The provided c3d feature are extracted after 5 iterations of training?

[PanXiebit]

Thanks for your impressive work and the open source code @Jevin754.

I am a little confused with the C3D features you provided in google/baidu drive. Its filename is "c3d_res_phoenix_body_iter5_120k.tar.gz" in google drive. Thus, is the feature extracted after 5 iterations of training(trained with pseudo-labels in your paper)?

I implement with your provided training scripts and C3D features. But I get worse results, "[Relaxation Evaluation] Epoch: 50, DEV WER: 40.3". Therefore, I feel puzzled about this C3D feature, do I need to implement alignment and iterative training by myself?

Hope for your reply.

[PanXiebit]

I find out where the problem is. The results of ctcdecoder executed on GPU and CPU are quite different. I changed the calculation of ctc_loss on cpu, and get desired result.

[DuckJ]

Do you know how to get the features of c3d？
@PanXiebit @Jevin754