Upload model (#1)

- Upload model (edba1de555b3ac3fbc0ad8165f8fa3d84fbee170)

attention.py

@@ -0,0 +1,295 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .utils import merge_splits, merge_splits_1d, split_feature, split_feature_1d
+
+
+def single_head_full_attention(q, k, v):
+    # q, k, v: [B, L, C]
+    assert q.dim() == k.dim() == v.dim() == 3
+
+    scores = torch.matmul(q, k.permute(0, 2, 1)) / (q.size(2) ** 0.5)  # [B, L, L]
+    attn = torch.softmax(scores, dim=2)  # [B, L, L]
+    out = torch.matmul(attn, v)  # [B, L, C]
+
+    return out
+
+
+def single_head_full_attention_1d(
+    q,
+    k,
+    v,
+    h=None,
+    w=None,
+):
+    # q, k, v: [B, L, C]
+
+    assert h is not None and w is not None
+    assert q.size(1) == h * w
+
+    b, _, c = q.size()
+
+    q = q.view(b, h, w, c)  # [B, H, W, C]
+    k = k.view(b, h, w, c)
+    v = v.view(b, h, w, c)
+
+    scale_factor = c**0.5
+
+    scores = torch.matmul(q, k.permute(0, 1, 3, 2)) / scale_factor  # [B, H, W, W]
+
+    attn = torch.softmax(scores, dim=-1)
+
+    out = torch.matmul(attn, v).view(b, -1, c)  # [B, H*W, C]
+
+    return out
+
+
+def single_head_split_window_attention(
+    q,
+    k,
+    v,
+    num_splits=1,
+    with_shift=False,
+    h=None,
+    w=None,
+    attn_mask=None,
+):
+    # ref: https://github.com/microsoft/Swin-Transformer/blob/main/models/swin_transformer.py
+    # q, k, v: [B, L, C]
+    assert q.dim() == k.dim() == v.dim() == 3
+
+    assert h is not None and w is not None
+    assert q.size(1) == h * w
+
+    b, _, c = q.size()
+
+    b_new = b * num_splits * num_splits
+
+    window_size_h = h // num_splits
+    window_size_w = w // num_splits
+
+    q = q.view(b, h, w, c)  # [B, H, W, C]
+    k = k.view(b, h, w, c)
+    v = v.view(b, h, w, c)
+
+    scale_factor = c**0.5
+
+    if with_shift:
+        assert attn_mask is not None  # compute once
+        shift_size_h = window_size_h // 2
+        shift_size_w = window_size_w // 2
+
+        q = torch.roll(q, shifts=(-shift_size_h, -shift_size_w), dims=(1, 2))
+        k = torch.roll(k, shifts=(-shift_size_h, -shift_size_w), dims=(1, 2))
+        v = torch.roll(v, shifts=(-shift_size_h, -shift_size_w), dims=(1, 2))
+
+    q = split_feature(q, num_splits=num_splits, channel_last=True)  # [B*K*K, H/K, W/K, C]
+    k = split_feature(k, num_splits=num_splits, channel_last=True)
+    v = split_feature(v, num_splits=num_splits, channel_last=True)
+
+    scores = (
+        torch.matmul(q.view(b_new, -1, c), k.view(b_new, -1, c).permute(0, 2, 1)) / scale_factor
+    )  # [B*K*K, H/K*W/K, H/K*W/K]
+
+    if with_shift:
+        scores += attn_mask.repeat(b, 1, 1)
+
+    attn = torch.softmax(scores, dim=-1)
+
+    out = torch.matmul(attn, v.view(b_new, -1, c))  # [B*K*K, H/K*W/K, C]
+
+    out = merge_splits(
+        out.view(b_new, h // num_splits, w // num_splits, c),
+        num_splits=num_splits,
+        channel_last=True,
+    )  # [B, H, W, C]
+
+    # shift back
+    if with_shift:
+        out = torch.roll(out, shifts=(shift_size_h, shift_size_w), dims=(1, 2))
+
+    out = out.view(b, -1, c)
+
+    return out
+
+
+def single_head_split_window_attention_1d(
+    q,
+    k,
+    v,
+    relative_position_bias=None,
+    num_splits=1,
+    with_shift=False,
+    h=None,
+    w=None,
+    attn_mask=None,
+):
+    # q, k, v: [B, L, C]
+
+    assert h is not None and w is not None
+    assert q.size(1) == h * w
+
+    b, _, c = q.size()
+
+    b_new = b * num_splits * h
+
+    window_size_w = w // num_splits
+
+    q = q.view(b * h, w, c)  # [B*H, W, C]
+    k = k.view(b * h, w, c)
+    v = v.view(b * h, w, c)
+
+    scale_factor = c**0.5
+
+    if with_shift:
+        assert attn_mask is not None  # compute once
+        shift_size_w = window_size_w // 2
+
+        q = torch.roll(q, shifts=-shift_size_w, dims=1)
+        k = torch.roll(k, shifts=-shift_size_w, dims=1)
+        v = torch.roll(v, shifts=-shift_size_w, dims=1)
+
+    q = split_feature_1d(q, num_splits=num_splits)  # [B*H*K, W/K, C]
+    k = split_feature_1d(k, num_splits=num_splits)
+    v = split_feature_1d(v, num_splits=num_splits)
+
+    scores = (
+        torch.matmul(q.view(b_new, -1, c), k.view(b_new, -1, c).permute(0, 2, 1)) / scale_factor
+    )  # [B*H*K, W/K, W/K]
+
+    if with_shift:
+        # attn_mask: [K, W/K, W/K]
+        scores += attn_mask.repeat(b * h, 1, 1)  # [B*H*K, W/K, W/K]
+
+    attn = torch.softmax(scores, dim=-1)
+
+    out = torch.matmul(attn, v.view(b_new, -1, c))  # [B*H*K, W/K, C]
+
+    out = merge_splits_1d(out, h, num_splits=num_splits)  # [B, H, W, C]
+
+    # shift back
+    if with_shift:
+        out = torch.roll(out, shifts=shift_size_w, dims=2)
+
+    out = out.view(b, -1, c)
+
+    return out
+
+
+class SelfAttnPropagation(nn.Module):
+    """
+    flow propagation with self-attention on feature
+    query: feature0, key: feature0, value: flow
+    """
+
+    def __init__(
+        self,
+        in_channels,
+        **kwargs,
+    ):
+        super().__init__()
+
+        self.q_proj = nn.Linear(in_channels, in_channels)
+        self.k_proj = nn.Linear(in_channels, in_channels)
+
+        for p in self.parameters():
+            if p.dim() > 1:
+                nn.init.xavier_uniform_(p)
+
+    def forward(
+        self,
+        feature0,
+        flow,
+        local_window_attn=False,
+        local_window_radius=1,
+        **kwargs,
+    ):
+        # q, k: feature [B, C, H, W], v: flow [B, 2, H, W]
+        if local_window_attn:
+            return self.forward_local_window_attn(
+                feature0, flow, local_window_radius=local_window_radius
+            )
+
+        b, c, h, w = feature0.size()
+
+        query = feature0.view(b, c, h * w).permute(0, 2, 1)  # [B, H*W, C]
+
+        # a note: the ``correct'' implementation should be:
+        # ``query = self.q_proj(query), key = self.k_proj(query)''
+        # this problem is observed while cleaning up the code
+        # however, this doesn't affect the performance since the projection is a linear operation,
+        # thus the two projection matrices for key can be merged
+        # so I just leave it as is in order to not re-train all models :)
+        query = self.q_proj(query)  # [B, H*W, C]
+        key = self.k_proj(query)  # [B, H*W, C]
+
+        value = flow.view(b, flow.size(1), h * w).permute(0, 2, 1)  # [B, H*W, 2]
+
+        scores = torch.matmul(query, key.permute(0, 2, 1)) / (c**0.5)  # [B, H*W, H*W]
+        prob = torch.softmax(scores, dim=-1)
+
+        out = torch.matmul(prob, value)  # [B, H*W, 2]
+        out = out.view(b, h, w, value.size(-1)).permute(0, 3, 1, 2)  # [B, 2, H, W]
+
+        return out
+
+    def forward_local_window_attn(
+        self,
+        feature0,
+        flow,
+        local_window_radius=1,
+    ):
+        assert flow.size(1) == 2 or flow.size(1) == 1  # flow or disparity or depth
+        assert local_window_radius > 0
+
+        b, c, h, w = feature0.size()
+
+        value_channel = flow.size(1)
+
+        feature0_reshape = self.q_proj(feature0.view(b, c, -1).permute(0, 2, 1)).reshape(
+            b * h * w, 1, c
+        )  # [B*H*W, 1, C]
+
+        kernel_size = 2 * local_window_radius + 1
+
+        feature0_proj = (
+            self.k_proj(feature0.view(b, c, -1).permute(0, 2, 1))
+            .permute(0, 2, 1)
+            .reshape(b, c, h, w)
+        )
+
+        feature0_window = F.unfold(
+            feature0_proj, kernel_size=kernel_size, padding=local_window_radius
+        )  # [B, C*(2R+1)^2), H*W]
+
+        feature0_window = (
+            feature0_window.view(b, c, kernel_size**2, h, w)
+            .permute(0, 3, 4, 1, 2)
+            .reshape(b * h * w, c, kernel_size**2)
+        )  # [B*H*W, C, (2R+1)^2]
+
+        flow_window = F.unfold(
+            flow, kernel_size=kernel_size, padding=local_window_radius
+        )  # [B, 2*(2R+1)^2), H*W]
+
+        flow_window = (
+            flow_window.view(b, value_channel, kernel_size**2, h, w)
+            .permute(0, 3, 4, 2, 1)
+            .reshape(b * h * w, kernel_size**2, value_channel)
+        )  # [B*H*W, (2R+1)^2, 2]
+
+        scores = torch.matmul(feature0_reshape, feature0_window) / (
+            c**0.5
+        )  # [B*H*W, 1, (2R+1)^2]
+
+        prob = torch.softmax(scores, dim=-1)
+
+        out = (
+            torch.matmul(prob, flow_window)
+            .view(b, h, w, value_channel)
+            .permute(0, 3, 1, 2)
+            .contiguous()
+        )  # [B, 2, H, W]
+
+        return out

