server, and some risks that may come with that is the rPi not being powerful Read more…
This week I worked on the web application for our project. I set up the application using Django, and set up the database using mySQL. The two objects we have right now are emotion and entry; “emotion” will be used to categorize the emotions of the journal entries as well as send the evaluated emotion to the raspberry pi. “entry” will be the journal entries that the raspberry pi sends to the web app. Database Read more…
Completed tasks: Verified performance of VGG19 Working on PCA of features from 4th pooling layer of VGG19 100000 features to 200 principal components Still working on writing SVM To-do: Finish PCA and SVM algorithm Download CK+ database and begin preliminary training Design presentation and report
Summary: Acquired datasets to train Worked on Design Review Presentation Decided on training algorithm Changes: We decided to use pre-trained convolutional neural networks to increase accuracy and reduce the time required in the training process. We will still be training the neural networks ourselves using the SVM model. Our schedule is still on track for this week. Next week, we hope to set up the Cloud to prepare to start training neural networks using the Read more…
This week I worked on the block diagrams for the Design Review presentation including the interactions between the servers and E-MO, the IoT devices and E-MO, and the url routing diagrams between the trained neural nets, the web application, and E-MO. (The main block digram for the server and E-MO can be found on the “About” page.) Also reached out to Professor Nace about receiving AWS credits. Also researched and learned more about training neural Read more…
Completed tasks: Installed Python3 and OpenCV on local machine. Installed Tensorflow Wrote face detection program Planned out algorithm for training facial emotion recognition Use pre-trained CNN (VGG19) to extract facial features VGG19 available in tensorflow Feed feature vector into SVM and classify into 6 emotions To-do: Verify face detection – OpenCV unable to open camera through WSL, need to find work-around Test facial feature extraction with VGG19 Write SVM and begin preliminary training Design review Read more…
I’ve finished setting up Python/OpenCV on my local machine and polished our block diagrams. I’ve also acquired the necessary data from the AffectNet and Cohn-Kahnade databases. The AffectNet database provides around 500,000 (~120gb of data) faces categorized into 11 emotions: The Cohn-Kahnade database is significantly smaller and also dated. Thus, Professor Marios advised us to stick to the AffectNet database for now as it has a sufficient amount of data for our training network. The Read more…