WorldModelForMaze / maze_vis_memory.py
Kalso42's picture
Upload folder using huggingface_hub
34e468d verified
Raw
History Blame Contribute Delete
17.5 kB
"""
Cross-architecture "memory length" figure.
Hold the readout fixed: for every sequence we always look at the SAME prediction
-- the distribution that predicts the token at --readout_pos. Then we perturb ONE
earlier path-step token at a time (positions from --flip_start up to readout_pos-1,
one position per trial) and measure KL(clean || perturbed) on that single fixed
readout.
Because the readout position (hence the context length) never changes, every
perturbation is measured under the same length, avoiding the 1/length dilution of
the older "flip-once read-everywhere" test (a flip in a length-20 context counts
~1/20, in length-80 ~1/80). A model that keeps a far token alive shows a high, flat
curve; a recent-steps shortcut decays fast.
The x-axis is the distance back from the readout, j = readout_pos - flipped position.
One run draws all six architectures together:
Transformer, Nextlat (transformer-nextlat), Mamba, Mamba-2, Gated-Delta, GRU
Per-architecture configs are set manually via --tf_config (transformer family) and
--rec_config (recurrent / SSM family). Supports Task A/C/E/H/I.
Gated-Delta needs the dedicated `fla` conda env (flash-linear-attention + triton).
To include it, run this whole script in that env, e.g.:
PYTHONNOUSERSITE=1 conda run -n fla python maze_vis_memory.py ...
Any model whose checkpoint is missing or fails to load is skipped with a warning.
Example (Task A):
conda run -n fla python maze_vis_memory.py --tasks A1 \
--tf_config 3_1_256 --rec_config 6_256 --num_train 500K \
--ckpt_iter 10000 --flip_start 10 --readout_pos 90
"""
import os
import sys
import glob
import pickle
import argparse
import importlib
import numpy as np
import torch
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from model.transformer import GPTConfig, GPT
from model.transformer_rope import GPTRoPEConfig, GPTRoPE
from model.transformer_nextlat import TransformerNextLatConfig, TransformerNextLat
from model.mamba import MambaConfig, Mamba
from model.mamba2 import Mamba2Config, Mamba2
from model.gated_deltanet import GatedDeltaNetConfig, GatedDeltaNet
from model.gru import GRUConfig, GRU
from cli_utils import parse_count, format_count
def _ensure_numpy_core_alias():
"""Alias numpy.core <-> numpy._core so pickled checkpoints load regardless of
the NumPy major version they were saved with. NumPy 2.0 renamed the private
``numpy.core`` package to ``numpy._core``; old pickles reference one name and
new ones the other, so we register whichever is missing."""
for src, dst in (('numpy._core', 'numpy.core'), ('numpy.core', 'numpy._core')):
try:
mod = importlib.import_module(src)
except Exception:
continue
sys.modules.setdefault(dst, mod)
for sub in ('multiarray', 'numeric', '_multiarray_umath', 'umath'):
try:
m = importlib.import_module(f'{src}.{sub}')
except Exception:
continue
sys.modules.setdefault(f'{dst}.{sub}', m)
def build_model(model_type, model_args):
"""Instantiate the architecture named by ``model_type`` from a model_args dict.
The model is returned on CPU; callers move it to the target device."""
if model_type == 'mamba':
return Mamba(MambaConfig(**model_args))
if model_type == 'mamba2':
return Mamba2(Mamba2Config(**model_args))
if model_type == 'gated-deltanet':
return GatedDeltaNet(GatedDeltaNetConfig(**model_args))
if model_type == 'gru':
return GRU(GRUConfig(**model_args))
if model_type == 'transformer-nextlat':
return TransformerNextLat(TransformerNextLatConfig(**model_args))
if model_type == 'transformer-rope':
return GPTRoPE(GPTRoPEConfig(**model_args))
return GPT(GPTConfig(**model_args))
def full_logits_any(model, idx):
"""Return full per-position logits (B, L, V) for any architecture.
