393ae336a7
This PR provides new features and updates for SAM 2: - We now support `torch.compile` of the entire SAM 2 model on videos, which can be turned on by setting `vos_optimized=True` in `build_sam2_video_predictor` (it uses the new `SAM2VideoPredictorVOS` predictor class in `sam2/sam2_video_predictor.py`). * Compared to the previous setting (which only compiles the image encoder backbone), the new full model compilation gives a major speedup in inference FPS. * In the VOS prediction script `tools/vos_inference.py`, you can specify this option in `tools/vos_inference.py` via the `--use_vos_optimized_video_predictor` flag. * Note that turning on this flag might introduce a small variance in the predictions due to numerical differences caused by `torch.compile` of the full model. * **PyTorch 2.5.1 is the minimum version for full support of this feature**. (Earlier PyTorch versions might run into compilation errors in some cases.) Therefore, we have updated the minimum PyTorch version to 2.5.1 accordingly in the installation scripts. - We also update the implementation of the `SAM2VideoPredictor` class for the SAM 2 video prediction in `sam2/sam2_video_predictor.py`, which allows for independent per-object inference. Specifically, in the new `SAM2VideoPredictor`: * Now **we handle the inference of each object independently** (as if we are opening a separate session for each object) while sharing their backbone features. * This change allows us to relax the assumption of prompting for multi-object tracking. Previously (due to the batching behavior in inference), if a video frame receives clicks for only a subset of objects, the rest of the (non-prompted) objects are assumed to be non-existent in this frame (i.e., in such frames, the user is telling SAM 2 that the rest of the objects don't appear). Now, if a frame receives clicks for only a subset of objects, we do not make any assumptions about the remaining (non-prompted) objects (i.e., now each object is handled independently and is not affected by how other objects are prompted). As a result, **we allow adding new objects after tracking starts** after this change (which was previously a restriction on usage). * We believe that the new version is a more natural inference behavior and therefore switched to it as the default behavior. The previous implementation of `SAM2VideoPredictor` is backed up to in `sam2/sam2_video_predictor_legacy.py`. All the VOS inference results using `tools/vos_inference.py` should remain the same after this change to the `SAM2VideoPredictor` class.
94 lines
2.9 KiB
Python
94 lines
2.9 KiB
Python
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
|
# All rights reserved.
|
|
|
|
# This source code is licensed under the license found in the
|
|
# LICENSE file in the root directory of this source tree.
|
|
|
|
"""Some utilities for backbones, in particular for windowing"""
|
|
|
|
from typing import Tuple
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
|
|
|
|
def window_partition(x, window_size):
|
|
"""
|
|
Partition into non-overlapping windows with padding if needed.
|
|
Args:
|
|
x (tensor): input tokens with [B, H, W, C].
|
|
window_size (int): window size.
|
|
Returns:
|
|
windows: windows after partition with [B * num_windows, window_size, window_size, C].
|
|
(Hp, Wp): padded height and width before partition
|
|
"""
|
|
B, H, W, C = x.shape
|
|
|
|
pad_h = (window_size - H % window_size) % window_size
|
|
pad_w = (window_size - W % window_size) % window_size
|
|
if pad_h > 0 or pad_w > 0:
|
|
x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
|
|
Hp, Wp = H + pad_h, W + pad_w
|
|
|
|
x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
|
|
windows = x.permute(0, 1, 3, 2, 4, 5).reshape(-1, window_size, window_size, C)
|
|
return windows, (Hp, Wp)
|
|
|
|
|
|
def window_unpartition(windows, window_size, pad_hw, hw):
|
|
"""
|
|
Window unpartition into original sequences and removing padding.
|
|
Args:
|
|
x (tensor): input tokens with [B * num_windows, window_size, window_size, C].
|
|
window_size (int): window size.
|
|
pad_hw (Tuple): padded height and width (Hp, Wp).
|
|
hw (Tuple): original height and width (H, W) before padding.
|
|
Returns:
|
|
x: unpartitioned sequences with [B, H, W, C].
|
|
"""
|
|
Hp, Wp = pad_hw
|
|
H, W = hw
|
|
B = windows.shape[0] // (Hp * Wp // window_size // window_size)
|
|
x = windows.reshape(
|
|
B, Hp // window_size, Wp // window_size, window_size, window_size, -1
|
|
)
|
|
x = x.permute(0, 1, 3, 2, 4, 5).reshape(B, Hp, Wp, -1)
|
|
|
|
if Hp > H or Wp > W:
|
|
x = x[:, :H, :W, :]
|
|
return x
|
|
|
|
|
|
class PatchEmbed(nn.Module):
|
|
"""
|
|
Image to Patch Embedding.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
kernel_size: Tuple[int, ...] = (7, 7),
|
|
stride: Tuple[int, ...] = (4, 4),
|
|
padding: Tuple[int, ...] = (3, 3),
|
|
in_chans: int = 3,
|
|
embed_dim: int = 768,
|
|
):
|
|
"""
|
|
Args:
|
|
kernel_size (Tuple): kernel size of the projection layer.
|
|
stride (Tuple): stride of the projection layer.
|
|
padding (Tuple): padding size of the projection layer.
|
|
in_chans (int): Number of input image channels.
|
|
embed_dim (int): embed_dim (int): Patch embedding dimension.
|
|
"""
|
|
super().__init__()
|
|
self.proj = nn.Conv2d(
|
|
in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding
|
|
)
|
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
|
x = self.proj(x)
|
|
# B C H W -> B H W C
|
|
x = x.permute(0, 2, 3, 1)
|
|
return x
|