XIV. Randomness & Reproducibility (随机性与可复现性)#

1. `torch.manual_seed()`#

Sets the global random seed to ensure consistent results across runs (实验可复现性).

1
import random, numpy as np
2

3
def set_seed(seed=42):
4
    torch.manual_seed(seed)
5
    torch.cuda.manual_seed_all(seed)
6
    np.random.seed(seed)
7
    random.seed(seed)
8

9
set_seed(42)

Note: Also set torch.backends.cudnn.deterministic=True for fully deterministic behavior.

2. `torch.Generator()`#

Creates an independent Random Number Generator (随机数生成器) object, avoiding interference with the global seed.

1
g = torch.Generator()
2
g.manual_seed(42)
3
x = torch.rand(3, generator=g)  # independent state
4

5
loader = DataLoader(ds, shuffle=True, generator=g)

Note: Safer than global seeds in multi-thread / multi-process scenarios.

3. `torch.distributions.*`#

Probability distribution library supporting sampling and log_prob computation. Foundation for VAE and Reinforcement Learning (强化学习).

1
from torch.distributions import Normal, Categorical
2

3
# VAE reparameterization trick (重参数化技巧)
4
dist = Normal(loc=mu, scale=torch.exp(logvar))
5
z = dist.rsample()      # differentiable sampling
6
log_p = dist.log_prob(z)

Note: rsample() is differentiable (reparameterization); sample() is not (for policy gradients).

💡 One-line Takeaway
Always call set_seed() at the start of every experiment and use rsample() for differentiable stochastic layers.

XIV. Randomness & Reproducibility (随机性与可复现性)#

1. torch.manual_seed()#

2. torch.Generator()#

3. torch.distributions.*#

1. `torch.manual_seed()`#

2. `torch.Generator()`#

3. `torch.distributions.*`#