Community
cancel
Showing results for 
Search instead for 
Did you mean: 
Highlighted
Community Manager
294 Views

Reshape layer followed by concat gives wrong result

Hello,

 

I found that following part of the network does not work well, until I insert fake pooling layer after reshape.

 

layer { name: "conv_a" type: "Convolution" bottom: "data" top: "conv_a" convolution_param { num_output: 256 bias_term: true kernel_size: 1 stride: 1 } } layer { name: "relu_a" type: "ReLU" bottom: "conv_a" top: "relu_a" } layer { name: "conv_b" type: "Convolution" bottom: "data" top: "conv_b" convolution_param { num_output: 64 bias_term: true kernel_size: 1 stride: 1 } } layer { name: "relu_b" type: "ReLU" bottom: "conv_b" top: "relu_b" } layer { name: "reshape" type: "Reshape" bottom: "relu_b" top: "reshape" reshape_param { shape { dim: -1 dim: 256 dim: 15 dim: 15 } } } layer { name: "pool_fake" type: "Pooling" bottom: "reshape" top: "pool_fake" pooling_param { pool: MAX kernel_size: 1 stride: 1 } } layer { name: "concat_a_b" type: "Concat" bottom: "pool_fake" bottom: "relu_b" top: "concat_a_b" }

 

thanks

 

Jan
0 Kudos
5 Replies
Highlighted
Community Manager
14 Views

@xhpohanka Interesting. Can you provide a link to your network?

0 Kudos
Highlighted
Community Manager
14 Views

@Tome_at_Intel sorry for delay - here is full prototxt https://pastebin.com/Zrdqm6wZ

 

It is slightly modified yolo detector, pool_xx is the fake layer there, without that I get wrong results
0 Kudos
Highlighted
Community Manager
14 Views

@xhpohanka Which weights are you using with this prototxt file?

0 Kudos
Highlighted
Community Manager
14 Views

0 Kudos
Highlighted
Community Manager
14 Views

@xhpohanka Thanks for reporting this. I was able to reproduce your issue and also confirm that your "fake" layer helped to produce successful results.

0 Kudos