added numpy requirements

requirements.txt

@@ -1,3 +1,4 @@
 # TODO: fix github to release
 git+https://github.com/huggingface/evaluate.git@505123230059f9605da8951880eddc9d1fbf4278
-datasets~=2.0
+datasets~=2.0
+numpy

triplet_margin_loss.py

@@ -15,14 +15,15 @@
 
 import datasets
 import evaluate
+import numpy as np
 
 
 _DESCRIPTION = """
-Triplet margin loss is a loss function that measures a relative similarity between the samples.
-A triplet is comprised of reference input 'anchor(a)', matching input 'positive examples(p)' and non-matching input 'negative examples(n)'.
-The loss function for each triplet is given by:
-L(a, p, n) = max{d(a,p) - d(a,n) + margin, 0}
-where d(x, y) is the 2nd order (Euclidean) pairwise distance between x and y
+Triplet margin loss is a loss function that measures a relative similarity between the samples.\n
+A triplet is comprised of reference input 'anchor (a)', matching input 'positive examples (p)' and non-matching input 'negative examples (n)'.
+The loss function for each triplet is given by:\n
+L(a, p, n) = max{d(a,p) - d(a,n) + margin, 0}\n
+where d(x, y) is the 2nd order (Euclidean) pairwise distance between x and y.
 """
 
 
@@ -97,8 +98,8 @@ class TripletMarginLoss(evaluate.EvaluationModule):
         )
 
     def _compute(self, anchor, positive, negative, margin=1.0):
-        d_a_p = sum((anchor - positive)**2)
-        d_a_n = sum((anchor - negative)**2)
+        d_a_p = np.sqrt(sum((anchor - positive)**2))
+        d_a_n = np.sqrt(sum((anchor - negative)**2))
         return {
             "accuracy": float(
                 max(d_a_p - d_a_n + margin, 0)