batchnorm with support for not using scale and shift parameters as well as inference values (u and s) and partial batchnorm (via a) will detect and use convolutional or fully connected version
(X, g=None, b=None, u=None, s=None, a=1., e=1e-8)
| 50 | return T.concatenate([x, y*T.ones((x.shape[0], y.shape[1], x.shape[2], x.shape[3]))], axis=1) |
| 51 | |
| 52 | def batchnorm(X, g=None, b=None, u=None, s=None, a=1., e=1e-8): |
| 53 | """ |
| 54 | batchnorm with support for not using scale and shift parameters |
| 55 | as well as inference values (u and s) and partial batchnorm (via a) |
| 56 | will detect and use convolutional or fully connected version |
| 57 | """ |
| 58 | if X.ndim == 4: |
| 59 | if u is not None and s is not None: |
| 60 | b_u = u.dimshuffle('x', 0, 'x', 'x') |
| 61 | b_s = s.dimshuffle('x', 0, 'x', 'x') |
| 62 | else: |
| 63 | b_u = T.mean(X, axis=[0, 2, 3]).dimshuffle('x', 0, 'x', 'x') |
| 64 | b_s = T.mean(T.sqr(X - b_u), axis=[0, 2, 3]).dimshuffle('x', 0, 'x', 'x') |
| 65 | if a != 1: |
| 66 | b_u = (1. - a)*0. + a*b_u |
| 67 | b_s = (1. - a)*1. + a*b_s |
| 68 | X = (X - b_u) / T.sqrt(b_s + e) |
| 69 | if g is not None and b is not None: |
| 70 | X = X*g.dimshuffle('x', 0, 'x', 'x') + b.dimshuffle('x', 0, 'x', 'x') |
| 71 | elif X.ndim == 2: |
| 72 | if u is None and s is None: |
| 73 | u = T.mean(X, axis=0) |
| 74 | s = T.mean(T.sqr(X - u), axis=0) |
| 75 | if a != 1: |
| 76 | u = (1. - a)*0. + a*u |
| 77 | s = (1. - a)*1. + a*s |
| 78 | X = (X - u) / T.sqrt(s + e) |
| 79 | if g is not None and b is not None: |
| 80 | X = X*g + b |
| 81 | else: |
| 82 | raise NotImplementedError |
| 83 | return X |
| 84 | |
| 85 | def deconv(X, w, subsample=(1, 1), border_mode=(0, 0), conv_mode='conv'): |
| 86 | """ |