object callImpl(IGraphNodeBase inputs, dynamic training)
{
IGraphNodeBase result = inputs;
var batchNormExtraArgs = new PythonDict <string, object>();
if (training != null)
{
batchNormExtraArgs["training"] = training;
}
for (int part = 0; part < PartCount; part++)
{
result = this.convs[part].apply(result);
result = this.batchNorms[part].apply(result, kwargs: batchNormExtraArgs);
if (part + 1 != PartCount)
{
result = tf.nn.relu(result);
}
}
result += (Tensor)result + inputs;
return(tf.nn.relu(result));
}
public string Sample(Session session, dynamic chars, IReadOnlyDictionary <char, int> vocabulary, int num = 200, string prime = "The ", int samplingType = 1)
{
dynamic state = this.CreateInitialState(session, vocabulary, prime);
int WeightedPick(IEnumerable <float32> weights)
{
double[] sums = weights.Aggregate((sum: 0.0, sums: new List <double>()),
(acc, value) => {
acc.sum += (double)value; acc.sums.Add(acc.sum);
return(acc.sum, acc.sums);
}).sums.ToArray();
int index = Array.BinarySearch(sums, this.random.NextDouble() * sums.Last());
return(index < 0 ? ~index : index);
}
string ret = prime;
char chr = prime.Last();
for (int i = 0; i < num; i++)
{
var x = np.zeros(new TensorShape(1, 1));
x[0, 0] = vocabulary[chr];
var feed = new PythonDict <dynamic, dynamic> {
[this.inputData] = x,
[this.initialState] = state,
};
var outputs = session.run(new dynamic[] { this.probs, this.finalState }, feed);
var probs = outputs[0];
state = outputs[1];
ndarray computedProbabilities = probs[0];
dynamic sample;
switch (samplingType)
{
case 1:
case 2 when chr == ' ':
sample = WeightedPick(computedProbabilities.Cast <ndarray>().SelectMany(s => s.Cast <float32>()));
break;
case 0:
case 2:
sample = computedProbabilities.argmax();
break;
default:
throw new NotSupportedException();
}
var pred = chars[sample];
ret += pred;
chr = pred;
}
return(ret);
}
public static HParams LoadHParams(string modelName)
{
var hParams = DefaultHParams;
string paramsOverridePath = Path.Combine("models", modelName, "hparams.json");
var overrides = JsonConvert.DeserializeObject <Dictionary <string, object> >(File.ReadAllText(paramsOverridePath));
var pyDict = new PythonDict <object, object>();
foreach (var entry in overrides)
{
pyDict.Add(entry.Key, entry.Value);
}
hParams.override_from_dict(pyDict);
return(hParams);
}
private dynamic CreateInitialState(Session session, IReadOnlyDictionary <char, int> vocabulary, string prime)
{
var state = session.run(this.rnn.zero_state(1, tf.float32));
foreach (char chr in prime.Substring(0, prime.Length - 1))
{
var x = np.zeros(new TensorShape(1, 1));
x[0, 0] = vocabulary[chr];
var feed = new PythonDict <dynamic, dynamic> {
[this.inputData] = x,
[this.initialState] = state,
};
state = Enumerable.First(session.run(new dynamic[] { this.finalState }, feed));
}
return(state);
}
static void Main()
{
GradientLog.OutputWriter = Console.Out;
GradientSetup.UseEnvironmentFromVariable();
var input = tf.placeholder(tf.float32, new TensorShape(null, 1), name: "x");
var output = tf.placeholder(tf.float32, new TensorShape(null, 1), name: "y");
var hiddenLayer = tf.layers.dense(input, hiddenSize,
activation: tf.sigmoid_fn,
kernel_initializer: new ones_initializer(),
bias_initializer: new random_uniform_initializer(minval: -x1, maxval: -x0),
name: "hidden");
var model = tf.layers.dense(hiddenLayer, units: 1, name: "output");
var cost = tf.losses.mean_squared_error(output, model);
var training = new GradientDescentOptimizer(learning_rate: learningRate).minimize(cost);
dynamic init = tf.global_variables_initializer();
new Session().UseSelf(session => {
session.run(new[] { init });
foreach (int iteration in Enumerable.Range(0, iterations))
{
var(trainInputs, trainOutputs) = GenerateTestValues();
var iterationDataset = new PythonDict <dynamic, object> {
[input] = trainInputs,
[output] = trainOutputs,
};
session.run(new[] { training }, feed_dict: iterationDataset);
if (iteration % 100 == 99)
{
Console.WriteLine($"cost = {session.run(new[] { cost }, feed_dict: iterationDataset)}");
}
}
var(testInputs, testOutputs) = GenerateTestValues();
