search

pertusa / InceptionBN-21K-for-Caffe

Inception-BN ImageNet (21K classes) model for Caffe
74 stars 47 forks source link

Some concat layers don't work #1

Closed crowsonkb closed 8 years ago

crowsonkb commented 8 years ago

An example error:

F1015 21:27:53.828809 3034956736 concat_layer.cpp:42] Check failed: top_shape[j] == bottom[i]->shape(j) (9 vs. 8) All inputs must have the same shape, except at concat_axis.
crowsonkb commented 8 years ago

Specifically anything from ch_concat_3c_chconcat on doesn't work for me.

crowsonkb commented 8 years ago

Also, the version of deploy.prototxt from before commit https://github.com/pertusa/InceptionBN-21K-for-Caffe/commit/cec7e092feef6e3aaf732c6fc30f4f70491747f5 doesn't work for me either.

pertusa commented 8 years ago

The initial commit of the model at http://www.dlsi.ua.es/~pertusa/deep/Inception21k.caffemodel had this issue, please try to download the model again (fixed on Sep. 16 in commit be8d78c). If the problem persists, it may be due to wrong image dimensions at the input layer. The input image should be resized to 224x224. Unlike other networks, the Inception21k architecture doesn't allow to use variable image sizes.