Make Emotion2vec support onnx (#2359)

* Make emotion2vec exportable to onnx

* Make export_meta of emotion2vec consistence with other models

* Include layer norm in the exported onnx model

.gitignore

@@ -27,3 +27,4 @@ GPT-SoVITS*
 modelscope_models
 examples/aishell/llm_asr_nar/*
 *egg-info
+env/

export.py

@@ -2,8 +2,9 @@
 from funasr import AutoModel
 
 model = AutoModel(
-    model="iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
+    model="iic/emotion2vec_base",
+    hub="ms"
 )
 
-res = model.export(type="onnx", quantize=False, opset_version=13, device='cuda')  # fp32 onnx-gpu
+res = model.export(type="onnx", quantize=False, opset_version=13, device='cpu')  # fp32 onnx-gpu
 # res = model.export(type="onnx_fp16", quantize=False, opset_version=13, device='cuda')  # fp16 onnx-gpu

funasr/models/emotion2vec/export_meta.py

@@ -0,0 +1,76 @@
+#!/usr/bin/env python3
+# -*- encoding: utf-8 -*-
+# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
+#  MIT License  (https://opensource.org/licenses/MIT)
+
+import types
+import torch
+import torch.nn.functional as F
+
+
+def export_rebuild_model(model, **kwargs):
+    model.device = kwargs.get("device")
+
+    # store original forward since self.extract_features is calling it
+    model._original_forward = model.forward
+
+    model.forward = types.MethodType(export_forward, model)
+    model.export_dummy_inputs = types.MethodType(export_dummy_inputs, model)
+    model.export_input_names = types.MethodType(export_input_names, model)
+    model.export_output_names = types.MethodType(export_output_names, model)
+    model.export_dynamic_axes = types.MethodType(export_dynamic_axes, model)
+    model.export_name = types.MethodType(export_name, model)
+
+    model.export_name = "emotion2vec"
+    return model
+
+
+def export_forward(
+    self, x: torch.Tensor
+):
+    with torch.no_grad():
+        if self.cfg.normalize:
+            mean = torch.mean(x, dim=1, keepdim=True)
+            var = torch.var(x, dim=1, keepdim=True, unbiased=False)
+            x = (x - mean) / torch.sqrt(var + 1e-5)
+            x = x.view(x.shape[0], -1)
+
+        # Call the original forward directly just like extract_features
+        # Cannot directly use self.extract_features since it is being replaced by export_forward
+        res = self._original_forward(
+            source=x,
+            padding_mask=None,
+            mask=False,
+            features_only=True,
+            remove_extra_tokens=True
+        )
+        
+        x = res["x"]
+
+        return x
+
+
+def export_dummy_inputs(self):
+    return (torch.randn(1, 16000),)
+
+
+def export_input_names(self):
+    return ["input"]
+
+
+def export_output_names(self):
+    return ["output"]
+
+
+def export_dynamic_axes(self):
+    return {
+        "input": {
+            0: "batch_size",
+            1: "sequence_length",
+        },
+        "output": {0: "batch_size", 1: "sequence_length"},
+    }
+
+
+def export_name(self):
+    return "emotion2vec"

funasr/models/emotion2vec/model.py

@@ -265,3 +265,9 @@ class Emotion2vec(torch.nn.Module):
             results.append(result_i)
 
         return results, meta_data
+
+    def export(self, **kwargs):
+        from .export_meta import export_rebuild_model
+
+        models = export_rebuild_model(model=self, **kwargs)
+        return models