File size: 9,657 Bytes
fd05d86 |
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 |
# src/infer_with_centroid.py
import argparse
from pathlib import Path
import torch
import torch.nn.functional as F
from PIL import Image, ImageDraw, ImageFont
import torchvision.transforms as T
from config import CFG
from prompt import load_kiln_embed
from model import build_model
from head import SimilarityHead
# -----------------------------
# Transforms (match training)
# -----------------------------
import matplotlib.pyplot as plt
import numpy as np
def save_heatmap(sal, out_path="outputs/heatmap.png"):
sal_np = sal.squeeze().cpu().numpy() # [H, W]
plt.imshow(sal_np, cmap="jet", interpolation="nearest")
plt.colorbar()
plt.title("Attention Saliency Heatmap")
plt.savefig(out_path, bbox_inches="tight")
plt.close()
def save_heatmap_overlay(pil_img, sal, out_path="outputs/heatmap_overlay.png", alpha=0.45):
W0, H0 = pil_img.size
# upsample saliency to image size
sal_up = F.interpolate(sal, size=(H0, W0), mode="bilinear", align_corners=False)
h = sal_up.squeeze().detach().cpu().numpy() # [H0, W0]
# normalize to 0..1
h = (h - h.min()) / (h.max() - h.min() + 1e-6)
# colormap -> RGBA [H0, W0, 4]
cmap = plt.get_cmap("jet")
h_rgba = (cmap(h) * 255).astype(np.uint8)
heat = Image.fromarray(h_rgba).convert("RGBA")
base = pil_img.convert("RGBA")
overlay = Image.blend(base, heat, alpha)
overlay.save(out_path)
def model_transform():
return T.Compose([
T.Resize((CFG.image_size, CFG.image_size)),
T.ToTensor(),
T.ConvertImageDtype(torch.float32),
T.Normalize(mean=CFG.mean, std=CFG.std),
])
def display_transform():
return T.Resize((CFG.image_size, CFG.image_size))
def load_checkpoint(ckpt_path: Path):
return torch.load(ckpt_path, map_location="cpu")
def build_model_and_head(device, ckpt):
args = ckpt.get("args", {})
last_k = args.get("last_k", 2)
model_name = args.get("model_name", "ViT-B-16")
pretrained = args.get("pretrained", "openai")
model = build_model(
model_name=model_name,
pretrained=pretrained,
train_last_k_blocks=last_k,
train_final_norm=True,
finetune_logit_scale=False,
device=device,
)
model.load_state_dict(ckpt["model_state"], strict=False)
head = SimilarityHead(init_alpha=10.0, init_beta=0.0).to(device)
head.load_state_dict(ckpt["head_state"], strict=False)
model.eval(); head.eval()
return model, head
def find_openclip_visual_tower(model: torch.nn.Module):
for m in model.modules():
if hasattr(m, "transformer") and hasattr(m.transformer, "resblocks"):
if hasattr(m, "class_embedding") and hasattr(m, "positional_embedding"):
return m
raise AttributeError("Could not find a ViT visual tower (transformer.resblocks) inside the model. Debugger")
def _patch_block_attention_capture(blk):
if hasattr(blk, "_orig_attention"):
return
blk._orig_attention = blk.attention
def attention_with_capture(q_x: torch.Tensor, k_x: torch.Tensor = None, v_x: torch.Tensor = None,
attn_mask: torch.Tensor = None, **kwargs):
k_x = k_x if k_x is not None else q_x
v_x = v_x if v_x is not None else q_x
#will have to keep dtype compatible with MHA here for future ref
attn_mask2 = attn_mask.to(q_x.dtype) if attn_mask is not None else None
out, w = blk.attn(
q_x, k_x, v_x,
need_weights=True,
average_attn_weights=False,
attn_mask=attn_mask2
)
if w.dim() == 4:
blk.__last_attn__ = w.detach()
elif w.dim() == 3:
# normalize to [B, T, S]
if w.shape[0] == q_x.shape[0]:
w_bts = w
elif w.shape[1] == q_x.shape[0]:
w_bts = w.permute(1, 0, 2)
else:
w_bts = w.unsqueeze(0) if w.shape[0] != 0 else w
blk.__last_attn__ = w_bts.unsqueeze(1).detach() # -> [B, 1, T, S] implicit single head
else:
blk.__last_attn__ = None
return out
blk.attention = attention_with_capture
def _enable_attn_capture(visual):
for blk in visual.transformer.resblocks:
_patch_block_attention_capture(blk)
@torch.no_grad()
def vit_centroid_from_attention(model, x):
B = x.shape[0]
visual = find_openclip_visual_tower(model)
_enable_attn_capture(visual)
_ = model.encode_image(x) # forward passing rn
A = []
for blk in visual.transformer.resblocks:
w = getattr(blk, "__last_attn__", None)
if w is None:
continue
if w.dim() == 4 and w.shape[0] in (1, B):
a = w.mean(1)
else:
if w.dim() == 3 and w.shape[0] == B:
a = w
elif w.dim() == 3 and w.shape[1] == B:
a = w.permute(1, 0, 2)
else:
continue
a = a / (a.sum(-1, keepdim=True) + 1e-6) # [B, T, S]
Ttok = a.size(-1)
I = torch.eye(Ttok, device=a.device).unsqueeze(0).expand(a.size(0), Ttok, Ttok)
a = 0.1 * I + 0.9 * a
A.append(a)
if not A:
raise RuntimeError("No attention captured. Backend likely changed shapes; print shapes inside attention_with_capture to inspect.")