backbone.py

@@ -0,0 +1,142 @@
+import torch.nn as nn
+
+from .trident_conv import MultiScaleTridentConv
+
+
+class ResidualBlock(nn.Module):
+    def __init__(
+        self,
+        in_planes,
+        planes,
+        norm_layer=nn.InstanceNorm2d,
+        stride=1,
+        dilation=1,
+    ):
+        super().__init__()
+
+        self.conv1 = nn.Conv2d(
+            in_planes,
+            planes,
+            kernel_size=3,
+            dilation=dilation,
+            padding=dilation,
+            stride=stride,
+            bias=False,
+        )
+        self.conv2 = nn.Conv2d(
+            planes, planes, kernel_size=3, dilation=dilation, padding=dilation, bias=False
+        )
+        self.relu = nn.ReLU(inplace=True)
+
+        self.norm1 = norm_layer(planes)
+        self.norm2 = norm_layer(planes)
+        if not stride == 1 or in_planes != planes:
+            self.norm3 = norm_layer(planes)
+
+        if stride == 1 and in_planes == planes:
+            self.downsample = None
+        else:
+            self.downsample = nn.Sequential(
+                nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride), self.norm3
+            )
+
+    def forward(self, x):
+        y = x
+        y = self.relu(self.norm1(self.conv1(y)))
+        y = self.relu(self.norm2(self.conv2(y)))
+
+        if self.downsample is not None:
+            x = self.downsample(x)
+
+        return self.relu(x + y)
+
+
+class CNNEncoder(nn.Module):
+    def __init__(
+        self,
+        output_dim=128,
+        norm_layer=nn.InstanceNorm2d,
+        num_output_scales=1,
+        **kwargs,
+    ):
+        super().__init__()
+        self.num_branch = num_output_scales
+
+        feature_dims = [64, 96, 128]
+
+        self.conv1 = nn.Conv2d(
+            3, feature_dims[0], kernel_size=7, stride=2, padding=3, bias=False
+        )  # 1/2
+        self.norm1 = norm_layer(feature_dims[0])
+        self.relu1 = nn.ReLU(inplace=True)
+
+        self.in_planes = feature_dims[0]
+        self.layer1 = self._make_layer(feature_dims[0], stride=1, norm_layer=norm_layer)  # 1/2
+        self.layer2 = self._make_layer(feature_dims[1], stride=2, norm_layer=norm_layer)  # 1/4
+
+        # highest resolution 1/4 or 1/8
+        stride = 2 if num_output_scales == 1 else 1
+        self.layer3 = self._make_layer(
+            feature_dims[2],
+            stride=stride,
+            norm_layer=norm_layer,
+        )  # 1/4 or 1/8
+
+        self.conv2 = nn.Conv2d(feature_dims[2], output_dim, 1, 1, 0)
+
+        if self.num_branch > 1:
+            if self.num_branch == 4:
+                strides = (1, 2, 4, 8)
+            elif self.num_branch == 3:
+                strides = (1, 2, 4)
+            elif self.num_branch == 2:
+                strides = (1, 2)
+            else:
+                raise ValueError
+
+            self.trident_conv = MultiScaleTridentConv(
+                output_dim,
+                output_dim,
+                kernel_size=3,
+                strides=strides,
+                paddings=1,
+                num_branch=self.num_branch,
+            )
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+            elif isinstance(m, (nn.BatchNorm2d, nn.InstanceNorm2d, nn.GroupNorm)):
+                if m.weight is not None:
+                    nn.init.constant_(m.weight, 1)
+                if m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+
+    def _make_layer(self, dim, stride=1, dilation=1, norm_layer=nn.InstanceNorm2d):
+        layer1 = ResidualBlock(
+            self.in_planes, dim, norm_layer=norm_layer, stride=stride, dilation=dilation
+        )
+        layer2 = ResidualBlock(dim, dim, norm_layer=norm_layer, stride=1, dilation=dilation)
+
+        layers = (layer1, layer2)
+
+        self.in_planes = dim
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.norm1(x)
+        x = self.relu1(x)
+
+        x = self.layer1(x)  # 1/2
+        x = self.layer2(x)  # 1/4
+        x = self.layer3(x)  # 1/8 or 1/4
+
+        x = self.conv2(x)
+
+        if self.num_branch > 1:
+            out = self.trident_conv([x] * self.num_branch)  # high to low res
+        else:
+            out = [x]
+
+        return out

config.json

@@ -0,0 +1,13 @@
+{
+  "architectures": [
+    "DoduoModel"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_doduo.DoduoConfig",
+    "AutoModel": "modeling_doduo.DoduoModel"
+  },
+  "dino_corr_mask_ratio": 0.99,
+  "model_type": "doduo",
+  "torch_dtype": "float32",
+  "transformers_version": "4.30.2"
+}

configuration_doduo.py

@@ -0,0 +1,13 @@
+from transformers import PretrainedConfig
+from typing import List
+
+class DoduoConfig(PretrainedConfig):
+    model_type = "doduo"
+
+    def __init__(
+        self,
+        dino_corr_mask_ratio: float = 0.99,
+        **kwargs,
+    ):
+        self.dino_corr_mask_ratio = dino_corr_mask_ratio
+        super().__init__(**kwargs)

dino.py

@@ -0,0 +1,420 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Mostly copy-paste from timm library.
+
+https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
+"""
+import math
+from functools import partial
+import warnings
+
+import torch
+import torch.nn as nn
+
+
+def drop_path(x, drop_prob: float = 0.0, training: bool = False):
+    if drop_prob == 0.0 or not training:
+        return x
+    keep_prob = 1 - drop_prob
+    shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+    random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
+    random_tensor.floor_()  # binarize
+    output = x.div(keep_prob) * random_tensor
+    return output
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks)."""
+
+    def __init__(self, drop_prob=None):
+        super().__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+
+
+class Mlp(nn.Module):
+    def __init__(
+        self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.0
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(
+        self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.0, proj_drop=0.0
+    ):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x):
+        B, N, C = x.shape
+        qkv = (
+            self.qkv(x)
+            .reshape(B, N, 3, self.num_heads, C // self.num_heads)
+            .permute(2, 0, 3, 1, 4)
+        )
+        q, k, v = qkv[0], qkv[1], qkv[2]
+
+        attn = (q @ k.transpose(-2, -1)) * self.scale
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x, attn
+
+
+class Block(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_heads,
+        mlp_ratio=4.0,
+        qkv_bias=False,
+        qk_scale=None,
+        drop=0.0,
+        attn_drop=0.0,
+        drop_path=0.0,
+        act_layer=nn.GELU,
+        norm_layer=nn.LayerNorm,
+    ):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            qk_scale=qk_scale,
+            attn_drop=attn_drop,
+            proj_drop=drop,
+        )
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(
+            in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop
+        )
+
+    def forward(self, x, return_attention=False):
+        y, attn = self.attn(self.norm1(x))
+        if return_attention:
+            return attn
+        x = x + self.drop_path(y)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+
+class PatchEmbed(nn.Module):
+    """Image to Patch Embedding."""
+
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
+        super().__init__()
+        num_patches = (img_size // patch_size) * (img_size // patch_size)
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.num_patches = num_patches
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+        x = self.proj(x).flatten(2).transpose(1, 2)
+        return x
+
+
+class VisionTransformer(nn.Module):
+    """Vision Transformer."""
+
+    def __init__(
+        self,
+        img_size=[224],
+        patch_size=16,
+        in_chans=3,
+        num_classes=0,
+        embed_dim=768,
+        depth=12,
+        num_heads=12,
+        mlp_ratio=4.0,
+        qkv_bias=False,
+        qk_scale=None,
+        drop_rate=0.0,
+        attn_drop_rate=0.0,
+        drop_path_rate=0.0,
+        norm_layer=nn.LayerNorm,
+        **kwargs
+    ):
+        super().__init__()
+        self.num_features = self.embed_dim = embed_dim
+
+        self.patch_embed = PatchEmbed(
+            img_size=img_size[0], patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim
+        )
+        num_patches = self.patch_embed.num_patches
+
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        dpr = [
+            x.item() for x in torch.linspace(0, drop_path_rate, depth)
+        ]  # stochastic depth decay rule
+        self.blocks = nn.ModuleList(
+            [
+                Block(
+                    dim=embed_dim,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    qk_scale=qk_scale,
+                    drop=drop_rate,
+                    attn_drop=attn_drop_rate,
+                    drop_path=dpr[i],
+                    norm_layer=norm_layer,
+                )
+                for i in range(depth)
+            ]
+        )
+        self.norm = norm_layer(embed_dim)
+
+        # Classifier head
+        self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+
+        trunc_normal_(self.pos_embed, std=0.02)
+        trunc_normal_(self.cls_token, std=0.02)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=0.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    def interpolate_pos_encoding(self, x, w, h):
+        npatch = x.shape[1] - 1
+        N = self.pos_embed.shape[1] - 1
+        if npatch == N and w == h:
+            return self.pos_embed
+        class_pos_embed = self.pos_embed[:, 0]
+        patch_pos_embed = self.pos_embed[:, 1:]
+        dim = x.shape[-1]
+        w0 = w // self.patch_embed.patch_size
+        h0 = h // self.patch_embed.patch_size
+        # we add a small number to avoid floating point error in the interpolation
+        # see discussion at https://github.com/facebookresearch/dino/issues/8
+        w0, h0 = w0 + 0.1, h0 + 0.1
+        patch_pos_embed = nn.functional.interpolate(
+            patch_pos_embed.reshape(1, int(math.sqrt(N)), int(math.sqrt(N)), dim).permute(
+                0, 3, 1, 2
+            ),
+            scale_factor=(w0 / math.sqrt(N), h0 / math.sqrt(N)),
+            mode="bicubic",
+        )
+        assert int(w0) == patch_pos_embed.shape[-2] and int(h0) == patch_pos_embed.shape[-1]
+        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
+        return torch.cat((class_pos_embed.unsqueeze(0), patch_pos_embed), dim=1)
+
+    def prepare_tokens(self, x):
+        B, nc, w, h = x.shape
+        x = self.patch_embed(x)  # patch linear embedding
+
+        # add the [CLS] token to the embed patch tokens
+        cls_tokens = self.cls_token.expand(B, -1, -1)
+        x = torch.cat((cls_tokens, x), dim=1)
+
+        # add positional encoding to each token
+        x = x + self.interpolate_pos_encoding(x, w, h)
+
+        return self.pos_drop(x)
+
+    def forward(self, x):
+        x = self.prepare_tokens(x)
+        for blk in self.blocks:
+            x = blk(x)
+        x = self.norm(x)
+        return x[:, 0]
+
+    def get_last_selfattention(self, x):
+        x = self.prepare_tokens(x)
+        for i, blk in enumerate(self.blocks):
+            if i < len(self.blocks) - 1:
+                x = blk(x)
+            else:
+                # return attention of the last block
+                return blk(x, return_attention=True)
+
+    def get_intermediate_layers(self, x, n=1):
+        x = self.prepare_tokens(x)
+        # we return the output tokens from the `n` last blocks
+        output = []
+        for i, blk in enumerate(self.blocks):
+            x = blk(x)
+            if len(self.blocks) - i <= n:
+                output.append(self.norm(x))
+        return output
+
+
+def vit_tiny(patch_size=16, **kwargs):
+    model = VisionTransformer(
+        patch_size=patch_size,
+        embed_dim=192,
+        depth=12,
+        num_heads=3,
+        mlp_ratio=4,
+        qkv_bias=True,
+        norm_layer=partial(nn.LayerNorm, eps=1e-6),
+        **kwargs
+    )
+    return model
+
+
+def vit_small(patch_size=16, **kwargs):
+    model = VisionTransformer(
+        patch_size=patch_size,
+        embed_dim=384,
+        depth=12,
+        num_heads=6,
+        mlp_ratio=4,
+        qkv_bias=True,
+        norm_layer=partial(nn.LayerNorm, eps=1e-6),
+        **kwargs
+    )
+    return model
+
+
+def vit_base(patch_size=16, **kwargs):
+    model = VisionTransformer(
+        patch_size=patch_size,
+        embed_dim=768,
+        depth=12,
+        num_heads=12,
+        mlp_ratio=4,
+        qkv_bias=True,
+        norm_layer=partial(nn.LayerNorm, eps=1e-6),
+        **kwargs
+    )
+    return model
+
+
+class DINOHead(nn.Module):
+    def __init__(
+        self,
+        in_dim,
+        out_dim,
+        use_bn=False,
+        norm_last_layer=True,
+        nlayers=3,
+        hidden_dim=2048,
+        bottleneck_dim=256,
+    ):
+        super().__init__()
+        nlayers = max(nlayers, 1)
+        if nlayers == 1:
+            self.mlp = nn.Linear(in_dim, bottleneck_dim)
+        else:
+            layers = [nn.Linear(in_dim, hidden_dim)]
+            if use_bn:
+                layers.append(nn.BatchNorm1d(hidden_dim))
+            layers.append(nn.GELU())
+            for _ in range(nlayers - 2):
+                layers.append(nn.Linear(hidden_dim, hidden_dim))
+                if use_bn:
+                    layers.append(nn.BatchNorm1d(hidden_dim))
+                layers.append(nn.GELU())
+            layers.append(nn.Linear(hidden_dim, bottleneck_dim))
+            self.mlp = nn.Sequential(*layers)
+        self.apply(self._init_weights)
+        self.last_layer = nn.utils.weight_norm(nn.Linear(bottleneck_dim, out_dim, bias=False))
+        self.last_layer.weight_g.data.fill_(1)
+        if norm_last_layer:
+            self.last_layer.weight_g.requires_grad = False
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=0.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, x):
+        x = self.mlp(x)
+        x = nn.functional.normalize(x, dim=-1, p=2)
+        x = self.last_layer(x)
+        return x
+
+def _no_grad_trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn(
+            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+            "The distribution of values may be incorrect.",
+            stacklevel=2,
+        )
+
+    with torch.no_grad():
+        # Values are generated by using a truncated uniform distribution and
+        # then using the inverse CDF for the normal distribution.
+        # Get upper and lower cdf values
+        lower = norm_cdf((a - mean) / std)
+        upper = norm_cdf((b - mean) / std)
+
+        # Uniformly fill tensor with values from [l, u], then translate to
+        # [2l-1, 2u-1].
+        tensor.uniform_(2 * lower - 1, 2 * upper - 1)
+
+        # Use inverse cdf transform for normal distribution to get truncated
+        # standard normal
+        tensor.erfinv_()
+
+        # Transform to proper mean, std
+        tensor.mul_(std * math.sqrt(2.0))
+        tensor.add_(mean)
+
+        # Clamp to ensure it's in the proper range
+        tensor.clamp_(min=a, max=b)
+        return tensor
+
+
+def trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0):
+    return _no_grad_trunc_normal_(tensor, mean, std, a, b)

