Giters

lzrobots / DeepEmbeddingModel_ZSL

Tensorflow code for CVPR 2017 paper: Learning a Deep Embedding Model for Zero-Shot Learning

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

visual feature size

GanAndreas opened this issue · comments

commented

HI, I'm currently try to implement the DeepEmbed with various CNN. I don't really understand the size of the visual feature. In the paper, it is stated that you are using 1024 dimension. But in your coding you are using 2048 which is doubled. I already check the Inception-V2. It also produce 1024 just right before the FC. Can anyone explain to me why it is 2048?

Thanks

commented

Hi please read the REAMME file, there are two setting,

The following are all 1024 dimension:
AwA_attribute.py will give you ZSL performance on AwA with attribute.
AwA_wordvector.py will give you ZSL performance on AwA with wordvector.
AwA_fusion.py will give you ZSL performance on AwA with attribute and wordvector fusion.
CUB_attribute.pywill give you ZSL performance on CUB with attribute.

While to faire compare the GBU setting methods which use ResNet feature (2048 dimension), we need to use 2048 dimension:

ZSL and GZSL performance evaluated under GBU setting [1]: ResNet feature, GBU split, averaged per class accuracy.
AwA1_GBU.py will give you ZSL and GZSL performance on AwA1 with attribute under GBU setting [1].
AwA2_GBU.py will give you ZSL and GZSL performance on AwA2 with attribute under GBU setting [1].
CUB1_GBU.py will give you ZSL and GZSL performance on CUB with attribute under GBU setting [1].
aPY_GBU.py will give you ZSL and GZSL performance on aPY with attribute under GBU setting [1].
SUN_GBU.py will give you ZSL and GZSL performance on SUN with attribute under GBU setting [1].

commented

I see.
thank you for the information.