AnonMLuser's picture
Anonymous artefact release (part 2)
038acee verified
Raw
History Blame Contribute Delete
9.71 kB
"""Reverse diffusion sampling with ReMDM remasking (Wang et al.).
Strategies (Section 4.1):
rescale: sigma = eta * sigma_max
cap: sigma = min(eta, sigma_max)
conf: per-token confidence-based remasking
Loop mode (Section 4.2, Algorithm 3):
Phase 1: standard MDLM decode, t in [1, t_on]
Phase 2: constant alpha(t_on), remasking active
Phase 3: standard MDLM decode, t in [t_off, 0]
"""
from __future__ import annotations
from collections.abc import Callable
from typing import Any
import jax
import jax.numpy as jnp
from .schedules import ScheduleFn
ModelApplyFn = Callable[
[Any, jnp.ndarray, jnp.ndarray, jnp.ndarray, Any | None], jnp.ndarray
]
# Stability guards; values must match the minihack twin exactly.
_SIGMA_DENOM_EPS = 1e-8 # sigma_max, posterior and temperature denominators
_NUCLEUS_EPS = 1e-12 # renormalisation and log guards in nucleus sampling
def _sigma_max(alpha_t: jnp.ndarray, alpha_s: jnp.ndarray) -> jnp.ndarray:
"""sigma_max = min(1, (1 - alpha_s) / alpha_t). [Eq. 7]"""
return jnp.minimum(1.0, (1.0 - alpha_s) / jnp.maximum(alpha_t, _SIGMA_DENOM_EPS))
def sigma_rescale(alpha_t, alpha_s, eta):
return eta * _sigma_max(alpha_t, alpha_s)
def sigma_cap(alpha_t, alpha_s, eta):
return jnp.minimum(eta, _sigma_max(alpha_t, alpha_s))
def sigma_conf(alpha_t, alpha_s, eta, psi, is_unmasked):
"""Per-token confidence remasking. Safe against all-masked rows."""
base = eta * _sigma_max(alpha_t, alpha_s)
any_unmasked = jnp.any(is_unmasked, axis=-1, keepdims=True)
neg_psi = jnp.where(is_unmasked, -psi, -jnp.inf)
safe_neg_psi = jnp.where(any_unmasked, neg_psi, 0.0)
eta_conf = jax.nn.softmax(safe_neg_psi, axis=-1)
return jnp.where(is_unmasked, eta_conf * base, 0.0)
_SIGMA_FNS = {
"rescale": lambda a_t, a_s, eta, *_: sigma_rescale(a_t, a_s, eta),
"cap": lambda a_t, a_s, eta, *_: sigma_cap(a_t, a_s, eta),
"conf": lambda a_t, a_s, eta, psi, unm: sigma_conf(a_t, a_s, eta, psi, unm),
}
def _nucleus_sample(rng, logits, top_p):
"""Top-p sampling from [B, H, V] logits -> [B, H] int32."""
probs = jax.nn.softmax(logits, axis=-1)
idx = jnp.argsort(-probs, axis=-1)
sorted_p = jnp.take_along_axis(probs, idx, axis=-1)
cum = jnp.cumsum(sorted_p, axis=-1)
cutoff = cum - sorted_p
sorted_p = jnp.where(cutoff >= top_p, 0.0, sorted_p)
sorted_p = sorted_p / jnp.maximum(sorted_p.sum(axis=-1, keepdims=True), _NUCLEUS_EPS)
B, H, V = logits.shape
flat = sorted_p.reshape(B * H, V)
tokens = jax.random.categorical(rng, jnp.log(flat + _NUCLEUS_EPS)).reshape(B, H)
return jnp.take_along_axis(idx, tokens[..., None], axis=-1).squeeze(-1)
def _decode(rng, logits, temperature, top_p):
"""Sample tokens from logits. Argmax only when temperature <= 0."""
