[WIP] Feature/ddp fixed (#401)

* Squashed commit of the following:

commit d738487089e41c22b3b1cd73aa7c1c40320a6ebf
Author: NanoCode012 <kevinvong@rocketmail.com>
Date:   Tue Jul 14 17:33:38 2020 +0700

    Adding world_size

    Reduce calls to torch.distributed. For use in create_dataloader.

commit e742dd9619d29306c7541821238d3d7cddcdc508
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Tue Jul 14 15:38:48 2020 +0800

    Make SyncBN a choice

commit e90d4004387e6103fecad745f8cbc2edc918e906
Merge: 5bf8beb cd90360
Author: yzchen <Chenyzsjtu@gmail.com>
Date:   Tue Jul 14 15:32:10 2020 +0800

    Merge pull request #6 from NanoCode012/patch-5

    Update train.py

commit cd9036017e7f8bd519a8b62adab0f47ea67f4962
Author: NanoCode012 <kevinvong@rocketmail.com>
Date:   Tue Jul 14 13:39:29 2020 +0700

    Update train.py

    Remove redundant `opt.` prefix.

commit 5bf8bebe8873afb18b762fe1f409aca116fac073
Merge: c9558a9 a1c8406
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Tue Jul 14 14:09:51 2020 +0800

    Merge branch 'master' of https://github.com/ultralytics/yolov5 into feature/DDP_fixed

commit c9558a9b51547febb03d9c1ca42e2ef0fc15bb31
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Tue Jul 14 13:51:34 2020 +0800

    Add device allocation for loss compute

commit 4f08c692fb5e943a89e0ee354ef6c80a50eeb28d
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Thu Jul 9 11:16:27 2020 +0800

    Revert drop_last

commit 1dabe33a5a223b758cc761fc8741c6224205a34b
Merge: a1ce9b1 4b8450b
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Thu Jul 9 11:15:49 2020 +0800

    Merge branch 'feature/DDP_fixed' of https://github.com/MagicFrogSJTU/yolov5 into feature/DDP_fixed

commit a1ce9b1e96b71d7fcb9d3e8143013eb8cebe5e27
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Thu Jul 9 11:15:21 2020 +0800

    fix lr warning

commit 4b8450b46db76e5e58cd95df965d4736077cfb0e
Merge: b9a50ae 02c63ef
Author: yzchen <Chenyzsjtu@gmail.com>
Date:   Wed Jul 8 21:24:24 2020 +0800

    Merge pull request #4 from NanoCode012/patch-4

    Add drop_last for multi gpu

commit 02c63ef81cf98b28b10344fe2cce08a03b143941
Author: NanoCode012 <kevinvong@rocketmail.com>
Date:   Wed Jul 8 10:08:30 2020 +0700

    Add drop_last for multi gpu

commit b9a50aed48ab1536f94d49269977e2accd67748f
Merge: ec2dc6c 121d90b
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Tue Jul 7 19:48:04 2020 +0800

    Merge branch 'master' of https://github.com/ultralytics/yolov5 into feature/DDP_fixed

commit ec2dc6cc56de43ddff939e14c450672d0fbf9b3d
Merge: d0326e3 82a6182
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Tue Jul 7 19:34:31 2020 +0800

    Merge branch 'feature/DDP_fixed' of https://github.com/MagicFrogSJTU/yolov5 into feature/DDP_fixed

commit d0326e398dfeeeac611ccc64198d4fe91b7aa969
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Tue Jul 7 19:31:24 2020 +0800

    Add SyncBN

commit 82a6182b3ad0689a4432b631b438004e5acb3b74
Merge: 96fa40a 050b2a5
Author: yzchen <Chenyzsjtu@gmail.com>
Date:   Tue Jul 7 19:21:01 2020 +0800

    Merge pull request #1 from NanoCode012/patch-2

    Convert BatchNorm to SyncBatchNorm

commit 050b2a5a79a89c9405854d439a1f70f892139b1c
Author: NanoCode012 <kevinvong@rocketmail.com>
Date:   Tue Jul 7 12:38:14 2020 +0700

    Add cleanup for process_group

commit 2aa330139f3cc1237aeb3132245ed7e5d6da1683
Author: NanoCode012 <kevinvong@rocketmail.com>
Date:   Tue Jul 7 12:07:40 2020 +0700

    Remove apex.parallel. Use torch.nn.parallel

    For future compatibility

commit 77c8e27e603bea9a69e7647587ca8d509dc1990d
Author: NanoCode012 <kevinvong@rocketmail.com>
Date:   Tue Jul 7 01:54:39 2020 +0700

    Convert BatchNorm to SyncBatchNorm

commit 96fa40a3a925e4ffd815fe329e1b5181ec92adc8
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Mon Jul 6 21:53:56 2020 +0800

