Closed codeboy5 closed 2 years ago
I am also changing so that we can read the data from .arff
files instead of .tsv
files since most datasets are available in that format.
This allows us to run the following code, currently only uni variate time series are supported. The block and container will work for multi variate too but the process to read the input would change somewhat.
using FastAI
path = "/home/saksham/Desktop/GSOC/Adiac"
recipe = TimeSeries.TimeSeriesDatasetRecipe(file="Adiac_TRAIN.arff")
data, blocks = loadrecipe(recipe, path)
println(Datasets.testrecipe(recipe, data, blocks))
sample = getobs(data,5)
println(checkblock(blocks, sample))
which returns
Test Passed
true
This allows us to do the following
path = datasetpath("adiac")
recipe = TimeSeries.TimeSeriesDatasetRecipe(file="Adiac_TRAIN.arff")
data, blocks = loadrecipe(recipe, path)
Looking very good so far! I've left some comments. Unfortunately I am not familiar with time series using deep learning, so I don't know what the observations should look like and what encodings will be needed. Do you have a reference that gives an overview of these things?
Other than that, some tests would be great :) (see the rest of the library as reference for how to write these inline, and check ReTest.jl for how to run these interactively); dataset doesn't need a test since we don't want to download that on every CI run)
Let me see if I can find a good reference for this. This has some (https://timeseriesai.github.io/tsai/data.preparation.html) which we can refer to for encoding. I'll work on adding tests and look at resolving some of the comments you mentioned. I think going further we would also need to decide some models that we want to work on first.
That's a good resource and probably a good reference for implementations here :+1:
The best way to proceed is to take an example task from tsai, e.g. classification, and implement the components (incl. the model) used there.
Hey sorry, I was busy with some school stuff. So i am currently working on this tutorial and add this functionality to FastAI.jl .
That seems like a great starting point!
Hey sorry, I was busy with some school stuff.
No worries, it's normal that PRs take time and are punctured by some inactivity 😄 👍
I was previously using ARFFFiles.jl. It was not working with the multivariate time series, so I used this function from tsai with some modifications. We are able to now read .ts for both univariate and multivariate time series.
Added Time Series Container and Block. Currently can only load univariate time series. This is work in progress for issue #155 . I was planning to add
loaddataset
function for such datasets. Currently all datasets have the same root URL :- "https://s3.amazonaws.com/fast-ai-" . For time series datasets the root url is different so i think we can proceed by addroot_url
field in theFastAIDataset
structure. How does this sound ?