def evaluate_embeddings(train_embedings, test_embedings, train_labels,
test_labels, k, n_cell_types):
# Pairwise euclidean distance between test_embedings and train_embedings
dist = np.expand_dims(test_embedings, axis=1) - np.expand_dims(
train_embedings, axis=0)
dist_mat = np.sqrt(np.sum(np.square(dist), axis=-1))
# Pick k nearest neighbors for each test example
knn_indices = np.argsort(dist_mat)[:, :k]
knn_dist = np.vstack(
[dist_mat[i, knn_indices[i, :]] for i in range(knn_indices.shape[0])])
knn_labels = np.vstack(
[train_labels[knn_indices[i, :]] for i in range(knn_indices.shape[0])])
# Convert distance to similarity
knn_sim_2d = 1 / (1 + knn_dist)
knn_sim_3d = np.expand_dims(knn_sim_2d, axis=-1)
# One-hot array to count occurances of each cell type among the k nearst neighbors
knn_labels_one_hot = LabelBinarizer().fit_transform(knn_labels.flatten())
knn_labels_one_hot = knn_labels_one_hot.reshape(
(knn_labels.shape[0], knn_labels.shape[1], -1))
# Use the similarity as weights. Count occurances of each cell type among the kompute overall mean average precisions (MAP)
knn_sim_cell_type = np.sum(knn_sim_3d * knn_labels_one_hot, axis=1)
# Compute accuracy
pred_labels = np.argmax(knn_sim_cell_type, axis=1)
accuracy = accuracy_score(test_labels, pred_labels)
# Compute overall mean average precisions (MAP)
pred_decision_val = np.max(knn_sim_cell_type, axis=1)
overall_map = mean_average_precision(test_labels, pred_labels,
pred_decision_val)
# Compute cell-type-specific MAP
cell_type_map = get_cell_type_map(n_cell_types, test_labels, pred_labels,
pred_decision_val)
return (accuracy, overall_map, cell_type_map)
