[Bug?] Error when input dimensions != 4D

[idata]

OK this not a serious bug, because there's a workaround. But the input dimensions HAVE to be 4D. For example, when I have:

input_shape {

dim: 1 dim: 1000}

I get the following error:

Traceback (most recent call last):

File "../../src/./mvNCCompile.py", line 99, in File "../../src/./mvNCCompile.py", line 81, in create_graph File "../../src/./Controllers/CaffeParser.py", line 879, in parse_caffeIndexError: tuple index out of range

Obviously this error is NOT really helpful ;)

But I quickly learned that this is the only type of input shape that works:

input_shape {

dim: 1 dim: 1 dim: 1 dim: 1000}

Maybe it is an idea to fix the error handling here, at least.

[idata]

@Bug80 Could you provide me with more details on the network you are using? Thanks.

[idata]

See the other thread (about using non-Softmax layers); unfortunately I cannot share the network with you but you could also download the dummy networks I provided there: https://www.dropbox.com/s/x63gsl9fz6pmgze/networks.zip?dl=0

Change any of the input dimensions to something other than 4D and you get errors!

[idata]

@Bug80 Thanks for reporting this and for your suggestion. mvNCCompile was designed with the conventional Caffe blob dimensions in mind (batch size, channel, height and width) so it expects 4 input dims. Per your discovery, if you are using a network that specifies less than 4 input dims, adding “dim: 1” as a filler will enable the network to be used with the NCS.

[idata]

@Tome_at_Intel OK thanks, I will use that as a workaround then. By the way, note that Caffe is very flexible when it comes to the number of input dimensions allowed. (batch size, channel, height, width) does make sense for image data but people may want to use the Movidius stick for other types of networks as well!