目录
- 前言
- run_nerf.py
* - config_parser()
- train()
- create_nerf()
- render()
- batchify_rays()
- render_rays()
- raw2outputs()
- render_path()
- run_nerf_helpers.py
* - class NeRF()
- get_rays_np()
- ndc_rays()
- load_llff.py
* - _load_data()
- _minify()
- load_llff_data()
- render_path_spiral()
前言
要想看懂instant-ngp的cuda代码，需要先对NeRF系列有足够深入的了解，原始的NeRF版本是基于tensorflow的，今天读的是MIT博士生Yen-Chen Lin实现的pytorch版本的代码。
代码链接：https://github.com/yenchenlin/nerf-pytorch
因为代码量比较大，所以我们先使用一个思维导图对项目逻辑进行梳理，然后逐个文件解析。为了保持思路连贯，我们会一次贴上整个函数的内容并逐行注释，然后贴相关的公式和示意图到代码段的下方。
; run_nerf.py
一切都从这个文件开始，让我们先来看看有哪些参数需要设置。
config_parser()
先是一些基本参数
parser.add_argument('--config', is_config_file=True,
help='config file path')
parser.add_argument("--expname", type=str,
help='experiment name')
parser.add_argument("--basedir", type=str, default='./logs/',
help='where to store ckpts and logs')
parser.add_argument("--datadir", type=str, default='./data/llff/fern',
help='input data directory')
然后是一些训练相关的参数
parser.add_argument("--netdepth", type=int, default=8,
help='layers in network')
parser.add_argument("--netwidth", type=int, default=256,
help='channels per layer')
parser.add_argument("--netdepth_fine", type=int, default=8,
help='layers in fine network')
parser.add_argument("--netwidth_fine", type=int, default=256,
help='channels per layer in fine network')
parser.add_argument("--N_rand", type=int, default=32*32*4,
help='batch size (number of random rays per gradient step)')
parser.add_argument("--lrate", type=float, default=5e-4,
help='learning rate')
parser.add_argument("--lrate_decay", type=int, default=250,
help='exponential learning rate decay (in 1000 steps)')
parser.add_argument("--chunk", type=int, default=1024*32,
help='number of rays processed in parallel, decrease if running out of memory')
parser.add_argument("--netchunk", type=int, default=1024*64,
help='number of pts sent through network in parallel, decrease if running out of memory')
parser.add_argument("--no_batching", action='store_true',
help='only take random rays from 1 image at a time')
parser.add_argument("--no_reload", action='store_true',
help='do not reload weights from saved ckpt')
parser.add_argument("--ft_path", type=str, default=None,
help='specific weights npy file to reload for coarse network')
然后是一些渲染时的参数
parser.add_argument("--N_samples", type=int, default=64,
help='number of coarse samples per ray')
parser.add_argument("--N_importance", type=int, default=0,
help='number of additional fine samples per ray')
parser.add_argument("--perturb", type=float, default=1.,
help='set to 0. for no jitter, 1. for jitter')
parser.add_argument("--use_viewdirs", action='store_true',
help='use full 5D input instead of 3D')
parser.add_argument("--i_embed", type=int, default=0,
help='set 0 for default positional encoding, -1 for none')
parser.add_argument("--multires", type=int, default=10,
help='log2 of max freq for positional encoding (3D location)')
parser.add_argument("--multires_views", type=int, default=4,
help='log2 of max freq for positional encoding (2D direction)')
parser.add_argument("--raw_noise_std", type=float, default=0.,
help='std dev of noise added to regularize sigma_a output, 1e0 recommended')
parser.add_argument("--render_only", action='store_true',
help='do not optimize, reload weights and render out render_poses path')
parser.add_argument("--render_test", action='store_true',
help='render the test set instead of render_poses path')
parser.add_argument("--render_factor", type=int, default=0,
help='downsampling factor to speed up rendering, set 4 or 8 for fast preview')
还有一些参数
parser.add_argument("--precrop_iters", type=int, default=0,
help='number of steps to train on central crops')
parser.add_argument("--precrop_frac", type=float,
default=.5, help='fraction of img taken for central crops')
parser.add_argument("--dataset_type", type=str, default='llff',
help='options: llff / blender / deepvoxels')
parser.add_argument("--testskip", type=int, default=8,
help='will load 1/N images from test/val sets, useful for large datasets like deepvoxels')
parser.add_argument("--shape", type=str, default='greek',
help='options : armchair / cube / greek / vase')
parser.add_argument("--white_bkgd", action='store_true',
help='set to render synthetic data on a white bkgd (always use for dvoxels)')
parser.add_argument("--half_res", action='store_true',
help='load blender synthetic data at 400x400 instead of 800x800')
parser.add_argument("--factor", type=int, default=8,
help='downsample factor for LLFF images')
parser.add_argument("--no_ndc", action='store_true',
help='do not use normalized device coordinates (set for non-forward facing scenes)')
parser.add_argument("--lindisp", action='store_true',
help='sampling linearly in disparity rather than depth')
parser.add_argument("--spherify", action='store_true',
help='set for spherical 360 scenes')
parser.add_argument("--llffhold", type=int, default=8,
help='will take every 1/N images as LLFF test set, paper uses 8')
parser.add_argument("--i_print", type=int, default=100,
help='frequency of console printout and metric loggin')
parser.add_argument("--i_img", type=int, default=500,
help='frequency of tensorboard image logging')
parser.add_argument("--i_weights", type=int, default=10000,
help='frequency of weight ckpt saving')
parser.add_argument("--i_testset", type=int, default=50000,
help='frequency of testset saving')
parser.add_argument("--i_video", type=int, default=50000,
help='frequency of render_poses video saving')
train()
训练过程的控制。开始训练，先把5D输入进行编码，然后交给MLP得到4D的数据（颜色和体素的密度），然后进行体渲染得到图片，再和真值计算L2 loss。
def train():
parser = config_parser()
args = parser.parse_args()
K = None
if args.dataset_type == 'llff':
images, poses, bds, render_poses, i_test = load_llff_data(args.datadir, args.factor,
recenter=True, bd_factor=.75,
spherify=args.spherify)
hwf = poses[0,:3,-1]
poses = poses[:,:3,:4]
print('Loaded llff', images.shape, render_poses.shape, hwf, args.datadir)
if not isinstance(i_test, list):
i_test = [i_test]
if args.llffhold > 0:
print('Auto LLFF holdout,', args.llffhold)
i_test = np.arange(images.shape[0])[::args.llffhold]
i_val = i_test
i_train = np.array([i for i in np.arange(int(images.shape[0])) if
(i not in i_test and i not in i_val)])
print('DEFINING BOUNDS')
if args.no_ndc:
near = np.ndarray.min(bds) * .9
far = np.ndarray.max(bds) * 1.
