The three key training parameters are the number of iterations, initial learning rate, and neural network architecture.
For the neural network architecture selection, you first need to understand how neural networks work:
and then decide on the best option for your problem:
The number of iterations is the number of passes (one pass corresponding to the forward of some images into the neural network and the back propagation of the error in the neural layers) through the neural network.
For each pass, the number of images that will be used is defined by the batch size. It can be found in the page Available architectures.
The relationship between number of iterations and number of epochs is the following: one epoch corresponds to the number of iterations that are necessary to go through all the images in a training set
Number of iterations per epoch = number of training images / batch size
We advise a number of iterations that is equivalent to 10 epochs maximum (10 x number of training images / batch size)
The optimizer is the algorithm technically responsible for training the model, in other words, the rule that is followed to update the parameters of the model in order to improve its performance.
In the literature, we can encounter different algorithms that use mainly the gradient of the loss as a rule to update the parameters.
In the platform we have set by default, for each architecture, a given optimizer with a given learning rate, which were a result of a benchmark campain.
In the case of Classification and Tagging, you have a choice between the following optimizers :
- Momentum (SGD)
- Rectified Adam
- RMS Prop
If you change the optimizer (in the case of Classification and Tagging) the value of the learning rate changes automatically. This value is a recommended value, but you can change it if you want to experiment
For Object Detection Tasks, changing the optimizer will not modify the learning rate. We recommend you to use the default optimizer per architecture, as well as its default learning rate