local SL = nerv.class('nerv.SelectLinearLayer', 'nerv.Layer')

--id: string
--global_conf: table
--layer_conf: table
--Get Parameters
function SL:__init(id, global_conf, layer_conf)
    nerv.Layer.__init(self, id, global_conf, layer_conf)

    self.vocab = layer_conf.vocab
 
    self:check_dim_len(1, 1)
    self:bind_params()
end

function SL:bind_params()
    self.ltp = self:find_param("ltp", self.lconf, self.gconf, nerv.LinearTransParam, {self.vocab, self.dim_out[1]}) --layer_conf.ltp
end

--Check parameter 
function SL:init(batch_size)
    if (self.dim_in[1] ~= 1) then --one word id 
        nerv.error("mismatching dimensions of ltp and input")
    end
    if (self.dim_out[1] ~= self.ltp.trans:ncol()) then
        nerv.error("mismatching dimensions of bp and output")
    end
    
    self.batch_size = bath_size
    self.ltp:train_init()
end

function SL:update()
    --use this to produce reproducable result, don't forget to set the dropout to zero!
    --for i = 1, input[1]:nrow(), 1 do
    --    local word_vec = self.ltp.trans[input[1][i - 1][0]]
    --    word_vec:add(word_vec, bp_err[1][i - 1], 1, - self.gconf.lrate / self.gconf.batch_size)
    --end 
    
    --I tried the update_select_rows kernel which uses atomicAdd, but it generates unreproducable result
    self.ltp:update_by_err_input()
end

function SL:propagate(input, output)
    --for i = 0, input[1]:ncol() - 1, 1 do
    --    if (input[1][0][i] > 0) then
    --        output[1][i]:copy_fromd(self.ltp.trans[input[1][0][i]])
    --    else
    --        output[1][i]:fill(0)
    --    end
    --end
    output[1]:copy_rows_fromd_by_colidx(self.ltp.trans, input[1])
end

function SL:back_propagate(bp_err, next_bp_err, input, output)
    --input is compressed, do nothing
    self.ltp:back_propagate_by_err_input(bp_err[1], input[1]:decompress(self.vocab))
end

function SL:get_params()
    local paramRepo = nerv.ParamRepo({self.ltp}, self.loc_type)
    return paramRepo
end