def embedding_separation_loss(y_embeddings, name='emb_sep_L'):
"""
Compute Embedding separation Loss on embedding vectors.
y_embeddings: shape (num_clusters, latent_dim) - for each cluster i, y_embeddings[i] is the
corresponding embedding of cluster in latent dimension.
return: A Embedding separation loss (how separate are the clusters in the latent space). Only L1
loss in latent space considered so far.
"""
embedding_column = tf.expand_dims(y_embeddings,
axis=1) # shape K, 1, latent_dim
embedding_row = tf.expand_dims(y_embeddings,
axis=0) # shape 1, K, latent_dim
# Compute L1 distance
pairwise_loss = tf.reduce_sum(squared_difference(embedding_column,
embedding_row),
axis=-1) # shape K, K
loss = -tf.reduce_sum(pairwise_loss, axis=None, name=name)
return loss
def update_state(self, y_true, y_pred, sample_weight=None):
y_true = tf.cast(y_true, self._dtype)
y_pred = tf.cast(y_pred, self._dtype)
error = math.squared_difference(y_true, y_pred)
return super(RMSE, self).update_state(error,
sample_weight=sample_weight)
def mse(pred, label):
# Mean Squared Error
loss = math.squared_difference(label, pred)
return math.reduce_sum(loss) / len(pred)
def nrmse(y_true, y_pred):
return tfm.sqrt(
tfm.reduce_mean(tfm.squared_difference(y_true, y_pred)) /
tfm.reduce_mean(y_true**2))