Search code examples
c++nlpdeep-learningcntk

CTF Reader throwing error for big files in CNTK

I am using a CTF reader function following the CNTK tutorials on Github.

def create_reader(path, is_training, input_dim, label_dim):
    return MinibatchSource(CTFDeserializer(path, StreamDefs(
        features = StreamDef(field='x', shape=input_dim, is_sparse=True),
        labels = StreamDef(field='y', shape=label_dim, is_sparse=False)
    )), randomize=is_training, epoch_size= INFINITELY_REPEAT if is_training else FULL_DATA_SWEEP)

This works completely fine except when the input file size is bigger than a certain size (unknown). Then it throws an error like this:

WARNING: Sparse index value (269) at offset 8923303 in the input file (C:\local\CNTK-2-0-beta6-0-Windows-64bit-CPU-Only\cntk\Examples\common\data_pos_train_balanced_ctf.txt) exceeds the maximum expected value (268).
attempt: Reached the maximum number of allowed errors while reading the input file (C:\local\CNTK-2-0-beta6-0-Windows-64bit-CPU-Only\cntk\Examples\common\data_pos_train_balanced_ctf.txt)., retrying 2-th time out of 5...
.
.
.

RuntimeError: Reached the maximum number of allowed errors while reading the input file (C:\local\CNTK-2-0-beta6-0-Windows-64bit-CPU-Only\cntk\Examples\common\data_pos_train_balanced_ctf.txt).

I identified that this kind of error is being thrown in the file TextParser.cpp https://github.com/Microsoft/CNTK/blob/5633e79febe1dc5147149af9190ad1944742328a/Source/Readers/CNTKTextFormatReader/TextParser.cpp

What is the solution to or work-around for this?

Solution

  • You need to know the dimensionality of your input and also know that indices start from 0. So if you created an input file mapping your vocabulary to the range 1 to 20000 the dimensionality is 20001.