You cannot select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
392 lines
15 KiB
Python
392 lines
15 KiB
Python
4 years ago
|
import torch
|
||
|
|
import torch.nn as nn
|
||
|
|
import torch.optim
|
||
|
|
import torch.distributed
|
||
|
|
from torch.nn.parallel import DistributedDataParallel as DDP
|
||
|
|
import torch.multiprocessing
|
||
|
|
import numpy as np
|
||
|
|
|
||
|
|
import os
|
||
|
|
from collections import OrderedDict
|
||
|
|
from ddp_model import NerfNet
|
||
|
|
import time
|
||
|
|
from data_loader_split import load_data_split
|
||
|
|
from utils import mse2psnr, img_HWC2CHW, colorize, colorize_np, TINY_NUMBER, to8b
|
||
|
|
import imageio
|
||
|
|
from ddp_run_nerf import config_parser
|
||
|
|
|
||
|
|
import logging
|
||
|
|
|
||
|
|
|
||
|
|
logger = logging.getLogger(__package__)
|
||
|
|
|
||
|
|
|
||
|
|
def setup_logger():
|
||
|
|
# create logger
|
||
|
|
logger = logging.getLogger(__package__)
|
||
|
|
logger.setLevel(logging.DEBUG)
|
||
|
|
|
||
|
|
# create console handler and set level to debug
|
||
|
|
ch = logging.StreamHandler()
|
||
|
|
ch.setLevel(logging.INFO)
|
||
|
|
|
||
|
|
# create formatter
|
||
|
|
formatter = logging.Formatter('%(asctime)s [%(levelname)s] %(name)s: %(message)s')
|
||
|
|
|
||
|
|
# add formatter to ch
|
||
|
|
ch.setFormatter(formatter)
|
||
|
|
|
||
|
|
# add ch to logger
|
||
|
|
logger.addHandler(ch)
|
||
|
|
|
||
|
|
|
||
|
|
def intersect_sphere(ray_o, ray_d):
|
||
|
|
'''
|
||
|
|
ray_o, ray_d: [..., 3]
|
||
|
|
compute the depth of the intersection point between this ray and unit sphere
|
||
|
|
'''
|
||
|
|
# note: d1 becomes negative if this mid point is behind camera
|
||
|
|
d1 = -torch.sum(ray_d * ray_o, dim=-1) / torch.sum(ray_d * ray_d, dim=-1)
|
||
|
|
p = ray_o + d1.unsqueeze(-1) * ray_d
|
||
|
|
# consider the case where the ray does not intersect the sphere
|
||
|
|
ray_d_cos = 1. / torch.norm(ray_d, dim=-1)
|
||
|
|
d2 = torch.sqrt(1. - torch.sum(p * p, dim=-1)) * ray_d_cos
|
||
|
|
|
||
|
|
return d1 + d2
|
||
|
|
|
||
|
|
|
||
|
|
def perturb_samples(z_vals):
|
||
|
|
# get intervals between samples
|
||
|
|
mids = .5 * (z_vals[..., 1:] + z_vals[..., :-1])
|
||
|
|
upper = torch.cat([mids, z_vals[..., -1:]], dim=-1)
|
||
|
|
lower = torch.cat([z_vals[..., 0:1], mids], dim=-1)
|
||
|
|
# uniform samples in those intervals
|
||
|
|
t_rand = torch.rand_like(z_vals)
|
||
|
|
z_vals = lower + (upper - lower) * t_rand # [N_rays, N_samples]
|
||
|
|
|
||
|
|
return z_vals
|
||
|
|
|
||
|
|
|
||
|
|
def sample_pdf(bins, weights, N_samples, det=False):
|
||
|
|
'''
|
||
|
|
:param bins: tensor of shape [..., M+1], M is the number of bins
|
||
|
|
:param weights: tensor of shape [..., M]
