Add transforms and presets for optical flow models

references/optical_flow/presets.py

@@ -0,0 +1,64 @@
+import torch
+import transforms as T
+
+
+class OpticalFlowPresetEval(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+        self.transforms = T.Compose(
+            [
+                T.PILToTensor(),
+                T.ConvertImageDtype(torch.float32),
+                T.Normalize(mean=0.5, std=0.5),  # map [0, 1] into [-1, 1]
+                T.ValidateModelInput(),
+            ]
+        )
+
+    def __call__(self, img1, img2, flow, valid):
+        return self.transforms(img1, img2, flow, valid)
+
+
+class OpticalFlowPresetTrain(torch.nn.Module):
+    def __init__(
+        self,
+        # MaybeRandomResizeAndCrop params
+        crop_size,
+        min_scale=-0.2,
+        max_scale=0.5,
+        stretch_prob=0.8,
+        # AsymmetricColorJitter params
+        brightness=0.4,
+        contrast=0.4,
+        saturation=0.4,
+        hue=0.5 / 3.14,
+        # Random[H,V]Flip params
+        asymmetric_jitter_prob=0.2,
+        do_flip=True,
+    ):
+        super().__init__()
+
+        transforms = [
+            T.PILToTensor(),
+            T.AsymmetricColorJitter(
+                brightness=brightness, contrast=contrast, saturation=saturation, hue=hue, p=asymmetric_jitter_prob
+            ),
+            T.RandomApply([T.RandomErase()], p=0.5),
+            T.MaybeResizeAndCrop(
+                crop_size=crop_size, min_scale=min_scale, max_scale=max_scale, stretch_prob=stretch_prob
+            ),
+        ]
+
+        if do_flip:
+            transforms += [T.RandomHorizontalFlip(p=0.5), T.RandomVerticalFlip(p=0.1)]
+
+        transforms += [
+            T.ConvertImageDtype(torch.float32),
+            T.Normalize(mean=0.5, std=0.5),  # map [0, 1] into [-1, 1]
+            T.MakeValidFlowMask(),
+            T.ValidateModelInput(),
+        ]
+        self.transforms = T.Compose(transforms)
+
+    def __call__(self, img1, img2, flow, valid):
+        return self.transforms(img1, img2, flow, valid)

references/optical_flow/transforms.py

@@ -0,0 +1,267 @@
+import torch
+import torchvision.transforms as T
+import torchvision.transforms.functional as F
+
+
+class ValidateModelInput(torch.nn.Module):
+    # Pass-through transform that checks the shape and dtypes to make sure the model gets what it expects
+    def __init__(self):
+        super().__init__()
+
+    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)
+
+        assert img1.shape == img2.shape
+        h, w = img1.shape[-2:]
+        if flow is not None:
+            assert flow.shape == (2, h, w)
+        if valid_flow_mask is not None:
+            assert valid_flow_mask.shape == (h, w)
+            assert valid_flow_mask.dtype == torch.bool
+
+        return img1, img2, flow, valid_flow_mask
+
+
+class MakeValidFlowMask(torch.nn.Module):
+    # This transform generates a valid_flow_mask if it doesn't exist.
+    # The flow is considered valid if ||flow||_inf < threshold
+    # This is a noop for Kitti and HD1K which already come with a built-in flow mask.
+    def __init__(self, threshold=1000):
+        super().__init__()
+        self.threshold = threshold
+
+    def forward(self, img1, img2, flow, valid_flow_mask):
+        if flow is not None and valid_flow_mask is None:
+            valid_flow_mask = (flow.abs() < self.threshold).all(axis=0)
+        return img1, img2, flow, valid_flow_mask
+
+
+class ConvertImageDtype(torch.nn.Module):
+    def __init__(self, dtype):
+        super().__init__()
+        self.dtype = dtype
+
+    def forward(self, img1, img2, flow, valid_flow_mask):
+        img1 = F.convert_image_dtype(img1, dtype=self.dtype)
+        img2 = F.convert_image_dtype(img2, dtype=self.dtype)
+
+        img1 = img1.contiguous()
+        img2 = img2.contiguous()
+
+        return img1, img2, flow, valid_flow_mask
+
+
+class Normalize(torch.nn.Module):
+    def __init__(self, mean, std):
+        super().__init__()
+        self.mean = mean
+        self.std = std
+
+    def forward(self, img1, img2, flow, valid_flow_mask):
+        img1 = F.normalize(img1, mean=self.mean, std=self.std)
+        img2 = F.normalize(img2, mean=self.mean, std=self.std)
+
+        return img1, img2, flow, valid_flow_mask
+
+
+class PILToTensor(torch.nn.Module):
+    # Converts all inputs to tensors
+    # Technically the flow and the valid mask are numpy arrays, not PIL images, but we keep that naming
+    # for consistency with the rest, e.g. the segmentation reference.
