91 lines
3.5 KiB
Python
91 lines
3.5 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.
|
|
|
|
import warnings
|
|
|
|
import numpy as np
|
|
import torch
|
|
from PIL import Image
|
|
|
|
|
|
def get_sdpa_settings():
|
|
if torch.cuda.is_available():
|
|
old_gpu = torch.cuda.get_device_properties(0).major < 7
|
|
# only use Flash Attention on Ampere (8.0) or newer GPUs
|
|
use_flash_attn = torch.cuda.get_device_properties(0).major >= 8
|
|
if not use_flash_attn:
|
|
warnings.warn(
|
|
"Flash Attention is disabled as it requires a GPU with Ampere (8.0) CUDA capability.",
|
|
category=UserWarning,
|
|
stacklevel=2,
|
|
)
|
|
# keep math kernel for PyTorch versions before 2.2 (Flash Attention v2 is only
|
|
# available on PyTorch 2.2+, while Flash Attention v1 cannot handle all cases)
|
|
pytorch_version = tuple(int(v) for v in torch.__version__.split(".")[:2])
|
|
if pytorch_version < (2, 2):
|
|
warnings.warn(
|
|
f"You are using PyTorch {torch.__version__} without Flash Attention v2 support. "
|
|
"Consider upgrading to PyTorch 2.2+ for Flash Attention v2 (which could be faster).",
|
|
category=UserWarning,
|
|
stacklevel=2,
|
|
)
|
|
math_kernel_on = pytorch_version < (2, 2) or not use_flash_attn
|
|
else:
|
|
old_gpu = True
|
|
use_flash_attn = False
|
|
math_kernel_on = True
|
|
|
|
return old_gpu, use_flash_attn, math_kernel_on
|
|
|
|
|
|
def mask_to_box(masks: torch.Tensor):
|
|
"""
|
|
compute bounding box given an input mask
|
|
|
|
Inputs:
|
|
- masks: [B, 1, H, W] boxes, dtype=torch.Tensor
|
|
|
|
Returns:
|
|
- box_coords: [B, 1, 4], contains (x, y) coordinates of top left and bottom right box corners, dtype=torch.Tensor
|
|
"""
|
|
B, _, h, w = masks.shape
|
|
device = masks.device
|
|
xs = torch.arange(w, device=device, dtype=torch.int32)
|
|
ys = torch.arange(h, device=device, dtype=torch.int32)
|
|
grid_xs, grid_ys = torch.meshgrid(xs, ys, indexing="xy")
|
|
grid_xs = grid_xs[None, None, ...].expand(B, 1, h, w)
|
|
grid_ys = grid_ys[None, None, ...].expand(B, 1, h, w)
|
|
min_xs, _ = torch.min(torch.where(masks, grid_xs, w).flatten(-2), dim=-1)
|
|
max_xs, _ = torch.max(torch.where(masks, grid_xs, -1).flatten(-2), dim=-1)
|
|
min_ys, _ = torch.min(torch.where(masks, grid_ys, h).flatten(-2), dim=-1)
|
|
max_ys, _ = torch.max(torch.where(masks, grid_ys, -1).flatten(-2), dim=-1)
|
|
bbox_coords = torch.stack((min_xs, min_ys, max_xs, max_ys), dim=-1)
|
|
|
|
return bbox_coords
|
|
|
|
|
|
def _load_img_as_tensor(img_path, image_size):
|
|
img_pil = Image.open(img_path)
|
|
img_np = np.array(img_pil.convert("RGB").resize((image_size, image_size)))
|
|
if img_np.dtype == np.uint8: # np.uint8 is expected for JPEG images
|
|
img_np = img_np / 255.0
|
|
else:
|
|
raise RuntimeError(f"Unknown image dtype: {img_np.dtype} on {img_path}")
|
|
img = torch.from_numpy(img_np).permute(2, 0, 1)
|
|
video_width, video_height = img_pil.size # the original video size
|
|
return img, video_height, video_width
|
|
|
|
|
|
def concat_points(old_point_inputs, new_points, new_labels):
|
|
"""Add new points and labels to previous point inputs (add at the end)."""
|
|
if old_point_inputs is None:
|
|
points, labels = new_points, new_labels
|
|
else:
|
|
points = torch.cat([old_point_inputs["point_coords"], new_points], dim=1)
|
|
labels = torch.cat([old_point_inputs["point_labels"], new_labels], dim=1)
|
|
|
|
return {"point_coords": points, "point_labels": labels}
|