|
||
|
|
:param N_samples: number of samples along each ray
|
||
|
|
:param det: if True, will perform deterministic sampling
|
||
|
|
:return: [..., N_samples]
|
||
|
|
'''
|
||
|
|
# Get pdf
|
||
|
|
weights = weights + TINY_NUMBER # prevent nans
|
||
|
|
pdf = weights / torch.sum(weights, dim=-1, keepdim=True) # [..., M]
|
||
|
|
cdf = torch.cumsum(pdf, dim=-1) # [..., M]
|
||
|
|
cdf = torch.cat([torch.zeros_like(cdf[..., 0:1]), cdf], dim=-1) # [..., M+1]
|
||
|
|
|
||
|
|
# Take uniform samples
|
||
|
|
dots_sh = list(weights.shape[:-1])
|
||
|
|
M = weights.shape[-1]
|
||
|
|
|
||
|
|
min_cdf = 0.00
|
||
|
|
max_cdf = 1.00 # prevent outlier samples
|
||
|
|
|
||
|
|
if det:
|
||
|
|
u = torch.linspace(min_cdf, max_cdf, N_samples, device=bins.device)
|
||
|
|
u = u.view([1]*len(dots_sh) + [N_samples]).expand(dots_sh + [N_samples,]) # [..., N_samples]
|
||
|
|
else:
|
||
|
|
sh = dots_sh + [N_samples]
|
||
|
|
u = torch.rand(*sh, device=bins.device) * (max_cdf - min_cdf) + min_cdf # [..., N_samples]
|
||
|
|
|
||
|
|
# Invert CDF
|
||
|
|
# [..., N_samples, 1] >= [..., 1, M] ----> [..., N_samples, M] ----> [..., N_samples,]
|
||
|
|
above_inds = torch.sum(u.unsqueeze(-1) >= cdf[..., :M].unsqueeze(-2), dim=-1).long()
|
||
|
|
|
||
|
|
# random sample inside each bin
|
||
|
|
below_inds = torch.clamp(above_inds-1, min=0)
|
||
|
|
inds_g = torch.stack((below_inds, above_inds), dim=-1) # [..., N_samples, 2]
|
||
|
|
|
||
|
|
cdf = cdf.unsqueeze(-2).expand(dots_sh + [N_samples, M+1]) # [..., N_samples, M+1]
|
||
|
|
cdf_g = torch.gather(input=cdf, dim=-1, index=inds_g) # [..., N_samples, 2]
|
||
|
|
|
||
|
|
bins = bins.unsqueeze(-2).expand(dots_sh + [N_samples, M+1]) # [..., N_samples, M+1]
|
||
|
|
bins_g = torch.gather(input=bins, dim=-1, index=inds_g) # [..., N_samples, 2]
|
||
|
|
|
||
|
|
# fix numeric issue
|
||
|
|
denom = cdf_g[..., 1] - cdf_g[..., 0] # [..., N_samples]
|
||
|
|
denom = torch.where(denom<TINY_NUMBER, torch.ones_like(denom), denom)
|
||
|
|
t = (u - cdf_g[..., 0]) / denom
|
||
|
|
|
||
|
|
samples = bins_g[..., 0] + t * (bins_g[..., 1] - bins_g[..., 0] + TINY_NUMBER)
|
||
|
|
|
||
|
|
return samples
|
||
|
|
|
||
|
|
|
||
|
|
def render_single_image(rank, world_size, models, ray_sampler, chunk_size):
|
||
|
|
##### parallel rendering of a single image
|
||
|
|
ray_batch = ray_sampler.get_all()
|
||
|
|
# split into ranks; make sure different processes don't overlap
|
||
|
|
rank_split_sizes = [ray_batch['ray_d'].shape[0] // world_size, ] * world_size
|
||
|
|
rank_split_sizes[-1] = ray_batch['ray_d'].shape[0] - sum(rank_split_sizes[:-1])
|
||
|
|
for key in ray_batch:
|
||
|
|
if torch.is_tensor(ray_batch[key]):
|
||
|
|