geometry.py

@@ -0,0 +1,229 @@
+import torch
+import torch.nn.functional as F
+
+
+def coords_grid(b, h, w, homogeneous=False, device=None):
+    y, x = torch.meshgrid(torch.arange(h), torch.arange(w))  # [H, W]
+
+    stacks = [x, y]
+
+    if homogeneous:
+        ones = torch.ones_like(x)  # [H, W]
+        stacks.append(ones)
+
+    grid = torch.stack(stacks, dim=0).float()  # [2, H, W] or [3, H, W]
+
+    grid = grid[None].repeat(b, 1, 1, 1)  # [B, 2, H, W] or [B, 3, H, W]
+
+    if device is not None:
+        grid = grid.to(device)
+
+    return grid
+
+
+def generate_window_grid(h_min, h_max, w_min, w_max, len_h, len_w, device=None):
+    assert device is not None
+
+    x, y = torch.meshgrid(
+        [
+            torch.linspace(w_min, w_max, len_w, device=device),
+            torch.linspace(h_min, h_max, len_h, device=device),
+        ],
+    )
+    grid = torch.stack((x, y), -1).transpose(0, 1).float()  # [H, W, 2]
+
+    return grid
+
+
+def normalize_coords(coords, h, w):
+    # coords: [B, H, W, 2]
+    c = torch.Tensor([(w - 1) / 2.0, (h - 1) / 2.0]).float().to(coords.device)
+    return (coords - c) / c  # [-1, 1]
+
+
+def bilinear_sample(img, sample_coords, mode="bilinear", padding_mode="zeros", return_mask=False):
+    # img: [B, C, H, W]
+    # sample_coords: [B, 2, H, W] in image scale
+    if sample_coords.size(1) != 2:  # [B, H, W, 2]
+        sample_coords = sample_coords.permute(0, 3, 1, 2)
+
+    b, _, h, w = sample_coords.shape
+
+    # Normalize to [-1, 1]
+    x_grid = 2 * sample_coords[:, 0] / (w - 1) - 1
+    y_grid = 2 * sample_coords[:, 1] / (h - 1) - 1
+
+    grid = torch.stack([x_grid, y_grid], dim=-1)  # [B, H, W, 2]
+
+    img = F.grid_sample(img, grid, mode=mode, padding_mode=padding_mode, align_corners=False)
+
+    if return_mask:
+        mask = (x_grid >= -1) & (y_grid >= -1) & (x_grid <= 1) & (y_grid <= 1)  # [B, H, W]
+
+        return img, mask
+
+    return img
+
+
+def flow_warp(feature, flow, mask=False, padding_mode="zeros"):
+    b, c, h, w = feature.size()
+    assert flow.size(1) == 2
+
+    grid = coords_grid(b, h, w).to(flow.device) + flow  # [B, 2, H, W]
+
+    return bilinear_sample(feature, grid, padding_mode=padding_mode, return_mask=mask)
+
+
+def forward_backward_consistency_check(fwd_flow, bwd_flow, alpha=0.01, beta=0.5):
+    # fwd_flow, bwd_flow: [B, 2, H, W]
+    # alpha and beta values are following UnFlow (https://arxiv.org/abs/1711.07837)
+    assert fwd_flow.dim() == 4 and bwd_flow.dim() == 4
+    assert fwd_flow.size(1) == 2 and bwd_flow.size(1) == 2
+    flow_mag = torch.norm(fwd_flow, dim=1) + torch.norm(bwd_flow, dim=1)  # [B, H, W]
+
+    warped_bwd_flow = flow_warp(bwd_flow, fwd_flow)  # [B, 2, H, W]
+    warped_fwd_flow = flow_warp(fwd_flow, bwd_flow)  # [B, 2, H, W]
+
+    diff_fwd = torch.norm(fwd_flow + warped_bwd_flow, dim=1)  # [B, H, W]
+    diff_bwd = torch.norm(bwd_flow + warped_fwd_flow, dim=1)
+
+    threshold = alpha * flow_mag + beta
+
+    fwd_occ = (diff_fwd > threshold).float()  # [B, H, W]
+    bwd_occ = (diff_bwd > threshold).float()
+
+    return fwd_occ, bwd_occ
+
+
+def back_project(depth, intrinsics):
+    # Back project 2D pixel coords to 3D points
+    # depth: [B, H, W]
+    # intrinsics: [B, 3, 3]
+    b, h, w = depth.shape
+    grid = coords_grid(b, h, w, homogeneous=True, device=depth.device)  # [B, 3, H, W]
+
+    intrinsics_inv = torch.inverse(intrinsics)  # [B, 3, 3]
+
+    points = intrinsics_inv.bmm(grid.view(b, 3, -1)).view(b, 3, h, w) * depth.unsqueeze(
+        1
+    )  # [B, 3, H, W]
+
+    return points
+
+
+def camera_transform(points_ref, extrinsics_ref=None, extrinsics_tgt=None, extrinsics_rel=None):
+    # Transform 3D points from reference camera to target camera
+    # points_ref: [B, 3, H, W]
+    # extrinsics_ref: [B, 4, 4]
+    # extrinsics_tgt: [B, 4, 4]
+    # extrinsics_rel: [B, 4, 4], relative pose transform
+    b, _, h, w = points_ref.shape
+
+    if extrinsics_rel is None:
+        extrinsics_rel = torch.bmm(extrinsics_tgt, torch.inverse(extrinsics_ref))  # [B, 4, 4]
+
+    points_tgt = (
+        torch.bmm(extrinsics_rel[:, :3, :3], points_ref.view(b, 3, -1))
+        + extrinsics_rel[:, :3, -1:]
+    )  # [B, 3, H*W]
+
+    points_tgt = points_tgt.view(b, 3, h, w)  # [B, 3, H, W]
+
+    return points_tgt
+
+
+def reproject(points_tgt, intrinsics, return_mask=False):
+    # reproject to target view
+    # points_tgt: [B, 3, H, W]
+    # intrinsics: [B, 3, 3]
+
+    b, _, h, w = points_tgt.shape
+
+    proj_points = torch.bmm(intrinsics, points_tgt.view(b, 3, -1)).view(b, 3, h, w)  # [B, 3, H, W]
+
+    X = proj_points[:, 0]
+    Y = proj_points[:, 1]
+    Z = proj_points[:, 2].clamp(min=1e-3)
+
+    pixel_coords = torch.stack([X / Z, Y / Z], dim=1).view(
+        b, 2, h, w
+    )  # [B, 2, H, W] in image scale
+
+    if return_mask:
+        # valid mask in pixel space
+        mask = (
+            (pixel_coords[:, 0] >= 0)
+            & (pixel_coords[:, 0] <= (w - 1))
+            & (pixel_coords[:, 1] >= 0)
+            & (pixel_coords[:, 1] <= (h - 1))
+        )  # [B, H, W]
+
+        return pixel_coords, mask
+
+    return pixel_coords
+
+
+def reproject_coords(
+    depth_ref,
+    intrinsics,
+    extrinsics_ref=None,
+    extrinsics_tgt=None,
+    extrinsics_rel=None,
+    return_mask=False,
+):
+    # Compute reprojection sample coords
+    points_ref = back_project(depth_ref, intrinsics)  # [B, 3, H, W]
+    points_tgt = camera_transform(
+        points_ref, extrinsics_ref, extrinsics_tgt, extrinsics_rel=extrinsics_rel
+    )
+
+    if return_mask:
+        reproj_coords, mask = reproject(
+            points_tgt, intrinsics, return_mask=return_mask
+        )  # [B, 2, H, W] in image scale
+
+        return reproj_coords, mask
+
+    reproj_coords = reproject(
+        points_tgt, intrinsics, return_mask=return_mask
+    )  # [B, 2, H, W] in image scale
+
+    return reproj_coords
+
+
+def compute_flow_with_depth_pose(
+    depth_ref,
+    intrinsics,
+    extrinsics_ref=None,
+    extrinsics_tgt=None,
+    extrinsics_rel=None,
+    return_mask=False,
+):
+    b, h, w = depth_ref.shape
+    coords_init = coords_grid(b, h, w, device=depth_ref.device)  # [B, 2, H, W]
+
+    if return_mask:
+        reproj_coords, mask = reproject_coords(
+            depth_ref,
+            intrinsics,
+            extrinsics_ref,
+            extrinsics_tgt,
+            extrinsics_rel=extrinsics_rel,
+            return_mask=return_mask,
+        )  # [B, 2, H, W]
+        rigid_flow = reproj_coords - coords_init
+
+        return rigid_flow, mask
+
+    reproj_coords = reproject_coords(
+        depth_ref,
+        intrinsics,
+        extrinsics_ref,
+        extrinsics_tgt,
+        extrinsics_rel=extrinsics_rel,
+        return_mask=return_mask,
+    )  # [B, 2, H, W]
+
+    rigid_flow = reproj_coords - coords_init
+
+    return rigid_flow

