Zipformer Training Issues:Gradient too small and cuda out of memory issue

[bsshruthi22]

Hello All,
we are training a zip former model for about 3400 hours of Tamil data.
We were facing this issue:

RuntimeError:
grad_scale is too small, exiting: 1.4901161193847656e-08

We have NVIDIA A6000 50GB GPU.
So as per the suggestion in the terminal,we reduced the max duration from 1000,600,500,400,300,150
and Learning rate changed from 0.045 to 0.04,0.035,0.02.

Finally we used max duration is 150 and learning rate is 0.02 and trained for 4 epochs.
Then we faced cuda out of memory issue as below.
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 6.86 GiB (GPU 0; 47.54 GiB total capacity; 42.23 GiB already allocated; 2.80 GiB free; 43.04 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

Kindly suggest if any parameter changes are required so that the training can be continued.

[csukuangfj]

Could you use
https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/local/display_manifest_statistics.py#L48
to get the statistics of your data?

Sample output is given in

icefall/egs/librispeech/ASR/local/display_manifest_statistics.py

Lines 48 to 70 in 329e34a

	## train-clean-100
	Cuts count: 85617
	Total duration (hours): 303.8
	Speech duration (hours): 303.8 (100.0%)
	***
	Duration statistics (seconds):
	mean 12.8
	std 3.8
	min 1.3
	0.1% 1.9
	0.5% 2.2
	1% 2.5
	5% 4.2
	10% 6.4
	25% 11.4
	50% 13.8
	75% 15.3
	90% 16.7
	95% 17.3
	99% 18.1
	99.5% 18.4
	99.9% 18.8
	max 27.2

[csukuangfj]

By the way, have you enabled

icefall/egs/librispeech/ASR/zipformer/train.py

Line 1339 in 329e34a

train_cuts = train_cuts.filter(remove_short_and_long_utt)

and if yes, what are the thresholds you are using?

[epicyclism]

I think I had the CUDA problem you report, reading the error and PyTorch documentation suggested adding

export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

before starting to train. This worked for me.

[bsshruthi22]

Could you use https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/local/display_manifest_statistics.py#L48 to get the statistics of your data?

Sample output is given in

icefall/egs/librispeech/ASR/local/display_manifest_statistics.py

Lines 48 to 70 in 329e34a

	## train-clean-100
	Cuts count: 85617
	Total duration (hours): 303.8
	Speech duration (hours): 303.8 (100.0%)
	***
	Duration statistics (seconds):
	mean 12.8
	std 3.8
	min 1.3
	0.1% 1.9
	0.5% 2.2
	1% 2.5
	5% 4.2
	10% 6.4
	25% 11.4
	50% 13.8
	75% 15.3
	90% 16.7
	95% 17.3
	99% 18.1
	99.5% 18.4
	99.9% 18.8
	max 27.2

Attached the manifest statistics file
Manifests_statistics.docx

[bsshruthi22]

I think I had the CUDA problem you report, reading the error and PyTorch documentation suggested adding

export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

before starting to train. This worked for me.

This command was not available in the version of pytorch which I was using.So I used export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512

[bsshruthi22]

By the way, have you enabled

icefall/egs/librispeech/ASR/zipformer/train.py

Line 1339 in 329e34a

train_cuts = train_cuts.filter(remove_short_and_long_utt)

and if yes, what are the thresholds you are using?

I have not explicitly set any thresholds. It must be taking the default values.Will check where it is being set in lhotse.

[csukuangfj]

I have not explicitly set any thresholds. It must be taking the default values.Will check where it is being set in lhotse.

Default values are fine for your dataset as long as you have used train_cuts.filter in your code; otherwise, some very long cut may lead to OOM.

[bsshruthi22]

After resuming from 4th epoch,I again faced this error. I had set PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 6.86 GiB (GPU 0; 47.54 GiB total capacity; 43.98 GiB already allocated; 914.00 MiB free; 44.95 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

Please advise whether we can train in NVIDIA A6000 50 GB GPU ?
Last year we had trained about 245 hours of data in Quadro RTX 4000 8GB GPU for 30 epochs.There was no issue.

[csukuangfj]

Please advise whether we can train in NVIDIA A6000 50 GB GPU ?

Yes, you can. We are using 32GB V100.

---

Which sampler are you using? Are you using

icefall/egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py

Line 309 in 6f1abd8

train_sampler = DynamicBucketingSampler(

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 6.86 GiB (GPU 0; 47.54 GiB total capacity; 43.98 GiB already allocated; 914.00 MiB free; 44.95 GiB reserved in total by PyTorch)

Have you changed train.py or any other files? If yes, could you post the git diff?

[bsshruthi22]

I have not changed sampler ,I am using all default settings.i have only changed the sampling rate to 8k and train, dev ,test folder names

[bsshruthi22]

Please advise whether we can train in NVIDIA A6000 50 GB GPU ?

Yes, you can. We are using 32GB V100.

Which sampler are you using? Are you using

icefall/egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py

Line 309 in 6f1abd8

train_sampler = DynamicBucketingSampler(

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 6.86 GiB (GPU 0; 47.54 GiB total capacity; 43.98 GiB already allocated; 914.00 MiB free; 44.95 GiB reserved in total by PyTorch)

Have you changed train.py or any other files? If yes, could you post the git diff?
I have not changed sampler ,I am using all default settings.i have only changed the sampling rate to 8k and train, dev ,test folder names.
Any thing you can advise on this.