My task requires me to extract features from the pool5 layer of vgg-19 net. Previously I was doing it using Caffe and then working on the extracted features using Tensorflow for further training. But now, I want to add a deconv layer on top and fine-tune the entire net for end-to-end training. Now my gradients are obtained using tensroflow while I had obtained the features from vgg-19 using caffe ,so I guess it will be difficult to backpropagate these gradients to caffe. So that is the reason I am looking for a vgg-19 model in tensorflow that can he used for extracting features from any layer. I found some projects online but they are not officially endorsed by Tensorflow.
Asked
Active
Viewed 424 times
0
-
Keras has a VGG-19 module in keras.applications. – Dr. Snoopy Aug 02 '17 at 17:59
-
Thanks Matias but I was looking for something in tensorflow because I have already coded my couple of layers and a compliated objective function in tensorflow and it will take more time to port all that code back to keras – HIMANSHU RAI Aug 02 '17 at 18:14
1 Answers
0
For tensorflow you can use the TF-Slim library and use the pretrained models which are available here. You can check their useful slim walk through notebook on how to use them.
Vijay Mariappan
- 16,921
- 3
- 40
- 59
-
Thanks vijay for your answer! So just a followup question. Are you aware of its performance as compared to caffe? – HIMANSHU RAI Aug 03 '17 at 01:52
-
Last time i checked(https://arxiv.org/pdf/1608.07249.pdf) tf performance is not as good as caffe, but i am not sure about the recent versions. – Vijay Mariappan Aug 03 '17 at 02:15
-
@vijaym unfortunately, the links you have provided is not active any more – osmancakirio Jul 23 '18 at 16:02