Normalizing embeddings ensures uniform vector scales, improving cosine similarity-based matching for unseen classes in One-Shot Learning.
Here is the code snippet you can refer to:

In the above code, we are using the following key points:
-
Uses L2 normalization to scale vectors to unit length.
-
Handles edge cases (e.g., zero vectors) safely using epsilon.
-
Enhances similarity comparison for matching queries and support samples.
Hence, normalization aligns embedding magnitudes, enabling reliable similarity comparisons in One-Shot Learning tasks.