public void Tune()
{
var hyperparams = new GptHParams(
embeddingDim: 16,
attentionHeads: 2,
encoderLayers: 2,
contextTokens: 16,
vocabularySize: TestEncoder.Count);
var encoder = new Gpt2Encoder(TestEncoder, TestBPE);
var dataset = Gpt2Dataset.FromTexts(encoder, new[] { EncoderJson });
var session = new Session();
using var _ = session.StartUsing();
int batchSize = 4;
var input = tf.placeholder(tf.int32, new TensorShape(batchSize, null));
var outputs = Gpt2Model.Model(hyperparams, input);
var tuner = new Gpt2Tuner(hyperparams, session,
inputPlaceholder: input,
outputs,
new GptTrainingSampler(dataset, new Random()),
batchSize: batchSize);
session.run(tf.global_variables_initializer());
float loss0 = tuner.FineTuneOnBatch();
float loss1 = tuner.FineTuneOnBatch();
Assert.True(loss1 < loss0);
}
public override int Run(string[] remainingArguments)
{
this.CheckRequiredArguments();
if (remainingArguments.Length < 1)
{
throw new ArgumentNullException("dataset");
}
string modelPath = CommonCommandOptions.ExpandModelNameToPathOrExit(this.ModelName);
string checkpoint = Gpt2Checkpoints.ProcessCheckpointConfig(
modelPath: modelPath,
checkpoint: this.Checkpoint,
runName: this.RunName);
var encoder = Gpt2Encoder.LoadEncoder(modelPath);
string searchPattern = this.Include ?? "*";
string datasetName = remainingArguments[0];
var dataset = searchPattern.EndsWith("*.csv")
? LoadCsv(encoder, root: datasetName, field: this.ColumnName ?? throw new ArgumentException("column must be specified for training on .csv files"))
: Gpt2Dataset.LoadDataset(encoder, path: datasetName, pattern: searchPattern);
if (dataset.Count == 0)
{
Console.Error.WriteLine("The dataset is empty!");
return(-1);
}
var hParams = Gpt2Model.LoadHParams(modelPath);
var random = this.Seed is null ? new Random() : new Random(this.Seed.Value);
tf.random.set_seed(this.Seed);
var stop = new CancellationTokenSource();
Console.CancelKeyPress += delegate { stop.Cancel(); };
dynamic config = config_pb2.ConfigProto.CreateInstance();
config.gpu_options.allow_growth = true;
var trainer = new Gpt2TunerLegacy(dataset, encoder, hParams, this.BatchSize, this.SampleLength, random)
{
SaveEvery = this.SaveEvery,
SampleNum = this.SampleNum,
SampleEvery = this.SampleEvery,
};
string checkpointOutputDirectory = Path.Combine(modelPath, Gpt2Checkpoints.CheckpointDir);
trainer.FineTune(
checkpointsDir: checkpointOutputDirectory, checkpoint: checkpoint,
run: this.RunName,
counter: checkpoint == "fresh" ? 1 : (int?)null,
sessionConfig: config,
cancellation: stop.Token);
return(0);
}
/// <summary>
/// Interactively run the model
/// </summary>
/// <param name="modelName">Which model to use</param>
/// <param name="checkpoint">Which checkpoint to load</param>
/// <param name="seed">Seed for random number generators, fix seed to reproduce results</param>
/// <param name="sampleCount">Number of samples to return total</param>
/// <param name="batchSize">Number of batches (only affects speed/memory). Must divide sampleCount.</param>
/// <param name="length">Number of tokens in generated text, if null (default), is
/// determined by model hyperparameters</param>
/// <param name="temperature">randomness in boltzmann distribution.
/// Lower temperature results in less random completions. As the
/// temperature approaches zero, the model will become deterministic and
/// repetitive. Higher temperature results in more random completions.</param>
/// <param name="topK">Controls diversity. 1 means only 1 word is
/// considered for each step (token), resulting in deterministic completions,
/// while 40 means 40 words are considered at each step. 0 (default) is a
/// special setting meaning no restrictions. 40 generally is a good value.
