class Trend[source]

Trend(breakpoints:int=None, date_range=None, mean=0, scale=1) :: Module

Broken Trend model, with breakpoints as defined by user. parameters:

• breakpoints (optional): number of breaks in trend
• date_range: Required if breakpoints is given. Start and stop date in days.
• mean: mean of time
• scale: standard deviation of time

Test with Dummy Data

N = 700
t = np.arange(N)

trend = 0.05*t + 0.1
trend[t>350] = -0.1*(t[t>350] - 350) + trend[t==350]

noise = np.random.randn(*trend.shape)
y = trend + noise
# y = np.clip(y, 1, 20)

plt.plot(t, y)
plt.plot(t, trend)
plt.show()

breakpoints = torch.Tensor([250, 500])
model = Trend(breakpoints)
learner = Learner(db, model, loss_func=F.mse_loss)
wd = 0
learner.lr_find(wd=wd)
learner.recorder.plot(skip_end=0)

LR Finder is complete, type {learner_name}.recorder.plot() to see the graph.

epochs = 10
learner.fit_one_cycle(epochs, 1e-1, wd=wd)

epoch	train_loss	valid_loss	time
0	15006.965820	48516.570312	00:00
1	10705.563477	993.336182	00:00
2	7057.067871	29231.976562	00:00
3	5744.917969	11459.618164	00:00
4	4433.373047	556.789673	00:00
5	3629.159668	733.305664	00:00
6	2988.094727	189.605576	00:00
7	2479.688965	1147.893921	00:00
8	2095.290527	1420.154663	00:00
9	1794.216064	1420.845947	00:00

y2 = learner.model(torch.Tensor(t)[:,None])
plt.plot(t, y, label='actual')
plt.plot(t[:train_len], y[:train_len], label='train set')
plt.plot(t, y2, label='prediction')
plt.legend()
plt.show()