Cat v Dog classifer and thoughts on the litter classifer.
Date: 2021-01-05
Author: Andrew
Trained the first cat v dog recogniser. 1 block VGG. First epoch of the 50 epochs took ages for some reason. As this was the first time running Keras/Tensorflow on this machine that might have had something to do with it. After the first epoch they took about a minute. Hmmm... was expecting faster. 16GB in RAM on motherboard and 8GB of VRAM on graphics card is going to work for a while but may need to offload to a S3/EC2 instance in future.
- 1 VGG - 72.949 accuracy on test set
- 2 VGG - 75.393 accuracy on test set
- 3 VGG - 79.545 accuracy on test set
- 3 VGG + Dropout - 81.548 accuracy on test set
- 3 VGG + Dropout + augment - 79.883 accuracy on test set
- 3 VGG + Dropout + augment (100 epochs) - 83.something accuracy on test set
Not world-beating standard but for very little investment a pretty good outcome. Now the questions is if I can achieve similar results with the litter data I've assembled to date. I can see I might face a couple of problems.
Firstly, I don't have a huge amount of images having only a few hundred classified so far. Secondly, the litter does vary quite a lot in shape and colour so it will take longer for the classifer to start identifying litter consistently. The frame size is 1280x720 - but I'll need to resize them down to something like 224x224 as that seems to be a common size used. The challenge there is the resize might actual lose small pieces of litter which would confuse the classifier. Finally, the focus on some of the pictures is not too good so I imagine that will also cause problems as there is no clearly defined edges to the litter to detect.
Before going to bed checked that all services on the BeachBot were enabled and found that I had disabled the camera service. Fixed that up and 100% battery so all good for tomorrow's field test.