matching.py

@@ -0,0 +1,505 @@
+import torch
+import torch.nn.functional as F
+
+from .geometry import coords_grid, generate_window_grid, normalize_coords
+
+
+def global_correlation_softmax_prototype(
+    feature0,
+    feature1,
+    value,
+    pred_bidir_flow=False,
+    corr_mask=None,
+):
+    """
+    feature0: [B, C, H, W]
+    feature1: [B, C, H, W]
+    value: [B, C1, H, W]
+    corr_mask: [B, H*W, H*W] or None, if not None, the value will be masked out
+    """
+    # global correlation
+    b, c, h, w = feature0.shape
+    c_value = value.size(1)
+    value = value.view(b, c_value, -1).permute(0, 2, 1)  # [B, H*W, C1]
+
+    feature0 = feature0.view(b, c, -1).permute(0, 2, 1)  # [B, H*W, C]
+    feature1 = feature1.view(b, c, -1)  # [B, C, H*W]
+
+    correlation = torch.matmul(feature0, feature1).view(b, h, w, h, w) / (
+        c**0.5
+    )  # [B, H, W, H, W]
+
+    correlation = correlation.view(b, h * w, h * w)  # [B, H*W, H*W]
+
+    if pred_bidir_flow:
+        correlation = torch.cat(
+            (correlation, correlation.permute(0, 2, 1)), dim=0
+        )  # [2*B, H*W, H*W]
+        value = value.repeat(2, 1, 1)  # [2*B, H*W, 2]
+        b = b * 2
+
+    if corr_mask is not None:
+        # mask out the correlation with corr_mask
+        if corr_mask.dtype == torch.bool:
+            # binary mask
+            correlation[corr_mask] = -65504.0
+            prob = F.softmax(correlation, dim=-1)  # [B, H*W, H*W]
+        else:
+            # float mask
+            # bin_mask = corr_mask < 0
+            # correlation[bin_mask] = -65504.0
+            prob = F.softmax(correlation, dim=-1)  # [B, H*W, H*W]
+            # soft_mask = torch.clamp(corr_mask, min=0)
+            prob = prob * corr_mask
+            # normalize
+            prob = prob / (prob.sum(dim=2, keepdim=True) + 1e-8)
+    else:
+        prob = F.softmax(correlation, dim=-1)  # [B, H*W, H*W]
+    #     if corr_mask.dtype == torch.bool:
+    #         # binary mask
+    #         bin_mask = corr_mask
+    #         num_true = bin_mask.sum(-1).unique().item()
+    #         soft_mask = torch.ones((b, h * w, num_true)).to(corr_mask.device)
+    #     else:
+    #         # float mask
+    #         bin_mask = corr_mask >= 0
+    #         num_true = bin_mask.sum(-1).unique().item()
+    #         soft_mask = corr_mask[bin_mask].view(b, h * w, num_true)
+    #         assert soft_mask.min() >= 0
+    #     correlation_seleted = correlation[bin_mask].view(b, h * w, num_true)
+    #     prob_seleted = F.softmax(correlation_seleted, dim=-1)  # [B, H*W, num_true]
+    #     prob_seleted = soft_mask * prob_seleted
+    #     prob_seleted = prob_seleted / (prob_seleted.sum(dim=2, keepdim=True) + 1e-8)
+    #     prob = torch.zeros_like(correlation)
+    #     prob[bin_mask] = prob_seleted.view(-1)
+    # else:
+    #     prob = F.softmax(correlation, dim=-1)
+
+    result = torch.matmul(prob, value).view(b, h, w, c_value).permute(0, 3, 1, 2)  # [B, 2, H, W]\
+    return result, correlation
+
+
+def local_correlation_softmax_prototype(
+    feature0,
+    feature1,
+    value,
+    radius=5,
+    pred_bidir_flow=False,
+    corr_mask=None,
+):
+    """
+    softmax around argmax point
+    feature0: [B, C, H, W]
+    feature1: [B, C, H, W]
+    value: [B, C1, H, W]
+    corr_mask: [B, H*W, H*W] or None, if not None, the value will be masked out
+    """
+    b, c, h, w = feature0.shape
+    c_value = value.size(1)
+    value = value.view(b, c_value, -1).permute(0, 2, 1)  # [B, H*W, C1]
+
+    feature0 = feature0.view(b, c, -1).permute(0, 2, 1)  # [B, H*W, C]
+    feature1 = feature1.view(b, c, -1)  # [B, C, H*W]
+
+    correlation = torch.matmul(feature0, feature1).view(b, h, w, h, w) / (
+        c**0.5
+    )  # [B, H, W, H, W]
+
+    correlation = correlation.view(b, h * w, h * w)  # [B, H*W, H*W]
+
+    if pred_bidir_flow:
+        correlation = torch.cat(
+            (correlation, correlation.permute(0, 2, 1)), dim=0
+        )  # [2*B, H*W, H*W]
+        value = value.repeat(2, 1, 1)  # [2*B, H*W, 2]
+        b = b * 2
+
+    if corr_mask is not None:
+        # mask out the correlation with corr_mask
+        if corr_mask.dtype == torch.bool:
+            # binary mask
+            correlation[corr_mask] = -65504.0
+            prob = F.softmax(correlation, dim=-1)  # [B, H*W, H*W]
+        else:
+            # float mask
+            # bin_mask = corr_mask < 0
+            # correlation[bin_mask] = -65504.0
+            prob = F.softmax(correlation, dim=-1)  # [B, H*W, H*W]
+            # soft_mask = torch.clamp(corr_mask, min=0)
+            prob = prob * corr_mask
+    else:
+        prob = F.softmax(correlation, dim=-1)  # [B, H*W, H*W]
+
+    # get local prob
+    # B, H*W, 2
+    coords = coords_grid(b, h, w, device=feature0.device).flatten(2).permute(0, 2, 1)
+    # B, H*W, H*W, 2
+    coords = coords.unsqueeze(1).repeat(1, h * w, 1, 1)
+    # B, H*W
+    argmax_pos = torch.argmax(prob, dim=2)
+    # B, H*W, 1, 2
+    argmax_pos = argmax_pos.view(b, h * w, 1, 1).repeat(1, 1, 1, 2)
+    # B, H*W, 1, 2
+    pos = torch.gather(coords, 2, argmax_pos)
+    # B, H*W, H*W
+    valid = ((coords - pos).square().sum(dim=-1) < (radius**2)).float()
+    prob = prob * valid
+
+    # normalize
+    prob = prob / (prob.sum(dim=2, keepdim=True) + 1e-8)
+    result = torch.matmul(prob, value).view(b, h, w, c_value).permute(0, 3, 1, 2)  # [B, 2, H, W]\
+    return result, correlation
+
+
+def global_correlation_softmax(
+    feature0,
+    feature1,
+    pred_bidir_flow=False,
+):
+    # global correlation
+    b, c, h, w = feature0.shape
+    feature0 = feature0.view(b, c, -1).permute(0, 2, 1)  # [B, H*W, C]
+    feature1 = feature1.view(b, c, -1)  # [B, C, H*W]
+
+    correlation = torch.matmul(feature0, feature1).view(b, h, w, h, w) / (
+        c**0.5
+    )  # [B, H, W, H, W]
+
+    # flow from softmax
+    init_grid = coords_grid(b, h, w).to(correlation.device)  # [B, 2, H, W]
+    grid = init_grid.view(b, 2, -1).permute(0, 2, 1)  # [B, H*W, 2]
+
+    correlation = correlation.view(b, h * w, h * w)  # [B, H*W, H*W]
+
+    if pred_bidir_flow:
+        correlation = torch.cat(
+            (correlation, correlation.permute(0, 2, 1)), dim=0
+        )  # [2*B, H*W, H*W]
+        init_grid = init_grid.repeat(2, 1, 1, 1)  # [2*B, 2, H, W]
+        grid = grid.repeat(2, 1, 1)  # [2*B, H*W, 2]
+        b = b * 2
+
+    prob = F.softmax(correlation, dim=-1)  # [B, H*W, H*W]
+
+    correspondence = torch.matmul(prob, grid).view(b, h, w, 2).permute(0, 3, 1, 2)  # [B, 2, H, W]
+
+    # when predicting bidirectional flow, flow is the concatenation of forward flow and backward flow
+    flow = correspondence - init_grid
+
+    return flow, prob
+
+
+def local_correlation_softmax(
+    feature0,
+    feature1,
+    local_radius,
+    padding_mode="zeros",
+):
+    b, c, h, w = feature0.size()
+    coords_init = coords_grid(b, h, w).to(feature0.device)  # [B, 2, H, W]
+    coords = coords_init.view(b, 2, -1).permute(0, 2, 1)  # [B, H*W, 2]
+
+    local_h = 2 * local_radius + 1
+    local_w = 2 * local_radius + 1
+
+    window_grid = generate_window_grid(
+        -local_radius,
+        local_radius,
+        -local_radius,
+        local_radius,
+        local_h,
+        local_w,
+        device=feature0.device,
+    )  # [2R+1, 2R+1, 2]
+    window_grid = window_grid.reshape(-1, 2).repeat(b, 1, 1, 1)  # [B, 1, (2R+1)^2, 2]
+    sample_coords = coords.unsqueeze(-2) + window_grid  # [B, H*W, (2R+1)^2, 2]
+
+    sample_coords_softmax = sample_coords
+
+    # exclude coords that are out of image space
+    valid_x = (sample_coords[:, :, :, 0] >= 0) & (
+        sample_coords[:, :, :, 0] < w
+    )  # [B, H*W, (2R+1)^2]
+    valid_y = (sample_coords[:, :, :, 1] >= 0) & (
+        sample_coords[:, :, :, 1] < h
+    )  # [B, H*W, (2R+1)^2]
+
+    valid = valid_x & valid_y  # [B, H*W, (2R+1)^2], used to mask out invalid values when softmax
+
+    # normalize coordinates to [-1, 1]
+    sample_coords_norm = normalize_coords(sample_coords, h, w)  # [-1, 1]
+    window_feature = F.grid_sample(
+        feature1, sample_coords_norm, padding_mode=padding_mode, align_corners=False
+    ).permute(
+        0, 2, 1, 3
+    )  # [B, H*W, C, (2R+1)^2]
+    feature0_view = feature0.permute(0, 2, 3, 1).view(b, h * w, 1, c)  # [B, H*W, 1, C]
+
+    corr = torch.matmul(feature0_view, window_feature).view(b, h * w, -1) / (
+        c**0.5
+    )  # [B, H*W, (2R+1)^2]
+
