[TTS]fix tacotron2 dygraph to static (#1414)

* fix tacotron2 dygraph to static , test=tts

* fix tacotron2 dygraph to static , test=tts

* simplify synthesize_e2e.py , test=tts

examples/csmsc/tts0/local/inference.sh

+#!/bin/bash
+
+train_output_path=$1
+
+stage=0
+stop_stage=0
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    python3 ${BIN_DIR}/../inference.py \
+        --inference_dir=${train_output_path}/inference \
+        --am=tacotron2_csmsc \
+        --voc=pwgan_csmsc \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/pd_infer_out \
+        --phones_dict=dump/phone_id_map.txt
+fi
+
+# for more GAN Vocoders
+# multi band melgan
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    python3 ${BIN_DIR}/../inference.py \
+        --inference_dir=${train_output_path}/inference \
+        --am=tacotron2_csmsc \
+        --voc=mb_melgan_csmsc \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/pd_infer_out \
+        --phones_dict=dump/phone_id_map.txt
+fi
+
+# style melgan
+# style melgan's Dygraph to Static Graph is not ready now
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    python3 ${BIN_DIR}/../inference.py \
+        --inference_dir=${train_output_path}/inference \
+        --am=tacotron2_csmsc \
+        --voc=style_melgan_csmsc \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/pd_infer_out \
+        --phones_dict=dump/phone_id_map.txt
+fi
+
+# hifigan
+if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+    python3 ${BIN_DIR}/../inference.py \
+        --inference_dir=${train_output_path}/inference \
+        --am=tacotron2_csmsc \
+        --voc=hifigan_csmsc \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/pd_infer_out \
+        --phones_dict=dump/phone_id_map.txt
+fi
\ No newline at end of file

examples/csmsc/tts0/local/synthesize_e2e.sh

@@ -22,8 +22,9 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
         --lang=zh \
         --text=${BIN_DIR}/../sentences.txt \
         --output_dir=${train_output_path}/test_e2e \
-        --inference_dir=${train_output_path}/inference \
-        --phones_dict=dump/phone_id_map.txt
+        --phones_dict=dump/phone_id_map.txt \
+        --inference_dir=${train_output_path}/inference
+        
 fi
 
 # for more GAN Vocoders

paddlespeech/t2s/exps/inference.py

@@ -33,7 +33,7 @@ def main():
         default='fastspeech2_csmsc',
         choices=[
             'speedyspeech_csmsc', 'fastspeech2_csmsc', 'fastspeech2_aishell3',
-            'fastspeech2_vctk'
+            'fastspeech2_vctk', 'tacotron2_csmsc'
         ],
         help='Choose acoustic model type of tts task.')
     parser.add_argument(

paddlespeech/t2s/exps/synthesize_e2e.py

@@ -178,10 +178,7 @@ def evaluate(args):
                 am_inference = jit.to_static(
                     am_inference,
                     input_spec=[InputSpec([-1], dtype=paddle.int64)])
-            paddle.jit.save(am_inference,
-                            os.path.join(args.inference_dir, args.am))
-            am_inference = paddle.jit.load(
-                os.path.join(args.inference_dir, args.am))
+
         elif am_name == 'speedyspeech':
             if am_dataset in {"aishell3", "vctk"} and args.speaker_dict:
                 am_inference = jit.to_static(
@@ -200,10 +197,13 @@ def evaluate(args):
                         InputSpec([-1], dtype=paddle.int64)
                     ])
 
-            paddle.jit.save(am_inference,
-                            os.path.join(args.inference_dir, args.am))
-            am_inference = paddle.jit.load(
-                os.path.join(args.inference_dir, args.am))
+        elif am_name == 'tacotron2':
+            am_inference = jit.to_static(
+                am_inference, input_spec=[InputSpec([-1], dtype=paddle.int64)])
+
+        paddle.jit.save(am_inference, os.path.join(args.inference_dir, args.am))
+        am_inference = paddle.jit.load(
+            os.path.join(args.inference_dir, args.am))
 
         # vocoder
         voc_inference = jit.to_static(

paddlespeech/t2s/models/new_tacotron2/tacotron2.py

@@ -432,6 +432,7 @@ class Tacotron2(nn.Layer):
 
