n3011/inception_v3_googlenet

GoogleNet Inception v3

Inception_v3_GoogLeNet

Implementation of Inception v3 in Caffe

Training and testing prototxt for Google Inception_v3 New Model

Network architecture was presented in : Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, Zbigniew Wojna, "Rethinking the Inception Architecture for Computer Vision", http://arxiv.org/abs/1512.00567

Implemented using Batch Normalization and Parametric rectified linear Unit (PReLU)

inception_v3_googlenet's Issues

Can you share your trained model?

Hi,

Can you share your trained .caffemodel file?

Thanks

Why I get the error "misaligned address" when training ?

However, I never get error like this when I train other nets like googLeNet. Here are logs:

F0925 14:07:56.489186 14091 math_functions.cu:79] Check failed: error == cudaSuccess (74 vs. 0)  misaligned address
*** Check failure stack trace: ***
    @     0x7f93c3722daa  (unknown)
    @     0x7f93c3722ce4  (unknown)
    @     0x7f93c37226e6  (unknown)
    @     0x7f93c3725687  (unknown)
    @     0x7f93c3ec36c8  caffe::caffe_gpu_memcpy()
    @     0x7f93c3e7e58e  caffe::SyncedMemory::to_gpu()
    @     0x7f93c3e7da39  caffe::SyncedMemory::gpu_data()
    @     0x7f93c3d2e522  caffe::Blob<>::gpu_data()
    @     0x7f93c3effe3e  caffe::BiasLayer<>::Backward_gpu()
    @     0x7f93c3f00917  caffe::ScaleLayer<>::Backward_gpu()
    @     0x7f93c3eaa097  caffe::Net<>::BackwardFromTo()
    @     0x7f93c3eaa201  caffe::Net<>::Backward()
    @     0x7f93c3e9df92  caffe::Solver<>::Step()
    @     0x7f93c3e9e849  caffe::Solver<>::Solve()
    @           0x40843e  train()
    @           0x405c8c  main
    @     0x7f93c227ff45  (unknown)
    @           0x40645d  (unknown)
    @              (nil)  (unknown)

I have checked my address, but nothing wrong.

Your last layers in training are wrong

The last layers should hold different names (or types) :

layer {
  name: "loss1/top-1" <== the name is loss
  type: "Accuracy" <== the type is accuracy
  bottom: "loss1/classifier"
  bottom: "label"
  top: "loss1/top-1"
  include {
    phase: TEST
  }
}
layer {
  name: "loss1/top-3" <== the name is loss
  type: "Accuracy" <== the type is accuracy
  bottom: "loss1/classifier"
  bottom: "label"
  top: "loss1/top-3"
  include {
    phase: TEST
  }
  accuracy_param {
    top_k: 3
  }

It should be like this :

layer {
  name: "acc/top-1"
  type: "Accuracy"
  bottom: "fc1"
  bottom: "label"
  top: "acc/top-1"
  include {
    phase: TEST
  }
}
layer {
  name: "acc/top-5"
  type: "Accuracy"
  bottom: "fc1"
  bottom: "label"
  top: "acc/top-5"
  include {
    phase: TEST
  }
  accuracy_param {
    top_k: 5
  }
}

deploy.prototxt

May you send your deploy.prototxt to me ? I want to test my model,but I encounter a problem with the file.

Input size

The paper specified using a 299x299 input, why do you use a 224x224 input, is there a performance difference compared to using 299x299?

Inception v3 implementation problem at test time

I have tried to implement and use your version of inception v3, (thanks alot for doing this work) I have attached my prototxt for the implementation of inception v3.
inceptionv3_prototxt.txt
I am om nvidias caffe-0.15 branch.

Training goes fine and i can see the accuracy on the validation set aswell. My problem occurs when i try test the model on a single image.

When i do this i get the following error:

insert_splits.cpp:29] Unknown bottom blob 'bn1' (layer 'conv1/3x3_s1_scale', bottom index 0)

Any ideas how to fix this issue, so i can use the model to classify new images.

Thanks alot.
Best regards,
William

BatchNormLayer problem?

Is the BN not complete? Because the prototxt file doesn't contain ScaleLayer after BatchNormLayer,

when I train Inception_v4 by caffe, it take a mistake.

I train your Inception_v3_GoogLeNet by caffe, it is successful. But I train Inception_v4 by caffe, it is wrong.

inception_v4_train_test.prototxt.txt

The mistake is
Check failed: top_shape[j] == bottom[i]->shape(j) (45 vs. 46) All input
s must have the same shape, except at concat_axis.
So,I want to konw how you modofy your Inception_v3_GoogLeNet through original Inception_v3.
thanks.

where deploy