timeseries_fastai

This repository aims to implement TimeSeries classification/regression algorithms. It makes extensive use of fastai V2!

Installation

You will need to install fastai V2 from here and then you can do from within the environment where you installed fastai V2:

pip install timeseries_fastai

and you are good to go.

TL;DR

git clone https://github.com/fastai/fastai
cd fastai
conda env create -f environment.yml
source activate fastai
pip install fastai timeseries_fastai

Time Series Classification from Scratch with Deep Neural Networks: A Strong Baseline

The original paper repo is here is implemented in Keras/Tf.

• Notebook 01: This is a basic notebook that implements the Deep Learning models proposed in Time Series Classification from Scratch with Deep Neural Networks: A Strong Baseline.

InceptionTime: Finding AlexNet for Time SeriesClassification

The original paper repo is here

• Notebook 02: Added InceptionTime architecture from InceptionTime: Finding AlexNet for Time SeriesClassification.

Results

You can run the benchmark using:

$python ucr.py --arch='inception' --tasks='all' --filename='inception.csv' --mixup=0.2

Default Values:

• lr = 1e-3
• opt = 'ranger'
• epochs = 40
• fp16 = True

import pandas as pd
from pathlib import Path

results_inception = pd.read_csv(Path.cwd().parent/'inception.csv', index_col=0)
results_inception.head(10)

<style scoped> .dataframe tbody tr th:only-of-type { vertical-align: middle; }

.dataframe tbody tr th {
    vertical-align: top;
}

.dataframe thead th {
    text-align: right;
}

</style>

	acc	acc_max	train_loss	val_loss
task
ACSF1	0.82	0.85	0.77	0.62
Adiac	0.77	0.77	0.81	0.89
ArrowHead	0.70	0.76	0.28	1.21
BME	0.85	0.88	0.21	0.79
Beef	0.77	0.83	0.50	0.53
BeetleFly	0.70	0.85	0.14	0.79
BirdChicken	0.95	0.95	0.14	0.20
CBF	0.95	0.97	0.22	0.24
Car	0.60	0.68	0.33	1.23
Chinatown	0.95	0.96	0.05	0.27

Getting Started

from timeseries_fastai.imports import *
from timeseries_fastai.core import *
from timeseries_fastai.data import *
from timeseries_fastai.models import *

PATH = Path.cwd().parent

df_train, df_test = load_df_ucr(PATH, 'Adiac')

Loading files from: /home/tcapelle/SteadySun/timeseries_fastai/Adiac

x_cols = df_train.columns[0:-2].to_list()

dls = TSDataLoaders.from_dfs(df_train, df_test, x_cols=x_cols, label_col='target', bs=16)

dls.show_batch()

inception = create_inception(1, len(dls.vocab))

learn = Learner(dls, inception, metrics=[accuracy])

learn.fit_one_cycle(5, 1e-3)

epoch	train_loss	valid_loss	accuracy	time
0	3.923817	3.641004	0.030691	00:01
1	3.678824	3.488305	0.107417	00:01
2	3.440019	3.159753	0.153453	00:01
3	3.089422	2.506488	0.301790	00:01
4	2.762558	2.371680	0.306905	00:01