Comments (12)
Hi @sjfleming
I have emailed you the file.
Best,
Devika
from cellbender.
Hi Stephen,
Thats completely understandable. I have now got access to a GPU to run my samples so i havent had a problem running them on the GPU. I did have to re-run some samples with different parameters as training dint converge, but no errors otherwise.
thanks,
Devika
from cellbender.
Hi Devika,
It does look like you are having some issues running the tool on this dataset. Running with the default z-dim and z-layers normally does not result in the kind of learning curve you have attached here. The big spike near epoch 75 and the overall wobbly look are bad signs. I have never seen this before with z-dim 20 and z-layers 500.
The other confusing thing is that cellbender should run deterministically, and there certainly should not be different behavior based on the number of epochs. Again, I haven't seen this before.
If you are willing to share this h5 file with me, I can take a look and try to debug.
What happens if you use --total-droplets-included 40000
?
from cellbender.
Hi @sjfleming
I reran the analysis with only -total-droplets-included 40000 reading previous posts perhaps i need to reduce this from 60,000 and even then it always fails with default z-dims and z-layers with the same error as before : "Encountered NaN loss" (see attached error report
MBR1_v3_error.txt
. But when i ran it with either increasing the z-dims and z-layers and or decreasing them it worked fine. However i dont know which result has converged the best ( i suspect decreasing the z-layers version) and also what was the problem.
I would be happy for you to have the h5 file. But do you mean the cell ranger output h5 file or do you mean the cellbender out put from my first version with the problematic result? I wasnt sure.
See attached my log and pdf report for try 1 (decrease z-layers with total cells 40,000
MaleBrainRep1_v2.log
MaleBrainRep1_v2.pdf
) and try 2(increased z-layers and z-dims with total cells as 40,000)
MaleBrainRep1_v4.log
MaleBrainRep1_v4.pdf
So would really like some advice on this.
Devika
from cellbender.
Yes, if you could email me the CellRanger raw output h5 file at
[email protected]
then I can take a quick look. I would like to understand what's causing these NaNs, because I don't usually see this behavior.
For now I would guess that decreasing z-layers (your try 1) would probably produce a better result, but I'll know more once I take a look.
from cellbender.
Sure thing. I did attach my reports and logs for each try as well in my previous post.
from cellbender.
I don't seem to be able to reproduce the issue you had with the NaN loss. It seems to run just fine for me. I'm not sure what has caused the issue here... The only thing I've done differently is to include the --cuda
flag, but that should make no difference on the computations that get carried out.
I will send you my version of the output h5 file, so you can compare.
from cellbender.
Hi Stephen,
Thank you for having a look on your end. I think maybe it might have to do with the way I had to install Cellbender in the conda environment on the server. As our server OS is Centos 6, i kept getting the GlibC library error even within the Conda environment for pytorch.
I used this fix detailed (link: https://gist.github.com/michaelchughes/85287f1c6f6440c060c3d86b4e7d764b) to compile my own GLIBC libraries and then get conda to recognize it.
I mean initially i thought it was my install. But considering other samples ran fine it couldnt be.
We should be getting an update in the coming months. But in the meanwhile , i dint find anyother solutions for the GLIBC errors for pytorch on Centos 6
if you could send me you output to compare that would be great.
Best,
Devika
from cellbender.
Hi @sjfleming
Just to give you an update. I ran my sample that was giving me trouble on a Centos 7 OS , where CellBender was installed using the manual install through a conda environment. I had no trouble while installing and everything worked fine. But the sample still gave the same error of UserWarning: Encountered NaN: loss".
maybe it needs to be run on a GPU. but on a clean install on a CPU system this error is recurring.
Thanks,
Devika
from cellbender.
Hi Devika,
I will look into this when I get a chance. So far, the GPU performance has been the priority since it takes so long to run on a CPU. But I definitely do not want it to error out on CPU. When I'm testing v2, I will test CPU performance as well.
from cellbender.
@deevdevil88 is this issue resolved?
from cellbender.
from cellbender.
Related Issues (20)
- cellbender v3.0 doesn't generate most of the output files, but doesn't have any errors
- "Trying to use CUDA, " \ AssertionError: Trying to use CUDA, but CUDA is not available.
- Number of cells after cellbender much more than number of cells from cellranger (filtered)
- New fileformat output from BD rhapsody HOT 1
- Should I keep decreasing the learning rate?
- Question about input h5 file HOT 1
- Computing the output in asynchronous chunks in parallel takes longer than 144 hours
- Unhandled division by zero
- Never mind
- Can't Computing target noise counts per gene for MCKP estimator HOT 1
- Importance of model loss
- Background Fraction HOT 2
- OOM posterior inference for chimeric sample even using --posterior-batch-size 1
- Not saving ckpt.tar.gz checkpoint HOT 7
- Fixed Single cell RNA seq HOT 1
- Cellbender on multiplexed chemistries
- RuntimeError: CUDA driver error: invalid argument - Google Container Registry (GCR)
- CellBender for .BAM file
- Increased number of UMIs per barcode after running remove-background HOT 1
- Mismatch between summary for algorithm convergence and learning curve
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from cellbender.