File size: 5,219 Bytes
63fba53 | 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 | """
visualize_attention.py
Visualize layer-wise attention results for all models, comparing COCO vs VGMED.
Usage:
python visualize_attention.py
python visualize_attention.py --metric js
python visualize_attention.py --output_dir my_plots
"""
import argparse
import os
import matplotlib.pyplot as plt
import numpy as np
import torch
ATTENTION_ROOT = "attention_results"
MODELS = {
"HuatuoGPT-V-7B": "HuatuoGPT-V-7B/HuatuoGPT-Vision-7B",
"LLaVA-7B": "LLaVA-7B/llava-1.5-7b-hf",
"LLaVA-13B": "LLaVA-13B/llava-1.5-13b-hf",
"LLaVA-Med-7B": "LLaVA-Med-7B/llava-med-v1.5-mistral-7b",
"MedGemma-4B-IT": "MedGemma-4B-IT/medgemma-4b-it",
"Qwen2.5VL-3B": "Qwen2.5VL-3B/Qwen2.5-VL-3B-Instruct",
"Qwen2.5VL-7B": "Qwen2.5VL-7B/Qwen2.5-VL-7B-Instruct",
"Qwen3VL-2B": "Qwen3VL-2B/Qwen3-VL-2B-Instruct",
"Qwen3VL-4B": "Qwen3VL-4B/Qwen3-VL-4B-Instruct",
"Qwen3VL-8B": "Qwen3VL-8B/Qwen3-VL-8B-Instruct",
}
METRICS = {
"js": "JS Divergence",
"kl": "KL Divergence",
"ratios": "Attention Ratios",
}
DATASETS = ["COCO", "VGMED"]
# ── shared style constants (kept in sync with visualize_grounding.py) ─────────
# Each model is saved as its own standalone PNG, so every figure must be readable
# on its own at a small size: large fonts, thick lines, high DPI.
COLORS = {"COCO": "#2196F3", "VGMED": "#F44336"}
DIFF_POS = "#F44336"
DIFF_NEG = "#2196F3"
FIG_WIDTH = 8.0
ROW_HEIGHT = 5.0
TITLE_FS = 22
AXIS_FS = 18
TICK_FS = 16
LEGEND_FS = 16
MARKER = "o"
MARKER_SIZE = 9
LINE_WIDTH = 3.5
GRID_ALPHA = 0.3
DPI = 600
plt.rcParams["font.weight"] = "bold"
plt.rcParams["axes.labelweight"] = "bold"
plt.rcParams["axes.titleweight"] = "bold"
plt.rcParams["axes.linewidth"] = 1.5
# ──────────────────────────────────────────────────────────────────────────────
def load_pt(path):
data = torch.load(path, map_location="cpu", weights_only=False)
arr = np.stack(data, axis=0) # [N, num_layers]
return arr.mean(axis=0) # [num_layers]
def plot_metric(metric_key, metric_label, output_dir):
"""One standalone PNG per model: top = metric curves, bottom = VGMED-COCO diff bars."""
for model_name, model_dir in MODELS.items():
fig, (ax_top, ax_bot) = plt.subplots(
2, 1, figsize=(FIG_WIDTH, ROW_HEIGHT * 2),
)
fig.suptitle(
f"{model_name} — {metric_label} (COCO vs VGMED)",
fontsize=TITLE_FS, fontweight="bold",
)
loaded = {}
for ds in DATASETS:
fname = f"{ds}_localization_attention_{metric_key}_normalized.pt"
fpath = os.path.join(ATTENTION_ROOT, model_dir, fname)
if not os.path.exists(fpath):
print(f"[skip] {fpath}")
continue
loaded[ds] = load_pt(fpath)
# top: curves
for ds, vals in loaded.items():
layers = np.arange(len(vals))
ax_top.plot(layers, vals, label=ds, color=COLORS[ds],
linewidth=LINE_WIDTH, marker=MARKER, markersize=MARKER_SIZE)
ax_top.set_ylabel(metric_label, fontsize=AXIS_FS)
ax_top.tick_params(labelsize=TICK_FS)
ax_top.grid(True, alpha=GRID_ALPHA)
if loaded:
ax_top.legend(fontsize=LEGEND_FS)
# bottom: VGMED − COCO diff bars
if "VGMED" in loaded and "COCO" in loaded:
vgmed = loaded["VGMED"]
coco = loaded["COCO"]
n = min(len(vgmed), len(coco))
diff = vgmed[:n] - coco[:n]
layers = np.arange(n)
ax_bot.bar(layers, diff,
color=[DIFF_POS if d >= 0 else DIFF_NEG for d in diff],
alpha=0.7, width=0.8)
ax_bot.axhline(0, color="black", linewidth=0.8, linestyle="--")
ax_bot.set_xlabel("Layer", fontsize=AXIS_FS)
ax_bot.set_ylabel("VGMED − COCO", fontsize=AXIS_FS)
ax_bot.tick_params(labelsize=TICK_FS)
ax_bot.grid(True, alpha=GRID_ALPHA, axis="y")
else:
ax_bot.set_visible(False)
plt.tight_layout()
model_dir_out = os.path.join(output_dir, metric_key)
os.makedirs(model_dir_out, exist_ok=True)
out_path = os.path.join(model_dir_out, f"{model_name}_{metric_key}.png")
plt.savefig(out_path, dpi=DPI, bbox_inches="tight")
plt.close()
print(f"Saved: {out_path}")
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--metric", choices=list(METRICS.keys()), default=None,
help="Plot only one metric (default: all)")
parser.add_argument("--output_dir", default="attention_plots")
args = parser.parse_args()
targets = {args.metric: METRICS[args.metric]} if args.metric else METRICS
for metric_key, metric_label in targets.items():
plot_metric(metric_key, metric_label, args.output_dir)
if __name__ == "__main__":
main()
|