+    # mask invalid locations
+    corr[~valid] = -1e9
+
+    prob = F.softmax(corr, -1)  # [B, H*W, (2R+1)^2]
+
+    correspondence = (
+        torch.matmul(prob.unsqueeze(-2), sample_coords_softmax)
+        .squeeze(-2)
+        .view(b, h, w, 2)
+        .permute(0, 3, 1, 2)
+    )  # [B, 2, H, W]
+
+    flow = correspondence - coords_init
+    match_prob = prob
+
+    return flow, match_prob
+
+
+def local_correlation_with_flow(
+    feature0,
+    feature1,
+    flow,
+    local_radius,
+    padding_mode="zeros",
+    dilation=1,
+):
+    b, c, h, w = feature0.size()
+    coords_init = coords_grid(b, h, w).to(feature0.device)  # [B, 2, H, W]
+    coords = coords_init.view(b, 2, -1).permute(0, 2, 1)  # [B, H*W, 2]
+
+    local_h = 2 * local_radius + 1
+    local_w = 2 * local_radius + 1
+
+    window_grid = generate_window_grid(
+        -local_radius,
+        local_radius,
+        -local_radius,
+        local_radius,
+        local_h,
+        local_w,
+        device=feature0.device,
+    )  # [2R+1, 2R+1, 2]
+    window_grid = window_grid.reshape(-1, 2).repeat(b, 1, 1, 1)  # [B, 1, (2R+1)^2, 2]
+    sample_coords = coords.unsqueeze(-2) + window_grid * dilation  # [B, H*W, (2R+1)^2, 2]
+
+    # flow can be zero when using features after transformer
+    if not isinstance(flow, float):
+        sample_coords = sample_coords + flow.view(b, 2, -1).permute(0, 2, 1).unsqueeze(
+            -2
+        )  # [B, H*W, (2R+1)^2, 2]
+    else:
+        assert flow == 0.0
+
+    # normalize coordinates to [-1, 1]
+    sample_coords_norm = normalize_coords(sample_coords, h, w)  # [-1, 1]
+    window_feature = F.grid_sample(
+        feature1, sample_coords_norm, padding_mode=padding_mode, align_corners=False
+    ).permute(
+        0, 2, 1, 3
+    )  # [B, H*W, C, (2R+1)^2]
+    feature0_view = feature0.permute(0, 2, 3, 1).view(b, h * w, 1, c)  # [B, H*W, 1, C]
+
+    corr = torch.matmul(feature0_view, window_feature).view(b, h * w, -1) / (
+        c**0.5
+    )  # [B, H*W, (2R+1)^2]
+
+    corr = corr.view(b, h, w, -1).permute(0, 3, 1, 2).contiguous()  # [B, (2R+1)^2, H, W]
+
+    return corr
+
+
+def global_correlation_softmax_stereo(
+    feature0,
+    feature1,
+):
+    # global correlation on horizontal direction
+    b, c, h, w = feature0.shape
+
+    x_grid = torch.linspace(0, w - 1, w, device=feature0.device)  # [W]
+
+    feature0 = feature0.permute(0, 2, 3, 1)  # [B, H, W, C]
+    feature1 = feature1.permute(0, 2, 1, 3)  # [B, H, C, W]
+
+    correlation = torch.matmul(feature0, feature1) / (c**0.5)  # [B, H, W, W]
+
+    # mask subsequent positions to make disparity positive
+    mask = torch.triu(torch.ones((w, w)), diagonal=1).type_as(feature0)  # [W, W]
+    valid_mask = (mask == 0).unsqueeze(0).unsqueeze(0).repeat(b, h, 1, 1)  # [B, H, W, W]
+
+    correlation[~valid_mask] = -1e9
+
+    prob = F.softmax(correlation, dim=-1)  # [B, H, W, W]
+
+    correspondence = (x_grid.view(1, 1, 1, w) * prob).sum(-1)  # [B, H, W]
+
+    # NOTE: unlike flow, disparity is typically positive
+    disparity = x_grid.view(1, 1, w).repeat(b, h, 1) - correspondence  # [B, H, W]
+
+    return disparity.unsqueeze(1), prob  # feature resolution
+
+
+def local_correlation_softmax_stereo(
+    feature0,
+    feature1,
+    local_radius,
+):
+    b, c, h, w = feature0.size()
+    coords_init = coords_grid(b, h, w).to(feature0.device)  # [B, 2, H, W]
+    coords = coords_init.view(b, 2, -1).permute(0, 2, 1).contiguous()  # [B, H*W, 2]
+
+    local_h = 1
+    local_w = 2 * local_radius + 1
+
+    window_grid = generate_window_grid(
+        0, 0, -local_radius, local_radius, local_h, local_w, device=feature0.device
+    )  # [1, 2R+1, 2]
+    window_grid = window_grid.reshape(-1, 2).repeat(b, 1, 1, 1)  # [B, 1, (2R+1), 2]
+    sample_coords = coords.unsqueeze(-2) + window_grid  # [B, H*W, (2R+1), 2]
+
+    sample_coords_softmax = sample_coords
+
+    # exclude coords that are out of image space
+    valid_x = (sample_coords[:, :, :, 0] >= 0) & (
+        sample_coords[:, :, :, 0] < w
+    )  # [B, H*W, (2R+1)^2]
+    valid_y = (sample_coords[:, :, :, 1] >= 0) & (
+        sample_coords[:, :, :, 1] < h
+    )  # [B, H*W, (2R+1)^2]
+
+    valid = valid_x & valid_y  # [B, H*W, (2R+1)^2], used to mask out invalid values when softmax
+
+    # normalize coordinates to [-1, 1]
+    sample_coords_norm = normalize_coords(sample_coords, h, w)  # [-1, 1]
+    window_feature = F.grid_sample(
+        feature1, sample_coords_norm, padding_mode="zeros", align_corners=False
+    ).permute(
+        0, 2, 1, 3
+    )  # [B, H*W, C, (2R+1)]
+    feature0_view = (
+        feature0.permute(0, 2, 3, 1).contiguous().view(b, h * w, 1, c)
+    )  # [B, H*W, 1, C]
+
+    corr = torch.matmul(feature0_view, window_feature).view(b, h * w, -1) / (
+        c**0.5
+    )  # [B, H*W, (2R+1)]
+
+    # mask invalid locations
+    corr[~valid] = -1e9
+
+    prob = F.softmax(corr, -1)  # [B, H*W, (2R+1)]
+
+    correspondence = (
+        torch.matmul(prob.unsqueeze(-2), sample_coords_softmax)
+        .squeeze(-2)
+        .view(b, h, w, 2)
+        .permute(0, 3, 1, 2)
+        .contiguous()
+    )  # [B, 2, H, W]
+
+    flow = correspondence - coords_init  # flow at feature resolution
+    match_prob = prob
+
+    flow_x = -flow[:, :1]  # [B, 1, H, W]
+
+    return flow_x, match_prob
+
+
+def correlation_softmax_depth(
+    feature0,
+    feature1,
+    intrinsics,
+    pose,
+    depth_candidates,
+    depth_from_argmax=False,
+    pred_bidir_depth=False,
+):
+    b, c, h, w = feature0.size()
+    assert depth_candidates.dim() == 4  # [B, D, H, W]
+    scale_factor = c**0.5
+
+    if pred_bidir_depth:
+        feature0, feature1 = torch.cat((feature0, feature1), dim=0), torch.cat(
+            (feature1, feature0), dim=0
+        )
+        intrinsics = intrinsics.repeat(2, 1, 1)
+        pose = torch.cat((pose, torch.inverse(pose)), dim=0)
+        depth_candidates = depth_candidates.repeat(2, 1, 1, 1)
+
+    # depth candidates are actually inverse depth
+    warped_feature1 = warp_with_pose_depth_candidates(
+        feature1,
+        intrinsics,
+        pose,
+        1.0 / depth_candidates,
+    )  # [B, C, D, H, W]
+
+    correlation = (feature0.unsqueeze(2) * warped_feature1).sum(1) / scale_factor  # [B, D, H, W]
+
+    match_prob = F.softmax(correlation, dim=1)  # [B, D, H, W]
+
+    # for cross-task transfer (flow -> depth), extract depth with argmax at test time
+    if depth_from_argmax:
+        index = torch.argmax(match_prob, dim=1, keepdim=True)
+        depth = torch.gather(depth_candidates, dim=1, index=index)
+    else:
+        depth = (match_prob * depth_candidates).sum(dim=1, keepdim=True)  # [B, 1, H, W]
+
+    return depth, match_prob
+
+
+def warp_with_pose_depth_candidates(
+    feature1,
+    intrinsics,
+    pose,
+    depth,
+    clamp_min_depth=1e-3,
+):
+    """
+    feature1: [B, C, H, W]
+    intrinsics: [B, 3, 3]
+    pose: [B, 4, 4]
+    depth: [B, D, H, W]
+    """
+
+    assert intrinsics.size(1) == intrinsics.size(2) == 3
+    assert pose.size(1) == pose.size(2) == 4
+    assert depth.dim() == 4
+
+    b, d, h, w = depth.size()
+    c = feature1.size(1)
+
+    with torch.no_grad():
+        # pixel coordinates
+        grid = coords_grid(b, h, w, homogeneous=True, device=depth.device)  # [B, 3, H, W]
+        # back project to 3D and transform viewpoint
+        points = torch.inverse(intrinsics).bmm(grid.view(b, 3, -1))  # [B, 3, H*W]
+        points = torch.bmm(pose[:, :3, :3], points).unsqueeze(2).repeat(1, 1, d, 1) * depth.view(
+            b, 1, d, h * w
+        )  # [B, 3, D, H*W]
+        points = points + pose[:, :3, -1:].unsqueeze(-1)  # [B, 3, D, H*W]
+        # reproject to 2D image plane
+        points = torch.bmm(intrinsics, points.view(b, 3, -1)).view(
+            b, 3, d, h * w
+        )  # [B, 3, D, H*W]
+        pixel_coords = points[:, :2] / points[:, -1:].clamp(min=clamp_min_depth)  # [B, 2, D, H*W]
+
+        # normalize to [-1, 1]
+        x_grid = 2 * pixel_coords[:, 0] / (w - 1) - 1
+        y_grid = 2 * pixel_coords[:, 1] / (h - 1) - 1
+
+        grid = torch.stack([x_grid, y_grid], dim=-1)  # [B, D, H*W, 2]
+
+    # sample features
+    warped_feature = F.grid_sample(
+        feature1,
+        grid.view(b, d * h, w, 2),
+        mode="bilinear",
+        padding_mode="zeros",
+        align_corners=False,
+    ).view(
+        b, c, d, h, w
+    )  # [B, C, D, H, W]
+
+    return warped_feature

