split op questions for model otimizer

[kaixuanliu]

Hi, I ran across a problem when I use mo to convert a caffemodel, it shows 'LayerParameter' object has no attribute 'num_split' for split layer. However, as far as I am concerned, split layer in caffe is just a copy operation. So I think this sentence _out_shape[node.axis] = np.int64(input_node.shape[node.axis] / _node.pb.num_split) in mo/ops/split.py should be like this:
out_shape[node.axis] = np.int64(input_node.shape[node.axis])
You may mistake split layer as slice layer, I guess.

[meshaun9]

Hi @kaixuanliu, It seems that you may be right based of what's currently implemented and what's in caffe's documentation for the split layer. I will do a test with the split layer coded as a copy instead of needing num_split value and check the results.

[meshaun9]

Hi @kaixuanliu, you are correct. A bug has been created for this and we plan on fixing this.

[lazarevevgeny]

Hi @kaixuanliu. It is not enough just to change the infer function for the Split layer. In fact, if you do this you break conversion of the models from other frameworks (TF, MXNet, Kaldi, etc). The right solution would be to eliminate Split layers during conversion of the Caffe models. Create a file "model_optimizer/extensions/front/caffe/remove_split.py" with the following content:

from mo.front.common.replacement import FrontReplacementOp

class RemoveSplit(FrontReplacementOp):
    op = "Split"
    enabled = True
    def replace_sub_graph(self, graph, match):
        split = match['op']
        split.op = 'Identity'
        for _, __, edge_attrs in graph.out_edges(split.id, data=True):
            edge_attrs['out'] = 0

[kaixuanliu]

Hi,@lazarevevgeny I agree with you. Sorry I am not familiar with TF or MXNet, and I have this concern before as you.