[fbsync] Add pretrained weights on Chairs and Things for raft_large (#5060)

Reviewed By: fmassa

Differential Revision: D33185004

fbshipit-source-id: bdd968bd22775c2f63a8e67877b6482bfb58cc5a

Author: NicolasHug

references/optical_flow/README.md

@@ -0,0 +1,57 @@
+# Optical flow reference training scripts
+
+This folder contains reference training scripts for optical flow.
+They serve as a log of how to train specific models, so as to provide baseline
+training and evaluation scripts to quickly bootstrap research.
+
+
+### RAFT Large
+
+The RAFT large model was trained on Flying Chairs and then on Flying Things.
+Both used 8 A100 GPUs and a batch size of 2 (so effective batch size is 16). The
+rest of the hyper-parameters are exactly the same as the original RAFT training
+recipe from https://github.com/princeton-vl/RAFT.
+
+```
+torchrun --nproc_per_node 8 --nnodes 1 train.py \
+    --dataset-root $dataset_root \
+    --name $name_chairs \
+    --model raft_large \
+    --train-dataset chairs \
+    --batch-size 2 \
+    --lr 0.0004 \
+    --weight-decay 0.0001 \
+    --num-steps 100000 \
+    --output-dir $chairs_dir
+```
+
+```
+torchrun --nproc_per_node 8 --nnodes 1 train.py \
+    --dataset-root $dataset_root \
+    --name $name_things \
+    --model raft_large \
+    --train-dataset things \
+    --batch-size 2 \
+    --lr 0.000125 \
+    --weight-decay 0.0001 \
+    --num-steps 100000 \
+    --freeze-batch-norm \
+    --output-dir $things_dir\
+    --resume $chairs_dir/$name_chairs.pth
+```
+
+
+### Evaluation
+
+```
+torchrun --nproc_per_node 1 --nnodes 1 train.py --val-dataset sintel --batch-size 1 --dataset-root $dataset_root --model raft_large --pretrained
+```
+
+This should give an epe of about 1.3822 on the clean pass and 2.7161 on the
+final pass of Sintel. Results may vary slightly depending on the batch size and
+the number of GPUs. For the most accurate resuts use 1 GPU and `--batch-size 1`:
+
+```
+Sintel val clean epe: 1.3822	1px: 0.9028	3px: 0.9573	5px: 0.9697	per_image_epe: 1.3822	f1: 4.0248
+Sintel val final epe: 2.7161	1px: 0.8528	3px: 0.9204	5px: 0.9392	per_image_epe: 2.7161	f1: 7.5964
+```

references/optical_flow/train.py

@@ -3,10 +3,16 @@
 from pathlib import Path
 
 import torch
+import torchvision.models.optical_flow
 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
+
+try:
+    from torchvision.prototype import models as PM
+    from torchvision.prototype.models import optical_flow as PMOF
+except ImportError:
+    PM = PMOF = None
 
 
 def get_train_dataset(stage, dataset_root):
@@ -125,6 +131,13 @@ def inner_loop(blob):
 
 def validate(model, args):
     val_datasets = args.val_dataset or []
+
+    if args.weights:
+        weights = PM.get_weight(args.weights)
+        preprocessing = weights.transforms()
+    else:
+        preprocessing = OpticalFlowPresetEval()
+
     for name in val_datasets:
         if name == "kitti":
             # Kitti has different image sizes so we need to individually pad them, we can't batch.
@@ -134,14 +147,14 @@ def validate(model, args):
                     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())
+            val_dataset = KittiFlow(root=args.dataset_root, split="train", transforms=preprocessing)
             _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()
+                    root=args.dataset_root, split="train", pass_name=pass_name, transforms=preprocessing
                 )
                 _validate(
                     model,
@@ -187,7 +200,11 @@ def train_one_epoch(model, optimizer, scheduler, train_loader, logger, current_s
 def main(args):
     utils.setup_ddp(args)
 
-    model = raft_small() if args.small else raft_large()
+    if args.weights:
+        model = PMOF.__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])
 
@@ -306,7 +323,12 @@ def get_args_parser(add_help=True):
         "--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(
+        "--model", type=str, default="raft_large", help="The name of the model to use - either raft_large or raft_small"
+    )
+    # TODO: resume, pretrained, and weights should be in an exclusive arg group
+    parser.add_argument("--pretrained", action="store_true", help="Whether to use pretrained weights")
+    parser.add_argument("--weights", default=None, type=str, help="the weights enum name to load.")
 
