File size: 5,608 Bytes
ef3d1e2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations
from pathlib import Path
from typing import Dict

import numpy as np
import pandas as pd
from PIL import Image

import torch
from torchvision import transforms
import timm
from timm.models.vision_transformer import resize_pos_embed
import joblib


# ----------------------- paths & device -----------------------
ROOT_DIR = Path(__file__).resolve().parent.parent      # AMLGroupSpaceFinal/
BASELINE_DIR = ROOT_DIR / "baseline"
LIST_DIR = ROOT_DIR / "list"

PLANT_CKPT_PATH = BASELINE_DIR / "plant_dinov2_patch14.pth"
LOGREG_PATH = BASELINE_DIR / "logreg_baseline.joblib"
SCALER_PATH = BASELINE_DIR / "scaler_baseline.joblib"
SPECIES_LIST_PATH = LIST_DIR / "species_list.txt"

DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")


# ----------------------- helpers (trimmed from evaluate.py) -----------------------
def read_species(p: Path):
    """Read species_list.txt and return list of species names in index order."""
    rows = []
    with open(p, "r", encoding="utf-8") as f:
        for ln in f:
            ln = ln.strip()
            if not ln or ln.startswith("#"):
                continue
            if ";" in ln:
                cid, name = ln.split(";", 1)
            else:
                parts = ln.split()
                cid, name = parts[0], " ".join(parts[1:]) if len(parts) > 1 else ""
            try:
                cid = int(cid)
            except ValueError:
                continue
            rows.append((cid, name))
    df = pd.DataFrame(rows, columns=["class_id", "species_name"])
    # same order as in training: iterrows order
    names = list(df["species_name"])
    return names


def pool_feats(out):
    feats = out
    if isinstance(out, dict):
        for key in ("pooled", "x_norm_clstoken", "cls_token", "x"):
            if key in out:
                feats = out[key]
                break
    if isinstance(feats, (list, tuple)):
        feats = feats[0]
    if isinstance(feats, torch.Tensor) and feats.dim() == 3:
        feats = feats[:, 0] if feats.size(1) > 1 else feats.mean(dim=1)
    if isinstance(feats, torch.Tensor) and feats.dim() > 2:
        feats = feats.flatten(1)
    return feats


def _unwrap_state_dict(obj):
    if isinstance(obj, dict):
        for k in ("state_dict", "model", "module", "ema", "shadow",
                  "backbone", "net", "student", "teacher"):
            if k in obj and isinstance(obj[k], dict):
                return obj[k]
    return obj


def _strip_prefixes(sd, prefixes=("module.", "backbone.", "model.", "student.")):
    out = {}
    for k, v in sd.items():
        for p in prefixes:
            if k.startswith(p):
                k = k[len(p):]
        out[k] = v
    return out


def maybe_load_plant_ckpt(model, ckpt_path: Path):
    if not ckpt_path.is_file():
        print(f"[baseline] plant ckpt not found at {ckpt_path}, using generic DINOv2 weights.")
        return
    try:
        sd = torch.load(ckpt_path, map_location="cpu")
        sd = _unwrap_state_dict(sd)
        sd = _strip_prefixes(sd)

        msd = model.state_dict()
        if "pos_embed" in sd and "pos_embed" in msd and sd["pos_embed"].shape != msd["pos_embed"].shape:
            sd["pos_embed"] = resize_pos_embed(sd["pos_embed"], msd["pos_embed"])
            print(f"[baseline] interpolated pos_embed to {tuple(msd['pos_embed'].shape)}")

        missing, unexpected = model.load_state_dict(sd, strict=False)
        print(f"[baseline] loaded plant ckpt; missing={len(missing)} unexpected={len(unexpected)}")
    except Exception as e:
        print(f"[baseline] failed to load '{ckpt_path}': {e}")


def build_backbone(size: int = 224):
    model = timm.create_model(
        "vit_base_patch14_dinov2",
        pretrained=True,    # generic DINOv2 as fallback
        num_classes=0,      # features only
        img_size=size,
        pretrained_cfg_overlay=dict(input_size=(3, size, size)),
    ).to(DEVICE)

    pe = getattr(model, "patch_embed", None)
    if pe is not None:
        if hasattr(pe, "img_size"):
            pe.img_size = (size, size)
        if hasattr(pe, "strict_img_size"):
            pe.strict_img_size = False

    maybe_load_plant_ckpt(model, PLANT_CKPT_PATH)
    model.eval()
    for p in model.parameters():
        p.requires_grad = False
    return model


# ----------------------- global objects (loaded once) -----------------------
IMAGE_SIZE = 224

species_names = read_species(SPECIES_LIST_PATH)
num_classes = len(species_names)

backbone = build_backbone(IMAGE_SIZE)

transform = transforms.Compose([
    transforms.Resize(int(IMAGE_SIZE * 1.12)),
    transforms.CenterCrop(IMAGE_SIZE),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406],
                         [0.229, 0.224, 0.225]),
])

scaler = joblib.load(SCALER_PATH)
logreg = joblib.load(LOGREG_PATH)


# ----------------------- public API for Gradio -----------------------
def predict_baseline(image: Image.Image, top_k: int = 5) -> Dict[str, float]:
    """
    Run DINOv2 + Logistic Regression baseline on a single PIL image.
    Returns {class_name: probability} for the top_k classes.
    """
    if image is None:
        return {}

    x = transform(image).unsqueeze(0).to(DEVICE)

    with torch.no_grad():
        out = backbone.forward_features(x)
        feats = pool_feats(out).cpu().numpy()

    feats_scaled = scaler.transform(feats)
    probs = logreg.predict_proba(feats_scaled)[0]  # shape [num_classes]

    top_idx = np.argsort(-probs)[:top_k]
    result = {species_names[i]: float(probs[i]) for i in top_idx}
    return result