initial commit of project

evaluation/helpers/approx_knn_classifier.py

@@ -0,0 +1,45 @@
+# adapted from:
+# Pellegrini., C., E. Özsoy., M. Wintergerst., and G. Groh. (2021).
+# “Exploiting Food Embeddings for Ingredient Substitution.”
+# In: Proceedings of the 14th International Joint Conference on Biomedical
+# Engineering Systems and Technologies - Volume 5: HEALTHINF, INSTICC.
+# SciTePress, pp. 67–77. isbn: 978-989-758-490-9. doi: 10.5220/0010202000670077.
+
+from pathlib import Path
+
+import numpy as np
+from annoy import AnnoyIndex
+from tqdm import tqdm
+
+
+# Full guide https://github.com/spotify/annoy
+class ApproxKNNClassifier:
+    def __init__(self, all_ingredient_embeddings, max_embedding_count,
+                 save_path=Path('data/eval/approx_knn_classifier.ann'), n_trees=10):
+
+        vector_length = all_ingredient_embeddings.shape[-1]
+        self.max_embedding_count = max_embedding_count
+        if save_path.exists():
+            print('Loading Existing Approx Classifier')
+            self.approx_knn_classifier = AnnoyIndex(vector_length, 'angular')
+            self.approx_knn_classifier.load(str(save_path))  # super fast, will just mmap the file
+        else:
+
+            # To make sure we don't just get ourselves: add max_embedding_count
+            self.approx_knn_classifier = AnnoyIndex(vector_length, 'angular')  # Length of item vector that will be indexed
+            for i in tqdm(range(len(all_ingredient_embeddings)), total=len(all_ingredient_embeddings), desc='Creating Approx Classifier'):
+                self.approx_knn_classifier.add_item(i, all_ingredient_embeddings[i])
+
+            self.approx_knn_classifier.build(n_trees)
+            print('Saving Approx Classifier')
+            self.approx_knn_classifier.save(str(save_path))
+
+    def k_nearest_neighbors(self, ingredient_embeddings):
+        all_indices, all_distances = [], []
+        for idx, ingredient_embedding in enumerate(
+                ingredient_embeddings):  # search_k gives you a run-time tradeoff between better accuracy and speed currently defaults
+            indices, distances = self.approx_knn_classifier.get_nns_by_vector(ingredient_embedding, self.max_embedding_count + 200, include_distances=True)
+            all_indices.append(indices)
+            all_distances.append(distances)
+
+        return np.stack(all_distances), np.stack(all_indices)