ray_batch[key] = torch.split(ray_batch[key], rank_split_sizes)[rank].to(rank)
|
||
|
|
|
||
|
|
# split into chunks and render inside each process
|
||
|
|
ray_batch_split = OrderedDict()
|
||
|
|
for key in ray_batch:
|
||
|
|
if torch.is_tensor(ray_batch[key]):
|
||
|
|
ray_batch_split[key] = torch.split(ray_batch[key], chunk_size)
|
||
|
|
|
||
|
|
# forward and backward
|
||
|
|
ret_merge_chunk = [OrderedDict() for _ in range(models['cascade_level'])]
|
||
|
|
for s in range(len(ray_batch_split['ray_d'])):
|
||
|
|
ray_o = ray_batch_split['ray_o'][s]
|
||
|
|
ray_d = ray_batch_split['ray_d'][s]
|
||
|
|
min_depth = ray_batch_split['min_depth'][s]
|
||
|
|
|
||
|
|
dots_sh = list(ray_d.shape[:-1])
|
||
|
|
for m in range(models['cascade_level']):
|
||
|
|
net = models['net_{}'.format(m)]
|
||
|
|
# sample depths
|
||
|
|
N_samples = models['cascade_samples'][m]
|
||
|
|
if m == 0:
|
||
|
|
# foreground depth
|
||
|
|
fg_far_depth = intersect_sphere(ray_o, ray_d) # [...,]
|
||
|
|
# fg_near_depth = 0.18 * torch.ones_like(fg_far_depth)
|
||
|
|
fg_near_depth = min_depth # [..., 3]
|
||
|
|
step = (fg_far_depth - fg_near_depth) / (N_samples - 1)
|
||
|
|
fg_depth = torch.stack([fg_near_depth + i * step for i in range(N_samples)], dim=-1) # [..., N_samples]
|
||
|
|
|
||
|
|
# background depth
|
||
|
|
bg_depth = torch.linspace(0., 1., N_samples).view(
|
||
|
|
[1, ] * len(dots_sh) + [N_samples,]).expand(dots_sh + [N_samples,]).to(rank)
|
||
|
|
|
||
|
|
# delete unused memory
|
||
|
|
del fg_near_depth
|
||
|
|
del step
|
||
|
|
torch.cuda.empty_cache()
|
||
|
|
else:
|
||
|
|
# sample pdf and concat with earlier samples
|
||
|
|
fg_weights = ret['fg_weights'].clone().detach()
|
||
|
|
fg_depth_mid = .5 * (fg_depth[..., 1:] + fg_depth[..., :-1]) # [..., N_samples-1]
|
||
|
|
fg_weights = fg_weights[..., 1:-1] # [..., N_samples-2]
|
||
|
|
fg_depth_samples = sample_pdf(bins=fg_depth_mid, weights=fg_weights,
|
||
|
|
N_samples=N_samples, det=True) # [..., N_samples]
|
||
|
|
fg_depth, _ = torch.sort(torch.cat((fg_depth, fg_depth_samples), dim=-1))
|
||
|
|
|
||
|
|
# sample pdf and concat with earlier samples
|
||
|
|
bg_weights = ret['bg_weights'].clone().detach()
|
||
|
|
bg_depth_mid = .5 * (bg_depth[..., 1:] + bg_depth[..., :-1])
|
||
|
|
bg_weights = bg_weights[..., 1:-1] # [..., N_samples-2]
|
||
|
|
bg_depth_samples = sample_pdf(bins=bg_depth_mid, weights=bg_weights,
|
||
|
|
N_samples=N_samples, det=True) # [..., N_samples]
|
||
|
|
bg_depth, _ = torch.sort(torch.cat((bg_depth, bg_depth_samples), dim=-1))
|
||
|
|
|
||
|
|
# delete unused memory
|
||
|
|
del fg_weights
|
||
|
|
del fg_depth_mid
|
||
|
|
del fg_depth_samples
|
||
|
|
del bg_weights
|
||
|
|
del bg_depth_mid
|
||
|
|
del bg_depth_samples
