search

BartyzalRadek / Multi-label-Inception-net

Multi-label image classification using pretrained Inception net.
329 stars 112 forks source link

Data overfitting and wrong classification result #9

Open vishalghor opened 6 years ago

vishalghor commented 6 years ago

hi @BartyzalRadek ,

I followed the code changes which you did for multi label classification for a custom dataset.

But the model tends to get overfit and also the output classification result for a images is not having individual probablities as you have depicted in the sample image i.e. one wiith car having 0.64 and accident as 0.39 probability. i have tried your code for flower photos dataset similar to the dataset used in single label implementation by tensorflow. but i am getting probability same as single label classifiation.

Kindly help me in resolving this issue.

tianyuecao commented 5 years ago

Hi, I have the similar question with you, I train on MS-COCO 2014, and find that the results only classify some of the pictures into 'person' but no other classes. Have you solved your problem? Tks.

vishalghor commented 5 years ago

Hi @tianyuecao i had resolved this issue, in my case the issue was due to not proper shuffling of data and the probabilty were resolved by using tf.sigmoid for activation.

civilman628 commented 5 years ago

@tianyuecao @tianyuecao I have the same issue. i think the activation function is already sigmoid. which line need to change?