From 532233175887d8a0411c5c573d16fe49b0d9a7d8 Mon Sep 17 00:00:00 2001 From: Jacob Date: Mon, 6 Jan 2025 22:38:18 -0600 Subject: [PATCH] Update electricity.ipynb --- docs/examples/electricity.ipynb | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) diff --git a/docs/examples/electricity.ipynb b/docs/examples/electricity.ipynb index 3af4337..076f4a6 100644 --- a/docs/examples/electricity.ipynb +++ b/docs/examples/electricity.ipynb @@ -444,7 +444,7 @@ "id": "4a82175c-56c8-454a-925f-9cabfbedd079" }, "source": [ - "The first thing you'll notice about this approach is that it's **incredibly slow to train**. The second problem is that the forecasts are **terrible**:" + "The problem is that the forecasts are **terrible**:" ] }, { @@ -495,10 +495,9 @@ "source": [ "### Attempt 2\n", "\n", - "Let's see if we can improve on this. We'll leave the model unchanged but make two changes:\n", + "Let's see if we can improve on this. We'll leave the model unchanged, but train it differently. We'll use the `n_step` argument to train our model on one-week ahead forecasts, instead of one-hour ahead. This improves the efficiency of training by encouraging the model to 'care about' longer range forecasts vs. over-focusing on the easier problem of forecasting the next hour.\n", "\n", - "- Use the `n_step` argument to train our model on one-week ahead forecasts, instead of one step (i.e. hour) ahead. This improves the efficiency of training by encouraging the model to 'care about' longer range forecasts vs. over-focusing on the easier problem of forecasting the next hour.\n", - "- Split our single series into multiple groups. This is helpful to speed up training, since pytorch has a non-trivial overhead for separate tensors: i.e., it scales well with an increasing batch-size (fewer, but bigger, tensors), but poorly with an increasing time-series length (smaller, but more, tensors)." + "Another thing we'll address is the fact that the simple (and lousy) model above was surprisingly slow to train. This is because Pytorch has a non-trivial overhead for separate tensors: i.e., it scales well with an increasing batch-size (fewer, but bigger, tensors), but poorly with an increasing time-series length (smaller, but more, tensors). So to speed things up, we'll split our single series into multiple groups." ] }, {