if top_p is not None:
return _nucleus_sample(rng, logits / jnp.maximum(temperature, _SIGMA_DENOM_EPS), top_p)
if temperature > _SIGMA_DENOM_EPS:
B, H, V = logits.shape
scaled = logits / temperature
return jax.random.categorical(rng, scaled.reshape(-1, V)).reshape(B, H)
return jnp.argmax(logits, axis=-1)
def sample_plan(
model_apply: ModelApplyFn,
params: Any,
rng: jax.Array,
obs: jnp.ndarray,
num_actions: int,
plan_horizon: int,
num_steps: int,
schedule_fn: ScheduleFn,
remask_strategy: str = "cap",
eta: float = 0.5,
use_loop: bool = False,
t_on: float = 0.55,
t_off: float = 0.05,
temperature: float = 1.0,
top_p: float | None = None,
history: jnp.ndarray | None = None,
hist_len: jnp.ndarray | None = None,
) -> jnp.ndarray:
"""Generate an action plan via reverse diffusion with ReMDM remasking.
Implements ReMDM Algorithm 1 (Wang et al.): per-token Bernoulli
unmasking via the approximate posterior, remasking w.p. sigma from the
Section 4.1 schedules, optional Section 4.2 loop mode. The
minihack twin is src/diffusion/sampling.py:remdm_sample.
When ``history``/``hist_len`` are given, positions ``0..hist_len[b]-1``
are locked to ``history`` (Diffuser-style inpainting): never remasked,
never re-decided. With both ``None`` the computation is identical to
the unlocked sampler.
Returns:
actions: [B, H] int32.
"""
B = obs.shape[0]
mask_id = num_actions
mask_val = jnp.array(mask_id, dtype=jnp.int32)
lock = history is not None
if lock:
pos = jnp.broadcast_to(jnp.arange(plan_horizon)[None, :], (B, plan_horizon))
lock_mask = pos < hist_len[:, None]
if remask_strategy not in _SIGMA_FNS:
raise ValueError(
f"Unknown strategy {remask_strategy!r}. Options: {list(_SIGMA_FNS)}"
)
get_sigma = _SIGMA_FNS[remask_strategy]
# Phase allocation for loop mode
if use_loop:
f1, f3 = 1.0 - t_on, t_off
denom = f1 + f3 + (t_on - t_off)
n1 = max(int(round(num_steps * f1 / denom)), 1)
n3 = max(int(round(num_steps * f3 / denom)), 1)
n2 = max(num_steps - n1 - n3, 1)
else:
n1, n2, n3 = num_steps, 0, 0
alpha_loop = schedule_fn(jnp.array(t_on))
z_init = jnp.full((B, plan_horizon), mask_id, dtype=jnp.int32)
if lock:
z_init = jnp.where(lock_mask, history, z_init)
psi_init = jnp.full((B, plan_horizon), jnp.inf)
def _step(carry, _unused, t_val, alpha_t, alpha_s, sigma_on):
z, rng, psi = carry
rng, s_rng, u_rng, r_rng = jax.random.split(rng, 4)
t_inp = jnp.full((B,), t_val)
logits = model_apply(params, obs, z, t_inp, None)
x_hat = _decode(s_rng, logits, temperature, top_p)
is_masked = z == mask_id
is_unmasked = ~is_masked
sigma = get_sigma(alpha_t, alpha_s, eta, psi, is_unmasked)
sigma = jnp.broadcast_to(sigma, z.shape)
sigma = jnp.where(sigma_on, sigma, 0.0)
if lock:
sigma = jnp.where(lock_mask, 0.0, sigma)
# Masked -> unmask probability
denom = jnp.maximum(1.0 - alpha_t, _SIGMA_DENOM_EPS)
p_unmask = jnp.clip((alpha_s - (1.0 - sigma) * alpha_t) / denom, 0.0, 1.0)
do_unmask = is_masked & (jax.random.uniform(u_rng, z.shape) < p_unmask)
