private static double evaluate(Tensor eval_data, TransformerModel model, Loss criterion, int bptt, int ntokens, torch.optim.Optimizer optimizer)
{
model.eval();
using (var d = torch.NewDisposeScope()) {
var src_mask = model.GenerateSquareSubsequentMask(bptt);
var total_loss = 0.0f;
var batch = 0;
for (int i = 0; i < eval_data.shape[0] - 1; batch++, i += bptt)
{
var(data, targets) = GetBatch(eval_data, i, bptt);
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);
total_loss += data.shape[0] * loss.to(torch.CPU).item <float>();
}
data.Dispose();
targets.Dispose();
d.DisposeEverythingBut(src_mask);
}
return(total_loss / eval_data.shape[0]);
}
}
private static void Train(
Model model,
torch.optim.Optimizer optimizer,
Loss loss,
Device device,
IEnumerable <(Tensor, Tensor)> dataLoader,
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);
}
}
}
