search

machrisaa / tensorflow-vgg

VGG19 and VGG16 on Tensorflow
2.21k stars 1.08k forks source link

I wonder if my vgg npy file broken #30

Closed ruiann closed 7 years ago

ruiann commented 7 years ago

The range of numbers in the output tensor generated by conv5_4 is very big, can up to 1e5, is this the correct output?

ruiann commented 7 years ago

or for different result the numbers have different range? when I try to compare two feature maps by Euclid distance, the loss is very big and able to hide other parts of loss function

machrisaa commented 7 years ago

Hi @ruiann, it could be big. It's difficult to tell if your npy file is broken or not. The only thing I will do is to run the test to check if it can classify the sample images (tiger and puzzle).

I added this line and check the magnitude of the output of conv5_4:

https://github.com/machrisaa/tensorflow-vgg/blob/master/test_vgg19.py

The max is just 300-ish for the tiger image. See if you can get the same result or you may need to download the npy again.

ruiann commented 7 years ago

@machrisaa yeah the result is ok, but its hard to reduce the Euclid loss of the feature map, I think I need to do more investment, thank you XD