add synthesis script for pwg

examples/parallelwave_gan/baker/synthesize.py

@@ -0,0 +1,81 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import sys
+import logging
+import argparse
+from pathlib import Path
+
+import yaml
+import jsonlines
+import paddle
+import numpy as np
+import soundfile as sf
+from paddle import distributed as dist
+
+from parakeet.datasets.data_table import DataTable
+from parakeet.models.parallel_wavegan import PWGGenerator
+
+from config import get_cfg_default
+
+parser = argparse.ArgumentParser(
+    description="synthesize with parallel wavegan.")
+parser.add_argument(
+    "--config", type=str, help="config file to overwrite default config")
+parser.add_argument("--params", type=str, help="generator parameter file")
+parser.add_argument("--test-metadata", type=str, help="dev data")
+parser.add_argument("--output-dir", type=str, help="output dir")
+parser.add_argument("--verbose", type=int, default=1, help="verbose")
+
+args = parser.parse_args()
+config = get_cfg_default()
+if args.config:
+    config.merge_from_file(args.config)
+
+print("========Args========")
+print(yaml.safe_dump(vars(args)))
+print("========Config========")
+print(config)
+print(
+    f"master see the word size: {dist.get_world_size()}, from pid: {os.getpid()}"
+)
+
+generator = PWGGenerator(**config["generator_params"])
+state_dict = paddle.load(args.params)
+generator.set_state_dict(state_dict)
+
+generator.remove_weight_norm()
+generator.eval()
+with jsonlines.open(args.test_metadata, 'r') as reader:
+    metadata = list(reader)
+
+test_dataset = DataTable(
+    metadata,
+    fields=['utt_id', 'feats'],
+    converters={
+        'utt_id': None,
+        'feats': np.load,
+    })
+output_dir = Path(args.output_dir)
+output_dir.mkdir(parents=True, exist_ok=True)
+
+for example in test_dataset:
+    utt_id = example['utt_id']
+    mel = example['feats']
+    mel = paddle.to_tensor(mel)  # (T, C)
+    wav = generator.inference(c=mel)
+    wav = wav.numpy()
+    print(f"{utt_id}, mel: {mel.shape}, wave: {wav.shape}")
+    sf.write(output_dir / (utt_id + ".wav"), wav, samplerate=24000)

examples/parallelwave_gan/baker/train.py

@@ -134,7 +134,8 @@ def train_sp(args, config):
         parameters=generator.parameters(),
         **config["generator_optimizer_params"])
     lr_schedule_d = StepDecay(**config["discriminator_scheduler_params"])
-    gradient_clip_d = nn.ClipGradByGlobalNorm(config["discriminator_grad_norm"])
+    gradient_clip_d = nn.ClipGradByGlobalNorm(config[
+        "discriminator_grad_norm"])
     optimizer_d = Adam(
         learning_rate=lr_schedule_d,
         grad_clip=gradient_clip_d,
@@ -180,7 +181,6 @@ def train_sp(args, config):
         },
         dataloader=dev_dataloader,
         lambda_adv=config.lambda_adv, )
-
     trainer = Trainer(
         updater,
         stop_trigger=(config.train_max_steps, "iteration"),

parakeet/models/parallel_wavegan.py

@@ -519,7 +519,7 @@ class PWGGenerator(nn.Layer):
 
         if c is not None:
             c = paddle.transpose(c, [1, 0]).unsqueeze(0)  # pseudo batch
-            c = nn.Pad1D(self.aux_context_window, mode='edge')(c)
+            c = nn.Pad1D(self.aux_context_window, mode='replicate')(c)
         out = self.forward(x, c).squeeze(0).transpose([1, 0])
         return out