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.
If the user has "sam2/sam2" in their path, they are likey importing the repo itself as "sam2" rather than importing the "sam2" python package (i.e. "sam2/sam2" directory). This typically happens because the user is running Python from the parent directory that contains the sam2 repo they cloned.
In general, the user should not run Python from the parent dir when the repo is cloned into (same is true for e.g. Numpy repo that contains names like `numpy/numpy` where the module and the repo have the same name), as the user encountered in https://github.com/facebookresearch/sam2/issues/346.
(close https://github.com/facebookresearch/sam2/issues/346)
sam2/build_sam.py:81:14: TOR102 [*] `torch.load` without `weights_only` parameter is unsafe. Explicitly set `weights_only` to False only if you trust the data you load and full pickle functionality is needed, otherwise set `weights_only=True`.
Found with https://github.com/pytorch-labs/torchfix/