Merge pull request #44 from Mayureshd-18/mayud

Added Ranking Losses

ML/Algorithms/Losses/RankingLosses/PairWiseRankingLoss/PairWiseRankingLoss.py

@@ -0,0 +1,27 @@
+import tensorflow as tf
+from typing import Tuple
+
+def pairwise_ranking_loss(y_true: tf.Tensor, y_pred: tf.Tensor, margin: float = 1.0) -> tf.Tensor:
+ """
+ Computes the pairwise ranking loss for a batch of pairs.
+
+ Args:
+ y_true: Tensor of true labels (0 for negative pairs, 1 for positive pairs).
+ y_pred: Tensor of predicted similarities/distances, expected to be a tensor of shape (batch_size, 2, embedding_dim) where 
+ y_pred[:, 0] is the anchor and y_pred[:, 1] is the positive/negative.
+ margin: Margin parameter for the pairwise ranking loss.
+
+ Returns:
+ loss: Computed pairwise ranking loss as a scalar tensor.
+ """
+ anchor, positive_or_negative = y_pred[:, 0], y_pred[:, 1]
+
+ distances = tf.reduce_sum(tf.square(anchor - positive_or_negative), axis=-1)
+ positive_loss = y_true * distances
+ negative_loss = (1 - y_true) * tf.maximum(margin - distances, 0.0)
+
+ loss = positive_loss + negative_loss
+ return tf.reduce_mean(loss)
+
+# Example usage:
+# model.compile(optimizer='adam', loss=pairwise_ranking_loss)

ML/Algorithms/Losses/RankingLosses/TripletLoss/TripletLoss.py

@@ -0,0 +1,26 @@
+import tensorflow as tf
+from typing import Tuple
+
+def triplet_loss_func(y_true: tf.Tensor, y_pred: tf.Tensor, alpha: float = 0.3) -> tf.Tensor:
+ """
+ Computes the triplet loss for a batch of triplets.
+
+ Args:
+ y_true: True values of classification (unused in this implementation, typically required for compatibility with Keras).
+ y_pred: Predicted values, expected to be a tensor of shape (batch_size, 3, embedding_dim) where 
+ y_pred[:, 0] is the anchor, y_pred[:, 1] is the positive, and y_pred[:, 2] is the negative.
+ alpha: Margin parameter for the triplet loss.
+
+ Returns:
+ loss: Computed triplet loss as a scalar tensor.
+ """
+ anchor, positive, negative = y_pred[:, 0], y_pred[:, 1], y_pred[:, 2]
+
+ positive_dist = tf.reduce_sum(tf.square(anchor - positive), axis=-1)
+ negative_dist = tf.reduce_sum(tf.square(anchor - negative), axis=-1)
+
+ loss = tf.maximum(positive_dist - negative_dist + alpha, 0.0)
+ return tf.reduce_mean(loss)
+
+# Example usage:
+# model.compile(optimizer='adam', loss=triplet_loss_func)

ML/README.md

@@ -21,7 +21,7 @@
 |-------|-----------|-------|-----------|-------|-----------|
 | 1 | [Mean Squared Error](./Algorithms/Losses/MeanSquaredError) | 2 | [R2 Squared](./Algorithms/Losses/R2Squared) | 3 | [Cross Entropy Loss](./Algorithms/Losses/CrossEntropyLoss) |
 | 4 | [Hinge Loss](./Algorithms/Losses/HingeLoss) | 5 | [Kullback Leibler (KL) Divergence Loss](./Algorithms/Losses/KullbackLeiblerDivergenceLoss) | 6 | [Mean Absolute Error](./Algorithms/Losses/MeanAbsoluteError) |
-| 7 | [Root Mean Squared Error](./Algorithms/Losses/RootMeanSquaredError) | 8 |  | 9 | |
+| 7 | [Root Mean Squared Error](./Algorithms/Losses/RootMeanSquaredError) | 8 | [Triplet Loss](./Algorithms/Losses/RankingLosses/TripletLoss) | 9 | [Pairwise Ranking Loss](./Algorithms/Losses/RankingLosses/PairWiseRankingLoss) |
 
 
 ## Available Documentations