diff options
Diffstat (limited to 'nerv/layer')
| -rw-r--r-- | nerv/layer/affine.lua | 91 | ||||
| -rw-r--r-- | nerv/layer/bias.lua | 28 | ||||
| -rw-r--r-- | nerv/layer/combiner.lua | 59 | ||||
| -rw-r--r-- | nerv/layer/init.lua | 79 | ||||
| -rw-r--r-- | nerv/layer/mse.lua | 52 | ||||
| -rw-r--r-- | nerv/layer/sigmoid.lua | 31 | ||||
| -rw-r--r-- | nerv/layer/softmax_ce.lua | 68 | ||||
| -rw-r--r-- | nerv/layer/window.lua | 28 |
8 files changed, 436 insertions, 0 deletions
diff --git a/nerv/layer/affine.lua b/nerv/layer/affine.lua new file mode 100644 index 0000000..00cbcfb --- /dev/null +++ b/nerv/layer/affine.lua @@ -0,0 +1,91 @@ +local MatrixParam = nerv.class('nerv.MatrixParam', 'nerv.Param') +local LinearTransParam = nerv.class('nerv.LinearTransParam', 'nerv.MatrixParam') +local BiasParam = nerv.class('nerv.BiasParam', 'nerv.MatrixParam') +local AffineLayer = nerv.class('nerv.AffineLayer', 'nerv.Layer') + +function MatrixParam:read(handle) + self.trans = self.gconf.cumat_type.new_from_host( + nerv.MMatrixFloat.load(handle)) +end + +function MatrixParam:write(handle) + self.trans:new_to_host():save(handle) +end + +function MatrixParam:train_init() + self.correction = self.trans:create() + self.correction:fill(0) +end + +function MatrixParam:update(gradient) + local gconf = self.gconf + self.correction:add(self.correction, gradient, gconf.momentum, 1.0) + -- momentum gain + local mmt_gain = 1.0 / (1.0 - gconf.momentum); + local n = self.gconf.batch_size * mmt_gain + -- perform update + self.trans:add(self.trans, self.correction, 1.0, -gconf.lrate / n) +end + +function LinearTransParam:update(gradient) + MatrixParam.update(self, gradient) + local gconf = self.gconf + -- weight decay + self.trans:add(self.trans, self.trans, 1.0, -gconf.lrate * gconf.wcost) +end + +function AffineLayer:__init(id, global_conf, layer_conf) + self.id = id + self.ltp = layer_conf.ltp + self.bp = layer_conf.bp + self.dim_in = layer_conf.dim_in + self.dim_out = layer_conf.dim_out + self.gconf = global_conf + self:check_dim_len(1, 1) -- exactly one input and one output + self.direct_update = layer_conf.direct_update +end + +function AffineLayer:init(batch_size) + if self.ltp.trans:ncol() ~= self.bp.trans:ncol() then + nerv.error("mismatching dimensions of linear transform and bias paramter") + end + if self.dim_in[1] ~= self.ltp.trans:nrow() then + nerv.error("mismatching dimensions of linear transform parameter and input") + end + if self.dim_out[1] ~= self.ltp.trans:ncol() then + nerv.error("mismatching dimensions of linear transform parameter and output") + end + self.ltp_grad = self.ltp.trans:create() + self.ltp:train_init() + self.bp:train_init() +end + +function AffineLayer:update(bp_err, input, output) + if self.direct_update then + self.ltp.correction:mul(input[1], bp_err[1], 1.0, gconf.momentum, 'T', 'N') + -- momentum gain + local mmt_gain = 1.0 / (1.0 - gconf.momentum); + local n = self.gconf.batch_size * mmt_gain + -- perform update + self.ltp.trans:add(self.ltp.trans, self.ltp.correction, 1.0, -gconf.lrate / n) + else + self.ltp_grad:mul(input[1], bp_err[1], 1.0, 0.0, 'T', 'N') + self.ltp:update(self.ltp_grad) + end + self.bp:update(bp_err[1]:colsum()) +end + +function AffineLayer:propagate(input, output) + -- apply linear transform + output[1]:mul(input[1], self.ltp.trans, 1.0, 0.0, 'N', 'N') + -- add bias + output[1]:add_row(self.bp.trans, 1.0) +end + +function AffineLayer:back_propagate(bp_err, next_bp_err, input, output) + next_bp_err[1]:mul(bp_err[1], self.ltp.trans, 