|
||
|
|
torch.cuda.empty_cache()
|
||
|
|
|
||
|
|
with torch.no_grad():
|
||
|
|
ret = net(ray_o, ray_d, fg_far_depth, fg_depth, bg_depth)
|
||
|
|
|
||
|
|
for key in ret:
|
||
|
|
if key not in ['fg_weights', 'bg_weights']:
|
||
|
|
if torch.is_tensor(ret[key]):
|
||
|
|
if key not in ret_merge_chunk[m]:
|
||
|
|
ret_merge_chunk[m][key] = [ret[key].cpu(), ]
|
||
|
|
else:
|
||
|
|
ret_merge_chunk[m][key].append(ret[key].cpu())
|
||
|
|
|
||
|
|
ret[key] = None
|
||
|
|
|
||
|
|
# clean unused memory
|
||
|
|
torch.cuda.empty_cache()
|
||
|
|
|
||
|
|
# merge results from different chunks
|
||
|
|
for m in range(len(ret_merge_chunk)):
|
||
|
|
for key in ret_merge_chunk[m]:
|
||
|
|
ret_merge_chunk[m][key] = torch.cat(ret_merge_chunk[m][key], dim=0)
|
||
|
|
|
||
|
|
# merge results from different processes
|
||
|
|
if rank == 0:
|
||
|
|
ret_merge_rank = [OrderedDict() for _ in range(len(ret_merge_chunk))]
|
||
|
|
for m in range(len(ret_merge_chunk)):
|
||
|
|
for key in ret_merge_chunk[m]:
|
||
|
|
# generate tensors to store results from other processes
|
||
|
|
sh = list(ret_merge_chunk[m][key].shape[1:])
|
||
|
|
ret_merge_rank[m][key] = [torch.zeros(*[size,]+sh, dtype=torch.float32) for size in rank_split_sizes]
|
||
|
|
torch.distributed.gather(ret_merge_chunk[m][key], ret_merge_rank[m][key])
|
||
|
|
ret_merge_rank[m][key] = torch.cat(ret_merge_rank[m][key], dim=0).reshape(
|
||
|
|
(ray_sampler.H, ray_sampler.W, -1)).squeeze()
|
||
|
|
# print(m, key, ret_merge_rank[m][key].shape)
|
||
|
|
else: # send results to main process
|
||
|
|
for m in range(len(ret_merge_chunk)):
|
||
|
|
for key in ret_merge_chunk[m]:
|
||
|
|
torch.distributed.gather(ret_merge_chunk[m][key])
|
||
|
|
|
||
|
|
# only rank 0 program returns
|
||
|
|
if rank == 0:
|
||
|
|
return ret_merge_rank
|
||
|
|
else:
|
||
|
|
return None
|
||
|
|
|
||
|
|
|
||
|
|
def setup(rank, world_size):
|
||
|
|
os.environ['MASTER_ADDR'] = 'localhost'
|
||
|
|
os.environ['MASTER_PORT'] = '12355'
|
||
|
|
# initialize the process group
|
||
|
|
torch.distributed.init_process_group("gloo", rank=rank, world_size=world_size)
|
||
|
|
|
||
|
|
|
||
|
|
def cleanup():
|
||
|
|
torch.distributed.destroy_process_group()
|
||
|
|
|
||
|
|
|
||
|
|
def ddp_test_nerf(rank, args):
|
||
|
|
###### set up multi-processing
|
||
|
|
setup(rank, args.world_size)
|
||
|
|
###### set up logger
|
||
|
|
logger = logging.getLogger(__package__)
|
||
|
|
setup_logger()
|
||
|
|
|
||
|
|
###### decide chunk size according to gpu memory
|
||
|
|
if torch.cuda.get_device_properties(rank).total_memory / 1e9 > 14:
|
||
|
|
logger.info('setting batch size according to 24G gpu')
|
||
|
|
args.N_rand = 1024
|
||
|
|
