- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
I compiled a network and tried to check if it can be downloaded to the USB stick for execution using mvNCCheck.
The error I get is _[Error 25] Myriad Error: "Cannot Allocate space for convolution."_. Reduced the size of the network by half and more and still have problems downloading the network to the USB stick.
1) Is there a way to determine what is the size of the allowable network during design phase?
2) What are the limitations in terms of number of operations allowed on a device ?
3) Can a large network be partitioned into several usb sticks?
Start ========= deploy.prototxt ========
name: "aaa"
input: "data"
input_shape {
dim: 1
dim: 3
dim: 256
dim: 512
}
layer {
name: "Convolution1"
type: "Convolution"
bottom: "data"
top: "Convolution1"
convolution_param {
num_output: 16
pad: 1
kernel_size: 3
stride: 1
weight_filler {
type: "gaussian"
std: 0.00999999977648
}
}
}
layer {
name: "BatchNorm1"
type: "BatchNorm"
bottom: "Convolution1"
top: "Convolution1"
}
layer {
name: "Scale1"
type: "Scale"
bottom: "Convolution1"
top: "Convolution1"
scale_param {
bias_term: true
}
}
layer {
name: "block11"
type: "ReLU"
bottom: "Convolution1"
top: "Convolution1"
}
layer {
name: "Convolution2"
type: "Convolution"
bottom: "Convolution1"
top: "Convolution2"
convolution_param {
num_output: 16
pad: 1
kernel_size: 3
stride: 1
weight_filler {
type: "gaussian"
std: 0.00999999977648
}
}
}
layer {
name: "BatchNorm2"
type: "BatchNorm"
bottom: "Convolution2"
top: "Convolution2"
}
layer {
name: "Scale2"
type: "Scale"
bottom: "Convolution2"
top: "Convolution2"
scale_param {
bias_term: true
}
}
layer {
name: "block12"
type: "ReLU"
bottom: "Convolution2"
top: "Convolution2"
}
layer {
name: "pool13"
type: "Pooling"
bottom: "Convolution2"
top: "pool13"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "Convolution3"
type: "Convolution"
bottom: "pool13"
top: "Convolution3"
convolution_param {
num_output: 32
pad: 1
kernel_size: 3
stride: 1
weight_filler {
type: "gaussian"
std: 0.00999999977648
}
}
}
layer {
name: "BatchNorm3"
type: "BatchNorm"
bottom: "Convolution3"
top: "Convolution3"
}
layer {
name: "Scale3"
type: "Scale"
bottom: "Convolution3"
top: "Convolution3"
scale_param {
bias_term: true
}
}
layer {
name: "block21"
type: "ReLU"
bottom: "Convolution3"
top: "Convolution3"
}
layer {
name: "Convolution4"
type: "Convolution"
bottom: "Convolution3"
top: "Convolution4"
convolution_param {
num_output: 32
pad: 1
kernel_size: 3
stride: 1
weight_filler {
type: "gaussian"
std: 0.00999999977648
}
}
}
layer {
name: "BatchNorm4"
type: "BatchNorm"
bottom: "Convolution4"
top: "Convolution4"
}
layer {
name: "Scale4"
type: "Scale"
bottom: "Convolution4"
top: "Convolution4"
scale_param {
bias_term: true
}
}
layer {
name: "block22"
type: "ReLU"
bottom: "Convolution4"
top: "Convolution4"
}
layer {
name: "pool23"
type: "Pooling"
bottom: "Convolution4"
top: "pool23"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "Convolution5"
type: "Convolution"
bottom: "pool23"
top: "Convolution5"
convolution_param {
num_output: 64
pad: 1
kernel_size: 3
stride: 1
weight_filler {
type: "gaussian"
std: 0.00999999977648
}
}
}
layer {
name: "BatchNorm5"
type: "BatchNorm"
bottom: "Convolution5"
top: "Convolution5"
}
layer {
name: "Scale5"
type: "Scale"
bottom: "Convolution5"
top: "Convolution5"
scale_param {
bias_term: true
}
}
layer {
name: "block31"
type: "ReLU"
bottom: "Convolution5"
top: "Convolution5"
}
layer {
name: "Convolution6"
type: "Convolution"
bottom: "Convolution5"
top: "Convolution6"
convolution_param {
num_output: 64
pad: 1
kernel_size: 3
stride: 1
weight_filler {
type: "gaussian"
std: 0.00999999977648
}
}
}
layer {
name: "BatchNorm6"
type: "BatchNorm"
bottom: "Convolution6"
top: "Convolution6"
}
layer {
name: "Scale6"
type: "Scale"
bottom: "Convolution6"
top: "Convolution6"
scale_param {
bias_term: true
}
}
layer {
name: "block32"
type: "ReLU"
bottom: "Convolution6"
