Closed lichuanqi closed 1 year ago
各位中秋快乐啊! :blush:
模型定义中的encoder.embed.out.0.weight与pre-trained model中的不同。 请检查模型的配置文件是否和pre-trained model一致以及 paddlespeech/s2t/modules/subsampling.py 中相关参数
感谢回复。模型文件model.yaml和权重文件wenetspeech.pdparams是从同一个的链接的压缩包下载下来的,应该是匹配的,model.yaml中有一个参数 input_layer
是可选参数,最终会传入paddlespeech/s2t/modules/subsampling.py,试了 conv2d, conv2d6 and conv2d8 都是一样的报错。
有可能是系统的问题吗,我目前是在Win上测试的,单个语音文件的预测是可以跑通的。我也试一下Ubuntu下是否有报错。
模型定义中的encoder.embed.out.0.weight与pre-trained model中的不同。 请检查模型的配置文件是否和pre-trained model一致以及 paddlespeech/s2t/modules/subsampling.py 中相关参数
在linux上测试吧,我们已ubuntu16的为准。
在linux上测试吧,我们已ubuntu16的为准。
好的,我测试下再来反馈,感谢回复。
在Ubuntu上测试了一下还是一样的报错,
环境如下:
* Ubuntu 20.04
* python 3.7
* paddlepaddle 2.3.2
* CUDA 10.2
* cuDNN 7.6
但是相同的环境,deepspeech2 的预测、模型导出测试都可以跑通。
请尝试一下develop版本是否会存在这个问题?
This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.
This issue is closed. Please re-open if needed.
Hello,我在使用conformer_wenetspeech模型测试时发生Skip loading for encoder.embed.out.0.weight.的报错, 猜测是某一层的权重维度与模型不匹配,研究了很长时间都没找到错误原因,请帮忙指点下啊,感谢。
报错和日志如下,
报错:
日志: py_eval.LAPTOP-CD0ILM5K.lichuan.2022-09-09_16-16-08_948182.log