Assert at least one proposal for inference

[beacandler]

problem

I encountered a problem when using the trained faster_rcnn model to inference a image:
roi_pooling_layer.cu:91 Check failed: error == cudaSuccess(9 vs. 0) invalid configuration arguments

reason

I found that the reason is that when the input image is too small (such as(10,10)), faster_rcnn will resize the image to (720, 720).
when inferencing, this image will be passed to proposal_layer.py，but there is no one roi at all.
then crashed .

code

1.roi_pooling layer calls CAFFE_GET_BLOCKS(N)
$CAFFE_ROOT/src/caffe/layers/roi_pooling_layer.cu 84:15
ROIPoolForward<Dtype><<<CAFFE_GET_BLOCKS(count), CAFFE_CUDA_NUM_THREADS>>>( count, bottom_data, spatial_scale_, channels_, height_, width_, pooled_height_, pooled_width_, bottom_rois, top_data, argmax_data);
2.roi_pooling layer will crash when call CAFFE_GET_BLOCKS(0):
in $CAFFE_ROOT/include/caffe/util/device_alternate.hpp 88:43
inline int CAFFE_GET_BLOCKS(const int N) { return (N + CAFFE_CUDA_NUM_THREADS - 1) / CAFFE_CUDA_NUM_THREADS; }
then the CUDA_POST_KERNEL_CHECK will throw error, then crashed.

solution

1. Assert N > 0, when calls CAFFE_GET_BLOCKS(N)
   return std::max(1, (N + CAFFE_CUDA_NUM_THREADS - 1) / CAFFE_CUDA_NUM_THREADS);
   but this method is not proposed.
2. Assert at least one proposal for inference
   modify $FASTER_RCNN_ROOT/lib/rpn/proposal_layer.py 150 - ,
   when len(keep) == 0
   if len(keep) == 0: keep = [0]
   this method is proposed.