-
Notifications
You must be signed in to change notification settings - Fork 2.5k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Make Theano's max_pool_2d support strides and support padding with 0 (to support size preserving version) #2196
Comments
This would be awesome, it seems like strided pooling is an essential component of many state of the art network architectures lately. One more addition that would be interesting is the support of implicit zero padding around the input. For example, in the GoogLeNet architecture, a component of the 'inception module' is a 3x3 pooling layer with stride 1, that has the same output dimensions as the input. This is only possible if the input is first padded with a border of zeros with width 1. Explicitly doing this padding would require a copy operation, which slows things down unnecessarily. I guess it comes down to substituting a zero whenever an index is out of bounds, so this should be straightforward to implement. |
The pooling routines from Caffe support these options if you can convince @f0k to pull them in. http://caffe.berkeleyvision.org/tutorial/layers.html |
… is greater than the pooling size
…and DownsampleFactorMaxGradGrad issue Theano#2196
… is greater than the pooling size
… is greater than the pooling size
DownsampleFactorMax support strides: issue #2196
… is greater than the pooling size
… is greater than the pooling size
Update, the support for stride on the CPU (with only python code) is merged. |
Padding was added in #2543. |
Allow max_pool_2d (in theano.tensor.signal.downsample) to work with strides that differ from the size of the pooling region.
The text was updated successfully, but these errors were encountered: