训练时，任意width、height的图片都不会报错么？

unsky / FPN

Feature Pyramid Networks for Object Detection

524 stars 263 forks source link

训练时，任意width、height的图片都不会报错么？ #65

Open shuishiwojiade opened 6 years ago

shuishiwojiade commented 6 years ago

根据网络的参数推算每层输出shape时，当我设置输入图片为h x w = 720 x 720时，计算出来，newC4的输出shape=45 x 45。而upp5的输出shape = 46 x 46。P4层为Eltwise，按理说应该会报错呀。因为没有实际运行网络，只是查看网络结果，用NetScope推算每层输出。望解答。

unsky commented 6 years ago

image_stride 的倍数

mantou22 commented 6 years ago

when I run the demo.py,there is a problem,could you help me??????

Loaded network /home/xmpu234/Documents/FPN-caffe_1/data/faster_rcnn_models/fpn_iter_25000.caffemodel Traceback (most recent call last): File "/home/xmpu234/Documents/FPN-caffe1/tools/demo.py", line 142, in , _= im_detect(net, im) File "/home/xmpu234/Documents/FPN-caffe_1/tools/../lib/fast_rcnn/test.py", line 158, in im_detect rois2 = net.blobs['rpn_rois/p2'].data.copy() KeyError: 'rpn_rois/p2