Merge pull request #2 from martinnovaak/403-quantization

martinnovaak · web-flow · commit 47584e960c28 · 2024-05-03T16:12:30.000+02:00
Change quantization to 403
diff --git a/trainer/model.py b/trainer/model.py
@@ -32,5 +32,5 @@ def forward(self, batch: Batch):
         return torch.sigmoid(self.output_layer(hidden_features))
 
     def clamp_weights(self):
-        self.feature_transformer.weight.data.clamp_(-2.0, 2.0)
-        self.output_layer.weight.data.clamp_(-2.0, 2.0)
+        self.feature_transformer.weight.data.clamp_(-1.27, 1.27)
+        self.output_layer.weight.data.clamp_(-1.27, 1.27)
diff --git a/trainer/quantize.py b/trainer/quantize.py
@@ -2,7 +2,7 @@
 import struct
 import numpy as np
 
-QA = 255
+QA = 403
 QB = 64
 QAB = QA * QB
 
