| |
| import argparse |
| import os |
| from pathlib import Path |
|
|
| import numpy as np |
| import torch |
|
|
|
|
| def _to_numpy(value): |
| if torch.is_tensor(value): |
| return value.detach().cpu().numpy() |
| return np.asarray(value) |
|
|
|
|
| def _load_matplotlib(): |
| import matplotlib |
|
|
| matplotlib.use("Agg") |
| import matplotlib.pyplot as plt |
|
|
| return plt |
|
|
|
|
| def save_heatmap(path, data, title, cmap="viridis"): |
| plt = _load_matplotlib() |
| fig, ax = plt.subplots(figsize=(8, 4.8), dpi=160) |
| im = ax.imshow(data, cmap=cmap, origin="upper") |
| ax.set_title(title) |
| ax.set_xlabel("token x") |
| ax.set_ylabel("token y") |
| fig.colorbar(im, ax=ax, fraction=0.046, pad=0.04) |
| fig.tight_layout() |
| fig.savefig(path) |
| plt.close(fig) |
|
|
|
|
| def save_flow(path, displacement_yx, margin, stride): |
| plt = _load_matplotlib() |
| h, w, _ = displacement_yx.shape |
| yy, xx = np.mgrid[0:h, 0:w] |
| sample = (slice(None, None, stride), slice(None, None, stride)) |
| dx = displacement_yx[..., 1][sample] |
| dy = displacement_yx[..., 0][sample] |
| confidence = margin[sample] |
|
|
| fig, ax = plt.subplots(figsize=(9, 5.4), dpi=160) |
| ax.imshow(margin, cmap="Greys", origin="upper", alpha=0.35) |
| quiver = ax.quiver( |
| xx[sample], |
| yy[sample], |
| dx, |
| dy, |
| confidence, |
| angles="xy", |
| scale_units="xy", |
| scale=1, |
| cmap="magma", |
| width=0.0035, |
| ) |
| ax.set_xlim(-0.5, w - 0.5) |
| ax.set_ylim(h - 0.5, -0.5) |
| ax.set_title(f"current token -> previous short token, stride={stride}") |
| ax.set_xlabel("token x") |
| ax.set_ylabel("token y") |
| fig.colorbar(quiver, ax=ax, fraction=0.046, pad=0.04, label="top1 - top2 score") |
| fig.tight_layout() |
| fig.savefig(path) |
| plt.close(fig) |
|
|
|
|
| def visualize(input_path, output_dir, stride): |
| artifact = torch.load(input_path, map_location="cpu") |
| output_dir = Path(output_dir) |
| output_dir.mkdir(parents=True, exist_ok=True) |
|
|
| displacement_yx = _to_numpy(artifact["displacement_yx"]) |
| margin = _to_numpy(artifact["margin"]) |
| magnitude = np.linalg.norm(displacement_yx.astype(np.float32), axis=-1) |
| top1_score = _to_numpy(artifact["top1_score"]) |
|
|
| stem = Path(input_path).stem |
| save_flow(output_dir / f"{stem}_flow_quiver.png", displacement_yx, margin, stride) |
| save_heatmap(output_dir / f"{stem}_displacement_heatmap.png", magnitude, "token displacement magnitude") |
| save_heatmap(output_dir / f"{stem}_confidence_heatmap.png", margin, "confidence: top1 - top2 score") |
| save_heatmap(output_dir / f"{stem}_top1_score_heatmap.png", top1_score, "top1 raw score") |
|
|
|
|
| def main(): |
| parser = argparse.ArgumentParser(description="Visualize short-history attention token matches.") |
| parser.add_argument("input", help="Path to a short_attn_*.pt artifact.") |
| parser.add_argument("--output-dir", default=None, help="Directory for PNG outputs. Defaults next to the artifact.") |
| parser.add_argument("--stride", type=int, default=4, help="Token stride for sparse flow arrows.") |
| args = parser.parse_args() |
|
|
| output_dir = args.output_dir or os.path.dirname(os.path.abspath(args.input)) |
| visualize(args.input, output_dir, args.stride) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
