public void Add(string name, torch.nn.Module submodule)
{
Debug.Assert(!handle.IsInvalid);
if (submodule.BoxedModule == null)
{
throw new InvalidOperationException("A Sequential or loaded module may not be added to a Sequential");
}
THSNN_Sequential_push_back(handle, name, submodule.BoxedModule.handle);
torch.CheckForErrors();
}
//public static void FreezeModuleParams(ModuleList modules)
//{
// foreach (var module in modules)
// {
// FreezeModuleParams(module);
// }
//}
public static void FreezeModuleParams(torch.nn.Module module)
{
if (module is null)
{
return;
}
foreach (var param in module.parameters())
{
param.requires_grad = false;
}
}
public ActivationFunction(string name) : base(name)
{
_function = name?.ToLower() switch
{
"relu" => torch.nn.ReLU(),
"gelu" => torch.nn.GELU(),
"gelu_fast" => new GeLUFast(),
"tanh" => torch.nn.Tanh(),
"linear" => torch.nn.Identity(),
_ => throw new NotSupportedException($"Activation function {name} not supported.")
};
}
internal void Add(string name, torch.nn.Module submodule)
{
Debug.Assert(!handle.IsInvalid);
if (submodule.BoxedModule == null)
{
throw new InvalidOperationException("A Sequential or loaded module may not be added to a Sequential");
}
THSNN_Sequential_push_back(handle, name, submodule.BoxedModule.handle);
torch.CheckForErrors();
// Keep the sub-module alive for at least as long as the Sequential object is alive.
_modules.Add(submodule);
_names.Add(name);
}
private static torch.Tensor ForwardOneLayer(torch.Tensor input, torch.Tensor selfAttentionPaddingMask,
torch.nn.Module convLayer, torch.nn.Module layerNorm)
{
using var disposeScope = torch.NewDisposeScope();
torch.Tensor x = selfAttentionPaddingMask.IsNull()
? input.alias()
: input.masked_fill(selfAttentionPaddingMask.T.unsqueeze(-1), 0);
var conv = convLayer.forward(x);
conv.add_(input);
var norm = layerNorm.forward(conv);
return(norm.MoveToOuterDisposeScope());
}
internal void Add(torch.nn.Module module)
{
var name = _modules.Count.ToString();
Add(name, module);
}
