private static void train(int epoch, Tensor train_data, TransformerModel model, Loss criterion, int bptt, int ntokens, torch.optim.Optimizer optimizer)
{
model.train();
using (var d = torch.NewDisposeScope()) {
var total_loss = 0.0f;
var batch = 0;
var log_interval = 200;
var src_mask = model.GenerateSquareSubsequentMask(bptt);
var tdlen = train_data.shape[0];
for (int i = 0; i < tdlen - 1; batch++, i += bptt)
{
var(data, targets) = GetBatch(train_data, i, bptt);
optimizer.zero_grad();
if (data.shape[0] != bptt)
{
src_mask = model.GenerateSquareSubsequentMask(data.shape[0]);
}
using (var output = model.forward(data, src_mask)) {
var loss = criterion(output.view(-1, ntokens), targets);
loss.backward();
torch.nn.utils.clip_grad_norm_(model.parameters().ToArray(), 0.5);
optimizer.step();
total_loss += loss.to(torch.CPU).item <float>();
}
if (batch % log_interval == 0 && batch > 0)
{
var cur_loss = total_loss / log_interval;
Console.WriteLine($"epoch: {epoch} | batch: {batch} / {tdlen / bptt} | loss: {cur_loss:0.00}");
total_loss = 0;
}
d.DisposeEverythingBut(src_mask);
}
}
}
private static void train(int epoch, TorchTensor train_data, TransformerModel model, Loss criterion, int bptt, int ntokens, Optimizer optimizer)
{
model.Train();
var total_loss = 0.0f;
var src_mask = model.GenerateSquareSubsequentMask(bptt);
var batch = 0;
var log_interval = 200;
var tdlen = train_data.shape[0];
for (int i = 0; i < tdlen - 1; batch++, i += bptt)
{
var(data, targets) = GetBatch(train_data, i, bptt);
optimizer.zero_grad();
if (data.shape[0] != bptt)
{
src_mask.Dispose();
src_mask = model.GenerateSquareSubsequentMask(data.shape[0]);
}
var output = model.forward(data, src_mask);
var loss = criterion(output.view(-1, ntokens), targets);
{
loss.backward();
model.parameters().clip_grad_norm(0.5);
optimizer.step();
total_loss += loss.to(Device.CPU).DataItem <float>();
}
GC.Collect();
if (batch % log_interval == 0 && batch > 0)
{
var cur_loss = total_loss / log_interval;
Console.WriteLine($"epoch: {epoch} | batch: {batch} / {tdlen/bptt} | loss: {cur_loss:0.00}");
total_loss = 0;
}
}
}
static void Main(string[] args)
{
Torch.SetSeed(1);
var cwd = Environment.CurrentDirectory;
var device = Torch.IsCudaAvailable() ? Device.CUDA : Device.CPU;
Console.WriteLine($"Running SequenceToSequence on {device.Type.ToString()}");
var vocab_iter = TorchText.Datasets.WikiText2("train", _dataLocation);
var tokenizer = TorchText.Data.Utils.get_tokenizer("basic_english");
var counter = new TorchText.Vocab.Counter <string>();
foreach (var item in vocab_iter)
{
counter.update(tokenizer(item));
}
var vocab = new TorchText.Vocab.Vocab(counter);
var(train_iter, valid_iter, test_iter) = TorchText.Datasets.WikiText2(_dataLocation);
var train_data = Batchify(ProcessInput(train_iter, tokenizer, vocab), batch_size).to(device);
var valid_data = Batchify(ProcessInput(valid_iter, tokenizer, vocab), eval_batch_size).to(device);
var test_data = Batchify(ProcessInput(test_iter, tokenizer, vocab), eval_batch_size).to(device);
var bptt = 32;
var(data, targets) = GetBatch(train_data, 0, bptt);
var ntokens = vocab.Count;
var model = new TransformerModel(ntokens, emsize, nhead, nhid, nlayers, dropout).to(device);
var loss = cross_entropy_loss();
var lr = 2.50;
var optimizer = NN.Optimizer.SGD(model.parameters(), lr);
var scheduler = NN.Optimizer.StepLR(optimizer, 1, 0.95, last_epoch: 15);
var totalTime = new Stopwatch();
totalTime.Start();
foreach (var epoch in Enumerable.Range(1, epochs))
{
var sw = new Stopwatch();
sw.Start();
train(epoch, train_data, model, loss, bptt, ntokens, optimizer);
var val_loss = evaluate(valid_data, model, loss, lr, bptt, ntokens, optimizer);
sw.Stop();
Console.WriteLine($"\nEnd of epoch: {epoch} | lr: {scheduler.LearningRate:0.00} | time: {sw.Elapsed.TotalSeconds:0.0}s | loss: {val_loss:0.00}\n");
scheduler.step();
}
var tst_loss = evaluate(test_data, model, loss, lr, bptt, ntokens, optimizer);
totalTime.Stop();
Console.WriteLine($"\nEnd of training | time: {totalTime.Elapsed.TotalSeconds:0.0}s | loss: {tst_loss:0.00}\n");
}