static DataSet Load(Gpt2Encoder encoder, IEnumerable <string> texts)
{
dynamic numpy = Py.Import("numpy");
var result = new DataSet();
string encodedEndOfText = encoder.EncodedEndOfText;
var chunk = new List <string>();
int chunkSize = 0;
void AddChunk()
{
PyObject tokens = numpy.stack(chunk);
chunk.Clear();
chunkSize = 0;
result.Add(ndarray.Wrap(tokens));
}
foreach (string text in texts)
{
if (string.IsNullOrWhiteSpace(text))
{
continue;
}
if (chunkSize + text.Length + encodedEndOfText.Length >= TrimAfter)
{
AddChunk();
}
else
{
chunkSize += text.Length + encodedEndOfText.Length;
var encoded = encoder.Encode(text);
chunk.AddRange(encoded);
chunk.Add(encodedEndOfText);
}
}
if (chunk.Count > 0)
{
AddChunk();
}
return(result);
}
internal static List <ndarray> LoadDataset(Gpt2Encoder encoder, List <string> fileNames)
{
if (encoder == null)
{
throw new ArgumentNullException(nameof(encoder));
}
var tokenChunks = new List <ndarray>();
foreach (string file in fileNames)
{
Debug.WriteLine($"Reading {file}");
if (Path.GetExtension(file) == ".npz")
{
// pre-encoded
dynamic npzObject = np.load(file);
var npz = npzObject.__enter__();
foreach (var item in npz.files)
{
tokenChunks.Add(npz[item]);
}
npzObject.__exit__();
}
else
{
string rawText = File.ReadAllText(file);
if (String.IsNullOrWhiteSpace(rawText))
{
continue;
}
dynamic numpy = Py.Import("numpy");
PyObject tokens = numpy.stack(encoder.Encode(rawText));
tokenChunks.Add(ndarray.Wrap(tokens));
}
}
return(tokenChunks);
}
