| 35 | |
| 36 | |
| 37 | class ANN(object): |
| 38 | def __init__(self, hidden_layer_sizes): |
| 39 | self.hidden_layer_sizes = hidden_layer_sizes |
| 40 | |
| 41 | def fit(self, X, Y, activation=T.tanh, learning_rate=1e-3, mu=0.5, reg=0, epochs=5000, batch_sz=None, print_period=100, show_fig=True): |
| 42 | X = X.astype(np.float32) |
| 43 | Y = Y.astype(np.float32) |
| 44 | |
| 45 | # initialize hidden layers |
| 46 | N, D = X.shape |
| 47 | self.hidden_layers = [] |
| 48 | M1 = D |
| 49 | count = 0 |
| 50 | for M2 in self.hidden_layer_sizes: |
| 51 | h = HiddenLayer(M1, M2, activation, count) |
| 52 | self.hidden_layers.append(h) |
| 53 | M1 = M2 |
| 54 | count += 1 |
| 55 | W = np.random.randn(M1) / np.sqrt(M1) |
| 56 | b = 0.0 |
| 57 | self.W = theano.shared(W, 'W_last') |
| 58 | self.b = theano.shared(b, 'b_last') |
| 59 | |
| 60 | if batch_sz is None: |
| 61 | batch_sz = N |
| 62 | |
| 63 | # collect params for later use |
| 64 | self.params = [self.W, self.b] |
| 65 | for h in self.hidden_layers: |
| 66 | self.params += h.params |
| 67 | |
| 68 | # for momentum |
| 69 | dparams = [theano.shared(np.zeros(p.get_value().shape)) for p in self.params] |
| 70 | |
| 71 | # set up theano functions and variables |
| 72 | thX = T.matrix('X') |
| 73 | thY = T.vector('Y') |
| 74 | Yhat = self.forward(thX) |
| 75 | |
| 76 | rcost = reg*T.mean([(p*p).sum() for p in self.params]) |
| 77 | cost = T.mean((thY - Yhat).dot(thY - Yhat)) + rcost |
| 78 | prediction = self.forward(thX) |
| 79 | grads = T.grad(cost, self.params) |
| 80 | |
| 81 | # momentum only |
| 82 | updates = [ |
| 83 | (p, p + mu*dp - learning_rate*g) for p, dp, g in zip(self.params, dparams, grads) |
| 84 | ] + [ |
| 85 | (dp, mu*dp - learning_rate*g) for dp, g in zip(dparams, grads) |
| 86 | ] |
| 87 | |
| 88 | train_op = theano.function( |
| 89 | inputs=[thX, thY], |
| 90 | outputs=[cost, prediction], |
| 91 | updates=updates, |
| 92 | ) |
| 93 | |
| 94 | self.predict_op = theano.function( |