else:
near = 0.
far = 1.
print('NEAR FAR', near, far)
elif args.dataset_type == 'blender':
images, poses, render_poses, hwf, i_split = load_blender_data(args.datadir, args.half_res, args.testskip)
print('Loaded blender', images.shape, render_poses.shape, hwf, args.datadir)
i_train, i_val, i_test = i_split
near = 2.
far = 6.
if args.white_bkgd:
images = images[...,:3]*images[...,-1:] + (1.-images[...,-1:])
else:
images = images[...,:3]
elif args.dataset_type == 'LINEMOD':
images, poses, render_poses, hwf, K, i_split, near, far = load_LINEMOD_data(args.datadir, args.half_res, args.testskip)
print(f'Loaded LINEMOD, images shape: {images.shape}, hwf: {hwf}, K: {K}')
print(f'[CHECK HERE] near: {near}, far: {far}.')
i_train, i_val, i_test = i_split
if args.white_bkgd:
images = images[...,:3]*images[...,-1:] + (1.-images[...,-1:])
else:
images = images[...,:3]
elif args.dataset_type == 'deepvoxels':
images, poses, render_poses, hwf, i_split = load_dv_data(scene=args.shape,
basedir=args.datadir,
testskip=args.testskip)
print('Loaded deepvoxels', images.shape, render_poses.shape, hwf, args.datadir)
i_train, i_val, i_test = i_split
hemi_R = np.mean(np.linalg.norm(poses[:,:3,-1], axis=-1))
near = hemi_R-1.
far = hemi_R+1.
else:
print('Unknown dataset type', args.dataset_type, 'exiting')
return
H, W, focal = hwf
H, W = int(H), int(W)
hwf = [H, W, focal]
if K is None:
K = np.array([
[focal, 0, 0.5*W],
[0, focal, 0.5*H],
[0, 0, 1]
])
if args.render_test:
render_poses = np.array(poses[i_test])
basedir = args.basedir
expname = args.expname
os.makedirs(os.path.join(basedir, expname), exist_ok=True)
f = os.path.join(basedir, expname, 'args.txt')
with open(f, 'w') as file:
for arg in sorted(vars(args)):
attr = getattr(args, arg)
file.write('{} = {}\n'.format(arg, attr))
if args.config is not None:
f = os.path.join(basedir, expname, 'config.txt')
with open(f, 'w') as file:
file.write(open(args.config, 'r').read())
render_kwargs_train, render_kwargs_test, start, grad_vars, optimizer = create_nerf(args)
global_step = start
bds_dict = {
'near' : near,
'far' : far,
}
render_kwargs_train.update(bds_dict)
render_kwargs_test.update(bds_dict)
render_poses = torch.Tensor(render_poses).to(device)
if args.render_only:
print('RENDER ONLY')
with torch.no_grad():
if args.render_test:
images = images[i_test]
else:
images = None
testsavedir = os.path.join(basedir, expname, 'renderonly_{}_{:06d}'.format('test' if args.render_test else 'path', start))
os.makedirs(testsavedir, exist_ok=True)
print('test poses shape', render_poses.shape)
rgbs, _ = render_path(render_poses, hwf, K, args.chunk, render_kwargs_test, gt_imgs=images, savedir=testsavedir, render_factor=args.render_factor)
print('Done rendering', testsavedir)
imageio.mimwrite(os.path.join(testsavedir, 'video.mp4'), to8b(rgbs), fps=30, quality=8)
return
N_rand = args.N_rand
use_batching = not args.no_batching
if use_batching:
print('get rays')
rays = np.stack([get_rays_np(H, W, K, p) for p in poses[:,:3,:4]], 0)
print('done, concats')
rays_rgb = np.concatenate([rays, images[:,None]], 1)
rays_rgb = np.transpose(rays_rgb, [0,2,3,1,4])
rays_rgb = np.stack([rays_rgb[i] for i in i_train], 0)
rays_rgb = np.reshape(rays_rgb, [-1,3,3])
rays_rgb = rays_rgb.astype(np.float32)
print('shuffle rays')
np.random.shuffle(rays_rgb)
print('done')
i_batch = 0
if use_batching:
images = torch.Tensor(images).to(device)
poses = torch.Tensor(poses).to(device)
if use_batching:
rays_rgb = torch.Tensor(rays_rgb).to(device)
N_iters = 200000 + 1
print('Begin')
print('TRAIN views are', i_train)
print('TEST views are', i_test)
print('VAL views are', i_val)
start = start + 1
for i in trange(start, N_iters):
time0 = time.time()
if use_batching:
batch = rays_rgb[i_batch:i_batch+N_rand]
batch = torch.transpose(batch, 0, 1)
batch_rays, target_s = batch[:2], batch[2]
i_batch += N_rand
if i_batch >= rays_rgb.shape[0]:
print("Shuffle data after an epoch!")
