Skywalker

4/5:

been gathering data for 5-8 class svm, 5-10 class svm, and rock paper scissors lizard spock
learned basic UDP client-server messaging today
will elaborate on this by sharing results from real-time RPS interpretation from one machine to the other, so that the other can determine the winner at a given time

4/15:

gather 2 rounds of 5-10 class svm data sets for both cy and leo on wed 4/17
2 graphs - don't average each round this time, put each round in its own graph (each graph has 2 lines)
send results + graph before thursday morning
prepare presentations for friday, share w keshav by thursday night (ppt on monitor)
keep in mind that the presentation is 1 minute long
debug connections between phones and probes DONE

3/1

Leo:

Goals: Skywalker code works on local machine, able to gather data

Methods:

Results:

Discussion:

Cy:

Goals:

get skywalker code set up on local machine

Methods:

meeting w leo mon night 8pm to set up

Results:

Discussion:

2/22 continued

Cy:

Goals:

load image data into svm
format new image data into h5py
downsample h5 files to improve computational speed
crop images to relevant sections

Methods:

split dataGlove.txt into 2 files, training labels and test labels (leo)
each line alternates being placed into training set or test set (same method with images) (leo)
svm and separated dataset are on github.gatech.edu/cheffley6/vipsvm
2/28 - going to attempt jason's method of considering 0.5 * avg(max, min) for each index of the list as the open/closed threshold
h5py function for typecasting
segment video to images + assign labels
manipulate h5py implementations directly

Results:

functional svm with proper parsing, takes ~30 minutes but is able to produce a result as of mon 2/25; currently is interpreting all as thumb down
2/28 svm interprpeting all inputs as final index in value

output of current code from github:

number correctly predicted: 177

total predictions made: 466

accuracy: 0.3798283261802575

[[ 0 29 0 0 0 0 0]

[ 0 177 0 0 0 0 0]

[ 0 59 0 0 0 0 0]

[ 0 34 0 0 0 0 0]

[ 0 56 0 0 0 0 0]

[ 0 49 0 0 0 0 0]

[ 0 62 0 0 0 0 0]]

output of new svm using image_data.h5:

number correctly predicted: 90

total predictions made: 180

accuracy: 0.5

[[ 0 0 0 0 20]

[ 0 0 0 0 24]

[ 0 0 0 0 22]

[ 0 0 0 0 90]]

2/22

Leo

Goals:

Find literature about probe placement (ultrasound, EEG, etc.)
Statics

Methods:

Results:

Discussion:

Goals:

CNN on toy dataset (anything from regression set on https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/)
externally read serial communication

Methods:

https://linux.die.net
- socat
CNN in Keras or Pytorch
Could not install torchvision with windows, attempting installation through ubuntu.
Keras installed without issue, going to use that for the time being. Requires tensorflow for backend.

Results:

running https://github.com/adventuresinML/adventures-in-ml-code/blob/master/keras_cnn.py
functional but not entirely sure what some of the data means. also unsure how to load datasets from anywhere other than the built-in data
not able to find much on serial reading

Discussion:

2/15

Leo

Goals:

Build SVM to classify data set
Find literature about probe placement (ultrasound, EEG, etc.)

Methods:

Results:

Built SVM using sklearn
- Classifies data sets from the libsvm dataset (1st link above)
- Returns accuracy (~80%)
Link to code: SVM

Discussion:

Goals:

pick data set, build svm for classifying
- return accuracy, confusion matrix
todo: build CNN for classifying data sets

Methods:

Results:

SVM built in IDLE using sklearn 's svm.svc() and numpy
can load any training set and test set from csie.ntu.edu.tw website's libsvm formats
results generally 80% accurate between a1a, a2a, a3a, and svmguide1 data sets
code accessible via https://github.com/cheffley6/samplesvm

Discussion:

Google Sites

Report abuse