As the man himself said "What I cannot create, I do not understand.".
So I try to built Neural Networks and training algorithms using only Numpy.
Train a MNIST model with [output_activation]_[Loss].
-
Sigmoid - MSE
$ python test/dense.py sigmoid_mse train
-
Sigmoid - Cross-entropy
$ python test/dense.py sigmoid_crossentropy train
-
Softmax - Log-likelihood
$ python test/dense.py softmax_loglikelihood train