merging main

This commit is contained in:
terrance committed 2023-05-10 21:26:04 +00:00
commit 80cdb73406
221 files changed
+50645 -2513

No files matched your search

+107
View File
@@ -0,0 +1,107 @@
# pylint: disable=all
from copy import copy
from typing import Any
import pytest
import torch
import yaml
from nerfstudio.cameras.cameras import Cameras
from nerfstudio.configs.base_config import InstantiateConfig
from nerfstudio.data.datamanagers.base_datamanager import (
DataManager,
DataparserOutputs,
VanillaDataManager,
VanillaDataManagerConfig,
)
from nerfstudio.data.datasets.base_dataset import InputDataset
from nerfstudio.data.datasets.depth_dataset import DepthDataset
class DummyDataParser:
def __init__(self, *args, **kwargs):
pass
def __getattr__(self, __name: str) -> Any:
if __name.startswith("_"):
return object.__getattribute__(self, __name)
return None
def get_dataparser_outputs(self, *args, **kwargs):
return DataparserOutputs(
[],
Cameras(
torch.ones((0, 3, 4)),
torch.ones((0, 1)),
torch.ones((0, 1)),
torch.ones((0, 1)),
torch.ones((0, 1)),
10,
10,
),
metadata={"depth_filenames": [], "depth_unit_scale_factor": 1.0},
)
@pytest.fixture
def config():
config = VanillaDataManagerConfig()
setattr(config, "dataparser", InstantiateConfig(_target=DummyDataParser))
setattr(config.dataparser, "data", None)
return config
def test_data_manager_type_inference(config):
# Mock for a faster test
assert VanillaDataManager(config).dataset_type is InputDataset
assert VanillaDataManager[DepthDataset](config).dataset_type is DepthDataset
class tmp(VanillaDataManager):
pass
assert tmp(config).dataset_type is InputDataset
class tmp2(VanillaDataManager[DepthDataset]):
pass
assert tmp2(config).dataset_type is DepthDataset
def test_data_manager_type_can_be_serialized(config):
# Mock for a faster test
assert VanillaDataManager(config).dataset_type is InputDataset
obj = yaml.load(yaml.dump(VanillaDataManager(config)), Loader=yaml.UnsafeLoader)
assert obj.dataset_type is InputDataset
assert isinstance(obj, VanillaDataManager)
assert VanillaDataManager[DepthDataset](config).dataset_type is DepthDataset
obj = yaml.load(yaml.dump(VanillaDataManager[DepthDataset](config)), Loader=yaml.UnsafeLoader)
assert obj.dataset_type is DepthDataset
assert isinstance(obj, VanillaDataManager)
class tmp(VanillaDataManager):
pass
try:
globals()["tmp"] = tmp
assert tmp(config).dataset_type is InputDataset
obj = yaml.load(yaml.dump(tmp(config)), Loader=yaml.UnsafeLoader)
assert obj.dataset_type is InputDataset
assert isinstance(obj, tmp)
finally:
globals().pop("tmp")
class tmp2(VanillaDataManager[DepthDataset]):
pass
try:
globals()["tmp2"] = tmp2
assert tmp2(config).dataset_type is DepthDataset
obj = yaml.load(yaml.dump(tmp2(config)), Loader=yaml.UnsafeLoader)
assert obj.dataset_type is DepthDataset
assert isinstance(obj, tmp2)
finally:
globals().pop("tmp2")
+26
View File
@@ -171,6 +171,32 @@ def test_tensor_hash_encoder():
assert encoded.shape == (10, out_dim)
def test_kplane_encoder():
"""Test K-Planes encoder"""
out_dim = 8
encoder = encodings.KPlanesEncoding(
resolution=(32, 32, 32, 16),
num_components=out_dim,
reduce="product",
)
assert encoder.get_out_dim() == out_dim
in_tensor = torch.randn((10, 4))
encoded = encoder(in_tensor)
assert encoded.shape == (10, out_dim)
encoder = encodings.KPlanesEncoding(
resolution=(32, 32, 32),
num_components=out_dim,
reduce="product",
)
in_tensor = torch.randn((10, 3))
encoded = encoder(in_tensor)
assert encoded.shape == (10, out_dim)
if __name__ == "__main__":
test_scaling_and_offset()
test_nerf_encoder()
+28
View File
@@ -0,0 +1,28 @@
"""
