From 8fecbb8e488569cd8e2f930075120e5f1b1b54fb Mon Sep 17 00:00:00 2001 From: Determinant Date: Wed, 20 Apr 2016 10:35:11 +0800 Subject: add doc for io --- nerv/io/frm_buffer.lua | 36 +++++++++++++++++++++++++++ nerv/io/init.lua | 67 ++++++++++++++++++++++++++++++++++++++++++++++++++ nerv/io/seq_buffer.lua | 34 +++++++++++++++++++++++++ nerv/nn/network.lua | 3 ++- 4 files changed, 139 insertions(+), 1 deletion(-) diff --git a/nerv/io/frm_buffer.lua b/nerv/io/frm_buffer.lua index 9761f16..45f73a0 100644 --- a/nerv/io/frm_buffer.lua +++ b/nerv/io/frm_buffer.lua @@ -1,5 +1,38 @@ +--- Implements a frame-level chopped and shuffled buffer which shall be used +-- for acyclic feed forward NNs (`chunk_size = 1`). +-- @author Ted Yin + +--- The class for a frame-level chopped and shuffled buffer +-- which shall be used for acyclic feed forward NNs +-- @type nerv.FrmBuffer + local FrmBuffer = nerv.class("nerv.FrmBuffer", "nerv.DataBuffer") +--- The constructor. +-- @param global_conf a table describing the computation state and providing +-- with some global settings +-- +-- The following fields in `global_conf` will be used: +-- +-- * `use_cpu`: whether to provide with the chunks/"mini-batches" stored in the +-- main memory on invocation of `get_data()` +-- * `mmat_type`: the class used for creating matrices in CPU computation +-- * `cumat_type` (if `use_cpu = false`): the class used for creating matrices +-- in GPU computation +-- +-- @param buffer_conf a table providing with settings dedicated for the buffer. +-- Available fields includes: +-- +-- * `readers`: an array of `nerv.DataReader` instances specifying the +-- readers used to read data +-- * `batch_size`: the number of rows for each batch matrix +-- * `buffer_size`: the number of frames to be buffered and shuffled at once +-- * `randomize`: shuffle the buffer after filled if true +-- * `consume`: drop the last frames which cannot make up a full `batch_size` +-- matrix if false +-- * `use_gpu`: the buffer space will be allocated on the device (graphics +-- card) if true + function FrmBuffer:__init(global_conf, buffer_conf) self.gconf = global_conf self.batch_size = buffer_conf.batch_size @@ -116,6 +149,9 @@ function FrmBuffer:saturate() return self.tail >= self.batch_size end +--- Get a batch group from the buffer. +-- See `nerv.DataBuffer` for reference + function FrmBuffer:get_data() local batch_size = self.batch_size if self.head >= self.tail then -- buffer is empty diff --git a/nerv/io/init.lua b/nerv/io/init.lua index d3ba27c..4ebbabf 100644 --- a/nerv/io/init.lua +++ b/nerv/io/init.lua @@ -1,3 +1,19 @@ +--- Implements parts of ChunkFile operations (methods) in Lua and define +-- the interface of DataReader and DataBuffer. +-- @author Ted Yin + +--- The class for on-disk chunk storage. +-- A *chunk* can be a group of parameters such as an instance of +-- `nerv.Param` (or `nerv.MatrixParam`, etc.), but can also be something else as long as it implements the following interface: +-- +-- * `.id` field: the unique identifier of the chunk +-- * `read(handle)`: define how to read from a file handle +-- * `write(handle)`: define how to write to a file handle +-- * `get_info()`: return a table of chunk metadata +-- * `set_info(info)`: set the metadata of the chunk +-- * for more information, please refer to `nerv.MatrixParam` as an example +-- @type nerv.ChunkFile + function nerv.ChunkFile:write_chunkdata(metadata, writer) if type(metadata) ~= "table" then nerv.error("metadata should be a Lua table") @@ -6,6 +22,9 @@ function nerv.ChunkFile:write_chunkdata(metadata, writer) return self._write_chunkdata(self.handle, table.tostring(metadata), writer) end +--- Write a chunk to the file. +-- @param chunk the chunk to be serialize. A *chunk* can be any Lua object +-- which implements the required interface. function nerv.ChunkFile:write_chunk(chunk) local id = chunk.id local type = chunk.__typename @@ -17,6 +36,11 @@ function nerv.ChunkFile:write_chunk(chunk) info = chunk:get_info()}, chunk) end +--- Read a chunk from the file. +-- @param id specifying the chunk to be read from a chunk file which may stores multiple chunks +-- @param global_conf a table describing the computation state and providing +-- with some global settings, which will be passed to the constructor of the chunk. + function nerv.ChunkFile:read_chunk(id, global_conf) if self.metadata == nil then nerv.error("wrong file opening mode") @@ -32,26 +56,69 @@ function nerv.ChunkFile:read_chunk(id, global_conf) return chunk end +--- Close a chunk file gracefully. function nerv.ChunkFile:close() self._close(self.handle) end +--- The abstract class which defines the interface of data readers. +-- Data readers, as the name suggests, take a duty of tackling with +-- task-specific low-level I/O, reading from on-disk data files and producing a +-- formalized data block once invoked by the data buffer (see +-- `nerv.DataBuffer`). The main interface is defined by `get_data` method. +-- @type nerv.DataReader + local DataReader = nerv.class("nerv.DataReader") +--- The constructor. +-- @param global_conf a table describing the computation state and providing +-- with some global settings +-- @param reader_conf a table providing with settings dedicated for the +-- reader + function DataReader:__init(global_conf, reader_conf) nerv.error_method_not_implemented() end +--- Get a data block from the reader +-- @return a table which maps data slot identifiers to data matrices. A data +-- slot identifier is a unique string naming one slot of data. Each identifier +-- maps to a matrix containing the data. (`{ = , ...