         # inference
         h = self.enc.inference(x)
+
         if self.spk_num is not None:
             sid_emb = self.sid_emb(spk_id.reshape([-1]))
             h = h + sid_emb

paddlespeech/t2s/modules/tacotron2/attentions.py

@@ -157,7 +157,7 @@ class AttLoc(nn.Layer):
         paddle.Tensor  
             previous attention weights (B, T_max)
         """
-        batch = len(enc_hs_pad)
+        batch = paddle.shape(enc_hs_pad)[0]
         # pre-compute all h outside the decoder loop
         if self.pre_compute_enc_h is None or self.han_mode:
             # (utt, frame, hdim)
@@ -172,33 +172,30 @@ class AttLoc(nn.Layer):
             dec_z = dec_z.reshape([batch, self.dunits])
 
         # initialize attention weight with uniform dist.
-        if att_prev is None:
+        if paddle.sum(att_prev) == 0:
             # if no bias, 0 0-pad goes 0
-
             att_prev = 1.0 - make_pad_mask(enc_hs_len)
             att_prev = att_prev / enc_hs_len.unsqueeze(-1)
 
         # att_prev: (utt, frame) -> (utt, 1, 1, frame)
         # -> (utt, att_conv_chans, 1, frame)
-
         att_conv = self.loc_conv(att_prev.reshape([batch, 1, 1, self.h_length]))
         # att_conv: (utt, att_conv_chans, 1, frame) -> (utt, frame, att_conv_chans)
         att_conv = att_conv.squeeze(2).transpose([0, 2, 1])
         # att_conv: (utt, frame, att_conv_chans) -> (utt, frame, att_dim)
         att_conv = self.mlp_att(att_conv)
-
-        # dec_z_tiled: (utt, frame, att_dim)
+        # dec_z_tiled: (utt, frame, att_dim)        
         dec_z_tiled = self.mlp_dec(dec_z).reshape([batch, 1, self.att_dim])
 
         # dot with gvec
         # (utt, frame, att_dim) -> (utt, frame)
-        e = self.gvec(
-            paddle.tanh(att_conv + self.pre_compute_enc_h +
-                        dec_z_tiled)).squeeze(2)
+        e = paddle.tanh(att_conv + self.pre_compute_enc_h + dec_z_tiled)
+        e = self.gvec(e).squeeze(2)
 
         # NOTE: consider zero padding when compute w.
         if self.mask is None:
             self.mask = make_pad_mask(enc_hs_len)
+
         e = masked_fill(e, self.mask, -float("inf"))
         # apply monotonic attention constraint (mainly for TTS)
         if last_attended_idx is not None:
@@ -211,7 +208,6 @@ class AttLoc(nn.Layer):
         # utt x hdim
         c = paddle.sum(
             self.enc_h * w.reshape([batch, self.h_length, 1]), axis=1)
-
         return c, w
 
 

paddlespeech/t2s/modules/tacotron2/decoder.py

@@ -15,7 +15,6 @@
 """Tacotron2 decoder related modules."""
 import paddle
 import paddle.nn.functional as F
-import six
 from paddle import nn
 
 from paddlespeech.t2s.modules.tacotron2.attentions import AttForwardTA
@@ -59,7 +58,7 @@ class Prenet(nn.Layer):
         super().__init__()
         self.dropout_rate = dropout_rate
         self.prenet = nn.LayerList()
-        for layer in six.moves.range(n_layers):
+        for layer in range(n_layers):
             n_inputs = idim if layer == 0 else n_units
             self.prenet.append(
                 nn.Sequential(nn.Linear(n_inputs, n_units), nn.ReLU()))
@@ -78,7 +77,7 @@ class Prenet(nn.Layer):
             Batch of output tensors (B, ..., odim).
 
         """
-        for i in six.moves.range(len(self.prenet)):
+        for i in range(len(self.prenet)):
             # F.dropout 引入了随机, tacotron2 的 dropout 是不能去掉的
             x = F.dropout(self.prenet[i](x))
         return x
@@ -129,7 +128,7 @@ class Postnet(nn.Layer):
         """
         super().__init__()
         self.postnet = nn.LayerList()
-        for layer in six.moves.range(n_layers - 1):
+        for layer in range(n_layers - 1):
             ichans = odim if layer == 0 else n_chans
             ochans = odim if layer == n_layers - 1 else n_chans
             if use_batch_norm:
@@ -196,7 +195,7 @@ class Postnet(nn.Layer):
             Batch of padded output tensor. (B, odim, Tmax).
 
