train_ngp_nerf.py

"""
Copyright (c) 2022 Ruilong Li, UC Berkeley.
"""

import argparse
import math
import pathlib
import time

import imageio
import numpy as np
import torch
import torch.nn.functional as F
import tqdm
from lpips import LPIPS
from radiance_fields.ngp import NGPRadianceField
from utils import (
    MIPNERF360_UNBOUNDED_SCENES,
    NERF_SYNTHETIC_SCENES,
    enlarge_aabb,
    render_image,
    set_random_seed,
)

from nerfacc import OccupancyGrid

parser = argparse.ArgumentParser()
parser.add_argument(
    "--data_root",
    type=str,
    # default=str(pathlib.Path.cwd() / "data/360_v2"),
    default=str(pathlib.Path.cwd() / "data/nerf_synthetic"),
    help="the root dir of the dataset",
)
parser.add_argument(
    "--train_split",
    type=str,
    default="train",
    choices=["train", "trainval"],
    help="which train split to use",
)
parser.add_argument(
    "--scene",
    type=str,
    default="lego",
    choices=NERF_SYNTHETIC_SCENES + MIPNERF360_UNBOUNDED_SCENES,
    help="which scene to use",
)
parser.add_argument(
    "--test_chunk_size",
    type=int,
    default=8192,
)
args = parser.parse_args()

device = "cuda:0"
set_random_seed(42)

if args.scene in MIPNERF360_UNBOUNDED_SCENES:
    from datasets.nerf_360_v2 import SubjectLoader

    # training parameters
    max_steps = 100000
    init_batch_size = 1024
    target_sample_batch_size = 1 << 18
    weight_decay = 0.0
    # scene parameters
    aabb = torch.tensor([-1.0, -1.0, -1.0, 1.0, 1.0, 1.0], device=device)
    near_plane = 0.02
    far_plane = None
    # dataset parameters
    train_dataset_kwargs = {"color_bkgd_aug": "random", "factor": 4}
    test_dataset_kwargs = {"factor": 4}
    # model parameters
    grid_resolution = 128
    grid_nlvl = 4
    # render parameters
    render_step_size = 1e-3
    alpha_thre = 1e-2
    cone_angle = 0.004

else:
    from datasets.nerf_synthetic import SubjectLoader

    # training parameters
    max_steps = 20000
    init_batch_size = 1024
    target_sample_batch_size = 1 << 18
    weight_decay = (
        1e-5 if args.scene in ["materials", "ficus", "drums"] else 1e-6
    )
    # scene parameters
    aabb = torch.tensor([-1.5, -1.5, -1.5, 1.5, 1.5, 1.5], device=device)
    near_plane = None
    far_plane = None
    # dataset parameters
    train_dataset_kwargs = {}
    test_dataset_kwargs = {}
    # model parameters
    grid_resolution = 128
    grid_nlvl = 1
    # render parameters
    render_step_size = 5e-3
    alpha_thre = 0.0
    cone_angle = 0.0

train_dataset = SubjectLoader(
    subject_id=args.scene,
    root_fp=args.data_root,
    split=args.train_split,
    num_rays=init_batch_size,
    device=device,
    **train_dataset_kwargs,
)

test_dataset = SubjectLoader(
    subject_id=args.scene,
    root_fp=args.data_root,
    split="test",
    num_rays=None,
    device=device,
    **test_dataset_kwargs,
)

# setup scene aabb
scene_aabb = enlarge_aabb(aabb, 1 << (grid_nlvl - 1))

# setup the radiance field we want to train.
grad_scaler = torch.cuda.amp.GradScaler(2**10)
radiance_field = NGPRadianceField(aabb=scene_aabb).to(device)
optimizer = torch.optim.Adam(
    radiance_field.parameters(), lr=1e-2, eps=1e-15, weight_decay=weight_decay
)
scheduler = torch.optim.lr_scheduler.ChainedScheduler(
    [
        torch.optim.lr_scheduler.LinearLR(
            optimizer, start_factor=0.01, total_iters=100
        ),
        torch.optim.lr_scheduler.MultiStepLR(
            optimizer,
            milestones=[
                max_steps // 2,
                max_steps * 3 // 4,
                max_steps * 9 // 10,
            ],
            gamma=0.33,
        ),
    ]
)
occupancy_grid = OccupancyGrid(
    roi_aabb=aabb, resolution=grid_resolution, levels=grid_nlvl
).to(device)

lpips_net = LPIPS(net="vgg").to(device)
lpips_norm_fn = lambda x: x[None, ...].permute(0, 3, 1, 2) * 2 - 1
lpips_fn = lambda x, y: lpips_net(lpips_norm_fn(x), lpips_norm_fn(y)).mean()