top: "Convolution6"
}
layer {
name: "pool34"
type: "Pooling"
bottom: "Convolution6"
top: "pool34"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "Convolution7"
type: "Convolution"
bottom: "pool34"
top: "Convolution7"
convolution_param {
num_output: 128
pad: 1
kernel_size: 3
stride: 1
weight_filler {
type: "gaussian"
std: 0.00999999977648
}
}
}
layer {
name: "BatchNorm7"
type: "BatchNorm"
bottom: "Convolution7"
top: "Convolution7"
}
layer {
name: "Scale7"
type: "Scale"
bottom: "Convolution7"
top: "Convolution7"
scale_param {
bias_term: true
}
}
layer {
name: "block41"
type: "ReLU"
bottom: "Convolution7"
top: "Convolution7"
}
layer {
name: "Convolution8"
type: "Convolution"
bottom: "Convolution7"
top: "Convolution8"
convolution_param {
num_output: 128
pad: 1
kernel_size: 3
stride: 1
weight_filler {
type: "gaussian"
std: 0.00999999977648
}
}
}
layer {
name: "BatchNorm8"
type: "BatchNorm"
bottom: "Convolution8"
top: "Convolution8"
}
layer {
name: "Scale8"
type: "Scale"
bottom: "Convolution8"
top: "Convolution8"
scale_param {
bias_term: true
}
}
layer {
name: "block42"
type: "ReLU"
bottom: "Convolution8"
top: "Convolution8"
}
layer {
name: "pool44"
type: "Pooling"
bottom: "Convolution8"
top: "pool44"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "Convolution9"
type: "Convolution"
bottom: "pool44"
top: "Convolution9"
convolution_param {
num_output: 128
pad: 1
kernel_size: 3
stride: 1
weight_filler {
type: "gaussian"
std: 0.00999999977648
}
}
}
layer {
name: "BatchNorm9"
type: "BatchNorm"
bottom: "Convolution9"
top: "Convolution9"
}
layer {
name: "Scale9"
type: "Scale"
bottom: "Convolution9"
top: "Convolution9"
scale_param {
bias_term: true
}
}
layer {
name: "block51"
type: "ReLU"
bottom: "Convolution9"
top: "Convolution9"
}
layer {
name: "block52"
type: "Convolution"
bottom: "Convolution9"
top: "block52"
convolution_param {
num_output: 2
pad: 0
kernel_size: 1
stride: 1
weight_filler {
type: "gaussian"
std: 0.00999999977648
}
}
}
layer {
name: "block53"
type: "Deconvolution"
bottom: "block52"
top: "block53"
convolution_param {
num_output: 2
bias_term: false
pad: 8
kernel_size: 32
stride: 16
weight_filler {
type: "bilinear"
}
}
}
layer {
name: "prob"
type: "Softmax"
bottom: "block53"
top: "prob"
}
End ========= deploy.prototxt =========
Thanks,
Yair
Link Copied
6 Replies
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@yairh We looked into this issue and we are currently testing a release candidate at the moment. I'll ping you and let you know as soon as we have something for you. Thanks.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@Tome,
Any update on this issue ?
Thanks,
Yair
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
As far as I can tell on this one, it isn't the size of the network that is the problem (that seems to be determined at compile time), it is the size of the individual convolutions. I fixed this in my case by cutting down the number of channels in my convolutions.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
The number of channels affect the amount of convolution units, which translates to network size. Can you please share the number of channels you used in each layer.
Thanks,
Yair
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Found out that is I reduce the image resolution to 128x256 the network is runnable.
Is there a limitation on the image size input / resolution? If yes then what is the limitation.
Thanks,
Yair
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@yairh The convolutions are just too big after the convolution5 layer. As a workaround, you can add a batch norm params eps value to each BatchNorm layer like so:
layer {
name: "BatchNorm1"
type: "BatchNorm"
bottom: "Convolution1"
top: "Convolution1"
batch_norm_param {
eps: 1e-2
}
}
Using this w/a, you can compile up to the block 53 layer, but the softmax layer will have to be done on the CPU.
Reply
Topic Options
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page