search

Azure / ObjectDetectionUsingCntk

Fast R-CNN Object Detection on Azure using CNTK
129 stars 61 forks source link

CNTK 2.1 error #25

Open dashesy opened 7 years ago

dashesy commented 7 years ago

When I run 3_runCntk.py I get this error:

TypeError: roipooling() missing 2 required positional arguments: 'roi_output_shape' and 'spatial_scale'

Is this some API breakage in version 2?

Edit: This is an API breakage in 2.1 because 2.0 works. Should we report to CNTK (if there is any API backwards compatibility assurance across major versions)?

pkranen commented 7 years ago

The signature of roi pooling changed from 2.0 to 2.1 (breaking change). The 2.1 version expects roi coordinates in absolute pixels, not relative to the image size anymore (now it is the same as in Caffe).

ghost commented 7 years ago

I am using cntk version 2.0 but still get the same msg

Herman-Wu commented 7 years ago

chang "from cntk.ops import input_variable, parameter, times, combine, roipooling " to to "from cntk.ops import input_variable, parameter, times, combine, roipooling,MAX_POOLING" and chang " roi_out=roipooling (conv_out, rois,(roi_dim, roi_dim)" to " roi_out=roipooling (conv_out, rois, MAX_POOLING,(roi_dim, roi_dim),1/16)". Then it should work.

michalmar commented 6 years ago

@Herman-Wu what does that change do? (meaning)