What's the mean file of the pretrained models?
kli-casia opened this issue · comments
I used global pixel mean as used here too: https://github.com/BVLC/caffe/blob/master/models/bvlc_googlenet/train_val.prototxt#L10-L16
Thanks @jay-mahadeokar
I am using your pretrained resnet50 model to test on imagenet val dataset.
The prototxt I use is https://github.com/jay-mahadeokar/pynetbuilder/blob/master/models/imagenet/resnet_50/test.prototxt
I use my own data layer as follows
layer {
name: "data"
type: "Data"
top: "data"
top: "label"
transform_param {
crop_size: 224
mean_value: 104
mean_value: 117
mean_value: 123
}
data_param {
source: "/home/kli/extra2/ImageNetLMDB/ilsvrc12_val_lmdb"
batch_size: 25
backend: LMDB
}
}
The model weights is download from https://www.dropbox.com/s/k382wr2bzi59m4c/resnet_50.caffemodel?dl=0
But caffe shows some error message
Cannot copy param 0 weights from layer 'conv_stage0_block0_proj_shortcut'; shape mismatch. Source param shape is 256 64 1 1 (16384); target param shape is 128 64 1 1 (8192). To learn this layer's parameters from scratch rather than copying from a saved net, rename the layer.
The test command I use is
~/local/caffe-master/build/tools/caffe test -model test.prototxt -weights resnet_50.caffemodel -gpu 2 -iterations 2000
Looks like I have made some error in uploading the train.prototxt and test.prototxts ( in resnet_50 I might have copied resnet_50_1by2 by mistake) Give me some time to check and commit the correct ones.
Thank you very much @jay-mahadeokar
@kli-nlpr can you check once again? I uploaded the correct file and hopefully it should work. Haven't run it myself so let me know if there is still some bug.
Great, I works this time
I0910 10:13:48.260939 10409 caffe.cpp:308] Batch 1998, loss = 1.35442
I0910 10:13:48.406500 10409 caffe.cpp:308] Batch 1999, accuracy = 0.68
I0910 10:13:48.406550 10409 caffe.cpp:308] Batch 1999, loss = 1.40331
I0910 10:13:48.406558 10409 caffe.cpp:313] Loss: 1.25629
I0910 10:13:48.406581 10409 caffe.cpp:325] accuracy = 0.697541
I0910 10:13:48.406594 10409 caffe.cpp:325] loss = 1.25629 (* 1 = 1.25629 loss)
The overall accuracy is slightly lower than yours 0.697541 vs 0.7175
.
Good to know you are able to run it! About accuracy im not sure what would be the difference. Maybe the way test images are resized/cropped? I have resized the images to 256x256, and I believe at test time center crop of 224x224 is taken. I used caffe on spark for my experiments.
Thank you very much for your timely help @jay-mahadeokar