rxnn 0.2.23__py3-none-any.whl → 0.2.24__py3-none-any.whl

rxnn/rxt/models.py

@@ -137,13 +137,13 @@ class RxTAlphaComponentBase(nn.Module, PyTorchModelHubMixin):
     def load_shared_memory(self, stm: ShortTermMemory):
         self.model.stm = stm
 
-    def freeze_without_memory(self):
+    def freeze_without_memory(self, unfreeze_norms: bool = True):
         for param in self.model.parameters():
             param.requires_grad_(False)
-        self.model.trainable_cross_attention_(True)
+        self.model.trainable_cross_attention_(True, with_norms=unfreeze_norms)
 
-    def freeze_memory(self):
-        self.model.trainable_cross_attention_(False)
+    def freeze_memory(self, with_norms: bool = True):
+        self.model.trainable_cross_attention_(False, with_norms=with_norms)
 
     def unfreeze_all(self):
         for param in self.model.parameters():
@@ -264,6 +264,14 @@ class RxTAlphaMemoryAttention(nn.Module, PyTorchModelHubMixin, license="apache-2
         attention_layers = nn.ModuleList([att_init() for _ in range(num_layers)])
         self.model = StmMemoryAttention(stm, attention_layers, memory_norm_layers)
 
+    def freeze(self):
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def unfreeze(self):
+        for param in self.parameters():
+            param.requires_grad = True
+
     def load_shared_memory(self, stm: ShortTermMemory):
         self.model.stm = stm
 

rxnn/training/models.py

@@ -78,24 +78,30 @@ class MrlActorModel(nn.Module):
     def freeze_components(self, stage: Literal['update', 'fetch', 'both'] = 'both'):
         """Freeze encoder/decoder except memory-related layers."""
         if self.encoder.freeze_without_memory is not None:
-            self.encoder.freeze_without_memory()
+            self.encoder.freeze_without_memory(unfreeze_norms=True)
             if stage == 'update':
-                self.encoder.freeze_memory()
+                self.encoder.freeze_memory(with_norms=True)
         else:
             for param in self.encoder.parameters():
                 param.requires_grad = False
-            self.encoder.model.trainable_cross_attention_(True if stage != 'update' else False)
+            self.encoder.model.trainable_cross_attention_(True if stage != 'update' else False, with_norms=True)
         if self.decoder.freeze_without_memory is not None:
-            self.decoder.freeze_without_memory()
+            self.decoder.freeze_without_memory(unfreeze_norms=True)
             if stage == 'update':
-                self.decoder.freeze_memory()
+                self.decoder.freeze_memory(with_norms=True)
         else:
             for param in self.decoder.parameters():
                 param.requires_grad = False
-            self.decoder.model.trainable_cross_attention_(True if stage != 'update' else False)
+            self.decoder.model.trainable_cross_attention_(True if stage != 'update' else False, with_norms=True)
         # Unfreeze memory attention
-        for param in self.memory_attention.parameters():
-            param.requires_grad = True if stage != 'fetch' else False
+        if self.memory_attention.freeze is not None:
+            if stage == 'fetch':
+                self.memory_attention.freeze()
+            else:
+                self.memory_attention.unfreeze()
+        else:
+            for param in self.memory_attention.parameters():
+                param.requires_grad = True if stage != 'fetch' else False
 
     def unfreeze_components(self):
         """Unfreeze all components after initial training."""
@@ -109,8 +115,11 @@ class MrlActorModel(nn.Module):
         else:
             for param in self.decoder.parameters():
                 param.requires_grad = True
-        for param in self.memory_attention.parameters():
-            param.requires_grad = True
+        if self.memory_attention.unfreeze is not None:
+            self.memory_attention.unfreeze()
+        else:
+            for param in self.memory_attention.parameters():
+                param.requires_grad = True
 
     def reset_memory(self):
         self.memory_attention.reset_memory()

rxnn/training/mrl.py

@@ -37,7 +37,7 @@ class CurriculumConfig(TypedDict):
     eval_dataset: Optional[MrlCurriculumDataset]
     callbacks: Optional[list[MrlTrainerCallback]]
     strategy: MrlStrategy
-    unfreeze_epoch: Optional[Union[int, tuple[int, int, int]]]
+    unfreeze_epoch: Optional[Union[int, tuple[int, int, int, int]]]
     random_resets: Optional[bool]
     random_resets_from: Optional[int]
     random_resets_ratio: Optional[float]
@@ -770,11 +770,7 @@ class MRLTrainer:
 