/// </param>
public static int Run(string modelName = "117M", string?checkpoint = null, int?seed = null,
int sampleCount = 1,
int batchSize = 1, int?length = null, float temperature = 1, int topK = 0)
{
if (sampleCount % batchSize != 0)
{
throw new ArgumentException();
}
string modelPath = CommonCommandOptions.ExpandModelNameToPathOrExit(modelName);
var encoder = Gpt2Encoder.LoadEncoder(modelPath);
var hParams = Gpt2Model.LoadHParams(modelPath);
int nCtx = hParams.ContextTokens;
if (length is null)
{
length = nCtx;
}
else if (length > nCtx)
{
throw new ArgumentException("Can't get samples longer than window size: " + nCtx);
}
foreach (var gpu in tf.config.list_physical_devices("gpu"))
{
tf.config.experimental.set_memory_growth(gpu, true);
}
var sess = new Session(graph: new Graph());
using var sessionContext = sess.StartUsing();
Tensor context = v1.placeholder(tf.int32, new TensorShape(batchSize, null));
tf.random.set_seed(seed);
Tensor output = Gpt2Sampler.SampleSequence(
hParams: hParams,
length: length.Value,
context: context,
batchSize: batchSize,
temperature: temperature,
topK: topK);
var saver = new Saver();
checkpoint ??= tf.train.latest_checkpoint(modelPath);
saver.restore(sess, checkpoint);
bool interrupted = false;
Console.CancelKeyPress += (object sender, ConsoleCancelEventArgs args) =>
Volatile.Write(ref interrupted, args.Cancel = true);
while (!interrupted)
{
string text;
do
{
Console.Write("Model prompt >>> ");
text = Console.ReadLine();
if (Volatile.Read(ref interrupted))
{
break;
}
if (string.IsNullOrEmpty(text))
{
Console.WriteLine("Prompt should not be empty");
}
} while (!Volatile.Read(ref interrupted) && string.IsNullOrEmpty(text));
if (Volatile.Read(ref interrupted))
{
break;
}
var contextTokens = encoder.Encode(text);
if (!tf.test.is_gpu_available() && contextTokens.Count >= length.Value)
{
Console.Error.WriteLine();
Console.Error.WriteLine("Prompt is too long.");
Console.Error.WriteLine();
continue;
}
int generated = 0;
foreach (int _ in Enumerable.Range(0, sampleCount / batchSize))
{
ndarray <int> @out;
try {
@out = sess.run(output, feed_dict: new Dictionary <object, object> {
[context] = Enumerable.Repeat(contextTokens, batchSize).ToArray(),
})[.., contextTokens.Count..];
} catch (InvalidArgumentError ex) {
throw new ArgumentOutOfRangeException(
"Unable to generate sequence of desired length. "
+ "Try lowering length by passing -l (-sample-length) parameter. "
+ "Current length: " + length.Value,
innerException: ex);
}
foreach (int i in Enumerable.Range(0, batchSize))
{
generated++;
var part = @out[i].AsArray();
text = encoder.Decode(part);
Console.WriteLine($"{Delimiter} SAMPLE {generated} {Delimiter}");
Console.WriteLine(text);
}
}
Console.Write(Delimiter);
Console.WriteLine(Delimiter);
}
return(0);
}
public void FineTune(string checkpointsDir, string checkpoint, string run, int?counter,
int topK = 40, float temperature = 1.0f,
dynamic?sessionConfig = null, CancellationToken cancellation = default)
{
Session session = sessionConfig is null
? Session.NewDyn(config : sessionConfig)
: new Session();
using var _ = session.StartUsing();
Tensor context = v1.placeholder(tf.int32, new TensorShape(this.batchSize, null));
var output = Gpt2Model.Model(this.hParams, input: context);
var sampler = new GptTrainingSampler(this.dataset, this.random);
var optimizer = new AdamOptimizer(learning_rate: 0.0002);
var tuner = new Gpt2Tuner(this.hParams, session, context, output, sampler, this.batchSize, optimizer);
Tensor sample = Gpt2Sampler.SampleSequence(
this.hParams,
length: this.sampleLength,
context: context,
batchSize: this.batchSize,
temperature: temperature,
topK: topK);
var saver = new Saver(
var_list: tuner.ModelVariables,
max_to_keep: 5,
keep_checkpoint_every_n_hours: 1);
session.run(v1.global_variables_initializer());
Console.WriteLine("Loading checkpoint " + checkpoint);
saver.restore(session, checkpoint);
Console.WriteLine("Loading dataset...");
Console.WriteLine($"Dataset has {sampler.TokenCount} tokens");
string counterFile = Path.Combine(checkpointsDir, run, "counter");
if (counter is null && File.Exists(counterFile))
{
counter = int.Parse(File.ReadAllText(counterFile), CultureInfo.InvariantCulture) + 1;
}
counter ??= 1;
string runCheckpointDir = Path.Combine(checkpointsDir, run);
string runSampleDir = Path.Combine(SampleDir, run);
void Save()
{
Directory.CreateDirectory(runCheckpointDir);
Console.WriteLine("Saving " + Path.Combine(runCheckpointDir, Invariant($"model-{counter}")));
saver.save(session,
Path.Combine(runCheckpointDir, "model"),
global_step: counter.Value);
File.WriteAllText(path: counterFile, contents: Invariant($"{counter}"));
}
void GenerateSamples()
{
var contextTokens = np.array(new[] { this.encoder.EncodedEndOfText });
var allText = new List <string>();
int index = 0;
string?text = null;
while (index < this.SampleNum)
{
ndarray <int> @out = session.run(sample, feed_dict: new Dictionary <object, object> {
[context] = Enumerable.Repeat(contextTokens, this.batchSize),
});
foreach (int i in Enumerable.Range(0, Math.Min(this.SampleNum - index, this.batchSize)))
{
text = this.encoder.Decode((ndarray <int>)@out[i]);
text = Invariant($"======== SAMPLE {index + 1} ========\n{text}\n");
allText.Add(text);
index++;
}
}
Console.WriteLine(text);
Directory.CreateDirectory(runSampleDir);
File.WriteAllLines(
path: Path.Combine(runSampleDir, Invariant($"samples-{counter}")),
contents: allText);
}
var avgLoss = (0.0, 0.0);
var startTime = DateTime.Now;
while (!cancellation.IsCancellationRequested)
{
if (counter % this.SaveEvery == 0)
{
Save();
}
if (counter % this.SampleEvery == 0)
{
GenerateSamples();
}
float lv = tuner.FineTuneOnBatch();
avgLoss = (avgLoss.Item1 * 0.99 + lv, avgLoss.Item2 * 0.99 + 1);
Console.WriteLine($"[{counter} | {DateTime.Now - startTime}] loss={lv} avg={avgLoss.Item1 / avgLoss.Item2}");
counter++;
}
Console.WriteLine("Interrupted");
Save();
}