rand_idx = torch.randperm(rays_rgb.shape[0])
rays_rgb = rays_rgb[rand_idx]
i_batch = 0
else:
img_i = np.random.choice(i_train)
target = images[img_i]
target = torch.Tensor(target).to(device)
pose = poses[img_i, :3,:4]
if N_rand is not None:
rays_o, rays_d = get_rays(H, W, K, torch.Tensor(pose))
if i < args.precrop_iters:
dH = int(H//2 * args.precrop_frac)
dW = int(W//2 * args.precrop_frac)
coords = torch.stack(
torch.meshgrid(
torch.linspace(H//2 - dH, H//2 + dH - 1, 2*dH),
torch.linspace(W//2 - dW, W//2 + dW - 1, 2*dW)
), -1)
if i == start:
print(f"[Config] Center cropping of size {2*dH} x {2*dW} is enabled until iter {args.precrop_iters}")
else:
coords = torch.stack(torch.meshgrid(torch.linspace(0, H-1, H), torch.linspace(0, W-1, W)), -1)
coords = torch.reshape(coords, [-1,2])
select_inds = np.random.choice(coords.shape[0], size=[N_rand], replace=False)
select_coords = coords[select_inds].long()
rays_o = rays_o[select_coords[:, 0], select_coords[:, 1]]
rays_d = rays_d[select_coords[:, 0], select_coords[:, 1]]
batch_rays = torch.stack([rays_o, rays_d], 0)
target_s = target[select_coords[:, 0], select_coords[:, 1]]
rgb, disp, acc, extras = render(H, W, K, chunk=args.chunk, rays=batch_rays,
verbose=i < 10, retraw=True,
**render_kwargs_train)
optimizer.zero_grad()
img_loss = img2mse(rgb, target_s)
trans = extras['raw'][...,-1]
loss = img_loss
psnr = mse2psnr(img_loss)
if 'rgb0' in extras:
img_loss0 = img2mse(extras['rgb0'], target_s)
loss = loss + img_loss0
psnr0 = mse2psnr(img_loss0)
loss.backward()
optimizer.step()
decay_rate = 0.1
decay_steps = args.lrate_decay * 1000
new_lrate = args.lrate * (decay_rate ** (global_step / decay_steps))
for param_group in optimizer.param_groups:
param_group['lr'] = new_lrate
dt = time.time()-time0
if i%args.i_weights==0:
path = os.path.join(basedir, expname, '{:06d}.tar'.format(i))
torch.save({
'global_step': global_step,
'network_fn_state_dict': render_kwargs_train['network_fn'].state_dict(),
'network_fine_state_dict': render_kwargs_train['network_fine'].state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
}, path)
print('Saved checkpoints at', path)
if i%args.i_video==0 and i > 0:
with torch.no_grad():
rgbs, disps = render_path(render_poses, hwf, K, args.chunk, render_kwargs_test)
print('Done, saving', rgbs.shape, disps.shape)
moviebase = os.path.join(basedir, expname, '{}_spiral_{:06d}_'.format(expname, i))
imageio.mimwrite(moviebase + 'rgb.mp4', to8b(rgbs), fps=30, quality=8)
imageio.mimwrite(moviebase + 'disp.mp4', to8b(disps / np.max(disps)), fps=30, quality=8)
if i%args.i_testset==0 and i > 0:
testsavedir = os.path.join(basedir, expname, 'testset_{:06d}'.format(i))
os.makedirs(testsavedir, exist_ok=True)
print('test poses shape', poses[i_test].shape)
with torch.no_grad():
render_path(torch.Tensor(poses[i_test]).to(device), hwf, K, args.chunk, render_kwargs_test, gt_imgs=images[i_test], savedir=testsavedir)
print('Saved test set')
if i%args.i_print==0:
tqdm.write(f"[TRAIN] Iter: {i} Loss: {loss.item()} PSNR: {psnr.item()}")
"""
print(expname, i, psnr.numpy(), loss.numpy(), global_step.numpy())
print('iter time {:.05f}'.format(dt))
with tf.contrib.summary.record_summaries_every_n_global_steps(args.i_print):
tf.contrib.summary.scalar('loss', loss)
tf.contrib.summary.scalar('psnr', psnr)
tf.contrib.summary.histogram('tran', trans)
if args.N_importance > 0:
tf.contrib.summary.scalar('psnr0', psnr0)
if i%args.i_img==0:
# Log a rendered validation view to Tensorboard
img_i=np.random.choice(i_val)
target = images[img_i]
pose = poses[img_i, :3,:4]
with torch.no_grad():
rgb, disp, acc, extras = render(H, W, focal, chunk=args.chunk, c2w=pose,
**render_kwargs_test)
psnr = mse2psnr(img2mse(rgb, target))
with tf.contrib.summary.record_summaries_every_n_global_steps(args.i_img):
tf.contrib.summary.image('rgb', to8b(rgb)[tf.newaxis])
tf.contrib.summary.image('disp', disp[tf.newaxis,...,tf.newaxis])
tf.contrib.summary.image('acc', acc[tf.newaxis,...,tf.newaxis])
tf.contrib.summary.scalar('psnr_holdout', psnr)
tf.contrib.summary.image('rgb_holdout', target[tf.newaxis])
if args.N_importance > 0:
with tf.contrib.summary.record_summaries_every_n_global_steps(args.i_img):
tf.contrib.summary.image('rgb0', to8b(extras['rgb0'])[tf.newaxis])
tf.contrib.summary.image('disp0', extras['disp0'][tf.newaxis,...,tf.newaxis])
tf.contrib.summary.image('z_std', extras['z_std'][tf.newaxis,...,tf.newaxis])
"""
global_step += 1
梳理完train，我们来重点看一下train当中调用过的几个函数
create_nerf()
先调用get_embedder获得一个对应的embedding函数，然后构建NeRF模型
def create_nerf(args):
"""Instantiate NeRF's MLP model.
