Implementation bug about unlabeled_matching_layer?
Closed this issue · 0 comments
serend1p1ty commented
https://github.com/Cysu/caffe/blob/aed38841ae4282d90575186b87c3287620913722/src/caffe/layers/unlabeled_matching_layer.cu#L48
caffe_gpu_memcpy
is based on bytes, so the code should be:
caffe_gpu_memcpy(K * sizeof(int), bottom_data + i * K, weight + queue_tail_ * K);
The shape of queue
that saves the features of unlabeled persons is (5000, 256)
But I find only the first 64 columns of the variable queue
have values.
Because of 64 = 256 / sizeof(int).