training error about facenet

[davidsandberg]

Hi,
It looks like there are too few images in the validation set.
The function "sample_people" tries to sample 45 persons with 40 images per person. But if some of the classes in the dataset contains less than 40 images more classes will be used to get 45*40 images.
One way to solve it would be to use a smaller "train_set_fraction", e.g. 0.8 or something. Then the validation set will be larger. Another (ugly) way could be to change the seed. Then a different set of classes will be included in the validation set, and perhaps these classes together will have enough images for validation.

[PoonamZ]

@davidsandberg I put --lfw_dir place empty as i don't want to see the validation accuracy. Then also I am getting the error as below,

Traceback (most recent call last):
File "src/train_tripletloss.py", line 486, in
main(parse_arguments(sys.argv[1:]))
File "src/train_tripletloss.py", line 186, in main
args.embedding_size, anchor, positive, negative, triplet_loss)
File "src/train_tripletloss.py", line 212, in train
image_paths, num_per_class = sample_people(dataset, args.people_per_batch, args.images_per_person)
File "src/train_tripletloss.py", line 327, in sample_people
class_index = class_indices[i]
IndexError: index 4 is out of bounds for axis 0 with size 4

[kaishijeng]

Can I reduce 40 images/person? What is the impact of doing this?
Thanks

On Thu, May 19, 2016 at 10:56 PM, David Sandberg [email protected]
wrote:

Hi,
It looks like there are too few images in the validation set.
The function "sample_people" tries to sample 45 persons with 40 images per
person. But if some of the classes in the dataset contains less than 40
images more classes will be used to get 45*40 images.
One way to solve it would be to use a smaller "train_set_fraction", e.g.
0.8 or something. Then the validation set will be larger. Another (ugly)
way could be to change the seed. Then a different set of classes will be
included in the validation set, and perhaps these classes together will
have enough images for validation.

—
You are receiving this because you authored the thread.
Reply to this email directly or view it on GitHub
#18 (comment)

[davidsandberg]

Yes, reducing either the number of images per person or the number of people would help to solve the problem. But i'm not sure how it will affect the training. Also, you need to make sure that you can get a whole number of batches, i.e. that people_per_batch*images_per_person = batch_size * n, where n is an integer.

[Shubhamcl]

Also making sure the batch_size and people_per_batch are multiples of 3 helps.

[KumailHussain]

@PoonamZ @davidsandberg any findings for the above error, I am getting the same error, Thanks in advance. I have training set of 21 images each class

[PoonamZ]

@KumailHussain I didn't get any solution for this.

[sassiaicha]

Hi, I am using "train_tripletloss.py" as training algorithm, but every time I try to run it, I get the error below and
I can't find any solution for that, please help me. Thank you in advance.

Restoring pretrained model: ./models/20170512-110547/model-20170512-110547.ckpt-250000
Traceback (most recent call last):
File "train_tripletloss.py", line 493, in
main(parse_arguments(sys.argv[1:]))
File "train_tripletloss.py", line 192, in main
args.embedding_size, anchor, positive, negative, triplet_loss)
File "train_tripletloss.py", line 218, in train
image_paths, num_per_class = sample_people(dataset, args.people_per_batch, args.images_per_person)
File "train_tripletloss.py", line 333, in sample_people
class_index = class_indices[i]
IndexError: index 2 is out of bounds for axis 0 with size 2

[RohanMathur17]

@sassiaicha , was facing the same error during training. Same Index error. Looks like you have chosen only 2 classes which is causing the problem.
I myself was using only 10 custom classes with 3 images for each. But when I increased my dataset (used LFW dataset for training), the code ran for training. Maybe try increasing your dataset.