CloneOptions
Options & Configuration · AiDotNet.Models
Controls what a Clone carries across from the original.
Implements: IEquatable<CloneOptions>
For Beginners
Cloning a model gives you a second, separate copy. This class lets you say
how much of the original the copy should bring with it. You almost never need to set any of
this — model.Clone() already does the sensible thing, giving you a complete independent
copy that behaves exactly like the original.
How It Works
The default is deliberately stronger than the equivalent in other libraries. In PyTorch,
copy.deepcopy(model) cannot carry optimizer state, because the optimizer is a separate
object that merely holds references to the model’s parameters — so a copy taken mid-training
restarts its optimizer from scratch. That is a structural limitation rather than a considered
choice, so this library does not reproduce it: Full carries optimizer state, and
a clone taken mid-training resumes as the original would.
The one thing the default does not share is the random number stream. Two models drawing
from the same stream produce identical dropout masks and identical shuffles forever, so their
training silently correlates and nothing in the results reveals it. The clone instead gets a
fresh stream derived from the original’s seed, which stays reproducible without the coupling.
Set ShareRandomState when you genuinely want a bit-identical twin.
The contract in one line: a clone trains as if it were the original, but its randomness does not track it.
Constructors
| Constructor | Summary |
|---|---|
CloneOptions() |
Properties
| Property | Summary |
|---|---|
CloneMode Mode { get; set; } |
Gets how much of the original’s storage the copy is allowed to share. |
CloneOptions Architecture { get; } |
Gets a configuration-only copy: same architecture and settings, nothing learned. |
CloneOptions CopyOnWrite { get; } |
Gets a copy that shares each weight tensor’s storage until either side writes to it. |
CloneOptions Full { get; } |
Gets the default: a complete, independent copy with a freshly derived random stream. |
CloneOptions Shared { get; } |
Gets an ALIAS: the copy points at the same parameters as the original. |
bool IncludeBuffers { get; set; } |
Gets a value indicating whether non-gradient learned state is carried. |
bool IncludeConfiguration { get; } |
Gets a value indicating whether configuration is carried. |
bool IncludeOptimizerState { get; set; } |
Gets a value indicating whether optimizer state is carried. |
bool IncludeParameters { get; set; } |
Gets a value indicating whether learned parameters are carried. |
bool ShareRandomState { get; set; } |
Gets a value indicating whether the clone shares the original’s random stream rather than deriving its own. |
Methods
| Method | Summary |
|---|---|
CloneOptions <Clone>$() |
|
bool Equals(CloneOptions other) |
|
bool Equals(object obj) |
|
bool op_Equality(CloneOptions left, CloneOptions right) |
|
bool op_Inequality(CloneOptions left, CloneOptions right) |
|
int GetHashCode() |
|
string ToString() |