+    def forward(self, img1, img2, flow, valid_flow_mask):
+        img1 = F.pil_to_tensor(img1)
+        img2 = F.pil_to_tensor(img2)
+        if flow is not None:
+            flow = torch.from_numpy(flow)
+        if valid_flow_mask is not None:
+            valid_flow_mask = torch.from_numpy(valid_flow_mask)
+
+        return img1, img2, flow, valid_flow_mask
+
+
+class AsymmetricColorJitter(T.ColorJitter):
+    def __init__(self, brightness=0, contrast=0, saturation=0, hue=0, p=0.2):
+        super().__init__(brightness=brightness, contrast=contrast, saturation=saturation, hue=hue)
+        self.p = p
+
+    def forward(self, img1, img2, flow, valid_flow_mask):
+
+        if torch.rand(1) < self.p:
+            # asymmetric: different transform for img1 and img2
+            img1 = super().forward(img1)
+            img2 = super().forward(img2)
+        else:
+            # symmetric: same transform for img1 and img2
+            batch = torch.stack([img1, img2])
+            batch = super().forward(batch)
+            img1, img2 = batch[0], batch[1]
+
+        return img1, img2, flow, valid_flow_mask
+
+
+class RandomErase(torch.nn.Module):
+    def forward(self, img1, img2, flow, valid_flow_mask):
+        bounds = [50, 100]
+        ht, wd = img2.shape[:2]
+
+        # Warning : This won't work with image values in [0, 1] because of round()
+        mean_color = img2.view(3, -1).float().mean(axis=-1).round()
+        for _ in range(torch.randint(1, 3, size=(1,)).item()):
+            x0 = torch.randint(0, wd, size=(1,)).item()
+            y0 = torch.randint(0, ht, size=(1,)).item()
+            dx, dy = torch.randint(bounds[0], bounds[1], size=(2,))
+            img2[:, y0 : y0 + dy, x0 : x0 + dx] = mean_color[:, None, None]
+
+        return img1, img2, flow, valid_flow_mask
+
+
+class RandomHorizontalFlip(T.RandomHorizontalFlip):
+    def forward(self, img1, img2, flow, valid_flow_mask):
+        if torch.rand(1) > self.p:
+            return img1, img2, flow, valid_flow_mask
+
+        img1 = F.hflip(img1)
+        img2 = F.hflip(img2)
+        flow = F.hflip(flow) * torch.tensor([-1, 1])[:, None, None]
+        if valid_flow_mask is not None:
+            valid_flow_mask = F.hflip(valid_flow_mask)
+        return img1, img2, flow, valid_flow_mask
+
+
+class RandomVerticalFlip(T.RandomVerticalFlip):
+    def forward(self, img1, img2, flow, valid_flow_mask):
+        if torch.rand(1) > self.p:
+            return img1, img2, flow, valid_flow_mask
+
+        img1 = F.vflip(img1)
+        img2 = F.vflip(img2)
+        flow = F.vflip(flow) * torch.tensor([1, -1])[:, None, None]
+        if valid_flow_mask is not None:
+            valid_flow_mask = F.vflip(valid_flow_mask)
+        return img1, img2, flow, valid_flow_mask
+
+
+class MaybeResizeAndCrop(torch.nn.Module):
+    # This transform will resize the input with a given proba, and then crop it.
+    # These are the reversed operations of the built-in RandomResizedCrop,
+    # although the order of the operations doesn't matter too much.
+    # The reason we don't rely on RandomResizedCrop is because of a significant
+    # difference in the parametrization of both transforms.
+    #
+    # There *is* a mapping between the inputs of MaybeResizeAndCrop and those of
+    # RandomResizedCrop, but the issue is that the parameters are sampled at
+    # random, with different distributions. Plotting (the equivalent of) `scale`
+    # and `ratio` from MaybeResizeAndCrop shows that the distributions of these
+    # parameters are very different from what can be obtained from the
+    # parametrization of RandomResizedCrop. I tried training RAFT by using
+    # RandomResizedCrop and tweaking the parameters a bit, but couldn't get
+    # an epe as good as with MaybeResizeAndCrop.