var testValues = session.run(new[] { model }, feed_dict: new PythonDict <dynamic, object> {
[input] = testInputs,
});
new variable_scope("hidden", reuse: true).UseSelf(_ => {
Variable w = tf.get_variable("kernel");
Variable b = tf.get_variable("bias");
Console.WriteLine("hidden:");
Console.WriteLine($"kernel= {w.eval()}");
Console.WriteLine($"bias = {b.eval()}");
});
new variable_scope("output", reuse: true).UseSelf(_ => {
Variable w = tf.get_variable("kernel");
Variable b = tf.get_variable("bias");
Console.WriteLine("hidden:");
Console.WriteLine($"kernel= {w.eval()}");
Console.WriteLine($"bias = {b.eval()}");
});
});
}
public void Train(string checkpoint, string run, int?counter, dynamic sessionConfig = null, CancellationToken cancellation = default)
{
Session sess = sessionConfig == null
? Session.NewDyn(config : sessionConfig)
: new Session();
sess.UseSelf(session => {
var context = tf.placeholder(tf.int32, new TensorShape(this.batchSize, null));
var output = Gpt2Model.Model(this.hParams, input: context);
Tensor labels = context[Range.All, Range.StartAt(1)];
Tensor logits = output["logits"][Range.All, Range.EndAt(new Index(1, fromEnd: true))];
var loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits_dyn(
labels: labels,
logits: logits));
var sample = Gpt2Sampler.SampleSequence(
this.hParams,
length: this.sampleLength,
context: context,
batchSize: this.batchSize,
temperature: 1.0f,
topK: 40);
var trainVars = tf.trainable_variables().Where((dynamic var) => var.name.Contains("model"));
var optimizer = new AdamOptimizer(learning_rate: 0.0002).minimize(loss, var_list: trainVars);
var saver = new Saver(
var_list: trainVars,
max_to_keep: 5,
keep_checkpoint_every_n_hours: 1);
session.run(tf.global_variables_initializer());
Console.WriteLine("Loading checkpoint " + checkpoint);
saver.restore(session, checkpoint);
Console.WriteLine("Loading dataset...");
var sampler = new TrainingSampler(this.dataset, this.random);
Console.WriteLine($"Dataset has {sampler.TokenCount} tokens");
string counterFile = Path.Combine(Gpt2Checkpoints.CheckpointDir, run, "counter");
if (counter == null && File.Exists(counterFile))
{
counter = int.Parse(File.ReadAllText(counterFile), CultureInfo.InvariantCulture) + 1;
}
counter = counter ?? 1;
string runCheckpointDir = Path.Combine(Gpt2Checkpoints.CheckpointDir, 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)
{
var @out = session.run(sample, feed_dict: new PythonDict <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(@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();
}
var batch = Enumerable.Range(0, this.batchSize)
.Select(_ => sampler.Sample(1024))
.ToArray();
var placeholderValues = new PythonDict <object, object> {
[context] = batch,
};
var tuple = session.run_dyn((optimizer, loss), feed_dict: placeholderValues);
var lv = tuple.Item2;
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();
});
}
public int Run()
{
dynamic datasets = Py.Import("sklearn.datasets");
dynamic slice = PythonEngine.Eval("slice");
var iris = datasets.load_iris();
dynamic firstTwoFeaturesIndex = new PyTuple(new PyObject[] {
slice(null),
slice(null, 2)
});
var input = iris.data.__getitem__(firstTwoFeaturesIndex);
IEnumerable target = iris.target;
var expectedOutput = target.Cast <dynamic>()
.Select(l => (int)l == 0 ? 1 : -1)
.ToArray();
int trainCount = expectedOutput.Length * 4 / 5;
var trainIn = numpy.np.array(((IEnumerable)input).Cast <dynamic>().Take(trainCount));
var trainOut = numpy.np.array(expectedOutput.Take(trainCount));
var testIn = numpy.np.array(((IEnumerable)input).Cast <dynamic>().Skip(trainCount));
var testOut = numpy.np.array(expectedOutput.Skip(trainCount));
var inPlace = tf.placeholder(shape: new TensorShape(null, input.shape[1]), dtype: tf.float32);
var outPlace = tf.placeholder(shape: new TensorShape(null, 1), dtype: tf.float32);
var w = new Variable(tf.random_normal(shape: new TensorShape((int)input.shape[1], 1)));
var b = new Variable(tf.random_normal(shape: new TensorShape(1, 1)));
var totalLoss = Loss(w, b, inPlace, outPlace);
var accuracy = Inference(w, b, inPlace, outPlace);
