It looks like you're new here. If you want to get involved, click one of these buttons!Sign In
It looks like you're new here. If you want to get involved, click one of these buttons!
Scale layer in Caffe can be used along with batch normalization to implement scale and shift operation, or it can be used to multiply 2 bottom blobs together. For example, in this paper it is used simply to multiple bottom blobs together and this layer doesn't have any weights and bias. I assume this possible utilization of
Scale layer was missed during writing CaffeParser.py.
Overall the bug is the following. If you take a squeeze-and-excitation caffe model (or any other model with
Scale layer as multiplier of bottom layers) and run
mvNCProfile deploy.prototxt -is 224 224 the code will fail at line 237
return blobs[layer.name].data.astype(dtype=data_type), None in CaffeParser.py since this layer neither has weights nor biases.