+    def __init__(self, crop_size, min_scale=-0.2, max_scale=0.5, stretch_prob=0.8):
+        super().__init__()
+        self.crop_size = crop_size
+        self.min_scale = min_scale
+        self.max_scale = max_scale
+        self.stretch_prob = stretch_prob
+        self.resize_prob = 0.8
+        self.max_stretch = 0.2
+
+    def forward(self, img1, img2, flow, valid_flow_mask):
+        # randomly sample scale
+        h, w = img1.shape[-2:]
+        # Note: in original code, they use + 1 instead of + 8 for sparse datasets (e.g. Kitti)
+        # It shouldn't matter much
+        min_scale = max((self.crop_size[0] + 8) / h, (self.crop_size[1] + 8) / w)
+
+        scale = 2 ** torch.FloatTensor(1).uniform_(self.min_scale, self.max_scale).item()
+        scale_x = scale
+        scale_y = scale
+        if torch.rand(1) < self.stretch_prob:
+            scale_x *= 2 ** torch.FloatTensor(1).uniform_(-self.max_stretch, self.max_stretch).item()
+            scale_y *= 2 ** torch.FloatTensor(1).uniform_(-self.max_stretch, self.max_stretch).item()
+
+        scale_x = max(scale_x, min_scale)
+        scale_y = max(scale_y, min_scale)
+
+        new_h, new_w = round(h * scale_y), round(w * scale_x)
+
+        if torch.rand(1).item() < self.resize_prob:
+            # rescale the images
+            img1 = F.resize(img1, size=(new_h, new_w))
+            img2 = F.resize(img2, size=(new_h, new_w))
+            if valid_flow_mask is None:
+                flow = F.resize(flow, size=(new_h, new_w))
+                flow = flow * torch.tensor([scale_x, scale_y])[:, None, None]
+            else:
+                flow, valid_flow_mask = self._resize_sparse_flow(
+                    flow, valid_flow_mask, scale_x=scale_x, scale_y=scale_y
+                )
+
+        # Note: For sparse datasets (Kitti), the original code uses a "margin"
+        # See e.g. https://github.com/princeton-vl/RAFT/blob/master/core/utils/augmentor.py#L220:L220
+        # We don't, not sure it matters much
+        y0 = torch.randint(0, img1.shape[1] - self.crop_size[0], size=(1,)).item()
+        x0 = torch.randint(0, img1.shape[2] - self.crop_size[1], size=(1,)).item()
+
+        img1 = img1[:, y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]
+        img2 = img2[:, y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]
+        flow = flow[:, y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]
+        if valid_flow_mask is not None:
+            valid_flow_mask = valid_flow_mask[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]
+
+        return img1, img2, flow, valid_flow_mask
+
+    def _resize_sparse_flow(self, flow, valid_flow_mask, scale_x=1.0, scale_y=1.0):
+        # This resizes both the flow and the valid_flow_mask mask (which is assumed to be reasonably sparse)
+        # There are as-many non-zero values in the original flow as in the resized flow (up to OOB)
+        # So for example if scale_x = scale_y = 2, the sparsity of the output flow is multiplied by 4
+
+        h, w = flow.shape[-2:]
+
+        h_new = int(round(h * scale_y))
+        w_new = int(round(w * scale_x))
+        flow_new = torch.zeros(size=[2, h_new, w_new], dtype=flow.dtype)
+        valid_new = torch.zeros(size=[h_new, w_new], dtype=valid_flow_mask.dtype)
+
+        jj, ii = torch.meshgrid(torch.arange(w), torch.arange(h), indexing="xy")
+
+        ii_valid, jj_valid = ii[valid_flow_mask], jj[valid_flow_mask]
+
+        ii_valid_new = torch.round(ii_valid.to(float) * scale_y).to(torch.long)
+        jj_valid_new = torch.round(jj_valid.to(float) * scale_x).to(torch.long)
+
+        within_bounds_mask = (0 <= ii_valid_new) & (ii_valid_new < h_new) & (0 <= jj_valid_new) & (jj_valid_new < w_new)
+
+        ii_valid = ii_valid[within_bounds_mask]
+        jj_valid = jj_valid[within_bounds_mask]
+        ii_valid_new = ii_valid_new[within_bounds_mask]
+        jj_valid_new = jj_valid_new[within_bounds_mask]
+
+        valid_flow_new = flow[:, ii_valid, jj_valid]
+        valid_flow_new[0] *= scale_x
+        valid_flow_new[1] *= scale_y
+
+        flow_new[:, ii_valid_new, jj_valid_new] = valid_flow_new
+        valid_new[ii_valid_new, jj_valid_new] = 1
+
+        return flow_new, valid_new
+
+
+class RandomApply(T.RandomApply):
+    def forward(self, img1, img2, flow, valid_flow_mask):
+        if self.p < torch.rand(1):
+            return img1, img2, flow, valid_flow_mask
+        for t in self.transforms:
+            img1, img2, flow, valid_flow_mask = t(img1, img2, flow, valid_flow_mask)
+        return img1, img2, flow, valid_flow_mask
+
+
+class Compose:
+    def __init__(self, transforms):
+        self.transforms = transforms
+
+    def __call__(self, img1, img2, flow, valid_flow_mask):
+        for t in self.transforms:
+            img1, img2, flow, valid_flow_mask = t(img1, img2, flow, valid_flow_mask)
+        return img1, img2, flow, valid_flow_mask