diff options
Diffstat (limited to 'nerv/layer/graph.lua')
| -rw-r--r-- | nerv/layer/graph.lua | 63 |
1 files changed, 63 insertions, 0 deletions
diff --git a/nerv/layer/graph.lua b/nerv/layer/graph.lua index 5790f95..5b5d4c7 100644 --- a/nerv/layer/graph.lua +++ b/nerv/layer/graph.lua @@ -1,5 +1,68 @@ +--- Implements a special kind of layers having an internal structure, a +-- directed graph of connected sub-level layers. + +--- The class describing the concept of a graph layer having an internal +-- structure, a directed graph of connected sub-level layers. Some of these +-- sub-level layers can again be graph layers, thus, it enables nested and +-- recursive layer declaration. The graph layer can be regarded as a container of +-- its sub-level layers. A layer other than a graph layer is also referenced as +-- "*primitive layer*". +-- @type nerv.GraphLayer + local GraphLayer = nerv.class('nerv.GraphLayer', 'nerv.Layer') +--- The constructor. +-- @param id the identifier +-- @param global_conf see `self.gconf` of `nerv.Layer.__init` +-- @param layer_conf a table providing with settings dedicated for the layer, +-- the following fields should be specified: +-- +-- * `lrepo`: the layer repo that should be used to find the sub-level layers +-- * `connections`: an array of 3-tuples describing the connections of +-- sub-level layers, the structure is as follow: +-- +-- { +-- {<from_port1>, <to_port1>, <time_shift1>}, -- tuple 1 +-- {<from_port2>, <to_port2>, <time_shift2>}, -- tuple 2 +-- {<from_port3>, <to_port3>, <time_shift3>}, -- tuple 3 +-- ... +-- } +-- Each tuple stands for a directed edge between two ports. The first two +-- elements in the tuple are called *port specification* which is a string +-- with the following format: +-- +-- <layer_id>[<port_idx>] +-- where the `<layer_id>` is a string that identifies the layer in +-- `lconf.lrepo`, and `<port_id>` is the input or output port index when used +-- in the first or second port specification respectively. +-- +-- The third element in the tuple is an integer specifying the time delay of +-- this connection. In most cases, it will be simply zero. But for an +-- recurrent network, a positive value `i` means the output from `<from_port>` +-- will be used as the input to `<to_port>` in `i`th computation of the future. +-- Negative values are also allowed to propagate the output to the past. +-- +-- Note that there are two possible strings of `<layer_id>` that have special +-- meanings: the string `"<input>"` and `"<output>"` are placeholders of the +-- the input and output ports of the outer graph layer. The input for the graph +-- layer as a whole can be used by establishing connections from +-- `"<input>[i]"`, and vice versa for the output. +-- +-- As an example, tuples: +-- +-- { +-- {"<input>[1]", "affine0[1]", 0}, +-- {"affine0[1]", "sigmoid0[1]", 0}, +-- {"sigmoid0[1]", "affine1[1]", 0}, +-- {"affine1[1]", "<output>[1]", 0} +-- } +-- Specify a graph layer that contains two stacked and fully connected linear +-- transformation sub-level layers. +-- +-- * `reversed`: optional, reverse the time shifting of all connections if true +-- +-- For other `layer_conf` fields that are shared by all layers, see `nerv.Layer.__init`. + function GraphLayer:__init(id, global_conf, layer_conf) nerv.Layer.__init(self, id, global_conf, layer_conf) self.lrepo = layer_conf.layer_repo |