mirror of
https://github.com/nerfstudio-project/nerfstudio.git
synced 2026-11-11 02:24:38 +08:00
242 lines
8.9 KiB
Python
242 lines
8.9 KiB
Python
"""
|
|
run_eval.py
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import dataclasses
|
|
import json
|
|
import logging
|
|
import os
|
|
from dataclasses import dataclass
|
|
from pathlib import Path
|
|
from typing import Dict, Literal, Tuple, Union
|
|
|
|
import dcargs
|
|
import mediapy as media
|
|
import torch
|
|
import yaml
|
|
from tqdm import tqdm
|
|
from typing_extensions import assert_never
|
|
|
|
from nerfactory.cameras.camera_paths import (
|
|
get_interpolated_camera_path,
|
|
get_spiral_path,
|
|
)
|
|
from nerfactory.cameras.cameras import Cameras
|
|
from nerfactory.configs import base as cfg
|
|
from nerfactory.pipelines.base import Pipeline
|
|
from nerfactory.utils.misc import human_format
|
|
from nerfactory.utils.writer import TimeWriter
|
|
|
|
logging.basicConfig(format="[%(filename)s:%(lineno)d] %(message)s", level=logging.INFO)
|
|
|
|
|
|
def _update_avg(prev_avg: float, new_val: float, step: int) -> float:
|
|
"""helper to calculate the running average
|
|
|
|
Args:
|
|
prev_avg (float): previous average value
|
|
new_val (float): new value to update the average with
|
|
step (int): current step number
|
|
|
|
Returns:
|
|
float: new updated average
|
|
"""
|
|
return (step * prev_avg + new_val) / (step + 1)
|
|
|
|
|
|
def _load_checkpoint(config: cfg.TrainerConfig, pipeline: Pipeline) -> Path:
|
|
"""Helper function to load checkpointed pipeline
|
|
|
|
Args:
|
|
config (DictConfig): Configuration of pipeline to load
|
|
pipeline (Pipeline): Pipeline instance of which to load weights
|
|
"""
|
|
assert config.load_dir is not None
|
|
if config.load_step is None:
|
|
print("Loading latest checkpoint from load_dir")
|
|
# NOTE: this is specific to the checkpoint name format
|
|
load_step = sorted(int(x[x.find("-") + 1 : x.find(".")]) for x in os.listdir(config.load_dir))[-1]
|
|
else:
|
|
load_step = config.load_step
|
|
load_path = config.load_dir / f"step-{load_step:09d}.ckpt"
|
|
assert load_path.exists(), f"Checkpoint {load_path} does not exist"
|
|
loaded_state = torch.load(load_path, map_location="cpu")
|
|
pipeline.load_pipeline(loaded_state["pipeline"])
|
|
print(f"done loading checkpoint from {load_path}")
|
|
return load_path
|
|
|
|
|
|
def _render_stats_dict(pipeline: Pipeline) -> Dict[str, float]:
|
|
"""Helper function to evaluate the pipeline on a dataloader.
|
|
|
|
Args:
|
|
pipeline (Pipeline): Pipeline to evaluate
|
|
|
|
Returns:
|
|
dict: returns the average psnr and average rays per second
|
|
"""
|
|
avg_psnr = 0
|
|
avg_rays_per_sec = 0
|
|
avg_fps = 0
|
|
for step, (camera_ray_bundle, batch) in tqdm(enumerate(pipeline.dataloader.eval_dataloader)):
|
|
with TimeWriter(writer=None, name=None, write=False) as t:
|
|
with torch.no_grad():
|
|
image_idx = int(camera_ray_bundle.camera_indices[0, 0])
|
|
outputs = pipeline.model.get_outputs_for_camera_ray_bundle(camera_ray_bundle)
|
|
psnr = pipeline.model.log_test_image_outputs(image_idx, step, batch, outputs)
|
|
avg_rays_per_sec = _update_avg(avg_rays_per_sec, camera_ray_bundle.origins.shape[0] / t.duration, step)
|
|
avg_psnr = _update_avg(avg_psnr, psnr, step)
|
|
avg_fps = _update_avg(avg_fps, 1 / t.duration, step)
|
|
return {"avg psnr": avg_psnr, "avg rays per sec": avg_rays_per_sec, "avg fps": avg_fps}
|
|
|
|
|
|
def _render_trajectory_video(
|
|
pipeline: Pipeline,
|
|
cameras: Cameras,
|
|
output_filename: Path,
|
|
rendered_output_name: str,
|
|
rendered_resolution_scaling_factor: float = 1.0,
|
|
num_rays_per_chunk: int = 4096,
|
|
) -> None:
|
|
"""Helper function to create a video of the spiral trajectory.
|
|
|
|
Args:
|
|
pipeline: Pipeline to evaluate with.
|
|
cameras: Cameras to render.
|
|
output_filename: Name of the output file.
|
|
rendered_output_name: Name of the renderer output to use.
|
|
rendered_resolution_scaling_factor: Scaling factor to apply to the camera image resolution.
|
|
Defaults to 1.0.
|
|
num_rays_per_chunk: Number of rays to use per chunk. Defaults to 4096.
|
|
"""
|
|
print("Creating trajectory video.")