}`) It +-- is a requirement that the number of rows in all the matrices in the returned +-- table stays the same. function DataReader:get_data() nerv.error_method_not_implemented() end +--- The abstract class which defines the interface of data buffers. +-- Buffer readers can be regarded as a data reorganizer which accepts variable +-- length data blocks and chops the stacked blocks into equi-length batch +-- groups used by network computation. The main interface is defined by +-- `get_data` method. +-- @type nerv.DataBuffer local DataBuffer = nerv.class("nerv.DataBuffer") +--- The constructor. +-- @param global_conf a table describing the computation +-- state and providing with some global settings +-- @param buffer_conf a table providing with settings +-- dedicated for the buffer function DataBuffer:__init(global_conf, buffer_conf) nerv.error_method_not_implemented() end +--- Get a batch group from the buffer. +-- @return a table containing the following fields: +-- +-- * `data`: a table which maps slot identifiers to chunks/"mini-batches". Each +-- chunk is an array of batch matrices (`{ = {, , ..., }, ...}`, see `nerv.Network`). +-- * `seq_length` : a table containing the length (number of frames) of each +-- sequence (utterance). (`{, , ..., }`) +-- * `new_seq`: a table containing the indices of batch matrix rows that are the +-- first frames of a sequence. (`{}` when there is no new appearing +-- sequence/utterance, at most `batch_size` elements when all the +-- sequences/utterances is on its first frame) + function DataBuffer:get_data() nerv.error_method_not_implemented() end diff --git a/nerv/io/seq_buffer.lua b/nerv/io/seq_buffer.lua index 029e7b8..65df617 100644 --- a/nerv/io/seq_buffer.lua +++ b/nerv/io/seq_buffer.lua @@ -1,5 +1,36 @@ +--- Implements a sequence-level chopped and shuffled buffer which +-- shall be used for cyclic NNs. +-- @author Qi Liu + +--- The class for a sequence-level chopped and shuffled buffer which +-- shall be used for cyclic NNs. +-- @type nerv.SeqBuffer + local SeqBuffer = nerv.class('nerv.SeqBuffer', 'nerv.DataBuffer') +--- The constructor. +-- @param global_conf a table describing the computation state and providing +-- with some global settings +-- +-- The following fields in `global_conf` will be used: +-- +-- * `use_cpu`: whether to provide with the chunks/"mini-batches" stored in the +-- main memory on invocation of `get_data()` +-- * `mmat_type`: the class used for creating matrices in CPU computation +-- * `cumat_type` (if `use_cpu = false`): the class used for creating matrices +-- in GPU computation +-- +-- @param buffer_conf a table providing with settings dedicated for the buffer. +-- Available fields includes: +-- +-- * `readers`: an array of `nerv.DataReader` instances specifying the +-- readers used to read data +-- * `batch_size`: the number of rows for each batch matrix +-- * `chunk_size`: the length of the BPTT context (number of batch +-- matrices to provide upon each invocation of `get_data()`) +-- * `nn_act_default`: the default value to fill into the "holes" (non-data +-- frames) + function SeqBuffer:__init(global_conf, buffer_conf) self.gconf = global_conf @@ -84,6 +115,9 @@ function SeqBuffer:saturate(batch) return true end +--- Get a batch group from the buffer. +-- See `nerv.DataBuffer` for reference + function SeqBuffer:get_data() local has_data = false for i = 1, self.batch_size do diff --git a/nerv/nn/network.lua b/nerv/nn/network.lua index 19fa9d3..5a6abb6 100644 --- a/nerv/nn/network.lua +++ b/nerv/nn/network.lua @@ -511,7 +511,8 @@ end -- * `output`: similar to `input`, but the matrices have different number of -- columns (depending on the width of the output, which is typically 1 for -- criteria, i.e. single column indicating the error), used to hold the output of the network --- * `seq_length` : a table containing the length (number of frames) of each sequence (utterance) +-- * `seq_length` : a table containing the length (number of frames) of each +-- sequence (utterance) -- * `new_seq`: a table containing the indices of batch matrix rows that are the -- first frames of a sequence -- * `do_train`: a bool value indicating whether to update the network -- cgit v1.2.3