modeling_doduo.py

@@ -0,0 +1,118 @@
+from typing import Dict, Tuple
+from PIL import Image
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from torchvision import transforms
+from transformers import PreTrainedModel
+
+from .dino import vit_small
+from .unimatch import UniMatch
+from .configuration_doduo import DoduoConfig
+
+class DoduoModel(PreTrainedModel):
+    config_class = DoduoConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = CorrSegFlowNet(
+            dino_corr_mask_ratio=config.dino_corr_mask_ratio
+        )
+
+    def forward(self, frame_src, frame_dst):
+        if isinstance(frame_src, Image.Image):
+            frame_src = self.model.process_frame(frame_src)
+            frame_dst = self.model.process_frame(frame_dst)
+        assert frame_src.shape == frame_dst.shape
+        return self.model(frame_src, frame_dst)
+    
+class CorrSegFlowNet(nn.Module):
+    def __init__(
+        self,
+        dino_corr_mask_ratio: float = 0.1,
+    ):
+        super().__init__()
+
+        self.dino_corr_mask_ratio = dino_corr_mask_ratio
+        self.unimatch = UniMatch(bilinear_upsample=True)
+        self.dino = vit_small(patch_size=8, num_classes=0)
+        for k in self.dino.parameters():
+            k.requires_grad = False
+
+        self.transform = transforms.Compose(
+            [
+                lambda x: transforms.ToTensor()(x)[:3],
+                transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
+            ]
+        )
+
+    def process_frame(self, frame):
+        device = next(self.parameters()).device
+        frame = self.transform(frame)
+        frame = frame.unsqueeze(0).to(device)
+        return frame
+
+    def forward(
+        self,
+        frame_src,
+        frame_dst,
+    ):
+        corr_mask = get_dino_corr_mask(
+            self.dino,
+            frame_src,
+            frame_dst,
+            mask_ratio=self.dino_corr_mask_ratio
+        )
+
+        flow, flow_low, correlation, feature0, feature1 = self.unimatch(
+            frame_src,
+            frame_dst,
+            return_feature=True,
+            bidirectional=False,
+            cycle_consistency=False,
+            corr_mask=corr_mask,
+        )
+        return flow
+
+@torch.no_grad()
+def extract_dino_feature(model, frame, return_h_w=False):
+    """frame: B, C, H, W"""
+    B = frame.shape[0]
+    out = model.get_intermediate_layers(frame, n=1)[0]
+    out = out[:, 1:, :]  # we discard the [CLS] token
+    h, w = int(frame.shape[2] / model.patch_embed.patch_size), int(
+        frame.shape[3] / model.patch_embed.patch_size
+    )
+    dim = out.shape[-1]
+    out = out.reshape(B, -1, dim)
+    if return_h_w:
+        return out, h, w
+    return out
+
+@torch.no_grad()
+def get_dino_corr_mask(
+    model, frame_src, frame_dst, mask_ratio
+):
+    # frame_src: B x C x H x W
+    # frame_dst: B x C x H x W
+    # mask_ratio: ratio of pixels to be masked
+    # return: B x h*w x h*w
+    feat_1, h, w = extract_dino_feature(model, frame_src, return_h_w=True)
+    feat_2 = extract_dino_feature(model, frame_dst)
+
+    feat_1_norm = F.normalize(feat_1, dim=2, p=2)
+    feat_2_norm = F.normalize(feat_2, dim=2, p=2)
+    aff_raw = torch.einsum("bnc,bmc->bnm", [feat_1_norm, feat_2_norm])
+
+    if mask_ratio <= 0:
+        # no corr mask
+        corr_mask = None
+    else:
+        if aff_raw.dtype == torch.float16:
+            aff_raw = aff_raw.float()
+        aff_percentile = torch.quantile(aff_raw, mask_ratio, 2, keepdim=True)
+        # True for masked
+        corr_mask = aff_raw < aff_percentile
+    return corr_mask

position.py

@@ -0,0 +1,49 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+# https://github.com/facebookresearch/detr/blob/main/models/position_encoding.py
+
+import math
+
+import torch
+import torch.nn as nn
+
+
+class PositionEmbeddingSine(nn.Module):
+    """This is a more standard version of the position embedding, very similar to the one used by
+    the Attention is all you need paper, generalized to work on images."""
+
+    def __init__(self, num_pos_feats=64, temperature=10000, normalize=True, scale=None):
+        super().__init__()
+        self.num_pos_feats = num_pos_feats
+        self.temperature = temperature
+        self.normalize = normalize
+        if scale is not None and normalize is False:
+            raise ValueError("normalize should be True if scale is passed")
+        if scale is None:
+            scale = 2 * math.pi
+        self.scale = scale
+
+    def forward(self, x):
+        # x = tensor_list.tensors  # [B, C, H, W]
+        # mask = tensor_list.mask  # [B, H, W], input with padding, valid as 0
+        b, c, h, w = x.size()
+        mask = torch.ones((b, h, w), device=x.device)  # [B, H, W]
+        y_embed = mask.cumsum(1, dtype=torch.float32)
+        x_embed = mask.cumsum(2, dtype=torch.float32)
+        if self.normalize:
+            eps = 1e-6
+            y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
+            x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
+
+        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
+
+        pos_x = x_embed[:, :, :, None] / dim_t
+        pos_y = y_embed[:, :, :, None] / dim_t
+        pos_x = torch.stack(
+            (pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4
+        ).flatten(3)
+        pos_y = torch.stack(
+            (pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4
+        ).flatten(3)
+        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+        return pos

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5340e547b103a76197fe408906930793aea54ed21a4c2f7845e73d1ca5810022
+size 103562927

transformer.py

@@ -0,0 +1,340 @@
+import torch
+import torch.nn as nn
+
+from .attention import (
+    single_head_full_attention,
+    single_head_full_attention_1d,
+    single_head_split_window_attention,
+    single_head_split_window_attention_1d,
+)
+from .utils import generate_shift_window_attn_mask, generate_shift_window_attn_mask_1d
+
+
+class TransformerLayer(nn.Module):
+    def __init__(
+        self,
+        d_model=128,
+        nhead=1,
+        no_ffn=False,
+        ffn_dim_expansion=4,
+    ):
+        super().__init__()
+
+        self.dim = d_model
+        self.nhead = nhead
+        self.no_ffn = no_ffn
+
+        # multi-head attention
+        self.q_proj = nn.Linear(d_model, d_model, bias=False)
+        self.k_proj = nn.Linear(d_model, d_model, bias=False)
+        self.v_proj = nn.Linear(d_model, d_model, bias=False)
+
+        self.merge = nn.Linear(d_model, d_model, bias=False)
+
+        self.norm1 = nn.LayerNorm(d_model)
+
+        # no ffn after self-attn, with ffn after cross-attn
+        if not self.no_ffn:
+            in_channels = d_model * 2
+            self.mlp = nn.Sequential(
+                nn.Linear(in_channels, in_channels * ffn_dim_expansion, bias=False),
+                nn.GELU(),
+                nn.Linear(in_channels * ffn_dim_expansion, d_model, bias=False),
+            )
+
+            self.norm2 = nn.LayerNorm(d_model)
+
+    def forward(
+        self,
+        source,
+        target,
+        height=None,
+        width=None,
+        shifted_window_attn_mask=None,
+        shifted_window_attn_mask_1d=None,
+        attn_type="swin",
+        with_shift=False,
+        attn_num_splits=None,
+    ):
+        # source, target: [B, L, C]
+        query, key, value = source, target, target
+
+        # for stereo: 2d attn in self-attn, 1d attn in cross-attn
+        is_self_attn = (query - key).abs().max() < 1e-6
+
+        # single-head attention
+        query = self.q_proj(query)  # [B, L, C]
+        key = self.k_proj(key)  # [B, L, C]
+        value = self.v_proj(value)  # [B, L, C]
+
+        if attn_type == "swin" and attn_num_splits > 1:  # self, cross-attn: both swin 2d
+            if self.nhead > 1:
+                # we observe that multihead attention slows down the speed and increases the memory consumption
+                # without bringing obvious performance gains and thus the implementation is removed
+                raise NotImplementedError
+            else:
+                message = single_head_split_window_attention(
+                    query,
+                    key,
+                    value,
+                    num_splits=attn_num_splits,
+                    with_shift=with_shift,
+                    h=height,
+                    w=width,
+                    attn_mask=shifted_window_attn_mask,
+                )
+
+        elif attn_type == "self_swin2d_cross_1d":  # self-attn: swin 2d, cross-attn: full 1d
+            if self.nhead > 1:
+                raise NotImplementedError
+            else:
+                if is_self_attn:
+                    if attn_num_splits > 1:
+                        message = single_head_split_window_attention(
+                            query,
+                            key,
+                            value,
+                            num_splits=attn_num_splits,
+                            with_shift=with_shift,
+                            h=height,
+                            w=width,
+                            attn_mask=shifted_window_attn_mask,
+                        )
+                    else:
+                        # full 2d attn
+                        message = single_head_full_attention(query, key, value)  # [N, L, C]
+
+                else:
+                    # cross attn 1d
+                    message = single_head_full_attention_1d(
+                        query,
+                        key,
+                        value,
+                        h=height,
+                        w=width,
+                    )
+
+        elif attn_type == "self_swin2d_cross_swin1d":  # self-attn: swin 2d, cross-attn: swin 1d
+            if self.nhead > 1:
+                raise NotImplementedError
+            else:
+                if is_self_attn:
+                    if attn_num_splits > 1:
+                        # self attn shift window
+                        message = single_head_split_window_attention(
+                            query,
+                            key,
+                            value,
+                            num_splits=attn_num_splits,
+                            with_shift=with_shift,
+                            h=height,
+                            w=width,
+                            attn_mask=shifted_window_attn_mask,
+                        )
+                    else:
+                        # full 2d attn
+                        message = single_head_full_attention(query, key, value)  # [N, L, C]
+                else:
+                    if attn_num_splits > 1:
+                        assert shifted_window_attn_mask_1d is not None
+                        # cross attn 1d shift
+                        message = single_head_split_window_attention_1d(
+                            query,
+                            key,
+                            value,
+                            num_splits=attn_num_splits,
+                            with_shift=with_shift,
+                            h=height,
+                            w=width,
+                            attn_mask=shifted_window_attn_mask_1d,
+                        )
+                    else:
+                        message = single_head_full_attention_1d(
+                            query,
+                            key,
+                            value,
+                            h=height,
+                            w=width,
+                        )
+
+        else:
+            message = single_head_full_attention(query, key, value)  # [B, L, C]
+
+        message = self.merge(message)  # [B, L, C]
+        message = self.norm1(message)
+
+        if not self.no_ffn:
+            message = self.mlp(torch.cat([source, message], dim=-1))
+            message = self.norm2(message)
+
+        return source + message
+
+
+class TransformerBlock(nn.Module):
+    """self attention + cross attention + FFN."""
+
+    def __init__(
+        self,
+        d_model=128,
+        nhead=1,
+        ffn_dim_expansion=4,
+    ):
+        super().__init__()
+
+        self.self_attn = TransformerLayer(
+            d_model=d_model,
+            nhead=nhead,
+            no_ffn=True,
+            ffn_dim_expansion=ffn_dim_expansion,
+        )
+
+        self.cross_attn_ffn = TransformerLayer(
+            d_model=d_model,
+            nhead=nhead,
+            ffn_dim_expansion=ffn_dim_expansion,
+        )
+
+    def forward(
+        self,
+        source,
+        target,
+        height=None,
+        width=None,
+        shifted_window_attn_mask=None,
+        shifted_window_attn_mask_1d=None,
+        attn_type="swin",
+        with_shift=False,
+        attn_num_splits=None,
+    ):
+        # source, target: [B, L, C]
+
+        # self attention
+        source = self.self_attn(
+            source,
+            source,
+            height=height,
+            width=width,
+            shifted_window_attn_mask=shifted_window_attn_mask,
+            attn_type=attn_type,
+            with_shift=with_shift,
+            attn_num_splits=attn_num_splits,
+        )
+
+        # cross attention and ffn
+        source = self.cross_attn_ffn(
+            source,
+            target,
+            height=height,
+            width=width,
+            shifted_window_attn_mask=shifted_window_attn_mask,
+            shifted_window_attn_mask_1d=shifted_window_attn_mask_1d,
+            attn_type=attn_type,
+            with_shift=with_shift,
+            attn_num_splits=attn_num_splits,
+        )
+
+        return source
+
+
+class FeatureTransformer(nn.Module):
+    def __init__(
+        self,
+        num_layers=6,
+        d_model=128,
+        nhead=1,
+        ffn_dim_expansion=4,
+    ):
+        super().__init__()
+
+        self.d_model = d_model
+        self.nhead = nhead
+
+        self.layers = nn.ModuleList(
+            [
+                TransformerBlock(
+                    d_model=d_model,
+                    nhead=nhead,
+                    ffn_dim_expansion=ffn_dim_expansion,
+                )
+                for i in range(num_layers)
+            ]
+        )
+
+        for p in self.parameters():
+            if p.dim() > 1:
+                nn.init.xavier_uniform_(p)
+
+    def forward(
+        self,
+        feature0,
+        feature1,
+        attn_type="swin",
+        attn_num_splits=None,
+        **kwargs,
+    ):
+
+        b, c, h, w = feature0.shape
+        assert self.d_model == c
+
+        feature0 = feature0.flatten(-2).permute(0, 2, 1)  # [B, H*W, C]
+        feature1 = feature1.flatten(-2).permute(0, 2, 1)  # [B, H*W, C]
+
+        # 2d attention
+        if "swin" in attn_type and attn_num_splits > 1:
+            # global and refine use different number of splits
+            window_size_h = h // attn_num_splits
+            window_size_w = w // attn_num_splits
+
+            # compute attn mask once
+            shifted_window_attn_mask = generate_shift_window_attn_mask(
+                input_resolution=(h, w),
+                window_size_h=window_size_h,
+                window_size_w=window_size_w,
+                shift_size_h=window_size_h // 2,
+                shift_size_w=window_size_w // 2,
+                device=feature0.device,
+            )  # [K*K, H/K*W/K, H/K*W/K]
+        else:
+            shifted_window_attn_mask = None
+
+        # 1d attention
+        if "swin1d" in attn_type and attn_num_splits > 1:
+            window_size_w = w // attn_num_splits
+
+            # compute attn mask once
+            shifted_window_attn_mask_1d = generate_shift_window_attn_mask_1d(
+                input_w=w,
+                window_size_w=window_size_w,
+                shift_size_w=window_size_w // 2,
+                device=feature0.device,
+            )  # [K, W/K, W/K]
+        else:
+            shifted_window_attn_mask_1d = None
+
+        # concat feature0 and feature1 in batch dimension to compute in parallel
+        concat0 = torch.cat((feature0, feature1), dim=0)  # [2B, H*W, C]
+        concat1 = torch.cat((feature1, feature0), dim=0)  # [2B, H*W, C]
+
+        for i, layer in enumerate(self.layers):
+            concat0 = layer(
+                concat0,
+                concat1,
+                height=h,
+                width=w,
+                attn_type=attn_type,
+                with_shift="swin" in attn_type and attn_num_splits > 1 and i % 2 == 1,
+                attn_num_splits=attn_num_splits,
+                shifted_window_attn_mask=shifted_window_attn_mask,
+                shifted_window_attn_mask_1d=shifted_window_attn_mask_1d,
+            )
+
+            # update feature1
+            concat1 = torch.cat(concat0.chunk(chunks=2, dim=0)[::-1], dim=0)
+
+        feature0, feature1 = concat0.chunk(chunks=2, dim=0)  # [B, H*W, C]
+
+        # reshape back
+        feature0 = feature0.view(b, h, w, c).permute(0, 3, 1, 2).contiguous()  # [B, C, H, W]
+        feature1 = feature1.view(b, h, w, c).permute(0, 3, 1, 2).contiguous()  # [B, C, H, W]
+
+        return feature0, feature1

