Spaces:
Running
on
Zero
Running
on
Zero
Upload network_inference.py
Browse files
src/pixel3dmm/network_inference.py
ADDED
@@ -0,0 +1,191 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import traceback
|
2 |
+
|
3 |
+
from tqdm import tqdm
|
4 |
+
import os
|
5 |
+
import torch
|
6 |
+
import numpy as np
|
7 |
+
from PIL import Image
|
8 |
+
from omegaconf import OmegaConf
|
9 |
+
from time import time
|
10 |
+
|
11 |
+
from pixel3dmm.utils.uv import uv_pred_to_mesh
|
12 |
+
from pixel3dmm.lightning.p3dmm_system import system as p3dmm_system
|
13 |
+
#from pixel3dmm.lightning.system_flame_params_legacy import system as system_flame_params_legacy
|
14 |
+
from pixel3dmm import env_paths
|
15 |
+
|
16 |
+
|
17 |
+
def pad_to_3_channels(img):
|
18 |
+
if img.shape[-1] == 3:
|
19 |
+
return img
|
20 |
+
elif img.shape[-1] == 1:
|
21 |
+
return np.concatenate([img, np.zeros_like(img[..., :1]), np.zeros_like(img[..., :1])], axis=-1)
|
22 |
+
elif img.shape[-1] == 2:
|
23 |
+
return np.concatenate([img, np.zeros_like(img[..., :1])], axis=-1)
|
24 |
+
else:
|
25 |
+
raise ValueError('too many dimensions in prediction type!')
|
26 |
+
|
27 |
+
def gaussian_fn(M, std):
|
28 |
+
n = torch.arange(0, M) - (M - 1.0) / 2.0
|
29 |
+
sig2 = 2 * std * std
|
30 |
+
w = torch.exp(-n ** 2 / sig2)
|
31 |
+
return w
|
32 |
+
|
33 |
+
|
34 |
+
def gkern(kernlen=256, std=128):
|
35 |
+
"""Returns a 2D Gaussian kernel array."""
|
36 |
+
gkern1d_x = gaussian_fn(kernlen, std=std * 5)
|
37 |
+
gkern1d_y = gaussian_fn(kernlen, std=std)
|
38 |
+
gkern2d = torch.outer(gkern1d_y, gkern1d_x)
|
39 |
+
return gkern2d
|
40 |
+
|
41 |
+
|
42 |
+
valid_verts = np.load(f'{env_paths.VALID_VERTICES_WIDE_REGION}')
|
43 |
+
|
44 |
+
|
45 |
+
def normals_n_uvs(cfg, model):
|
46 |
+
if cfg.model.prediction_type == 'flame_params':
|
47 |
+
cfg.data.mirror_aug = False
|
48 |
+
|
49 |
+
# data loader
|
50 |
+
if cfg.model.feature_map_type == 'DINO':
|
51 |
+
feature_map_size = 32
|
52 |
+
elif cfg.model.feature_map_type == 'sapiens':
|
53 |
+
feature_map_size = 64
|
54 |
+
|
55 |
+
batch_size = 1 # cfg.inference_batch_size
|
56 |
+
|
57 |
+
prediction_types = cfg.model.prediction_type.split(',')
|
58 |
+
|
59 |
+
conv = torch.nn.Conv2d(in_channels=1, out_channels=1, kernel_size=11, bias=False, padding='same')
|
60 |
+
g_weights = gkern(11, 2)
|
61 |
+
g_weights /= torch.sum(g_weights)
|
62 |
+
conv.weight = torch.nn.Parameter(g_weights.unsqueeze(0).unsqueeze(0))
|
63 |
+
|
64 |
+
OUT_NAMES = str(cfg.video_name).split(',')
|
65 |
+
|
66 |
+
print(f"""
|
67 |
+
<<<<<<<< STARTING PIXEL3DMM INFERENCE for {cfg.video_name} in {prediction_types} MODE >>>>>>>>
|
68 |
+
""")
|
69 |
+
|
70 |
+
for OUT_NAME in OUT_NAMES:
|
71 |
+
folder = f'{env_paths.PREPROCESSED_DATA}/{OUT_NAME}/'
|
72 |
+
IMAGE_FOLDER = f'{folder}/cropped'
|
73 |
+
SEGMENTATION_FOLDER = f'{folder}/seg_og/'
|
74 |
+
|
75 |
+
out_folders = {}
|
76 |
+
out_folders_wGT = {}
|
77 |
+
out_folders_viz = {}
|
78 |
+
|
79 |
+
for prediction_type in prediction_types:
|
80 |
+
out_folders[prediction_type] = f'{env_paths.PREPROCESSED_DATA}/{OUT_NAME}/p3dmm/{prediction_type}/'
|
81 |
+
out_folders_wGT[prediction_type] = f'{env_paths.PREPROCESSED_DATA}/{OUT_NAME}/p3dmm_wGT/{prediction_type}/'
|
82 |
+
os.makedirs(out_folders[prediction_type], exist_ok=True)
|
83 |
+
os.makedirs(out_folders_wGT[prediction_type], exist_ok=True)
|
84 |
+
out_folders_viz[prediction_type] = f'{env_paths.PREPROCESSED_DATA}/{OUT_NAME}/p3dmm_extraViz/{prediction_type}/'
|
85 |
+
os.makedirs(out_folders_viz[prediction_type], exist_ok=True)
|
86 |
+
|
87 |
+
|
88 |
+
image_names = os.listdir(f'{IMAGE_FOLDER}')
|
89 |
+
image_names.sort()
|
90 |
+
|
91 |
+
if os.path.exists(out_folders[prediction_type]):
|
92 |
+
if len(os.listdir(out_folders[prediction_type])) == len(image_names):
|
93 |
+
return
|
94 |
+
|
95 |
+
for i in tqdm(range(len(image_names))):
|
96 |
+
try:
|
97 |
+
img = np.array(Image.open(f'{IMAGE_FOLDER}/{image_names[i]}').resize((512, 512))) / 255 # need 512,512 images as input; normalize to [0, 1] range
|
98 |
+
img = torch.from_numpy(img)[None, None].float().cuda() # 1,1,512,512,3
|
99 |
+
img_seg = np.array(Image.open(f'{SEGMENTATION_FOLDER}/{image_names[i][:-4]}.png').resize((512, 512), Image.NEAREST))
|
100 |
+
if len(img_seg.shape) == 3:
|
101 |
+
img_seg = img_seg[..., 0]
|
102 |
+
#img_seg = np.array(Image.open(f'{SEGEMNTATION_FOLDER}/{int(image_names[i][:-4])*3:05d}.png').resize((512, 512), Image.NEAREST))
|
103 |
+
mask = ((img_seg == 2) | ((img_seg > 3) & (img_seg < 14)) ) & ~(img_seg==11)
|
104 |
+
mask = torch.from_numpy(mask).long().cuda()[None, None] # 1, 1, 512, 512
|
105 |
+
#mask = torch.ones_like(img[..., 0]).cuda().bool()
|
106 |
+
batch = {
|
107 |
+
'tar_msk': mask,
|
108 |
+
'tar_rgb': img,
|
109 |
+
}
|
110 |
+
batch_mirrored = {
|
111 |
+
'tar_rgb': torch.flip(batch['tar_rgb'], dims=[3]).cuda(),
|
112 |
+
'tar_msk': torch.flip(batch['tar_msk'], dims=[3]).cuda(),
|
113 |
+
}
|
114 |
+
|
115 |
+
with torch.no_grad():
|
116 |
+
output, conf = model.net(batch)
|
117 |
+
output_mirrored, conf = model.net(batch_mirrored)
|
118 |
+
|
119 |
+
if 'uv_map' in output:
|
120 |
+
fliped_uv_pred = torch.flip(output_mirrored['uv_map'], dims=[4])
|
121 |
+
fliped_uv_pred[:, :, 0, :, :] *= -1
|
122 |
+
fliped_uv_pred[:, :, 0, :, :] += 2*0.0075
|
123 |
+
output['uv_map'] = (output['uv_map'] + fliped_uv_pred)/2
|
124 |
+
if 'normals' in output:
|
125 |
+
fliped_uv_pred = torch.flip(output_mirrored['normals'], dims=[4])
|
126 |
+
fliped_uv_pred[:, :, 0, :, :] *= -1
|
127 |
+
output['normals'] = (output['normals'] + fliped_uv_pred)/2
|
128 |
+
if 'disps' in output:
|
129 |
+
fliped_uv_pred = torch.flip(output_mirrored['disps'], dims=[4])
|
130 |
+
fliped_uv_pred[:, :, 0, :, :] *= -1
|
131 |
+
output['disps'] = (output['disps'] + fliped_uv_pred)/2
|
132 |
+
|
133 |
+
|
134 |
+
|
135 |
+
for prediction_type in prediction_types:
|
136 |
+
for i_batch in range(batch_size):
|
137 |
+
|
138 |
+
i_view = 0
|
139 |
+
gt_rgb = batch['tar_rgb']
|
140 |
+
|
141 |
+
# normalize to [0,1] range
|
142 |
+
if prediction_type == 'uv_map':
|
143 |
+
tmp_output = torch.clamp((output[prediction_type][i_batch, i_view] + 1) / 2, 0, 1)
|
144 |
+
elif prediction_type == 'disps':
|
145 |
+
tmp_output = torch.clamp((output[prediction_type][i_batch, i_view] + 50) / 100, 0, 1)
|
146 |
+
elif prediction_type in ['normals', 'normals_can']:
|
147 |
+
tmp_output = output[prediction_type][i_batch, i_view]
|
148 |
+
tmp_output = tmp_output / torch.norm(tmp_output, dim=0).unsqueeze(0)
|
149 |
+
tmp_output = torch.clamp((tmp_output + 1) / 2, 0, 1)
|
150 |
+
# undo "weird" convention of normals that I used for preprocessing
|
151 |
+
tmp_output = torch.stack(
|
152 |
+
[tmp_output[0, ...], 1 - tmp_output[2, ...], 1 - tmp_output[1, ...]],
|
153 |
+
dim=0)
|
154 |
+
|
155 |
+
|
156 |
+
content = [
|
157 |
+
gt_rgb[i_batch, i_view].detach().cpu().numpy(),
|
158 |
+
pad_to_3_channels(tmp_output.permute(1, 2, 0).detach().cpu().float().numpy()),
|
159 |
+
]
|
160 |
+
|
161 |
+
catted = (np.concatenate(content, axis=1) * 255).astype(np.uint8)
|
162 |
+
Image.fromarray(catted).save(f'{out_folders_wGT[prediction_type]}/{image_names[i]}')
|
163 |
+
|
164 |
+
|
165 |
+
Image.fromarray(
|
166 |
+
pad_to_3_channels(
|
167 |
+
tmp_output.permute(1, 2, 0).detach().cpu().float().numpy() * 255).astype(
|
168 |
+
np.uint8)).save(
|
169 |
+
f'{out_folders[prediction_type]}/{image_names[i][:-4]}.png')
|
170 |
+
|
171 |
+
|
172 |
+
# this visulization is quite slow, therefore disable it per default
|
173 |
+
if prediction_type == 'uv_map' and cfg.viz_uv_mesh:
|
174 |
+
to_show_non_mirr = uv_pred_to_mesh(
|
175 |
+
output[prediction_type][i_batch:i_batch + 1, ...],
|
176 |
+
batch['tar_msk'][i_batch:i_batch + 1, ...],
|
177 |
+
batch['tar_rgb'][i_batch:i_batch + 1, ...],
|
178 |
+
right_ear = [537, 1334, 857, 554, 941],
|
179 |
+
left_ear = [541, 476, 237, 502, 286],
|
180 |
+
)
|
181 |
+
|
182 |
+
Image.fromarray(to_show_non_mirr).save(f'{out_folders_viz[prediction_type]}/{image_names[i]}')
|
183 |
+
|
184 |
+
|
185 |
+
except Exception:
|
186 |
+
traceback.print_exc()
|
187 |
+
|
188 |
+
print(f"""
|
189 |
+
<<<<<<<< FINISHED PIXEL3DMM INFERENCE for {cfg.video_name} in {prediction_types} MODE >>>>>>>>
|
190 |
+
""")
|
191 |
+
|