SAM 2 Update 12/11/2024 -- full model compilation for a major VOS speedup and a new SAM2VideoPredictor to better handle multi-object tracking (#486)

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.
This commit is contained in:
Ronghang Hu
2024-12-11 15:00:55 -08:00
committed by GitHub
parent c2ec8e14a1
commit 393ae336a7
27 changed files with 1794 additions and 443 deletions

View File

@@ -25,6 +25,11 @@ class PositionEmbeddingSine(nn.Module):
temperature: int = 10000,
normalize: bool = True,
scale: Optional[float] = None,
# Following settings only relevant
# for warmping up cache for compilation
warmup_cache: bool = True,
image_size: int = 1024,
strides: Tuple[int] = (4, 8, 16, 32),
):
super().__init__()
assert num_pos_feats % 2 == 0, "Expecting even model width"
@@ -38,6 +43,12 @@ class PositionEmbeddingSine(nn.Module):
self.scale = scale
self.cache = {}
if warmup_cache and torch.cuda.is_available():
# Warmup cache for cuda, to help with compilation
device = torch.device("cuda")
for stride in strides:
cache_key = (image_size // stride, image_size // stride)
self._pe(1, device, *cache_key)
def _encode_xy(self, x, y):
# The positions are expected to be normalized
@@ -76,19 +87,20 @@ class PositionEmbeddingSine(nn.Module):
return pos
@torch.no_grad()
def forward(self, x: torch.Tensor):
cache_key = (x.shape[-2], x.shape[-1])
def _pe(self, B, device, *cache_key):
H, W = cache_key
if cache_key in self.cache:
return self.cache[cache_key][None].repeat(x.shape[0], 1, 1, 1)
return self.cache[cache_key].to(device)[None].repeat(B, 1, 1, 1)
y_embed = (
torch.arange(1, x.shape[-2] + 1, dtype=torch.float32, device=x.device)
torch.arange(1, H + 1, dtype=torch.float32, device=device)
.view(1, -1, 1)
.repeat(x.shape[0], 1, x.shape[-1])
.repeat(B, 1, W)
)
x_embed = (
torch.arange(1, x.shape[-1] + 1, dtype=torch.float32, device=x.device)
torch.arange(1, W + 1, dtype=torch.float32, device=device)
.view(1, 1, -1)
.repeat(x.shape[0], x.shape[-2], 1)
.repeat(B, H, 1)
)
if self.normalize:
@@ -96,7 +108,7 @@ class PositionEmbeddingSine(nn.Module):
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 = torch.arange(self.num_pos_feats, dtype=torch.float32, device=device)
dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
pos_x = x_embed[:, :, :, None] / dim_t
@@ -111,6 +123,12 @@ class PositionEmbeddingSine(nn.Module):
self.cache[cache_key] = pos[0]
return pos
@torch.no_grad()
def forward(self, x: torch.Tensor):
B = x.shape[0]
cache_key = (x.shape[-2], x.shape[-1])
return self._pe(B, x.device, *cache_key)
class PositionEmbeddingRandom(nn.Module):
"""