Merge remote-tracking branch 'upstream/multiweight' into multiweight_tests

Author: datumbox

README.rst

@@ -21,7 +21,9 @@ supported Python versions.
 +--------------------------+--------------------------+---------------------------------+
 | ``torch``                | ``torchvision``          | ``python``                      |
 +==========================+==========================+=================================+
-| ``main`` / ``nightly``   | ``main`` / ``nightly``   | ``>=3.7``, ``<=3.9``            |
+| ``main`` / ``nightly``   | ``main`` / ``nightly``   | ``>=3.7``, ``<=3.10``           |
++--------------------------+--------------------------+---------------------------------+
+| ``1.11.0``               | ``0.12.0``               | ``>=3.7``, ``<=3.10``           |
 +--------------------------+--------------------------+---------------------------------+
 | ``1.10.2``               | ``0.11.3``               | ``>=3.6``, ``<=3.9``            |
 +--------------------------+--------------------------+---------------------------------+

references/classification/transforms.py

@@ -21,8 +21,14 @@ class RandomMixup(torch.nn.Module):
 
     def __init__(self, num_classes: int, p: float = 0.5, alpha: float = 1.0, inplace: bool = False) -> None:
         super().__init__()
-        assert num_classes > 0, "Please provide a valid positive value for the num_classes."
-        assert alpha > 0, "Alpha param can't be zero."
+
+        if num_classes < 1:
+            raise ValueError(
+                f"Please provide a valid positive value for the num_classes. Got num_classes={num_classes}"
+            )
+
+        if alpha <= 0:
+            raise ValueError("Alpha param can't be zero.")
 
         self.num_classes = num_classes
         self.p = p
@@ -99,8 +105,10 @@ class RandomCutmix(torch.nn.Module):
 
     def __init__(self, num_classes: int, p: float = 0.5, alpha: float = 1.0, inplace: bool = False) -> None:
         super().__init__()
-        assert num_classes > 0, "Please provide a valid positive value for the num_classes."
-        assert alpha > 0, "Alpha param can't be zero."
+        if num_classes < 1:
+            raise ValueError("Please provide a valid positive value for the num_classes.")
+        if alpha <= 0:
+            raise ValueError("Alpha param can't be zero.")
 
         self.num_classes = num_classes
         self.p = p

references/detection/coco_eval.py

@@ -12,7 +12,8 @@
 
 class CocoEvaluator:
     def __init__(self, coco_gt, iou_types):
-        assert isinstance(iou_types, (list, tuple))
+        if not isinstance(iou_types, (list, tuple)):
+            raise TypeError(f"This constructor expects iou_types of type list or tuple, instead  got {type(iou_types)}")
         coco_gt = copy.deepcopy(coco_gt)
         self.coco_gt = coco_gt
 

references/detection/coco_utils.py

@@ -126,7 +126,10 @@ def _has_valid_annotation(anno):
             return True
         return False
 
-    assert isinstance(dataset, torchvision.datasets.CocoDetection)
+    if not isinstance(dataset, torchvision.datasets.CocoDetection):
+        raise TypeError(
+            f"This function expects dataset of type torchvision.datasets.CocoDetection, instead  got {type(dataset)}"
+        )
     ids = []
     for ds_idx, img_id in enumerate(dataset.ids):
         ann_ids = dataset.coco.getAnnIds(imgIds=img_id, iscrowd=None)

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 inner_loop(blob):
         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 inner_loop(blob):
         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:
@@ -145,21 +151,21 @@ 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":
             for pass_name in ("clean", "final"):
                 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/transforms.py

@@ -7,16 +7,21 @@ class ValidateModelInput(torch.nn.Module):
     # Pass-through transform that checks the shape and dtypes to make sure the model gets what it expects
     def forward(self, img1, img2, flow, valid_flow_mask):
 
-        assert all(isinstance(arg, torch.Tensor) for arg in (img1, img2, flow, valid_flow_mask) if arg is not None)
-        assert all(arg.dtype == torch.float32 for arg in (img1, img2, flow) if arg is not None)
+        if not all(isinstance(arg, torch.Tensor) for arg in (img1, img2, flow, valid_flow_mask) if arg is not None):
+            raise TypeError("This method expects all input arguments to be of type torch.Tensor.")
+        if not all(arg.dtype == torch.float32 for arg in (img1, img2, flow) if arg is not None):
+            raise TypeError("This method expects the tensors img1, img2 and flow of be of dtype torch.float32.")
 