"""
embed_fn, input_ch = get_embedder(args.multires, args.i_embed)
input_ch_views = 0
embeddirs_fn = None
if args.use_viewdirs:
embeddirs_fn, input_ch_views = get_embedder(args.multires_views, args.i_embed)
output_ch = 5 if args.N_importance > 0 else 4
skips = [4]
model = NeRF(D=args.netdepth, W=args.netwidth,
input_ch=input_ch, output_ch=output_ch, skips=skips,
input_ch_views=input_ch_views, use_viewdirs=args.use_viewdirs).to(device)
grad_vars = list(model.parameters())
model_fine = None
if args.N_importance > 0:
model_fine = NeRF(D=args.netdepth_fine, W=args.netwidth_fine,
input_ch=input_ch, output_ch=output_ch, skips=skips,
input_ch_views=input_ch_views, use_viewdirs=args.use_viewdirs).to(device)
grad_vars += list(model_fine.parameters())
network_query_fn = lambda inputs, viewdirs, network_fn : run_network(inputs, viewdirs, network_fn,
embed_fn=embed_fn,
embeddirs_fn=embeddirs_fn,
netchunk=args.netchunk)
optimizer = torch.optim.Adam(params=grad_vars, lr=args.lrate, betas=(0.9, 0.999))
start = 0
basedir = args.basedir
expname = args.expname
if args.ft_path is not None and args.ft_path!='None':
ckpts = [args.ft_path]
else:
ckpts = [os.path.join(basedir, expname, f) for f in sorted(os.listdir(os.path.join(basedir, expname))) if 'tar' in f]
print('Found ckpts', ckpts)
if len(ckpts) > 0 and not args.no_reload:
ckpt_path = ckpts[-1]
print('Reloading from', ckpt_path)
ckpt = torch.load(ckpt_path)
start = ckpt['global_step']
optimizer.load_state_dict(ckpt['optimizer_state_dict'])
model.load_state_dict(ckpt['network_fn_state_dict'])
if model_fine is not None:
model_fine.load_state_dict(ckpt['network_fine_state_dict'])
render_kwargs_train = {
'network_query_fn' : network_query_fn,
'perturb' : args.perturb,
'N_importance' : args.N_importance,
'network_fine' : model_fine,
'N_samples' : args.N_samples,
'network_fn' : model,
'use_viewdirs' : args.use_viewdirs,
'white_bkgd' : args.white_bkgd,
'raw_noise_std' : args.raw_noise_std,
}
if args.dataset_type != 'llff' or args.no_ndc:
print('Not ndc!')
render_kwargs_train['ndc'] = False
render_kwargs_train['lindisp'] = args.lindisp
render_kwargs_test = {k : render_kwargs_train[k] for k in render_kwargs_train}
render_kwargs_test['perturb'] = False
render_kwargs_test['raw_noise_std'] = 0.
return render_kwargs_train, render_kwargs_test, start, grad_vars, optimizer
render()
接下来我们看一下如何渲染，render函数返回的是光束对应的rgb图、视差图、不透明度，以及raw
def render(H, W, K, chunk=1024*32, rays=None, c2w=None, ndc=True,
near=0., far=1.,
use_viewdirs=False, c2w_staticcam=None,
**kwargs):
"""Render rays
Args:
H: int. Height of image in pixels.
W: int. Width of image in pixels.
focal: float. Focal length of pinhole camera.
chunk: int. Maximum number of rays to process simultaneously. Used to
control maximum memory usage. Does not affect final results.
rays: array of shape [2, batch_size, 3]. Ray origin and direction for
each example in batch.
c2w: array of shape [3, 4]. Camera-to-world transformation matrix.
ndc: bool. If True, represent ray origin, direction in NDC coordinates.
near: float or array of shape [batch_size]. Nearest distance for a ray.
far: float or array of shape [batch_size]. Farthest distance for a ray.
use_viewdirs: bool. If True, use viewing direction of a point in space in model.
c2w_staticcam: array of shape [3, 4]. If not None, use this transformation matrix for
camera while using other c2w argument for viewing directions.
Returns:
rgb_map: [batch_size, 3]. Predicted RGB values for rays.
disp_map: [batch_size]. Disparity map. Inverse of depth.
acc_map: [batch_size]. Accumulated opacity (alpha) along a ray.
extras: dict with everything returned by render_rays().
