-
Notifications
You must be signed in to change notification settings - Fork 17
/
Copy pathevaluate.py
109 lines (102 loc) · 5.58 KB
/
evaluate.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
import argparse
from models import AFE, CKD, HPE_EDE, MFE, Generator
import numpy as np
import torch
import torch.nn.functional as F
import imageio
import os
from skimage import io, img_as_float32
from utils import transform_kp, transform_kp_with_new_pose
@torch.no_grad()
def eval(args):
g_models = {"afe": AFE(), "ckd": CKD(), "hpe_ede": HPE_EDE(), "mfe": MFE(), "generator": Generator()}
ckp_path = os.path.join(args.ckp_dir, "%s-checkpoint.pth.tar" % str(args.ckp).zfill(8))
checkpoint = torch.load(ckp_path, map_location=torch.device("cpu"))
for k, v in g_models.items():
v.cuda()
v.load_state_dict(checkpoint[k])
v.eval()
output_frames = []
if args.source == "r":
frames = sorted(os.listdir(args.driving))[: args.num_frames]
num_frames = len(frames)
video_array = [img_as_float32(io.imread(os.path.join(args.driving, frames[idx]))) for idx in range(num_frames)]
s = np.array(video_array[0], dtype="float32").transpose((2, 0, 1))
s = torch.from_numpy(s).cuda().unsqueeze(0)
fs = g_models["afe"](s)
kp_c = g_models["ckd"](s)
yaw_s, pitch_s, roll_s, t_s, delta_s = g_models["hpe_ede"](s)
kp_s, Rs = transform_kp(kp_c, yaw_s, pitch_s, roll_s, t_s, delta_s)
for img in video_array[1:]:
img = np.array(img, dtype="float32").transpose((2, 0, 1))
img = torch.from_numpy(img).cuda().unsqueeze(0)
yaw, pitch, roll, t, delta = g_models["hpe_ede"](img)
kp_d, Rd = transform_kp(kp_c, yaw, pitch, roll, t, delta)
deformation, occlusion = g_models["mfe"](fs, kp_s, kp_d, Rs, Rd)
generated_d = g_models["generator"](fs, deformation, occlusion)
generated_d = torch.cat((img, generated_d), dim=3)
# generated_d = F.interpolate(generated_d, scale_factor=0.5)
generated_d = generated_d.squeeze(0).data.cpu().numpy()
generated_d = np.transpose(generated_d, [1, 2, 0])
generated_d = generated_d.clip(0, 1)
generated_d = (255 * generated_d).astype(np.uint8)
output_frames.append(generated_d)
elif args.source == "f":
frames = sorted(os.listdir(args.driving))[: args.num_frames]
num_frames = len(frames)
video_array = [img_as_float32(io.imread(os.path.join(args.driving, frames[idx]))) for idx in range(num_frames)]
for img in video_array:
img = np.array(img, dtype="float32").transpose((2, 0, 1))
img = torch.from_numpy(img).cuda().unsqueeze(0)
fs = g_models["afe"](img)
kp_c = g_models["ckd"](img)
yaw, pitch, roll, t, delta = g_models["hpe_ede"](img)
kp_s, Rs = transform_kp(kp_c, yaw, pitch, roll, t, delta)
kp_d, Rd = transform_kp_with_new_pose(kp_c, yaw, pitch, roll, t, delta, 0 * yaw, 0 * pitch, 0 * roll)
deformation, occlusion = g_models["mfe"](fs, kp_s, kp_d, Rs, Rd)
generated_d = g_models["generator"](fs, deformation, occlusion)
generated_d = torch.cat((img, generated_d), dim=3)
# generated_d = F.interpolate(generated_d, scale_factor=0.5)
generated_d = generated_d.squeeze(0).data.cpu().numpy()
generated_d = np.transpose(generated_d, [1, 2, 0])
generated_d = generated_d.clip(0, 1)
generated_d = (255 * generated_d).astype(np.uint8)
output_frames.append(generated_d)
else:
s = img_as_float32(io.imread(args.source))[:, :, :3]
s = np.array(s, dtype="float32").transpose((2, 0, 1))
s = torch.from_numpy(s).cuda().unsqueeze(0)
s = F.interpolate(s, size=(256, 256))
fs = g_models["afe"](s)
kp_c = g_models["ckd"](s)
yaw, pitch, roll, t, delta = g_models["hpe_ede"](s)
kp_s, Rs = transform_kp(kp_c, yaw, pitch, roll, t, delta)
frames = sorted(os.listdir(args.driving))[: args.num_frames]
num_frames = len(frames)
video_array = [img_as_float32(io.imread(os.path.join(args.driving, frames[idx]))) for idx in range(num_frames)]
for img in video_array:
img = np.array(img, dtype="float32").transpose((2, 0, 1))
img = torch.from_numpy(img).cuda().unsqueeze(0)
yaw, pitch, roll, t, delta = g_models["hpe_ede"](img)
kp_d, Rd = transform_kp(kp_c, yaw, pitch, roll, t, delta)
deformation, occlusion = g_models["mfe"](fs, kp_s, kp_d, Rs, Rd)
generated_d = g_models["generator"](fs, deformation, occlusion)
generated_d = torch.cat((img, generated_d), dim=3)
generated_d = generated_d.squeeze(0).data.cpu().numpy()
generated_d = np.transpose(generated_d, [1, 2, 0])
generated_d = generated_d.clip(0, 1)
generated_d = (255 * generated_d).astype(np.uint8)
output_frames.append(generated_d)
imageio.mimsave(args.output, output_frames)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="face-vid2vid")
def str2bool(s):
return s.lower().startswith("t")
parser.add_argument("--ckp_dir", type=str, default="ckp", help="Checkpoint dir")
parser.add_argument("--output", type=str, default="output.gif", help="Output video")
parser.add_argument("--ckp", type=int, default=0, help="Checkpoint epoch")
parser.add_argument("--source", type=str, default="r", help="Source image, f for face frontalization, r for reconstruction")
parser.add_argument("--driving", type=str, help="Driving dir")
parser.add_argument("--num_frames", type=int, default=90, help="Number of frames")
args = parser.parse_args()
eval(args)