         """
-        for i in six.moves.range(len(self.postnet)):
+        for i in range(len(self.postnet)):
             xs = self.postnet[i](xs)
         return xs
 
@@ -360,7 +359,7 @@ class Decoder(nn.Layer):
         # define lstm network
         prenet_units = prenet_units if prenet_layers != 0 else odim
         self.lstm = nn.LayerList()
-        for layer in six.moves.range(dlayers):
+        for layer in range(dlayers):
             iunits = idim + prenet_units if layer == 0 else dunits
             lstm = nn.LSTMCell(iunits, dunits)
             if zoneout_rate > 0.0:
@@ -437,47 +436,50 @@ class Decoder(nn.Layer):
         # initialize hidden states of decoder
         c_list = [self._zero_state(hs)]
         z_list = [self._zero_state(hs)]
-        for _ in six.moves.range(1, len(self.lstm)):
-            c_list += [self._zero_state(hs)]
-            z_list += [self._zero_state(hs)]
+        for _ in range(1, len(self.lstm)):
+            c_list.append(self._zero_state(hs))
+            z_list.append(self._zero_state(hs))
         prev_out = paddle.zeros([paddle.shape(hs)[0], self.odim])
 
         # initialize attention
-        prev_att_w = None
+        prev_att_ws = []
+        prev_att_w = paddle.zeros(paddle.shape(hlens))
+        prev_att_ws.append(prev_att_w)
         self.att.reset()
 
         # loop for an output sequence
         outs, logits, att_ws = [], [], []
         for y in ys.transpose([1, 0, 2]):
             if self.use_att_extra_inputs:
-                att_c, att_w = self.att(hs, hlens, z_list[0], prev_att_w,
+                att_c, att_w = self.att(hs, hlens, z_list[0], prev_att_ws[-1],
                                         prev_out)
             else:
-                att_c, att_w = self.att(hs, hlens, z_list[0], prev_att_w)
+                att_c, att_w = self.att(hs, hlens, z_list[0], prev_att_ws[-1])
             prenet_out = self.prenet(
                 prev_out) if self.prenet is not None else prev_out
             xs = paddle.concat([att_c, prenet_out], axis=1)
             # we only use the second output of LSTMCell in paddle
             _, next_hidden = self.lstm[0](xs, (z_list[0], c_list[0]))
             z_list[0], c_list[0] = next_hidden
-            for i in six.moves.range(1, len(self.lstm)):
+            for i in range(1, len(self.lstm)):
                 # we only use the second output of LSTMCell in paddle
                 _, next_hidden = self.lstm[i](z_list[i - 1],
                                               (z_list[i], c_list[i]))
                 z_list[i], c_list[i] = next_hidden
             zcs = (paddle.concat([z_list[-1], att_c], axis=1)
                    if self.use_concate else z_list[-1])
-            outs += [
-                self.feat_out(zcs).reshape([paddle.shape(hs)[0], self.odim, -1])
-            ]
-            logits += [self.prob_out(zcs)]
-            att_ws += [att_w]
+            outs.append(
+                self.feat_out(zcs).reshape([paddle.shape(hs)[0], self.odim, -1
+                                            ]))
+            logits.append(self.prob_out(zcs))
+            att_ws.append(att_w)
             # teacher forcing
             prev_out = y
-            if self.cumulate_att_w and prev_att_w is not None:
+            if self.cumulate_att_w and paddle.sum(prev_att_w) != 0:
                 prev_att_w = prev_att_w + att_w  # Note: error when use +=
             else:
                 prev_att_w = att_w
+            prev_att_ws.append(prev_att_w)
         # (B, Lmax)
         logits = paddle.concat(logits, axis=1)
         # (B, odim, Lmax) 
@@ -552,6 +554,7 @@ class Decoder(nn.Layer):
     .. _`Deep Voice 3`: https://arxiv.org/abs/1710.07654
         """
         # setup
+
         assert len(paddle.shape(h)) == 2
         hs = h.unsqueeze(0)
         ilens = paddle.shape(h)[0]
@@ -561,13 +564,16 @@ class Decoder(nn.Layer):
         # initialize hidden states of decoder
         c_list = [self._zero_state(hs)]
         z_list = [self._zero_state(hs)]
-        for _ in six.moves.range(1, len(self.lstm)):
-            c_list += [self._zero_state(hs)]
-            z_list += [self._zero_state(hs)]
+        for _ in range(1, len(self.lstm)):
+            c_list.append(self._zero_state(hs))
+            z_list.append(self._zero_state(hs))
         prev_out = paddle.zeros([1, self.odim])
 