"""
if c2w is not None:
rays_o, rays_d = get_rays(H, W, K, c2w)
else:
rays_o, rays_d = rays
if use_viewdirs:
viewdirs = rays_d
if c2w_staticcam is not None:
rays_o, rays_d = get_rays(H, W, K, c2w_staticcam)
viewdirs = viewdirs / torch.norm(viewdirs, dim=-1, keepdim=True)
viewdirs = torch.reshape(viewdirs, [-1,3]).float()
sh = rays_d.shape
if ndc:
rays_o, rays_d = ndc_rays(H, W, K[0][0], 1., rays_o, rays_d)
rays_o = torch.reshape(rays_o, [-1,3]).float()
rays_d = torch.reshape(rays_d, [-1,3]).float()
near, far = near * torch.ones_like(rays_d[...,:1]), far * torch.ones_like(rays_d[...,:1])
rays = torch.cat([rays_o, rays_d, near, far], -1)
if use_viewdirs:
rays = torch.cat([rays, viewdirs], -1)
all_ret = batchify_rays(rays, chunk, **kwargs)
for k in all_ret:
k_sh = list(sh[:-1]) + list(all_ret[k].shape[1:])
all_ret[k] = torch.reshape(all_ret[k], k_sh)
k_extract = ['rgb_map', 'disp_map', 'acc_map']
ret_list = [all_ret[k] for k in k_extract]
ret_dict = {k : all_ret[k] for k in all_ret if k not in k_extract}
return ret_list + [ret_dict]
batchify_rays()
将光束作为一个batch，chunk是并行处理的光束数量，ret是一个chunk（1024×32=32768）的结果，all_ret是一个batch的结果
def batchify_rays(rays_flat, chunk=1024*32, **kwargs):
"""Render rays in smaller minibatches to avoid OOM.
"""
all_ret = {}
for i in range(0, rays_flat.shape[0], chunk):
ret = render_rays(rays_flat[i:i+chunk], **kwargs)
for k in ret:
if k not in all_ret:
all_ret[k] = []
all_ret[k].append(ret[k])
all_ret = {k : torch.cat(all_ret[k], 0) for k in all_ret}
return all_ret
render_rays()
def render_rays(ray_batch,
network_fn,
network_query_fn,
N_samples,
retraw=False,
lindisp=False,
perturb=0.,
N_importance=0,
network_fine=None,
white_bkgd=False,
raw_noise_std=0.,
verbose=False,
pytest=False):
"""Volumetric rendering.
Args:
ray_batch: array of shape [batch_size, ...]. All information necessary
for sampling along a ray, including: ray origin, ray direction, min
dist, max dist, and unit-magnitude viewing direction.
network_fn: function. Model for predicting RGB and density at each point
in space. 用于预测每个点的 RGB 和密度的模型
network_query_fn: function used for passing queries to network_fn.
N_samples: int. Number of different times to sample along each ray.每条射线上的采样次数
retraw: bool. If True, include model's raw, unprocessed predictions.
lindisp: bool. If True, sample linearly in inverse depth rather than in depth.
perturb: float, 0 or 1. If non-zero, each ray is sampled at stratified
random points in time.
N_importance: int. Number of additional times to sample along each ray.
These samples are only passed to network_fine.
network_fine: "fine" network with same spec as network_fn.
white_bkgd: bool. If True, assume a white background.
raw_noise_std: ...
verbose: bool. If True, print more debugging info.
Returns:
rgb_map: [num_rays, 3]. Estimated RGB color of a ray. Comes from fine model.
disp_map: [num_rays]. Disparity map. 1 / depth.
acc_map: [num_rays]. Accumulated opacity along each ray. Comes from fine model.
raw: [num_rays, num_samples, 4]. Raw predictions from model.
rgb0: See rgb_map. Output for coarse model.
disp0: See disp_map. Output for coarse model.
acc0: See acc_map. Output for coarse model.
z_std: [num_rays]. Standard deviation of distances along ray for each
sample.
"""
N_rays = ray_batch.shape[0]
rays_o, rays_d = ray_batch[:,0:3], ray_batch[:,3:6]
viewdirs = ray_batch[:,-3:] if ray_batch.shape[-1] > 8 else None
bounds = torch.reshape(ray_batch[...,6:8], [-1,1,2])
near, far = bounds[...,0], bounds[...,1]
t_vals = torch.linspace(0., 1., steps=N_samples)
if not lindisp:
z_vals = near * (1.-t_vals) + far * (t_vals)
else:
z_vals = 1./(1./near * (1.-t_vals) + 1./far * (t_vals))
z_vals = z_vals.expand([N_rays, N_samples])
if perturb > 0.:
mids = .5 * (z_vals[...,1:] + z_vals[...,:-1])
upper = torch.cat([mids, z_vals[...,-1:]], -1)
lower = torch.cat([z_vals[...,:1], mids], -1)
t_rand = torch.rand(z_vals.shape)
if pytest:
np.random.seed(0)
t_rand = np.random.rand(*list(z_vals.shape))
t_rand = torch.Tensor(t_rand)
z_vals = lower + (upper - lower) * t_rand
pts = rays_o[...,None,:] + rays_d[...,None,:] * z_vals[...,:,None]
raw = network_query_fn(pts, viewdirs, network_fn)
rgb_map, disp_map, acc_map, weights, depth_map = raw2outputs(raw, z_vals, rays_d, raw_noise_std, white_bkgd, pytest=pytest)
if N_importance > 0:
rgb_map_0, disp_map_0, acc_map_0 = rgb_map, disp_map, acc_map
z_vals_mid = .5 * (z_vals[...,1:] + z_vals[...,:-1])
z_samples = sample_pdf(z_vals_mid, weights[...,1:-1], N_importance, det=(perturb==0.), pytest=pytest)
z_samples = z_samples.detach()
z_vals, _ = torch.sort(torch.cat([z_vals, z_samples], -1), -1)
pts = rays_o[...,None,:] + rays_d[...,None,:] * z_vals[...,:,None]
run_fn = network_fn if network_fine is None else network_fine
raw = network_query_fn(pts, viewdirs, run_fn)
rgb_map, disp_map, acc_map, weights, depth_map = raw2outputs(raw, z_vals, rays_d, raw_noise_std, white_bkgd, pytest=pytest)
ret = {'rgb_map' : rgb_map, 'disp_map' : disp_map, 'acc_map' : acc_map}
if retraw:
ret['raw'] = raw
if N_importance > 0:
ret['rgb0'] = rgb_map_0
ret['disp0'] = disp_map_0
ret['acc0'] = acc_map_0
ret['z_std'] = torch.std(z_samples, dim=-1, unbiased=False)
for k in ret:
if (torch.isnan(ret[k]).any() or torch.isinf(ret[k]).any()) and DEBUG:
print(f"! [Numerical Error] {k} contains nan or inf.")
