File size: 7,089 Bytes
1ea7ba6 | 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 | """
Evaluate all baselines from existing checkpoints — no re-training.
Outputs:
outputs/baseline_comparison.json
outputs/baseline_comparison.png (bar chart)
Usage:
python eval_baselines.py
"""
import sys, os
_base = "/mnt/d/SpiceNet" if os.path.exists("/mnt/d/SpiceNet") else "D:/SpiceNet"
sys.path.insert(0, _base)
import json
import joblib
import numpy as np
import torch
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import config
from src.dataset import get_dataloaders, build_splits
from src.baselines import (
make_resnet50, make_efficientnet_b4, make_vit_base,
predict_nn, build_svm_features,
)
from src.model import load_checkpoint
from src.utils import (
set_seed, compute_and_print_metrics, save_metrics,
topk_accuracy,
)
from sklearn.metrics import accuracy_score, f1_score
import torch
@torch.no_grad()
def eval_nn_baseline(name, model_fn, ckpt_path, test_loader, device):
print(f"\n{'-'*50} {name}")
model = model_fn()
state = torch.load(ckpt_path, map_location=device, weights_only=True)
model.load_state_dict(state)
model.to(device).eval()
all_preds, all_labels, all_logits = [], [], []
for imgs, tex, col, labels in test_loader:
imgs = imgs.to(device)
logits = model(imgs)
all_logits.append(logits.cpu())
all_preds.extend(logits.argmax(1).cpu().tolist())
all_labels.extend(labels.tolist())
logits_cat = torch.cat(all_logits, dim=0)
labels_t = torch.tensor(all_labels)
top5 = topk_accuracy(logits_cat, labels_t, k=min(5, config.NUM_CLASSES))
acc = accuracy_score(all_labels, all_preds)
f1_m = f1_score(all_labels, all_preds, average="macro")
f1_w = f1_score(all_labels, all_preds, average="weighted")
print(f" Top-1 : {acc:.4f} | Top-5 : {top5:.4f} | F1-macro : {f1_m:.4f}")
return {"model": name, "top1": acc, "top5": top5, "f1_macro": f1_m, "f1_weighted": f1_w}
def eval_svm(pkl_path, x_te, y_te):
print(f"\n{'-'*50} SVM (HOG + Color Histogram)")
pipe = joblib.load(pkl_path)
X_te = build_svm_features(x_te)
y_pred = pipe.predict(X_te)
acc = accuracy_score(y_te, y_pred)
f1_m = f1_score(y_te, y_pred, average="macro")
f1_w = f1_score(y_te, y_pred, average="weighted")
print(f" Top-1 : {acc:.4f} | F1-macro : {f1_m:.4f}")
return {"model": "SVM (HOG + Color)", "top1": acc, "top5": None, "f1_macro": f1_m, "f1_weighted": f1_w}
@torch.no_grad()
def eval_spicefusionnet(ckpt_path, mode, test_loader, device):
label = "SpiceFusionNet (ours)" if mode == "fusion" else "EfficientNet-B4 image-only (Phase 1)"
print(f"\n{'-'*50} {label}")
model, _, _, _ = load_checkpoint(ckpt_path, device)
model.eval()
all_preds, all_labels, all_logits = [], [], []
for imgs, tex, col, labels in test_loader:
imgs, labels = imgs.to(device), labels.to(device)
if mode == "fusion":
tex, col = tex.to(device), col.to(device)
logits, _ = model.forward_fusion(imgs, tex, col)
else:
logits = model.forward_image(imgs)
all_logits.append(logits.cpu())
all_preds.extend(logits.argmax(1).cpu().tolist())
all_labels.extend(labels.cpu().tolist())
logits_cat = torch.cat(all_logits, dim=0)
labels_t = torch.tensor(all_labels)
top5 = topk_accuracy(logits_cat, labels_t, k=min(5, config.NUM_CLASSES))
