[fbsync] Add training reference for optical flow models (#5027)

Reviewed By: NicolasHug

Differential Revision: D32950938

fbshipit-source-id: 0f271d45026c821c109493d9aa7f404b5373012d

Author: Vincent Moens

references/optical_flow/train.py

@@ -0,0 +1,334 @@
+import argparse
+import warnings
+from pathlib import Path
+
+import torch
+import utils
+from presets import OpticalFlowPresetTrain, OpticalFlowPresetEval
+from torchvision.datasets import KittiFlow, FlyingChairs, FlyingThings3D, Sintel, HD1K
+from torchvision.models.optical_flow import raft_large, raft_small
+
+
+def get_train_dataset(stage, dataset_root):
+    if stage == "chairs":
+        transforms = OpticalFlowPresetTrain(crop_size=(368, 496), min_scale=0.1, max_scale=1.0, do_flip=True)
+        return FlyingChairs(root=dataset_root, split="train", transforms=transforms)
+    elif stage == "things":
+        transforms = OpticalFlowPresetTrain(crop_size=(400, 720), min_scale=-0.4, max_scale=0.8, do_flip=True)
+        return FlyingThings3D(root=dataset_root, split="train", pass_name="both", transforms=transforms)
+    elif stage == "sintel_SKH":  # S + K + H as from paper
+        crop_size = (368, 768)
+        transforms = OpticalFlowPresetTrain(crop_size=crop_size, min_scale=-0.2, max_scale=0.6, do_flip=True)
+
+        things_clean = FlyingThings3D(root=dataset_root, split="train", pass_name="clean", transforms=transforms)
+        sintel = Sintel(root=dataset_root, split="train", pass_name="both", transforms=transforms)
+
+        kitti_transforms = OpticalFlowPresetTrain(crop_size=crop_size, min_scale=-0.3, max_scale=0.5, do_flip=True)
+        kitti = KittiFlow(root=dataset_root, split="train", transforms=kitti_transforms)
+
+        hd1k_transforms = OpticalFlowPresetTrain(crop_size=crop_size, min_scale=-0.5, max_scale=0.2, do_flip=True)
+        hd1k = HD1K(root=dataset_root, split="train", transforms=hd1k_transforms)
+
+        # As future improvement, we could probably be using a distributed sampler here
+        # The distribution is S(.71), T(.135), K(.135), H(.02)
+        return 100 * sintel + 200 * kitti + 5 * hd1k + things_clean
+    elif stage == "kitti":
+        transforms = OpticalFlowPresetTrain(
+            # resize and crop params
+            crop_size=(288, 960),
+            min_scale=-0.2,
+            max_scale=0.4,
+            stretch_prob=0,
+            # flip params
+            do_flip=False,
+            # jitter params
+            brightness=0.3,
+            contrast=0.3,
+            saturation=0.3,
+            hue=0.3 / 3.14,
+            asymmetric_jitter_prob=0,
+        )
+        return KittiFlow(root=dataset_root, split="train", transforms=transforms)
+    else:
+        raise ValueError(f"Unknown stage {stage}")
+
+
+@torch.no_grad()
+def _validate(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
+
+    model.eval()
+
+    sampler = torch.utils.data.distributed.DistributedSampler(val_dataset, shuffle=False, drop_last=True)
+    val_loader = torch.utils.data.DataLoader(
+        val_dataset,
+        sampler=sampler,
+        batch_size=batch_size,
+        pin_memory=True,
+        num_workers=args.num_workers,
+    )
+
+    num_flow_updates = num_flow_updates or args.num_flow_updates
+
+    def inner_loop(blob):
+        if blob[0].dim() == 3:
+            # input is not batched so we add an extra dim for consistency
+            blob = [x[None, :, :, :] if x is not None else None for x in blob]
+
+        image1, image2, flow_gt = blob[:3]
+        valid_flow_mask = None if len(blob) == 3 else blob[-1]
+
+        image1, image2 = image1.cuda(), image2.cuda()
+
+        padder = utils.InputPadder(image1.shape, mode=padder_mode)
+        image1, image2 = padder.pad(image1, image2)
+
+        flow_predictions = model(image1, image2, num_flow_updates=num_flow_updates)
+        flow_pred = flow_predictions[-1]
+        flow_pred = padder.unpad(flow_pred).cpu()
+
+        metrics, num_pixels_tot = utils.compute_metrics(flow_pred, flow_gt, valid_flow_mask)
+
+        # We compute per-pixel epe (epe) and per-image epe (called f1-epe in RAFT paper).
+        # per-pixel epe: average epe of all pixels of all images
+        # per-image epe: average epe on each image independently, then average over images
+        for name in ("epe", "1px", "3px", "5px", "f1"):  # f1 is called f1-all in paper
+            logger.meters[name].update(metrics[name], n=num_pixels_tot)
+        logger.meters["per_image_epe"].update(metrics["epe"], n=batch_size)
+
+    logger = utils.MetricLogger()
+    for meter_name in ("epe", "1px", "3px", "5px", "per_image_epe", "f1"):
+        logger.add_meter(meter_name, fmt="{global_avg:.4f}")
+
+    num_processed_samples = 0
+    for blob in logger.log_every(val_loader, header=header, print_freq=None):
+        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.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()
+    print(header, logger)
+
+
+def validate(model, args):
+    val_datasets = args.val_dataset or []
+    for name in val_datasets:
+        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:
+                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=OpticalFlowPresetEval())
+            _validate(
+                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=OpticalFlowPresetEval()
+                )
+                _validate(
+                    model,
+                    args,
+                    val_dataset,
+                    num_flow_updates=32,
+                    padder_mode="sintel",
+                    header=f"Sintel val {pass_name}",