do_remask = is_unmasked & (jax.random.uniform(r_rng, z.shape) < sigma)
z_new = jnp.where(do_unmask, x_hat, z)
z_new = jnp.where(do_remask, mask_val, z_new)
# Update confidence history
probs = jax.nn.softmax(logits, axis=-1)
decode_prob = jnp.take_along_axis(probs, x_hat[..., None], axis=-1).squeeze(-1)
psi_new = jnp.where(do_unmask, decode_prob, psi)
psi_new = jnp.where(do_remask, jnp.inf, psi_new)
if lock:
z_new = jnp.where(lock_mask, history, z_new)
psi_new = jnp.where(lock_mask, jnp.inf, psi_new)
return (z_new, rng, psi_new), None
def _phase1_step(carry, idx):
t = 1.0 - idx * (1.0 - t_on) / n1
s = jnp.maximum(1.0 - (idx + 1) * (1.0 - t_on) / n1, t_on)
return _step(carry, idx, t, schedule_fn(t), schedule_fn(s), False)
def _phase2_step(carry, idx):
return _step(carry, idx, t_on, alpha_loop, alpha_loop, True)
def _phase3_step(carry, idx):
t = t_off - idx * t_off / n3
s = jnp.maximum(t_off - (idx + 1) * t_off / n3, 0.0)
return _step(carry, idx, t, schedule_fn(t), schedule_fn(s), False)
def _simple_step(carry, idx):
t = (num_steps - idx) / num_steps
s = jnp.maximum((num_steps - idx - 1) / num_steps, 0.0)
return _step(carry, idx, t, schedule_fn(t), schedule_fn(s), True)
carry = (z_init, rng, psi_init)
if use_loop:
carry, _ = jax.lax.scan(_phase1_step, carry, jnp.arange(n1))
carry, _ = jax.lax.scan(_phase2_step, carry, jnp.arange(n2))
if n3 > 0:
carry, _ = jax.lax.scan(_phase3_step, carry, jnp.arange(n3))
else:
carry, _ = jax.lax.scan(_simple_step, carry, jnp.arange(num_steps))
z_final = carry[0]
# Final greedy cleanup for any remaining masks
final_logits = model_apply(params, obs, z_final, jnp.zeros((B,)), None)
fallback = jnp.argmax(final_logits, axis=-1)
return jnp.where(z_final == mask_id, fallback, z_final)
def sample_plan_inpainting(
apply_fn: ModelApplyFn,
params: Any,
rng: jax.Array,
obs: jnp.ndarray,
history: jnp.ndarray,
hist_len: jnp.ndarray,
num_actions: int,
plan_horizon: int,
diffusion_steps: int,
schedule_fn: ScheduleFn,
remask_strategy: str,
eta: float,
use_loop: bool,
t_on: float,
t_off: float,
temperature: float,
top_p: float | None,
) -> jnp.ndarray:
"""ReMDM sampling with a locked historical prefix (inpainting).
Positions ``0 .. hist_len[b] - 1`` are fixed to ``history`` (planning
as inpainting, Diffuser Sec. 3.3); the remainder are generated by the
full ReMDM Algorithm 1 sampler with the configured Section 4.1
remasking schedule.
Shares the conforming ReMDM sampler core: the configured
``remask_strategy``, ``eta`` and loop parameters govern remasking,
and committed tokens are carried over per SUBS rather than
re-decided at every step.
Returns:
``[B, plan_horizon]`` int32 completed action plan.
"""
return sample_plan(
apply_fn,
params,
rng,
obs,
num_actions,
plan_horizon,
diffusion_steps,
schedule_fn,
remask_strategy=remask_strategy,
eta=eta,
use_loop=use_loop,
t_on=t_on,
t_off=t_off,
temperature=temperature,
top_p=top_p,
history=history,
hist_len=hist_len,
)