    Fix the datset inconsistency problem

commit 16e7c269d062c8d16c4d4ff70cc80fd87935dc95
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Mon Jul 6 11:34:03 2020 +0800

    Add loss multiplication to preserver the single-process performance

commit e83805563065ffd2e38f85abe008fc662cc17909
Merge: 625bb49 3bdea3f
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Fri Jul 3 20:56:30 2020 +0800

    Merge branch 'master' of https://github.com/ultralytics/yolov5 into feature/DDP_fixed

commit 625bb49f4e52d781143fea0af36d14e5be8b040c
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Thu Jul 2 22:45:15 2020 +0800

    DDP established

* Squashed commit of the following:

commit 94147314e559a6bdd13cb9de62490d385c27596f
Merge: 65157e2 37acbdc
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Thu Jul 16 14:00:17 2020 +0800

    Merge branch 'master' of https://github.com/ultralytics/yolov4 into feature/DDP_fixed

commit 37acbdc0b6
Author: Glenn Jocher <glenn.jocher@ultralytics.com>
Date:   Wed Jul 15 20:03:41 2020 -0700

    update test.py --save-txt

commit b8c2da4a0d
Author: Glenn Jocher <glenn.jocher@ultralytics.com>
Date:   Wed Jul 15 20:00:48 2020 -0700

    update test.py --save-txt

commit 65157e2fc97d371bc576e18b424e130eb3026917
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Wed Jul 15 16:44:13 2020 +0800

    Revert the README.md removal

commit 1c802bfa503623661d8617ca3f259835d27c5345
Merge: cd55b44 0f3b8bb
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Wed Jul 15 16:43:38 2020 +0800

    Merge branch 'feature/DDP_fixed' of https://github.com/MagicFrogSJTU/yolov5 into feature/DDP_fixed

commit cd55b445c4dcd8003ff4b0b46b64adf7c16e5ce7
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Wed Jul 15 16:42:33 2020 +0800

    fix the DDP performance deterioration bug.

commit 0f3b8bb1fae5885474ba861bbbd1924fb622ee93
Author: Glenn Jocher <glenn.jocher@ultralytics.com>
Date:   Wed Jul 15 00:28:53 2020 -0700

    Delete README.md

commit f5921ba1e35475f24b062456a890238cb7a3cf94
Merge: 85ab2f3 bd3fdbb
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Wed Jul 15 11:20:17 2020 +0800

    Merge branch 'feature/DDP_fixed' of https://github.com/MagicFrogSJTU/yolov5 into feature/DDP_fixed

commit bd3fdbbf1b08ef87931eef49fa8340621caa7e87
Author: Glenn Jocher <glenn.jocher@ultralytics.com>
Date:   Tue Jul 14 18:38:20 2020 -0700

    Update README.md

commit c1a97a7767ccb2aa9afc7a5e72fd159e7c62ec02
Merge: 2bf86b8 f796708
Author: Glenn Jocher <glenn.jocher@ultralytics.com>
Date:   Tue Jul 14 18:36:53 2020 -0700

    Merge branch 'master' into feature/DDP_fixed

commit 2bf86b892fa2fd712f6530903a0d9b8533d7447a
Author: NanoCode012 <kevinvong@rocketmail.com>
Date:   Tue Jul 14 22:18:15 2020 +0700

    Fixed world_size not found when called from test

commit 85ab2f38cdda28b61ad15a3a5a14c3aafb620dc8
Merge: 5a19011 c8357ad
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Tue Jul 14 22:19:58 2020 +0800

    Merge branch 'feature/DDP_fixed' of https://github.com/MagicFrogSJTU/yolov5 into feature/DDP_fixed

commit 5a19011949398d06e744d8d5521ab4e6dfa06ab7
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Tue Jul 14 22:19:15 2020 +0800

    Add assertion for <=2 gpus DDP

commit c8357ad5b15a0e6aeef4d7fe67ca9637f7322a4d
Merge: e742dd9 787582f
Author: yzchen <Chenyzsjtu@gmail.com>
Date:   Tue Jul 14 22:10:02 2020 +0800

    Merge pull request #8 from MagicFrogSJTU/NanoCode012-patch-1

    Modify number of dataloaders' workers

commit 787582f97251834f955ef05a77072b8c673a8397
Author: NanoCode012 <kevinvong@rocketmail.com>
Date:   Tue Jul 14 20:38:58 2020 +0700

    Fixed issue with single gpu not having world_size

commit 63648925288d63a21174a4dd28f92dbfebfeb75a
Author: NanoCode012 <kevinvong@rocketmail.com>
Date:   Tue Jul 14 19:16:15 2020 +0700

