remove pandas-related FutureWarning

[lostella]

Issue #, if available: #87

Description of changes: addressed pandas complaints.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[szha]

Job PR-88/1 is complete.
Docs are uploaded to http://gluon-ts-staging.s3-accelerate.dualstack.amazonaws.com/PR-88/1/index.html

[szha]

Job PR-88/2 is complete.
Docs are uploaded to http://gluon-ts-staging.s3-accelerate.dualstack.amazonaws.com/PR-88/2/index.html

[alexw91]

Codecov Report

Merging #88 into master will increase coverage by <.01%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master      #88      +/-   ##
==========================================
+ Coverage   78.63%   78.64%   +<.01%     
==========================================
  Files         110      110              
  Lines        6282     6283       +1     
==========================================
+ Hits         4940     4941       +1     
  Misses       1342     1342

Impacted Files	Coverage Δ
src/gluonts/dataset/artificial/_base.py	32.16% <ø> (ø)	⬆️
src/gluonts/evaluation/_base.py	99.38% <100%> (ø)	⬆️
src/gluonts/model/seasonal_naive/_predictor.py	96.66% <100%> (ø)	⬆️
src/gluonts/evaluation/backtest.py	87.5% <100%> (ø)	⬆️

[rherbrich74]

Is this fix already installable via pip install --upgrade gluonts? I am running python 3.7.3 and I am still getting this issue when I am trying to train a feedforward model.

[lostella]

@rherbrich74 not yet, we will make a new release tomorrow probably. In the meantime you can

pip install git+https://github.com/awslabs/gluon-ts.git  

[rherbrich74]

@lostella: That worked - thanks! I have one more question on the semantics of make_evaluation_predictions and the parameter num_eval_samples: If I follow the tutorial, the first element of the forecasts array contains 100 sample paths and they are all around the target values:

forecasts[0].samples
array([[ 678.3231 , 576.42914, 561.24054, ..., 800.3523 , 901.3257 ,
916.05774],
[ 690.817 , 593.2066 , 1075.1001 , ..., 852.2531 , 808.0277 ,
665.4199 ],
[ 684.40814, 601.9213 , 523.0439 , ..., 847.2773 , 773.35095,
712.46436],
...,
[ 676.9368 , 563.7203 , 405.0648 , ..., 679.5283 , 732.5705 ,
680.9393 ],
[ 655.64545, 642.0739 , 524.64435, ..., 819.95447, 777.0701 ,
768.63086],
[ 741.4884 , 668.1726 , 518.23926, ..., 840.6607 , 858.57776,
774.19196]], dtype=float32)

But the second element of this array is a factor of five too high:

forecasts[1].samples
array([[3119.2852, 2552.321 , 2850.9817, ..., 3502.2512, 3602.2888,
2993.7017],
[2985.6204, 2787.8015, 2402.0273, ..., 3434.465 , 2011.0767,
3282.1006],
[2831.7378, 2684.3398, 2630.1077, ..., 1831.2285, 3129.4214,
2915.8489],
...,
[2978.8643, 2787.955 , 2404.7107, ..., 3094.3599, 3225.1023,
3651.5989],
[2733.5464, 2616.9783, 2448.9888, ..., 3635.505 , 3800.9192,
3365.458 ],
[3255.8765, 2840.524 , 2388.526 , ..., 3959.009 , 3052.5603,
2729.477 ]], dtype=float32)

Also, I do not understand the logic of this parameter. The number of sample paths for the predictor are initialized in the constructor

estimator = SimpleFeedForwardEstimator(
num_hidden_dimensions=[10],
prediction_length=dataset.metadata.prediction_length,
context_length=100,
freq=dataset.metadata.time_granularity,
num_eval_samples=1000,
trainer=Trainer(ctx="cpu", epochs=5, learning_rate=1E-3, hybridize=True, num_batches_per_epoch=200,),
)

Can you help clarify?

[lostella]

@rherbrich74 sure:

1. The forecasts[1] I get looks OK to me. Are you comparing it to the corresponding time series? In the notebook, you should change

tss[0][-plot_length:].plot(ax=ax)  # plot the time series
forecasts[0].plot(prediction_intervals=prediction_intervals, color='g')

into

tss[1][-plot_length:].plot(ax=ax)  # plot the time series
forecasts[1].plot(prediction_intervals=prediction_intervals, color='g')

to do the right comparison.

2. I believe the num_eval_samples argument of make_evaluation_predictions is unused by some models (e.g. SimpleFeedForward, as you noticed) which have it hard-coded since their creation. But maybe other types of predictors (like the R-based one) allow customizing this at inference time. This is definitely a potentially confusing point, we'll think about how to streamline the API wrt this. Thanks for the feedback!

[rherbrich74]

Thanks for the quick response, @lostella ! I still don't get the right output; this is what I see

I made the code publicly available at https://github.com/rherbrich74/gluon-ts-experiments/; maybe there is something obvious I am doing wrong.

[vafl]

That does look funny. Thanks for the sample.

FYI num_eval_samples should be fixed by #103