- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi, I am working on a PWC-Net model and when I load the model converted to IF by Model Optimizer, the program exits with error "Unsupported primitive of type: DataAugmentation name: img0s_aug".
I've found data_augmentation.py file in extentions folder of Model Optimizer, which is supposed to indicate that I have already registered the DataAugmentation layers as a custom model.
I'd like to know how could I fix this error and what is the cause.
Thanks!
Link Copied
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Dear Songlin,
if I understand well your problem, the model successfully converted through the Model Optimizer and you face an issue during inference. On which hardware are you targeting?
For your information, for a custom layer, you need to write an extension for the MO but also for the plugin of your target hardware. This is only possible for CPU and GPU for the moment. The data_augmentation.py file only applies for the support through the MO. For custom layer support on CPU and GPU, you can read more information at this link: https://software.intel.com/en-us/articles/OpenVINO-Custom-Layers-Support-in-Inference-Engine (search CPU kernels).
Best,
Severine
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Dear Severine,
Thanks for your advice.
Yes, as you said, I successfully converted my model and I am targeting on Intel CPU.
It seems that for every custom layers, I need to both write an extension for the MO and also write an implementation of this layer for the Inference Engine, even though it is already implemented as a caffe layer. Is that true?
Best regards,
Songlin
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi Songlin,
yes, to support this layer in the IE for CPU, you need to write an extension too.
Also, for your information, this layer DataAugmentation was added to the MO to support the FlowNet model and it may well be the case that the layer is doing something different even if the model is passing. Therefore, I advise you to look at the MO extension and verify it has the expected behavior.
Best,
Severine
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page