search

david-husx / crossview_localisation

82 stars 30 forks source link

How to initialize from a pre-trained VGG16 model #2

Open Liumouliu opened 6 years ago

Liumouliu commented 6 years ago

Hi,

Thank you for this interesting work. Here I have a silly question: How to initialize the VGG16 weights from a pre-trained model.

Since the VGG16 of your method is defined by yourself, I have used slim to initialize the weights, but failed. Could you please add that part?

Thank you

david-husx commented 6 years ago

Hello,

Thank you for your interests.

I will upload a tensorflow model to initialise the VGG16 weights. There are some problems of our server so it's late. I will use google drive to upload it in this week.

Originally I use python to manully load vgg16 model (dict type saved in numpy). It is complicated and a little bit messy so I didn't release this part. I recommend you use my tensorflow model to initialise the network.

Sorry for the trouble. I will upload the model soon.