jeong-tae/deepfashion-tensorflow

deepfashion-tensorflow's Issues

test acc still low

Hi,
As you mentioned in text, I got a same validation acc as you, but the test acc is really low and sames do not make sense.
Have you got the idea why it work not well?

Another question.
Why feature dimensions extract from vgg16/fc7 is 1536-d?

attributes loss 계산할 때

제가 잘은 모르지만..
논문을 보면 아래 내용이 있던데

Third, a weighted cross-entropy loss is utilized to predict attributes .....

attributes loss 계산할 때 tf.nn.weighted_cross_entropy_with_logits 사용해야하는거 아닌가요?

triplet 데이터 만들 때

q, p, n 을 만드는데
n 이야 같은 카테고리의 다른 옷이면 되지만
q 와 p 는 같은 옷인데 다른 사진으로 해야하지 않나요?