     parser.add_argument(
         "--num_flow_updates",

test/test_prototype_models.py

@@ -91,7 +91,8 @@ def test_naming_conventions(model_fn):
     + TM.get_models_from_module(models.detection)
     + TM.get_models_from_module(models.quantization)
     + TM.get_models_from_module(models.segmentation)
-    + TM.get_models_from_module(models.video),
+    + TM.get_models_from_module(models.video)
+    + TM.get_models_from_module(models.optical_flow),
 )
 def test_schema_meta_validation(model_fn):
     classification_fields = ["size", "categories", "acc@1", "acc@5"]
@@ -102,6 +103,7 @@ def test_schema_meta_validation(model_fn):
         "quantization": classification_fields + ["backend", "quantization", "unquantized"],
         "segmentation": ["categories", "mIoU", "acc"],
         "video": classification_fields,
+        "optical_flow": [],
     }
     module_name = model_fn.__module__.split(".")[-2]
     fields = set(defaults["all"] + defaults[module_name])
@@ -201,13 +203,18 @@ def test_old_vs_new_factory(model_fn, dev):
     if module_name == "detection":
         x = [x]
 
+    if module_name == "optical_flow":
+        args = [x, x]  # RAFT model requires img1, img2 as input
+    else:
+        args = [x]
+
     # compare with new model builder parameterized in the old fashion way
     try:
         model_old = _build_model(_get_original_model(model_fn), **kwargs).to(device=dev)
         model_new = _build_model(model_fn, **kwargs).to(device=dev)
     except ModuleNotFoundError:
         pytest.skip(f"Model '{model_name}' not available in both modules.")
-    torch.testing.assert_close(model_new(x), model_old(x), rtol=0.0, atol=0.0, check_dtype=False)
+    torch.testing.assert_close(model_new(*args), model_old(*args), rtol=0.0, atol=0.0, check_dtype=False)
 
 
 def test_smoke():

torchvision/models/optical_flow/raft.py

@@ -8,6 +8,7 @@
 from torch.nn.modules.instancenorm import InstanceNorm2d
 from torchvision.ops import ConvNormActivation
 
+from ..._internally_replaced_utils import load_state_dict_from_url
 from ...utils import _log_api_usage_once
 from ._utils import grid_sample, make_coords_grid, upsample_flow
 
@@ -19,6 +20,9 @@
 )
 
 
+_MODELS_URLS = {"raft_large": "https://download.pytorch.org/models/raft_large_C_T_V2-1bb1363a.pth"}
+
+
 class ResidualBlock(nn.Module):
     """Slightly modified Residual block with extra relu and biases."""
 
@@ -474,8 +478,8 @@ def forward(self, image1, image2, num_flow_updates: int = 12):
         hidden_state = torch.tanh(hidden_state)
         context = F.relu(context)
 
-        coords0 = make_coords_grid(batch_size, h // 8, w // 8).cuda()
-        coords1 = make_coords_grid(batch_size, h // 8, w // 8).cuda()
+        coords0 = make_coords_grid(batch_size, h // 8, w // 8).to(fmap1.device)
+        coords1 = make_coords_grid(batch_size, h // 8, w // 8).to(fmap1.device)
 
         flow_predictions = []
         for _ in range(num_flow_updates):
@@ -496,6 +500,9 @@ def forward(self, image1, image2, num_flow_updates: int = 12):
 
 def _raft(
     *,
+    arch=None,
+    pretrained=False,
+    progress=False,
     # Feature encoder
     feature_encoder_layers,
     feature_encoder_block,
@@ -560,14 +567,19 @@ def _raft(
             multiplier=0.25,  # See comment in MaskPredictor about this
         )
 
-    return RAFT(
+    model = RAFT(
         feature_encoder=feature_encoder,
         context_encoder=context_encoder,
         corr_block=corr_block,
         update_block=update_block,
         mask_predictor=mask_predictor,
         **kwargs,  # not really needed, all params should be consumed by now
     )
+    if pretrained:
+        state_dict = load_state_dict_from_url(_MODELS_URLS[arch], progress=progress)
+        model.load_state_dict(state_dict)
+
+    return model
 
 
 def raft_large(*, pretrained=False, progress=True, **kwargs):
@@ -584,10 +596,10 @@ def raft_large(*, pretrained=False, progress=True, **kwargs):
         nn.Module: The model.
     """
 
-    if pretrained:
-        raise ValueError("No checkpoint is available for raft_large")
-
     return _raft(
+        arch="raft_large",
+        pretrained=pretrained,
+        progress=progress,
         # Feature encoder
         feature_encoder_layers=(64, 64, 96, 128, 256),
         feature_encoder_block=ResidualBlock,
@@ -629,11 +641,13 @@ def raft_small(*, pretrained=False, progress=True, **kwargs):
         nn.Module: The model.
 
