RAFT training reference Improvement (#5590)

* Change optical flow train.py function name from validate to evaluate so it is similar to other references

* Add --device as parameter and enable to run in non distributed mode

* Format with ufmt

* Fix unneccessary param and bug

* Enable saving the optimizer and scheduler on the checkpoint

* Fix bug when evaluate before resume and save or load model without ddp

* Fix case where --train-dataset is None
Co-authored-by: Nicolas Hug <nicolashug@fb.com>

references/optical_flow/train.py

@@ -60,16 +60,21 @@ def get_train_dataset(stage, dataset_root):
 
 
 @torch.no_grad()
-def _validate(model, args, val_dataset, *, padder_mode, num_flow_updates=None, batch_size=None, header=None):
+def _evaluate(model, args, val_dataset, *, padder_mode, num_flow_updates=None, batch_size=None, header=None):
     """Helper function to compute various metrics (epe, etc.) for a model on a given dataset.
 
     We process as many samples as possible with ddp, and process the rest on a single worker.
     """
     batch_size = batch_size or args.batch_size
+    device = torch.device(args.device)
 
     model.eval()
 
-    sampler = torch.utils.data.distributed.DistributedSampler(val_dataset, shuffle=False, drop_last=True)
+    if args.distributed:
+        sampler = torch.utils.data.distributed.DistributedSampler(val_dataset, shuffle=False, drop_last=True)
+    else:
+        sampler = torch.utils.data.SequentialSampler(val_dataset)
+
     val_loader = torch.utils.data.DataLoader(
         val_dataset,
         sampler=sampler,
@@ -88,7 +93,7 @@ def _validate(model, args, val_dataset, *, padder_mode, num_flow_updates=None, b
         image1, image2, flow_gt = blob[:3]
         valid_flow_mask = None if len(blob) == 3 else blob[-1]
 
-        image1, image2 = image1.cuda(), image2.cuda()
+        image1, image2 = image1.to(device), image2.to(device)
 
         padder = utils.InputPadder(image1.shape, mode=padder_mode)
         image1, image2 = padder.pad(image1, image2)
@@ -115,21 +120,22 @@ def _validate(model, args, val_dataset, *, padder_mode, num_flow_updates=None, b
         inner_loop(blob)
         num_processed_samples += blob[0].shape[0]  # batch size
 
-    num_processed_samples = utils.reduce_across_processes(num_processed_samples)
-    print(
-        f"Batch-processed {num_processed_samples} / {len(val_dataset)} samples. "
-        "Going to process the remaining samples individually, if any."
-    )
+    if args.distributed:
+        num_processed_samples = utils.reduce_across_processes(num_processed_samples)
+        print(
+            f"Batch-processed {num_processed_samples} / {len(val_dataset)} samples. "
+            "Going to process the remaining samples individually, if any."
+        )
+        if args.rank == 0:  # we only need to process the rest on a single worker
+            for i in range(num_processed_samples, len(val_dataset)):
+                inner_loop(val_dataset[i])
 
-    if args.rank == 0:  # we only need to process the rest on a single worker
-        for i in range(num_processed_samples, len(val_dataset)):
-            inner_loop(val_dataset[i])
+        logger.synchronize_between_processes()
 
-    logger.synchronize_between_processes()
     print(header, logger)
 
 
-def validate(model, args):
+def evaluate(model, args):
     val_datasets = args.val_dataset or []
 
     if args.prototype:
@@ -145,13 +151,13 @@ def validate(model, args):
         if name == "kitti":
             # Kitti has different image sizes so we need to individually pad them, we can't batch.
             # see comment in InputPadder
-            if args.batch_size != 1 and args.rank == 0:
+            if args.batch_size != 1 and (not args.distributed or args.rank == 0):
                 warnings.warn(
                     f"Batch-size={args.batch_size} was passed. For technical reasons, evaluating on Kitti can only be done with a batch-size of 1."
                 )
 
             val_dataset = KittiFlow(root=args.dataset_root, split="train", transforms=preprocessing)
-            _validate(
+            _evaluate(
                 model, args, val_dataset, num_flow_updates=24, padder_mode="kitti", header="Kitti val", batch_size=1
             )
         elif name == "sintel":
@@ -159,7 +165,7 @@ def validate(model, args):
                 val_dataset = Sintel(
                     root=args.dataset_root, split="train", pass_name=pass_name, transforms=preprocessing
                 )
-                _validate(
+                _evaluate(
                     model,
                     args,
                     val_dataset,
@@ -172,11 +178,12 @@ def validate(model, args):
 
 
 def train_one_epoch(model, optimizer, scheduler, train_loader, logger, args):
+    device = torch.device(args.device)
     for data_blob in logger.log_every(train_loader):
 
         optimizer.zero_grad()
 
-        image1, image2, flow_gt, valid_flow_mask = (x.cuda() for x in data_blob)
+        image1, image2, flow_gt, valid_flow_mask = (x.to(device) for x in data_blob)
         flow_predictions = model(image1, image2, num_flow_updates=args.num_flow_updates)
 
         loss = utils.sequence_loss(flow_predictions, flow_gt, valid_flow_mask, args.gamma)
@@ -200,36 +207,68 @@ def main(args):
         raise ValueError("The weights parameter works only in prototype mode. Please pass the --prototype argument.")
     utils.setup_ddp(args)
 
+    if args.distributed and args.device == "cpu":
+        raise ValueError("The device must be cuda if we want to run in distributed mode using torchrun")
+    device = torch.device(args.device)
+
     if args.prototype:
         model = prototype.models.optical_flow.__dict__[args.model](weights=args.weights)
     else:
         model = torchvision.models.optical_flow.__dict__[args.model](pretrained=args.pretrained)
 
