i207m / pinnacle Goto Github PK
View Code? Open in Web Editor NEWCodebase for PINNacle: A Comprehensive Benchmark of Physics-Informed Neural Networks for Solving PDEs.
Home Page: https://arxiv.org/abs/2306.08827
Codebase for PINNacle: A Comprehensive Benchmark of Physics-Informed Neural Networks for Solving PDEs.
Home Page: https://arxiv.org/abs/2306.08827
Hi, could you please also provide your ref data for high dim?
Hi, we meet a problem when we are running the cloned code with the primary settings except changing the default method from 'adam' to 'laaf' or 'gaaf'. Here is the content of logerr.txt:
Traceback (most recent call last):
File "/scratch/tpang/yuanzhe_hu/TBv2-PINNacle/benchmark.py", line 153, in <module>
trainer.train_all()
File "/scratch/tpang/yuanzhe_hu/TBv2-PINNacle/trainer.py", line 98, in train_all
model = get_model()
File "/scratch/tpang/yuanzhe_hu/TBv2-PINNacle/benchmark.py", line 104, in get_model_dde
net = DNN_LAAF(len(parse_hidden_layers(command_args))-1, parse_hidden_layers[0], pde.input_dim, pde.output_dim)
TypeError: 'function' object is not subscriptable
Besides, how can we run the method 'gpinn' and 'hp-vpinn' since I cannot find the specific settings for these two methods in benchmark.py file ?
PINNacle is getting quite some attention from the community!
To allow contributions and clarify reusability, it would be highly beneficial to include a license. See here.
Hello Authors,
Thank you for the hard work on the benchmarks.
The demo code did not include any benchmark using FBPINN on different cases nor utilised multiAdam. Can you provide more documentation, please?
Thank you.
Good luck with NIPs!
Hello Authors,
How to open the dat file under the ref file?
Thank you.
Hello,
I am currently working on creating new sampling/resampling methods for PINNs. I decided to use your benchmark, because it has great number of PDEs and methods to compare!
However, I have already encountered many minor and major flaws in your code, because, of course, it is a very new work. And I am happy to help :) (i also hope that your work will be accepted for ICLR).
The main issue I encountering currently is about RAR-wrapper. I can't find myself the "root" or the inter-dependency of it, in order to correct the code. In becnhmark.py I change RAR parameter as:
line 120: model.train = rar_wrapper(pde, model, {"interval": 10, "count": 1})
When I change argument interval
to a number lower than log-every
, the following error occurs:
$ python benchmark_fast.py --method rar --log-every 100 --device cpu
Using backend: pytorch
Set the default float type to float32
***** Begin #0-0 *****
Compiling model...
'compile' took 0.000140 s
PDE Class Name: Burgers1D
Training model...
Step Train loss Test loss Test metric
0 [1.45e-02, 6.48e-01, 4.15e-02] [1.58e-02, 6.48e-01, 4.15e-02] []
10 [4.48e-04, 4.10e-01, 1.98e-02] [4.47e-04, 4.10e-01, 1.98e-02] []
Traceback (most recent call last):
File "/home/dymchens-ext/PINNacle/benchmark_fast.py", line 149, in <module>
trainer.train_all()
File "/home/dymchens-ext/PINNacle/trainer.py", line 99, in train_all
model.train(**train_args, model_save_path=save_path)
File "/home/dymchens-ext/PINNacle/src/utils/rar.py", line 19, in wrapper
train(*args, **kwargs)
File "/home/dymchens-ext/PINNacle/deepxde/utils/internal.py", line 22, in wrapper
result = f(*args, **kwargs)
File "/home/dymchens-ext/PINNacle/deepxde/model.py", line 603, in train
self.callbacks.on_train_end()
File "/home/dymchens-ext/PINNacle/deepxde/callbacks.py", line 94, in on_train_end
callback.on_train_end()
File "/home/dymchens-ext/PINNacle/src/utils/callbacks.py", line 208, in on_train_end
self.frmses[:, 0], self.frmses[:, 1], self.frmses[:, 2]]).T,
IndexError: too many indices for array: array is 1-dimensional, but 2 were indexed
I guess it has to do with the fact, that wrapper changes model
attribute iterations
to interval, and therefore the callback on_train_end
is called earlier than any logging was done, so there is no metrics calculated yet. I, though, don't know whether it is better to change the wrapper behaviour or the callbacks, or just to not log anything, or to restrict calculating default metrics (because, for example, I don't care about any except mse and residual). Currently, I just log every 10 iterations, and, of course, it is eating a lot of storage and as well logger is messy. In case of method i want to implement (writing a new wrapper, similar to rar), I have to, actually, update training data every iteration, so, my --interval 1
argument is surely impossible to use without code correction.
Can you either direct me how to resolve this issue, and I can do merge commit for you, or maybe you have time to resolve it yourself so I can continue my work? Thanks in advance.
Apart from this issue, I wanted to express few things (if needed, I can open issues, just tell me):
Argparse:
benchmark.py
help comment for each argument, their usage and intentend values are not obvious without looking inside the code.the "full documentation" website is lacking many descriptions, even when there is a subsection "header" in a content tree, but content item is not clickable (for example, RAR), but I gues you are aware of that.
weird one: if I would like to use summary()
function from src.utils.summary
standalone, deepxde uses tensorflow backend (at least the logger says so, and then I get error "no module tensorflow"), and there is no tensorflow in requirements.txt
file... (I work in isolated environment with Nix, so it was easy to catch, but maybe it is not intended for user to use plot/summary functions).
RAR implementation and choice of RAR:
X_train
:errors.txt
through runs/
? I think this kind of benchmark script would be amazing.Thanks a lot in advance for your attention and possible answer. You can also reply me in email: sofya (dot) dymchenko -at- gmail com.
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.