Every model here only projects the last position when called without targets
(an inference-time optimization), so we pass ``targets=idx`` to force the
lm_head over all positions. The returned loss is ignored."""
out = model(idx, targets=idx)
return out[0] if isinstance(out, (tuple, list)) else out
def collect_sequences(seq_path, stoi, readout_pos, num_seqs):
"""Load up to ``num_seqs`` tokenized sequences that are at least ``readout_pos``
tokens long. Returns a list of (ids_list, colon_index) where colon_index is the
position of the ':' separator token in the sequence."""
seqs = []
with open(seq_path) as f:
for line in f:
parts = line.split()
if ':' not in parts:
continue
colon = parts.index(':')
try:
ids = [stoi[t] for t in parts]
except KeyError:
continue
if len(ids) < readout_pos:
continue
seqs.append((ids, colon))
if len(seqs) >= num_seqs:
break
return seqs
_ensure_numpy_core_alias()
# --------- manually set the six models here (config picked by --tf_config/--rec_config) ---------
# (display label, model_type, config_kind 'tf'|'rec', checkpoint suffix)
MODELS = [
('Transformer', 'transformer', 'tf', ''),
('Nextlat', 'transformer-nextlat', 'tf', 'NL'),
('Mamba', 'mamba', 'rec', ''),
('Mamba-2', 'mamba2', 'rec', ''),
('Gated-Delta', 'gated-deltanet', 'rec', ''),
('GRU', 'gru', 'rec', ''),
]
def parse_args():
p = argparse.ArgumentParser(description='Cross-architecture memory-length figure (Task A/C/E/H/I).')
p.add_argument('--tasks', type=str, default='A1', help='Task tag, e.g. A1, C1, E1, H1, I1.')
p.add_argument('--tf_config', type=str, default='3_1_256',
help='Config for the transformer family (layers_heads_dim), e.g. 3_1_256 or 12_12_576.')
p.add_argument('--rec_config', type=str, default='6_256',
help='Config for the recurrent/SSM family (layers_dim), e.g. 6_256 or 24_576.')
p.add_argument('--num_train', type=parse_count, default='500K')
p.add_argument('--ckpt_iter', type=int, default=10000)
p.add_argument('--path_type', type=str, default='RWs')
p.add_argument('--num_nodes', type=int, default=100)
p.add_argument('--dataset', type=str, default='maze')
p.add_argument('--device', type=str, default='cuda:0')
p.add_argument('--split', type=str, default='train', choices=['test', 'train'])
p.add_argument('--test_size', type=str, default='10K')
# ---- the two knobs of the test ----
p.add_argument('--flip_start', type=int, default=10,
help='Start perturbing from this token position, then sweep one position '
'at a time up to readout_pos-1.')
p.add_argument('--readout_pos', type=int, default=90,
help='Always read the prediction of the token at this position; only this '
'one fixed prediction is measured.')
p.add_argument('--num_seqs', type=int, default=400)
p.add_argument('--batch_size', type=int, default=384)
p.add_argument('--seed', type=int, default=0)
p.add_argument('--init_seed', type=int, default=1337,
help='Seed for the random init used when --ckpt_iter 0 (untrained baseline).')
p.add_argument('--out_dir', type=str, default='out/plot')
return p.parse_args()
def load_model(model_type, config, suffix, args, device):
out_dir = f"out/{model_type.replace('-', '_')}/{args.dataset}_{config}_{args.num_nodes}"
train_label = format_count(args.num_train)
tag = f'{args.tasks}_{args.path_type}' + (f'_{suffix}' if suffix else '')
if args.ckpt_iter == 0:
return load_untrained_model(out_dir, tag, train_label, model_type, args, device)
ckpt_path = os.path.join(out_dir, f'{args.ckpt_iter}_ckpt_maze_{tag}_{train_label}.pt')
print(f'[{model_type}/{config}] loading {ckpt_path}')
ckpt = torch.load(ckpt_path, map_location=device, weights_only=False)
mt = ckpt.get('model_type', model_type)
model = build_model(mt, ckpt['model_args']).to(device)
model.load_state_dict({k.replace('_orig_mod.', ''): v for k, v in ckpt['model'].items()})
model.eval()
return model
def load_untrained_model(out_dir, tag, train_label, model_type, args, device):
"""Build a randomly-initialized (iter-0) baseline from a reference checkpoint's model_args."""
pattern = os.path.join(out_dir, f'*_ckpt_maze_{tag}_{train_label}.pt')
candidates = [p for p in glob.glob(pattern)
if not os.path.basename(p).startswith('0_ckpt_')]
if not candidates:
raise FileNotFoundError(
f'No reference checkpoint matching {pattern} to infer model_args for iter 0.')
ref_path = sorted(candidates)[0]
print(f'[{model_type}/{tag}] iter 0 baseline from model_args of {ref_path}')
ckpt = torch.load(ref_path, map_location='cpu', weights_only=False)
mt = ckpt.get('model_type', model_type)
torch.manual_seed(args.init_seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(args.init_seed)
model = build_model(mt, ckpt['model_args']).to(device)
model.eval()
return model
def fixed_readout_perturb(model, seqs, args, device, index_ids):
"""Hold the readout position fixed and perturb ONE earlier token at a time.
