PT gradient_noise

returnn/torch/updater.py

@@ -5,7 +5,7 @@
 
 from __future__ import annotations
 
-from typing import Optional, Union, Any, Type, Sequence, Set, Dict, List, Tuple
+from typing import Optional, Union, Any, Type, Sequence, Iterable, Set, Dict, List, Tuple
 import os
 import gc
 import torch
@@ -118,11 +118,11 @@ def __init__(self, *, config, network, device, initial_learning_rate=1.0):
         self.optimizer = None  # type: typing.Optional[torch.optim.Optimizer]
 
         self._grad_clip_global_norm = self.config.float("gradient_clip_global_norm", 0.0)
+        self._grad_noise = self.config.float("gradient_noise", 0.0)
 
         # Check other options we have in TF updater, which we might support here later as well,
         # but currently do not support.
         for opt_name in [
-            "gradient_noise",
             "gradient_clip",
             "gradient_clip_norm",
             "gradient_clip_avg_norm",
@@ -183,6 +183,9 @@ def step(self, *, grad_scaler: Optional[torch.cuda.amp.GradScaler] = None):
         """
         if self._grad_clip_global_norm:
             torch.nn.utils.clip_grad_norm_(self.network.parameters(), self._grad_clip_global_norm)
+        if self._grad_noise:
+            gradient_noise_(self.network.parameters(), self._grad_noise)
+
         if grad_scaler is not None:
             grad_scaler.step(self.optimizer)
             grad_scaler.update()
@@ -532,3 +535,15 @@ def _wrap_user_blacklist_wd_modules(
         assert issubclass(mod, (rf.Module, torch.nn.Module)), f"invalid blacklist_weight_decay_modules {mods!r}"
         res.append(mod)
     return tuple(res)
+
+
+def gradient_noise_(params: Iterable[torch.nn.Parameter], std: float):
+    """
+    Add gradient noise to parameters, using a truncated normal distribution.
+    """
+    a, b = -2 * std, 2 * std
+    for param in params:
+        if param.requires_grad and param.grad is not None:
+            noise = torch.empty_like(param.grad)
+            torch.nn.init.trunc_normal_(noise, std=std, a=a, b=b)
+            param.grad += noise