Deeplab v2 deployment

120 views
Skip to first unread message

srikanth sridhar

unread,
Jul 5, 2017, 8:30:17 AM7/5/17
to Caffe Users
Hello all,

I am trying to use the deeplab v2 codes.I was able to train on a dataset but when I want to run it for a single image,i am facing some issues.


When I am training the input layer type is ImageSegData.Since i cannot find the deploy file,the attached deploy file is the one I created from the train.prototxt file.

My doubts are as follows,

1.Is it necessary the input data layer in deploy and train prototxts,should be same?
2.I tried using a MemoryData layer where I will do the preprocessing of the image and i will load the image like

net.blobs['data'].data[...]=image
net.forward()

 but i am getting the following error "MemoryDataLayer needs to be initialized by calling Reset"


The documentation for memory data layer is not available available and it seems soon it will replaced by "input" layer.

Any suggestions on this will be helpful



deploy.prototxt
train.prototxt

nzif

unread,
Jan 2, 2018, 11:11:37 AM1/2/18
to Caffe Users
Hello,

Was wondering if you have figured out this problem? I facing same issue and what I want to is extracting fc7 features for visualization.
Is there any way to visualize DeepLab intermediate features?
Your insights would be helpful.

TIA
Reply all
Reply to author
Forward
0 new messages