Update onnx_export.py

onnx_export.py

@@ -1,9 +1,51 @@
 import torch
+from torchaudio.models.wav2vec2.utils import import_fairseq_model
+from fairseq import checkpoint_utils
 from onnxexport.model_onnx import SynthesizerTrn
 import utils
 
-def main(NetExport):
+def get_hubert_model():
+    vec_path = "hubert/checkpoint_best_legacy_500.pt"
+    print("load model(s) from {}".format(vec_path))
+    models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(
+        [vec_path],
+        suffix="",
+    )
+    model = models[0]
+    model.eval()
+    return model
+
+
+def main(HubertExport, NetExport):
     path = "SoVits4.0"
+
+    '''if HubertExport:
+        device = torch.device("cpu")
+        vec_path = "hubert/checkpoint_best_legacy_500.pt"
+        models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(
+            [vec_path],
+            suffix="",
+        )
+        original = models[0]
+        original.eval()
+        model = original
+        test_input = torch.rand(1, 1, 16000)
+        model(test_input)
+        torch.onnx.export(model,
+                          test_input,
+                          "hubert4.0.onnx",
+                          export_params=True,
+                          opset_version=16,
+                          do_constant_folding=True,
+                          input_names=['source'],
+                          output_names=['embed'],
+                          dynamic_axes={
+                              'source':
+                                  {
+                                      2: "sample_length"
+                                  },
+                          }
+                          )'''
     if NetExport:
         device = torch.device("cpu")
         hps = utils.get_hparams_from_file(f"checkpoints/{path}/config.json")
@@ -15,17 +57,15 @@ def main(NetExport):
         _ = SVCVITS.eval().to(device)
         for i in SVCVITS.parameters():
             i.requires_grad = False
-        
-        n_frame = 10
-        test_hidden_unit = torch.rand(1, n_frame, 256)
-        test_pitch = torch.rand(1, n_frame)
-        test_mel2ph = torch.arange(0, n_frame, dtype=torch.int64)[None] # torch.LongTensor([0, 1, 2, 3, 4, 5, 6, 7, 8, 9]).unsqueeze(0)
-        test_uv = torch.ones(1, n_frame, dtype=torch.float32)
-        test_noise = torch.randn(1, 192, n_frame)
+        test_hidden_unit = torch.rand(1, 10, 256)
+        test_pitch = torch.rand(1, 10)
+        test_mel2ph = torch.LongTensor([0, 1, 2, 3, 4, 5, 6, 7, 8, 9]).unsqueeze(0)
+        test_uv = torch.ones(1, 10, dtype=torch.float32)
+        test_noise = torch.randn(1, 192, 10)
         test_sid = torch.LongTensor([0])
         input_names = ["c", "f0", "mel2ph", "uv", "noise", "sid"]
         output_names = ["audio", ]
-        
+        SVCVITS.eval()
         torch.onnx.export(SVCVITS,
                           (
                               test_hidden_unit.to(device),
@@ -51,4 +91,4 @@ def main(NetExport):
 
 
 if __name__ == '__main__':
-    main(True)
+    main(False, True)