File size: 10,521 Bytes
ec7f44e |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 |
import os
import json
import glob
import xml.etree.ElementTree as ET
import numpy as np
from PIL import Image
from torch.utils.data import Dataset, DataLoader
import torchvision.transforms as T
import torch
import torch.nn as nn
import torch.optim as optim
from shapely.geometry import Polygon
from pathlib import Path
# =====================
# Data Utils
# # =====================
import numpy as np
import json
def flat_corners_from_mockup(mockup_path):
"""
Returns 4 corners of print area from mockup.json
ordered TL, TR, BR, BL and normalized [0,1] w.r.t background.
"""
d = json.loads(Path(mockup_path).read_text())
bg_w = d["background"]["width"]
bg_h = d["background"]["height"]
area = d["printAreas"][0]
x, y = area["position"]["x"], area["position"]["y"]
w, h = area["width"], area["height"]
angle = area["rotation"]
cx, cy = x + w/2.0, y + h/2.0
# corners in px (TL,TR,BR,BL)
dx, dy = w/2.0, h/2.0
corners = np.array([[-dx, -dy], [dx, -dy], [dx, dy], [-dx, dy]], dtype=np.float32)
theta = np.deg2rad(angle)
R = np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]], dtype=np.float32)
rot = (corners @ R.T) + np.array([cx, cy], dtype=np.float32)
# normalize
norm = np.zeros_like(rot)
norm[:,0] = rot[:,0] / bg_w
norm[:,1] = rot[:,1] / bg_h
return rot.astype(np.float32), norm.astype(np.float32)
def parse_xml_points(xml_path):
"""
Parse the 4 corner points from the XML (FourPoint transform).
Returns normalized coordinates (TL, TR, BR, BL).
"""
tree = ET.parse(xml_path)
root = tree.getroot()
points = []
bg_w = int(root.find("background").get("width"))
bg_h = int(root.find("background").get("height"))
for transform in root.findall(".//transform"):
if transform.get("type") == "FourPoint":
for pt in ["TopLeft", "TopRight", "BottomRight", "BottomLeft"]:
node = transform.find(f".//point[@type='{pt}']")
if node is not None:
x = float(node.get("x")) / bg_w
y = float(node.get("y")) / bg_h
points.append([x, y])
break # only first transform
return np.array(points, dtype=np.float32) # (4,2)
class KP4Dataset(Dataset):
def __init__(self, root, img_size=512):
self.root = Path(root)
self.img_size = img_size
self.samples = []
# Transform pipeline (resize + tensor + normalize)
self.transform = T.Compose([
T.Resize((img_size, img_size)),
T.ToTensor(),
T.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
])
# Walk recursively
for xml_file in self.root.rglob("*.xml"):
if "_visual" not in xml_file.stem:
continue
# Find matching perspective image
base = xml_file.stem
img_file = None
for ext in [".png", ".jpg", ".jpeg"]:
cand = xml_file.with_suffix(ext)
if cand.exists():
img_file = cand
break
if img_file is None:
continue
# Flat image (background)
flat_img = xml_file.parent / (base.replace("_visual", "_background") + ".png")
if not flat_img.exists():
flat_img = xml_file.parent / (base.replace("_visual", "_background") + ".jpg")
if not flat_img.exists():
continue
# Mockup.json
json_file = xml_file.parent / "mockup.json"
if not json_file.exists():
continue
self.samples.append((img_file, xml_file, flat_img, json_file))
if not self.samples:
raise RuntimeError(f"No valid samples found under {root}")
def __len__(self):
return len(self.samples)
def __getitem__(self, idx):
img_file, xml_file, flat_img, json_file = self.samples[idx]
img = self.transform(Image.open(img_file).convert("RGB"))
flat = self.transform(Image.open(flat_img).convert("RGB"))
# flat points
_, flat_norm = flat_corners_from_mockup(json_file)
flat_pts = torch.tensor(flat_norm, dtype=torch.float32)
# perspective points
persp_norm = parse_xml_points(xml_file)
persp_pts = torch.tensor(persp_norm, dtype=torch.float32)
return {
"persp_img": img,
"flat_img": flat,
"flat_pts": flat_pts,
"persp_pts": persp_pts,
"xml": str(xml_file),
"json": str(json_file),
}
# =====================
# Model
# =====================
class SimpleTransformer(nn.Module):
def __init__(self, d_model=128, nhead=4, num_layers=2):
super().__init__()
self.fc_in = nn.Linear(8, d_model) # 4 corners * 2
encoder_layer = nn.TransformerEncoderLayer(d_model=d_model, nhead=nhead, batch_first=True)
self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
self.fc_out = nn.Linear(d_model, 8) # predict 4 corners (x,y)*4
def forward(self, x):
