File size: 5,253 Bytes
0c0ff0e | 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 | """Plot per-layer probe accuracy for the architectures on one figure.
Data is loaded automatically from the kdetour result files saved by
maze_kstep_detour_test.py (out/maze_kdetour/kdetour_*.npz, key
'layer_probe_acc'). You only need to set TASK / DATASET / configs below.
Horizontal axis = absolute layer number. Because the RNN/SSM models here have
twice as many layers as the transformers, one transformer layer is aligned to
two Mamba/GRU layers: transformer layer i is drawn at x = 2*i. The bottom of
the plot uses two rows of tick labels -- top row = Mamba/GRU layer number,
bottom row = Transformer layer number.
"""
import os
import argparse
import numpy as np
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
# ---------------------------------------------------------------------------
# Defaults (overridable on the command line, see --help).
# ---------------------------------------------------------------------------
TASK = 'I1'
DATASET = '10M'
# Model configs: transformer-family uses TF_CONFIG, RNN/SSM-family uses RNN_CONFIG.
TF_CONFIG = '6_6_384'
RNN_CONFIG = '12_384'
# kdetour result settings (match how the .npz files were produced).
CKPT_ITER = 10000
PATH_TYPE = 'RWs'
KDETOUR_DIR = 'out/maze_kdetour'
OUT_DIR = 'out/plot'
def parse_args():
p = argparse.ArgumentParser(description='Plot per-layer probe accuracy from kdetour npz files.')
p.add_argument('--task', default=TASK)
p.add_argument('--dataset', default=DATASET)
p.add_argument('--tf_config', default=TF_CONFIG)
p.add_argument('--rnn_config', default=RNN_CONFIG)
p.add_argument('--ckpt_iter', type=int, default=CKPT_ITER)
p.add_argument('--path_type', default=PATH_TYPE)
p.add_argument('--kdetour_dir', default=KDETOUR_DIR)
p.add_argument('--out_dir', default=OUT_DIR)
return p.parse_args()
# (display name, file model key, is_transformer_family). Configs come from args.
MODEL_SPECS = [
('Transformer', 'transformer', True),
('Transformer-NextLat', 'transformer_nextlat', True),
('Mamba', 'mamba', False),
('Mamba-2', 'mamba2', False),
('Gated-DeltaNet', 'gated_deltanet', False),
('GRU', 'gru', False),
]
MARKERS = ['o', 's', '^', 'D', 'P', 'v']
def load_layer_probe_acc(args, model_key, config):
"""Return (layer_probe_acc array, best_layer int) from the kdetour npz, or
(None, None) if the file is missing / has no layer_probe_acc."""
fname = f'kdetour_{args.task}_{args.path_type}_{args.ckpt_iter}_{args.dataset}_{model_key}_{config}.npz'
path = os.path.join(args.kdetour_dir, fname)
if not os.path.exists(path):
print(f"[missing] {path}")
return None, None
d = np.load(path, allow_pickle=True)
if 'layer_probe_acc' not in d:
print(f"[no layer_probe_acc] {path}")
return None, None
vals = np.asarray(d['layer_probe_acc'], dtype=float)
best = int(d['best_layer']) if 'best_layer' in d else int(np.argmax(vals) + 1)
return vals, best
def main():
args = parse_args()
os.makedirs(args.out_dir, exist_ok=True)
out = os.path.join(
args.out_dir,
f'layer_probe_acc_{args.task}_{args.dataset}_{args.tf_config}_{args.rnn_config}.png')
plt.figure(figsize=(10, 5.5))
# Transformer family = is_tf True; 1 transformer layer spans 2 RNN layers,
# so transformer layer i is placed at x = 2*i to align with Mamba/GRU.
parsed = []
maxx = 1
for (name, key, is_tf), marker in zip(MODEL_SPECS, MARKERS):
config = args.tf_config if is_tf else args.rnn_config
vals, best = load_layer_probe_acc(args, key, config)
if vals is None or len(vals) == 0:
parsed.append(None)
continue
n = len(vals)
x = [2 * i for i in range(1, n + 1)] if is_tf else list(range(1, n + 1))
label = f'{name} ({config})'
parsed.append((label, x, vals, marker, best))
maxx = max(maxx, max(x))
plotted = 0
for item in parsed:
if item is None:
continue
label, x, vals, marker, best = item
plt.plot(x, vals, marker=marker, markersize=6, linewidth=2, label=label)
print(f"{label}: {len(vals)} layers, best = L{best} ({vals[best - 1]:.1f}%)")
plotted += 1
if plotted == 0:
print("Nothing to plot: no kdetour npz files found for these settings.")
return
# Two-row tick labels: top row = Mamba/GRU layer, bottom row = Transformer layer.
ticks = list(range(1, maxx + 1))
labels = []
for t in ticks:
tf = str(t // 2) if t % 2 == 0 else ''
labels.append(f"{t}\n{tf}")
plt.xticks(ticks, labels, fontsize=8)
plt.xlabel('layer (top: Mamba/GRU layer, bottom: Transformer layer)', fontsize=11)
plt.ylabel('probe accuracy (%)', fontsize=12)
plt.title(f'Per-layer probe accuracy (Task {args.task}, {args.dataset})', fontsize=13)
plt.ylim(-2, 105)
plt.grid(True, alpha=0.3)
plt.legend(fontsize=10, framealpha=0.9)
plt.tight_layout()
plt.savefig(out, dpi=150)
plt.close()
print(f"Saved figure to {out}")
if __name__ == '__main__':
main()
|