return ret
raw2outputs()
把模型的预测转化为有实际意义的表达，输入预测、时间和光束方向，输出光束颜色、视差、密度、每个采样点的权重和深度
def raw2outputs(raw, z_vals, rays_d, raw_noise_std=0, white_bkgd=False, pytest=False):
"""Transforms model's predictions to semantically meaningful values.
Args:
raw: [num_rays, num_samples along ray, 4]. Prediction from model.
z_vals: [num_rays, num_samples along ray]. Integration time.
rays_d: [num_rays, 3]. Direction of each ray.
Returns:
rgb_map: [num_rays, 3]. Estimated RGB color of a ray.
disp_map: [num_rays]. Disparity map. Inverse of depth map.
acc_map: [num_rays]. Sum of weights along each ray.
weights: [num_rays, num_samples]. Weights assigned to each sampled color.
depth_map: [num_rays]. Estimated distance to object.
"""
raw2alpha = lambda raw, dists, act_fn=F.relu: 1.-torch.exp(-act_fn(raw)*dists)
dists = z_vals[...,1:] - z_vals[...,:-1]
dists = torch.cat([dists, torch.Tensor([1e10]).expand(dists[...,:1].shape)], -1)
dists = dists * torch.norm(rays_d[...,None,:], dim=-1)
rgb = torch.sigmoid(raw[...,:3])
noise = 0.
if raw_noise_std > 0.:
noise = torch.randn(raw[...,3].shape) * raw_noise_std
if pytest:
np.random.seed(0)
noise = np.random.rand(*list(raw[...,3].shape)) * raw_noise_std
noise = torch.Tensor(noise)
alpha = raw2alpha(raw[...,3] + noise, dists)
weights = alpha * torch.cumprod(torch.cat([torch.ones((alpha.shape[0], 1)), 1.-alpha + 1e-10], -1), -1)[:, :-1]
rgb_map = torch.sum(weights[...,None] * rgb, -2)
depth_map = torch.sum(weights * z_vals, -1)
disp_map = 1./torch.max(1e-10 * torch.ones_like(depth_map), depth_map / torch.sum(weights, -1))
acc_map = torch.sum(weights, -1)
if white_bkgd:
rgb_map = rgb_map + (1.-acc_map[...,None])
return rgb_map, disp_map, acc_map, weights, depth_map
render_path()
根据pose等信息获得颜色和视差
def render_path(render_poses, hwf, K, chunk, render_kwargs, gt_imgs=None, savedir=None, render_factor=0):
H, W, focal = hwf
if render_factor!=0:
H = H//render_factor
W = W//render_factor
focal = focal/render_factor
rgbs = []
disps = []
t = time.time()
for i, c2w in enumerate(tqdm(render_poses)):
print(i, time.time() - t)
t = time.time()
rgb, disp, acc, _ = render(H, W, K, chunk=chunk, c2w=c2w[:3,:4], **render_kwargs)
rgbs.append(rgb.cpu().numpy())
disps.append(disp.cpu().numpy())
if i==0:
print(rgb.shape, disp.shape)
"""
if gt_imgs is not None and render_factor==0:
p = -10. * np.log10(np.mean(np.square(rgb.cpu().numpy() - gt_imgs[i])))
print(p)
"""
if savedir is not None:
rgb8 = to8b(rgbs[-1])
filename = os.path.join(savedir, '{:03d}.png'.format(i))
imageio.imwrite(filename, rgb8)
rgbs = np.stack(rgbs, 0)
disps = np.stack(disps, 0)
return rgbs, disps
run_nerf_helpers.py
这个里面写了一些必要的函数
class NeRF()
这个类用于创建model，alpha输出的是密度，rgb是颜色，一个batch是1024个光束，也就是一个光束采样64个点
class NeRF(nn.Module):
def __init__(self, D=8, W=256, input_ch=3, input_ch_views=3, output_ch=4, skips=[4], use_viewdirs=False):
"""
"""
super(NeRF, self).__init__()
self.D = D
self.W = W
self.input_ch = input_ch
self.input_ch_views = input_ch_views
self.skips = skips
self.use_viewdirs = use_viewdirs
self.pts_linears = nn.ModuleList(
[nn.Linear(input_ch, W)] + [nn.Linear(W, W) if i not in self.skips else nn.Linear(W + input_ch, W) for i in range(D-1)])
self.views_linears = nn.ModuleList([nn.Linear(input_ch_views + W, W//2)])
if use_viewdirs:
self.feature_linear = nn.Linear(W, W)
self.alpha_linear = nn.Linear(W, 1)
self.rgb_linear = nn.Linear(W//2, 3)
else:
self.output_linear = nn.Linear(W, output_ch)
def forward(self, x):
input_pts, input_views = torch.split(x, [self.input_ch, self.input_ch_views], dim=-1)
h = input_pts
for i, l in enumerate(self.pts_linears):
h = self.pts_linears[i](h)
h = F.relu(h)
if i in self.skips:
h = torch.cat([input_pts, h], -1)
if self.use_viewdirs:
alpha = self.alpha_linear(h)
feature = self.feature_linear(h)
h = torch.cat([feature, input_views], -1)
for i, l in enumerate(self.views_linears):
h = self.views_linears[i](h)
h = F.relu(h)
rgb = self.rgb_linear(h)
