ConvRF

Yu-Chung Peng and Madi Kusmanov

Yu-Chungs temp ConvRF repo: https://github.com/ypeng22/ConvRFClassifier

Notes on another conv-rf-file: https://github.com/tpsatish95/deep-conv-rf/blob/master/experiments/random_forest/deep_conv_rf.py Chops the images the same way Takes a "type" variable which can take values unshared, shared, rerf_shared if unshared: convolution is just kernel_forest does predict_proba, there are out_height * out_length different RF's if shared: convolution is kernel_forest does predict proba, except this time there is only one huge forest. Think of shared like one kernel across all segment chunks, and unshared as a kernel for each segment chunk if rerf_shared: convolution is like above except the huge forest is a Rerf (I think it is a sporf?)

Old repo experiment notes: Under deep-conv-rf/experiments/results For Mnist: No convRF setup did substantially better than naive RF, most of the time RF did better For Cifar: 1vs9 showed best improvements, others marginal to none For SVHN: Same as Mnist

Under deep-conv-rf/notebooks No substantial evidence of ConvRF being much better than naive, maybe 65% accuracy vs 67%, but no difference more than 10% (so 60% vs 66%)