Looks like the Caffe layer "Silence" is not supported... is this correct?

idata · ‎11-03-2017

Hey Guys,

I am using a deploy.prototxt definition that has a "Silence" layer… as a result, I am seeing the following error with the example environment =>

mdouglas_95136:~/workspace/ncsdk/examples/caffe/GoogLeNet$ make run

making prereqs

(cd ../../data/ilsvrc12; make)

make[1]: Entering directory '/home/user/workspace/ncsdk/examples/data/ilsvrc12'

make[1]: Leaving directory '/home/user/workspace/ncsdk/examples/data/ilsvrc12'

making prototxt

Prototxt file already exists

making caffemodel

caffemodel file already exists

making compile

mvNCCompile -w deploy.caffemodel -s 12 deploy.prototxt

mvNCCompile v02.00, Copyright @ Movidius Ltd 2016

[Warning: 37] Output layer's name (instance_ignore_layer) must match its top (category_score)

[Error 4] Toolkit Error: Stage Type Not Supported: Silence

Anyone know if:

a) this is a legitimate error and "Silence" layer is not supported?

and

b) when is the support coming?

c) if it's in Beta, can I be a Beta tester? :-)

Thanks a bunch!

Mark Douglas

idata · ‎11-04-2017

@mdouglas_95136 Hi, in order to help us debug your issue, can you provide a link to your prototxt file? Thanks

idata · ‎11-05-2017

Tome! Thanks for asking.

I would need your email address for my SharePoint box.

my email is mark.douglas@nxp.com

idata · ‎11-06-2017

@ Tome_at_Intel

so my deploy.prototxt file is just too long and I get errors trying to place it in these postings.

mark.douglas@nxp.com

Thanks,

Mark D.

idata · ‎11-06-2017

@mdouglas_95136 When typing a comment, there is an option to link a URL or add a file attachment in the forum toolbar above the comment box.

idata · ‎11-06-2017

ok thanks @Tome_at_Intel

I need to open another discussion as if you look at the comment here there is no option to add a file… but I can from a "new" discussion.

Thanks,

Mark

idata · ‎11-06-2017

Getting "file not allowed" Argh…

when I try and add a file by either dropping in or by browsing. I can only see this option from a "new discussion" and even then it don't work. I tried under PaleMoon, FireFox, Chrome and IE browsers… no different.

So this is working for you guys?

MD

idata · ‎11-06-2017

here is a "try" to add the prototxt file that's stumbling on the "Silence" layer

idata · ‎11-06-2017

[file:///e:/Movidius/deploy_not_compile.prototxt](file:///e:/Movidius/deploy_not_compile.prototxt "file:///e:/Movidius/deploy_not_compile.prototxt")

idata · ‎11-06-2017

sorry… for me this feature is just not usable :-(

idata · ‎11-06-2017

@mdouglas_95136, alternatively you can copy paste the entire prototxt file within the Code formatting template (between ` symbol - it'll be on the same key as ~ tilde symbol). Like so:

input: "data" input_shape { dim: 1 dim: 3 dim: 227 dim: 227 }

idata · ‎11-06-2017

Thanks @AshwinVijayakumar ; however, same error about content length:

"Body is 3093 characters too long."

it's just too long or too many characters (multi-layered net config)

I may try in pieces

MD

idata · ‎11-06-2017

layer {

type: "Input"

top: "data"

input_param {

shape {

dim: 1

dim: 8

dim: 512

}

layer {

type: "Slice"

bottom: "data"

top: "dump_blob"

top: "mask"

slice_param {

slice_point: 7

axis: 1

}

layer {

type: "Silence"

bottom: "dump_blob"

