- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi everyone!.
I have a model trained in Caffe that works perfectly on CPU, nevertheless, although the conversion to graph works correctly and I can run inference with no problems I always get very bad results at the output layer.
I use an input of dimensions (1x3x100x10) with kernels of (3x3). I am not sure if this could be the problem, I read in the documentation that only squared kernels could be used, but I didn't find information about the inputs.
I use a convolutional layer at the first step and when I compare both outputs (from my caffe model, and from the converted graph file) I find both very different.
Is this a bug or a not supported feature?.
PD: I tested other models (with square inputs, for example, yolo) and they all work fine.
Thanks for your answers!
Link Copied
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@edgar_mg Thanks for reporting this. To answer your question, the NCSDK should support non-square input resolutions. Can you provide your prototxt and caffeweights model file for testing purposes? Thanks again.
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page