         # initialize attention
-        prev_att_w = None
+        prev_att_ws = []
+        prev_att_w = paddle.zeros([ilens])
+        prev_att_ws.append(prev_att_w)
+
         self.att.reset()
 
         # setup for attention constraint
@@ -579,6 +585,7 @@ class Decoder(nn.Layer):
         # loop for an output sequence
         idx = 0
         outs, att_ws, probs = [], [], []
+        prob = paddle.zeros([1])
         while True:
             # updated index
             idx += self.reduction_factor
@@ -589,7 +596,7 @@ class Decoder(nn.Layer):
                     hs,
                     ilens,
                     z_list[0],
-                    prev_att_w,
+                    prev_att_ws[-1],
                     prev_out,
                     last_attended_idx=last_attended_idx,
                     backward_window=backward_window,
@@ -599,19 +606,20 @@ class Decoder(nn.Layer):
                     hs,
                     ilens,
                     z_list[0],
-                    prev_att_w,
+                    prev_att_ws[-1],
                     last_attended_idx=last_attended_idx,
                     backward_window=backward_window,
                     forward_window=forward_window, )
 
-            att_ws += [att_w]
+            att_ws.append(att_w)
             prenet_out = self.prenet(
                 prev_out) if self.prenet is not None else prev_out
             xs = paddle.concat([att_c, prenet_out], axis=1)
             # we only use the second output of LSTMCell in paddle
             _, next_hidden = self.lstm[0](xs, (z_list[0], c_list[0]))
+
             z_list[0], c_list[0] = next_hidden
-            for i in six.moves.range(1, len(self.lstm)):
+            for i in range(1, len(self.lstm)):
                 # we only use the second output of LSTMCell in paddle
                 _, next_hidden = self.lstm[i](z_list[i - 1],
                                               (z_list[i], c_list[i]))
@@ -619,38 +627,38 @@ class Decoder(nn.Layer):
             zcs = (paddle.concat([z_list[-1], att_c], axis=1)
                    if self.use_concate else z_list[-1])
             # [(1, odim, r), ...]
-            outs += [self.feat_out(zcs).reshape([1, self.odim, -1])]
+            outs.append(self.feat_out(zcs).reshape([1, self.odim, -1]))
+
+            prob = F.sigmoid(self.prob_out(zcs))[0]
+            probs.append(prob)
 
-            # [(r), ...]
-            probs += [F.sigmoid(self.prob_out(zcs))[0]]
             if self.output_activation_fn is not None:
                 prev_out = self.output_activation_fn(
                     outs[-1][:, :, -1])  # (1, odim)
             else:
                 prev_out = outs[-1][:, :, -1]  # (1, odim)
-            if self.cumulate_att_w and prev_att_w is not None:
+            if self.cumulate_att_w and paddle.sum(prev_att_w) != 0:
                 prev_att_w = prev_att_w + att_w  # Note: error when use +=
             else:
                 prev_att_w = att_w
+            prev_att_ws.append(prev_att_w)
             if use_att_constraint:
                 last_attended_idx = int(att_w.argmax())
 
