Add OnnxInference

onnx_infer.py

@@ -0,0 +1,71 @@
+from onnx_modules.V220_OnnxInference import OnnxInferenceSession
+import numpy as np
+Session = OnnxInferenceSession(
+        {
+        "enc" : "onnx/BertVits2.2PT/BertVits2.2PT_enc_p.onnx",   
+        "emb_g" : "onnx/BertVits2.2PT/BertVits2.2PT_emb.onnx",
+        "dp" : "onnx/BertVits2.2PT/BertVits2.2PT_dp.onnx",
+        "sdp" : "onnx/BertVits2.2PT/BertVits2.2PT_sdp.onnx",
+        "flow" : "onnx/BertVits2.2PT/BertVits2.2PT_flow.onnx",
+        "dec" : "onnx/BertVits2.2PT/BertVits2.2PT_dec.onnx"
+        },
+        Providers = ["CPUExecutionProvider"]
+    )
+
+#这里的输入和原版是一样的，只需要在原版预处理结果出来之后加上.numpy()即可
+x = np.expand_dims(
+    np.array(
+        [
+            0,
+            97,
+            0,
+            8,
+            0,
+            78,
+            0,
+            8,
+            0,
+            76,
+            0,
+            37,
+            0,
+            40,
+            0,
+            97,
+            0,
+            8,
+            0,
+            23,
+            0,
+            8,
+            0,
+            74,
+            0,
+            26,
+            0,
+            104,
+            0,
+        ]
+    ),
+    0
+)
+tone = np.zeros_like(x)
+language = np.zeros_like(x)
+sid = np.array([0])
+bert = np.random.randn(x.shape[1], 1024)
+ja_bert = np.random.randn(x.shape[1], 1024)
+en_bert = np.random.randn(x.shape[1], 1024)
+emo = np.random.randn(512, 1)
+
+audio = Session(
+    x,
+    tone,
+    language,
+    bert,
+    ja_bert,
+    en_bert,
+    emo,
+    sid
+)
+
+print(audio)

onnx_modules/V220_OnnxInference/__init__.py

@@ -0,0 +1,92 @@
+import numpy as np
+import onnxruntime as ort
+
+def convert_pad_shape(pad_shape):
+    layer = pad_shape[::-1]
+    pad_shape = [item for sublist in layer for item in sublist]
+    return pad_shape
+
+def sequence_mask(length, max_length=None):
+    if max_length is None:
+        max_length = length.max()
+    x = np.arange(max_length, dtype=length.dtype)
+    return np.expand_dims(x, 0) < np.expand_dims(length, 1)
+
+def generate_path(duration, mask):
+    """
+    duration: [b, 1, t_x]
+    mask: [b, 1, t_y, t_x]
+    """
+
+    b, _, t_y, t_x = mask.shape
+    cum_duration = np.cumsum(duration, -1)
+
+    cum_duration_flat = cum_duration.reshape(b * t_x)
+    path = sequence_mask(cum_duration_flat, t_y)
+    path = path.reshape(b, t_x, t_y)
+    path = path ^ np.pad(path, ((0, 0), (1, 0), (0, 0)))[:, :-1]
+    path = np.expand_dims(path, 1).transpose(0, 1, 3, 2)
+    return path
+
+class OnnxInferenceSession():
+    def __init__(self, path, Providers = ["CPUExecutionProvider"]):
+        self.enc = ort.InferenceSession(path["enc"], providers=Providers)
+        self.emb_g = ort.InferenceSession(path["emb_g"], providers=Providers)
+        self.dp = ort.InferenceSession(path["dp"], providers=Providers)
+        self.sdp = ort.InferenceSession(path["sdp"], providers=Providers)
+        self.flow = ort.InferenceSession(path["flow"], providers=Providers)
+        self.dec = ort.InferenceSession(path["dec"], providers=Providers)
+
+    def __call__(
+            self, 
+            seq,
+            tone,
+            language,
+            bert_zh,
+            bert_jp,
+            bert_en,
+            emo,
+            sid,
+            seed = 114514,
+            seq_noise_scale = 0.8,
+            sdp_noise_scale = 0.6,
+            length_scale = 1.,
+            sdp_ratio = 0.
+        ):
+        g = self.emb_g.run(None, {'sid': sid.astype(np.int64),})[0]
+        g = np.expand_dims(g, -1)
+        enc_rtn = self.enc.run(
+            None,
+            {
+                "x" : seq.astype(np.int64),
+                "t" : tone.astype(np.int64),
+                "language" : language.astype(np.int64),
+                "bert_0" : bert_zh.astype(np.float32),
+                "bert_1" : bert_jp.astype(np.float32),
+                "bert_2" : bert_en.astype(np.float32),
+                "emo" : emo.astype(np.float32),
+                "g" : g.astype(np.float32)
+            })
+        x, m_p, logs_p, x_mask = enc_rtn[0], enc_rtn[1], enc_rtn[2], enc_rtn[3]
+        np.random.seed(seed)
+        zinput = np.random.randn(x.shape[0], 2, x.shape[2]) * sdp_noise_scale
+        logw = self.sdp.run(None, {"x" : x, "x_mask" : x_mask, "zin" : zinput.astype(np.float32), "g" : g})[0] * (sdp_ratio) + \
+            self.dp.run(None, {"x" : x, "x_mask" : x_mask, "g" : g})[0] * (1 - sdp_ratio)
+        w = np.exp(logw) * x_mask * length_scale
+        w_ceil = np.ceil(w)
+        y_lengths = np.clip(np.sum(w_ceil, (1, 2)), a_min=1., a_max=100000).astype(np.int64)
+        y_mask = np.expand_dims(sequence_mask(y_lengths, None), 1)
+        attn_mask = np.expand_dims(x_mask, 2) * np.expand_dims(y_mask, -1)
+        attn = generate_path(w_ceil, attn_mask)
+        m_p = np.matmul(attn.squeeze(1), m_p.transpose(0, 2, 1)).transpose(
+            0, 2, 1
+        )  # [b, t', t], [b, t, d] -> [b, d, t']
+        logs_p = np.matmul(attn.squeeze(1), logs_p.transpose(0, 2, 1)).transpose(
+            0, 2, 1
+        )  # [b, t', t], [b, t, d] -> [b, d, t']
+        
+        z_p = m_p + np.random.randn(m_p.shape[0], m_p.shape[1], m_p.shape[2]) * np.exp(logs_p) * seq_noise_scale
+
+        z = self.flow.run(None, {"z_p" : z_p.astype(np.float32), "y_mask" : y_mask.astype(np.float32), "g": g})[0]
+
+        return self.dec.run(None, {"z_in" : z.astype(np.float32), "g": g})[0]