ColamanAI's picture
Upload 169 files
b74998d verified
# Copyright (c) Meta Platforms, Inc. and affiliates.
# This source code is licensed under the Apache License, Version 2.0
# found in the LICENSE file in the root directory of this source tree.
"""
Inference wrapper for MUSt3R
"""
import datetime
import os
import numpy as np
import torch
from dust3r.viz import rgb
from must3r.demo.inference import SceneState
from must3r.engine.inference import inference_multi_ar, postprocess
from must3r.model import get_pointmaps_activation, load_model
from mapanything.models.external.vggt.utils.rotation import mat_to_quat
def must3r_inference(
views,
filelist,
model,
retrieval,
device,
amp,
num_mem_images,
max_bs,
init_num_images=2,
batch_num_views=1,
render_once=False,
is_sequence=False,
viser_server=None,
num_refinements_iterations=2,
verbose=True,
):
if amp == "fp16":
dtype = torch.float16
elif amp == "bf16":
assert torch.cuda.is_bf16_supported()
dtype = torch.bfloat16
else:
assert not amp
dtype = torch.float32
max_bs = None if max_bs == 0 else max_bs
encoder, decoder = model
pointmaps_activation = get_pointmaps_activation(decoder, verbose=verbose)
def post_process_function(x):
return postprocess(
x, pointmaps_activation=pointmaps_activation, compute_cam=True
)
if verbose:
print("loading images")
time_start = datetime.datetime.now()
nimgs = len(views)
ellapsed = datetime.datetime.now() - time_start
if verbose:
print(f"loaded in {ellapsed}")
print("running inference")
time_start = datetime.datetime.now()
if viser_server is not None:
viser_server.reset(nimgs)
imgs = [b["img"].to("cpu") for b in views]
true_shape = [torch.from_numpy(b["true_shape"]).to("cpu") for b in views]
true_shape = torch.stack(true_shape, dim=0)
nimgs = true_shape.shape[0]
# Use all images as keyframes
keyframes = np.linspace(0, len(imgs) - 1, num_mem_images, dtype=int).tolist()
encoder_precomputed_features = None
not_keyframes = sorted(set(range(nimgs)).difference(set(keyframes)))
assert (len(keyframes) + len(not_keyframes)) == nimgs
# reorder images
views = [views[i] for i in keyframes] + [views[i] for i in not_keyframes]
imgs = [b["img"].to(device) for b in views]
true_shape = [torch.from_numpy(b["true_shape"]).to(device) for b in views]
filenames = [filelist[i] for i in keyframes + not_keyframes]
img_ids = [torch.tensor(v) for v in keyframes + not_keyframes]
if encoder_precomputed_features is not None:
x_start, pos_start = encoder_precomputed_features
x = [x_start[i] for i in keyframes] + [x_start[i] for i in not_keyframes]
pos = [pos_start[i] for i in keyframes] + [pos_start[i] for i in not_keyframes]
encoder_precomputed_features = (x, pos)
mem_batches = [init_num_images]
while (sum_b := sum(mem_batches)) != max(num_mem_images, init_num_images):
size_b = min(batch_num_views, num_mem_images - sum_b)
mem_batches.append(size_b)
if render_once:
to_render = list(range(num_mem_images, nimgs))
else:
to_render = None
with torch.autocast("cuda", dtype=dtype):
x_out_0, x_out = inference_multi_ar(
encoder,
decoder,
imgs,
img_ids,
true_shape,
mem_batches,
max_bs=max_bs,
verbose=verbose,
to_render=to_render,
encoder_precomputed_features=encoder_precomputed_features,
device=device,
preserve_gpu_mem=True,
post_process_function=post_process_function,
viser_server=viser_server,
num_refinements_iterations=num_refinements_iterations,
)
if to_render is not None:
x_out = x_out_0 + x_out
ellapsed = datetime.datetime.now() - time_start
if verbose:
print(f"inference in {ellapsed}")
try:
print(str(int(torch.cuda.max_memory_reserved(device) / (1024**2))) + " MB")
except Exception:
pass
if viser_server is not None:
viser_server.reset_cam_visility()
viser_server.send_message("Finished")
if verbose:
print("preparing pointcloud")
time_start = datetime.datetime.now()
focals = []
cams2world = []
for i in range(nimgs):
focals.append(float(x_out[i]["focal"].cpu()))
cams2world.append(x_out[i]["c2w"].cpu())
# x_out to cpu
for i in range(len(x_out)):
for k in x_out[i].keys():
x_out[i][k] = x_out[i][k].cpu()
rgbimg = [rgb(imgs[i], true_shape[i]) for i in range(nimgs)]
scene = SceneState(x_out, rgbimg, true_shape, focals, cams2world, filenames)
ellapsed = datetime.datetime.now() - time_start
if verbose:
print(f"pointcloud prepared in {ellapsed}")
return scene
class MUSt3RWrapper(torch.nn.Module):
def __init__(
self,
name,
ckpt_path,
retrieval_ckpt_path,
img_size=512,
amp="bf16",
max_bs=1,
**kwargs,
):
super().__init__()
self.name = name
self.ckpt_path = ckpt_path
self.retrieval_ckpt_path = retrieval_ckpt_path
self.amp = amp
self.max_bs = max_bs
# Init the model and load the checkpoint
self.model = load_model(self.ckpt_path, img_size=512)
def forward(self, views):
"""
Forward pass wrapper for MUSt3R.
