def train_epoch(self, network):
self._setup(network)
losses = []
X_batch = batch_iterator(network.X, network.batch_size)
y_batch = batch_iterator(network.y, network.batch_size)
for X, y in tqdm(zip(X_batch, y_batch), 'Epoch progress'):
loss = np.mean(network.update(X, y))
self.update(network)
losses.append(loss)
epoch_loss = np.mean(losses)
return epoch_loss
def _train(self):
for i in range(self.max_epochs):
error = 0
for batch in batch_iterator(self.X, batch_size=self.batch_size):
positive_hidden = sigmoid(np.dot(batch, self.W) + self.bias_h)
hidden_states = self._sample(positive_hidden)
positive_associations = np.dot(batch.T, positive_hidden)
negative_visible = sigmoid(
np.dot(hidden_states, self.W.T) + self.bias_v)
negative_hidden = sigmoid(
np.dot(negative_visible, self.W) + self.bias_h)
negative_associations = np.dot(negative_visible.T,
negative_hidden)
lr = self.lr / float(batch.shape[0])
self.W += lr * (
(positive_associations - negative_associations) /
float(self.batch_size))
self.bias_h += lr * (negative_hidden.sum(axis=0) -
negative_associations.sum(axis=0))
self.bias_v += lr * (np.asarray(batch.sum(axis=0)).squeeze() -
negative_visible.sum(axis=0))
error += np.sum((batch - negative_visible)**2)
self.errors.append(error)
logging.info('Iteration %s, error %s' % (i, error))
logging.debug('Weights: %s' % self.W)
logging.debug('Hidden bias: %s' % self.bias_h)
logging.debug('Visible bias: %s' % self.bias_v)
def _train(self):
"""Use CD-1 training procedure, basically an exact inference for `positive_associations`,
followed by a "non burn-in" block Gibbs Sampling for the `negative_associations`."""
for i in range(self.max_epochs):
error = 0
for batch in batch_iterator(self.X, batch_size=self.batch_size):
positive_hidden = sigmoid(np.dot(batch, self.W) + self.bias_h)
hidden_states = self._sample(positive_hidden) # sample hidden state h1
positive_associations = np.dot(batch.T, positive_hidden)
negative_visible = sigmoid(np.dot(hidden_states, self.W.T) + self.bias_v)
negative_visible = self._sample(negative_visible) # use the samped hidden state h1 to sample v1
negative_hidden = sigmoid(np.dot(negative_visible, self.W) + self.bias_h)
negative_associations = np.dot(negative_visible.T, negative_hidden)
lr = self.lr / float(batch.shape[0])
self.W += lr * ((positive_associations - negative_associations) / float(self.batch_size))
self.bias_h += lr * (negative_hidden.sum(axis=0) - negative_associations.sum(axis=0))
self.bias_v += lr * (np.asarray(batch.sum(axis=0)).squeeze() - negative_visible.sum(axis=0))
error += np.sum((batch - negative_visible) ** 2)
self.errors.append(error)
logging.info("Iteration %s, error %s" % (i, error))
logging.debug("Weights: %s" % self.W)
logging.debug("Hidden bias: %s" % self.bias_h)
logging.debug("Visible bias: %s" % self.bias_v)
def _predict(self, X=None):
if not self._initialized:
self._setup_layers(X.shape)
y = []
X_batch = batch_iterator(X, self.batch_size)
for Xb in X_batch:
y.append(self.fprop(Xb))
return np.concatenate(y)
def train_epoch(self, network):
losses = []
# Create batch iterator
X_batch = batch_iterator(network.X, network.batch_size)
y_batch = batch_iterator(network.y, network.batch_size)
batch = zip(X_batch, y_batch)
if network.verbose:
batch = tqdm(batch, total=int(np.ceil(network.n_samples / network.batch_size)))
for X, y in batch:
loss = np.mean(network.update(X, y))
self.update(network)
losses.append(loss)
epoch_loss = np.mean(losses)
return epoch_loss
def _predict(self, X=None):
if not self._initialized:
self._setup_layers(X.shape)
y = []
X_batch = batch_iterator(X, self.batch_size)
for Xb in X_batch:
y.append(self.fprop(Xb))
return np.concatenate(y)
def train_epoch(self, network):
losses = []
# Create batch iterator
X_batch = batch_iterator(network.X, network.batch_size)
y_batch = batch_iterator(network.y, network.batch_size)
batch = zip(X_batch, y_batch)
if network.verbose:
batch = tqdm(batch)
for X, y in batch:
loss = np.mean(network.update(X, y))
self.update(network)
losses.append(loss)
epoch_loss = np.mean(losses)
return epoch_loss
def _predict(self, X=None):
y = []
X_batch = batch_iterator(X, self.batch_size)
for Xb in X_batch:
y.append(self.fprop(Xb))
return np.concatenate(y)
