[WIP] Blog about working with Time Series Data using FastAI.jl

blog/_posts/2022-09-08-Adding-Time-Series-Support-to-FastAI.jl.md

@@ -0,0 +1,103 @@
+---
+title: Adding Time Series Support to FastAI.jl
+author: Saksham
+layout: blog
+---
+
+[FastAI.jl](https://github.com/FluxML/FastAI.jl) is a Julia library inspired by [fastai](https://github.com/fastai/fastai), and its goal is to create state-of-the-art deep learning models easily. FastAI.jl simplifies training fast and accurate neural nets using modern best practices.
+
+Models for time series data constitute an integral part of any machine learning stack. This blog post will demonstrate how to start working with time-series data with FastAI.jl and the [FastTimeSeries](https://github.com/FluxML/FastAI.jl/tree/master/FastTimeSeries) submodule. The FastTimeSeries submodule has been inspired by [tsai](https://timeseriesai.github.io/tsai/), a package built on top of fastai for time series tasks. The work presented here was done as part of [GSoC'22](https://summerofcode.withgoogle.com/programs/2022/projects/Q9GVFW33) under the mentorship of Brian Chen, Kyle Daruwalla, and Lorenz Ohly.
+
+
+## Loading the data in a container
+
+To start off, we'll load the [ECG5000](http://timeseriesclassification.com/description.php?Dataset=ECG5000) dataset.
+
+```julia
+julia> using FastAI, FastTimeSeries, Flux
+julia> data, blocks = load(datarecipes()["ecg5000"]);
+```
+
+Our easy to use interface allows to load an input time series along with it's label at any index using `getobs(data, index)`. It also allows us to check the total number of observations using `numobs(data)`.
+
+```julia
+julia> input, class = sample = getobs(data, 25)
+(Float32[-0.28834122 -2.2725453 … 1.722784 1.2959242], "1")
+julia> numobs(data)
+5000
+```
+
+## Tasks
+
+The library supports `TSClassificationSingle` and `TSRegression` tasks--used for single label time-series classification and single label time-series regression, respectively. We will pass our `data` and `blocks` from the previous step into the task:
+
+```julia
+julia> task = TSClassificationSingle(blocks, data);
+```
+
+## Data Preprocessing
+
+Although `data` can already be passed to `DataLoader` for loading during training, we would often like to perform transformations on it. We can encode a sample input using `encodesample(task, Phase(), sample)` where `Phase` is a [FastAI.jl Context](https://fluxml.ai/FastAI.jl/dev/references/FastAI.Context).
+
+```julia
+julia> input, class = sample = getobs(data, 25)
+(Float32[-0.28834122 -2.2725453 … 1.722784 1.2959242], "1")
+julia> encodesample(task, Training(), (input, class))
+(Float32[-0.28937635 -2.2807038 … 1.7289687 1.3005764], Bool[1, 0, 0, 0, 0])
+```
+
+## Models
+
+The library contains implementation of the following models.
+
+- Basic stacked RNNs.  
+  Stacked RNN network. Feeds the data through a chain of RNN layers, where the hidden state
+  of the previous layer gets fed to the next one. The Model has the following arguments.
+  - `c_in` : The number of input channels.
+  - `c_out` : The number of output classes.
+  - `hiddensize` : The number of "hidden channels" to use.
+  - `layers` : The number of RNN layers to use in the stacked network. 
+
+```julia
+julia> backbone = FastTimeSeries.Models.StackedLSTM(1, 16, 10, 2);
+julia> model = FastAI.taskmodel(task, backbone);
+```
+
+- [InceptionTime](https://arxiv.org/abs/1909.04939) 
+  An implementation of the InceptionTime Model. The Model has the following arguments. 
+  - `c_in` : The number of input channels.
+  - `c_out` : The number of output classes.
+  - `nf` : The number of "hidden channels" to use.
+
+```julia
+julia> model = FastTimeSeries.Models.InceptionTime(1, 5);
+```
+
+## Training
+
+We create a pair of training and validation data loaders using `taskdataloaders` . They take care of batching and loading the data in parallel in the background. With the addition of an optimizer and a loss function, we can create a `Learner` and start training.
+
+```julia
+julia> traindl, validdl = taskdataloaders(data, task, 16);
+julia> learner = Learner(model, tasklossfn(task); data=(traindl, validdl), optimizer=ADAM(), callbacks = [ToGPU(), Metrics(accuracy)]);
+julia> fitonecycle!(learner, 10, 0.033);
+```
+
+We can view the loss and accuracy on the training and validation data after the training is compelete.
+
+<p float="middle">
+  <img src="/assets/2022-09-08-FastAI-time-series/train_result.png" height="250">
+</p>
+
+## Conclusion
+
+We saw how we could work on time-series data using FastAI.jl. The library now supports the following features. 
+
+- Load time series classification datasets and regression datasets from [UCR Classification](http://timeseriesclassification.com/index.php) and [Monash Regression](http://tseregression.org/) respectively.
+- Perform data transformations on the loaded data.
+- Perform classification and regression tasks using InceptionTime or stacked RNNs.
+
+## Future Work
+
+Future work would involve adding tools to the library to improve the experience of working with time series. These would include tools for visualizations and analysis. 
+FastTimeSeries would also be expanded to work on more complex tasks such as multi-class classification and single and multi-step forecasting. We would also implement more complex models such as [Transformers](https://arxiv.org/abs/2010.02803).