angus924 / rocket Goto Github PK
View Code? Open in Web Editor NEWROCKET: Exceptionally fast and accurate time series classification using random convolutional kernels
License: GNU General Public License v3.0
ROCKET: Exceptionally fast and accurate time series classification using random convolutional kernels
License: GNU General Public License v3.0
Hi,
I have learnt your paper these day, which is very enlightening to me. I have a question about Results_ucr_resamples. I do not how to reproduce the results and what does resample mean. Could you please give me some help?
First of all: Thank you very much for providing this new method! If it works, it works very well and leads to good results (i.e. leads to results comparable to other methods within minutes instead of weeks).
Unfortunately, it doesn't work all the time for me. Especially if I use 10,000 kernels as recommended in your paper. This is what I end up with on a Windows 10 machine equipped with an NVIDIA 1080 TI GPU and 32 GB RAM:
Algorithms.trainAndEvaluateROCKET(...)
File "PathToMyProject\MyClass.py", line 180, in trainAndEvaluateROCKET
classifier.fit(X_training_transform,trainY)
File "PathToMyAnacondaFolder\lib\site-packages\sklearn\linear_model\ridge.py", line 1815, in fit
_BaseRidgeCV.fit(self, X, Y, sample_weight=sample_weight)
File "PathToMyAnacondaFolder\lib\site-packages\sklearn\linear_model\ridge.py", line 1528, in fit
estimator.fit(X, y, sample_weight=sample_weight)
File "PathToMyAnacondaFolder\lib\site-packages\sklearn\linear_model\ridge.py", line 1436, in fit
X_mean, *decomposition = decompose(X, y, sqrt_sw)
File "PathToMyAnacondaFolder\lib\site-packages\sklearn\linear_model\ridge.py", line 1348, in _svd_decompose_design_matrix
U, s, _ = linalg.svd(X, full_matrices=0)
File "PathToMyAnacondaFolder\lib\site-packages\scipy\linalg\decomp_svd.py", line 129, in svd
full_matrices=full_matrices, overwrite_a=overwrite_a)
MemoryError
My testing data-set consists of 3,278 time series (with a length of 1041 amplitude values each) and the training data-set consists of 43,264 time series (again with a length of 1041 amplitude values each). Any help is highly appreciated.
// edit:I forgot to mention that it works perfectly fine for 100 or 1,000 kernels.
Hi,
I notice that that the ROCKET transform implemented in sktime now supports the transformation of multivariate datasets.
I was wondering whether you could detail how it is applied?
Many thanks.
Hi,
How can I use it for Multivariate time-series (multi-dimension time-series)?
Thanks,
Hi I'd like to reproduce the scalability experiment using the SITS dataset mentioned in the paper. Is this dataset publicly available? If so, where can I find it?
Thanks =)
Hi,
I would like to know if, instead of the predicted class, I can know the probability of being in that class. Also, if it allows for time series with different lengths.
regards,
Ferran
Hello, I am a master's student and I am working on your article. I have a problem about the final accuracy, even though I actually implemented your code, I get much less accuracy. I have put the link related to the code below. If you tell me the error of the code, I will be grateful.
https://colab.research.google.com/drive/1vH9qAI2lgF_0htYPhU8UQb45aGUBEGk5?usp=share_link
Hi, I wonder if a multivariant implementation of the rocket is available?
Hi,
I just wanted to let you know that it is possible to fix the random seed with Numba. The only change would be to add a seed
parameter to the generate_kernels function:
@njit
def generate_kernels(input_length, num_kernels, seed=42):
np.random.seed(seed)
This way, the results would be perfectly reproducible (as it is, there is randomness when the kernels are generated that cannot be fixed).
Best,
Johann
Hi Angus,
First of all congrats on your new TSC method. I've used it and the results are very good, and super fast compared to other methods!
I've noticed a discrepancy between the bake-off and the scalable codes you've shared.
In the scalable, you perform a per-feature normalization while you don't do that in the bake-off.
Is there any reason for that? I've run a few tests with some bake-off datasets and have seen that this per-feature normalization hurts performance.
Hi,
I found that in multi-rocket the features are not normalized when fed into the ridge classifier, While they are normalized in rocket and mini-rocket. I wonder if there is any reason the features are normalized, and in the case of the features being normalized, I wonder why the MAXs and PPVs are normalized together instead of normalized individually.
Thank you in advance.
Yunrui
Hello,
I just learn the amazing Rocket and MiniRocket recently. When I used them to classify some UCRArchive2018 datasets (such as 'DodgerLoopGame'), NaN values representing missing values caused Errors. Therefore, I wonder how you handle these values? Is it approriate to directly delete samples with NaN values?
Thanks for your help!
A declarative, efficient, and flexible JavaScript library for building user interfaces.
๐ 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. ๐๐๐
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 โค๏ธ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.