migrate metric learning example to keras 3

Refactor loss function and tensor handling in metric learning for compatibility with Keras 3.

Author: Shi-pra-19

examples/vision/metric_learning_tf_similarity.py

@@ -54,7 +54,7 @@
 import numpy as np
 
 import tensorflow as tf
-from tensorflow import keras
+import keras
 
 import tensorflow_similarity as tfsim
 
@@ -216,13 +216,14 @@
 val_steps = 50
 
 # init similarity loss
-loss = tfsim.losses.MultiSimilarityLoss()
+loss = tfsim.losses.MultiSimilarityLoss(reduction='sum_over_batch_size')
 
 # compiling and training
 model.compile(
     optimizer=keras.optimizers.Adam(learning_rate),
     loss=loss,
     steps_per_execution=10,
+    run_eagerly=True, 
 )
 history = model.fit(
     train_ds, epochs=epochs, validation_data=val_ds, validation_steps=val_steps
@@ -321,7 +322,7 @@
 for idx in np.argsort(y_display):
     tfsim.visualization.viz_neigbors_imgs(
         x_display[idx],
-        y_display[idx],
+        y_display[idx].numpy(),
         nns[idx],
         class_mapping=class_mapping,
         fig_size=(16, 2),
@@ -394,7 +395,7 @@
 """
 
 idx_no_match = np.where(np.array(matches) == 10)
-no_match_queries = x_confusion[idx_no_match]
+no_match_queries = no_match_queries = keras.ops.take(x_confusion, keras.ops.cast(idx_no_match[0], dtype="int32"), axis=0)
 if len(no_match_queries):
     plt.imshow(no_match_queries[0])
 else: