-
Notifications
You must be signed in to change notification settings - Fork 4
/
render_lightning.py
202 lines (163 loc) · 8.84 KB
/
render_lightning.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import os
from glob import glob
from pathlib import Path
import time
from typing import Any
import cv2
import imageio
from natsort import natsorted
import hydra
import numpy as np
import pandas as pd
from omegaconf import DictConfig, OmegaConf
import torch
import torchvision
import lightning as L
from tqdm import tqdm
from moviepy.editor import ImageSequenceClip, clips_array
from scene.dataset_readers import SceneType
from model.unc_2d_unet import Unc2DUnet
from utils.routines import load_from_model_path
from utils.pca import compute_feat_pca_from_renders
from utils.vis import resize_image
@hydra.main(version_base=None, config_path="conf", config_name="render")
def main(cfg_render : DictConfig) -> None:
model_path = cfg_render.model_path
model, scene, cfg = load_from_model_path(
model_path, source_path=cfg_render.source_path, simple_scene=True)
save_root = model_path
if cfg_render.eval_on_gg:
from model.gaussian_grouping import GaussianGrouping
from gaussian_grouping.render import visualize_obj
assert cfg_render.gg_ckpt_folder is not None, "gg_ckpt_folder must be specified"
print(f"Loading GaussainGrouping model from {cfg_render.gg_ckpt_folder}")
cfg.model.name = "gaussian_grouping"
cfg.model.gg_ckpt_folder = cfg_render.gg_ckpt_folder
model = GaussianGrouping(cfg, scene)
save_root = Path(cfg_render.gg_ckpt_folder)
if cfg_render.no_render_feat:
do_pca_render = False
else:
do_pca_render = cfg.model.dim_extra > 0
L.seed_everything(cfg.seed)
# Downsample a subset of the data and compute a PCA in the space of rendered features
if do_pca_render:
pca = compute_feat_pca_from_renders(scene, cfg_render.render_subset, [model])
pca = pca[0]
loader = scene.get_data_loader(cfg_render.render_subset, shuffle=False)
save_folder = os.path.join(save_root, f"render_{cfg_render.render_subset}")
gt_folder = os.path.join(save_folder, "gt")
os.makedirs(gt_folder, exist_ok=True)
render_folder = os.path.join(save_folder, "render")
os.makedirs(render_folder, exist_ok=True)
if do_pca_render:
feat_folder = os.path.join(save_folder, "feature")
os.makedirs(feat_folder, exist_ok=True)
if cfg_render.eval_on_gg:
pred_obj_folder = os.path.join(save_folder, "pred_obj")
os.makedirs(pred_obj_folder, exist_ok=True)
if isinstance(model, Unc2DUnet):
trans_mask_folder = os.path.join(save_folder, "unc_mask")
os.makedirs(trans_mask_folder, exist_ok=True)
print(f"Rendering to {save_folder}")
render_times = []
start = torch.cuda.Event(enable_timing=True)
end = torch.cuda.Event(enable_timing=True)
for batch_idx, batch in tqdm(enumerate(loader), total = len(loader)):
gt_image = batch['image'].to("cuda")[0] # (3, H, W)
subset = batch['subset'][0]
viewpoint_cam = scene.get_camera(batch['idx'].item(), subset=subset)
with torch.no_grad():
start.record()
render_pkg = model(viewpoint_cam, render_feature = do_pca_render)
end.record()
torch.cuda.synchronize()
render_times.append(start.elapsed_time(end))
render, gt_image_processed = scene.postProcess(render_pkg["render"], gt_image, viewpoint_cam)
render = render.clamp(0.0, 1.0) # (3, H, W)
# print(render.shape)
if scene.scene_type == SceneType.ARIA:
render_vis = np.rot90(render.permute(1, 2, 0).cpu().numpy(), k=-1) # (H, W, 3)
else:
render_vis = render.permute(1, 2, 0).cpu().numpy()
# print(render_vis.shape)
render_vis = resize_image(render_vis, cfg_render.image_size)
# print(render_vis.shape)
render_vis = (render_vis * 255).astype(np.uint8)
imageio.imwrite(os.path.join(render_folder, '{0:05d}'.format(batch_idx) + f".{cfg_render.save_ext}"), render_vis)
if do_pca_render:
render_feat = render_pkg["render_features"].permute(1, 2, 0).cpu().numpy() # (H, W, D)
feat_shape = render_feat.shape[:2]
render_feat = render_feat.reshape(-1, render_feat.shape[-1]) # (N, D)
render_pca = pca.transform(render_feat) # (N, 3)
render_pca = render_pca.reshape(feat_shape[0], feat_shape[1], 3) # (H, W, 3)
if scene.scene_type == SceneType.ARIA:
render_pca_vis = np.rot90(render_pca, k=-1) # (W, H, 3)
else:
render_pca_vis = render_pca
render_pca_vis = resize_image(render_pca_vis, cfg_render.image_size)
render_pca_vis = (render_pca_vis * 255).astype(np.uint8)
imageio.imwrite(os.path.join(feat_folder, '{0:05d}'.format(batch_idx) + f".{cfg_render.save_ext}"), render_pca_vis)
if cfg_render.eval_on_gg:
logits = model.classifier(render_pkg["render_features"])
pred_obj = torch.argmax(logits,dim=0)
pred_obj_mask = visualize_obj(pred_obj.cpu().numpy().astype(np.uint8))
if scene.scene_type == SceneType.ARIA:
pred_obj_mask = np.rot90(pred_obj_mask, k=-1) # (W, H, 3)
else:
pred_obj_mask = pred_obj_mask
pred_obj_mask = resize_image(pred_obj_mask, cfg_render.image_size)
imageio.imwrite(os.path.join(pred_obj_folder, '{0:05d}'.format(batch_idx) + f".{cfg_render.save_ext}"), pred_obj_mask)
# Convert tensor to numpy and rotate
gt_image_vis = gt_image_processed.permute(1, 2, 0).cpu().numpy()
if scene.scene_type == SceneType.ARIA:
gt_image_vis = np.rot90(gt_image_vis, k=-1)
else:
gt_image_vis = gt_image_vis
gt_image_vis = resize_image(gt_image_vis, cfg_render.image_size)
if isinstance(model, Unc2DUnet):
with torch.no_grad():
trans_mask = model.get_unc_mask(batch)
if scene.scene_type == SceneType.ARIA:
trans_mask_vis = np.rot90(trans_mask.squeeze(0).cpu().numpy(), k=-1)
else:
trans_mask_vis = trans_mask.squeeze(0).cpu().numpy()
trans_mask_vis = resize_image(trans_mask_vis, cfg_render.image_size)
trans_mask_vis = cv2.applyColorMap(255 - (trans_mask_vis * 255).astype(np.uint8), cv2.COLORMAP_TURBO)
trans_mask_vis = trans_mask_vis.astype(gt_image_vis.dtype) / 255.0
trans_mask_vis = cv2.addWeighted(gt_image_vis, 0.5, trans_mask_vis, 0.5, 0)
trans_mask_vis = (trans_mask_vis * 255).astype(np.uint8)
imageio.imwrite(os.path.join(trans_mask_folder, '{0:05d}'.format(batch_idx) + f".{cfg_render.save_ext}"), trans_mask_vis)
gt_image_vis = (gt_image_vis * 255).astype(np.uint8)
imageio.imwrite(os.path.join(gt_folder, '{0:05d}'.format(batch_idx) + f".{cfg_render.save_ext}"), gt_image_vis)
# # Instead of duplicate the images, create a soft link
# src = viewpoint_cam.image_path
# gt_ext = src.split(".")[-1]
# dst = os.path.join(gt_folder, '{0:05d}'.format(batch_idx) + f".{gt_ext}")
# if not os.path.exists(dst):
# os.symlink(src, dst)
# torchvision.utils.save_image(render_rgb, os.path.join(render_folder, '{0:05d}'.format(batch_idx) + f".{cfg_render.save_ext}"))
print(f"Average render time: {np.mean(render_times):.2f} ms")
# Concatenate the saved images into a video
print(f"Concating to video")
gt_paths = natsorted(glob(os.path.join(gt_folder, f"*.{cfg_render.save_ext}")))
clip_gt = ImageSequenceClip(gt_paths, fps=cfg_render.fps)
render_paths = natsorted(glob(os.path.join(render_folder, f"*.{cfg_render.save_ext}")))
clip_render = ImageSequenceClip(render_paths, fps=cfg_render.fps)
clips = [[clip_gt, clip_render]]
if do_pca_render:
feat_paths = natsorted(glob(os.path.join(feat_folder, f"*.{cfg_render.save_ext}")))
clip_feat = ImageSequenceClip(feat_paths, fps=cfg_render.fps)
clips[0].append(clip_feat)
if cfg_render.eval_on_gg:
pred_obj_paths = natsorted(glob(os.path.join(pred_obj_folder, f"*.{cfg_render.save_ext}")))
clip_pred_obj = ImageSequenceClip(pred_obj_paths, fps=cfg_render.fps)
clips[0].append(clip_pred_obj)
clip_combined = clips_array(clips)
clip_combined.write_videofile(os.path.join(save_folder, f"combined.mp4"))
if __name__ == "__main__":
main()