CUDA error - Githubissues

jiay7 commented 2 years ago

I have the following problems when using rnnt_loss_simple:

1655195372(1)

I checked the shape of the input tensor： 1655195453(1)

and the output of the code： 1655195474(1)

python=3.8.11 torch verison=1.10.1 cudatoolkit=10.2.89 CUDA version=10.2.

Is this a known issue? How can it be debugged and solved?

Thank you!

csukuangfj commented 2 years ago

What is your vocabulary size ?

jiay7 commented 2 years ago

I use this loss in my own RNNT model. The dictionary size is 4233

pkufool commented 2 years ago

could you post your vocab size?

From
Date	06/14/2022 16:32
To
Cc
Subject	[danpovey/fast_rnnt] CUDA error (Issue #7)

I have the following problems when using rnnt_loss_simple:

I checked the shape of the input tensor：

and the output of the code：

python=3.8.11 torch verison=1.10.1 cudatoolkit=10.2.89 CUDA version=10.2.

Is this a known issue? How can it be debugged and solved?

Thank you!

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.

I use char as the modeling unit, like:

1655196217(1)

dec_out_smoothed.shape MUST be (N, U, vocab_size)

Also, enc_out_smoothed.shape MUST be (N, T, vocab_size).

You are using 512 for the last dim of dec_out_smoothed and enc_out_smoothed, which is not correct.

Yes, you are right. I changed the output dim of lin_dec_smoothed and lin_enc_smoothed, and it works. Thank you.

k2-fsa / fast_rnnt