Spaces:
Running
on
Zero
Running
on
Zero
File size: 8,065 Bytes
178f950 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 |
import os
import copy
import sys
import json
import importlib
import argparse
import torch
import torch.nn.functional as F
import numpy as np
import pandas as pd
import utils3d
from tqdm import tqdm
from easydict import EasyDict as edict
from concurrent.futures import ThreadPoolExecutor
from queue import Queue
from torchvision import transforms
from PIL import Image
torch.set_grad_enabled(False)
def get_data(frames, sha256):
with ThreadPoolExecutor(max_workers=16) as executor:
def worker(view):
image_path = os.path.join(opt.output_dir, 'renders', sha256, view['file_path'])
try:
image = Image.open(image_path)
except:
print(f"Error loading image {image_path}")
return None
image = image.resize((518, 518), Image.Resampling.LANCZOS)
image = np.array(image).astype(np.float32) / 255
image = image[:, :, :3] * image[:, :, 3:]
image = torch.from_numpy(image).permute(2, 0, 1).float()
c2w = torch.tensor(view['transform_matrix'])
c2w[:3, 1:3] *= -1
extrinsics = torch.inverse(c2w)
fov = view['camera_angle_x']
intrinsics = utils3d.torch.intrinsics_from_fov_xy(torch.tensor(fov), torch.tensor(fov))
return {
'image': image,
'extrinsics': extrinsics,
'intrinsics': intrinsics
}
datas = executor.map(worker, frames)
for data in datas:
if data is not None:
yield data
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--output_dir', type=str, required=True,
help='Directory to save the metadata')
parser.add_argument('--filter_low_aesthetic_score', type=float, default=None,
help='Filter objects with aesthetic score lower than this value')
parser.add_argument('--model', type=str, default='dinov2_vitl14_reg',
help='Feature extraction model')
parser.add_argument('--instances', type=str, default=None,
help='Instances to process')
parser.add_argument('--batch_size', type=int, default=16)
parser.add_argument('--rank', type=int, default=0)
parser.add_argument('--world_size', type=int, default=1)
opt = parser.parse_args()
opt = edict(vars(opt))
feature_name = opt.model
os.makedirs(os.path.join(opt.output_dir, 'features', feature_name), exist_ok=True)
# load model
dinov2_model = torch.hub.load('facebookresearch/dinov2', opt.model)
dinov2_model.eval().cuda()
transform = transforms.Compose([
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
n_patch = 518 // 14
# get file list
if os.path.exists(os.path.join(opt.output_dir, 'metadata.csv')):
metadata = pd.read_csv(os.path.join(opt.output_dir, 'metadata.csv'))
else:
raise ValueError('metadata.csv not found')
if opt.instances is not None:
with open(opt.instances, 'r') as f:
instances = f.read().splitlines()
metadata = metadata[metadata['sha256'].isin(instances)]
else:
if opt.filter_low_aesthetic_score is not None:
metadata = metadata[metadata['aesthetic_score'] >= opt.filter_low_aesthetic_score]
if f'feature_{feature_name}' in metadata.columns:
metadata = metadata[metadata[f'feature_{feature_name}'] == False]
metadata = metadata[metadata['voxelized'] == True]
metadata = metadata[metadata['rendered'] == True]
start = len(metadata) * opt.rank // opt.world_size
end = len(metadata) * (opt.rank + 1) // opt.world_size
metadata = metadata[start:end]
records = []
# filter out objects that are already processed
sha256s = list(metadata['sha256'].values)
for sha256 in copy.copy(sha256s):
if os.path.exists(os.path.join(opt.output_dir, 'features', feature_name, f'{sha256}.npz')):
records.append({'sha256': sha256, f'feature_{feature_name}' : True})
sha256s.remove(sha256)
# extract features
load_queue = Queue(maxsize=4)
try:
with ThreadPoolExecutor(max_workers=8) as loader_executor, \
ThreadPoolExecutor(max_workers=8) as saver_executor:
def loader(sha256):
try:
with open(os.path.join(opt.output_dir, 'renders', sha256, 'transforms.json'), 'r') as f:
metadata = json.load(f)
frames = metadata['frames']
data = []
for datum in get_data(frames, sha256):
datum['image'] = transform(datum['image'])
data.append(datum)
positions = utils3d.io.read_ply(os.path.join(opt.output_dir, 'voxels', f'{sha256}.ply'))[0]
load_queue.put((sha256, data, positions))
except Exception as e:
print(f"Error loading data for {sha256}: {e}")
loader_executor.map(loader, sha256s)
def saver(sha256, pack, patchtokens, uv):
pack['patchtokens'] = F.grid_sample(
patchtokens,
uv.unsqueeze(1),
mode='bilinear',
align_corners=False,
).squeeze(2).permute(0, 2, 1).cpu().numpy()
pack['patchtokens'] = np.mean(pack['patchtokens'], axis=0).astype(np.float16)
save_path = os.path.join(opt.output_dir, 'features', feature_name, f'{sha256}.npz')
np.savez_compressed(save_path, **pack)
records.append({'sha256': sha256, f'feature_{feature_name}' : True})
for _ in tqdm(range(len(sha256s)), desc="Extracting features"):
sha256, data, positions = load_queue.get()
positions = torch.from_numpy(positions).float().cuda()
indices = ((positions + 0.5) * 64).long()
assert torch.all(indices >= 0) and torch.all(indices < 64), "Some vertices are out of bounds"
n_views = len(data)
N = positions.shape[0]
pack = {
'indices': indices.cpu().numpy().astype(np.uint8),
}
patchtokens_lst = []
uv_lst = []
for i in range(0, n_views, opt.batch_size):
batch_data = data[i:i+opt.batch_size]
bs = len(batch_data)
batch_images = torch.stack([d['image'] for d in batch_data]).cuda()
batch_extrinsics = torch.stack([d['extrinsics'] for d in batch_data]).cuda()
batch_intrinsics = torch.stack([d['intrinsics'] for d in batch_data]).cuda()
features = dinov2_model(batch_images, is_training=True)
uv = utils3d.torch.project_cv(positions, batch_extrinsics, batch_intrinsics)[0] * 2 - 1
patchtokens = features['x_prenorm'][:, dinov2_model.num_register_tokens + 1:].permute(0, 2, 1).reshape(bs, 1024, n_patch, n_patch)
patchtokens_lst.append(patchtokens)
uv_lst.append(uv)
patchtokens = torch.cat(patchtokens_lst, dim=0)
uv = torch.cat(uv_lst, dim=0)
# save features
saver_executor.submit(saver, sha256, pack, patchtokens, uv)
saver_executor.shutdown(wait=True)
except:
print("Error happened during processing.")
records = pd.DataFrame.from_records(records)
records.to_csv(os.path.join(opt.output_dir, f'feature_{feature_name}_{opt.rank}.csv'), index=False)
|