From 4c30cd6eb180998c4d01ac60b23ea3d2a6729fd7 Mon Sep 17 00:00:00 2001
From: 0x45f <wangzhen45@baidu.com>
Date: Thu, 23 Feb 2023 11:47:16 +0000
Subject: [PATCH] Support dy2st for VITS

---
 paddlespeech/t2s/exps/vits/synthesize_e2e.py |  4 ++--
 paddlespeech/t2s/models/vits/transform.py    | 19 +++++++++++--------
 2 files changed, 13 insertions(+), 10 deletions(-)

diff --git a/paddlespeech/t2s/exps/vits/synthesize_e2e.py b/paddlespeech/t2s/exps/vits/synthesize_e2e.py
index f9c8b3f71..eb3cad034 100644
--- a/paddlespeech/t2s/exps/vits/synthesize_e2e.py
+++ b/paddlespeech/t2s/exps/vits/synthesize_e2e.py
@@ -64,7 +64,6 @@ def evaluate(args):
     vits = VITS(idim=vocab_size, odim=odim, **config["model"])
     vits.set_state_dict(paddle.load(args.ckpt)["main_params"])
     vits.eval()
-    VITSInference
 
     vits_inference = VITSInference(vits)
     # whether dygraph to static
@@ -108,7 +107,8 @@ def evaluate(args):
                     spk_id = None
                     if spk_num is not None:
                         spk_id = paddle.to_tensor(args.spk_id)
-                    wav = vits_inference(text=part_phone_ids, sids=spk_id)
+                    # wav = vits_inference(text=part_phone_ids, sids=spk_id)
+                    wav = vits_inference(part_phone_ids)
                     if flags == 0:
                         wav_all = wav
                         flags = 1
diff --git a/paddlespeech/t2s/models/vits/transform.py b/paddlespeech/t2s/models/vits/transform.py
index 8c6227a6a..360453cca 100644
--- a/paddlespeech/t2s/models/vits/transform.py
+++ b/paddlespeech/t2s/models/vits/transform.py
@@ -82,8 +82,10 @@ def unconstrained_rational_quadratic_spline(
     inside_interval_mask = (inputs >= -tail_bound) & (inputs <= tail_bound)
     outside_interval_mask = ~inside_interval_mask
 
-    outputs = paddle.zeros(paddle.shape(inputs))
-    logabsdet = paddle.zeros(paddle.shape(inputs))
+    # outputs = paddle.zeros(paddle.shape(inputs))
+    # logabsdet = paddle.zeros(paddle.shape(inputs))
+    outputs = paddle.zeros(inputs.shape)
+    logabsdet = paddle.zeros(inputs.shape)
     if tails == "linear":
         unnormalized_derivatives = F.pad(
             unnormalized_derivatives,
@@ -140,15 +142,16 @@ def rational_quadratic_spline(
         min_bin_width=1e-3,
         min_bin_height=1e-3,
         min_derivative=1e-3, ):
-    if paddle.min(inputs) < left or paddle.max(inputs) > right:
-        raise ValueError("Input to a transform is not within its domain")
+
+    # if paddle.min(inputs) < left or paddle.max(inputs) > right:
+    #     raise ValueError("Input to a transform is not within its domain")
 
     num_bins = unnormalized_widths.shape[-1]
 
-    if min_bin_width * num_bins > 1.0:
-        raise ValueError("Minimal bin width too large for the number of bins")
-    if min_bin_height * num_bins > 1.0:
-        raise ValueError("Minimal bin height too large for the number of bins")
+    # if min_bin_width * num_bins > 1.0:
+    #     raise ValueError("Minimal bin width too large for the number of bins")
+    # if min_bin_height * num_bins > 1.0:
+    #     raise ValueError("Minimal bin height too large for the number of bins")
 
     widths = F.softmax(unnormalized_widths, axis=-1)
     widths = min_bin_width + (1 - min_bin_width * num_bins) * widths