-    model = model.to(args.local_rank)
-    model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank])
+    if args.distributed:
+        model = model.to(args.local_rank)
+        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank])
+        model_without_ddp = model.module
+    else:
+        model.to(device)
+        model_without_ddp = model
 
     if args.resume is not None:
-        d = torch.load(args.resume, map_location="cpu")
-        model.load_state_dict(d, strict=True)
+        checkpoint = torch.load(args.resume, map_location="cpu")
+        model_without_ddp.load_state_dict(checkpoint["model"])
 
     if args.train_dataset is None:
         # Set deterministic CUDNN algorithms, since they can affect epe a fair bit.
         torch.backends.cudnn.benchmark = False
         torch.backends.cudnn.deterministic = True
-        validate(model, args)
+        evaluate(model, args)
         return
 
     print(f"Parameter Count: {sum(p.numel() for p in model.parameters() if p.requires_grad)}")
 
+    train_dataset = get_train_dataset(args.train_dataset, args.dataset_root)
+
+    optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr, weight_decay=args.weight_decay, eps=args.adamw_eps)
+
+    scheduler = torch.optim.lr_scheduler.OneCycleLR(
+        optimizer=optimizer,
+        max_lr=args.lr,
+        epochs=args.epochs,
+        steps_per_epoch=ceil(len(train_dataset) / (args.world_size * args.batch_size)),
+        pct_start=0.05,
+        cycle_momentum=False,
+        anneal_strategy="linear",
+    )
+
+    if args.resume is not None:
+        optimizer.load_state_dict(checkpoint["optimizer"])
+        scheduler.load_state_dict(checkpoint["scheduler"])
+        args.start_epoch = checkpoint["epoch"] + 1
+    else:
+        args.start_epoch = 0
+
     torch.backends.cudnn.benchmark = True
 
     model.train()
     if args.freeze_batch_norm:
         utils.freeze_batch_norm(model.module)
 
-    train_dataset = get_train_dataset(args.train_dataset, args.dataset_root)
+    if args.distributed:
+        sampler = torch.utils.data.distributed.DistributedSampler(train_dataset, shuffle=True, drop_last=True)
+    else:
+        sampler = torch.utils.data.RandomSampler(train_dataset)
 
-    sampler = torch.utils.data.distributed.DistributedSampler(train_dataset, shuffle=True, drop_last=True)
     train_loader = torch.utils.data.DataLoader(
         train_dataset,
         sampler=sampler,
@@ -238,25 +277,15 @@ def main(args):
         num_workers=args.num_workers,
     )
 
-    optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr, weight_decay=args.weight_decay, eps=args.adamw_eps)
-
-    scheduler = torch.optim.lr_scheduler.OneCycleLR(
-        optimizer=optimizer,
-        max_lr=args.lr,
-        epochs=args.epochs,
-        steps_per_epoch=ceil(len(train_dataset) / (args.world_size * args.batch_size)),
-        pct_start=0.05,
-        cycle_momentum=False,
-        anneal_strategy="linear",
-    )
-
     logger = utils.MetricLogger()
 
     done = False
-    for current_epoch in range(args.epochs):
+    for current_epoch in range(args.start_epoch, args.epochs):
         print(f"EPOCH {current_epoch}")
+        if args.distributed:
+            # needed on distributed mode, otherwise the data loading order would be the same for all epochs
+            sampler.set_epoch(current_epoch)
 
-        sampler.set_epoch(current_epoch)  # needed, otherwise the data loading order would be the same for all epochs
         train_one_epoch(
             model=model,
             optimizer=optimizer,
@@ -269,13 +298,19 @@ def main(args):
         # Note: we don't sync the SmoothedValues across processes, so the printed metrics are just those of rank 0
         print(f"Epoch {current_epoch} done. ", logger)
 
-        if args.rank == 0:
-            # TODO: Also save the optimizer and scheduler
-            torch.save(model.state_dict(), Path(args.output_dir) / f"{args.name}_{current_epoch}.pth")
-            torch.save(model.state_dict(), Path(args.output_dir) / f"{args.name}.pth")
+        if not args.distributed or args.rank == 0:
+            checkpoint = {
+                "model": model_without_ddp.state_dict(),
+                "optimizer": optimizer.state_dict(),
+                "scheduler": scheduler.state_dict(),
+                "epoch": current_epoch,
+                "args": args,
+            }
+            torch.save(checkpoint, Path(args.output_dir) / f"{args.name}_{current_epoch}.pth")
+            torch.save(checkpoint, Path(args.output_dir) / f"{args.name}.pth")
 
         if current_epoch % args.val_freq == 0 or done:
-            validate(model, args)
+            evaluate(model, args)
             model.train()
             if args.freeze_batch_norm:
                 utils.freeze_batch_norm(model.module)
@@ -349,6 +384,7 @@ def get_args_parser(add_help=True):
         action="store_true",
     )
     parser.add_argument("--weights", default=None, type=str, help="the weights enum name to load.")
+    parser.add_argument("--device", default="cuda", type=str, help="device (Use cuda or cpu, Default: cuda)")
 
     return parser
 

references/optical_flow/utils.py

@@ -256,7 +256,12 @@ def setup_ddp(args):
         # if we're here, the script was called by run_with_submitit.py
         args.local_rank = args.gpu
     else:
-        raise ValueError(r"Sorry, I can't set up the distributed training ¯\_(ツ)_/¯.")
+        print("Not using distributed mode!")
+        args.distributed = False
+        args.world_size = 1
+        return
+
+    args.distributed = True
 
     _redefine_print(is_main=(args.rank == 0))