acc = accuracy_score(all_labels, all_preds)
f1_m = f1_score(all_labels, all_preds, average="macro")
f1_w = f1_score(all_labels, all_preds, average="weighted")
print(f" Top-1 : {acc:.4f} | Top-5 : {top5:.4f} | F1-macro : {f1_m:.4f}")
return {"model": label, "top1": acc, "top5": top5, "f1_macro": f1_m, "f1_weighted": f1_w}
def plot_comparison(results, output_path):
models = [r["model"] for r in results]
top1_vals = [r["top1"] * 100 for r in results]
f1_vals = [r["f1_macro"] * 100 for r in results]
x = np.arange(len(models))
width = 0.35
fig, ax = plt.subplots(figsize=(13, 6))
bars1 = ax.bar(x - width/2, top1_vals, width, label="Top-1 Accuracy (%)", color="#4C72B0", zorder=3)
bars2 = ax.bar(x + width/2, f1_vals, width, label="Macro F1-Score (%)", color="#DD8452", zorder=3)
for bar in bars1:
ax.text(bar.get_x() + bar.get_width()/2, bar.get_height() + 0.3,
f"{bar.get_height():.2f}", ha="center", va="bottom", fontsize=8.5, fontweight="bold")
for bar in bars2:
ax.text(bar.get_x() + bar.get_width()/2, bar.get_height() + 0.3,
f"{bar.get_height():.2f}", ha="center", va="bottom", fontsize=8.5, fontweight="bold")
ax.set_ylabel("Score (%)", fontsize=12)
ax.set_title("Baseline Comparison — SpiceNet", fontsize=14, fontweight="bold")
ax.set_xticks(x)
ax.set_xticklabels(models, rotation=18, ha="right", fontsize=10)
ax.set_ylim(0, 108)
ax.yaxis.grid(True, linestyle="--", alpha=0.6, zorder=0)
ax.legend(fontsize=11)
plt.tight_layout()
plt.savefig(output_path, dpi=150)
plt.close()
print(f"\nChart saved → {output_path}")
def print_table(results):
print(f"\n{'='*75}")
print(f" {'Model':<40} {'Top-1':>8} {'Top-5':>8} {'F1-Macro':>9}")
print(f"{'='*75}")
for r in results:
top5_str = f"{r['top5']*100:7.2f}%" if r["top5"] is not None else " N/A"
flag = " << ours" if "ours" in r["model"] else ""
print(f" {r['model']:<40} {r['top1']*100:7.2f}% {top5_str} {r['f1_macro']*100:8.2f}%{flag}")
print(f"{'='*75}")
def main():
set_seed(config.RANDOM_SEED)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Device: {device}\n")
ckpt = config.CHECKPOINT_DIR
out = config.OUTPUT_DIR
out.mkdir(parents=True, exist_ok=True)
# Loaders — image-only for baselines, multimodal for SpiceFusionNet fusion
_, _, test_loader_img, x_te, y_te = get_dataloaders(multimodal=False)
_, _, test_loader_fuse, _, _ = get_dataloaders(multimodal=True)
results = []
# SVM
results.append(eval_svm(ckpt / "svm_best.pkl", x_te, y_te))
# ResNet-50
results.append(eval_nn_baseline(
"ResNet-50 (fine-tuned)",
make_resnet50, ckpt / "resnet50_best.pth",
test_loader_img, device,
))
# EfficientNet-B4 image-only baseline
results.append(eval_nn_baseline(
"EfficientNet-B4 (image-only)",
make_efficientnet_b4, ckpt / "efficientnet_b4_best.pth",
test_loader_img, device,
))
# ViT-Base
results.append(eval_nn_baseline(
"ViT-Base (fine-tuned)",
make_vit_base, ckpt / "vit_base_best.pth",
test_loader_img, device,
))
# SpiceFusionNet — full fusion
results.append(eval_spicefusionnet(
ckpt / "best.pth", "fusion",
test_loader_fuse, device,
))
print_table(results)
save_metrics({"baselines": results}, out / "baseline_comparison.json")
plot_comparison(results, out / "baseline_comparison.png")
if __name__ == "__main__":
main()
|