[fbsync] Add SimpleCopyPaste augmentation (pytorch#5825)

Summary:
* added simple POC

* added jitter and crop options

* added references

* moved simplecopypaste to detection module

* working POC for simple copy paste in detection

* added comments

* remove transforms from class
updated the labels
added gaussian blur

* removed loop for mask calculation

* replaced Gaussian blur with functional api

* added inplace operations

* added changes to accept tuples instead of tensors

* - make copy paste functional
- make only one copy of batch and target

* add inplace support within copy paste functional

* Updated code for copy-paste transform

* Fixed code formatting

* [skip ci] removed manual thresholding

* Replaced cropping by resizing data to paste

* Removed inplace arg (as useless) and put a check on iscrowd target

* code-formatting

* Updated copypaste op to make it torch scriptable
Added fallbacks to support LSJ

* Fixed flake8

* Updates according to the review

Differential Revision: D37212651

fbshipit-source-id: 467b670164150dd5cc424f4d616d436295ce818d

Co-authored-by: vfdev-5 <vfdev.5@gmail.com>
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

references/detection/train.py

@@ -31,6 +31,8 @@
 from coco_utils import get_coco, get_coco_kp
 from engine import train_one_epoch, evaluate
 from group_by_aspect_ratio import GroupedBatchSampler, create_aspect_ratio_groups
+from torchvision.transforms import InterpolationMode
+from transforms import SimpleCopyPaste
 
 
 def get_dataset(name, image_set, transform, data_path):
@@ -145,6 +147,13 @@ def get_args_parser(add_help=True):
     # Mixed precision training parameters
     parser.add_argument("--amp", action="store_true", help="Use torch.cuda.amp for mixed precision training")
 
+    # Use CopyPaste augmentation training parameter
+    parser.add_argument(
+        "--use-copypaste",
+        action="store_true",
+        help="Use CopyPaste data augmentation. Works only with data-augmentation='lsj'.",
+    )
+
     return parser
 
 
@@ -180,8 +189,20 @@ def main(args):
     else:
         train_batch_sampler = torch.utils.data.BatchSampler(train_sampler, args.batch_size, drop_last=True)
 
+    train_collate_fn = utils.collate_fn
+    if args.use_copypaste:
+        if args.data_augmentation != "lsj":
+            raise RuntimeError("SimpleCopyPaste algorithm currently only supports the 'lsj' data augmentation policies")
+
+        copypaste = SimpleCopyPaste(resize_interpolation=InterpolationMode.BILINEAR, blending=True)
+
+        def copypaste_collate_fn(batch):
+            return copypaste(*utils.collate_fn(batch))
+
+        train_collate_fn = copypaste_collate_fn
+
     data_loader = torch.utils.data.DataLoader(
-        dataset, batch_sampler=train_batch_sampler, num_workers=args.workers, collate_fn=utils.collate_fn
+        dataset, batch_sampler=train_batch_sampler, num_workers=args.workers, collate_fn=train_collate_fn
     )
 
     data_loader_test = torch.utils.data.DataLoader(

references/detection/transforms.py

@@ -3,6 +3,7 @@
 import torch
 import torchvision
 from torch import nn, Tensor
+from torchvision import ops
 from torchvision.transforms import functional as F
 from torchvision.transforms import transforms as T, InterpolationMode
 
@@ -437,3 +438,157 @@ def forward(
                 )
 