     """
-
     if pretrained:
         raise ValueError("No checkpoint is available for raft_small")
 
     return _raft(
+        arch="raft_small",
+        pretrained=pretrained,
+        progress=progress,
         # Feature encoder
         feature_encoder_layers=(32, 32, 64, 96, 128),
         feature_encoder_block=BottleneckBlock,

torchvision/prototype/models/optical_flow/raft.py

@@ -4,12 +4,11 @@
 from torch.nn.modules.instancenorm import InstanceNorm2d
 from torchvision.models.optical_flow import RAFT
 from torchvision.models.optical_flow.raft import _raft, BottleneckBlock, ResidualBlock
-
-# from torchvision.prototype.transforms import RaftEval
+from torchvision.prototype.transforms import RaftEval
+from torchvision.transforms.functional import InterpolationMode
 
 from .._api import WeightsEnum
-
-# from .._api import Weights
+from .._api import Weights
 from .._utils import handle_legacy_interface
 
 
@@ -22,17 +21,33 @@
 )
 
 
+_COMMON_META = {"interpolation": InterpolationMode.BILINEAR}
+
+
 class Raft_Large_Weights(WeightsEnum):
-    pass
-    # C_T_V1 = Weights(
-    #     # Chairs + Things
-    #     url="",
-    #     transforms=RaftEval,
-    #     meta={
-    #         "recipe": "",
-    #         "epe": -1234,
-    #     },
-    # )
+    C_T_V1 = Weights(
+        # Chairs + Things, ported from original paper repo (raft-things.pth)
+        url="https://download.pytorch.org/models/raft_large_C_T_V1-22a6c225.pth",
+        transforms=RaftEval,
+        meta={
+            **_COMMON_META,
+            "recipe": "https://github.com/princeton-vl/RAFT",
+            "sintel_train_cleanpass_epe": 1.4411,
+            "sintel_train_finalpass_epe": 2.7894,
+        },
+    )
+
+    C_T_V2 = Weights(
+        # Chairs + Things
+        url="https://download.pytorch.org/models/raft_large_C_T_V2-1bb1363a.pth",
+        transforms=RaftEval,
+        meta={
+            **_COMMON_META,
+            "recipe": "https://github.com/pytorch/vision/tree/main/references/optical_flow",
+            "sintel_train_cleanpass_epe": 1.3822,
+            "sintel_train_finalpass_epe": 2.7161,
+        },
+    )
 
     # C_T_SKHT_V1 = Weights(
     #     # Chairs + Things + Sintel fine-tuning, i.e.:
@@ -59,7 +74,7 @@ class Raft_Large_Weights(WeightsEnum):
     #     },
     # )
 
-    # default = C_T_V1
+    default = C_T_V2
 
 
 class Raft_Small_Weights(WeightsEnum):
@@ -75,13 +90,13 @@ class Raft_Small_Weights(WeightsEnum):
     # default = C_T_V1
 
 
-@handle_legacy_interface(weights=("pretrained", None))
+@handle_legacy_interface(weights=("pretrained", Raft_Large_Weights.C_T_V2))
 def raft_large(*, weights: Optional[Raft_Large_Weights] = None, progress=True, **kwargs):
     """RAFT model from
     `RAFT: Recurrent All Pairs Field Transforms for Optical Flow <https://arxiv.org/abs/2003.12039>`_.
 
     Args:
-        weights(Raft_Large_weights, optinal): TODO not implemented yet
+        weights(Raft_Large_weights, optional): pretrained weights to use.
         progress (bool): If True, displays a progress bar of the download to stderr
         kwargs (dict): Parameters that will be passed to the :class:`~torchvision.models.optical_flow.RAFT` class
             to override any default.
@@ -92,7 +107,7 @@ def raft_large(*, weights: Optional[Raft_Large_Weights] = None, progress=True, *
 
     weights = Raft_Large_Weights.verify(weights)
 
-    return _raft(
+    model = _raft(
         # Feature encoder
         feature_encoder_layers=(64, 64, 96, 128, 256),
         feature_encoder_block=ResidualBlock,
@@ -119,6 +134,11 @@ def raft_large(*, weights: Optional[Raft_Large_Weights] = None, progress=True, *
         **kwargs,
     )
 
+    if weights is not None:
+        model.load_state_dict(weights.get_state_dict(progress=progress))
+
+    return model
+
 
 @handle_legacy_interface(weights=("pretrained", None))
 def raft_small(*, weights: Optional[Raft_Small_Weights] = None, progress=True, **kwargs):
@@ -138,7 +158,7 @@ def raft_small(*, weights: Optional[Raft_Small_Weights] = None, progress=True, *
 
     weights = Raft_Small_Weights.verify(weights)
 
-    return _raft(
+    model = _raft(
         # Feature encoder
         feature_encoder_layers=(32, 32, 64, 96, 128),
         feature_encoder_block=BottleneckBlock,
@@ -164,3 +184,7 @@ def raft_small(*, weights: Optional[Raft_Small_Weights] = None, progress=True, *
         use_mask_predictor=False,
         **kwargs,
     )
+
+    if weights is not None:
+        model.load_state_dict(weights.get_state_dict(progress=progress))
+    return model