File size: 3,514 Bytes
0a2c21d | 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 | #!/usr/bin/env python3
import argparse
from pathlib import Path
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import torch
def main():
parser = argparse.ArgumentParser(description="Visualize matched token change-rate curves.")
parser.add_argument("input", help="Path to token_dynamics_*.pt artifact.")
parser.add_argument("--output-dir", default=None)
parser.add_argument("--stride", type=int, default=16, help="Plot every N-th noise token.")
parser.add_argument("--max-pairs", type=int, default=32, help="Maximum number of pair plots.")
args = parser.parse_args()
artifact = torch.load(args.input, map_location="cpu")
output_dir = Path(args.output_dir or Path(args.input).parent)
output_dir.mkdir(parents=True, exist_ok=True)
stem = Path(args.input).stem
match_indices = artifact["match_indices"].long()
history_rates = artifact["history_change_rates"].float()
noise_rates = artifact["noise_change_rates"].float()
grid_h, grid_w = artifact["grid"]
step_indices = artifact.get("step_indices", list(range(history_rates.shape[0])))
timesteps = artifact.get("timesteps", step_indices)
if history_rates.numel() == 0 or noise_rates.numel() == 0:
raise RuntimeError("Artifact has no change-rate records.")
x = list(timesteps)
plotted = 0
for noise_idx in range(match_indices.numel()):
if noise_idx % args.stride != 0:
continue
if plotted >= args.max_pairs:
break
hist_idx = int(match_indices[noise_idx].item())
noise_y = noise_idx // grid_w
noise_x = noise_idx % grid_w
hist_y = hist_idx // grid_w
hist_x = hist_idx % grid_w
fig, ax = plt.subplots(figsize=(8, 4.5), dpi=160)
ax.plot(x, history_rates[:, hist_idx].numpy(), label=f"history ({hist_y},{hist_x})", linewidth=2)
ax.plot(x, noise_rates[:, noise_idx].numpy(), label=f"noise ({noise_y},{noise_x})", linewidth=2)
ax.set_xlabel("denoise step / timestep")
ax.set_ylabel("token change rate")
hist_src = artifact.get("history_source_chunk")
hist_src_frame = artifact.get("history_source_frame")
src_note = ""
if hist_src is not None and hist_src_frame is not None:
src_note = f"\nhistory rates: chunk {hist_src} frame {hist_src_frame}"
ax.set_title(
f"matched pair: noise ({noise_y},{noise_x}) -> history ({hist_y},{hist_x}){src_note}"
)
ax.grid(True, alpha=0.3)
ax.legend()
fig.tight_layout()
out = output_dir / f"{stem}_pair_noise{noise_y}_{noise_x}_hist{hist_y}_{hist_x}.png"
fig.savefig(out)
plt.close(fig)
plotted += 1
summary_fig, ax = plt.subplots(figsize=(9, 5), dpi=160)
mean_history = history_rates.mean(dim=1).numpy()
mean_noise = noise_rates.mean(dim=1).numpy()
ax.plot(x, mean_history, label="history mean", linewidth=2)
ax.plot(x, mean_noise, label="noise mean", linewidth=2)
ax.set_xlabel("denoise step / timestep")
ax.set_ylabel("mean token change rate")
ax.set_title("mean change rate over all tokens")
ax.grid(True, alpha=0.3)
ax.legend()
summary_fig.tight_layout()
summary_path = output_dir / f"{stem}_mean_summary.png"
summary_fig.savefig(summary_path)
plt.close(summary_fig)
print(f"Saved {plotted} pair plots to {output_dir}")
print(f"Saved summary: {summary_path}")
if __name__ == "__main__":
main()
|