search

microsoft / CNTK

Microsoft Cognitive Toolkit (CNTK), an open source deep-learning toolkit
https://docs.microsoft.com/cognitive-toolkit/
Other
17.52k stars 4.28k forks source link

Model changes from 2.1 to 2.2? #2611

Open mortengryning opened 6 years ago

mortengryning commented 6 years ago

Hi.

When I used CNTK 2.1 to train a Fast RCNN model the model contained 3 inputs arguments and 3 outputs

Function modelFunc = Function.Load(modelFile, device);
modelFunc.Arguments.ElementAt(X)
modelFunc.Outputs.ElementAt(Y);

The inputs were named "features", "rois", and "roiLabels". The outputs were named "ce", "errs", and "z". I used the third argument to evaluate the model in c#

After having trained a model in CNTK 2.2, the model only contains 2 input arguments and 2 output arguments. The input arguments are named "data" and "roiproposals". The output arguments are named "cls_pred", and "bbox_regression".

How should I use the new parameters to evaluate a model in C#. Do you have any examples for RCNN for a 2.2 CNTK model?

mortengryning commented 6 years ago

Here's the difference in Shapes: (2500=ROI Suggestions)

2.1 IN Features 1000 1000 3 IN Rois 4 2500 IN RoiLabels 10 2500

OUT Ce 1 250 OUT Errs 1 1 OUT z 10 2500

2.2 IN data 1000 1000 3 IN Roi_proposals 4 2500

OUT cls_pred 4 2500 OUT bbox_regr 16 2500

Are labels no longer required? How do I check which labels was recognized for each prediction?