For every sequence we always read the SAME prediction -- the distribution that
predicts the token at --readout_pos (conditioned on the tokens before it). We
then flip a single path-step token at position p (for p from --flip_start up to
readout_pos-1) and measure KL(clean || perturbed) on that one fixed readout.
Because the readout position (hence the context length) never changes, this
removes the 1/length dilution confound of the persistence test. Returns
(js, kl_mean) with j = readout_pos - p (distance back from the readout) and
kl_mean[j] = mean KL over sequences and alternative flips.
All perturbed contexts (across every sequence, position and alternative) are
pooled into one big set and run through the model in batches of --batch_size,
rather than one sequence at a time -- so the GPU stays saturated.
"""
R = args.readout_pos
start = args.flip_start
index_set = set(index_ids)
# ---- gather the clean (unperturbed) contexts, one per usable sequence ----
bases = [] # (S, R) int64 base contexts
seq_meta = [] # (colon, ids_list) per kept sequence
for ids_list, colon in seqs:
if len(ids_list) < R:
continue
bases.append(ids_list[:R])
seq_meta.append((colon, ids_list))
if not bases:
return np.arange(R), np.full(R, np.nan)
bases = np.asarray(bases, dtype=np.int64) # (S, R)
S = bases.shape[0]
def _forward_batched(rows_np):
"""rows_np: (N, R) int64 -> softmax over the readout position, (N, V)."""
outs = []
for b in range(0, rows_np.shape[0], args.batch_size):
chunk = torch.from_numpy(rows_np[b:b + args.batch_size]).to(device)
with torch.no_grad():
lg = full_logits_any(model, chunk)
outs.append(torch.softmax(lg[:, R - 1, :], dim=-1))
return torch.cat(outs, 0) # (N, V)
# clean readout distribution for each sequence
clean_all = _forward_batched(bases) # (S, V) on device
# ---- build every perturbation (seq, position, alternative) up front ----
seq_idx, pair_of, pair_p = [], [], []
triples = [] # (seq i, position p, alt token)
pid = -1
for i, (colon, ids_list) in enumerate(seq_meta):
lo = max(start, colon + 1)
for p in range(lo, R):
orig = ids_list[p]
if orig not in index_set:
continue
pid += 1
pair_p.append(p)
for alt in index_ids:
if alt == orig:
continue
triples.append((i, p, alt))
seq_idx.append(i)
pair_of.append(pid)
if not triples:
return np.arange(R), np.full(R, np.nan)
rows_np = bases[[t[0] for t in triples]].copy() # (N, R)
for k, (_, p, alt) in enumerate(triples):
rows_np[k, p] = alt
seq_idx = np.asarray(seq_idx)
pair_of = np.asarray(pair_of)
pair_p = np.asarray(pair_p)
n_pairs = pid + 1
# ---- run all perturbed contexts and accumulate KL per (seq, position) ----
pair_sum = np.zeros(n_pairs)
pair_cnt = np.zeros(n_pairs)
for b in range(0, rows_np.shape[0], args.batch_size):
chunk = torch.from_numpy(rows_np[b:b + args.batch_size]).to(device)
with torch.no_grad():
lg = full_logits_any(model, chunk)
pert = torch.softmax(lg[:, R - 1, :], dim=-1) # (n, V)
cl = clean_all[seq_idx[b:b + args.batch_size]] # (n, V)
kl = (cl * (torch.log(cl + 1e-12) - torch.log(pert + 1e-12))).sum(-1).cpu().numpy()
po = pair_of[b:b + args.batch_size]
np.add.at(pair_sum, po, kl)
np.add.at(pair_cnt, po, 1)
# mean over the alternative flips for each (seq, position), then bin by j = R - p
pair_mean = pair_sum / np.maximum(pair_cnt, 1)
kl_sum = np.zeros(R)
cnt = np.zeros(R)
for pid_i in range(n_pairs):
j = R - pair_p[pid_i]
kl_sum[j] += pair_mean[pid_i]