args.chunk_size = 8192
|
||
|
|
else:
|
||
|
|
logger.info('setting batch size according to 12G gpu')
|
||
|
|
args.N_rand = 512
|
||
|
|
args.chunk_size = 4096
|
||
|
|
|
||
|
|
###### create network and wrap in ddp; each process should do this
|
||
|
|
# fix random seed just to make sure the network is initialized with same weights at different processes
|
||
|
|
torch.manual_seed(777)
|
||
|
|
# very important!!! otherwise it might introduce extra memory in rank=0 gpu
|
||
|
|
torch.cuda.set_device(rank)
|
||
|
|
|
||
|
|
models = OrderedDict()
|
||
|
|
models['cascade_level'] = args.cascade_level
|
||
|
|
models['cascade_samples'] = [int(x.strip()) for x in args.cascade_samples.split(',')]
|
||
|
|
for m in range(models['cascade_level']):
|
||
|
|
net = NerfNet(args).to(rank)
|
||
|
|
net = DDP(net, device_ids=[rank], output_device=rank)
|
||
|
|
optim = torch.optim.Adam(net.parameters(), lr=args.lrate)
|
||
|
|
models['net_{}'.format(m)] = net
|
||
|
|
models['optim_{}'.format(m)] = optim
|
||
|
|
|
||
|
|
start = -1
|
||
|
|
|
||
|
|
###### load pretrained weights; each process should do this
|
||
|
|
if (args.ckpt_path is not None) and (os.path.isfile(args.ckpt_path)):
|
||
|
|
ckpts = [args.ckpt_path]
|
||
|
|
else:
|
||
|
|
ckpts = [os.path.join(args.basedir, args.expname, f)
|
||
|
|
for f in sorted(os.listdir(os.path.join(args.basedir, args.expname))) if f.endswith('.pth')]
|
||
|
|
def path2iter(path):
|
||
|
|
tmp = os.path.basename(path)[:-4]
|
||
|
|
idx = tmp.rfind('_')
|
||
|
|
return int(tmp[idx + 1:])
|
||
|
|
ckpts = sorted(ckpts, key=path2iter)
|
||
|
|
logger.info('Found ckpts: {}'.format(ckpts))
|
||
|
|
if len(ckpts) > 0 and not args.no_reload:
|
||
|
|
fpath = ckpts[-1]
|
||
|
|
logger.info('Reloading from: {}'.format(fpath))
|
||
|
|
start = path2iter(fpath)
|
||
|
|
# configure map_location properly for different processes
|
||
|
|
map_location = {'cuda:%d' % 0: 'cuda:%d' % rank}
|
||
|
|
to_load = torch.load(fpath, map_location=map_location)
|
||
|
|
for m in range(models['cascade_level']):
|
||
|
|
for name in ['net_{}'.format(m), 'optim_{}'.format(m)]:
|
||
|
|
models[name].load_state_dict(to_load[name])
|
||
|
|
models[name].load_state_dict(to_load[name])
|
||
|
|
|
||
|
|
render_splits = [x.strip() for x in args.render_splits.strip().split(',')]
|
||
|
|
# start testing
|
||
|
|
for split in render_splits:
|
||
|
|
out_dir = os.path.join(args.basedir, args.expname,
|
||
|
|
'render_{}_{:06d}'.format(split, start))
|
||
|
|
if rank == 0:
|
||
|
|
os.makedirs(out_dir, exist_ok=True)
|
||
|
|
|
||
|
|
###### load data and create ray samplers; each process should do this
|
||
|
|
ray_samplers = load_data_split(args.datadir, args.scene, split, try_load_min_depth=args.load_min_depth)
