Community
cancel
Showing results for 
Search instead for 
Did you mean: 
idata
Community Manager
286 Views

incorrect shape/size set for 2d output tensor

i have a unet style model that includes an 2d output tensor. my minimal example case is the following…

 

imgs (1, 64, 64, 3) #12288 e1 (1, 31, 31, 8) #7688 conv, stride=2 e2 (1, 15, 15, 16) #3600 conv, stride=2 d1 (1, 31, 31, 8) #7688 deconv, stride=2 output (1, 31, 31, 1) #961 1x1 conv, stride=1

 

this model works with one small quirk; when i call queue_inference_with_fifo_elem i get an output tensor of shape (29791,) whereas i expect a shape of (31,31,1)

 

it turns out though the i can slice off the first 31*31 and reshape and it's ok…

 

ncs_positive_prediction = ncs_positive_prediction[:31*31].reshape((31,31,1))

 

…so the data is _there_, but the shape/size is just not being set correctly.

 

this model, with full reproduction of this case, is described in my movidius_bug_reports repo on github with additional notes in this github issue

Tags (1)
0 Kudos
2 Replies
idata
Community Manager
90 Views

@matpalm Thanks for reporting this. We'll take a look at this and see why this is happening. Have you tried using mvNCCheck with your model to see if the shape from TensorFlow matches the expected shape (31, 31, 1)?

idata
Community Manager
90 Views

though haven't used mvNCCheck, i do have a parallel run-inference-on-frozen-graph using just tensorflow code and that is all fine.

Reply