-
Notifications
You must be signed in to change notification settings - Fork 28
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
Showing
2 changed files
with
85 additions
and
0 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 1,85 @@ | ||
| import numpy as np | ||
| import tensorflow as tf | ||
| from a_nice_mc.objectives import Energy | ||
| from a_nice_mc.utils.evaluation import effective_sample_size, acceptance_rate | ||
| from a_nice_mc.utils.logger import save_ess, create_logger | ||
|
|
||
| logger = create_logger(__name__) | ||
|
|
||
|
|
||
| class NN(Energy): | ||
| def __init__(self, data, labels, arch, act=tf.nn.tanh, prec=1.0): | ||
| """ | ||
| Bayesian Neural Network Model (assumes factored Normal prior) | ||
| :param data: data for Regression task | ||
| :param labels: label for Regression task | ||
| :param scale: std of the Normal prior | ||
| :param arch: list of layer widths for feed-forward network | ||
| """ | ||
| super(NN, self).__init__() | ||
| self.arch = arch | ||
| self.theta_dim = np.sum([arch[i] * arch[i 1] for i in range(len(arch) - 1)]) | ||
| self.act = act | ||
| self.x_dim = data.shape[1] | ||
| self.y_dim = labels.shape[1] | ||
| self.prec_prior = prec | ||
|
|
||
| self.data = tf.constant(data, tf.float32) | ||
| self.labels = tf.constant(labels, tf.float32) | ||
|
|
||
| def _unflatten(self, theta): | ||
| """theta is assumed to have shape (num_chains, target_dim)""" | ||
| m = tf.shape(theta)[0] # num chains | ||
| weights = [] | ||
| start = 0 | ||
| for i in range(len(self.arch) - 1): | ||
| size = self.arch[i] * self.arch[i 1] | ||
| w = tf.reshape(theta[:, start:start size], | ||
| (m, self.arch[i], self.arch[i 1])) | ||
| weights.append(w) | ||
| start = size | ||
| return weights | ||
|
|
||
| def energy_fn(self, theta, x, y): | ||
| """ theta has shape (num_chains, target_dim) | ||
| We subsume the biases into the weight matrices by appending ones to | ||
| the hidden state.""" | ||
| h = tf.expand_dims(x, 0) | ||
| h = tf.concat([h, tf.ones((1, h.shape[1], 1))], axis=2) | ||
| h = tf.tile(h, [tf.shape(theta)[0], 1, 1]) | ||
| weights = self._unflatten(theta) | ||
| for W in weights[:-1]: | ||
| h = self.act(h @ W) | ||
| mean = h @ weights[-1] | ||
| mahalob = 0.5 * tf.reduce_sum((y - mean) ** 2, axis=2) | ||
| prior = 0.5 * tf.reduce_sum(theta ** 2, axis=1, keepdims=True) | ||
|
|
||
| return tf.reduce_sum(mahalob self.prec_prior * prior, axis=1) | ||
|
|
||
| def __call__(self, v): | ||
| return self.energy_fn(v, self.data, self.labels) | ||
|
|
||
| def evaluate(self, zv, path=None): | ||
| z, v = zv | ||
| z_ = np.reshape(z, [-1, z.shape[-1]]) | ||
| m = np.mean(z_, axis=0, dtype=np.float64) | ||
| v = np.std(z_, axis=0, dtype=np.float64) | ||
| print('mean: {}'.format(m)) | ||
| print('std: {}'.format(v)) | ||
| logger.info('Acceptance rate %.4f' % (acceptance_rate(z))) | ||
| ess = effective_sample_size( | ||
| z, | ||
| self.mean(), self.std() * self.std(), | ||
| logger=logger | ||
| ) | ||
| if path: | ||
| save_ess(ess, path) | ||
| np.save(path '/trajectory.npy', z) | ||
|
|
||
| @staticmethod | ||
| def mean(): | ||
| return None | ||
|
|
||
| @staticmethod | ||
| def std(): | ||
| return None |
Empty file.