             # 4. Freeze all components except memory attention and memory cross-attention layers in decoder/encoder
             if unfreeze_epoch != 0:
-                is_staged_unfreeze = isinstance(unfreeze_epoch, tuple)
-                if is_staged_unfreeze:
-                    self.actor.freeze_components('update')
-                else:
-                    self.actor.freeze_components()
+                self.actor.freeze_components('both')
 
             # 5. Setup train DataLoader
             if self.use_ddp:
@@ -817,8 +813,10 @@ class MRLTrainer:
                 # 11. Unfreeze all components before selected epoch
                 is_staged_unfreeze = isinstance(unfreeze_epoch, tuple)
                 if is_staged_unfreeze:
-                    fetch_epoch, both_epoch, all_epoch = unfreeze_epoch
-                    if epoch == fetch_epoch:
+                    update_epoch, fetch_epoch, both_epoch, all_epoch = unfreeze_epoch
+                    if epoch == update_epoch:
+                        self.actor.freeze_components('update')
+                    elif epoch == fetch_epoch:
                         self.actor.freeze_components('fetch')
                     elif epoch == both_epoch:
                         self.actor.freeze_components('both')

rxnn/transformers/layers.py

@@ -57,9 +57,12 @@ class ReactiveTransformerLayer(nn.Module):
         self.use_moe = use_moe
         self.use_moe_att = use_moe_att
 
-    def trainable_cross_attention_(self, is_trainable: bool):
+    def trainable_cross_attention_(self, is_trainable: bool, with_norms: bool = True):
         for param in self.memory_cross_attention.parameters():
             param.requires_grad_(is_trainable)
+        if with_norms:
+            for param in self.norm2.parameters():
+                param.requires_grad_(is_trainable)
 
     def update_max_len(self, max_seq_len: int):
         if self.attention.rope is not None:

rxnn/transformers/models.py

@@ -33,11 +33,11 @@ class ReactiveTransformerBase(nn.Module):
         self.num_shared_layers = len(shared_layers) if shared_layers else 0
         self.num_own_layers = len(own_layers) if own_layers else 0
 
-    def trainable_cross_attention_(self, is_trainable: bool):
+    def trainable_cross_attention_(self, is_trainable: bool, with_norms: bool = True):
         for i in range(self.num_shared_layers):
-            self.shared_layers[i].trainable_cross_attention_(is_trainable)
+            self.shared_layers[i].trainable_cross_attention_(is_trainable, with_norms)
         for i in range(self.num_own_layers):
-            self.layers[i].trainable_cross_attention_(is_trainable)
+            self.layers[i].trainable_cross_attention_(is_trainable, with_norms)
 
     def moe_router_loss(self):
         return torch.stack([self.layers[i].moe_router_loss() for i in range(self.num_own_layers) if self.layers[i].use_moe or self.layers[i].use_moe_att] + [

{rxnn-0.2.23.dist-info → rxnn-0.2.24.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: rxnn
-Version: 0.2.23
+Version: 0.2.24
 Summary: RxNN: Reactive Neural Networks Platform
 License: Apache-2.0
 Keywords: deep-learning,ai,machine-learning

