I am working on object detection. My training set is 5984 and test set is 1496 (KITTI Tracking training dataset).
How big should batch size and number of epochs be when fitting a model in Tensorflow? Is there any sort of rule of thumb to use based on data input size? Should I split it to train/val/test or stick with my current split train/val?
Try as high a batch size as your memory allow - this will be evaluated more efficiently on GPUs and you can usually cope with a higher learning rate making learning faster. Try a high number of epochs and babysit the learning process until you see convergence and stop before you overfit. Have a look here: http://cs231n.github.io/neural-networks-3/ to get ideas on how to babysit the learning process.
Split your dataset into train/val/test - evaluate your model on the validation set after each epoch and use the test set for estimating your generalization error.