outputs = torch.cat([rgb, alpha], -1)
else:
outputs = self.output_linear(h)
return outputs
def load_weights_from_keras(self, weights):
assert self.use_viewdirs, "Not implemented if use_viewdirs=False"
for i in range(self.D):
idx_pts_linears = 2 * i
self.pts_linears[i].weight.data = torch.from_numpy(np.transpose(weights[idx_pts_linears]))
self.pts_linears[i].bias.data = torch.from_numpy(np.transpose(weights[idx_pts_linears+1]))
idx_feature_linear = 2 * self.D
self.feature_linear.weight.data = torch.from_numpy(np.transpose(weights[idx_feature_linear]))
self.feature_linear.bias.data = torch.from_numpy(np.transpose(weights[idx_feature_linear+1]))
idx_views_linears = 2 * self.D + 2
self.views_linears[0].weight.data = torch.from_numpy(np.transpose(weights[idx_views_linears]))
self.views_linears[0].bias.data = torch.from_numpy(np.transpose(weights[idx_views_linears+1]))
idx_rbg_linear = 2 * self.D + 4
self.rgb_linear.weight.data = torch.from_numpy(np.transpose(weights[idx_rbg_linear]))
self.rgb_linear.bias.data = torch.from_numpy(np.transpose(weights[idx_rbg_linear+1]))
idx_alpha_linear = 2 * self.D + 6
self.alpha_linear.weight.data = torch.from_numpy(np.transpose(weights[idx_alpha_linear]))
self.alpha_linear.bias.data = torch.from_numpy(np.transpose(weights[idx_alpha_linear+1]))
get_rays_np()
获得光束的方法
def get_rays_np(H, W, K, c2w):
i, j = np.meshgrid(np.arange(W, dtype=np.float32), np.arange(H, dtype=np.float32), indexing='xy')
dirs = np.stack([(i-K[0][2])/K[0][0], -(j-K[1][2])/K[1][1], -np.ones_like(i)], -1)
rays_d = np.sum(dirs[..., np.newaxis, :] * c2w[:3,:3], -1)
rays_o = np.broadcast_to(c2w[:3,-1], np.shape(rays_d))
return rays_o, rays_d
ndc_rays()
把光线的原点移动到near平面
def ndc_rays(H, W, focal, near, rays_o, rays_d):
t = -(near + rays_o[...,2]) / rays_d[...,2]
rays_o = rays_o + t[...,None] * rays_d
o0 = -1./(W/(2.*focal)) * rays_o[...,0] / rays_o[...,2]
o1 = -1./(H/(2.*focal)) * rays_o[...,1] / rays_o[...,2]
o2 = 1. + 2. * near / rays_o[...,2]
d0 = -1./(W/(2.*focal)) * (rays_d[...,0]/rays_d[...,2] - rays_o[...,0]/rays_o[...,2])
d1 = -1./(H/(2.*focal)) * (rays_d[...,1]/rays_d[...,2] - rays_o[...,1]/rays_o[...,2])
d2 = -2. * near / rays_o[...,2]
rays_o = torch.stack([o0,o1,o2], -1)
rays_d = torch.stack([d0,d1,d2], -1)
return rays_o, rays_d
接下来我们了解一下数据是怎么读取的
load_llff.py
_load_data()
def _load_data(basedir, factor=None, width=None, height=None, load_imgs=True):
poses_arr = np.load(os.path.join(basedir, 'poses_bounds.npy'))
poses = poses_arr[:, :-2].reshape([-1, 3, 5]).transpose([1,2,0])
bds = poses_arr[:, -2:].transpose([1,0])
img0 = [os.path.join(basedir, 'images', f) for f in sorted(os.listdir(os.path.join(basedir, 'images'))) \
if f.endswith('JPG') or f.endswith('jpg') or f.endswith('png')][0]
sh = imageio.imread(img0).shape
sfx = ''
if factor is not None:
sfx = '_{}'.format(factor)
_minify(basedir, factors=[factor])
factor = factor
elif height is not None:
factor = sh[0] / float(height)
width = int(sh[1] / factor)
_minify(basedir, resolutions=[[height, width]])
sfx = '_{}x{}'.format(width, height)
elif width is not None:
factor = sh[1] / float(width)
height = int(sh[0] / factor)
_minify(basedir, resolutions=[[height, width]])
sfx = '_{}x{}'.format(width, height)
else:
factor = 1
imgdir = os.path.join(basedir, 'images' + sfx)
if not os.path.exists(imgdir):
print( imgdir, 'does not exist, returning' )
return
imgfiles = [os.path.join(imgdir, f) for f in sorted(os.listdir(imgdir)) if f.endswith('JPG') or f.endswith('jpg') or f.endswith('png')]
if poses.shape[-1] != len(imgfiles):
print( 'Mismatch between imgs {} and poses {} !!!!'.format(len(imgfiles), poses.shape[-1]) )
return
sh = imageio.imread(imgfiles[0]).shape
poses[:2, 4, :] = np.array(sh[:2]).reshape([2, 1])
poses[2, 4, :] = poses[2, 4, :] * 1./factor
if not load_imgs:
return poses, bds
def imread(f):
if f.endswith('png'):
return imageio.imread(f, ignoregamma=True)
else:
