fix vits reduce_sum's input/output dtype, test=tts (#3028)

Author: yt605155624

paddlespeech/t2s/models/vits/duration_predictor.py

@@ -155,21 +155,19 @@ def forward(
             z_u, z1 = paddle.split(z_q, [1, 1], 1)
             u = F.sigmoid(z_u) * x_mask
             z0 = (w - u) * x_mask
-            logdet_tot_q += paddle.sum(
-                (F.log_sigmoid(z_u) + F.log_sigmoid(-z_u)) * x_mask, [1, 2])
-            logq = (paddle.sum(-0.5 *
-                               (math.log(2 * math.pi) +
-                                (e_q**2)) * x_mask, [1, 2]) - logdet_tot_q)
-
+            tmp1 = (F.log_sigmoid(z_u) + F.log_sigmoid(-z_u)) * x_mask
+            logdet_tot_q += paddle.sum(tmp1, [1, 2])
+            tmp2 = -0.5 * (math.log(2 * math.pi) + (e_q**2)) * x_mask
+            logq = (paddle.sum(tmp2, [1, 2]) - logdet_tot_q)
             logdet_tot = 0
             z0, logdet = self.log_flow(z0, x_mask)
             logdet_tot += logdet
             z = paddle.concat([z0, z1], 1)
             for flow in self.flows:
                 z, logdet = flow(z, x_mask, g=x, inverse=inverse)
                 logdet_tot = logdet_tot + logdet
-            nll = (paddle.sum(0.5 * (math.log(2 * math.pi) +
-                                     (z**2)) * x_mask, [1, 2]) - logdet_tot)
+            tmp3 = 0.5 * (math.log(2 * math.pi) + (z**2)) * x_mask
+            nll = (paddle.sum(tmp3, [1, 2]) - logdet_tot)
             # (B,)
             return nll + logq
         else:

paddlespeech/t2s/models/vits/generator.py

@@ -371,8 +371,9 @@ def forward(
             # (B, H, T_text)
             s_p_sq_r = paddle.exp(-2 * logs_p)
             # (B, 1, T_text)
+            tmp1 = -0.5 * math.log(2 * math.pi) - logs_p
             neg_x_ent_1 = paddle.sum(
-                -0.5 * math.log(2 * math.pi) - logs_p,
+                tmp1,
                 [1],
                 keepdim=True, )
             # (B, T_feats, H) x (B, H, T_text) = (B, T_feats, T_text)
@@ -384,8 +385,9 @@ def forward(
                 z_p.transpose([0, 2, 1]),
                 (m_p * s_p_sq_r), )
             # (B, 1, T_text)
+            tmp2 = -0.5 * (m_p**2) * s_p_sq_r
             neg_x_ent_4 = paddle.sum(
-                -0.5 * (m_p**2) * s_p_sq_r,
+                tmp2,
                 [1],
                 keepdim=True, )
             # (B, T_feats, T_text)
@@ -403,7 +405,6 @@ def forward(
         w = attn.sum(2)
         dur_nll = self.duration_predictor(x, x_mask, w=w, g=g)
         dur_nll = dur_nll / paddle.sum(x_mask)
-
         # expand the length to match with the feature sequence
         # (B, T_feats, T_text) x (B, T_text, H) -> (B, H, T_feats)
         m_p = paddle.matmul(attn.squeeze(1),
@@ -511,8 +512,9 @@ def inference(
             # (B, H, T_text)
             s_p_sq_r = paddle.exp(-2 * logs_p)
             # (B, 1, T_text)
+            tmp3 = -0.5 * math.log(2 * math.pi) - logs_p
             neg_x_ent_1 = paddle.sum(
-                -0.5 * math.log(2 * math.pi) - logs_p,
+                tmp3,
                 [1],
                 keepdim=True, )
             # (B, T_feats, H) x (B, H, T_text) = (B, T_feats, T_text)
@@ -524,8 +526,9 @@ def inference(
                 z_p.transpose([0, 2, 1]),
                 (m_p * s_p_sq_r), )
             # (B, 1, T_text)
+            tmp4 = -0.5 * (m_p**2) * s_p_sq_r
             neg_x_ent_4 = paddle.sum(
-                -0.5 * (m_p**2) * s_p_sq_r,
+                tmp4,
                 [1],
                 keepdim=True, )
             # (B, T_feats, T_text)

paddlespeech/t2s/models/vits/transform.py

@@ -61,8 +61,12 @@ def piecewise_rational_quadratic_transform(
 
 
 def mask_preprocess(x, mask):
+    # bins.dtype = int32
     B, C, T, bins = paddle.shape(x)
-    new_x = paddle.zeros([mask.sum(), bins])
+    mask_int = paddle.cast(mask, dtype='int64')
+    # paddle.sum 输入是 int32 或 bool 的时候，输出是 int64
+    # paddle.zeros (fill_constant) 的 shape 会被强制转成 int32 类型
+    new_x = paddle.zeros([paddle.sum(mask_int), bins])
     for i in range(bins):
         new_x[:, i] = x[:, :, :, i][mask]
     return new_x
@@ -240,4 +244,7 @@ def rational_quadratic_spline(
 
 def _searchsorted(bin_locations, inputs, eps=1e-6):
     bin_locations[..., -1] += eps
-    return paddle.sum(inputs[..., None] >= bin_locations, axis=-1) - 1
+    mask = inputs[..., None] >= bin_locations
+    mask_int = paddle.cast(mask, 'int64')
+    out = paddle.sum(mask_int, axis=-1) - 1
+    return out