devlyn37 / deepfire Goto Github PK

For centralized machine learning

To improve our training data, we could introduce superpixel segmentation into our pipeline, which would involve grouping together similar pixels and separating them into multiple images. This should probably be done after we have a better idea of the kind of accuracy we can expect from our models to see if this results in an improvement.

Currently our datasets have a lot of photos that are either not aerial from UAVs or photos of fire that are not wildfires. We need a way to filter out these photos.

From my understanding there are many different ways to 'train' models in terms of how the training dataset is used to ensure that the model can generalize better to an independent dataset and avoid overfitting, etc.

@mjfortier probably knows more about this than any of us given that he has taken a machine learning theory course.

See https://en.wikipedia.org/wiki/Cross-validation_(statistics). The one methodology I have come across in datamining was k-fold cross validation.

It would be good if we could define how we plan to use our data in terms of training samples/validation samples, etc.

If we are going to be comparing various models it is a good idea to identify some measures that we can use to compare them.

For binary classifers see (https://en.wikipedia.org/wiki/Evaluation_of_binary_classifiers). There seems to be 20+ metrics listed on that wikipage.

I imagine most of them are configurable to be recorded through tensorflow/pytorch.

Would be good to define this.

Based on: Dual Deep Learning Model for Image Based Smoke Detection

In order to create an effective model with three classifications, we need more photos that contain only smoke.

• Setup XServer from remote image viewing
• Move datasets on remote server

As discussed in meetings, a large part of our report is going to be comparisons between models we use for transfer learning. In addition, the professor suggested that we experiment with three-way classification (smoke, fire, forest) and ensemble models. The codebase should make this easier by enabling different base models to be used with different training methods and model types without the need to copy and paste a bunch of code each time.