search

PaddlePaddle / Parakeet

PAddle PARAllel text-to-speech toolKIT (supporting Tacotron2, Transformer TTS, FastSpeech2/FastPitch, SpeedySpeech, WaveFlow and Parallel WaveGAN)
Other
598 stars 83 forks source link

训练deepvoice3时，用了自己batch_size=4训练的waveflow,报了tensor的shape错误。请问下应该在哪个地方修改shape #29

Closed Ficery1 closed 3 years ago

Ficery1 commented 4 years ago

Traceback (most recent call last): File "E:/AI/Parakeet/examples/deepvoice3/train.py", line 187, in train(args, config) File "E:/AI/Parakeet/examples/deepvoice3/train.py", line 137, in train vocoder = WaveflowVocoder() File "E:\AI\Parakeet\examples\deepvoice3\vocoder.py", line 24, in init load_parameters(self.model, checkpoint_path=checkpoint_path) File "E:\AI\Parakeet\parakeet\utils\io.py", line 135, in load_parameters model.set_dict(model_dict) File "E:\setup\anaconda\envs\Parakeet\lib\site-packages\paddle\fluid\dygraph\layers.py", line 665, in set_dict use_structured_name=use_structured_name) File "E:\setup\anaconda\envs\Parakeet\lib\site-packages\paddle\fluid\dygraph\layers.py", line 705, in load_dict para.set_value(stat_dict[key_name]) File "", line 2, in set_value File "E:\setup\anaconda\envs\Parakeet\lib\site-packages\paddle\fluid\wrapped_decorator.py", line 25, in impl return wrapped_func(*args, *kwargs) File "E:\setup\anaconda\envs\Parakeet\lib\site-packages\paddle\fluid\framework.py", line 216, in impl return func(args, **kwargs) File "E:\setup\anaconda\envs\Parakeet\lib\site-packages\paddle\fluid\dygraph\varbase_patch_methods.py", line 66, in set_value self.name, self_tensor_np.shape, value_np.shape) AssertionError: Variable Shape not match, Variable [ weight_norm_wrapper_2.w_0 ] need tensor with shape (128, 1, 1, 1) but load set tensor with shape (64, 1, 1, 1) W0728 16:56:40.347012 11192 device_context.cc:252] Please NOTE: device: 0, CUDA Capability: 52, Driver API Version: 10.1, Runtime API Version: 10.0 W0728 16:56:40.487767 11192 device_context.cc:260] device: 0, cuDNN Version: 7.6.

iclementine commented 4 years ago

看来应该是 deep voice 3 的 train代码中用的 waveflow 是 residual channel = 128 的。你目前训练的 是 residual channel 为 64 的。

因为它在加载参数的时候出了问题。

WaveflowVocoder 的初始化方式是通过一个配置文件构建模型，然后加载训练好的参数。这个要对应才能正确加载参数。可以查看 examples/deepvoice3/vocoder.py

图片
Ficery1 commented 4 years ago

看来应该是deep voice 3的trains中用的waveflow是剩余频道= 128的。你目前训练的是剩余频道为64的。

因为它在加载参数的时候出了问题。

WaveflowVocoder的初始化方式是通过一个配置文件构建模型，然后加载训练好的参数。这个要对应的能力正确加载参数。可以查看 examples/deepvoice3/vocoder.py

图片

谢谢。找到了问题。项目中wavefolw的配置文件中n_channels默认是64。我在使用deepvoice3时 是下载的预训练好的参数和配置文件，该配置文件中n_channels默认是128.