cnt[j] += 1
js = np.arange(R)
kl_mean = np.where(cnt > 0, kl_sum / np.maximum(cnt, 1), np.nan)
return js, kl_mean
def main():
args = parse_args()
device = args.device if torch.cuda.is_available() else 'cpu'
data_path = f'data/{args.dataset}/{args.num_nodes}'
with open(f'{data_path}/meta_{args.tasks}_{args.path_type}.pkl', 'rb') as f:
meta = pickle.load(f)
stoi = meta['stoi']
train_label = format_count(args.num_train)
if args.split == 'train':
seq_path = f'{data_path}/train_{args.tasks}_{args.path_type}_{train_label}.txt'
else:
seq_path = f'{data_path}/test_{args.tasks}_{args.path_type}_{args.test_size}.txt'
seqs = collect_sequences(seq_path, stoi, args.readout_pos, args.num_seqs)
print(f'Using {len(seqs)} {args.split} sequences (> {args.readout_pos} tokens)')
if not seqs:
raise SystemExit(f'No sequences longer than readout_pos={args.readout_pos} in {seq_path}; '
f'lower --readout_pos.')
# Derive the move/direction alphabet from the actual path region (after the colon).
# Different tasks encode steps differently (e.g. H uses digits 1-4, A uses N/S/E/W),
# so read the real step tokens from the data rather than hard-coding them.
move_ids = set()
for ids_list, colon in seqs:
move_ids.update(ids_list[colon + 1:args.readout_pos])
index_ids = sorted(move_ids)
itos = {v: k for k, v in stoi.items()}
print(f'Path-step alphabet ({len(index_ids)}): {[itos.get(i, i) for i in index_ids]}')
results = {}
for display, model_type, kind, suffix in MODELS:
config = args.tf_config if kind == 'tf' else args.rec_config
label = f'{display} {config}'
try:
model = load_model(model_type, config, suffix, args, device)
except FileNotFoundError:
print(f' ! skip {label}: checkpoint not found')
continue
except ImportError as e:
print(f' ! skip {label}: {e}')
continue
js, kl_mean = fixed_readout_perturb(model, seqs, args, device, index_ids)
kl1 = float(np.nan_to_num(kl_mean)[1]) if kl_mean.size > 1 else float('nan')
results[label] = dict(js=js, kl=np.nan_to_num(kl_mean), kl1=kl1)
print(f' {label}: KL@(j=1)={kl1:.3f}')
del model
if device.startswith('cuda'):
torch.cuda.empty_cache()
if not results:
raise SystemExit('No models loaded; nothing to plot.')
os.makedirs(args.out_dir, exist_ok=True)
tag = (f'{args.tasks}_{train_label}_{args.tf_config}_{args.rec_config}'
f'_{args.path_type}_ckpt{args.ckpt_iter}_read{args.readout_pos}_{args.split}')
fig, ax = plt.subplots(figsize=(8.5, 5.5))
colors = plt.cm.tab10(np.linspace(0, 1, max(len(results), 3)))
jmax = args.readout_pos - args.flip_start
for (label, r), c in zip(results.items(), colors):
m = (r['js'] >= 1) & (r['js'] <= jmax)
ax.plot(r['js'][m], r['kl'][m], '-', color=c, lw=2, label=label)
ax.set_xlabel(f'distance back from the fixed readout at token {args.readout_pos} '
f'(j = {args.readout_pos} - flipped position)')
ax.set_ylabel('effect on the fixed prediction: KL(clean || perturbed)')
ax.set_title(f'Single-token influence on a fixed readout, Task {args.tasks} '
f'({train_label}, {args.split})')
ax.set_xlim(1, jmax)
ax.set_ylim(bottom=0)
ax.legend()
ax.grid(alpha=0.3)
fig.tight_layout()
png = os.path.join(args.out_dir, f'memory_{tag}.png')
fig.savefig(png, dpi=130)
print(f'Wrote {png}')
npz = os.path.join(args.out_dir, f'memory_{tag}.npz')
np.savez(npz, **{label: np.stack([r['js'], r['kl']]) for label, r in results.items()})
print(f'Wrote {npz}')
if __name__ == '__main__':
main()