x = self.fc_in(x).unsqueeze(1) # (B,1,8)->(B,1,d_model)
x = self.transformer(x)
x = self.fc_out(x).squeeze(1) # (B,d_model)->(B,8)
return x
# =====================
# Metrics
# =====================
def mse_loss(pred, gt):
return ((pred-gt)**2).mean()
def mean_corner_error(pred, gt, img_w, img_h):
pred_px = pred * torch.tensor([img_w,img_h], device=pred.device)
gt_px = gt * torch.tensor([img_w,img_h], device=gt.device)
err = torch.norm(pred_px-gt_px, dim=-1).mean().item()
return err
def iou_quad(pred, gt):
pred_poly = Polygon(pred.tolist())
gt_poly = Polygon(gt.tolist())
if not pred_poly.is_valid or not gt_poly.is_valid:
return 0.0
inter = pred_poly.intersection(gt_poly).area
union = pred_poly.union(gt_poly).area
return inter/union if union > 0 else 0.0
# =====================
# Training
# =====================
def train_model(
train_root,
test_root,
epochs=20,
batch_size=8,
lr=1e-3,
img_size=256,
save_dir="Transformer/checkpoints",
resume_path=None
):
device = "cuda" if torch.cuda.is_available() else "cpu"
train_ds = KP4Dataset(train_root, img_size=img_size)
val_ds = KP4Dataset(test_root, img_size=img_size)
train_loader = DataLoader(train_ds, batch_size=batch_size, shuffle=True)
val_loader = DataLoader(val_ds, batch_size=1, shuffle=False)
model = SimpleTransformer().to(device)
optimizer = optim.Adam(model.parameters(), lr=lr)
start_epoch = 0
os.makedirs(save_dir, exist_ok=True)
# Resume Training
if resume_path is not None and os.path.exists(resume_path):
print(f"Loading checkpoint from {resume_path}")
checkpoint = torch.load(resume_path, map_location=device)
model.load_state_dict(checkpoint["model_state"])
optimizer.load_state_dict(checkpoint["optimizer_state"])
start_epoch = checkpoint["epoch"]
print(f"Resumed from epoch {start_epoch}")
# ===================== Track Best Model =====================
best_iou = -1.0
best_model_path = os.path.join(save_dir, "best_model.pth")
for epoch in range(start_epoch, epochs):
# -------- Training --------
model.train()
total_loss = 0
for batch in train_loader:
flat_pts = batch["flat_pts"].to(device)
persp_pts = batch["persp_pts"].to(device)
flat_pts_in = flat_pts.view(flat_pts.size(0), -1)
target = persp_pts.view(persp_pts.size(0), -1)
pred = model(flat_pts_in)
loss = mse_loss(pred, target)
optimizer.zero_grad()
loss.backward()
optimizer.step()
total_loss += loss.item()
print(f"Epoch {epoch+1}/{epochs} - Train Loss: {total_loss/len(train_loader):.6f}")
# -------- Validation --------
model.eval()
mse_all, ce_all, iou_all = [], [], []
with torch.no_grad():
for batch in val_loader:
flat_pts = batch["flat_pts"].to(device)
persp_pts = batch["persp_pts"].to(device)
flat_pts_in = flat_pts.view(1, -1)
target = persp_pts.view(1, -1)
pred = model(flat_pts_in)
mse_all.append(mse_loss(pred, target).item())
pred_quad = pred.view(4,2).cpu()
gt_quad = persp_pts.view(4,2).cpu()
w,h = batch["persp_img"].shape[2], batch["persp_img"].shape[1]
ce_all.append(mean_corner_error(pred_quad, gt_quad, w, h))
iou_all.append(iou_quad(pred_quad, gt_quad))
val_mse = np.mean(mse_all)
val_ce = np.mean(ce_all)
val_iou = np.mean(iou_all)
print(f" Val MSE: {val_mse:.6f}, CornerErr(px): {val_ce:.2f}, IoU: {val_iou:.3f}")
if (epoch + 1) % 100 == 0:
# -------- Save Epoch Checkpoint (like before) --------
checkpoint_path = os.path.join(save_dir, f"epoch_{epoch+1}.pth")
torch.save({
"epoch": epoch+1,
"model_state": model.state_dict(),
"optimizer_state": optimizer.state_dict(),
"val_iou": val_iou,
}, checkpoint_path)
print(f"Checkpoint saved: {checkpoint_path}")
# -------- Save Best Model --------
if val_iou > best_iou:
best_iou = val_iou
torch.save({
"epoch": epoch+1,
"model_state": model.state_dict(),
"optimizer_state": optimizer.state_dict(),
"best_iou": best_iou,
}, best_model_path)
print(f"Best model updated at epoch {epoch+1} (IoU={val_iou:.3f})")
# Save final model weights
final_path = os.path.join(save_dir, "final_model.pth")
torch.save(model.state_dict(), final_path)
print(f"Final model saved at {final_path}")
print(f"Best model saved at {best_model_path} with IoU={best_iou:.3f}")
return model
# =====================
# Main
# =====================
if __name__ == "__main__":
model = train_model(
train_root="Transformer/train",
test_root="Transformer/test",
epochs=3000,
batch_size=4,
lr=1e-3,
img_size=256,
resume_path=None
)
|