-            # check whether to finish generation
-            if sum(paddle.cast(probs[-1] >= threshold,
-                               'int64')) > 0 or idx >= maxlen:
+            if prob >= threshold or idx >= maxlen:
                 # check mininum length
                 if idx < minlen:
                     continue
-                # (1, odim, L)
-                outs = paddle.concat(outs, axis=2)
-                if self.postnet is not None:
-                    # (1, odim, L)
-                    outs = outs + self.postnet(outs)
-                # (L, odim)
-                outs = outs.transpose([0, 2, 1]).squeeze(0)
-                probs = paddle.concat(probs, axis=0)
-                att_ws = paddle.concat(att_ws, axis=0)
                 break
+        # (1, odim, L)
+        outs = paddle.concat(outs, axis=2)
+        if self.postnet is not None:
+            # (1, odim, L)
+            outs = outs + self.postnet(outs)
+        # (L, odim)
+        outs = outs.transpose([0, 2, 1]).squeeze(0)
+        probs = paddle.concat(probs, axis=0)
+        att_ws = paddle.concat(att_ws, axis=0)
 
         if self.output_activation_fn is not None:
             outs = self.output_activation_fn(outs)
@@ -685,9 +693,9 @@ class Decoder(nn.Layer):
         # initialize hidden states of decoder
         c_list = [self._zero_state(hs)]
         z_list = [self._zero_state(hs)]
-        for _ in six.moves.range(1, len(self.lstm)):
-            c_list += [self._zero_state(hs)]
-            z_list += [self._zero_state(hs)]
+        for _ in range(1, len(self.lstm)):
+            c_list.append(self._zero_state(hs))
+            z_list.append(self._zero_state(hs))
         prev_out = paddle.zeros([paddle.shape(hs)[0], self.odim])
 
         # initialize attention
@@ -702,14 +710,14 @@ class Decoder(nn.Layer):
                                         prev_out)
             else:
                 att_c, att_w = self.att(hs, hlens, z_list[0], prev_att_w)
-            att_ws += [att_w]
+            att_ws.append(att_w)
             prenet_out = self.prenet(
                 prev_out) if self.prenet is not None else prev_out
             xs = paddle.concat([att_c, prenet_out], axis=1)
             # we only use the second output of LSTMCell in paddle
             _, next_hidden = self.lstm[0](xs, (z_list[0], c_list[0]))
             z_list[0], c_list[0] = next_hidden
-            for i in six.moves.range(1, len(self.lstm)):
+            for i in range(1, len(self.lstm)):
                 z_list[i], c_list[i] = self.lstm[i](z_list[i - 1],
                                                     (z_list[i], c_list[i]))
             # teacher forcing

paddlespeech/t2s/modules/tacotron2/encoder.py

@@ -14,7 +14,6 @@
 # Modified from espnet(https://github.com/espnet/espnet)
 """Tacotron2 encoder related modules."""
 import paddle
-import six
 from paddle import nn
 
 
@@ -88,7 +87,7 @@ class Encoder(nn.Layer):
 
         if econv_layers > 0:
             self.convs = nn.LayerList()
-            for layer in six.moves.range(econv_layers):
+            for layer in range(econv_layers):
                 ichans = (embed_dim if layer == 0 and input_layer == "embed"
                           else econv_chans)
                 if use_batch_norm:
@@ -130,6 +129,7 @@ class Encoder(nn.Layer):
                 direction='bidirectional',
                 bias_ih_attr=True,
                 bias_hh_attr=True)
+            self.blstm.flatten_parameters()
         else:
             self.blstm = None
 
@@ -157,7 +157,7 @@ class Encoder(nn.Layer):
         """
         xs = self.embed(xs).transpose([0, 2, 1])
         if self.convs is not None:
-            for i in six.moves.range(len(self.convs)):
+            for i in range(len(self.convs)):
                 if self.use_residual:
                     xs += self.convs[i](xs)
                 else:
@@ -167,7 +167,8 @@ class Encoder(nn.Layer):
         if not isinstance(ilens, paddle.Tensor):
             ilens = paddle.to_tensor(ilens)
         xs = xs.transpose([0, 2, 1])
-        self.blstm.flatten_parameters()
+        # for dygraph to static graph
+        # self.blstm.flatten_parameters()
         # (B, Tmax, C)
         # see https://www.paddlepaddle.org.cn/documentation/docs/zh/faq/train_cn.html#paddletorch-nn-utils-rnn-pack-padded-sequencetorch-nn-utils-rnn-pad-packed-sequenceapi
         xs, _ = self.blstm(xs, sequence_length=ilens)
@@ -191,6 +192,6 @@ class Encoder(nn.Layer):
 
         """
         xs = x.unsqueeze(0)
-        ilens = paddle.to_tensor([x.shape[0]])
+        ilens = paddle.shape(x)[0]
 
         return self.forward(xs, ilens)[0][0]