trident_conv.py

@@ -0,0 +1,96 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# https://github.com/facebookresearch/detectron2/blob/main/projects/TridentNet/tridentnet/trident_conv.py
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+from torch.nn.modules.utils import _pair
+
+
+class MultiScaleTridentConv(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride=1,
+        strides=1,
+        paddings=0,
+        dilations=1,
+        dilation=1,
+        groups=1,
+        num_branch=1,
+        test_branch_idx=-1,
+        bias=False,
+        norm=None,
+        activation=None,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.kernel_size = _pair(kernel_size)
+        self.num_branch = num_branch
+        self.stride = _pair(stride)
+        self.groups = groups
+        self.with_bias = bias
+        self.dilation = dilation
+        if isinstance(paddings, int):
+            paddings = [paddings] * self.num_branch
+        if isinstance(dilations, int):
+            dilations = [dilations] * self.num_branch
+        if isinstance(strides, int):
+            strides = [strides] * self.num_branch
+        self.paddings = [_pair(padding) for padding in paddings]
+        self.dilations = [_pair(dilation) for dilation in dilations]
+        self.strides = [_pair(stride) for stride in strides]
+        self.test_branch_idx = test_branch_idx
+        self.norm = norm
+        self.activation = activation
+
+        assert len({self.num_branch, len(self.paddings), len(self.strides)}) == 1
+
+        self.weight = nn.Parameter(
+            torch.Tensor(out_channels, in_channels // groups, *self.kernel_size)
+        )
+        if bias:
+            self.bias = nn.Parameter(torch.Tensor(out_channels))
+        else:
+            self.bias = None
+
+        nn.init.kaiming_uniform_(self.weight, nonlinearity="relu")
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0)
+
+    def forward(self, inputs):
+        num_branch = self.num_branch if self.training or self.test_branch_idx == -1 else 1
+        assert len(inputs) == num_branch
+
+        if self.training or self.test_branch_idx == -1:
+            outputs = [
+                F.conv2d(
+                    input, self.weight, self.bias, stride, padding, self.dilation, self.groups
+                )
+                for input, stride, padding in zip(inputs, self.strides, self.paddings)
+            ]
+        else:
+            outputs = [
+                F.conv2d(
+                    inputs[0],
+                    self.weight,
+                    self.bias,
+                    self.strides[self.test_branch_idx]
+                    if self.test_branch_idx == -1
+                    else self.strides[-1],
+                    self.paddings[self.test_branch_idx]
+                    if self.test_branch_idx == -1
+                    else self.paddings[-1],
+                    self.dilation,
+                    self.groups,
+                )
+            ]
+
+        if self.norm is not None:
+            outputs = [self.norm(x) for x in outputs]
+        if self.activation is not None:
+            outputs = [self.activation(x) for x in outputs]
+        return outputs

unimatch.py

@@ -0,0 +1,213 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .backbone import CNNEncoder
+from .geometry import coords_grid
+from .matching import (
+    global_correlation_softmax_prototype,
+    local_correlation_softmax_prototype,
+)
+from .transformer import FeatureTransformer
+from .utils import feature_add_position
+
+
+class UniMatch(nn.Module):
+    def __init__(
+        self,
+        num_scales=1,
+        feature_channels=128,
+        upsample_factor=8,
+        num_head=1,
+        ffn_dim_expansion=4,
+        num_transformer_layers=6,
+        bilinear_upsample=False,
+        corr_fn="global",
+    ):
+        super().__init__()
+
+        self.feature_channels = feature_channels
+        self.num_scales = num_scales
+        self.upsample_factor = upsample_factor
+        self.bilinear_upsample = bilinear_upsample
+        if corr_fn == "global":
+            self.corr_fn = global_correlation_softmax_prototype
+        elif corr_fn == "local":
+            self.corr_fn = local_correlation_softmax_prototype
+        else:
+            raise NotImplementedError(f"Correlation function {corr_fn} not implemented")
+
+        # CNN
+        self.backbone = CNNEncoder(output_dim=feature_channels, num_output_scales=num_scales)
+
+        # Transformer
+        self.transformer = FeatureTransformer(
+            num_layers=num_transformer_layers,
+            d_model=feature_channels,
+            nhead=num_head,
+            ffn_dim_expansion=ffn_dim_expansion,
+        )
+
+        # convex upsampling similar to RAFT
+        # concat feature0 and low res flow as input
+        if not bilinear_upsample:
+            self.upsampler = nn.Sequential(
+                nn.Conv2d(2 + feature_channels, 256, 3, 1, 1),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(256, upsample_factor**2 * 9, 1, 1, 0),
+            )
+
+    def extract_feature(self, img0, img1):
+        concat = torch.cat((img0, img1), dim=0)  # [2B, C, H, W]
+        features = self.backbone(concat)  # list of [2B, C, H, W], resolution from high to low
+
+        # reverse: resolution from low to high
+        features = features[::-1]
+
+        feature0, feature1 = [], []
+
+        for i in range(len(features)):
+            feature = features[i]
+            chunks = torch.chunk(feature, 2, 0)  # tuple
+            feature0.append(chunks[0])
+            feature1.append(chunks[1])
+
+        return feature0, feature1
+
+    def correlate_feature(self, feature0, feature1, attn_splits=2, attn_type="swin"):
+        feature0, feature1 = feature_add_position(
+            feature0, feature1, attn_splits, self.feature_channels
+        )
+        feature0, feature1 = self.transformer(
+            feature0,
+            feature1,
+            attn_type=attn_type,
+            attn_num_splits=attn_splits,
+        )
+        b, c, h, w = feature0.shape
+        feature0 = feature0.view(b, c, -1).permute(0, 2, 1)  # [B, H*W, C]
+        feature1 = feature1.view(b, c, -1)  # [B, C, H*W]
+        correlation = torch.matmul(feature0, feature1).view(b, h, w, h, w) / (
+            c**0.5
+        )  # [B, H, W, H, W]
+        correlation = correlation.view(b, h * w, h * w)  # [B, H*W, H*W]
+        return correlation
+
+    def forward(
+        self,
+        img0,
+        img1,
+        attn_type="swin",
+        attn_splits=2,
+        return_feature=False,
+        bidirectional=False,
+        cycle_consistency=False,
+        corr_mask=None,
+    ):
+        # list of features, resolution low to high
+        feature0_list, feature1_list = self.extract_feature(img0, img1)  # list of features
+        assert self.num_scales == 1  # multi-scale depth model is not supported yet
+        scale_idx = 0
+        feature0, feature1 = feature0_list[scale_idx], feature1_list[scale_idx]
+
+        if cycle_consistency:
+            # get both directions of features
+            feature0, feature1 = torch.cat((feature0, feature1), dim=0), torch.cat(
+                (feature1, feature0), dim=0
+            )
+
+        # add position to features
+        feature0, feature1 = feature_add_position(
+            feature0, feature1, attn_splits, self.feature_channels
+        )
+
+        # Transformer
+        feature0, feature1 = self.transformer(
+            feature0,
+            feature1,
+            attn_type=attn_type,
+            attn_num_splits=attn_splits,
+        )
+        b, c, h, w = feature0.shape
+        # downsampled_img0 = F.interpolate(img0, size=(h, w), mode="bilinear", align_corners=False)
+        flow_coords = coords_grid(b, h, w).to(feature0.device)  # [B, 2, H, W]
+        # values = torch.cat((downsampled_img0, flow_coords), dim=1)  # [B, 5, H, W]
+        # correlation and softmax
+        query_results, correlation = self.corr_fn(
+            feature0, feature1, flow_coords, pred_bidir_flow=bidirectional, corr_mask=corr_mask
+        )
+        if bidirectional:
+            flow_coords = torch.cat((flow_coords, flow_coords), dim=0)
+            up_feature = torch.cat((feature0, feature1), dim=0)
+        else:
+            up_feature = feature0
+        flow = query_results - flow_coords
+        flow_up = self.upsample_flow(flow, up_feature, bilinear=self.bilinear_upsample)
+        if return_feature:
+            return flow_up, flow, correlation, feature0, feature1
+        else:
+            return flow_up, flow, correlation
+
+    def forward_features(
+        self,
+        img0,
+        img1,
+        attn_type="swin",
+        attn_splits=2,
+    ):
+
+        feature0_list, feature1_list = self.extract_feature(img0, img1)  # list of features
+        assert self.num_scales == 1  # multi-scale depth model is not supported yet
+        scale_idx = 0
+        feature0, feature1 = feature0_list[scale_idx], feature1_list[scale_idx]
+        # add position to features
+        feature0, feature1 = feature_add_position(
+            feature0, feature1, attn_splits, self.feature_channels
+        )
+
+        # Transformer
+        feature0, feature1 = self.transformer(
+            feature0,
+            feature1,
+            attn_type=attn_type,
+            attn_num_splits=attn_splits,
+        )
+        return feature0, feature1
+
+    def upsample_flow(self, flow, feature, bilinear=False, upsample_factor=8, is_depth=False):
+        if bilinear:
+            multiplier = 1 if is_depth else upsample_factor
+            up_flow = (
+                F.interpolate(
+                    flow, scale_factor=upsample_factor, mode="bilinear", align_corners=False
+                )
+                * multiplier
+            )
+        else:
+            concat = torch.cat((flow, feature), dim=1)
+            mask = self.upsampler(concat)
+            up_flow = upsample_flow_with_mask(
+                flow, mask, upsample_factor=self.upsample_factor, is_depth=is_depth
+            )
+        return up_flow
+
+
+def upsample_flow_with_mask(flow, up_mask, upsample_factor, is_depth=False):
+    # convex upsampling following raft
+
+    mask = up_mask
+    b, flow_channel, h, w = flow.shape
+    mask = mask.view(b, 1, 9, upsample_factor, upsample_factor, h, w)  # [B, 1, 9, K, K, H, W]
+    mask = torch.softmax(mask, dim=2)
+
+    multiplier = 1 if is_depth else upsample_factor
+    up_flow = F.unfold(multiplier * flow, [3, 3], padding=1)
+    up_flow = up_flow.view(b, flow_channel, 9, 1, 1, h, w)  # [B, 2, 9, 1, 1, H, W]
+
+    up_flow = torch.sum(mask * up_flow, dim=2)  # [B, 2, K, K, H, W]
+    up_flow = up_flow.permute(0, 1, 4, 2, 5, 3)  # [B, 2, K, H, K, W]
+    up_flow = up_flow.reshape(
+        b, flow_channel, upsample_factor * h, upsample_factor * w
+    )  # [B, 2, K*H, K*W]
+
+    return up_flow

