rxnn 0.2.51__tar.gz → 0.2.52__tar.gz

{rxnn-0.2.51 → rxnn-0.2.52}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: rxnn
-Version: 0.2.51
+Version: 0.2.52
 Summary: RxNN: Reactive Neural Networks Platform
 License: Apache-2.0
 Keywords: deep-learning,ai,machine-learning

{rxnn-0.2.51 → rxnn-0.2.52}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "poetry.core.masonry.api"
 
 [tool.poetry]
 name = "rxnn"
-version = "0.2.51"
+version = "0.2.52"
 description = "RxNN: Reactive Neural Networks Platform"
 
 license = "Apache-2.0"

{rxnn-0.2.51 → rxnn-0.2.52}/src/rxnn/memory/attention.py

@@ -12,6 +12,7 @@ class StmMemoryAttention(nn.Module):
             per_slot_gate: bool = False,
             init_gate: float = 0.0,
             use_dynamic_gate: bool = False,
+            use_tanh_gate: bool = False,
             *args,
             **kwargs
     ):
@@ -24,6 +25,7 @@ class StmMemoryAttention(nn.Module):
         self.use_gated_residual = use_gated_residual
         self.per_slot_gate = per_slot_gate
         self.use_dynamic_gate = use_dynamic_gate
+        self.use_tanh_gate = use_tanh_gate
         if self.use_gated_residual:
             gate_shape = (self.num_layers, self.stm.stm_size, 1) if self.per_slot_gate else (self.num_layers,)
             self.gate = nn.Parameter(torch.full(gate_shape, init_gate))
@@ -37,10 +39,13 @@ class StmMemoryAttention(nn.Module):
         if self.use_dynamic_gate:
             mean_dim = -1 if self.per_slot_gate else [1, 2]
             gate_input = gate * (new_layer_stm + layer_stm).mean(dim=mean_dim, keepdim=True)
-            layer_gate = torch.sigmoid(gate_input)
+            layer_gate = torch.tanh(gate_input) if self.use_tanh_gate else torch.sigmoid(gate_input)
         else:
-            layer_gate = torch.sigmoid(gate)
-        return layer_gate * new_layer_stm + (1 - layer_gate) * layer_stm
+            layer_gate = torch.tanh(gate) if self.use_tanh_gate else torch.sigmoid(gate)
+        if self.use_tanh_gate:
+            return (1 + layer_gate) * new_layer_stm + (1 - layer_gate) * layer_stm
+        else:
+            return layer_gate * new_layer_stm + (1 - layer_gate) * layer_stm
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         new_stm = torch.zeros_like(self.stm.memory)

{rxnn-0.2.51 → rxnn-0.2.52}/src/rxnn/training/mrl.py

@@ -9,7 +9,7 @@ import random, os
 from ..transformers.sampler import BatchSampler
 from .callbacks import MrlTrainerCallback
 from .dataset import MrlCurriculumDataset
-from .utils import smart_concat, smart_concat_critic_states, TokenizedDict
+from .utils import smart_concat, smart_concat_critic_states, TokenizedDict, get_gradient_norms
 from .rl import RlAlgorithm
 from .reward import MrlRewardMode, MrlRewardModel
 from .models import MrlActorAction, MrlActorModel, MrlCriticModel
@@ -109,6 +109,7 @@ class MRLTrainer:
             use_ddp: bool = False,
             use_amp: bool = False,
             dtype: torch.dtype = torch.float32,
+            debug_mode: bool = False,
     ):
         """
         Trainer for Memory Reinforcement Learning (MRL) algorithm for reactive models and Attention-Based Memory System.
@@ -139,6 +140,7 @@ class MRLTrainer:
         self.shared_freeze_embeddings = config.get('freeze_embeddings', False)
         self.freeze_embeddings = self.shared_freeze_embeddings
         self.use_memory_warmup = config.get('use_memory_warmup', False)
+        self.debug_mode = debug_mode
         # Internal update epochs config
         self.shared_update_epochs = config.get('update_epochs', 10)
         self.update_epochs = self.shared_update_epochs
@@ -566,6 +568,14 @@ class MRLTrainer:
         else:
             return main_loss
 
