Is it possible to change the downsample rate of the model？

[KleinXin]

For example, if I want to downsample the feature map 3 times in ResNet50 instead of 4 which means a 14x14 feature map can be acquired through a 224 input image instead of the traditional 7x7 feature map.

This will be very helpfull for those people who focus on segmentation development.

[rwightman]

@KleinXin several of the models support that, including the resnet.py based ones...add arg output_stride=16 for 1/16 * input == 14x14 , and you can use output_stride=8 for 1/8 * input res

[KleinXin]

Thank you for your reply！
After setting output_stride=16, I get the model I want.
By printting out the feature_info list, I guess the recduction in layer4 is removed. I also have a question about the position where the reduction is removed.
Is it possible to set this position? For example, if I want to set the reduction in layer 1/2/4 instead of 1/2/3？

[rwightman]

There is no support for specifying the stride control in that manner, you could manually override the dilation/stride calculations to do that (a bit of work), but I've rarely seen it being done that way...