1.0, 0.0, 'N', 'T') +end + +function AffineLayer:get_params() + return nerv.ParamRepo({self.ltp, self.bp}) +end diff --git a/nerv/layer/bias.lua b/nerv/layer/bias.lua new file mode 100644 index 0000000..c99274d --- /dev/null +++ b/nerv/layer/bias.lua @@ -0,0 +1,28 @@ +local BiasLayer = nerv.class("nerv.BiasLayer", "nerv.Layer") + +function BiasLayer:__init(id, global_conf, layer_conf) + self.id = id + self.gconf = global_conf + self.bias = layer_conf.bias + self.dim_in = layer_conf.dim_in + self.dim_out = layer_conf.dim_out + self:check_dim_len(1, 1) +end + +function BiasLayer:init() + if self.dim_in[1] ~= self.bias.trans:ncol() then + nerv.error("mismatching dimensions of input and bias parameter") + end + if self.dim_out[1] ~= self.bias.trans:ncol() then + nerv.error("mismatching dimensions of output and bias parameter") + end +end + +function BiasLayer:propagate(input, output) + output[1]:copy_fromd(input[1]) + output[1]:add_row(self.bias.trans, 1.0) +end + +function BiasLayer:get_params() + return nerv.ParamRepo({self.bias}) +end diff --git a/nerv/layer/combiner.lua b/nerv/layer/combiner.lua new file mode 100644 index 0000000..7bd7617 --- /dev/null +++ b/nerv/layer/combiner.lua @@ -0,0 +1,59 @@ +local CombinerLayer = nerv.class('nerv.CombinerLayer', 'nerv.Layer') + +function CombinerLayer:__init(id, global_conf, layer_conf) + self.id = id + self.lambda = layer_conf.lambda + self.dim_in = layer_conf.dim_in + self.dim_out = layer_conf.dim_out + self.gconf = global_conf + self:check_dim_len(#self.lambda, -1) + if #self.dim_in < 1 then + nerv.error("no input specified") + end + if #self.dim_out < 1 then + nerv.error("no output specified") + end +end + +function CombinerLayer:init(batch_size) + local dim = self.dim_in[1] + for i = 2, #self.dim_in do + if self.dim_in[i] ~= dim then + nerv.error("mismatching dimensions of inputs") + end + end + for i = 1, #self.dim_out do + if self.dim_out[i] ~= dim then + nerv.error("mismatching dimensions of inputs/outputs") + end + end + self.sum = self.gconf.cumat_type(batch_size, dim) +end + +function CombinerLayer:update(bp_err, input, output) +end + +function CombinerLayer:propagate(input, output) + output[1]:fill(0) + for i = 1, #self.dim_in do + output[1]:add(output[1], input[i], 1.0, self.lambda[i]) + end + for i = 2, #self.dim_out do + output[i]:copy_fromd(output[1]) + end +end + +function CombinerLayer:back_propagate(bp_err, next_bp_err, input, output) + local sum = self.sum + sum:copy_fromd(bp_err[1]) + for i = 2, #self.dim_out do + sum:add(sum, bp_err[i], 1.0, 1.0) + end + for i = 1, #self.dim_in do + next_bp_err[i]:add(next_bp_err[i], sum, 0.0, self.lambda[i]) + end +end + +function CombinerLayer:get_params() + return nerv.ParamRepo({}) +end diff --git a/nerv/layer/init.lua b/nerv/layer/init.lua new file mode 100644 index 0000000..3c55a94 --- /dev/null +++ b/nerv/layer/init.lua @@ -0,0 +1,79 @@ +-- The following methods must be implemented to let a layer work properly + +local Param = nerv.class('nerv.Param') + +function Param:__init(id, global_conf) + self.id = id + self.gconf = global_conf +end + +function Param:get_info() + return self.info +end + +function Param:set_info(info) + self.info = info +end + +function Param:read(handle) + nerv.error_method_not_implemented() +end + +function Param:write(handle) + nerv.error_method_not_implemented() +end + +function Param:update(gradient) + nerv.error_method_not_implemented() +end + +local Layer = nerv.class('nerv.Layer') + +function