File size: 5,625 Bytes
d487538 | 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 | import os
import cv2
import numpy as np
import torch
import torch.nn.functional as F
import face_alignment
import lpips
import pandas as pd
from unet_acc import DenseMotion, UNetGenerator, warp_image
device = "cuda" if torch.cuda.is_available() else "cpu"
dense_motion = DenseMotion(kp_channels=68).to(device)
generator = UNetGenerator(in_channels=4).to(device)
ckpt = torch.load("checkpoints/best.pth", map_location=device)
dense_motion.load_state_dict(ckpt["dense_motion"])
generator.load_state_dict(ckpt["generator"])
dense_motion.eval()
generator.eval()
lpips_fn = lpips.LPIPS(net="alex").to(device)
fa = face_alignment.FaceAlignment(
face_alignment.LandmarksType.TWO_D,
device=device
)
# Metrics
def landmark_distance(pred, gt):
pred = cv2.cvtColor(pred, cv2.COLOR_GRAY2RGB)
gt = cv2.cvtColor(gt, cv2.COLOR_GRAY2RGB)
pl = fa.get_landmarks(pred)
gl = fa.get_landmarks(gt)
if pl is None or gl is None:
return None
pl, gl = pl[0], gl[0]
eye_dist = np.linalg.norm(gl[36] - gl[45]) + 1e-6
return np.mean(np.linalg.norm(pl - gl, axis=1)) / eye_dist
def lpips_score(pred, gt):
pred = pred.repeat(1, 3, 1, 1)
gt = gt.repeat(1, 3, 1, 1)
return lpips_fn(pred, gt).item()
def l1_score(pred, gt):
return F.l1_loss(pred, gt).item()
def temporal_jitter(frames):
diffs = []
for i in range(1, len(frames)):
diffs.append(torch.mean(torch.abs(frames[i] - frames[i - 1])).item())
return np.std(diffs), np.mean(diffs)
LOCK_IDXS = list(range(36, 48)) + list(range(48, 68))
def infer_no_warp(src):
B, _, H, W = src.shape
flow = torch.zeros(B, 2, H, W).to(device)
occ = torch.ones(B, 1, H, W).to(device)
return torch.clamp(generator(torch.cat([src, flow, occ], 1)), 0, 1)
def infer_warp(src, src_kp, drv_kp):
flow, occ = dense_motion(src_kp, drv_kp)
warped = warp_image(src, flow)
return torch.clamp(generator(torch.cat([warped, flow, occ], 1)), 0, 1)
def infer_warp_lock(src, src_kp, drv_kp):
kp = src_kp.clone()
kp[:, LOCK_IDXS] = drv_kp[:, LOCK_IDXS]
flow, occ = dense_motion(src_kp, kp)
warped = warp_image(src, flow)
return torch.clamp(generator(torch.cat([warped, flow, occ], 1)), 0, 1)
def infer_warp_lock_mask(src, src_kp, drv_kp, mask):
kp = src_kp.clone()
kp[:, LOCK_IDXS] = drv_kp[:, LOCK_IDXS]
flow, occ = dense_motion(src_kp, kp)
warped = warp_image(src, flow)
pred = generator(torch.cat([warped, flow, occ], 1))
return torch.clamp(pred * mask + src * (1 - mask), 0, 1)
def evaluate_sequence(src, src_kp, drv_kps, gt_frames, mask, mode):
preds_torch = []
lmd, lp, l1 = [], [], []
with torch.no_grad():
for t, drv_kp in enumerate(drv_kps):
if mode == "no_warp":
pred = infer_no_warp(src)
elif mode == "warp":
pred = infer_warp(src, src_kp, drv_kp)
elif mode == "warp_lock":
pred = infer_warp_lock(src, src_kp, drv_kp)
elif mode == "warp_lock_mask":
pred = infer_warp_lock_mask(src, src_kp, drv_kp, mask)
else:
raise ValueError
gt = gt_frames[t]
pred_np = (pred.detach().cpu().squeeze().numpy() * 255).astype(np.uint8)
gt_np = (gt.detach().cpu().squeeze().numpy() * 255).astype(np.uint8)
lm = landmark_distance(pred_np, gt_np)
if lm is not None:
lmd.append(lm)
lp.append(lpips_score(pred, gt))
l1.append(l1_score(pred, gt))
preds_torch.append(pred)
jit_std, _ = temporal_jitter(preds_torch)
return {
"LMD": np.mean(lmd) if len(lmd) > 0 else np.nan,
"LPIPS": np.mean(lp),
"Jitter": jit_std
}
def run_all(src, src_kp, drv_kps, gt_frames, mask):
rows = []
for mode in ["no_warp", "warp", "warp_lock", "warp_lock_mask"]:
print(f"Evaluating {mode}")
res = evaluate_sequence(src, src_kp, drv_kps, gt_frames, mask, mode)
res["Method"] = mode
rows.append(res)
df = pd.DataFrame(rows)
df = df[["Method", "LMD", "LPIPS", "Jitter"]]
df.to_csv("ablation_results.csv", index=False)
print(df)
if __name__ == "__main__":
src_img = cv2.imread(r"motion_transfer\new_dataset\test\dataset\87\frames\00000.jpg", cv2.IMREAD_GRAYSCALE)
src = torch.tensor(
src_img / 255.0,
dtype=torch.float32
).unsqueeze(0).unsqueeze(0).to(device)
src_kp = torch.tensor(
np.load(r"motion_transfer\new_dataset\test\dataset\87\combined\00000.npy"),
dtype=torch.float32
).permute(2, 0, 1).unsqueeze(0).to(device)
drv_kps = []
gt_frames = []
for f in sorted(os.listdir(r"motion_transfer\new_dataset\test\dataset\87\frames")):
gt = cv2.imread(os.path.join(r"motion_transfer\new_dataset\test\dataset\87\frames", f), cv2.IMREAD_GRAYSCALE)
gt_frames.append(
torch.tensor(
gt / 255.0,
dtype=torch.float32
).unsqueeze(0).unsqueeze(0).to(device)
)
kp = torch.tensor(
np.load(os.path.join(r"motion_transfer\new_dataset\test\dataset\87\combined", f.replace(".jpg", ".npy"))),
dtype=torch.float32
).permute(2, 0, 1).unsqueeze(0).to(device)
drv_kps.append(kp)
mask = torch.ones_like(src)
run_all(src, src_kp, drv_kps, gt_frames, mask)
|