}

layer {

type: "Convolution"

bottom: "data"

top: "conv1_1"

param {

lr_mult: 1.0

decay_mult: 0.0

}

param {

lr_mult: 1.0

decay_mult: 0.0

}

convolution_param {

num_output: 48

pad: 1

kernel_size: 3

stride: 2

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

value: 0.0

}

layer {

type: "ReLU"

bottom: "conv1_1"

top: "conv1_1"

relu_param {

negative_slope: 0.0

}

layer {

type: "Convolution"

bottom: "conv1_1"

top: "conv1_2"

param {

lr_mult: 1.0

decay_mult: 0.0

}

param {

lr_mult: 1.0

decay_mult: 0.0

}

convolution_param {

num_output: 48

pad: 1

kernel_size: 3

stride: 1

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

value: 0.0

}

layer {

type: "ReLU"

bottom: "conv1_2"

top: "conv1_2"

relu_param {

negative_slope: 0.0

}

layer {

type: "Convolution"

bottom: "conv1_2"

top: "conv2_1"

param {

lr_mult: 1.0

decay_mult: 0.0

}

param {

lr_mult: 1.0

decay_mult: 0.0

}

convolution_param {

num_output: 64

pad: 1

kernel_size: 3

stride: 2

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

value: 0.0

}

layer {

idata · ‎11-06-2017

type: "ReLU"

bottom: "conv2_1"

top: "conv2_1"

relu_param {

negative_slope: 0.0

}

layer {

type: "Convolution"

bottom: "conv2_1"

top: "conv2_2"

param {

lr_mult: 1.0

decay_mult: 0.0

}

param {

lr_mult: 1.0

decay_mult: 0.0

}

convolution_param {

num_output: 64

pad: 1

kernel_size: 3

stride: 1

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

value: 0.0

}

layer {

type: "ReLU"

bottom: "conv2_2"

top: "conv2_2"

relu_param {

negative_slope: 0.0

}

layer {

type: "Convolution"

bottom: "conv2_2"

top: "conv3_1"

param {

lr_mult: 1.0

decay_mult: 0.0

}

param {

lr_mult: 1.0

decay_mult: 0.0

}

convolution_param {

num_output: 96

pad: 1

kernel_size: 3

stride: 2

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

value: 0.0

}

layer {

type: "ReLU"

bottom: "conv3_1"

top: "conv3_1"

relu_param {

negative_slope: 0.0

}

layer {

type: "Convolution"

bottom: "conv3_1"

top: "conv4_1"

param {

lr_mult: 1.0

decay_mult: 0.0

}

param {

lr_mult: 1.0

decay_mult: 0.0

}

convolution_param {

num_output: 128

pad: 1

kernel_size: 3

stride: 2

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

value: 0.0

}

layer {

type: "ReLU"

bottom: "conv4_1"

top: "conv4_1"

relu_param {

negative_slope: 0.0

}

layer {

type: "Convolution"

bottom: "conv4_1"

top: "conv5_1"

param {

lr_mult: 1.0

decay_mult: 0.0

}

param {

lr_mult: 1.0

decay_mult: 0.0

}

convolution_param {

num_output: 192

pad: 1

kernel_size: 3

stride: 2

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

value: 0.0

}

layer {

type: "ReLU"

bottom: "conv5_1"

top: "conv5_1"

relu_param {

negative_slope: 0.0

}

layer {

type: "Deconvolution"

bottom: "conv5_1"

top: "refine5_deconv"

param {

lr_mult: 1.0

decay_mult: 0.0

}

param {

lr_mult: 1.0

decay_mult: 0.0

}

convolution_param {

num_output: 128

pad: 1

kernel_size: 4

group: 1

stride: 2

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

}

layer {

type: "ReLU"

bottom: "refine5_deconv"

top: "refine5_deconv"

relu_param {

negative_slope: 0.0

}

idata · ‎11-06-2017

layer {

type: "Concat"

bottom: "conv4_1"

bottom: "refine5_deconv"

top: "refine4_concat"

