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