- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi, I'm trying to use a network that contains layers with output channels < 8, which is forbidden in NCS. Is there any workaround to solver this problem instead of changing the network?
Another problem is that any network contains the PReLU layer can't pass mvNCCheck, and seems produce wrong result.
Thanks!
Link Copied
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@ms2308 Can you provide a link to your files for testing and debugging purposes?
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@Tome_at_Intel Thanks for the reply. The models can be get from https://github.com/kpzhang93/MTCNN_face_detection_alignment/tree/master/code/codes/MTCNNv2/model .
The prototxt runs fine without the model, so my guess is the implementation might be correct but data order in NCS-PReLU is wrong.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@Tome_at_Intel FYI:
WITH MODEL
mvNCCheck rnet.prototxt -on conv2
mvNCCheck v02.00, Copyright @ Movidius Ltd 2016
/usr/local/bin/ncsdk/Controllers/FileIO.py:52: UserWarning: You are using a large type. Consider reducing your data sizes for best performance
"Consider reducing your data sizes for best performance\033[0m")
USB: Transferring Data…
USB: Myriad Execution Finished
USB: Myriad Connection Closing.
USB: Myriad Connection Closed.
Result: (48, 9, 9)
1) 3085 2.834
2) 3084 2.5684
3) 3412 2.4023
4) 2183 2.334
5) 3094 2.3008
Expected: (48, 9, 9)
1) 2657 2.4727
2) 2649 2.3184
3) 861 2.3086
4) 3085 2.2715
5) 2669 2.2422
Obtained values
Obtained Min Pixel Accuracy: 16.109853982925415% (max allowed=2%), Fail
Obtained Average Pixel Accuracy: 1.488905493170023% (max allowed=1%), Fail
Obtained Percentage of wrong values: 26.080246913580247% (max allowed=0%), Fail
Obtained Pixel-wise L2 error: 2.0507845160736085% (max allowed=1%), Fail
Obtained Global Sum Difference: 412.45660400390625
WITH ZERO VALUE MODEL:
mvNCCheck rnet.prototxt -w zero_weights.caffemodel -on conv2
mvNCCheck v02.00, Copyright @ Movidius Ltd 2016
/usr/local/bin/ncsdk/Controllers/FileIO.py:52: UserWarning: You are using a large type. Consider reducing your data sizes for best performance
"Consider reducing your data sizes for best performance\033[0m")
USB: Transferring Data…
USB: Myriad Execution Finished
USB: Myriad Connection Closing.
USB: Myriad Connection Closed.
Result: (48, 9, 9)
1) 2592 2.7715
2) 2610 2.6523
3) 1133 2.5801
4) 2675 2.5762
5) 2674 2.5723
Expected: (48, 9, 9)
1) 2592 2.7754
2) 2610 2.6484
3) 2675 2.5742
4) 2674 2.5723
5) 1133 2.5449
Obtained values
Obtained Min Pixel Accuracy: 1.6977481544017792% (max allowed=2%), Pass
Obtained Average Pixel Accuracy: 0.13053794391453266% (max allowed=1%), Pass
Obtained Percentage of wrong values: 0.0% (max allowed=0%), Pass
Obtained Pixel-wise L2 error: 0.23660713157300328% (max allowed=1%), Pass
Obtained Global Sum Difference: 14.085982322692871
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@ms2308 Thanks for providing your model files. I'll check this out and get back to you.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@Tome_at_Intel Thanks. There's another issue in the implementation of SSD's 'Normalize' layer. (This layer seems not appear in the Mobile-SSD). The pixel value error between Caffe and NCS in these layers are high enough to make impact on the final output. I couldn't provide the model for testing for now but please have a check.
Since the parameter for 'Normalize' and 'PReLU' layers are similar (one value per input channel), the bug could be same.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@ms2308 Thank you very much for reporting this. We are investigating to confirm if there is a bug/error in the PReLu and Normalize layers.
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page