diff options
Diffstat (limited to 'nerv/layer/affine.lua')
| -rw-r--r-- | nerv/layer/affine.lua | 75 |
1 files changed, 58 insertions, 17 deletions
diff --git a/nerv/layer/affine.lua b/nerv/layer/affine.lua index 16250fd..b68cf3d 100644 --- a/nerv/layer/affine.lua +++ b/nerv/layer/affine.lua @@ -48,6 +48,10 @@ function MatrixParam:_update(alpha, beta) -- momentum gain local mmt_gain = 1.0 / (1.0 - gconf.momentum) local n = gconf.batch_size * mmt_gain + -- clip gradient + if gconf.clip then + self.correction_acc:clip(-gconf.clip, gconf.clip) + end -- perform update if gconf.momentum > 0 then self.correction:add(self.correction, self.correction_acc, gconf.momentum, 1.0) @@ -87,7 +91,11 @@ local AffineLayer = nerv.class('nerv.AffineLayer', 'nerv.Layer') -- @param global_conf see `self.gconf` of `nerv.Layer.__init` -- @param layer_conf a table providing with settings dedicated for the layer, -- for `layer_conf` fields that are shared by all layers, see --- `nerv.Layer.__init`. The affine layer requires parameters to be bound, the +-- `nerv.Layer.__init`. This fields can be specified: +-- * `activation`: the type of the activation function layer, also known as \sigma in \sigma(Wx + b). The activation function layer must gurantee not use parameter `input` in its `back_propagate` function. Default value none (no activation function). +-- * `no_bias`: a bool value indicates use bias parameter or not. Default value false. +-- * `param_type`: a string table has the same length with `dim_in`, indicates the parameter type for every input. 'D' for diagonal weight matrix, 'N' for normal weight matrix. Default 'N' for every input. +-- The affine layer requires parameters to be bound, the -- following parameter names will be looked up while binding: -- -- * `ltp`: the linear transformation parameter, also known as the weight matrix, W in Wx + b @@ -96,6 +104,11 @@ local AffineLayer = nerv.class('nerv.AffineLayer', 'nerv.Layer') function AffineLayer:__init(id, global_conf, layer_conf) nerv.Layer.__init(self, id, global_conf, layer_conf) self:check_dim_len(-1, 1) -- exactly one output, allow multiple inputs + self.param_type = layer_conf.param_type or table.vector(#self.dim_in, 'N') + if layer_conf.activation then + self.activation = layer_conf.activation('', global_conf, {dim_in = {self.dim_out[1]}, dim_out = {self.dim_out[1]}}) + end + self.no_bias = layer_conf.no_bias self:bind_params() end @@ -108,24 +121,29 @@ function AffineLayer:bind_params() self["ltp" .. i] = self:find_param(pid_list, lconf, self.gconf, nerv.LinearTransParam, {self.dim_in[i], self.dim_out[1]}) + if self.param_type[i] == 'D' then + self['ltp' .. i].trans:diagonalize() + end local no_update = lconf["no_update_ltp" .. i] if (no_update ~= nil) and no_update or lconf.no_update_all then self["ltp" .. i].no_update = true end end self.ltp = self.ltp1 -- alias of ltp1 - self.bp = self:find_param("bp", lconf, self.gconf, - nerv.BiasParam, - {1, self.dim_out[1]}, - nerv.Param.gen_zero) - local no_update = lconf["no_update_bp"] - if (no_update ~= nil) and no_update or lconf.no_update_all then - self.bp.no_update = true + if not self.no_bias then + self.bp = self:find_param("bp", lconf, self.gconf, + nerv.BiasParam, + {1, self.dim_out[1]}, + nerv.Param.gen_zero) + local no_update = lconf["no_update_bp"] + if (no_update ~= nil) and no_update or lconf.no_update_all then + self.bp.no_update = true + end end end function AffineLayer:init(batch_size) - if self.dim_out[1] ~= self.bp.trans:ncol() then + if not self.no_bias and self.dim_out[1] ~= self.bp.trans:ncol() then nerv.error("mismatching dimensions of linear transform and bias paramter") end for i = 1, #self.dim_in do @@ -137,7 +155,13 @@ function AffineLayer:init(batch_size) end self["ltp" .. i]:train_init() end - self.bp:train_init() + if not self.no_bias then + self.bp:train_init() + end + if self.activation then + self.bak_mat = self.mat_type(batch_size, self.dim_out[1]) + self.bak_mat:fill(0) + end end function AffineLayer:batch_resize(batch_size) @@ -147,26 +171,43 @@ end function AffineLayer:update() for i = 1, #self.dim_in do self["ltp" .. i]:update_by_err_input() + if self.param_type[i] == 'D' then + self['ltp' .. i].trans:diagonalize() + end + end + if not self.no_bias then + self.bp:update_by_gradient() end - self.bp:update_by_gradient() end function AffineLayer:propagate(input, output) + local result = self.activation and self.bak_mat or output[1] -- apply linear transform - output[1]:mul(input[1], self.ltp1.trans, 1.0, 0.0, 'N', 'N') + result:mul(input[1], self.ltp1.trans, 1.0, 0.0, 'N', 'N') for i = 2, #self.dim_in do - output[1]:mul(input[i], self["ltp" .. i].trans, 1.0, 1.0, 'N', 'N') + result:mul(input[i], self["ltp" .. i].trans, 1.0, 1.0, 'N', 'N') end -- add bias - output[1]:add_row(self.bp.trans, 1.0) + if not self.no_bias then + result:add_row(self.bp.trans, 1.0) + end + if self.activation then + self.activation:propagate({result}, output) + end end function AffineLayer:back_propagate(bp_err, next_bp_err, input, output) + local result = self.activation and self.bak_mat or bp_err[1] + if self.activation then + self.activation:back_propagate(bp_err, {result}, {result}, output) + end for i = 1, #self.dim_in do - next_bp_err[i]:mul(bp_err[1], self["ltp" .. i].trans, 1.0, 0.0, 'N', 'T') - self["ltp" .. i]:back_propagate_by_err_input(bp_err[1], input[i]) + next_bp_err[i]:mul(result, self["ltp" .. i].trans, 1.0, 0.0, 'N', 'T') + self["ltp" .. i]:back_propagate_by_err_input(result, input[i]) + end + if not self.no_bias then + self.bp:back_propagate_by_gradient(result:colsum()) end - self.bp:back_propagate_by_gradient(bp_err[1]:colsum()) end function AffineLayer:get_params() |