fix: overfit test (#137)

* fix: overfit test

I realized I put the wrong sign in the overfit test.
I will also change the margin in it to 0.5

* fix: keras euclidean triplet loss

* fix: reduce distance margin

* fix: sqrt-induced possible nans

finetuner/tuner/keras/losses.py

@@ -48,7 +48,13 @@ def call(self, inputs, **kwargs):
         # https://github.com/tensorflow/tensorflow/issues/12071
         dist_pos = tf.reduce_sum(tf.math.squared_difference(anchor, positive), axis=-1)
         dist_neg = tf.reduce_sum(tf.math.squared_difference(anchor, negative), axis=-1)
-        return tf.reduce_mean(tf.nn.relu(dist_pos - dist_neg + self._margin))
+
+        dist_pos = tf.maximum(dist_pos, 1e-9)
+        dist_neg = tf.maximum(dist_neg, 1e-9)
+
+        return tf.reduce_mean(
+            tf.nn.relu(tf.sqrt(dist_pos) - tf.sqrt(dist_neg) + self._margin)
+        )
 
 
 class CosineTripletLoss(BaseLoss, Layer):

tests/integration/keras/test_overfit.py

@@ -52,7 +52,7 @@ def test_overfit_keras(
     vec_embedings = embed_model(vecs).numpy()
 
     # Compute distances between embeddings
-    metric = 'sqeuclidean' if loss.startswith('Euclidean') else 'cosine'
+    metric = 'euclidean' if loss.startswith('Euclidean') else 'cosine'
     dists = squareform(pdist(vec_embedings, metric=metric))
 
     # Make sure that for each class, the two instances are closer than
@@ -62,4 +62,4 @@ def test_overfit_keras(
         dist_other = dists[2 * i : 2 * i + 2, :].copy()
         dist_other[:, 2 * i : 2 * i + 2] = 10_000
 
-        assert cls_dist < dist_other.min() + 1
+        assert cls_dist < dist_other.min() - 0.1

tests/integration/paddle/test_overfit.py

@@ -54,7 +54,7 @@ def test_overfit_paddle(
     vec_embedings = embed_model(paddle.Tensor(vecs)).numpy()
 
     # Compute distances between embeddings
-    metric = 'sqeuclidean' if 'Euclidean' in loss else 'cosine'
+    metric = 'euclidean' if 'Euclidean' in loss else 'cosine'
     dists = squareform(pdist(vec_embedings, metric=metric))
 
     # Make sure that for each class, the two instances are closer than
@@ -64,4 +64,4 @@ def test_overfit_paddle(
         dist_other = dists[2 * i : 2 * i + 2, :].copy()
         dist_other[:, 2 * i : 2 * i + 2] = 10_000
 
-        assert cls_dist < dist_other.min() + 1
+        assert cls_dist < dist_other.min() - 0.1

tests/integration/torch/test_overfit.py

@@ -54,7 +54,7 @@ def test_overfit_pytorch(
         vec_embedings = embed_model(torch.Tensor(vecs)).numpy()
 
     # Compute distances between embeddings
-    metric = 'sqeuclidean' if 'Euclidean' in loss else 'cosine'
+    metric = 'euclidean' if 'Euclidean' in loss else 'cosine'
     dists = squareform(pdist(vec_embedings, metric=metric))
 
     # Make sure that for each class, the two instances are closer than
@@ -64,4 +64,4 @@ def test_overfit_pytorch(
         dist_other = dists[2 * i : 2 * i + 2, :].copy()
         dist_other[:, 2 * i : 2 * i + 2] = 10_000
 
-        assert cls_dist < dist_other.min() + 1
+        assert cls_dist < dist_other.min() - 0.1