I am trying to understand the deploy.prototxt example under Caffe's blvc_Googlenet
While trying to solve this problem I have: Caffe error while testing: Check failed: count_ == proto.data_size() (9408 vs. 0) I have gone through the deploy.prototxt already provided.
- Why does the deploy example have dropout layer? Is not it something we need just for training?
-
Since it is Googlenet, we have 3 losses (train_val.prototxt), I basically got rid of them all in deploy.prototxt but the last, as I only need the final prediction:
layer { name: "loss3/classifier" type: "InnerProduct" bottom: "pool5/7x7_s1" top: "loss3/classifier" inner_product_param { num_output: 7 weight_filler { type: "xavier" } bias_filler { type: "constant" value: 0 } } }
layer { name: "prob" type: "Softmax" bottom: "loss3/classifier" top: "prob" }
But I now wonder if I should keep the previous loss layers?
I have done training/testing with caffe using deploy.prototxt before but I seem to have trouble with Googlenet so I think I might be missing something. Sorry I can't articulate the problem better as I don't have much clue why it fails.
Aucun commentaire:
Enregistrer un commentaire