+                )
+        else:
+            warnings.warn(f"Can't validate on {val_dataset}, skipping.")
+
+
+def train_one_epoch(model, optimizer, scheduler, train_loader, logger, current_step, args):
+    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)
+        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)
+        metrics, _ = utils.compute_metrics(flow_predictions[-1], flow_gt, valid_flow_mask)
+
+        metrics.pop("f1")
+        logger.update(loss=loss, **metrics)
+
+        loss.backward()
+
+        torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1)
+
+        optimizer.step()
+        scheduler.step()
+
+        current_step += 1
+
+        if current_step == args.num_steps:
+            return True, current_step
+
+    return False, current_step
+
+
+def main(args):
+    utils.setup_ddp(args)
+
+    model = raft_small() if args.small else raft_large()
+    model = model.to(args.local_rank)
+    model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank])
+
+    if args.resume is not None:
+        d = torch.load(args.resume, map_location="cpu")
+        model.load_state_dict(d, strict=True)
+
+    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)
+        return
+
+    print(f"Parameter Count: {sum(p.numel() for p in model.parameters() if p.requires_grad)}")
+
+    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)
+
+    sampler = torch.utils.data.distributed.DistributedSampler(train_dataset, shuffle=True, drop_last=True)
+    train_loader = torch.utils.data.DataLoader(
+        train_dataset,
+        sampler=sampler,
+        batch_size=args.batch_size,
+        pin_memory=True,
+        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,
+        total_steps=args.num_steps + 100,
+        pct_start=0.05,
+        cycle_momentum=False,
+        anneal_strategy="linear",
+    )
+
+    logger = utils.MetricLogger()
+
+    done = False
+    current_epoch = current_step = 0
+    while not done:
+        print(f"EPOCH {current_epoch}")
+
+        sampler.set_epoch(current_epoch)  # needed, otherwise the data loading order would be the same for all epochs
+        done, current_step = train_one_epoch(
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            train_loader=train_loader,
+            logger=logger,
+            current_step=current_step,
+            args=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)
+
+        current_epoch += 1
+
+        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 current_epoch % args.val_freq == 0 or done:
+            validate(model, args)
+            model.train()
+            if args.freeze_batch_norm:
+                utils.freeze_batch_norm(model.module)
+
+
+def get_args_parser(add_help=True):
+    parser = argparse.ArgumentParser(add_help=add_help, description="Train or evaluate an optical-flow model.")
+    parser.add_argument(
+        "--name",
+        default="raft",
+        type=str,
+        help="The name of the experiment - determines the name of the files where weights are saved.",
+    )
+    parser.add_argument(
+        "--output-dir", default="checkpoints", type=str, help="Output dir where checkpoints will be stored."
+    )
+    parser.add_argument(
+        "--resume",
+        type=str,
+        help="A path to previously saved weights. Used to re-start training from, or evaluate a pre-saved model.",
+    )
+
+    parser.add_argument("--num-workers", type=int, default=12, help="Number of workers for the data loading part.")
+
+    parser.add_argument(
+        "--train-dataset",
+        type=str,
+        help="The dataset to use for training. If not passed, only validation is performed (and you probably want to pass --resume).",
+    )
+    parser.add_argument("--val-dataset", type=str, nargs="+", help="The dataset(s) to use for validation.")
+    parser.add_argument("--val-freq", type=int, default=2, help="Validate every X epochs")
+    # TODO: eventually, it might be preferable to support epochs instead of num_steps.
+    # Keeping it this way for now to reproduce results more easily.
+    parser.add_argument("--num-steps", type=int, default=100000, help="The total number of steps (updates) to train.")
+    parser.add_argument("--batch-size", type=int, default=6)
+
+    parser.add_argument("--lr", type=float, default=0.00002, help="Learning rate for AdamW optimizer")
+    parser.add_argument("--weight-decay", type=float, default=0.00005, help="Weight decay for AdamW optimizer")
+    parser.add_argument("--adamw-eps", type=float, default=1e-8, help="eps value for AdamW optimizer")
+
+    parser.add_argument(
+        "--freeze-batch-norm", action="store_true", help="Set BatchNorm modules of the model in eval mode."
+    )
+
+    parser.add_argument("--small", action="store_true", help="Use the 'small' RAFT architecture.")
+
+    parser.add_argument(
+        "--num_flow_updates",
+        type=int,
+        default=12,
+        help="number of updates (or 'iters') in the update operator of the model.",
+    )
+
+    parser.add_argument("--gamma", type=float, default=0.8, help="exponential weighting for loss. Must be < 1.")
+
+    parser.add_argument("--dist-url", default="env://", help="URL used to set up distributed training")
+
+    parser.add_argument(
+        "--dataset-root",
+        help="Root folder where the datasets are stored. Will be passed as the 'root' parameter of the datasets.",
+        required=True,
+    )
+
+    return parser
+
+
+if __name__ == "__main__":
+    args = get_args_parser().parse_args()
+    Path(args.output_dir).mkdir(exist_ok=True)
+    main(args)