Add files via upload

More efficient MDX23C memory usage.

separate.py

@@ -727,32 +727,20 @@ class SeperateMDXC(SeperateAttributes):
         mix = torch.cat([torch.zeros(2, C - H), mix, torch.zeros(2, pad_size + C - H)], 1)
         mix = mix.to(self.device)
 
-        chunks = []
+        chunks = mix.unfold(1, C, H).transpose(0, 1)
-        i = 0
+        batches = [chunks[i : i + batch_size] for i in range(0, len(chunks), batch_size)]
-        while i + C <= mix.shape[1]:
-            chunks.append(mix[:, i:i + C])
-            i += H
-        chunks = torch.stack(chunks)
         
-        batches = []
+        X = torch.zeros(S, *mix.shape).to(self.device) if S > 1 else torch.zeros_like(mix) 
-        i = 0
-        while i < len(chunks):
-            batches.append(chunks[i:i + batch_size])
-            i = i + batch_size
-
-        X = torch.zeros(S, 2, C - H) if S > 1 else torch.zeros(2, C - H)
-        X = X.to(self.device)
 
         #with torch.cuda.amp.autocast():
         with torch.no_grad():
+            cnt = 0
             for batch in batches:
                 self.running_inference_progress_bar(len(batches))
                 x = model(batch)
                 for w in x:
-                    a = X[..., :-(C - H)]
+                    X[..., cnt * H : cnt * H + C] += w
-                    b = X[..., -(C - H):] + w[..., :(C - H)]
+                    cnt += 1
-                    c = w[..., (C - H):]
-                    X = torch.cat([a, b, c], -1)
 
         estimated_sources = X[..., C - H:-(pad_size + C - H)] / N