Assumption:
- The batch size of input views is 1.
Args:
views (List[dict]): List of dictionaries containing the input views' images and instance information.
Each dictionary should contain the following keys, where B is the batch size and is 1:
"img" (tensor): Image tensor of shape (B, C, H, W).
"data_norm_type" (list): ["dust3r"]
"label" (list): ["scene_name"]
"instance" (list): ["image_name"]
Returns:
List[dict]: A list containing the final outputs for the input views.
"""
# Check the batch size of input views
batch_size_per_view, _, height, width = views[0]["img"].shape
device = views[0]["img"].device
num_views = len(views)
assert batch_size_per_view == 1, (
f"Batch size of input views should be 1, but got {batch_size_per_view}."
)
# Check the data norm type
data_norm_type = views[0]["data_norm_type"][0]
assert data_norm_type == "dust3r", (
"MUSt3R expects a normalized image with the DUSt3R normalization scheme applied"
)
# Convert the input views to the expected input format
images = []
image_paths = []
for view in views:
images.append(
dict(
img=view["img"][0].cpu(),
idx=len(images),
instance=str(len(images)),
true_shape=np.int32([view["img"].shape[-2], view["img"].shape[-1]]),
)
)
view_name = os.path.join(view["label"][0], view["instance"][0])
image_paths.append(view_name)
# Run MUSt3R inference
scene = must3r_inference(
images,
image_paths,
self.model,
self.retrieval_ckpt_path,
device,
self.amp,
num_views,
self.max_bs,
verbose=False,
)
# Make sure scene is not None
if scene is None:
raise RuntimeError("MUSt3R failed.")
# Get the predictions
predictions = scene.x_out
# Convert the output to the MapAnything format
with torch.autocast("cuda", enabled=False):
res = []
for view_idx in range(num_views):
# Get the current view predictions
curr_view_prediction = predictions[view_idx]
curr_view_conf = curr_view_prediction["conf"]
curr_view_pose = curr_view_prediction["c2w"].unsqueeze(0)
# Convert the pose to quaternions and translation
curr_view_cam_translations = curr_view_pose[..., :3, 3]
curr_view_cam_quats = mat_to_quat(curr_view_pose[..., :3, :3])
# Get the camera frame pointmaps
curr_view_pts3d_cam = curr_view_prediction["pts3d_local"].unsqueeze(0)
# Get the depth along ray and ray directions
curr_view_depth_along_ray = torch.norm(
curr_view_pts3d_cam, dim=-1, keepdim=True
)
curr_view_ray_dirs = curr_view_pts3d_cam / curr_view_depth_along_ray
# Get the pointmaps
curr_view_pts3d = curr_view_prediction["pts3d"].unsqueeze(0)
# Append the outputs to the result list
res.append(
{
"pts3d": curr_view_pts3d.to(device),
"pts3d_cam": curr_view_pts3d_cam.to(device),
"ray_directions": curr_view_ray_dirs.to(device),
"depth_along_ray": curr_view_depth_along_ray.to(device),
"cam_trans": curr_view_cam_translations.to(device),
"cam_quats": curr_view_cam_quats.to(device),
"conf": curr_view_conf.to(device),
}
)
return res