supported layers

[idata]

Hi,

I'm testing few models with NCS at the moment. I found out the layers supported by mvNCCompile is quite limited.Can you provide what full list of layers supported by your product in each framework (caffe/tensorflow)?I wasn't able to find it in the dev documents.Also, I see 'concat' layer is not supported atm, Is there any time-plan when it is going to be updated?Thanks

[idata]

Supported layers are listed in the latest release notes: https://github.com/movidius/ncsdk/releases/tag/v1.10.01.01

Currently includes (according to the release notes)

1. Convolutions
   • NxN Convolution with Stride S.
   • The following cases have been extensively tested: 1x1s1,3x3s1,5x5s1,7x7s1, 7x7s2, 7x7s4
   • Group convolution
   • Depth Convolution
   • Dilated convolution
2. Max Pooling Radix NxM with Stride S
3. Average Pooling: Radix NxM with Stride S, Global average pooling
4. Local Response Normalization
5. Relu, Relu-X, Prelu (see erattum #10)
6. Softmax
7. Sigmoid
8. Tanh (see erratum #10)
9. Deconvolution
10. Slice
11. Scale
12. ElmWise unit : supported operations - sum, prod, max
13. Fully Connected Layers (limited support -- see erratum #8)
14. Reshape
15. Flatten
16. Power
17. Crop
18. ELU
19. Batch Normalization

[idata]

@tf_movidius Thank you very much!

[idata]

Hi,

My net uses a tf.slice operator on a tensor and I get [Error 5] Toolkit Error: Stage Details Not Supported: Slice type not supported while running mvNCCompile on it.

Could you help me with this?

[idata]

@gokul_uf Yes. We're working to add as much support as we can for all of the different neural networks out there, so if you can provide a link to your network, we can help with debugging issues.

[idata]

@Tome_at_Intel The net is at model.py here https://github.com/SpaceML/GalaxyGAN_python/

I had modified the net to have both image and cond passed in as single tensor and slice it inside to get around the only one input limitation. That's where I got the Slice issue.

I later realized that I needed only cond so I reverted back to the original code and I get the following error.

[Error 5] Toolkit Error: Stage Details Not Supported: Unsupported Mean operation

It'd be great if you could help me with this, because I have a demo coming up :smile:

[idata]

@gokul_uf Since I cannot find the network's meta file from the link., it looks like I am required to download their data and train to obtain the network. It would save me a lot of time if you could provide me the model's meta file.

[idata]

@Tome_at_Intel Apologies, my bad! The meta files are at https://drive.google.com/drive/folders/1yoqOVZIAfm3UpVAqWq3Ps-6cmpgd_1Cr?usp=sharing

the non_bnorm one has the bnorm op removed (because we use just 1 image per training so no point having bnorm)

the non_unet one has the skip links removed in the generator method (this uses concat and I saw that concat is buggy in ncsdk)

the input node name is cond and output is gen/gen_op.

Thanks!

[idata]

@gokul_uf I need the checkpoint file also if you could provide it. I am getting an error saying : Failed to find any matching files for non_bnorm_model_1.ckpt. Thanks!

[idata]

Hi, are you planning to release support for 3dconvolutions/ 3d Max pooling any time soon?

Thanks!

[idata]

@gokul_uf Can you try your network with the latest SDK (1.11) and let me know if this fixes your issue?

@JMat I can't give a specific timeline for you regarding the support for 3d convolutions, but as of now, it is something we are considering and looking into. Thanks.

[idata]

@gokul_uf I had issues too, according to /opt/movidius/NCSDK/ncsdk-x86_64/tk/Controllers/TensorFlowParser.py, line 1489:

if (len(input_shape) != 4 or len(slicingbegin) != 4 or len(slicingsize) != 4 or

slicingbegin[0] != 0 or slicingbegin[1] != 0 or slicingbegin[2] != 0 or input_shape[0] != slicingsize[0] or input_shape[1] != slicingsize[1] or input_shape[2] != slicingsize[2]):

You can slice only on 4th dimension (image channels). Moreover, you can't use "-1" indexing you can use on base Tensorflow.

[idata]

Is GRU supported?

[idata]

@sreerag_ibtl,

I don't think GRU is supported as currently the NCSDK does not have support for RCNNs, hope this helps.Regards,@Luis_at_Intel