# training
tic = time.time()
for step in range(max_steps + 1):
    radiance_field.train()

    i = torch.randint(0, len(train_dataset), (1,)).item()
    data = train_dataset[i]

    render_bkgd = data["color_bkgd"]
    rays = data["rays"]
    pixels = data["pixels"]

    def occ_eval_fn(x):
        density = radiance_field.query_density(x)
        return density * render_step_size

    # update occupancy grid
    occupancy_grid.every_n_step(
        step=step,
        occ_eval_fn=occ_eval_fn,
        occ_thre=1e-2,
    )

    # render
    rgb, acc, depth, n_rendering_samples = render_image(
        radiance_field,
        occupancy_grid,
        rays,
        scene_aabb=scene_aabb,
        # rendering options
        near_plane=near_plane,
        render_step_size=render_step_size,
        render_bkgd=render_bkgd,
        cone_angle=cone_angle,
        alpha_thre=alpha_thre,
    )
    if n_rendering_samples == 0:
        continue

    if target_sample_batch_size > 0:
        # dynamic batch size for rays to keep sample batch size constant.
        num_rays = len(pixels)
        num_rays = int(
            num_rays * (target_sample_batch_size / float(n_rendering_samples))
        )
        train_dataset.update_num_rays(num_rays)

    # compute loss
    loss = F.smooth_l1_loss(rgb, pixels)

    optimizer.zero_grad()
    # do not unscale it because we are using Adam.
    grad_scaler.scale(loss).backward()
    optimizer.step()
    scheduler.step()

    if step % 5000 == 0:
        elapsed_time = time.time() - tic
        loss = F.mse_loss(rgb, pixels)
        psnr = -10.0 * torch.log(loss) / np.log(10.0)
        print(
            f"elapsed_time={elapsed_time:.2f}s | step={step} | "
            f"loss={loss:.5f} | psnr={psnr:.2f} | "
            f"n_rendering_samples={n_rendering_samples:d} | num_rays={len(pixels):d} | "
            f"max_depth={depth.max():.3f} | "
        )

    if step > 0 and step % max_steps == 0:
        # evaluation
        radiance_field.eval()

        psnrs = []
        lpips = []
        with torch.no_grad():
            for i in tqdm.tqdm(range(len(test_dataset))):
                data = test_dataset[i]
                render_bkgd = data["color_bkgd"]
                rays = data["rays"]
                pixels = data["pixels"]

                # rendering
                rgb, acc, depth, _ = render_image(
                    radiance_field,
                    occupancy_grid,
                    rays,
                    scene_aabb=scene_aabb,
                    # rendering options
                    near_plane=near_plane,
                    render_step_size=render_step_size,
                    render_bkgd=render_bkgd,
                    cone_angle=cone_angle,
                    alpha_thre=alpha_thre,
                    # test options
                    test_chunk_size=args.test_chunk_size,
                )
                mse = F.mse_loss(rgb, pixels)
                psnr = -10.0 * torch.log(mse) / np.log(10.0)
                psnrs.append(psnr.item())
                lpips.append(lpips_fn(rgb, pixels).item())
                # if i == 0:
                #     imageio.imwrite(
                #         "rgb_test.png",
                #         (rgb.cpu().numpy() * 255).astype(np.uint8),
                #     )
                #     imageio.imwrite(
                #         "rgb_error.png",
                #         (
                #             (rgb - pixels).norm(dim=-1).cpu().numpy() * 255
                #         ).astype(np.uint8),
                #     )
        psnr_avg = sum(psnrs) / len(psnrs)
        lpips_avg = sum(lpips) / len(lpips)
        print(f"evaluation: psnr_avg={psnr_avg}, lpips_avg={lpips_avg}")