utils.py

@@ -0,0 +1,257 @@
+import torch
+import torch.nn.functional as F
+
+from .position import PositionEmbeddingSine
+
+
+def generate_window_grid(h_min, h_max, w_min, w_max, len_h, len_w, device=None):
+    assert device is not None
+
+    x, y = torch.meshgrid(
+        [
+            torch.linspace(w_min, w_max, len_w, device=device),
+            torch.linspace(h_min, h_max, len_h, device=device),
+        ],
+    )
+    grid = torch.stack((x, y), -1).transpose(0, 1).float()  # [H, W, 2]
+
+    return grid
+
+
+def normalize_coords(coords, h, w):
+    # coords: [B, H, W, 2]
+    c = torch.Tensor([(w - 1) / 2.0, (h - 1) / 2.0]).float().to(coords.device)
+    return (coords - c) / c  # [-1, 1]
+
+
+def normalize_img(img0, img1):
+    # loaded images are in [0, 255]
+    # normalize by ImageNet mean and std
+    mean = torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1).to(img1.device)
+    std = torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1).to(img1.device)
+    img0 = (img0 / 255.0 - mean) / std
+    img1 = (img1 / 255.0 - mean) / std
+
+    return img0, img1
+
+
+def split_feature(
+    feature,
+    num_splits=2,
+    channel_last=False,
+):
+    if channel_last:  # [B, H, W, C]
+        b, h, w, c = feature.size()
+        assert h % num_splits == 0 and w % num_splits == 0
+
+        b_new = b * num_splits * num_splits
+        h_new = h // num_splits
+        w_new = w // num_splits
+
+        feature = (
+            feature.view(b, num_splits, h // num_splits, num_splits, w // num_splits, c)
+            .permute(0, 1, 3, 2, 4, 5)
+            .reshape(b_new, h_new, w_new, c)
+        )  # [B*K*K, H/K, W/K, C]
+    else:  # [B, C, H, W]
+        b, c, h, w = feature.size()
+        assert h % num_splits == 0 and w % num_splits == 0
+
+        b_new = b * num_splits * num_splits
+        h_new = h // num_splits
+        w_new = w // num_splits
+
+        feature = (
+            feature.view(b, c, num_splits, h // num_splits, num_splits, w // num_splits)
+            .permute(0, 2, 4, 1, 3, 5)
+            .reshape(b_new, c, h_new, w_new)
+        )  # [B*K*K, C, H/K, W/K]
+
+    return feature
+
+
+def merge_splits(
+    splits,
+    num_splits=2,
+    channel_last=False,
+):
+    if channel_last:  # [B*K*K, H/K, W/K, C]
+        b, h, w, c = splits.size()
+        new_b = b // num_splits // num_splits
+
+        splits = splits.view(new_b, num_splits, num_splits, h, w, c)
+        merge = (
+            splits.permute(0, 1, 3, 2, 4, 5)
+            .contiguous()
+            .view(new_b, num_splits * h, num_splits * w, c)
+        )  # [B, H, W, C]
+    else:  # [B*K*K, C, H/K, W/K]
+        b, c, h, w = splits.size()
+        new_b = b // num_splits // num_splits
+
+        splits = splits.view(new_b, num_splits, num_splits, c, h, w)
+        merge = (
+            splits.permute(0, 3, 1, 4, 2, 5)
+            .contiguous()
+            .view(new_b, c, num_splits * h, num_splits * w)
+        )  # [B, C, H, W]
+
+    return merge
+
+
+def generate_shift_window_attn_mask(
+    input_resolution,
+    window_size_h,
+    window_size_w,
+    shift_size_h,
+    shift_size_w,
+    device=torch.device("cuda"),
+):
+    # ref: https://github.com/microsoft/Swin-Transformer/blob/main/models/swin_transformer.py
+    # calculate attention mask for SW-MSA
+    h, w = input_resolution
+    img_mask = torch.zeros((1, h, w, 1)).to(device)  # 1 H W 1
+    h_slices = (
+        slice(0, -window_size_h),
+        slice(-window_size_h, -shift_size_h),
+        slice(-shift_size_h, None),
+    )
+    w_slices = (
+        slice(0, -window_size_w),
+        slice(-window_size_w, -shift_size_w),
+        slice(-shift_size_w, None),
+    )
+    cnt = 0
+    for h in h_slices:
+        for w in w_slices:
+            img_mask[:, h, w, :] = cnt
+            cnt += 1
+
+    mask_windows = split_feature(
+        img_mask, num_splits=input_resolution[-1] // window_size_w, channel_last=True
+    )
+
+    mask_windows = mask_windows.view(-1, window_size_h * window_size_w)
+    attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+    attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(
+        attn_mask == 0, float(0.0)
+    )
+
+    return attn_mask
+
+
+def feature_add_position(feature0, feature1, attn_splits, feature_channels):
+    pos_enc = PositionEmbeddingSine(num_pos_feats=feature_channels // 2)
+
+    if attn_splits > 1:  # add position in splited window
+        feature0_splits = split_feature(feature0, num_splits=attn_splits)
+        feature1_splits = split_feature(feature1, num_splits=attn_splits)
+
+        position = pos_enc(feature0_splits)
+
+        feature0_splits = feature0_splits + position
+        feature1_splits = feature1_splits + position
+
+        feature0 = merge_splits(feature0_splits, num_splits=attn_splits)
+        feature1 = merge_splits(feature1_splits, num_splits=attn_splits)
+    else:
+        position = pos_enc(feature0)
+
+        feature0 = feature0 + position
+        feature1 = feature1 + position
+
+    return feature0, feature1
+
+
+def upsample_flow_with_mask(flow, up_mask, upsample_factor, is_depth=False):
+    # convex upsampling following raft
+
+    mask = up_mask
+    b, flow_channel, h, w = flow.shape
+    mask = mask.view(b, 1, 9, upsample_factor, upsample_factor, h, w)  # [B, 1, 9, K, K, H, W]
+    mask = torch.softmax(mask, dim=2)
+
+    multiplier = 1 if is_depth else upsample_factor
+    up_flow = F.unfold(multiplier * flow, [3, 3], padding=1)
+    up_flow = up_flow.view(b, flow_channel, 9, 1, 1, h, w)  # [B, 2, 9, 1, 1, H, W]
+
+    up_flow = torch.sum(mask * up_flow, dim=2)  # [B, 2, K, K, H, W]
+    up_flow = up_flow.permute(0, 1, 4, 2, 5, 3)  # [B, 2, K, H, K, W]
+    up_flow = up_flow.reshape(
+        b, flow_channel, upsample_factor * h, upsample_factor * w
+    )  # [B, 2, K*H, K*W]
+
+    return up_flow
+
+
+def split_feature_1d(
+    feature,
+    num_splits=2,
+):
+    # feature: [B, W, C]
+    b, w, c = feature.size()
+    assert w % num_splits == 0
+
+    b_new = b * num_splits
+    w_new = w // num_splits
+
+    feature = feature.view(b, num_splits, w // num_splits, c).view(
+        b_new, w_new, c
+    )  # [B*K, W/K, C]
+
+    return feature
+
+
+def merge_splits_1d(
+    splits,
+    h,
+    num_splits=2,
+):
+    b, w, c = splits.size()
+    new_b = b // num_splits // h
+
+    splits = splits.view(new_b, h, num_splits, w, c)
+    merge = splits.view(new_b, h, num_splits * w, c)  # [B, H, W, C]
+
+    return merge
+
+
+def window_partition_1d(x, window_size_w):
+    """
+    Args:
+        x: (B, W, C)
+        window_size (int): window size
+
+    Returns:
+        windows: (num_windows*B, window_size, C)
+    """
+    B, W, C = x.shape
+    x = x.view(B, W // window_size_w, window_size_w, C).view(-1, window_size_w, C)
+    return x
+
+
+def generate_shift_window_attn_mask_1d(
+    input_w, window_size_w, shift_size_w, device=torch.device("cuda")
+):
+    # calculate attention mask for SW-MSA
+    img_mask = torch.zeros((1, input_w, 1)).to(device)  # 1 W 1
+    w_slices = (
+        slice(0, -window_size_w),
+        slice(-window_size_w, -shift_size_w),
+        slice(-shift_size_w, None),
+    )
+    cnt = 0
+    for w in w_slices:
+        img_mask[:, w, :] = cnt
+        cnt += 1
+
+    mask_windows = window_partition_1d(img_mask, window_size_w)  # nW, window_size, 1
+    mask_windows = mask_windows.view(-1, window_size_w)
+    attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(
+        2
+    )  # nW, window_size, window_size
+    attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(
+        attn_mask == 0, float(0.0)
+    )
+
+    return attn_mask