    Add assert message for clarification

    Clarify why assertion was thrown to users

commit 69364d6050e048d0d8834e0f30ce84da3f6a13f3
Author: NanoCode012 <kevinvong@rocketmail.com>
Date:   Tue Jul 14 17:36:48 2020 +0700

    Changed number of workers check

commit d738487089e41c22b3b1cd73aa7c1c40320a6ebf
Author: NanoCode012 <kevinvong@rocketmail.com>
Date:   Tue Jul 14 17:33:38 2020 +0700

    Adding world_size

    Reduce calls to torch.distributed. For use in create_dataloader.

commit e742dd9619d29306c7541821238d3d7cddcdc508
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Tue Jul 14 15:38:48 2020 +0800

    Make SyncBN a choice

commit e90d4004387e6103fecad745f8cbc2edc918e906
Merge: 5bf8beb cd90360
Author: yzchen <Chenyzsjtu@gmail.com>
Date:   Tue Jul 14 15:32:10 2020 +0800

    Merge pull request #6 from NanoCode012/patch-5

    Update train.py

commit cd9036017e7f8bd519a8b62adab0f47ea67f4962
Author: NanoCode012 <kevinvong@rocketmail.com>
Date:   Tue Jul 14 13:39:29 2020 +0700

    Update train.py

    Remove redundant `opt.` prefix.

commit 5bf8bebe8873afb18b762fe1f409aca116fac073
Merge: c9558a9 a1c8406
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Tue Jul 14 14:09:51 2020 +0800

    Merge branch 'master' of https://github.com/ultralytics/yolov5 into feature/DDP_fixed

commit c9558a9b51547febb03d9c1ca42e2ef0fc15bb31
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Tue Jul 14 13:51:34 2020 +0800

    Add device allocation for loss compute

commit 4f08c692fb5e943a89e0ee354ef6c80a50eeb28d
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Thu Jul 9 11:16:27 2020 +0800

    Revert drop_last

commit 1dabe33a5a223b758cc761fc8741c6224205a34b
Merge: a1ce9b1 4b8450b
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Thu Jul 9 11:15:49 2020 +0800

    Merge branch 'feature/DDP_fixed' of https://github.com/MagicFrogSJTU/yolov5 into feature/DDP_fixed

commit a1ce9b1e96b71d7fcb9d3e8143013eb8cebe5e27
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Thu Jul 9 11:15:21 2020 +0800

    fix lr warning

commit 4b8450b46db76e5e58cd95df965d4736077cfb0e
Merge: b9a50ae 02c63ef
Author: yzchen <Chenyzsjtu@gmail.com>
Date:   Wed Jul 8 21:24:24 2020 +0800

    Merge pull request #4 from NanoCode012/patch-4

    Add drop_last for multi gpu

commit 02c63ef81cf98b28b10344fe2cce08a03b143941
Author: NanoCode012 <kevinvong@rocketmail.com>
Date:   Wed Jul 8 10:08:30 2020 +0700

    Add drop_last for multi gpu

commit b9a50aed48ab1536f94d49269977e2accd67748f
Merge: ec2dc6c 121d90b
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Tue Jul 7 19:48:04 2020 +0800

    Merge branch 'master' of https://github.com/ultralytics/yolov5 into feature/DDP_fixed

commit ec2dc6cc56de43ddff939e14c450672d0fbf9b3d
Merge: d0326e3 82a6182
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Tue Jul 7 19:34:31 2020 +0800

    Merge branch 'feature/DDP_fixed' of https://github.com/MagicFrogSJTU/yolov5 into feature/DDP_fixed

commit d0326e398dfeeeac611ccc64198d4fe91b7aa969
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Tue Jul 7 19:31:24 2020 +0800

    Add SyncBN

commit 82a6182b3ad0689a4432b631b438004e5acb3b74
Merge: 96fa40a 050b2a5
Author: yzchen <Chenyzsjtu@gmail.com>
Date:   Tue Jul 7 19:21:01 2020 +0800

    Merge pull request #1 from NanoCode012/patch-2

    Convert BatchNorm to SyncBatchNorm

commit 050b2a5a79a89c9405854d439a1f70f892139b1c
Author: NanoCode012 <kevinvong@rocketmail.com>
Date:   Tue Jul 7 12:38:14 2020 +0700

    Add cleanup for process_group

commit 2aa330139f3cc1237aeb3132245ed7e5d6da1683
Author: NanoCode012 <kevinvong@rocketmail.com>
Date:   Tue Jul 7 12:07:40 2020 +0700

    Remove apex.parallel. Use torch.nn.parallel

    For future compatibility

commit 77c8e27e603bea9a69e7647587ca8d509dc1990d
Author: NanoCode012 <kevinvong@rocketmail.com>
Date:   Tue Jul 7 01:54:39 2020 +0700