Test losses
"""
import torch
from nerfstudio.model_components.losses import tv_loss
def test_tv_loss():
"""Test total variation loss"""
grids = torch.ones([3, 64, 128, 128])
assert tv_loss(grids).item() == 0
grids = torch.zeros([1, 10, 100, 100])
for i in range(100):
for j in range(100):
grid_location = (i + j) % 2
for k in range(10):
grids[0, k, i, j] = grid_location
# TV_row = 1, TV_col = 1. Totol tv should be 2 * (TV_row + TV_col)
assert tv_loss(grids).item() == 4.0
if __name__ == "__main__":
test_tv_loss()
+1 -1
View File
@@ -40,7 +40,7 @@ def test_sh_renderer():
sh_renderer = renderers.SHRenderer()
rgb = sh_renderer(sh=sh, directions=directions, weights=weights)
assert torch.max(rgb) > 0.9
assert torch.max(rgb) > 0.7
def test_acc_renderer():
+20
View File
@@ -57,6 +57,9 @@ def test_load_state_dict():
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.register_parameter("param", nn.Parameter(torch.ones((3,))))
self.register_module(
"module", Model(*args, **kwargs)
) # make sure that non-DDP checkpoint won't be processed as DDP one
def load_state_dict(self, state_dict, strict=True):
"""Mocked load_state_dict"""
@@ -64,6 +67,7 @@ def test_load_state_dict():
was_called = True
assert strict
assert "param" in state_dict
assert "module.device_indicator_param" in state_dict
config = VanillaPipelineConfig(
datamanager=VanillaDataManagerConfig(
@@ -77,3 +81,19 @@ def test_load_state_dict():
pipeline.load_pipeline(state_dict, 0)
assert was_called
assert pipeline.model.param[0].item() == 2
# preparation for another test
state_dict["_model.param"].mul_(2) # pylint: disable=unsubscriptable-object
was_called = False
# pretends to be a DDP checkpoint
ddp_state_dict = {}
for key, value in state_dict.items():
if key.startswith("_model."):
# replace the prefix "_model" with "_model.module"
ddp_state_dict["_model.module." + key[len("_model.") :]] = value
else:
ddp_state_dict[key] = value
# load DDP checkpoint
pipeline.load_pipeline(ddp_state_dict, 0)
assert was_called
assert pipeline.model.param[0].item() == 4
+21
View File
@@ -1,6 +1,7 @@
"""
Tests for the nerfstudio.plugins.registry module.
"""
import os
import sys
from nerfstudio.engine.trainer import TrainerConfig
@@ -49,3 +50,23 @@ def test_discover_methods():
finally:
# Revert mock
registry.entry_points = entry_points_backup
def test_discover_methods_from_environment_variable():
"""Tests if a custom method from env variable gets properly registered using the discover_methods method"""
old_env = None
try:
old_env = os.environ.get("NERFSTUDIO_METHOD_CONFIGS", None)
os.environ["NERFSTUDIO_METHOD_CONFIGS"] = "test-method-env=test_registry:TestConfig"
# Discover plugins
methods, _ = registry.discover_methods()
assert "test-method-env" in methods
config = methods["test-method-env"]
assert isinstance(config, TrainerConfig)
finally:
# Revert mock
if old_env is not None:
os.environ["NERFSTUDIO_METHOD_CONFIGS"] = old_env
else:
del os.environ["NERFSTUDIO_METHOD_CONFIGS"]
+64
View File
@@ -12,10 +12,37 @@ from nerfstudio.data.dataparsers.nerfstudio_dataparser import NerfstudioDataPars
from nerfstudio.data.utils.colmap_parsing_utils import Camera
from nerfstudio.data.utils.colmap_parsing_utils import Image as ColmapImage
from nerfstudio.data.utils.colmap_parsing_utils import (
<<<<<<< HEAD
write_cameras_binary,
write_images_binary,
)
from scripts.process_data import ProcessImages
=======
qvec2rotmat,
write_cameras_binary,
write_images_binary,
)
from nerfstudio.process_data.images_to_nerfstudio_dataset import (
ImagesToNerfstudioDataset,
)
def random_quaternion(num_poses: int):
"""
Generates random rotation quatenion.
