import os
import numpy as np
import time
from . import Dataset
from sklearn.datasets import make_moons, make_circles
from sklearn.model_selection import train_test_split
def load_path(N=1000, std=0.03, option=0, seed=None):
X = np.linspace(0, 1, N)
if option == 0:
# simple cosine
X = np.stack([X * 9, np.cos(X * 9)], 1)
elif option == 1:
# circle
X -= 0.5
X *= 2
Y = np.concatenate([np.sqrt(1 - X[::2] ** 2), -np.sqrt(1 - X[::2] ** 2)], 0)
X = np.concatenate([X[::2], X[::2]], 0)
X = np.stack([X, Y], 1)
elif option == 2:
# swiss roll
t = np.linspace(0, 2, N)
X = np.stack([t * np.cos(3.14 * t), t * np.sin(3.14 * t)], 1)
elif option == 3:
# square
left_side = np.stack([np.ones(N // 4) * X[0], X[::4]], 1)
right_side = np.stack([np.ones(N // 4) * X[-1], X[::4]], 1)
upper_side = np.stack([X[::4], np.ones(N // 4) * X[0]], 1)
lower_side = np.stack([X[::4], np.ones(N // 4) * X[-1]], 1)
X = np.concatenate([left_side, upper_side, right_side, lower_side[::-1]], 0)
elif option == 4:
# simple cosine
X = np.stack([X * 9, np.random.rand(*X.shape), np.cos(X * 9)], 1)
elif option == 5:
# X = np.random.randn(N, 3)
# X /= np.linalg.norm(X, 2, axis=1, keepdims=True)
# X[:,0] = np.abs(X[:, 0])
import sklearn
X, _ = sklearn.datasets.make_swiss_roll(N, 0)
X += np.random.RandomState(seed).randn(N, X.shape[1]) * std
X -= X.mean(0, keepdims=True)
X /= X.max(0, keepdims=True)
X = X.astype("float32")
dataset = Dataset()
dataset["inputs/train_set"] = X
return dataset
def load_mini(N=1000):
X, y = make_moons(N, noise=0.035, random_state=20)
x_, y_ = make_circles(N, noise=0.02, random_state=20)
x_[:, 1] += 2.0
y_ += 2
X = np.concatenate([X, x_], axis=0)
y = np.concatenate([y, y_])
X -= X.mean(0, keepdims=True)
X /= X.max(0, keepdims=True)
X = X.astype("float32")
y = y.astype("int32")
dict_init = [
("datum_shape", (2,)),
("n_classes", 4),
("name", "mini"),
("classes", [str(u) for u in range(4)]),
]
dataset = Dataset(**dict(dict_init))
dataset["inputs/train_set"] = X
dataset["outputs/train_set"] = y
return dataset
def load_chirp2D(N, seed=0):
X1, X2 = np.meshgrid(np.linspace(0, 4, N), np.linspace(0, 4, N))
np.random.seed(seed)
M = np.array([[1.4, -0.4], [-0.4, 0.6]])
X = np.stack([X1.flatten(), X2.flatten()], 1)
y = np.sin((X * np.dot(X, M)).sum(1))
y -= y.mean()
y /= y.max()
X = X.astype("float32")
y = y.astype("float32")
dict_init = [("datum_shape", (2,)), ("name", "chirp2D")]
dataset = Dataset(**dict(dict_init))
images = {"train_set": X}
labels = {"train_set": y}
dataset.add_variable({"input": images, "output": labels})
return dataset
