From 500ee411758c5f208c1698e0c3383580691f792a Mon Sep 17 00:00:00 2001
From: Miao Zheng <76149310+MeowZheng@users.noreply.github.com>
Date: Sat, 20 Nov 2021 22:04:52 +0800
Subject: [PATCH] [Enhancement] set the same seed for all rank when distributed
 training (#27)

---
 mmflow/apis/__init__.py |  4 ++--
 mmflow/apis/train.py    | 33 ++++++++++++++++++++++++++++++++-
 tools/train.py          | 14 +++++++-------
 3 files changed, 41 insertions(+), 10 deletions(-)

diff --git a/mmflow/apis/__init__.py b/mmflow/apis/__init__.py
index 396ffbe..3c10c92 100644
--- a/mmflow/apis/__init__.py
+++ b/mmflow/apis/__init__.py
@@ -1,9 +1,9 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 from .inference import inference_model, init_model
 from .test import multi_gpu_test, single_gpu_test
-from .train import set_random_seed, train_model
+from .train import init_random_seed, set_random_seed, train_model
 
 __all__ = [
     'set_random_seed', 'train_model', 'init_model', 'inference_model',
-    'multi_gpu_test', 'single_gpu_test'
+    'multi_gpu_test', 'single_gpu_test', 'init_random_seed'
 ]
diff --git a/mmflow/apis/train.py b/mmflow/apis/train.py
index 1a193de..44d4090 100644
--- a/mmflow/apis/train.py
+++ b/mmflow/apis/train.py
@@ -5,9 +5,10 @@
 
 import numpy as np
 import torch
+import torch.distributed as dist
 from mmcv.parallel import MMDataParallel, MMDistributedDataParallel
 from mmcv.runner import (HOOKS, Fp16OptimizerHook, OptimizerHook,
-                         build_optimizer, build_runner)
+                         build_optimizer, build_runner, get_dist_info)
 from mmcv.utils import Config, build_from_cfg
 
 from mmflow.core import DistEvalHook, EvalHook
@@ -18,6 +19,36 @@
 Dataset = torch.utils.data.Dataset
 
 
+def init_random_seed(seed: Optional[int] = None, device: str = 'cuda') -> int:
+    """Initialize random seed.
+
+    If the seed is not set, the seed will be automatically randomized,
+    and then broadcast to all processes to avoid some potential bugs.
+    Args:
+        seed (int, Optional): The seed. Default to None.
+        device (str): The device where the seed will be put on.
+            Default to 'cuda'.
+    Returns:
+        int: Seed to be used.
+    """
+    if seed is not None:
+        return seed
+
+    # Make sure all ranks share the same random seed to avoid
+    # some potential bugs. Please refer to
+    # https://github.com/open-mmlab/mmdetection/issues/6339
+    rank, world_size = get_dist_info()
+    seed = np.random.randint(2**31)
+    if world_size == 1:
+        return seed
+    if rank == 0:
+        random_num = torch.tensor(seed, dtype=torch.int32, device=device)
+    else:
+        random_num = torch.tensor(0, dtype=torch.int32, device=device)
+    dist.broadcast(random_num, src=0)
+    return random_num.item()
+
+
 def set_random_seed(seed: int, deterministic: bool = False) -> None:
     """Set random seed.
 
diff --git a/tools/train.py b/tools/train.py
index ebc6548..0461d8d 100644
--- a/tools/train.py
+++ b/tools/train.py
@@ -12,7 +12,7 @@
 from mmcv.utils import get_git_hash
 
 from mmflow import __version__
-from mmflow.apis import set_random_seed, train_model
+from mmflow.apis import init_random_seed, set_random_seed, train_model
 from mmflow.datasets import build_dataset
 from mmflow.models import build_flow_estimator
 from mmflow.utils import collect_env, get_root_logger
@@ -130,12 +130,12 @@ def main():
     logger.info(f'Config:\n{cfg.pretty_text}')
 
     # set random seeds
-    if args.seed is not None:
-        logger.info(f'Set random seed to {args.seed}, deterministic: '
-                    f'{args.deterministic}')
-        set_random_seed(args.seed, deterministic=args.deterministic)
-    cfg.seed = args.seed
-    meta['seed'] = args.seed
+    seed = init_random_seed(args.seed)
+    logger.info(f'Set random seed to {seed}, '
+                f'deterministic: {args.deterministic}')
+    set_random_seed(seed, deterministic=args.deterministic)
+    cfg.seed = seed
+    meta['seed'] = seed
     meta['exp_name'] = osp.basename(args.config)
 
     model = build_flow_estimator(cfg.model)