|
|
images = []
|
|
cameras.rescale_output_resolution(rendered_resolution_scaling_factor)
|
|
for camera_idx in tqdm(range(cameras.size)):
|
|
camera_ray_bundle = cameras.generate_rays(camera_indices=camera_idx).to(pipeline.device)
|
|
camera_ray_bundle.num_rays_per_chunk = num_rays_per_chunk
|
|
with torch.no_grad():
|
|
outputs = pipeline.model.get_outputs_for_camera_ray_bundle(camera_ray_bundle)
|
|
image = outputs[rendered_output_name].cpu().numpy()
|
|
images.append(image)
|
|
|
|
seconds = 5.0
|
|
fps = len(images) / seconds
|
|
media.write_video(output_filename, images, fps=fps)
|
|
|
|
|
|
def _eval_setup(config_path: Path) -> Tuple[cfg.Config, Pipeline, Path]:
|
|
"""Shared setup for loading a saved pipeline for evaluation.
|
|
|
|
Args:
|
|
config_path: Path to config YAML file.
|
|
|
|
Returns:
|
|
Loaded config, pipeline module, and corresponding checkpoint.
|
|
"""
|
|
# load save config
|
|
config = yaml.load(config_path.read_text(), Loader=yaml.Loader)
|
|
assert isinstance(config, cfg.Config)
|
|
|
|
# load checkpoints from wherever they were saved
|
|
# TODO: expose the ability to choose an arbitrary checkpoint
|
|
config.trainer.load_dir = config.trainer.model_dir
|
|
config.pipeline.dataloader.eval_image_indices = None
|
|
|
|
# setup pipeline (which includes the dataloaders)
|
|
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
pipeline = config.pipeline.setup(device=device, test_mode=True)
|
|
assert isinstance(pipeline, Pipeline)
|
|
pipeline.eval()
|
|
|
|
# load checkpointed information
|
|
checkpoint_path = _load_checkpoint(config.trainer, pipeline)
|
|
|
|
return config, pipeline, checkpoint_path
|
|
|
|
|
|
@dataclass
|
|
class ComputePSNR:
|
|
"""Load a checkpoint, compute some PSNR metrics, and save to a JSON."""
|
|
|
|
# Path to config YAML file.
|
|
load_config: Path
|
|
# Name of the output file.
|
|
output_path: Path = Path("output.json")
|
|
|
|
def main(self) -> None:
|
|
"""Main function."""
|
|
config, pipeline, checkpoint_path = _eval_setup(self.load_config)
|
|
|
|
assert self.output_path.suffix == ".json"
|
|
stats_dict = _render_stats_dict(pipeline)
|
|
avg_psnr = stats_dict["avg psnr"]
|
|
avg_rays_per_sec = stats_dict["avg rays per sec"]
|
|
avg_fps = stats_dict["avg fps"]
|
|
print(f"Avg. PSNR: {avg_psnr:0.4f}")
|
|
print(f"Avg. Rays / Sec: {human_format(avg_rays_per_sec)}")
|
|
print(f"Avg. FPS: {avg_fps:0.4f}")
|
|
# save output to some file
|
|
self.output_path.parent.mkdir(parents=True, exist_ok=True)
|
|
benchmark_info = {
|
|
"experiment_name": config.experiment_name,
|
|
"method_name": config.method_name,
|
|
"checkpoint": str(checkpoint_path),
|
|
"results": stats_dict,
|
|
}
|
|
self.output_path.write_text(json.dumps(benchmark_info, indent=2), "utf8")
|
|
print(f"Saved results to: {self.output_path}")
|
|
|
|
|
|
@dataclasses.dataclass
|
|
class RenderTrajectory:
|
|
"""Load a checkpoint, render a trajectory, and save to a video file."""
|
|
|
|
# Path to config YAML file.
|
|
load_config: Path
|
|
# Name of the renderer output to use. rgb, depth, etc.
|
|
rendered_output_name: str = "rgb"
|
|
# Trajectory to render.
|
|
traj: Literal["spiral", "interp"] = "spiral"
|
|
# Scaling factor to apply to the camera image resolution.
|
|
rendered_resolution_scaling_factor: float = 1.0
|
|
# Name of the output file.
|
|
output_path: Path = Path("output.mp4")
|
|
|
|
def main(self) -> None:
|
|
"""Main function."""
|
|
config, pipeline, _ = _eval_setup(self.load_config)
|
|
|
|
# TODO(ethan): use camera information from parsing args
|
|
if self.traj == "spiral":
|
|
camera_start = pipeline.dataloader.eval_dataloader.get_camera(image_idx=0)
|
|
# TODO(ethan): pass in the up direction of the camera
|
|
camera_path = get_spiral_path(camera_start, steps=30, radius=0.1)
|
|
elif self.traj == "interp":
|
|
cameras = pipeline.dataloader.eval_dataloader.get_camera(image_idx=[0, 10])
|
|
camera_path = get_interpolated_camera_path(cameras, steps=30)
|
|
else:
|
|
assert_never(self.traj)
|
|
|
|
_render_trajectory_video(
|
|
pipeline,
|
|
camera_path,
|
|
output_filename=self.output_path,
|
|
rendered_output_name=self.rendered_output_name,
|
|
rendered_resolution_scaling_factor=self.rendered_resolution_scaling_factor,
|
|
num_rays_per_chunk=config.pipeline.dataloader.eval_num_rays_per_chunk,
|
|
)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
# A Union over dataclass types will create a subcommand for each type.
|
|
#
|
|
# TODO: it would make sense to split this script up into separate scripts.
|
|
# - To reduce duplicate code, some "upstream" refactor could also simplify the
|
|
# shared `_load_checkpoint()` and `_eval_setup()` helpers. The high-level
|
|
# operations implemented by each seem fairly universal; ideally the checkpoint
|
|
# loading logic, for example, would be the same as what's used for loading a
|
|
# checkpoint when resuming a training run.
|
|
dcargs.cli(Union[ComputePSNR, RenderTrajectory]).main()
|