0

I want to try CaffeNet (from caffe/models/bvlc_reference_caffenet) for image classification but my data set has only 3 labels. I see that last layer in CaffeNet has 1000 outputs. May I just change it to 3? And more general question: is it ok to use such complex network for image classification if you have train data with ~1500 samples and 3 labels?

John Tracid
  • 3,836
  • 3
  • 22
  • 33
  • You can try replacing the last layer with 1000 nodes to a layer with 3 nodes. Freeze the rest of the layers and train only that layer. – Autonomous May 21 '17 at 18:34
  • i think its fine to map to 3 labels...ideally you could map to 4 labels and ignore the last one...i am not sure why but sometimes caffe refuses to work if you have unlabbeled objects – Eliethesaiyan May 22 '17 at 12:15
  • You can change it to 3 labels. I also used Caffe and I changed the output to 2. For more information, please visit this link [Caffe Layers](https://stackoverflow.com/questions/36841158/fine-tuning-of-googlenet-model) – Ashutosh Singla May 14 '18 at 16:39

0 Answers0