train.py

   1 #!/usr/bin/python3
   2
   3 import mnist
   4 import nnet
   5
   6 train_images, train_labels, rows, cols = mnist.load('train-images-idx3-ubyte', 'train-labels-idx1-ubyte')
   7 test_images, test_labels, rows2, cols2 = mnist.load('t10k-images-idx3-ubyte', 't10k-labels-idx1-ubyte')
   8 assert rows == rows2
   9 assert cols == cols2
  10 num_train = train_images.shape[1]
  11 nnet_batch = 10000
  12
  13 # neural network structure: two hidden layers, one output layer
  14 #                   (input)--> [Linear->Sigmoid] -> [Linear->Sigmoid] -->(output)
  15 # handle 10,000 vectors at a time
  16 Z1 = nnet.LinearLayer(input_shape=(rows * cols, nnet_batch), n_out=20)
  17 A1 = nnet.SigmoidLayer(Z1.Z.shape)
  18 Z2 = nnet.LinearLayer(input_shape=A1.A.shape, n_out=16)
  19 A2 = nnet.SigmoidLayer(Z2.Z.shape)
  20 ZO = nnet.LinearLayer(input_shape=A2.A.shape, n_out=10)
  21 AO = nnet.SigmoidLayer(ZO.Z.shape)
  22 net = (Z1, A1, Z2, A2, ZO, AO)
  23
  24 res = nnet.forward(net, train_images[:, 0:10000])
  25 print(f'output vector of first image: {res[:, 0]}')
  26 digit, conf = nnet.classify(res[:, 0])
  27 print(f'classification of first image: {digit} with confidence {conf}; real label {test_labels[0]}')
  28 print(f'correctly recognized images after initialization: {nnet.accuracy(net, test_images, test_labels)}%')
  29
  30 train_y = nnet.label_vectors(train_labels, 10)
  31 for i in range(100):
  32     for batch in range(0, num_train, nnet_batch):
  33         cost = nnet.train(net, train_images[:, batch:(batch + nnet_batch)], train_y[:, batch:(batch + nnet_batch)])
  34     print(f'cost after training round {i}: {cost}')
  35 print(f'correctly recognized images after training: {nnet.accuracy(net, test_images, test_labels)}%')