    Convert BatchNorm to SyncBatchNorm

commit 96fa40a3a925e4ffd815fe329e1b5181ec92adc8
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Mon Jul 6 21:53:56 2020 +0800

    Fix the datset inconsistency problem

commit 16e7c269d062c8d16c4d4ff70cc80fd87935dc95
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Mon Jul 6 11:34:03 2020 +0800

    Add loss multiplication to preserver the single-process performance

commit e83805563065ffd2e38f85abe008fc662cc17909
Merge: 625bb49 3bdea3f
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Fri Jul 3 20:56:30 2020 +0800

    Merge branch 'master' of https://github.com/ultralytics/yolov5 into feature/DDP_fixed

commit 625bb49f4e52d781143fea0af36d14e5be8b040c
Author: yizhi.chen <chenyzsjtu@outlook.com>
Date:   Thu Jul 2 22:45:15 2020 +0800

    DDP established

* Fixed destroy_process_group in DP mode

* Update torch_utils.py

* Update utils.py

Revert build_targets() to current master.

* Update datasets.py

* Fixed world_size attribute not found

Co-authored-by: NanoCode012 <kevinvong@rocketmail.com>
Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>

utils/utils.py

@@ -8,6 +8,7 @@ import time
 from copy import copy
 from pathlib import Path
 from sys import platform
+from contextlib import contextmanager
 
 import cv2
 import matplotlib
@@ -31,6 +32,18 @@ matplotlib.rc('font', **{'size': 11})
 cv2.setNumThreads(0)
 
 
+@contextmanager
+def torch_distributed_zero_first(local_rank: int):
+    """
+    Decorator to make all processes in distributed training wait for each local_master to do something.
+    """
+    if local_rank not in [-1, 0]:
+        torch.distributed.barrier()
+    yield
+    if local_rank == 0:
+        torch.distributed.barrier()
+
+
 def init_seeds(seed=0):
     random.seed(seed)
     np.random.seed(seed)
@@ -424,15 +437,16 @@ class BCEBlurWithLogitsLoss(nn.Module):
 
 
 def compute_loss(p, targets, model):  # predictions, targets, model
+    device = targets.device
     ft = torch.cuda.FloatTensor if p[0].is_cuda else torch.Tensor
-    lcls, lbox, lobj = ft([0]), ft([0]), ft([0])
+    lcls, lbox, lobj = ft([0]).to(device), ft([0]).to(device), ft([0]).to(device)
     tcls, tbox, indices, anchors = build_targets(p, targets, model)  # targets
     h = model.hyp  # hyperparameters
     red = 'mean'  # Loss reduction (sum or mean)
 
     # Define criteria
-    BCEcls = nn.BCEWithLogitsLoss(pos_weight=ft([h['cls_pw']]), reduction=red)
-    BCEobj = nn.BCEWithLogitsLoss(pos_weight=ft([h['obj_pw']]), reduction=red)
+    BCEcls = nn.BCEWithLogitsLoss(pos_weight=ft([h['cls_pw']]), reduction=red).to(device)
+    BCEobj = nn.BCEWithLogitsLoss(pos_weight=ft([h['obj_pw']]), reduction=red).to(device)
 
     # class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
     cp, cn = smooth_BCE(eps=0.0)
@@ -448,7 +462,7 @@ def compute_loss(p, targets, model):  # predictions, targets, model
     balance = [1.0, 1.0, 1.0]
     for i, pi in enumerate(p):  # layer index, layer predictions
         b, a, gj, gi = indices[i]  # image, anchor, gridy, gridx
-        tobj = torch.zeros_like(pi[..., 0])  # target obj
+        tobj = torch.zeros_like(pi[..., 0]).to(device)  # target obj
 
         nb = b.shape[0]  # number of targets
         if nb:
@@ -458,7 +472,7 @@ def compute_loss(p, targets, model):  # predictions, targets, model
             # GIoU
             pxy = ps[:, :2].sigmoid() * 2. - 0.5
             pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]
-            pbox = torch.cat((pxy, pwh), 1)  # predicted box
+            pbox = torch.cat((pxy, pwh), 1).to(device)  # predicted box
             giou = bbox_iou(pbox.t(), tbox[i], x1y1x2y2=False, GIoU=True)  # giou(prediction, target)
             lbox += (1.0 - giou).sum() if red == 'sum' else (1.0 - giou).mean()  # giou loss
 
@@ -467,7 +481,7 @@ def compute_loss(p, targets, model):  # predictions, targets, model
 
             # Class
             if model.nc > 1:  # cls loss (only if multiple classes)
-                t = torch.full_like(ps[:, 5:], cn)  # targets
+                t = torch.full_like(ps[:, 5:], cn).to(device)  # targets
                 t[range(nb), tcls[i]] = cp
                 lcls += BCEcls(ps[:, 5:], t)  # BCE