{rxnn-0.2.23.dist-info → rxnn-0.2.24.dist-info}/RECORD

@@ -9,14 +9,14 @@ rxnn/memory/attention.py,sha256=p-r8DK3iVhNn-JAESVzIXDCG8gk1R_-x5xHclZ5jgb0,1813
 rxnn/memory/norm.py,sha256=Ofl8Q5NYEF9GQeO0bhM43tkTW91J0y6TSvTAOYMgloM,6278
 rxnn/memory/stm.py,sha256=IH_3INw7FdI013t56ui3Zq9GPUq-k3HeZGjx6BerS4g,3888
 rxnn/rxt/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-rxnn/rxt/models.py,sha256=qlYgU002VE21ZOlcxEM9iv9tAvsbe4mngcMI2sw3j9k,12078
+rxnn/rxt/models.py,sha256=3gCYD_OXvQc8GaXQvRCSj1OcYOSHayWlpP5lsg9wMMk,12389
 rxnn/training/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 rxnn/training/base.py,sha256=_xik1GXE4RJ_nxwqLQ1ccXA5pRtBCi-jL-jeRFBdHBU,11851
 rxnn/training/bml.py,sha256=FJszaQXOLx2ZHBa1CQpyMrG8i4Kj14E-gzDAEK_Ei5k,17272
 rxnn/training/callbacks.py,sha256=-N0MQPpZQaUWCINdTOsjul4bDGbGr2JgQBqOIXBLS6o,35053
 rxnn/training/dataset.py,sha256=7hTilFWPpqUEc6zNcMqBPjxFKxCfvTKKF3E8tVlwccQ,51250
-rxnn/training/models.py,sha256=wf98gYKKm9-ZY3zwdX9NIeJ-lvh7Ro1SoAijmQxYM28,5599
-rxnn/training/mrl.py,sha256=7flR_ZvgSIWvY6JjbRLYJwZnD6cO55N1-zwJIf5VHQ8,43334
+rxnn/training/models.py,sha256=5fl1hESVj2Hakqz5to8ZJzw5Q4_RKZAUq2bn6nRiPV8,6045
+rxnn/training/mrl.py,sha256=14wx3pVha15B7eRWPRgoxRtV5dPtBI0yadIHOYZjX6k,43275
 rxnn/training/reward.py,sha256=7MTVdNm5HnWmt6zFDi3TAYmnVSL_-24riOoY2F7z4x8,11290
 rxnn/training/rl.py,sha256=j-KNLoZjhaEKasYNOc8DxHtwvknAgAJFwvXKot6otFA,3272
 rxnn/training/scheduler.py,sha256=LcjU35mEwz2U5x3U6tLfeeYlBqMxbFSxYzJYuXkWbSY,1408
@@ -25,14 +25,14 @@ rxnn/training/utils.py,sha256=Bw8nZLKIt7NQpUVCYkb_79kWKChVFOYgYXwODo4SvNc,5718
 rxnn/transformers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 rxnn/transformers/attention.py,sha256=d0Igo1Nrn76BphbHrzekiKJfT3RCy4iSXSB6FLAOTgc,16247
 rxnn/transformers/ff.py,sha256=jJnuBDsnnX5uYC_WZH8cXAYrMnz0P-iX7MwcPivjRtI,2533
-rxnn/transformers/layers.py,sha256=MbOIX4PurbTbYxcXSavyFsNpTHCm26K_Ssk_VUCzKIE,7469
+rxnn/transformers/layers.py,sha256=UQZbrAg1UAttPASeqS7BP1a4JalktThmRMzX99Qghss,7618
 rxnn/transformers/mask.py,sha256=J0cfLVLt3SzS2ra3KcY4khrkhI975Dw4CjpUi3Sn25s,419
-rxnn/transformers/models.py,sha256=VvP7r7E6tj7OWsYKlJLCy2vsQ3xSSnlNez6QxR-jBAA,8276
+rxnn/transformers/models.py,sha256=_2qO1SASHtKvTW3dW-Dy9HEmAvoNVC1_addm2tM9Zbs,8325
 rxnn/transformers/moe.py,sha256=j6jEx6Ip0zttlUZKKn82azxo95lkLZs-H2GLSMD88hY,5859
 rxnn/transformers/positional.py,sha256=1PjcJybUzeQlIKJI4tahAGZcYgCRCL0otxs7mpsNuzM,4410
 rxnn/transformers/sampler.py,sha256=t6iiQTdLQ0TakUWnnhKkb5DKF2F_9-thXHBydDF3fxg,17389
 rxnn/utils.py,sha256=ihb6OTyDtPiocB_lOvnq7eOkjjpCkgs8wxvXUBNQ7mM,996
-rxnn-0.2.23.dist-info/LICENSE,sha256=C8coDFIUYuOcke4JLPwTqahQUCyXyGq6WOaigOkx8tY,11275
-rxnn-0.2.23.dist-info/METADATA,sha256=CqyER_fR5UiM00I8t2x41ORbwzagxq8pv-Eduy2RAng,25960
-rxnn-0.2.23.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
-rxnn-0.2.23.dist-info/RECORD,,
+rxnn-0.2.24.dist-info/LICENSE,sha256=C8coDFIUYuOcke4JLPwTqahQUCyXyGq6WOaigOkx8tY,11275
+rxnn-0.2.24.dist-info/METADATA,sha256=PrVfcCd8NBFtFnD8lAJqU7UW3lLEc-Tr7MQhK6obvuo,25960
+rxnn-0.2.24.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
+rxnn-0.2.24.dist-info/RECORD,,