return imageio.imread(f)
imgs = imgs = [imread(f)[...,:3]/255. for f in imgfiles]
imgs = np.stack(imgs, -1)
print('Loaded image data', imgs.shape, poses[:,-1,0])
return poses, bds, imgs
_minify()
这个函数主要负责创建目标分辨率的数据集
def _minify(basedir, factors=[], resolutions=[]):
needtoload = False
for r in factors:
imgdir = os.path.join(basedir, 'images_{}'.format(r))
if not os.path.exists(imgdir):
needtoload = True
for r in resolutions:
imgdir = os.path.join(basedir, 'images_{}x{}'.format(r[1], r[0]))
if not os.path.exists(imgdir):
needtoload = True
if not needtoload:
return
from shutil import copy
from subprocess import check_output
imgdir = os.path.join(basedir, 'images')
imgs = [os.path.join(imgdir, f) for f in sorted(os.listdir(imgdir))]
imgs = [f for f in imgs if any([f.endswith(ex) for ex in ['JPG', 'jpg', 'png', 'jpeg', 'PNG']])]
imgdir_orig = imgdir
wd = os.getcwd()
for r in factors + resolutions:
if isinstance(r, int):
name = 'images_{}'.format(r)
resizearg = '{}%'.format(100./r)
else:
name = 'images_{}x{}'.format(r[1], r[0])
resizearg = '{}x{}'.format(r[1], r[0])
imgdir = os.path.join(basedir, name)
if os.path.exists(imgdir):
continue
print('Minifying', r, basedir)
os.makedirs(imgdir)
check_output('cp {}/* {}'.format(imgdir_orig, imgdir), shell=True)
ext = imgs[0].split('.')[-1]
args = ' '.join(['mogrify', '-resize', resizearg, '-format', 'png', '*.{}'.format(ext)])
print(args)
os.chdir(imgdir)
check_output(args, shell=True)
os.chdir(wd)
if ext != 'png':
check_output('rm {}/*.{}'.format(imgdir, ext), shell=True)
print('Removed duplicates')
print('Done')
load_llff_data()
def load_llff_data(basedir, factor=8, recenter=True, bd_factor=.75, spherify=False, path_zflat=False):
poses, bds, imgs = _load_data(basedir, factor=factor)
print('Loaded', basedir, bds.min(), bds.max())
poses = np.concatenate([poses[:, 1:2, :], -poses[:, 0:1, :], poses[:, 2:, :]], 1)
poses = np.moveaxis(poses, -1, 0).astype(np.float32)
imgs = np.moveaxis(imgs, -1, 0).astype(np.float32)
images = imgs
bds = np.moveaxis(bds, -1, 0).astype(np.float32)
sc = 1. if bd_factor is None else 1./(bds.min() * bd_factor)
poses[:,:3,3] *= sc
bds *= sc
if recenter:
poses = recenter_poses(poses)
if spherify:
poses, render_poses, bds = spherify_poses(poses, bds)
else:
c2w = poses_avg(poses)
print('recentered', c2w.shape)
print(c2w[:3,:4])
up = normalize(poses[:, :3, 1].sum(0))
close_depth, inf_depth = bds.min()*.9, bds.max()*5.
dt = .75
mean_dz = 1./(((1.-dt)/close_depth + dt/inf_depth))
focal = mean_dz
shrink_factor = .8
zdelta = close_depth * .2
tt = poses[:,:3,3]
rads = np.percentile(np.abs(tt), 90, 0)
c2w_path = c2w
N_views = 120
N_rots = 2
if path_zflat:
zloc = -close_depth * .1
c2w_path[:3,3] = c2w_path[:3,3] + zloc * c2w_path[:3,2]
rads[2] = 0.
N_rots = 1
N_views/=2
render_poses = render_path_spiral(c2w_path, up, rads, focal, zdelta, zrate=.5, rots=N_rots, N=N_views)
render_poses = np.array(render_poses).astype(np.float32)
c2w = poses_avg(poses)
print('Data:')
print(poses.shape, images.shape, bds.shape)
dists = np.sum(np.square(c2w[:3,3] - poses[:,:3,3]), -1)
i_test = np.argmin(dists)
print('HOLDOUT view is', i_test)
images = images.astype(np.float32)
poses = poses.astype(np.float32)
return images, poses, bds, render_poses, i_test
render_path_spiral()
def render_path_spiral(c2w, up, rads, focal, zdelta, zrate, rots, N):
render_poses = []
rads = np.array(list(rads) + [1.])
hwf = c2w[:,4:5]
for theta in np.linspace(0., 2. * np.pi * rots, N+1)[:-1]:
c = np.dot(c2w[:3,:4], np.array([np.cos(theta), -np.sin(theta), -np.sin(theta*zrate), 1.]) * rads)
z = normalize(c - np.dot(c2w[:3,:4], np.array([0,0,-focal, 1.])))
render_poses.append(np.concatenate([viewmatrix(z, up, c), hwf], 1))
return render_poses
Original: https://blog.csdn.net/YuhsiHu/article/details/124676445
Author: YuhsiHu
Title: 【代码详解】nerf-pytorch代码逐行分析
原创文章受到原创版权保护。转载请注明出处：https://www.johngo689.com/648123/
转载文章受原作者版权保护。转载请注明原作者出处！