File size: 4,285 Bytes
d4bcd5c | 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 | """
Poster/logbook figures for the SPLIT-VLM reproduction:
1. fig_mechanism_heatmaps.png — temporal-shift importance vs CLS attention over
the 24x24 patch grid for a real image (illustrates position bias, Claim 3),
plus which tokens SPLIT vs attention-topk keep at 64 tokens.
2. fig_mechanism_bars.png — summary bars (center-of-mass row, Gini, Spearman).
Reuses the cached CLIP vision tower.
"""
import os, sys, json
import numpy as np
import torch
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from PIL import Image
sys.path.insert(0, os.path.dirname(__file__))
from split_prune import (temporal_shift_importance, region_ids_grid, split_select,
attention_select)
MODEL = "openai/clip-vit-large-patch14-336"
DEVICE = os.environ.get("FIGDEV","cpu")
GRID = (24, 24)
REGION = (4, 4)
def load_feats(path):
from transformers import CLIPVisionModel, CLIPImageProcessor
proc = CLIPImageProcessor.from_pretrained(MODEL)
model = CLIPVisionModel.from_pretrained(MODEL, torch_dtype=torch.float32,
attn_implementation="eager").to(DEVICE).eval()
img = Image.open(path).convert("RGB")
px = proc(images=img, return_tensors="pt")["pixel_values"].to(DEVICE)
with torch.no_grad():
out = model(px, output_hidden_states=True, output_attentions=True)
hs = [h[0, 1:, :].float().cpu() for h in out.hidden_states]
att = torch.stack([a[0, :, 0, 1:].mean(0) for a in out.attentions]).mean(0).float().cpu()
imp = temporal_shift_importance(hs)
emb = hs[-1]
return img, hs, emb, imp, att
def grid_img(v):
return v.numpy().reshape(GRID)
def main():
path = "outputs/sample_images/cats.jpg"
img, hs, emb, imp, att = load_feats(path)
N = imp.shape[0]
keep_split = split_select(hs, emb, 64, GRID, REGION).numpy()
keep_attn = attention_select(att, 64).numpy()
def mask_grid(keep):
m = np.zeros(N); m[keep] = 1.0
return m.reshape(GRID)
fig, ax = plt.subplots(1, 5, figsize=(18, 4.2))
ax[0].imshow(img.resize((336, 336))); ax[0].set_title("Input (LLaVA 336x336)", fontsize=12)
im1 = ax[1].imshow(grid_img(imp), cmap="viridis"); ax[1].set_title("Temporal-shift importance\n(attention-free, centered)", fontsize=12)
im2 = ax[2].imshow(grid_img(att), cmap="magma"); ax[2].set_title("CLS attention\n(position bias -> lower rows)", fontsize=12)
ax[3].imshow(mask_grid(keep_split), cmap="Greens"); ax[3].set_title("SPLIT keeps @64\n(all 16 regions covered)", fontsize=12)
ax[4].imshow(mask_grid(keep_attn), cmap="Oranges"); ax[4].set_title("Attention-topk keeps @64\n(regions emptied)", fontsize=12)
for a in ax: a.set_xticks([]); a.set_yticks([])
plt.tight_layout()
plt.savefig("outputs/fig_mechanism_heatmaps.png", dpi=130, bbox_inches="tight")
print("wrote outputs/fig_mechanism_heatmaps.png")
# summary bars from mechanism_validation.json
with open("outputs/mechanism_validation.json") as f:
mv = json.load(f)
imgs = list(mv["images"].values())
imp_com = np.mean([x["imp_com_row"] for x in imgs])
att_com = np.mean([x["att_com_row"] for x in imgs])
imp_g = np.mean([x["imp_gini"] for x in imgs])
att_g = np.mean([x["att_gini"] for x in imgs])
spear = np.mean([x["imp_att_spearman"] for x in imgs])
fig, ax = plt.subplots(1, 3, figsize=(13, 3.6))
ax[0].bar(["temporal\nshift", "CLS\nattention"], [imp_com, att_com], color=["#2D5F8B", "#C1666B"])
ax[0].axhline(11.5, ls="--", c="gray", lw=1); ax[0].text(1.05, 11.5, "grid center", fontsize=8, va="bottom")
ax[0].set_title("Center-of-mass row\n(11.5 = unbiased)"); ax[0].set_ylim(10, 14)
ax[1].bar(["temporal\nshift", "CLS\nattention"], [imp_g, att_g], color=["#2D5F8B", "#C1666B"])
ax[1].set_title("Gini (concentration)\nlower = more spread")
ax[2].bar(["Spearman(imp, attn)"], [spear], color=["#6B4E8B"])
ax[2].axhline(0, c="k", lw=0.8); ax[2].set_ylim(-1, 1)
ax[2].set_title("Temporal-shift vs attention\n(anti-correlated => different signal)")
plt.tight_layout()
plt.savefig("outputs/fig_mechanism_bars.png", dpi=130, bbox_inches="tight")
print("wrote outputs/fig_mechanism_bars.png")
if __name__ == "__main__":
main()
|