function mycnn(char_dim, num_widths, basic_width, dw, seq_length, c2w_seq_length, c2w_dropout, batch_size)
    local inputs = {}
    local outputs = {}
    table.insert(inputs, nn.Identity()())
    local conv1, _conv1, mp1, _mp1, proj, _proj = {}, {}, {}, {}, {}, {}

    for i = 1, num_widths do
        local nkernels = math.min(200, basic_width * i)
        -- input : seq_length*batch_size X c2w_seq_length X char_dim
        --conv1[i] = nn.TemporalConvolutionFB(char_dim, nkernels, i, dw)(inputs[1])
        local input = nn.View(1, -1, char_dim):setNumInputDims(2)(inputs[1])
        conv1[i] = cudnn.SpatialConvolution(1, nkernels, char_dim, i)(input)
        -- output : seq_length*batch_size X [(c2w_seq_length - i + 1) / dw + 1] X nkernels

        _conv1[i] = nn.Tanh()(conv1[i])

        --mp1[i] = nn.TemporalMaxPooling( outputFrame(c2w_seq_length, i, dw), dw )(_conv1[i])
        mp1[i] = cudnn.SpatialMaxPooling(1, outputFrame(c2w_seq_length, i, dw))(_conv1[i])
        -- output : seq_length*batch_size X 1 X nkernels

        --_mp1[i] = nn.Reshape(seq_length*batch_size, nkernels)(mp1[i])
        _mp1[i] = nn.Squeeze()(mp1[i])

        --_proj[i] = nn.Dropout(c2w_dropout)(_mp1[i]) -- TODO : add dropout?

        table.insert(outputs, _mp1[i])
    end

    local real_output = nn.JoinTable(2)(outputs)

    local module = nn.gModule(inputs, {real_output})
    return transfer_data(module)
end