Layer:__init(id, global_conf, layer_conf) + nerv.error_method_not_implemented() +end + +function Layer:init(batch_size) + nerv.error_method_not_implemented() +end + +function Layer:update(bp_err, input, output) + nerv.error_method_not_implemented() +end + +function Layer:propagate(input, output) + nerv.error_method_not_implemented() +end + +function Layer:back_propagate(bp_err, next_bp_err, input, output) + nerv.error_method_not_implemented() +end + +function Layer:check_dim_len(len_in, len_out) + local expected_in = #self.dim_in + local expected_out = #self.dim_out + if len_in > 0 and expected_in ~= len_in then + nerv.error("layer %s expects %d inputs, %d given", + self.id, len_in, expected_in) + end + if len_out > 0 and expected_out ~= len_out then + nerv.error("layer %s expects %d outputs, %d given", + self.id, len_out, expected_out) + end +end + +function Layer:get_params() + nerv.error_method_not_implemented() +end + +function Layer:get_dim() + return self.dim_in, self.dim_out +end + +nerv.include('affine.lua') +nerv.include('sigmoid.lua') +nerv.include('softmax_ce.lua') +nerv.include('bias.lua') +nerv.include('window.lua') +nerv.include('mse.lua') +nerv.include('combiner.lua') diff --git a/nerv/layer/mse.lua b/nerv/layer/mse.lua new file mode 100644 index 0000000..9a97add --- /dev/null +++ b/nerv/layer/mse.lua @@ -0,0 +1,52 @@ +local MSELayer = nerv.class("nerv.MSELayer", "nerv.Layer") + +function MSELayer:__init(id, global_conf, layer_conf) + self.id = id + self.dim_in = layer_conf.dim_in + self.dim_out = layer_conf.dim_out + self.gconf = global_conf + self:check_dim_len(2, -1) +end + +function MSELayer:init(batch_size) + if self.dim_in[1] ~= self.dim_in[2] then + nerv.error("mismatching dimensions of previous network output and labels") + end + self.scale = 1 / self.dim_in[1] + self.total_mse = 0.0 + self.total_frames = 0 + self.mse = self.gconf.cumat_type(batch_size, self.dim_in[1]) + self.mse_sum = self.gconf.cumat_type(batch_size, 1) + self.diff = self.mse:create() +end + +function MSELayer:update(bp_err, input, output) + -- no params, therefore do nothing +end + +function MSELayer:propagate(input, output) + local mse = self.mse + local mse_sum = self.mse_sum + mse:add(input[1], input[2], 1.0, -1.0) + self.diff:copy_fromd(mse) + mse:mul_elem(mse, mse) + mse_sum:add(mse_sum, mse:rowsum(mse), 0.0, self.scale) + if output[1] ~= nil then + output[1]:copy_fromd(mse_sum) + end + self.total_mse = self.total_mse + mse_sum:colsum()[0] + self.total_frames = self.total_frames + mse_sum:nrow() +end + +-- NOTE: must call propagate before back_propagate +function MSELayer:back_propagate(bp_err, next_bp_err, input, output) + local nbe = next_bp_err[1] + nbe:add(nbe, self.diff, 0.0, 2 * self.scale) + if bp_err[1] ~= nil then + nbe:scale_rows_by_col(bp_err[1]) + end +end + +function MSELayer:get_params() + return nerv.ParamRepo({}) +end diff --git a/nerv/layer/sigmoid.lua b/nerv/layer/sigmoid.lua new file mode 100644 index 0000000..dfd09eb --- /dev/null +++ b/nerv/layer/sigmoid.lua @@ -0,0 +1,31 @@ +local SigmoidLayer = nerv.class("nerv.SigmoidLayer", "nerv.Layer") + +function SigmoidLayer:__init(id, global_conf, layer_conf) + self.id = id + self.gconf = global_conf + self.dim_in = layer_conf.dim_in + self.dim_out = layer_conf.dim_out + self:check_dim_len(1, 1) +end + +function SigmoidLayer:init() + if self.dim_in[1] ~= self.dim_out[1] then + nerv.error("mismatching dimensions of input and output") + end +end + +function