| 30 | |
| 31 | |
| 32 | class ANN(object): |
| 33 | def __init__(self, hidden_layer_sizes): |
| 34 | self.hidden_layer_sizes = hidden_layer_sizes |
| 35 | |
| 36 | def fit(self, X, Y, learning_rate=0.01, mu=0.99, epochs=30, batch_sz=100): |
| 37 | # cast to float32 |
| 38 | learning_rate = np.float32(learning_rate) |
| 39 | mu = np.float32(mu) |
| 40 | |
| 41 | N, D = X.shape |
| 42 | K = len(set(Y)) |
| 43 | |
| 44 | self.hidden_layers = [] |
| 45 | mi = D |
| 46 | for mo in self.hidden_layer_sizes: |
| 47 | h = HiddenLayer(mi, mo) |
| 48 | self.hidden_layers.append(h) |
| 49 | mi = mo |
| 50 | |
| 51 | # initialize logistic regression layer |
| 52 | W = init_weights((mo, K)) |
| 53 | b = np.zeros(K, dtype=np.float32) |
| 54 | self.W = theano.shared(W) |
| 55 | self.b = theano.shared(b) |
| 56 | |
| 57 | self.params = [self.W, self.b] |
| 58 | self.allWs = [] |
| 59 | for h in self.hidden_layers: |
| 60 | self.params += h.params |
| 61 | self.allWs.append(h.W) |
| 62 | self.allWs.append(self.W) |
| 63 | |
| 64 | X_in = T.matrix('X_in') |
| 65 | targets = T.ivector('Targets') |
| 66 | pY = self.forward(X_in) |
| 67 | |
| 68 | cost = -T.mean( T.log(pY[T.arange(pY.shape[0]), targets]) ) |
| 69 | prediction = self.predict(X_in) |
| 70 | |
| 71 | updates = momentum_updates(cost, self.params, mu, learning_rate) |
| 72 | train_op = theano.function( |
| 73 | inputs=[X_in, targets], |
| 74 | outputs=[cost, prediction], |
| 75 | updates=updates, |
| 76 | ) |
| 77 | |
| 78 | n_batches = N // batch_sz |
| 79 | costs = [] |
| 80 | lastWs = [W.get_value() for W in self.allWs] |
| 81 | W_changes = [] |
| 82 | print("supervised training...") |
| 83 | for i in range(epochs): |
| 84 | print("epoch:", i) |
| 85 | X, Y = shuffle(X, Y) |
| 86 | for j in range(n_batches): |
| 87 | Xbatch = X[j*batch_sz:(j*batch_sz + batch_sz)] |
| 88 | Ybatch = Y[j*batch_sz:(j*batch_sz + batch_sz)] |
| 89 | c, p = train_op(Xbatch, Ybatch) |