Does nvdla support object detection

[Adithyak1998]

I trained a model on DIGITS for object detection. I followed this website: https://github.com/NVIDIA/DIGITS/tree/master/examples/object-detection I then downloaded the model which included deploy.prototxt and a .caffemodel file.

I tried to use nvdla to compile but encountered the following error: Command: ./nvdla_compiler --prototxt obj_detect.prototxt --caffemodel obj_detect.caffemodel Output:

creating new wisdom context...
opening wisdom context...
parsing caffe network...
libnvdla<1> error: parsing layer type Power index 0
(DLA) Error 0x00000004: Unable to parse caffemodel: "obj_detect.prototxt" (in ParseTest.cpp, function parseCaffeNetwork(), line 91)
(DLA) Error 0x00000004: (propagating from ParseTest.cpp, function parseAndCompile(), line 164)
(DLA) Error 0x00000004: (propagating from main.cpp, function launchTest(), line 87)

I did see issues with similar errors for mobilenet (which is put under "future release"). This it the same case with object detection?

[prasshantg]

@Adithyak1998 power layer is not yet supported. Adding this feature to future release of compiler

[Adithyak1998]

May I know the ETA for this

[prasshantg]

We will try to get it released in next compiler release coming by end of this week or early next week.

[Adithyak1998]

Ok, but NVDLA verilog HW IP supports object detection right?

[prasshantg]

Please share exact prototxt you are using to confirm the layers used.

[Adithyak1998]

These ars the prototxt files I used: https://raw.githubusercontent.com/NVIDIA/caffe/caffe-0.15/examples/kitti/detectnet_network.prototxt https://raw.githubusercontent.com/NVIDIA/caffe/caffe-0.15/examples/kitti/detectnet_network-2classes.prototxt

Both of them have a power layer in them. I just wanted to clarify if the NVDLA hardware supports that layer

[prasshantg]

In fact, layer for which you are getting error is not for deploy phase but just for training or validation phase. Currently we are supporting layers for deploy phase only. Prototxt that you are using is all-in-one and looks like parser is not detecting that this layer is not for deploy phase.

Checking more how it can be resolved. You can try using prototxt for deploy phase instead of all-in-one.

[Adithyak1998]

I'm sorry but I don't understand what you mean by prototxt for deploy phase Where can I find one?

[prasshantg]

Refer https://github.com/BVLC/caffe/wiki/Using-a-Trained-Network:-Deploy to understand what is prototxt for deploy

[prasshantg]

Confirmed no issue in parser and this layer is marked for deploy phase, will plan support for it in next release

layer { name: "deploy_transform" type: "Power" bottom: "data" top: "transformed_data" power_param { shift: -127 } include: { phase: TEST not_stage: "val" } }

[Adithyak1998]

Is power layer supported in the parser now?

[DrVijayK]

An update on ETA for Power layer support in compiler will be highly appreciated.

[raghavgurbaxani]

Yes,

I was able to run the RCNN caffemodel without ROI Pooling :)

Worked fine for me.

[zaddan]

@raghavgurbaxani where did you get the RCNN caffemodel from?

[raghavgurbaxani]

@zaddan I found it on the Caffe model zoo. Hope this helps

[zaddan]

@raghavgurbaxani I looked inside caffe model zoo, and the only RCNN that I found was faster-rcnn and if I am not mistaken their models have ROI Pooling in them. Could you help me find the exact model you tried?

[Adithyak1998]

@raghavgurbaxani Do you mean to say that we have to locate ROI before giving input to NVDLA?

If not, what is the alternative for ROI Pooling layer which NVDLA supports?

[raghavgurbaxani]

@zaddan @Adithyak1998

I used this model, worked fine for me

https://github.com/BVLC/caffe/tree/master/models/bvlc_reference_rcnn_ilsvrc13

[arvindhbti]

@zaddan @Adithyak1998

I used this model, worked fine for me

https://github.com/BVLC/caffe/tree/master/models/bvlc_reference_rcnn_ilsvrc13

Yeah it is working fine for me also.

@raghavgurbaxani can you tell me that how to visualize the output because it is giving like-

-1.89844 -2.11719 -1.95312 -1.87207 -2.08789 -2.13086 -2.07617 -3.13086 .........

@raghavgurbaxani how can i map the exact output.Thanks.

[neroangelo296]

@zaddan @Adithyak1998 I used this model, worked fine for me https://github.com/BVLC/caffe/tree/master/models/bvlc_reference_rcnn_ilsvrc13

Yeah it is working fine for me also.

@raghavgurbaxani can you tell me that how to visualize the output because it is giving like-

-1.89844 -2.11719 -1.95312 -1.87207 -2.08789 -2.13086 -2.07617 -3.13086 .........

@raghavgurbaxani how can i map the exact output.Thanks.

@arvindhbti, Have you found a solution to the visualization problem yet?