Fixed up accuracy layer.

[BlGene]

This PR extends the accuracy layer to behave more like the SoftmaxWithLoss layer. It allows bottom datum to have a width and height, as well as the ignore_label option. Original functionality of the layer remains unchanged.

[BlGene]

This is added because it allows one to define databases that store several Datums to one keyword.

[sguada]

Leave that for another PR, where it is used.

[shelhamer]

Seconding @sguada. Note also that there is a BlobProtoVector that could be
used for packing multiple pieces of data too.

On Sun, Feb 22, 2015 at 4:01 PM, Sergio Guadarrama <notifications@github.com

wrote:

In src/caffe/proto/caffe.proto
#1942 (comment):

@@ -30,6 +30,10 @@ message Datum {
optional bool encoded = 7 [default = false];
}

+message DatumVector {

Leave that for another PR, where it is used.

—
Reply to this email directly or view it on GitHub
https://github.com/BVLC/caffe/pull/1942/files#r25138930.

[BlGene]

The travis test timed out, it AFAIK it builds and tests fine.

[BlGene]

@sguada @shelhamer I concede DatumVector is out.

[philkr]

Nice PR. I'm currently using it and it works as advertised.
One small remark: As this now operates on larger blobs it might make sense to also have a GPU implementation similar to SoftmaxWithLoss.

[shelhamer]

Thanks for the accuracy layer extension. This needs a rebase after the generalization to N-D blobs in #1970.

[BlGene]

@shelhamer @jeffdonahue
Hi Guys, I updated the PR to include the changes from #2076.

@philkr Thanks. I don't have a GPU implementation planned ATM, but I will push one as soon as I have one.

[jeffdonahue]

Thanks @BlGene, in general please rebase your code rather than merging. I squashed and cherry-picked your changes in #2076.