CODEWITHSUNDEEP
pythonpytorchtensor
cfive
cfive

Reputation: 61

Sorting a tensor list in ascending order

I am working on a facial comparison app that will give me the closest n number of faces to my target face.

I have done this with dlib/face_recognition as it uses numpy arrays, however i am now trying to do the same thing with facenet/pytorch and running into an issue because it uses tensors.

I have created a database of embeddings and I am giving the function one picture to compare to them. What i would like is for it to sort the list from lowest distances to highest, and give me the lowest 5 results or so.

here is the code I am working on that is doing the comparison. at this point i am feeding it a photo and asking it to compare against the embedding database.

def face_match(img_path, data_path): # img_path= location of photo, data_path= location of data.pt 
    # getting embedding matrix of the given img
    img_path = (os.getcwd()+'/1.jpg')
    img = Image.open(img_path)
    face = mtcnn(img) # returns cropped face and probability
    emb = resnet(face.unsqueeze(0)).detach() # detech is to make required gradient false

    saved_data = torch.load('data.pt') # loading data.pt file
    embedding_list = saved_data[0] # getting embedding data
    name_list = saved_data[1] # getting list of names
    dist_list = [] # list of matched distances, minimum distance is used to identify the person
    
    for idx, emb_db in enumerate(embedding_list):
        dist = torch.dist(emb, emb_db)
        dist_list.append(dist)
    
    namestodistance = list(zip(name_list,dist_list))
    
    print(namestodistance)

face_match('1.jpg', 'data.pt')

This results in giving me all the names and their distance from the target photo in alphabetical order of the names, in the form of (Adam Smith, tensor(1.2123432)), Brian Smith, tensor(0.6545464) etc. If the 'tensor' wasn't part of every entry I think it would be no problem to sort it. I don't quite understand why its being appended to the entries. I can cut this down to the best 5 by adding [0:5] at the end of dist_list, but I can't figure out how to sort the list, I think the problem is the word tensor being in every entry.

I have tried for idx, emb_db in enumerate(embedding_list): dist = torch.dist(emb, emb_db) sorteddist = torch.sort(dist) but for whatever reason this only returns one distance value, and it isn't the smallest one.

idx_min = dist_list.index(min(dist_list)), this works fine in giving me the lowest value and then matching it to a name using namelist[idx_min], therefore giving the best match, but I would like the best 5 matches in order as opposed to just the best match.

Anyone able to solve this ?

Upvotes: 1

Views: 836

Answers (1)

AlexNe
AlexNe

Reputation: 959

Unfortunately I cannot test your code, but to me it seems like you are operation on a python list of tuples. You can sort that by using a key:

namestodistance = [('Alice', .1), ('Bob', .3), ('Carrie', .2)]
names_top = sorted(namestodistance, key=lambda x: x[1])
print(names_top[:2])

Of course you have to modify the anonymous function in key to return a sortable value instead of e.g. a torch.tensor.

This can be done by using key = lambda x: x[1].item().

Edit: To answer the question that crept up in the comments, we can refactor our code a little. Namely

namestodistance = list(map(lambda x: (x[0], x[1].item()), namestodistance)
names_top = sorted(namestodistance, key=lambda x: x[1])
print(names_top[:2])

Upvotes: 1

Related Questions