updated setup

pyforecaster/diagnostics.py

@@ -0,0 +1,67 @@
+import pandas as pd
+import numpy as np
+import matplotlib.pyplot as plt
+import umap
+from scipy.interpolate import griddata
+from pyforecaster.formatter import Formatter
+from pyforecaster.forecaster import LinearForecaster, LGBForecaster
+
+data = pd.read_pickle('tests/data/test_data.zip').droplevel(0, 1)
+data = data.resample('1h').mean()
+formatter = Formatter().add_transform(['all'], lags=np.arange(24), relative_lags=True)
+formatter.add_transform(['all'], ['min', 'max'], agg_bins=[1, 2, 15, 20])
+formatter.add_target_transform(['all'], lags=-np.arange(1, 12))
+x, y = formatter.transform(data)
+n_tr = int(len(x) * 0.8)
+x_tr, x_te, y_tr, y_te = [x.iloc[:n_tr, :].copy(), x.iloc[n_tr:, :].copy(), y.iloc[:n_tr].copy(),
+                          y.iloc[n_tr:].copy()]
+m_lgb = LinearForecaster().fit(x_tr, y_tr)
+y_hat = m_lgb.predict(x_te)
+
+for i in range(0, 24*10, 24):
+    plt.figure()
+    plt.plot(y_te.iloc[i], label='y_te')
+    plt.plot(y_hat.iloc[i], label='y_hat')
+    plt.show()
+
+tr = umap.UMAP(n_neighbors=20,learning_rate=0.1).fit(y_tr)
+y_te_tr = tr.transform(y_te)
+y_hat_tr = tr.transform(y_hat)
+
+
+plt.figure()
+xy = y_hat_tr
+uv = y_te_tr - y_hat_tr
+plt.quiver(*xy.T, *uv.T,
+           np.sum(uv ** 2, axis=1) ** 0.5, angles='xy')
+grid_x, grid_y = np.meshgrid(np.linspace(xy[:, 0].min(), xy[:, 0].max(), 20),
+                             np.linspace(xy[:, 1].min(), xy[:, 1].max(), 20))
+# Interpolate components
+u_grid = griddata(xy, uv[:, 0], (grid_x, grid_y), method='cubic')
+v_grid = griddata(xy, uv[:, 1], (grid_x, grid_y), method='cubic')
+plt.streamplot(grid_x, grid_y, u_grid, v_grid, color=np.sqrt(u_grid ** 2 + v_grid ** 2), cmap='viridis')
+plt.show()
+
+
+
+# plot a vectorial field using y_te.iloc[:, t] and y_hat.iloc[:, t] as starts and ends of the vectors
+# color it proportionally to the error (length of vector)
+tuples = np.random.randint(0, y_tr.shape[1], (10, 2))
+for t in tuples:
+    plt.figure()
+    #plt.scatter(*y_te.iloc[:, t].values.T, s=1, c='b')
+    #plt.scatter(*y_hat.iloc[:, t].values.T, s=10, c='r')
+    xy = y_hat.iloc[:, t].values
+    uv = y_te.iloc[:, t].values - y_hat.iloc[:, t].values
+    plt.quiver(*xy.T, *uv.T, y_hat.index.hour, angles='xy', alpha=0.5)
+
+    grid_x, grid_y = np.meshgrid(np.linspace(xy[:, 0].min(), xy[:, 0].max(), 20),
+                                 np.linspace(xy[:, 1].min(), xy[:, 1].max(), 20))
+    # Interpolate components
+    u_grid = griddata(xy, uv[:, 0], (grid_x, grid_y), method='cubic')
+    v_grid = griddata(xy, uv[:, 1], (grid_x, grid_y), method='cubic')
+    plt.streamplot(grid_x, grid_y, u_grid, v_grid, color=np.sqrt(u_grid ** 2 + v_grid ** 2), cmap='viridis')
+
+    plt.title('Error, tuple: {}'.format(t))
+    plt.show()
+

setup.py

@@ -37,8 +37,8 @@
                       'jaxlib>=0.4.1',
                       'quantile-forest>=1.2.0',
                       "optax>=0.1.7",
-                      "flax>=0.7.4",
+                      "flax>=0.10",
                       "statsmodels>=0.14.2",
                       ],
-    python_requires='>=3.8',
+    python_requires='>=3.11',
 )