public static Tensor dropout(Tensor x, double p = 0.5, bool training = true)
{
if (!training)
{
return(x);
}
switch (x.dtype)
{
case torch.float16:
{
var y = new Tensor(x.__shape, x.dtype, (!torch.autograd.grad_mode.no_grad.prev) && x.requires_grad);
var dropmap = new bool[x.__half.Length];
MKL.Dropout(x.__half, p, y.__half, dropmap);
if (y.requires_grad)
{
y.__backward_fn = () =>
{
MKL.dDropout(x.grad.__half, p, dropmap, y.grad.__half);
if (x.__backward_fn != null)
{
x.__backward_fn();
}
};
}
return(y);
}
case torch.float32:
{
var y = new Tensor(x.__shape, x.dtype, (!torch.autograd.grad_mode.no_grad.prev) && x.requires_grad);
var dropmap = new bool[x.__float.Length];
MKL.Dropout(x.__float, p, y.__float, dropmap);
if (y.requires_grad)
{
y.__backward_fn = () =>
{
MKL.dDropout(x.grad.__float, p, dropmap, y.grad.__float);
if (x.__backward_fn != null)
{
x.__backward_fn();
}
};
}
return(y);
}
case torch.float64:
{
var y = new Tensor(x.__shape, x.dtype, (!torch.autograd.grad_mode.no_grad.prev) && x.requires_grad);
var dropmap = new bool[x.__double.Length];
MKL.Dropout(x.__double, p, y.__double, dropmap);
if (y.requires_grad)
{
y.__backward_fn = () =>
{
MKL.dDropout(x.grad.__double, p, dropmap, y.grad.__double);
if (x.__backward_fn != null)
{
x.__backward_fn();
}
};
}
return(y);
}
case torch.int8:
case torch.uint8:
case torch.int16:
case torch.int32:
case torch.int64:
{
throw new TorchException("TorchException: nn.functional.dropout is not implemented for integer tensors.");
}
case torch.@bool:
{
throw new TorchException("TorchException: nn.functional.dropout is not implemented for bool tensors.");
}
default:
{
return(null);
}
}
}
