A summary of seeding in some packages

random_seeding.py

import random
import functools

import numpy as np
import tensorflow as tf
import torch
import gym

import tianshou as ts
import pfrl

RANDOM_SEED = 0

# python
random.seed(RANDOM_SEED)

# numpy
np.random.seed(RANDOM_SEED)

# TF
tf.random.set_seed(RANDOM_SEED)

# pytorch
torch.manual_seed(RANDOM_SEED)

# gym
env = gym.make("Pendulum-v0")
env.seed(RANDOM_SEED)
env.action_space.seed(RANDOM_SEED)

# Tianshou (based on Pytorch)
vector_env_t = ts.env.DummyVectorEnv([lambda: gym.make("Pendulum-v0") for _ in range(3)])
vector_env_t.seed(RANDOM_SEED)

# PFRL (based on Pytorch)
# pfrl.utils.set_random_seed(RANDOM_SEED)  # same with the seeding above

def make_env(idx):
  return gym.make("Pendulum-v0")

vector_env_p = pfrl.envs.MultiprocessVectorEnv([functools.partial(make_env, idx) for idx in range(3)])
vector_env_p.seed(seeds=[RANDOM_SEED + i for i in range(3)])


# print result:
print("python: ",random.random())
print("numpy: ", np.random.rand())
print("tf: ", tf.random.normal((1,)))
print("torch: ", torch.rand(1))
print("gym obs: ", env.reset())
print("gym action: ", env.action_space.sample())
print("vector tianshou: ", vector_env_t.reset())
print("vector pfrl: ", vector_env_p.reset())

""" You'll get those results exactly by running this code.
python:  0.8444218515250481
numpy:  0.5488135039273248
tf:  tf.Tensor([1.5110626], shape=(1,), dtype=float32)
torch:  tensor([0.4963])
gym obs:  [-0.94223519 -0.33495202  0.93078187]
gym action:  [-1.7825598]
vector tianshou:  [[-0.94223519 -0.33495202  0.93078187]
 [-0.35283079  0.93568715  0.02900014]
 [-0.57182853  0.82037317  0.30004494]]
vector pfrl:  [array([-0.94223519, -0.33495202,  0.93078187]), array([-0.35283079,  0.93568715,  0.02900014]), array([-0.57182853,  0.82037317,  0.30004494])]
"""