var trainOp = new GradientDescentOptimizer(this.flags.InitialLearningRate).minimize(totalLoss);
var expectedTrainOut = trainOut.reshape((trainOut.Length, 1));
var expectedTestOut = testOut.reshape((testOut.Length, 1));
new Session().UseSelf(sess =>
{
var init = tensorflow.tf.global_variables_initializer();
sess.run(init);
for (int step = 0; step < this.flags.StepCount; step++)
{
(numpy.ndarray @in, numpy.ndarray @out) = NextBatch(trainIn, trainOut, sampleCount: this.flags.BatchSize);
var feed = new PythonDict <object, object> {
[inPlace] = @in,
[outPlace] = @out,
};
sess.run(trainOp, feed_dict: feed);
var loss = sess.run(totalLoss, feed_dict: feed);
var trainAcc = sess.run(accuracy, new PythonDict <object, object>
{
[inPlace] = trainIn,
[outPlace] = expectedTrainOut,
});
var testAcc = sess.run(accuracy, new PythonDict <object, object>
{
[inPlace] = testIn,
[outPlace] = expectedTestOut,
});
if ((step + 1) % 100 == 0)
{
Console.WriteLine($"Step{step}: test acc {testAcc}, train acc {trainAcc}");
}
}
//if (this.flags.IsEvaluation)
//{
//}
});
return(0);
}
static int Train(CharRNNTrainingParameters args)
{
var dataLoader = new TextLoader(args.dataDir, args.BatchSize, args.SeqLength);
args.VocabularySize = dataLoader.vocabularySize;
dynamic checkpoint = null;
if (!string.IsNullOrEmpty(args.initFrom))
{
checkpoint = tf.train.latest_checkpoint(args.initFrom);
var savedArgs =
JsonConvert.DeserializeObject <CharRNNModelParameters>(
File.ReadAllText(Path.Combine(args.initFrom, ConfigFileName)));
Trace.Assert(savedArgs.ModelType == args.ModelType);
Trace.Assert(savedArgs.RNNSize == args.RNNSize);
Trace.Assert(savedArgs.LayerCount == args.LayerCount);
Trace.Assert(savedArgs.SeqLength == args.SeqLength);
var(chars, vocabulary) = LoadCharsVocabulary(Path.Combine(args.saveDir, CharsVocabularyFileName));
Trace.Assert(dataLoader.chars.SequenceEqual(chars));
Trace.Assert(dataLoader.vocabulary.SequenceEqual(vocabulary));
}
Directory.CreateDirectory(args.saveDir);
File.WriteAllText(Path.Combine(args.saveDir, ConfigFileName), JsonConvert.SerializeObject(args));
File.WriteAllText(Path.Combine(args.saveDir, CharsVocabularyFileName), JsonConvert.SerializeObject((dataLoader.chars, dataLoader.vocabulary)));
var model = new CharRNNModel(args, training: true);
new Session().UseSelf(session => {
var summaries = tf.summary.merge_all();
var writer = new FileWriter(Path.Combine(args.logDir, DateTime.Now.ToString("s").Replace(':', '-')));
writer.add_graph(session.graph);
session.run(new dynamic[] { tf.global_variables_initializer() });
var globals = tf.global_variables();
var saver = new Saver(globals);
if (checkpoint != null)
{
saver.restore(session, checkpoint);
}
int totalNumberOfBatches = args.epochs * dataLoader.batchCount;
for (int epoch = 0; epoch < args.epochs; epoch++)
{
session.run(new[] { tf.assign(
model.learningRate,
tf.constant(args.learningRate * Math.Pow(args.decayRate, epoch))) });
dataLoader.ResetBatchPointer();
var state = session.run(model.initialState.Items().Cast <object>());
var stopwatch = Stopwatch.StartNew();
for (int batch = 0; batch < dataLoader.batchCount; batch++)
{
stopwatch.Restart();
var(input, targets) = dataLoader.NextBatch();
var feed = new PythonDict <dynamic, dynamic> {
[model.inputData] = input,
[model.targets] = targets,
};
foreach (var(i, tuple) in model.initialState.Items().Enumerate())
{
feed[tuple.c] = state[i].c;
feed[tuple.h] = state[i].h;
}
var step = session.run(new dynamic[] { summaries, model.cost, model.finalState, model.trainOp }, feed);
int currentBatch = (epoch * dataLoader.batchCount) + batch;
writer.add_summary(step[0], currentBatch);
var time = stopwatch.Elapsed;
Console.WriteLine(
$"{currentBatch}/{totalNumberOfBatches} " +
$"(epoch {epoch}), " +
$"train loss = {step[1]}, time/batch = {time}");
if ((currentBatch % args.saveEvery == 0) ||
(epoch == args.epochs - 1 && batch == dataLoader.batchCount - 1))
{
string checkpointPath = Path.Combine(args.saveDir, "model.ckpt");
saver.save(session, checkpointPath, global_step: currentBatch);
Console.WriteLine("model saved to " + checkpointPath);
}
}
}
});
return(0);
}