Threshold layer to binarize features

[sguada]

It just binarize the inputs according to a predefined threshold the default is threshold = 0 what means any number greater than 0 would output 1, otherwise would output 0.

Currently there is only a CPU version, and there is not backward available.

[sergeyk]

@sguada care to rebase on dev and include docstring in the manner of other layers in neuron_layers.hpp?

[kloudkl]

Can this be implemented as an operation for the EltwiseLayer?

[sguada]

Threshold layer only takes one blob, EltwiseLayer perform operations
between two blobs. Although in theory one could use two blobs to compare
them, not sure how useful could be that.

That functionality could be added after this PR is merged.

Sergio

2014-05-23 19:13 GMT-07:00 kloudkl [email protected]:

Can this be implemented as an operation for the EltwiseLayer?

—
Reply to this email directly or view it on GitHubhttps://github.com//pull/422#issuecomment-44075151
.

[shelhamer]

It seems reasonable to me to add unary elementwise operations to the
EltwiseLayer. Thresholding as done here and the soft threshold [1] would be
useful. Blob scaling could be done as unary multiplication with a
coefficient too.

@jeffdonahue @sguada what do you think? I could branch off this PR, rebase,
and incorporate these operations into EltwiseLayer.

[1] http://www.simonlucey.com/soft-thresholding/

On Fri, May 23, 2014 at 7:18 PM, Sergio Guadarrama <[email protected]

wrote:

Threshold layer only takes one blob, EltwiseLayer perform operations
between two blobs. Although in theory one could use two blobs to compare
them, not sure how useful could be that.

That functionality could be added after this PR is merged.

Sergio

2014-05-23 19:13 GMT-07:00 kloudkl [email protected]:

Can this be implemented as an operation for the EltwiseLayer?

—
Reply to this email directly or view it on GitHub<
https://github.com/BVLC/caffe/pull/422#issuecomment-44075151>
.

—
Reply to this email directly or view it on GitHubhttps://github.com//pull/422#issuecomment-44075242
.

[sguada]

I think it makes total sense to have unary elementwise operations, so go ahead pull my PR and adapt it to EltwiseLayer.

[kloudkl]

Almost all the neuron layers can be refactored into very simple operations.

The implementations of the non-unary operations are too complex to be placed in the switch cases.

[sguada]

Agreed, neuron layers could be transformed into unary eltement wise
operations

Sergio

2014-05-27 1:44 GMT-07:00 kloudkl [email protected]:

Almost all the neuron layers can be refactored into very simple operationshttps://github.com/antinucleon/cxxnet/blob/master/cxxnet/core/cxxnet_op.h
.

—
Reply to this email directly or view it on GitHubhttps://github.com//pull/422#issuecomment-44249329
.

[shelhamer]

For coherence and modularity it might make sense to have

• a neuron layer, that collects all the activation types
• a threshold layer, for cutoff binarization (as in this PR) and soft
  thresholding
• the eltwise layer we have now, but add unary multiplication for scaling

Threshold could be folded into Eltwise as already discussed or kept
distinct–reconsidering it feels like thresholds deserve their own layer,
even if the implementation fits the Eltwise scheme.

Sounds like proto v2 is brewing...

On Tue, May 27, 2014 at 8:06 AM, Sergio Guadarrama <[email protected]

wrote:

Agreed, neuron layers could be transformed into unary eltement wise
operations

Sergio

2014-05-27 1:44 GMT-07:00 kloudkl [email protected]:

Almost all the neuron layers can be refactored into very simple
operations<
https://github.com/antinucleon/cxxnet/blob/master/cxxnet/core/cxxnet_op.h>

.

—
Reply to this email directly or view it on GitHub<
https://github.com/BVLC/caffe/pull/422#issuecomment-44249329>
.

—
Reply to this email directly or view it on GitHubhttps://github.com//pull/422#issuecomment-44288713
.

[shelhamer]

@sguada I'm stalling on this 'til after NIPS. If you want to rebase and merge now go for it, or else I'll revisit this after the deadline and hack whatever decision we come to.

[sguada]

Thanks, @shelhamer I'm going to rebase and merge it now, we can refactor it later if we decide so.