}

layer {

type: "Convolution"

bottom: "refine4_concat"

top: "refine4_conv_3x3_2"

param {

lr_mult: 1.0

decay_mult: 0.0

}

param {

lr_mult: 1.0

decay_mult: 0.0

}

convolution_param {

num_output: 128

pad: 1

kernel_size: 3

stride: 1

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

value: 0.0

}

layer {

type: "ReLU"

bottom: "refine4_conv_3x3_2"

top: "refine4_conv_3x3_2"

relu_param {

negative_slope: 0.0

}

layer {

type: "Deconvolution"

bottom: "refine4_conv_3x3_2"

top: "refine4_deconv"

param {

lr_mult: 1.0

decay_mult: 0.0

}

param {

lr_mult: 1.0

decay_mult: 0.0

}

convolution_param {

num_output: 96

pad: 1

kernel_size: 4

group: 1

stride: 2

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

}

layer {

type: "ReLU"

bottom: "refine4_deconv"

top: "refine4_deconv"

relu_param {

negative_slope: 0.0

}

layer {

type: "Concat"

bottom: "conv3_1"

bottom: "refine4_deconv"

top: "refine3_concat"

}

layer {

type: "Convolution"

bottom: "refine3_concat"

top: "refine3_conv_3x3_2"

param {

lr_mult: 1.0

decay_mult: 0.0

}

param {

lr_mult: 1.0

decay_mult: 0.0

}

convolution_param {

num_output: 96

pad: 1

kernel_size: 3

stride: 1

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

value: 0.0

}

layer {

type: "ReLU"

bottom: "refine3_conv_3x3_2"

top: "refine3_conv_3x3_2"

relu_param {

negative_slope: 0.0

}

layer {

type: "Deconvolution"

bottom: "refine3_conv_3x3_2"

top: "refine3_deconv"

param {

lr_mult: 1.0

decay_mult: 0.0

}

param {

lr_mult: 1.0

decay_mult: 0.0

}

convolution_param {

num_output: 64

pad: 1

kernel_size: 4

group: 1

stride: 2

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

}

layer {

type: "ReLU"

bottom: "refine3_deconv"

top: "refine3_deconv"

relu_param {

negative_slope: 0.0

}

layer {

type: "Concat"

bottom: "conv2_2"

bottom: "refine3_deconv"

top: "refine2_concat"

}

layer {

type: "Convolution"

bottom: "refine2_concat"

top: "refine2_conv_3x3_2"

param {

lr_mult: 1.0

decay_mult: 0.0

}

param {

lr_mult: 1.0

decay_mult: 0.0

}

convolution_param {

num_output: 64

pad: 1

kernel_size: 3

stride: 1

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

value: 0.0

}

layer {

type: "ReLU"

bottom: "refine2_conv_3x3_2"

top: "refine2_conv_3x3_2"

relu_param {

negative_slope: 0.0

}

layer {

type: "Deconvolution"

bottom: "refine2_conv_3x3_2"

top: "refine2_deconv"

param {

lr_mult: 1.0

decay_mult: 0.0

}

param {

lr_mult: 1.0

decay_mult: 0.0

}

convolution_param {

num_output: 48

pad: 1

kernel_size: 4

group: 1

stride: 2

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

}

layer {

type: "ReLU"

bottom: "refine2_deconv"

top: "refine2_deconv"

relu_param {

negative_slope: 0.0

}

layer {

type: "Concat"

bottom: "conv1_2"

bottom: "refine2_deconv"

top: "refine1_concat"

}

layer {

type: "Convolution"

bottom: "refine1_concat"

top: "refine1_conv_3x3_2"

param {

lr_mult: 1.0

decay_mult: 0.0

}

param {

lr_mult: 1.0

decay_mult: 0.0

}

convolution_param {

num_output: 48

pad: 1

kernel_size: 3

stride: 1

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

value: 0.0

}

layer {

type: "ReLU"

bottom: "refine1_conv_3x3_2"

top: "refine1_conv_3x3_2"

relu_param {

negative_slope: 0.0

}

layer {

type: "Deconvolution"

bottom: "refine1_conv_3x3_2"

top: "predict"