         return image, target
+
+
+def _copy_paste(
+    image: torch.Tensor,
+    target: Dict[str, Tensor],
+    paste_image: torch.Tensor,
+    paste_target: Dict[str, Tensor],
+    blending: bool = True,
+    resize_interpolation: F.InterpolationMode = F.InterpolationMode.BILINEAR,
+) -> Tuple[torch.Tensor, Dict[str, Tensor]]:
+
+    # Random paste targets selection:
+    num_masks = len(paste_target["masks"])
+
+    if num_masks < 1:
+        # Such degerante case with num_masks=0 can happen with LSJ
+        # Let's just return (image, target)
+        return image, target
+
+    # We have to please torch script by explicitly specifying dtype as torch.long
+    random_selection = torch.randint(0, num_masks, (num_masks,), device=paste_image.device)
+    random_selection = torch.unique(random_selection).to(torch.long)
+
+    paste_masks = paste_target["masks"][random_selection]
+    paste_boxes = paste_target["boxes"][random_selection]
+    paste_labels = paste_target["labels"][random_selection]
+
+    masks = target["masks"]
+
+    # We resize source and paste data if they have different sizes
+    # This is something we introduced here as originally the algorithm works
+    # on equal-sized data (for example, coming from LSJ data augmentations)
+    size1 = image.shape[-2:]
+    size2 = paste_image.shape[-2:]
+    if size1 != size2:
+        paste_image = F.resize(paste_image, size1, interpolation=resize_interpolation)
+        paste_masks = F.resize(paste_masks, size1, interpolation=F.InterpolationMode.NEAREST)
+        # resize bboxes:
+        ratios = torch.tensor((size1[1] / size2[1], size1[0] / size2[0]), device=paste_boxes.device)
+        paste_boxes = paste_boxes.view(-1, 2, 2).mul(ratios).view(paste_boxes.shape)
+
+    paste_alpha_mask = paste_masks.sum(dim=0) > 0
+
+    if blending:
+        paste_alpha_mask = F.gaussian_blur(
+            paste_alpha_mask.unsqueeze(0),
+            kernel_size=(5, 5),
+            sigma=[
+                2.0,
+            ],
+        )
+
+    # Copy-paste images:
+    image = (image * (~paste_alpha_mask)) + (paste_image * paste_alpha_mask)
+
+    # Copy-paste masks:
+    masks = masks * (~paste_alpha_mask)
+    non_all_zero_masks = masks.sum((-1, -2)) > 0
+    masks = masks[non_all_zero_masks]
+
+    # Do a shallow copy of the target dict
+    out_target = {k: v for k, v in target.items()}
+
+    out_target["masks"] = torch.cat([masks, paste_masks])
+
+    # Copy-paste boxes and labels
+    boxes = ops.masks_to_boxes(masks)
+    out_target["boxes"] = torch.cat([boxes, paste_boxes])
+
+    labels = target["labels"][non_all_zero_masks]
+    out_target["labels"] = torch.cat([labels, paste_labels])
+
+    # Update additional optional keys: area and iscrowd if exist
+    if "area" in target:
+        out_target["area"] = out_target["masks"].sum((-1, -2)).to(torch.float32)
+
+    if "iscrowd" in target and "iscrowd" in paste_target:
+        # target['iscrowd'] size can be differ from mask size (non_all_zero_masks)
+        # For example, if previous transforms geometrically modifies masks/boxes/labels but
+        # does not update "iscrowd"
+        if len(target["iscrowd"]) == len(non_all_zero_masks):
+            iscrowd = target["iscrowd"][non_all_zero_masks]
+            paste_iscrowd = paste_target["iscrowd"][random_selection]
+            out_target["iscrowd"] = torch.cat([iscrowd, paste_iscrowd])
+
+    # Check for degenerated boxes and remove them
+    boxes = out_target["boxes"]
+    degenerate_boxes = boxes[:, 2:] <= boxes[:, :2]
+    if degenerate_boxes.any():
+        valid_targets = ~degenerate_boxes.any(dim=1)
+
+        out_target["boxes"] = boxes[valid_targets]
+        out_target["masks"] = out_target["masks"][valid_targets]
+        out_target["labels"] = out_target["labels"][valid_targets]
+
+        if "area" in out_target:
+            out_target["area"] = out_target["area"][valid_targets]
+        if "iscrowd" in out_target and len(out_target["iscrowd"]) == len(valid_targets):
+            out_target["iscrowd"] = out_target["iscrowd"][valid_targets]
+
+    return image, out_target
+
+
+class SimpleCopyPaste(torch.nn.Module):
+    def __init__(self, blending=True, resize_interpolation=F.InterpolationMode.BILINEAR):
+        super().__init__()
+        self.resize_interpolation = resize_interpolation
+        self.blending = blending
+
+    def forward(
+        self, images: List[torch.Tensor], targets: List[Dict[str, Tensor]]
+    ) -> Tuple[List[torch.Tensor], List[Dict[str, Tensor]]]:
+        torch._assert(
+            isinstance(images, (list, tuple)) and all([isinstance(v, torch.Tensor) for v in images]),
+            "images should be a list of tensors",
+        )
+        torch._assert(
+            isinstance(targets, (list, tuple)) and len(images) == len(targets),
+            "targets should be a list of the same size as images",
+        )
+        for target in targets:
+            # Can not check for instance type dict with inside torch.jit.script
+            # torch._assert(isinstance(target, dict), "targets item should be a dict")
+            for k in ["masks", "boxes", "labels"]:
+                torch._assert(k in target, f"Key {k} should be present in targets")
+                torch._assert(isinstance(target[k], torch.Tensor), f"Value for the key {k} should be a tensor")
+
+        # images = [t1, t2, ..., tN]
+        # Let's define paste_images as shifted list of input images
+        # paste_images = [t2, t3, ..., tN, t1]
+        # FYI: in TF they mix data on the dataset level
+        images_rolled = images[-1:] + images[:-1]
+        targets_rolled = targets[-1:] + targets[:-1]
+
+        output_images: List[torch.Tensor] = []
+        output_targets: List[Dict[str, Tensor]] = []
+
+        for image, target, paste_image, paste_target in zip(images, targets, images_rolled, targets_rolled):
+            output_image, output_data = _copy_paste(
+                image,
+                target,
+                paste_image,
+                paste_target,
+                blending=self.blending,
+                resize_interpolation=self.resize_interpolation,
+            )
+            output_images.append(output_image)
+            output_targets.append(output_data)
+
+        return output_images, output_targets
+
+    def __repr__(self) -> str:
+        s = f"{self.__class__.__name__}(blending={self.blending}, resize_interpolation={self.resize_interpolation})"
+        return s