|
||
|
|
for idx in range(len(ray_samplers)):
|
||
|
|
### each process should do this; but only main process merges the results
|
||
|
|
fname = '{:06d}.png'.format(idx)
|
||
|
|
if ray_samplers[idx].img_path is not None:
|
||
|
|
fname = os.path.basename(ray_samplers[idx].img_path)
|
||
|
|
|
||
|
|
if os.path.isfile(os.path.join(out_dir, fname)):
|
||
|
|
logger.info('Skipping {}'.format(fname))
|
||
|
|
continue
|
||
|
|
|
||
|
|
time0 = time.time()
|
||
|
|
ret = render_single_image(rank, args.world_size, models, ray_samplers[idx], args.chunk_size)
|
||
|
|
dt = time.time() - time0
|
||
|
|
if rank == 0: # only main process should do this
|
||
|
|
|
||
|
|
logger.info('Rendered {} in {} seconds'.format(fname, dt))
|
||
|
|
|
||
|
|
# only save last level
|
||
|
|
im = ret[-1]['rgb'].numpy()
|
||
|
|
# compute psnr if ground-truth is available
|
||
|
|
if ray_samplers[idx].img_path is not None:
|
||
|
|
gt_im = ray_samplers[idx].get_img()
|
||
|
|
psnr = mse2psnr(np.mean((gt_im - im) * (gt_im - im)))
|
||
|
|
logger.info('{}: psnr={}'.format(fname, psnr))
|
||
|
|
|
||
|
|
im = to8b(im)
|
||
|
|
imageio.imwrite(os.path.join(out_dir, fname), im)
|
||
|
|
|
||
|
|
# im = ret[-1]['diffuse_rgb'].numpy()
|
||
|
|
# im = to8b(im)
|
||
|
|
# imageio.imwrite(os.path.join(out_dir, 'diffuse_' + fname), im)
|
||
|
|
|
||
|
|
im = ret[-1]['fg_rgb'].numpy()
|
||
|
|
im = to8b(im)
|
||
|
|
imageio.imwrite(os.path.join(out_dir, 'fg_' + fname), im)
|
||
|
|
|
||
|
|
im = ret[-1]['bg_rgb'].numpy()
|
||
|
|
im = to8b(im)
|
||
|
|
imageio.imwrite(os.path.join(out_dir, 'bg_' + fname), im)
|
||
|
|
|
||
|
|
im = ret[-1]['fg_depth'].numpy()
|
||
|
|
im = colorize_np(im, cmap_name='jet', append_cbar=True)
|
||
|
|
im = to8b(im)
|
||
|
|
imageio.imwrite(os.path.join(out_dir, 'fg_depth_' + fname), im)
|
||
|
|
|
||
|
|
im = ret[-1]['bg_depth'].numpy()
|
||
|
|
im = colorize_np(im, cmap_name='jet', append_cbar=True)
|
||
|
|
im = to8b(im)
|
||
|
|
imageio.imwrite(os.path.join(out_dir, 'bg_depth_' + fname), im)
|
||
|
|
|
||
|
|
torch.cuda.empty_cache()
|
||
|
|
|
||
|
|
# clean up for multi-processing
|
||
|
|
cleanup()
|
||
|
|
|
||
|
|
|
||
|
|
def test():
|
||
|
|
parser = config_parser()
|
||
|
|
args = parser.parse_args()
|
||
|
|
logger.info(parser.format_values())
|
||
|
|
|
||
|
|
if args.world_size == -1:
|
||
|
|
args.world_size = torch.cuda.device_count()
|
||
|
|
logger.info('Using # gpus: {}'.format(args.world_size))
|
||
|
|
torch.multiprocessing.spawn(ddp_test_nerf,
|
||
|
|
args=(args,),
|
||
|
|
nprocs=args.world_size,
|
||
|
|
join=True)
|
||
|
|
|
||
|
|
|
||
|
|
if __name__ == '__main__':
|
||
|
|
setup_logger()
|
||
|
|
test()
|
||
|
|
|
||
|
|
|