# joint attention will be here
R = A[0]
for a in A[1:]:
R = a @ R
cls_to_tokens = R[:, 0, 1:]
if hasattr(visual, "patch_size"):
ph, pw = (visual.patch_size if isinstance(visual.patch_size, tuple) else (visual.patch_size, visual.patch_size))
else:
ph = pw = 16
B, _, H_img, W_img = x.shape
H = H_img // ph
W = W_img // pw
sal = cls_to_tokens.view(B, 1, H, W)
sal = F.relu(sal)
sal = sal / (sal.sum(dim=[2, 3], keepdim=True) + 1e-6)
ys = torch.arange(H, device=sal.device).view(1, 1, H, 1).float()
xs = torch.arange(W, device=sal.device).view(1, 1, 1, W).float()
cy = (sal * ys).sum(dim=[2, 3])
cx = (sal * xs).sum(dim=[2, 3])
cx_px = (cx + 0.5) * pw
cy_px = (cy + 0.5) * ph
return cx_px.squeeze(1), cy_px.squeeze(1), sal
def annotate_text(pil_img, text):
img = pil_img.convert("RGB").copy()
draw = ImageDraw.Draw(img)
try:
font = ImageFont.truetype("arial.ttf", 18)
except:
font = ImageFont.load_default()
tw, th = draw.textbbox((0, 0), text, font=font)[2:]
pad = 6
box = [pad, pad, pad + tw + 2 * pad, pad + th + 2 * pad]
draw.rectangle(box, fill=(0, 0, 0, 180))
draw.text((pad * 2, pad * 2), text, fill=(255, 255, 255), font=font)
return img
def draw_centroid(pil_img, cx, cy, r=8, color=(255, 0, 0), w=2):
img = pil_img.convert("RGB").copy()
d = ImageDraw.Draw(img)
d.ellipse([cx - r, cy - r, cx + r, cy + r], outline=color, width=w)
d.line([(cx - 2 * r, cy), (cx + 2 * r, cy)], fill=color, width=w)
d.line([(cx, cy - 2 * r), (cx, cy + 2 * r)], fill=color, width=w)
return img
@torch.no_grad()
def predict_prob_and_centroid(pil_img, device, t_kiln, model, head, thresh=0.5):
x = model_transform()(pil_img).unsqueeze(0).to(device)
# classification prob
z = model.encode_image(x)
z = z / (z.norm(dim=-1, keepdim=True) + 1e-8)
s = torch.einsum("bd,d->b", z, t_kiln)
logits = head(s)
prob = torch.sigmoid(logits)[0].item()
cx_r, cy_r, _ = vit_centroid_from_attention(model, x)
return prob, float(cx_r[0]), float(cy_r[0]), x.shape[-1], x.shape[-2]
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--ckpt", type=str, default="checkpoints/best.pt")
ap.add_argument("--image", type=str, required=True)
ap.add_argument("--out", type=str, default="outputs/preview_centroid.jpg")
ap.add_argument("--threshold", type=float, default=0.5)
ap.add_argument("--annotate_on_original", action="store_true")
args = ap.parse_args()
ckpt_path = Path(args.ckpt)
img_path = Path(args.image)
out_path = Path(args.out)
out_path.parent.mkdir(parents=True, exist_ok=True)
assert ckpt_path.exists(), f"Checkpoint not found: {ckpt_path}"
assert img_path.exists(), f"Image not found: {img_path}"
device = "cuda" if torch.cuda.is_available() else "cpu"
t_kiln = load_kiln_embed(to_device=device)
t_kiln = t_kiln / (t_kiln.norm() + 1e-8)
ckpt = load_checkpoint(ckpt_path)
model, head = build_model_and_head(device, ckpt)
pil = Image.open(img_path).convert("RGB")
prob, cx_r, cy_r, W_r, H_r = predict_prob_and_centroid(pil, device, t_kiln, model, head, args.threshold)
_, _, sal = vit_centroid_from_attention(model, model_transform()(pil).unsqueeze(0).to(device))
save_heatmap(sal, "outputs/heatmap.png")
save_heatmap_overlay(pil, sal, "outputs/heatmap_overlay.png", alpha=0.45)
pred = int(prob >= args.threshold)
disp = display_transform()(pil)
disp = annotate_text(disp, f"pred: {'KILN' if pred==1 else 'NOT'} p={prob:.3f}")
if pred == 1:
if args.annotate_on_original:
W0, H0 = pil.size
sx = W0 / W_r
sy = H0 / H_r
cx0, cy0 = cx_r * sx, cy_r * sy
out_img = draw_centroid(pil, cx0, cy0)
print(f"centroid_px(original): ({cx0:.1f}, {cy0:.1f}) prob={prob:.3f}")
else:
out_img = draw_centroid(disp, cx_r, cy_r)
print(f"centroid_px(resized {W_r}x{H_r}): ({cx_r:.1f}, {cy_r:.1f}) prob={prob:.3f}")
else:
out_img = disp
print(f"no kiln (p={prob:.3f})")
out_img.save(out_path)
print(f"[saved] {out_path}")
if __name__ == "__main__":
main()
|