Merge pull request #125 from opentensor/staging

v1.1.6

CHANGELOG.md

@@ -1,4 +1,7 @@
 # Changelog
+## 1.1.6 / 2023-08-10
+### What’s Changed
+- Diversity regularization by @isabella618033 in https://github.com/opentensor/validators/pull/124
 
 ## 1.1.5 / 2023-08-08
 ### What’s Changed

openvalidators/__init__.py

@@ -28,6 +28,6 @@
 from . import weights
 from . import event
 
-__version__ = "1.1.5"
+__version__ = "1.1.6"
 version_split = __version__.split(".")
 __spec_version__ = (1000 * int(version_split[0])) + (10 * int(version_split[1])) + (1 * int(version_split[2]))

openvalidators/reward/diversity.py

@@ -55,7 +55,7 @@ def __init__( self, device: str ):
         self.device = device
         self.tokenizer = AutoTokenizer.from_pretrained( DiversityRewardModel.diversity_model_path )
         self.model = AutoModel.from_pretrained( DiversityRewardModel.diversity_model_path ).to(self.device)
-        self.reward_quantile = torch.tensor(0.1).to(self.device)
+        self.reward_bottom_k = 3
         self.history_reward_bottom_k = 2
         self.historic_embeddings = torch.tensor([]).to(self.device)
         self.history_range = (500, 15500)
@@ -116,18 +116,22 @@ def regularise( rewards ):
         similarity = pairwise_cosine_similarity( embeddings, self.historic_embeddings[self.history_range[0]:] )
 
         # Reward to be at the bottom_k smallest of the 1 - similarity score.
-        rewards = torch.topk((1 - similarity), self.history_reward_bottom_k, largest = False)[0][:, -1]
+        rewards = torch.topk((1 - torch.abs(similarity)), self.history_reward_bottom_k, largest = False)[0][:, -1]
 
         return regularise(rewards) 
 
     def get_batch_rewards( self, embeddings: torch.FloatTensor ) -> torch.FloatTensor:
+        def regularise( rewards ):
+            # sigmoid function that maps 0.07 -> 0.23; 0.1 -> 0.5; 0.2 -> 0.98
+            return 1/(1 + torch.exp(-40 * rewards + 4))
+
         # Calculate the pairwise cosine similarity.
         similarity = pairwise_cosine_similarity( embeddings, embeddings )
 
         # Reward to be at the 10% quantile of the 1 - similarity score.
-        rewards = (1 - similarity).quantile(self.reward_quantile, dim = 1 )
+        rewards = torch.topk((1 - torch.abs(similarity)), self.reward_bottom_k, largest = False)[0][:, -1]
 
-        return rewards 
+        return regularise(rewards) 
 
     def get_rewards( self, prompt: str, completions: List[str], name: str ) -> torch.FloatTensor:
         # Check if completions are empty, return 0 if so
@@ -149,4 +153,7 @@ def get_rewards( self, prompt: str, completions: List[str], name: str ) -> torch
         if historic_rewards != None:
             return batch_rewards * historic_rewards
         else:
-            return batch_rewards
+            return batch_rewards
+
+    def normalize_rewards( self, rewards: torch.FloatTensor ) -> torch.FloatTensor:
+        return rewards

openvalidators/reward/relevance.py

@@ -176,4 +176,4 @@ def reward( self, prompt: str, completion: str ) -> torch.FloatTensor:
         # Calculate the pairwise cosine similarity.
         similarity = pairwise_cosine_similarity( prompt_embed, embeddings )
 
-        return similarity
+        return torch.abs(similarity)