SigmoidLayer:update(bp_err, input, output) + -- no params, therefore do nothing +end + +function SigmoidLayer:propagate(input, output) + output[1]:sigmoid(input[1]) +end + +function SigmoidLayer:back_propagate(bp_err, next_bp_err, input, output) + next_bp_err[1]:sigmoid_grad(bp_err[1], output[1]) +end + +function SigmoidLayer:get_params() + return nerv.ParamRepo({}) +end diff --git a/nerv/layer/softmax_ce.lua b/nerv/layer/softmax_ce.lua new file mode 100644 index 0000000..daf891e --- /dev/null +++ b/nerv/layer/softmax_ce.lua @@ -0,0 +1,68 @@ +local SoftmaxCELayer = nerv.class("nerv.SoftmaxCELayer", "nerv.Layer") + +function SoftmaxCELayer:__init(id, global_conf, layer_conf) + self.id = id + self.gconf = global_conf + self.dim_in = layer_conf.dim_in + self.dim_out = layer_conf.dim_out + self.compressed = layer_conf.compressed + if self.compressed == nil then + self.compressed = false + end + self:check_dim_len(2, -1) -- two inputs: nn output and label +end + +function SoftmaxCELayer:init(batch_size) + if not self.compressed and (self.dim_in[1] ~= self.dim_in[2]) then + nerv.error("mismatching dimensions of previous network output and labels") + end + self.total_ce = 0.0 + self.total_correct = 0 + self.total_frames = 0 + self.softmax = self.gconf.cumat_type(batch_size, self.dim_in[1]) + self.ce = self.softmax:create() +end + +function SoftmaxCELayer:update(bp_err, input, output) + -- no params, therefore do nothing +end + +function SoftmaxCELayer:propagate(input, output) + local softmax = self.softmax + local ce = self.ce + local classified = softmax:softmax(input[1]) + local label = input[2] + ce:log_elem(softmax) + if self.compressed then + label = label:decompress(input[1]:ncol()) + end + ce:mul_elem(ce, label) + ce = ce:rowsum() + if output[1] ~= nil then + output[1]:copy_fromd(ce) + end + -- add total ce + self.total_ce = self.total_ce - ce:colsum()[0] + self.total_frames = self.total_frames + softmax:nrow() + -- TODO: add colsame for uncompressed label + if self.compressed then + self.total_correct = self.total_correct + classified:colsame(input[2])[0] + end +end + +function SoftmaxCELayer:back_propagate(bp_err, next_bp_err, input, output) + -- softmax output - label + local label = input[2] + if self.compressed then + label = label:decompress(input[1]:ncol()) + end + local nbe = next_bp_err[1] + nbe:add(self.softmax, label, 1.0, -1.0) + if bp_err[1] ~= nil then + nbe:scale_rows_by_col(bp_err[1]) + end +end + +function SoftmaxCELayer:get_params() + return nerv.ParamRepo({}) +end diff --git a/nerv/layer/window.lua b/nerv/layer/window.lua new file mode 100644 index 0000000..4e9a3b1 --- /dev/null +++ b/nerv/layer/window.lua @@ -0,0 +1,28 @@ +local WindowLayer = nerv.class("nerv.WindowLayer", "nerv.Layer") + +function WindowLayer:__init(id, global_conf, layer_conf) + self.id = id + self.gconf = global_conf + self.window = layer_conf.window + self.dim_in = layer_conf.dim_in + self.dim_out = layer_conf.dim_out + self:check_dim_len(1, 1) +end + +function WindowLayer:init() + if self.dim_in[1] ~= self.window.trans:ncol() then + nerv.error("mismatching dimensions of input and window parameter") + end + if self.dim_out[1] ~= self.window.trans:ncol() then + nerv.error("mismatching dimensions of output and window parameter") + end +end + +function WindowLayer:propagate(input, output) + output[1]:copy_fromd(input[1]) + output[1]:scale_rows_by_row(self.window.trans) +end + +function WindowLayer:get_params() + return nerv.ParamRepo({self.window}) +end |