-        assert img1.shape == img2.shape
+        if img1.shape != img2.shape:
+            raise ValueError("img1 and img2 should have the same shape.")
         h, w = img1.shape[-2:]
-        if flow is not None:
-            assert flow.shape == (2, h, w)
+        if flow is not None and flow.shape != (2, h, w):
+            raise ValueError(f"flow.shape should be (2, {h}, {w}) instead of {flow.shape}")
         if valid_flow_mask is not None:
-            assert valid_flow_mask.shape == (h, w)
-            assert valid_flow_mask.dtype == torch.bool
+            if valid_flow_mask.shape != (h, w):
+                raise ValueError(f"valid_flow_mask.shape should be ({h}, {w}) instead of {valid_flow_mask.shape}")
+            if valid_flow_mask.dtype != torch.bool:
+                raise TypeError("valid_flow_mask should be of dtype torch.bool instead of {valid_flow_mask.dtype}")
 
         return img1, img2, flow, valid_flow_mask
 
@@ -109,7 +114,8 @@ class RandomErasing(T.RandomErasing):
     def __init__(self, p=0.5, scale=(0.02, 0.33), ratio=(0.3, 3.3), value=0, inplace=False, max_erase=1):
         super().__init__(p=p, scale=scale, ratio=ratio, value=value, inplace=inplace)
         self.max_erase = max_erase
-        assert self.max_erase > 0
+        if self.max_erase <= 0:
+            raise ValueError("max_raise should be greater than 0")
 
     def forward(self, img1, img2, flow, valid_flow_mask):
         if torch.rand(1) > self.p:

references/optical_flow/utils.py

@@ -71,7 +71,10 @@ def update(self, **kwargs):
         for k, v in kwargs.items():
             if isinstance(v, torch.Tensor):
                 v = v.item()
-            assert isinstance(v, (float, int))
+            if not isinstance(v, (float, int)):
+                raise TypeError(
+                    f"This method expects the value of the input arguments to be of type float or int, instead  got {type(v)}"
+                )
             self.meters[k].update(v)
 
     def __getattr__(self, attr):
@@ -256,7 +259,12 @@ 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))
 

references/segmentation/coco_utils.py

@@ -68,7 +68,11 @@ def _has_valid_annotation(anno):
         # if more than 1k pixels occupied in the image
         return sum(obj["area"] for obj in anno) > 1000
 
-    assert isinstance(dataset, torchvision.datasets.CocoDetection)
+    if not isinstance(dataset, torchvision.datasets.CocoDetection):
+        raise TypeError(
+            f"This function expects dataset of type torchvision.datasets.CocoDetection, instead  got {type(dataset)}"
+        )
+
     ids = []
     for ds_idx, img_id in enumerate(dataset.ids):
         ann_ids = dataset.coco.getAnnIds(imgIds=img_id, iscrowd=None)

references/segmentation/utils.py

@@ -118,7 +118,10 @@ def update(self, **kwargs):
         for k, v in kwargs.items():
             if isinstance(v, torch.Tensor):
                 v = v.item()
-            assert isinstance(v, (float, int))
+            if not isinstance(v, (float, int)):
+                raise TypeError(
+                    f"This method expects the value of the input arguments to be of type float or int, instead  got {type(v)}"
+                )
             self.meters[k].update(v)
 
     def __getattr__(self, attr):