Albert Faucher, Pierre Jacquier, Michael Segev (Group 70) — COMP551
Basic fully-connected model with two hidden layers. ReLU activation. Very low accuracy. Run SimpleNN_bench.py to replicate results.
Two convolutionnal layers, with batchnorm, max-pooling and ReLU activation, followed by three fully-connected layers. SGD optimizer. Reached 0.77066 on Kaggle (batchnorm made a big difference, dropout not at all). Run ConvNN_bench.py to replicate results.
Attempt at creating a deeper convolutional network. 6 conv layers, got almost 0.96 on Kaggle. Run DeepConvNN_bench.py to replicate results.
Experiment with 7 conv layer model, got Adam to work properly but only gets 92% or so. Run DeeperConvNN_bench.py to replicate results.
Same as DeepConvNN but using bounding-box preprocessing. Three methods were proposed, with CROP_TIGHT being the default and best-performing one. Run DigitDeepConvNN_bench.py to replicate results.
digitfind_test1.py and digitfind_test2 are useful to test out the output of the preprocessing, and they respectively plot images of the training set and export preprocessed images as png files for debugging purposes.
timeitnumpytorchrandommatplotlibosmathpandaspicklecv2