CNTK/Source/Readers/DataReader_v2.txt

thoughts on DataReader redesign
===============================

current basic usage pattern:

 - StartMinibatchLoop() sets the start and MB size
 - GetMinibatch() fills matrices in a dictionary of named matrices
 - sample code:

    std::map<std::wstring, Matrix<ElemType>*> matrices;
    matrices[featureNames[0]] = &featuresMatrix;
    matrices[labelNames[0]] = &labelsMatrix;

    dataReader.StartMinibatchLoop(mbSize, epoch, epochSize);
    while (dataReader.GetMinibatch(matrices))
    {
        Matrix<ElemType>& features = *matrices[featureNames[0]];
        Matrix<ElemType>& labels = *matrices[labelNames[0]];
    // no function called at end, implied in GetMinibatch()

issues with current data reader design:

 - monolithic, combines all (or not) of these:
    - paging in data (incl. format parsing)
    - randomization
    - caching
    - packing of parallel streams
    - prefetch (in the original DBN.exe version of the HTK reader)
    - minibatch decimation in presence of MPI (done in a way that avoids to read data that is not needed by a node)
 - multiple streams must match in their timing frame-by-frame
    - kills sequence-to-sequence
    - currently circumvented by paring multiple networks

goals:

 - remove time-synchronity limitation
    - which means that the interface must separate the notion of frames and utterances
 - break into composable blocks
    - hopefully, people in the future will only have to implement the data paging
    - note: packing is not a reader function; nodes themselves may determine packing for each minibatch
 - more abstract notion of 'utterance' e.g. include variable-size images (2D) and video (3D)
 - seems we canb keep the existing DataReader interface, but with extensions and a new implementation

feature details that must be considered/covered:

 - augmentation of context frames
 - some utterances are missing
    - multi-lingual training (multi-task learning where each utterance only has labels for one task)
    - realignment may fail on some utterances 
 - should support non-matrix data, e.g. lattices
 - maybe we can improve efficiency of decimated minibatch reading (current approach from DBN.exe is not optimally load-balanced)

thinking out loud on how we may proceed (high level):

 - basic unit of thought is the utterance, not the minibatch
    - a minibatch is a set of utterances
    - framewise CE training: each frame is an utterance; the N frames are batched into N streams of 1 frame
       - note: design must be non-wasteful in this important special case
    - an utterance should be understood more generally as a fixed or variable-dimension N-dimensional tensor,
      including images (2D tensor, of possibly variable size) and even video (3D tensor).
      And 'utterance length' generalizes to image dimensions as well. Everything that's variable length.
 - interface Sequencer
    - determines the sequence of utterances and grouping into minibatches
       - by driving on utterance level, different feature streams with mismatching timing are not a concern of the Sequencer
    - owns knowlegde of blocks
       - provides caching control information, that is, when to release data from memory
       - for frame mode, there must be some form of translation between utterances and frames, so that we can cache utterances while randomizing over frames
    - does NOT actually read data; only provides descriptors of what to read, which are passed to pagers
       - DataReader class does the reading
       - in eval mode, there is also a DataWriter
    - class RandomSequencer
       - performs block randomization, based on one user-selected data pager
       - for SGD
    - class RandomFrameSequencer
       - treats frames of utterances into individual utterances and randomizes those (for CE training of DNNs or other windows models)
    - class LinearSequencer
       - returns data in original sequence
       - for evaluation
 - interface DataPager
    - random access to page in utterances
       - specified by a descriptor obtained from the Sequencer
       - knowledge of how to parse input data formats is in these pagers
       - data assumed immutable
    - examples:
       - HTK features
       - HTK labels from MLF (also: labels stored in feature format, for reduced startup time)
       - Python adapter
    - lightweight agreement between DataPager and Sequencer:
       - pager provides block-forming relevant information, such that the reading of data consecutively in each block will be optimal;
         sequencer will ask one user-selected pager to provide this information as a basis for block randomization
    - class CachedDataPager
       - TODO: think this through:
          - are DataPagers driven in blocks? That would be the unit of caching
          - releasing a block from cache must be an explicit function
          - maybe that helper class needs to do that
          - or we use a ref count for utterances to control releasing of blocks? Could be expensive, since invidiual speech frames can be utterances (DNN/CE). It's only a refcount of 0 or 1
    - should we call this DataPageReader or DataBlockReader?
    - let's call them Pager for now (need to change because name has a problem with reading vs. writing)
 - class DataReader
    - outer layer of the new structure
       - designed to support reading data streams that that have mismatching utterance lengths
       - there is only one DataReader instance that handles all utterance-data streams (including mismatching lenths)
    - takes a reference to one user-specified sequencer
    - takes ownership of one or more user-supplied pagers
    - after construction, the above are only accessed through the DataReader
    - a nested hierarchy of DataReaders implement specific functionaliaty
      class CachingDataReader
       - wraps a DataReader with caching--this is what one would use when randomizing (not needed for evaluation)
      class PrefetchingDataReader
       - wraps a DataReader and reads ahead on a parallel thread
          - TODO: so where does the sequencer run?? Or does sequencer provides a FIFO of minibatches (lookahead)?
                  maybe sequence info is routed through the prefetch for everything? Consider that we also need to do writing, so this becomes weird
                  in that one would always have to access the sequencer through the DataReader (in order to get the correctly delayed sequencing information)
      class BatchingDataReader?
       - meant to batch utterances into streams
       - NO: this should not be a DataReader, as this is a network function. But we may have supporting code in the reader interface or a reader helper class
       - instead, there should be a Minibatch batcher class that (re-)batches and reshuffles minibatches (this could be a ComputeNode, actually)
    - this new DataReader differs (extends) the current DataReader as:
       - GetMinibatch() has to return utterance and length/packing information for every minibatch
       - minibatches must also carry their own sequencing information (utterance ids); this can then be used for data writing
       - we may want to rethink the glue between reading and Input nodes. Maybe Input nodes can know about readers?
 - how to set up the whole thing:
    - create all desired data pagers
    - create a Sequencer of the desired type
       - e.g. random utterance, random frame, non-random
    - pass it one user-selected data pager to let it determining how data is grouped in blocks
    - create the desired DataReader by passing it the readers and the sequencer
       - there may be a hierarchy of nested DataReaders, to do caching and prefetching
    - use it mostly as before
       - Note: sequencer information can only be accessed through the DataReader.

on writing:

 - writing is used for evaluation, and also for data conversion
 - DataReader returns minibatches that carry their utterance information (i.e. utterance ids)
 - class DataWriter
    - new SaveData() overload takes an output minibatch complete with utterance information
      TODO: we could make the two interfaces a little more symmetric w.r.t. function naming

[fseide 9/2015]