+    def _log_gradients(self):
+        encoder_total, encoder_mean = get_gradient_norms(self.actor.encoder)
+        decoder_total, decoder_mean = get_gradient_norms(self.actor.decoder)
+        mem_att_total, mem_att_mean = get_gradient_norms(self.actor.memory_attention)
+        print(f"Encoder grad norm - total: {encoder_total:.4f}, mean: {encoder_mean:.4f}")
+        print(f"Decoder grad norm - total: {decoder_total:.4f}, mean: {decoder_mean:.4f}")
+        print(f"Memory attention grad norm - total: {mem_att_total:.4f}, mean: {mem_att_mean:.4f}")
+
     def update_actor(self, state: tuple[TokenizedDict, TokenizedDict, TokenizedDict], action: TokenizedDict,
                      advantages: torch.Tensor, old_log_probs: torch.Tensor, epoch: int) -> float:
         # 1. Reset actor gradients
@@ -596,6 +606,8 @@ class MRLTrainer:
             self.scaler.unscale_(self.optimizer)
             torch.nn.utils.clip_grad_norm_(self.actor.unique_parameters(), max_norm=1.0,
                                            error_if_nonfinite=False)
+            if self.debug_mode:
+                self._log_gradients()
             # 4.5 Run scaled optimization step
             self.scaler.step(self.optimizer)
             self.scaler.update()
@@ -613,6 +625,8 @@ class MRLTrainer:
             # 4.4 Clip gradient norms
             torch.nn.utils.clip_grad_norm_(self.actor.unique_parameters(), max_norm=1.0,
                                            error_if_nonfinite=False)
+            if self.debug_mode:
+                self._log_gradients()
             # 4.5 Run scaled optimization step
             self.optimizer.step()
         # 5. Get float loss value for callbacks/writer

{rxnn-0.2.51 → rxnn-0.2.52}/src/rxnn/training/rl.py

@@ -36,7 +36,7 @@ class PPOConfig(TypedDict):
 
 
 class PPOAlgorithm(RlAlgorithm):
-    def __init__(self, config: Optional[PPOConfig] = None):
+    def __init__(self, config: Optional[PPOConfig] = None, debug_mode: bool = False):
         super(PPOAlgorithm, self).__init__()
 
         if config is None:
@@ -49,7 +49,8 @@ class PPOAlgorithm(RlAlgorithm):
         self.entropy_coef = config.get('entropy_coef', 0.01)
         self.use_distributed_advantage_norm = config.get('use_distributed_advantage_norm', False)
         self.clip_critic_values = config.get('clip_critic_values', True)
-        self.critic_value_clip = config.get('critic_value_clip', 10.0)
+        self.critic_value_clip = config.get('critic_value_clip', 20.0)
+        self.debug_mode = debug_mode
 
     def critic_loss(self, values: torch.Tensor, ref_values: torch.Tensor) -> torch.Tensor:
         # Critic loss with clipped values
@@ -96,6 +97,14 @@ class PPOAlgorithm(RlAlgorithm):
 
         advantages = advantages.unsqueeze(-1)
 
+        if self.debug_mode:
+            print(
+                f"Logits stats: min={new_logits.min().item():.4f}, max={new_logits.max().item():.4f}, mean={new_logits.mean().item():.4f}")
+            print(
+                f"Ratio stats: min={ratio.min().item():.4f}, max={ratio.max().item():.4f}, mean={ratio.mean().item():.4f}")
+            print(
+                f"Advantage stats: min={advantages.min().item():.4f}, max={advantages.max().item():.4f}, mean={advantages.mean().item():.4f}")
+
         # c) Clipped surrogate loss
         surr1 = ratio * advantages
         surr2 = torch.clamp(ratio, 1.0 - self.clip_eps, 1.0 + self.clip_eps) * advantages

{rxnn-0.2.51 → rxnn-0.2.52}/src/rxnn/training/utils.py

@@ -1,4 +1,5 @@
 import torch
+import torch.nn as nn
 from typing import TypedDict
 
 
@@ -142,3 +143,13 @@ def smart_concat(query: TokenizedDict, answer: TokenizedDict, max_length: int, p
         'input_ids': combined_ids,
         'attention_mask': combined_mask
     }
+
+def get_gradient_norms(model: nn.Module):
+    total_norm = 0
+    for p in model.parameters():
+        if p.grad is not None:
+            param_norm = p.grad.data.norm(2)
+            total_norm += param_norm.item() ** 2
+    total_norm = total_norm ** 0.5
+    mean_norm = total_norm / len(list(model.parameters()))
+    return total_norm, mean_norm

{rxnn-0.2.51 → rxnn-0.2.52}/src/rxnn/transformers/layers.py

@@ -110,7 +110,7 @@ class ReactiveTransformerLayer(nn.Module):
         residual = x
         if not self.use_post_norm:
             x = self.norm2(x)
-        x = self.memory_cross_attention(x, stm, stm)
+        x = self.memory_cross_attention(x, stm, stm, mask=mask)
         x = residual + x
         if self.use_post_norm:
             x = self.norm2(x)