"""
u, v, w = np.random.uniform(size=(3, num_poses))
quaternion = np.stack(
(
np.sqrt(1 - u) * np.sin(2 * np.pi * v),
np.sqrt(1 - u) * np.cos(2 * np.pi * v),
np.sqrt(u) * np.sin(2 * np.pi * w),
np.sqrt(u) * np.cos(2 * np.pi * w),
),
-1,
)
return quaternion
>>>>>>> b1ddb9e695b789fae47e0f760a5c7f59f25af0bb
def test_process_images_skip_colmap(tmp_path: Path):
@@ -34,6 +61,7 @@ def test_process_images_skip_colmap(tmp_path: Path):
)
frames = {}
num_frames = 10
<<<<<<< HEAD
qvecs = np.random.uniform(size=(num_frames, 4))
tvecs = np.random.uniform(size=(num_frames, 3))
# original_poses = np.concatenate(
@@ -49,6 +77,23 @@ def test_process_images_skip_colmap(tmp_path: Path):
# ),
# -2,
# )
=======
qvecs = random_quaternion(num_frames)
tvecs = np.random.uniform(size=(num_frames, 3))
original_poses = np.concatenate(
(
np.concatenate(
(
np.stack(list(map(qvec2rotmat, qvecs))),
tvecs[:, :, None],
),
-1,
),
np.array([[[0, 0, 0, 1]]], dtype=qvecs.dtype).repeat(num_frames, 0),
),
-2,
)
>>>>>>> b1ddb9e695b789fae47e0f760a5c7f59f25af0bb
for i in range(num_frames):
frames[i + 1] = ColmapImage(i + 1, qvecs[i], tvecs[i], 1, f"image_{i}.png", [], [])
Image.new("RGB", (width, height)).save(tmp_path / "images" / f"image_{i}.png")
@@ -61,8 +106,15 @@ def test_process_images_skip_colmap(tmp_path: Path):
(tmp_path / "mocked_bin" / "colmap").touch(mode=0o777)
(tmp_path / "mocked_bin" / "ffmpeg").touch(mode=0o777)
<<<<<<< HEAD
# Run ProcessImages
cmd = ProcessImages(tmp_path / "images", tmp_path / "nerfstudio", colmap_model_path=sparse_path, skip_colmap=True)
=======
# Convert images into a NerfStudio dataset
cmd = ImagesToNerfstudioDataset(
data=tmp_path / "images", output_dir=tmp_path / "nerfstudio", colmap_model_path=sparse_path, skip_colmap=True
)
>>>>>>> b1ddb9e695b789fae47e0f760a5c7f59f25af0bb
cmd.main()
os.environ["PATH"] = old_path
@@ -77,4 +129,16 @@ def test_process_images_skip_colmap(tmp_path: Path):
outputs = parser.get_dataparser_outputs("train")
assert len(outputs.image_filenames) == 9
assert torch.is_tensor(outputs.dataparser_transform)
<<<<<<< HEAD
# TODO @jkulhanek: Add tests if the loaded poses are the same
=======
# Test if the original poses can be obtained back
gt_poses = original_poses[[0, 1, 2, 3, 4, 5, 6, 7, 9]]
dataparser_poses = outputs.transform_poses_to_original_space(outputs.cameras.camera_to_worlds, "opencv").numpy()
dataparser_poses = np.concatenate(
(dataparser_poses, np.array([[[0, 0, 0, 1]]]).repeat(len(dataparser_poses), 0)), 1
)
dataparser_poses = np.linalg.inv(dataparser_poses)
np.testing.assert_allclose(gt_poses, dataparser_poses, rtol=0, atol=1e-5)
>>>>>>> b1ddb9e695b789fae47e0f760a5c7f59f25af0bb
+7 -6
View File
@@ -13,7 +13,7 @@ from nerfstudio.configs.method_configs import method_configs
from nerfstudio.data.dataparsers.blender_dataparser import BlenderDataParserConfig
from nerfstudio.data.dataparsers.minimal_dataparser import MinimalDataParserConfig
from nerfstudio.engine.trainer import TrainerConfig
from scripts.train import train_loop
from nerfstudio.scripts.train import train_loop
BLACKLIST = [
"base",
@@ -21,13 +21,18 @@ BLACKLIST = [
"instant-ngp",
"instant-ngp-bounded",
"nerfacto",
"nerfacto-big",
"volinga",
"phototourism",
"depth-nerfacto",
"nerfplayer-ngp",
"nerfplayer-nerfacto",
"neus",
<<<<<<< HEAD
"dreamfusion",
=======
"neus-facto",
>>>>>>> b1ddb9e695b789fae47e0f760a5c7f59f25af0bb
]
@@ -50,11 +55,7 @@ def set_reduced_config(config: TrainerConfig):
if hasattr(config.pipeline.model, "num_importance_samples"):
config.pipeline.model.num_importance_samples = 4
# remove viewer
config.viewer.enable = False
# model specific config settings
if config.method_name == "instant-ngp":
config.pipeline.model.field_implementation = "torch"
config.viewer.quit_on_train_completion = True
return config
+5 -1
View File
@@ -168,7 +168,11 @@ def test_equall_nerfacc():
# time2 = time.time()
# time3 = time.time()
t_min_nerfacc, t_max_nerfacc = nerfacc.ray_aabb_intersect(origins, directions, aabb)
t_min_nerfacc, t_max_nerfacc, _ = nerfacc.ray_aabb_intersect(
origins, directions, aabb[None, :], near_plane=0, far_plane=1e10, miss_value=1e10
)
t_max_nerfacc = t_max_nerfacc.squeeze(-1)
t_min_nerfacc = t_min_nerfacc.squeeze(-1)
# time4 = time.time()
# print("pytorch ", time2-time1)