gradientsky / fastseq

A way to use fastai with sequence data

Fastseq

A way to use fastai with sequence data

Installing

Please install fastai2 according to the instructions.

Then install Fastseq by:

git clone https://github.com/takotab/fastseq.git
cd fastseq
pip install -e .

If you have tried it, please let me know. Even if you got stuck and werent able to install. You can send me an email, leave a comment in the (excelent) fastai timeseries forum, or make an issue in github. I would love to know what you think of libary any tips or comments. Do you work on these kind of data, would want more explanation of the model, would you like to see the approuch on multivariate ts. Where do think I should go with this.

Thank you

How to use

from fastai2.basics import *
from fastseq.all import *
from fastseq.nbeats.model import *
from fastseq.nbeats.learner import *
from fastseq.nbeats.callbacks import *

horizon, lookback = 7, 35    

Getting the data fastai style:

path = untar_data(URLs.m4_daily)
data = TSDataLoaders.from_folder(path, horizon = horizon, lookback = lookback, nrows = 300, step=3)

torch.Size([1, 1020])
Train:70707; Valid: 1200; Test 300

data.show_batch()

learn = nbeats_learner(data, season = lookback+horizon)   

from fastai2.callback.all import *
learn.lr_find()

learn.fit_flat_cos(5, 2e-2)
learn.recorder.plot_loss()
learn.recorder.plot_sched()

epoch	train_loss	valid_loss	mae	smape	theta	b_loss	f_loss	f_smape	time
0	1.608365	1.429745	0.468662	0.756240	2.085228	nan	nan	0.103745	01:26
1	1.543519	1.398041	0.449884	0.728823	2.033593	nan	nan	0.104480	01:25
2	1.557428	1.470988	0.467814	0.739414	1.833764	nan	nan	0.108116	01:26
3	1.507902	1.488569	0.459715	0.723740	1.953411	nan	nan	0.104374	01:28
4	1.425810	1.338848	0.421528	0.684144	1.802997	nan	nan	0.099711	01:32

learn.show_results(0)

learn.show_results(1)

learn.fit_flat_cos(1,5e-4)
learn.recorder.plot_loss()
learn.recorder.plot_sched()

epoch	train_loss	valid_loss	mae	smape	theta	b_loss	f_loss	f_smape	time
0	1.424263	1.343198	0.420948	0.683976	1.796664	nan	nan	0.100251	01:28

Interperation

learn.n_beats_attention.means()

<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>

	trend0_0	trend0_1	seasonality1_0	seasonality1_1	seasonality1_2	seasonality1_3
theta_0_mean	0.084346004	-0.15566854	0.33333334	0.0	0.10788739	-0.8549633
theta_0_std	0.9808921	1.0191917	0.40388188	0.0	0.17689645	0.5912291
theta_1_mean	0.002812498	-0.00619044	0.083333336	0.18620752	-0.05470855	-0.26863918
theta_1_std	0.03269631	0.035362627	0.10097047	0.23295449	0.4260137	0.22262059
theta_2_mean	-0.0012419696	0.0013037563	0.015625	-0.125	0.033530425	-0.013941522
theta_2_std	0.002147075	0.0010560098	0.061155755	0.21879749	0.06385607	0.1387485
theta_3_mean	-0.005537528	6.2500553e-06	NaN	0.30989584	-0.052088413	-0.3953999
theta_3_std	0.024471516	1.6711083e-05	NaN	0.25557557	0.18564244	0.502909
theta_4_mean	0.37125885	-0.0177321	NaN	NaN	-0.0139180375	-0.16807629
theta_4_std	0.81073123	0.077905715	NaN	NaN	0.060866524	0.21030271
theta_5_mean	0.28519723	0.29930714	NaN	NaN	NaN	0.046367485
theta_5_std	0.8494713	0.58685046	NaN	NaN	NaN	0.124893904
att_mean	0.484153	0.808544	0.6875	0.692708	0.833238	0.935759
att_std	0.45344	0.332966	0.463512	0.461371	0.368922	0.242476
theta_6_mean	NaN	0.27313724	NaN	NaN	NaN	NaN
theta_6_std	NaN	0.64940923	NaN	NaN	NaN	NaN
theta_7_mean	NaN	0.237134	NaN	NaN	NaN	NaN
theta_7_std	NaN	0.6669893	NaN	NaN	NaN	NaN

interp = NBeatsInterpretation.from_learner(learn)

interp.plot_top_losses(3, combine_stack=True)