Logit Transmission via gRPC: Relocating scores normalization to local…

… client

alfred/fm/remote/grpc.py

@@ -4,6 +4,7 @@
 import json
 import logging
 import socket
+import torch.nn.functional as F
 from concurrent import futures
 from typing import Optional, Union, Iterable, Tuple, Any, List
 
@@ -122,11 +123,19 @@ def _run_req_gen():
         output = []
         for response in self.stub.Run(_run_req_gen()):
             if response.ranked:
+                logits = ast.literal_eval(response.logit)
+                candidates = list(logits.keys())
+                logit_values = torch.tensor(list(logits.values()))
+                probabilities = F.softmax(logit_values, dim=0)
+                scores = {
+                    candidate: prob.item() for candidate, prob in zip(candidates, probabilities)
+                }
                 output.append(
                     RankedResponse(
                         **{
                             "prediction": response.message,
-                            "scores": ast.literal_eval(response.logit),
+                            "scores": scores,
+                            "logit": logits,
                             "embeddings": bytes_to_tensor(response.embedding),
                         }
                     )
@@ -241,7 +250,7 @@ def Run(self, request_iterator, context):
                 yield query_pb2.RunResponse(
                     message=response.prediction,
                     ranked=True,
-                    logit=str(response.scores),
+                    logit=str(response.logits),
                     embedding=tensor_to_bytes(response.embeddings),
                 )
             else: