torchbox.ml package

Submodules

torchbox.ml.evolutionary module

class torchbox.ml.evolutionary.GeneticAlgorithm(pop_size: int = 50, n_gen: int = 100, crossover_prob: float = 0.8, mutate_prob: float = 0.1, elitism: bool = True, elitism_num: int = 2, device: str = 'cpu')

Bases: ABC

PyTorch 实现的通用可扩展遗传算法基类 只需重写: initialize_population, crossover, mutate, selection 四个方法 主进化流程通用，支持 GPU 加速

abstractmethod crossover(parent1: Tensor, parent2: Tensor) → Tensor

【必须重写】交叉操作 输入: 两个父代个体 (chromosome_length,) 返回: 子代个体 (chromosome_length,)

evaluate_fitness(population: Tensor) → Tensor

【可选重写】适应度函数（默认需要用户重写，也可内置） 输入: 种群张量 [pop_size, chrom_len] 返回: 适应度张量 [pop_size] (值越大表示个体越优)

evolve() → Tuple[Tensor, Tensor]

通用主进化流程（核心！无需修改） 返回: 最优个体, 最优适应度历史

abstractmethod initialize_population() → Tensor

【必须重写】种群初始化 返回: 种群张量 (shape: [pop_size, chromosome_length])

abstractmethod mutate(individual: Tensor) → Tensor

【必须重写】变异操作 输入: 单个个体 (chromosome_length,) 返回: 变异后个体

plot_fitness_curve(save_path: str | None = None)

绘制适应度进化曲线（可视化）

abstractmethod selection() → Tensor

【必须重写】选择操作 根据适应度从种群中选择父代 返回: 选中的父代种群张量 [pop_size, chrom_len]

torchbox.ml.reduction_pca module

torchbox.ml.reduction_pca.pca(x, sdim=-2, fdim=-1, npcs='all', eigbkd='svd')

Principal Component Analysis (pca) on raw data

Parameters:

  • x (Tensor) – the input data

  • sdim (int, optional) – the dimension index of sample, by default -2

  • fdim (int, optional) – the dimension index of feature, by default -1

  • npcs (int or str, optional) – the number of components, by default 'all'

  • eigbkd (str, optional) – the backend of eigen decomposition, 'svd' (default) or 'eig'

Returns:

U, S, K (if npcs is integer)

Return type:

tensor

Examples

_images/MNISTPCA_ORIG.png _images/MNISTPCA_K90.png

The results shown in the above figure can be obtained by the following codes.

rootdir, dataset = '/mnt/d/DataSets/oi/dgi/mnist/official/', 'test'
x, _ = tb.read_mnist(rootdir=rootdir, dataset=dataset, fmt='ubyte')
print(x.shape)
N, M2, _ = x.shape
x = x.to(th.float32)
pcr = 0.9

u, s = tb.pca(x, sdim=0, fdim=(1, 2), eigbkd='svd')
k = tb.pcapc(s, pcr=pcr)
print(u.shape, s.shape, k)
u = u[..., :k]
y = x.reshape(N, -1) @ u  # N-k
z = y @ u.T.conj()
# z[z<0] = 0
z = z.reshape(N, M2, M2)
print(tb.nmse(x, z, dim=(1, 2)))
xp = th.nn.functional.pad(x[:35], (1, 1, 1, 1, 0, 0), 'constant', 255)
zp = th.nn.functional.pad(z[:35], (1, 1, 1, 1, 0, 0), 'constant', 255)
plt = tb.imshow(tb.patch2tensor(xp, (5*(M2+2), 7*(M2+2)), dim=(1, 2)), titles=['Orignal'])
plt = tb.imshow(tb.patch2tensor(zp, (5*(M2+2), 7*(M2+2)), dim=(1, 2)), titles=['Reconstructed with %d PCs(%.2f%%)' % (k, 100*k/u.shape[0])])
plt.show()

u, s = tb.pca(x.reshape(N, -1), sdim=0, fdim=1, npcs=2, eigbkd='svd')
print(u.shape, s.shape)
y = x.reshape(N, -1) @ u  # N-k
z = y @ u.T.conj()
z = z.reshape(N, M2, M2)
print(tb.nmse(x, z, dim=(1, 2)))
torchbox.ml.reduction_pca.pcapc(s, pcr=0.9)

get principal component according to the ratio of variance

Parameters:

  • s (Tensor) – eigenvalues

  • pcr (float, optional) – the ratio of variance, by default 0.9

torchbox.ml.reduction_pca.pcat(x, sdim=-2, fdim=-1, isnorm=True, eigbkd='svd')

gets Principal Component Analysis transformation

Parameters:

  • x (Tensor) – the input data

  • sdim (int, optional) – the dimension index of sample, by default -2

  • fdim (int or tuple, optional) – the dimension index of feature, by default -1

  • isnorm (bool, optional) – whether to normalize covariance matric with the number of samples, by default True

  • eigbkd (str, optional) – the backend of eigen decomposition, 'svd' (default) or 'eig'

Returns:

the PCA transformation matrix, the eigenvalues

Return type:

tensor

Module contents