how to implement SVM inplace of KNN in face recognition

[008karan]

I tried to implement SVM in place of KNN but got stuck. In KNN we compare the distance from the nearest neighbor with a threshold value. How to do these in SVM? I made changes here. svm_clf = SVC(gamma='auto', kernel='rbf', C=20) svm_clf.fit(X, y)

but don't know what changes i need to do here: closest_distances= knn_clf.kneighbors(faces_encodings, n_neighbors=1)

is_recognized = [closest_distances[0][i][0] <= DIST_THRESH for i in range(len(X_faces_loc))] return [(pred, loc) if rec else ("N/A", loc) for pred, loc, rec in zip(knn_clf.predict(faces_encodings), X_faces_loc, is_recognized)]

[donaldxu]

I think you can use predict_proba instead

[Roosh27]

@ageitgey Can you please help on this?

[lanzani]

Hi! In SVM there no something like "closest_distances", but you can calculate the "probability" (the prediction confidence). If you combine this with a threshold you can obtain the "Unknown label". To do so:

Step 1:

Add the parameter "probability=True" svm_clf = SVC(gamma='auto', kernel='rbf', probability=True, C=20, ) svm_clf.fit(X, y)

Step 2:

When you do the prediction use: svm_clf.predict_proba(face_encodings) This function returns the value of confidence for each trained element.

I have done a fast implementation about this. It works with only one face for img, I've not tested it with multiple faces. Try by yourself: def predict_SVM(faces_encodings):

    known_faces_names = ["face1", "face2"]
    DIST_THRESH = 0.7

    recognized_faces = []

    face_distances = svm_clf.predict_proba(faces_encodings)
    for i in range(len(faces_encodings)):
        name = "N/A"
        max_index = np.argmax(face_distances[i])
        if face_distances[i][max_index] >= DIST_THRESH:
            name = known_faces_names[max_index]
        recognized_faces.append(name)

    return recognized_faces

I suggest to read this: https://prateekvjoshi.com/2015/12/15/how-to-compute-confidence-measure-for-svm-classifiers/

It's very clear and explain how this works and also how the implementation works. Bye 👋