mvNCCompile taking too much time to compile a SSD mobilenet pb file using tf-ssd-config file

[idata]

It always keeps getting stuck after this,

mvNCCompile v02.00, Copyright @ Intel Corporation 2017

final_output_graph.pb

2019-03-12 09:04:43.544935: W tensorflow/core/framework/allocator.cc:113] Allocation of 8640000 exceeds 10% of system memory. 2019-03-12 09:04:45.343218: W tensorflow/core/framework/allocator.cc:113] Allocation of 8640000 exceeds 10% of system memory. 2019-03-12 09:04:47.008148: W tensorflow/core/framework/allocator.cc:113] Allocation of 8640000 exceeds 10% of system memory. 2019-03-12 09:04:49.956965: W tensorflow/core/framework/allocator.cc:113] Allocation of 8640000 exceeds 10% of system memory. 2019-03-12 09:04:51.681470: W tensorflow/core/framework/allocator.cc:113] Allocation of 8640000 exceeds 10% of system memory. terminate called after throwing an instance of 'std::bad_alloc' what(): std::bad_alloc Aborted

Download link for the model,

drive.google.com/file/d/17HdXplYI6OYINULMBuaYf_aH-ewt9B0c/view?usp=sharing

Its config file,

drive.google.com/open?id=1w0BEfq0hohJ5I3wxfqYJe9CPyng5cAZI

The command I'm using,

mvNCCompile -s 12 final_output_graph.pb --tf-ssd-config mymodel.config -o myoutput.graph

I've also tried with this,

mvNCCompile -s 6 final_output_graph.pb --tf-ssd-config mymodel.config -o myoutput.graph

@Tome_at_Intel @Sahira_at_Intel @tf_movidius

It would be great if you could help me out with this.

[idata]

Hi @shujath

Thank you for reaching out! I was able to compile the model you link with the configuration file you provided. I have seen a similar issue when the system does not have enough memory. How much memory does your system have? I am running a virtual machine with 8 GB of memory allocated.

The following command took about 1 minute to generate the graph.

mvNCCompile -s12 final_output_graph.pb --tf-ssd-config mymodel.config

Regards,

Jesus

[idata]

Hi @Jesus_at_Intel ,

Thank you for replying, I am using it on Raspberry pi3, with an 8GB SD Card, and Raspbian Stretch and ncsdk2.It is always giving me a bad alloc error, now its giving me a OOM Resource exhausted error.

[idata]

Hi @shujath

The Neural Compute SDK is can be installed in two modes, Full SDK mode and API-only mode. We recomend installing the Full SDK mode on a laptop/desktop to compile the networks and API-only on the Raspberry Pi. However, API-only mode will not allow you to profile, check or compile neural networks. Could you try to install the NCSDK on a laptop/desktop and compile the neural network into a graph and then run the application on the Raspberry Pi?

You can find additional information about NCS applications on the Raspberry Pi at the following link:

https://movidius.github.io/blog/ncs-apps-on-rpi/

Regards,

Jesus

[idata]

HI @Jesus_at_Intel

Thank you, I'll look into it. Is there any way to setup NCSDK on Windows 10 without using VirtualBox but by using Hyper-V or setting it up on Windows itself as the host platform?

[idata]

Hi @shujath

The Neural Compute SDK is not supported on Windows. However, there has been discussions by others that have the NCSDK working on windows. Take a look at the following thread:

https://ncsforum.movidius.com/discussion/comment/1548/#Comment_1548

The other option would be running a VM but that may require Hyper-V as you mentioned.

Regards,

Jesus

[idata]

Hi @Jesus_at_Intel

Okay, I've got the graph compiled on ubuntu, now I'm facing an issue with returned output result from the graph.

I'm getting an array of values, How do I determine the num_detections, detection_scores, detection_boxes from the array?I've tried using this utility wrapper for my ssd mobilenet graph, It keeps giving me 1.0 for its first index and the rest of the values are negative, exponential and float values.It would be great if you could help me out with this.Thank you.

Regards,

Shujath

[idata]

Also, @Jesus_at_Intel ,

I am getting the exact same plots when i'm plotting the array values for when there is object and when there is none.This is basically a graph of victordibia's handtracking model.