OffTopic: Training sets of modified images

TensorFlow, Keras

It was an idea to check training process on changed images
Original pixels were masked by nearest pixel (y+1, x+1) - XOR operation
One pix set was XORed 100%, second - on half, by 50% RGB value of near pixel


As shown below fully masked image set has some better result on ~8th epoch

My PC has not GPU, so impossible to make more training with different parameters.
Maybe this idea with masking has some sense

Interesting! Thanks for sharing this idea and your results. So this would be an interesting example of “data augmentation”: figuring out ways to generate more training data based on your existing data.

Just to make sure I understand the meaning of your data: the accuracy figures you are showing are on whatever the training data is in each particular case, right? Or is that on a “test” or “cross validation” set? If it’s the former, it seems like also showing the prediction accuracy on your test dataset would be important to really understand the full effect here.

In the “big picture”, it looks like the earlier advantages of the fully masked data seem to be in the noise by the time you get out to 25 epochs. Meaning that all three methods give equivalently good accuracy (~98%) by the time you run that many epochs.

Sorry for my English.
It was 3 separate trainings - first one with original JPEG files.
Second - with the same files but processed by direct XOR
Third - XOR by 50% value of nearby pixel

Result of each test - Loss, Accuracy, Validation Loss and Validation Accuracy for every epoch
data file - XLS

On previous chart - Loss & Accuracy

Here - total mess of validation data

One set training on my PC - about 6 hours of calculations.
So, whole 24 hours for 3 sets :slight_smile:

I have saved checkpoints for every epoch - don’t know yet how to load that to know prediction accuracy

There are more than 20 000 pix of dogs and cats in original set.
For me it is impossible to train so big set.

“Masked” images are more contours than are regular pictures.
Maybe for such contoured pictures some another model construction could work much better.
Or other hyperparameters.
Model that works on contour lines.

“Masking” could be kinda first layer prior layers of “conventional” model