Added Temporal Score functions

AUTHORS.rst

@@ -14,4 +14,5 @@ Contributors
 * Rohit Patil <[email protected]>
 * Parantak Singh <[email protected]>
 * Abheesht Sharma <[email protected]>
+* Harshit Pandey <[email protected]>
 

decepticonlp/metrics/char_metrics.py

@@ -13,7 +13,7 @@ class CharacterMetrics(metaclass=abc.ABCMeta):
         An abstract class used to represent the character metrics. Subclasses implement the calculate method.
         Methods
         -------
-        apply(self, text1: str, text2: str, **kwargs)
+        calculate(self, text1: str, text2: str, **kwargs)
             - calculates the similarity/distance between two strings using the appropriate metric.
     """
 

decepticonlp/metrics/temporal_metrics.py

@@ -0,0 +1,109 @@
+import torch
+import torch.nn as nn
+
+
+class RankWords:
+    """
+        Accepts a feature vector torch tensor and outputs a temporal ranking of each word
+        inputs of shape (number of words,feature_vector_size)
+        Methods
+        --------
+        temporal_score(self, model, sentence)
+            - calculates the temporal score each word in a sentence(considering the first i words) 
+        temportal_tail_score(self, model, sentence)
+            - calculates the temporal score each word in a sentence(considering n-i words)
+        combined_score(self, model, sentence, lambda_)
+            - calculates the combined score taking into consideration temporal score and tailed temporal score with lambda as the weight
+    """
+
+    def temporal_score(self, model, sentence):
+        """ 
+            Considering a input sequence x1,x2,...,xn
+            we calculate T(xi) = F(x1,x2,...,xi) - F(x1,x2,...,xi-1)
+            where F is one pass through a RNN cell 
+
+            Example:
+
+            >> import temporal_metrics
+            >> rw = temporal_metrics.RankWords()
+            >> embedded_word = embed("i like dogs") #embed should convert the sentence to a (sent_length * embedding_dim) shape tensor
+            >> print(rw.temporal_score(model,embedded_word))  #seed = 99 
+            tensor([[1.1921e-07],[2.2054e-05],[8.4886e-01]])
+
+            :params
+            :model : The model must output torch tensor of dimension (1,1,1) as its output
+            :sentence : An tensor of shape (sent_length * embedding_dim) where each word has been embedded 
+
+            returns temporal score of each word
+            :return type: torch.Tensor
+        """
+        inputs = sentence.reshape(len(sentence), 1, -1)
+        with torch.no_grad():
+            pred = model(inputs)
+            losses = torch.zeros(inputs.shape[0:2])
+            for i in range(inputs.size()[0]):
+                tempinputs = inputs.clone()
+                tempinputs[i, :, :].zero_()
+                tempoutput = model(tempinputs)
+                losses[i] = torch.dist(
+                    tempoutput.squeeze(1), pred.squeeze(1)
+                )  # L2 Norm
+        return losses
+
+    def temportal_tail_score(self, model, sentence):
+        """ 
+            Considering a input sequence x1,x2,...,xn
+            we calculate T(xi) = F(xi,xi+1,...,xn) - F(xi+1,xi+2...,xn)
+            where F is one pass through a RNN cell
+
+            Example:
+
+            >> import temporal_metrics
+            >> rw = temporal_metrics.RankWords()
+            >> embedded_word = embed("i like dogs") #embed should convert the sentence to a (sent_length * embedding_dim) shape tensor
+            >> print(rw.temportal_tail_score(model,embedded_word)) #seed = 99
+
+            tensor([[5.9605e-08],[1.9789e-05],[0.0000e+00]])
+
+            :params
+            :model : The model must output torch tensor of dimension (1,1,1) as its output
+            :sentence : An tensor of shape (sent_length * embedding_dim) where each word has been embedded 
+
+            returns temporal score of each word
+            :return type: torch.Tensor
+        """
+        inputs = sentence.reshape(len(sentence), 1, -1)
+        losses = torch.zeros(inputs.shape[0:2])
+        with torch.no_grad():
+            for i in range(inputs.size()[0] - 1):
+                pred = model(inputs[i:, :, :])
+                tempinputs = inputs[i + 1 :, :,].clone()
+                tempoutput = model(tempinputs)
+                losses[i] = torch.dist(tempoutput, pred)  # L2 Norm
+        return losses
+
+    def combined_score(self, model, sentence, lambda_):
+        """ 
+            Computes combined temporal_score, temportal_tail_score
+            Combined Score = temporal_score + λ(temportal_tail_score)
+
+            Example:
+
+            >> import temporal_metrics
+            >> rw = temporal_metrics.RankWords()
+            >> embedded_word = embed("i like dogs") #embed should convert the sentence to a (sent_length * embedding_dim) shape tensor
+            >> print(rw.combined_score(model,embedded_word,0.5)) #seed = 99
+
+            tensor([[6.8545e-07], [3.8507e-03], [8.4886e-01]])
+
+            :params
+            :model : The model must output torch tensor of dimension (1,1,1) as its output
+            :sentence : An tensor of shape (sent_length * embedding_dim) where each word has been embedded 
+            :lambda_ : (float) , 0 <= lambda_ <= 1 
+
+            returns temporal score of each word
+            :return type: torch.Tensor
+        """
+        return self.temporal_score(
+            model, sentence
+        ) + lambda_ * self.temportal_tail_score(model, sentence)

requirements.txt

@@ -9,7 +9,7 @@ Sphinx==1.8.5
 twine==1.14.0
 tensorflow==2.2.0
 tensorflow-hub==0.8.0
-
+torch==1.5.0
 setuptools
 pytest==5.4.2
 pytest-runner==5.2