param {

lr_mult: 1.0

decay_mult: 0.0

}

param {

lr_mult: 1.0

decay_mult: 0.0

}

convolution_param {

num_output: 12

pad: 1

kernel_size: 4

group: 1

stride: 2

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

}

idata · ‎11-06-2017

layer {

name: "Slice_[\'category_pt\', \'instance_pt\', \'confidence_pt\', \'classify_pt\', \'heading_pt\', \'height_pt\']"

type: "Slice"

bottom: "predict"

top: "category_pt"

top: "instance_pt"

top: "confidence_pt"

top: "classify_pt"

top: "heading_pt"

top: "height_pt"

slice_param {

slice_point: 1

slice_point: 3

slice_point: 4

slice_point: 9

slice_point: 11

}

layer {

type: "Sigmoid"

bottom: "category_pt"

top: "all_category_score"

propagate_down: false

}

layer {

type: "Eltwise"

bottom: "all_category_score"

bottom: "mask"

top: "category_score"

eltwise_param {

operation: PROD

}

layer {

type: "Sigmoid"

bottom: "confidence_pt"

top: "confidence_score"

propagate_down: false

}

layer {

type: "Sigmoid"

bottom: "classify_pt"

top: "class_score"

propagate_down: false

}

idata · ‎11-06-2017

well… there it is boys… I like four(4) pieces.

--- I'm not sure this form tool is adequate in it's current capacity to accommodate files this size.

So you'd need to join each of these pieces and this is the file I can't see to get to build (with the error above)

Thanks!

Mark

idata · ‎11-06-2017

@mdouglas_95136 Do you have a specific purpose for the Silence layer? The Silence layer is useful in testing and training, however I'm not sure if it was meant for use with the deployment of a network.

idata · ‎11-06-2017

Thanks @Tome_at_Intel

I have asked about this layer… I've asked the owners if they could remove it. No response just yet.

--- do you know if the mvNCCompile supports "Silence" anyway? why is it barfing?

Thanks,

Mark

idata · ‎11-08-2017

@Tome_at_Intel Just checking in here :)

===> were you able to re-assemble this deploy.prototxt file (sorry I had to send in 4 pieces)

and give it a try with mvNCCompile?

Thanks so much for the help. If the SDK supports "Silence" then this should work.

well, keep me posted.

Mark Douglas

idata · ‎11-08-2017

@Tome_at_Intel just FYI: there are some known issues with your Caffe implementation listed sort of in your release notes (or this page)

https://movidius.github.io/ncsdk/Caffe.html

if you look down at

"Known Issues

Caffe Input Layer"

"Limitation: We don’t support this “input_param” format for the input layer"

so if I fix this to match the GoogLeNet same one you guys use ->

input: "data"

input_shape {

dim: 1

dim: 8

dim: 512

}

and… I remove the "Silence" layer I get different errors ->

____ start of make compile error ______*

user@user-PowerEdge-R720:~/workspace/ncsdk/examples/caffe/Apollo_GoogLeNet$ make compile

making prereqs

(cd ../../data/ilsvrc12; make)

make[1]: Entering directory '/home/user/workspace/ncsdk/examples/data/ilsvrc12'

make[1]: Leaving directory '/home/user/workspace/ncsdk/examples/data/ilsvrc12'

making prototxt

Prototxt file already exists

making caffemodel

caffemodel file already exists

making compile

mvNCCompile -w deploy.caffemodel -s 12 deploy.prototxt

[Warning: 37] Output layer's name (instance_ignore_layer) must match its top (category_score)

[Error 17] Toolkit Error: Internal Error: Could not build graph. Missing link: data

user@user-PowerEdge-R720:~/workspace/ncsdk/examples/caffe/Apollo_GoogLeNet$

______** end of "make compile" error ______***

So it looks like there is sufficient Caffe Input issues where you should have a guide to what might be the acceptable deploy.prototxt file.

please advise based on my last delopy.prototxt located here ->

https://github.com/ApolloAuto/apollo/blob/master/modules/perception/model/cnn_segmentation/deploy.prototxt

this happens to be the default Apollo deploy.prototxt file.

please advise.

Thanks,

Mark