scripts/ablation.py

import os
import sys
# Ensure repo root is on path (ablation.py lives in scripts/)
_script_dir = os.path.dirname(os.path.abspath(__file__))
_repo_root = os.path.dirname(_script_dir)
if _repo_root not in sys.path:
    sys.path.insert(0, _repo_root)

os.environ["KERAS_BACKEND"] = "jax"

# JAX 0.6.x compat: spmd_mode removed
import jax
if not hasattr(jax, "spmd_mode"):
    import contextlib
    @contextlib.contextmanager
    def _spmd_mode(_disabled=False):
        yield
    jax.spmd_mode = _spmd_mode

import json
from datetime import datetime
from time import time
import itertools

import jax
import jax.numpy as jnp
import keras
import numpy as np
import pgx
import pytz
from omegaconf import OmegaConf
from pydantic import BaseModel
from pgx.experimental import auto_reset
from typing import get_args

from resnet import PQNet
from util import KeyGenerator


class Config(BaseModel):
    env_id: str = "connect_four"
    seed: int = 0
    mode: str = "full"
    selfplay_vmap: int = 1024
    selfplay_step: int = 2048
    alpha: float = 0.03
    beta: float = 0.1
    tau: float = 8.0
    fitting_batch_size: int = 4096
    fitting_epochs: int = 1
    limit_simulator_evaluations: int = 10 ** 9
    num_channels: int = 128
    num_blocks: int = 6
    num_params: int = 0
    zero_init: bool = True
    learning_rate: float = 1e-3
    optimizer: str = "Adam"
    host_name: str = os.uname().nodename
    device_kind: str = jax.local_devices()[0].device_kind
    evaluation_vmap: int = 1024
    save_interval: int = 50
    checkpoints_dir: str = "./checkpoints"
    wandb_on: bool = True
    comment: str = ""


def network(observations, params):
    observations = jax.vmap(boolify)(observations)
    outputs, _ = model.stateless_call(**params, inputs=observations, training=False)
    return outputs["logits"], outputs["qvalue"]


def encode(observation):
    return jnp.packbits(observation.astype(jnp.bool).flatten())


def decode(code):
    return jnp.unpackbits(code)[:prod_observation_shape].reshape(env.observation_shape).astype(jnp.float32)


def boolify(observation):
    return decode(encode(observation))


def calculate_targets(R, V, T):
    V_next = jnp.concatenate([V[1:], jnp.array([jnp.nan])])
    RVT = jnp.stack([R, V_next, T], axis=1)
    if config.mode == "td0":
        lambda_ = 0.0
    elif config.mode == "mc":
        lambda_ = 1.0
    else:
        lambda_ = jnp.exp(-1 / jnp.clip(config.tau, min=1e-12))
    gamma = -1

    def body_fn(carry, rvt):
        r, v, terminated = rvt
        target = r + gamma * ((1 - lambda_) * v + lambda_ * carry)
        return jax.lax.cond((terminated > 0).all(), lambda: (r, r), lambda: (target, target))

    _, G = jax.lax.scan(body_fn, jnp.nan, RVT, reverse=True)
    return G


@jax.jit
def selfplay(rng_key, params):
    n_vmap = config.selfplay_vmap
    n_step = config.selfplay_step
    step_fn = auto_reset(env.step, env.init)

    def body_fn(i_step, loop_state):
        states, C, A, R, T, P, V, STATS, rng_key = loop_state
        rng_key, key1, key2 = jax.random.split(rng_key, 3)

        a, b = config.alpha, config.beta
        logits, qvalue = network(states.observation, params)

        if config.mode == "kl_only":
            improved_logits = (b * logits + qvalue) / (b + 1e-12)
        elif config.mode == "entropy_only":
            improved_logits = qvalue / (a + 1e-12)
        else:
            improved_logits = (b * logits + qvalue) / (a + b + 1e-12)

        improved_logits = improved_logits + jnp.log(states.legal_action_mask)
        improved_policies = jax.nn.softmax(improved_logits, axis=-1)

        actions = jax.vmap(lambda p, key: jax.random.choice(key, a=env.num_actions, p=p))(
            improved_policies, jax.random.split(key1, n_vmap)
        )

        C = C.at[i_step].set(jax.vmap(encode)(states.observation))
        A = A.at[i_step].set(actions)
        P = P.at[i_step].set(improved_policies)
        V = V.at[i_step].set(jnp.sum(improved_policies * qvalue, axis=1))
        current_player = states.current_player
        states = jax.vmap(step_fn)(states, actions, jax.random.split(key2, n_vmap))
        R = R.at[i_step].set(jax.vmap(lambda R, c: R[c])(states.rewards, current_player))
        T = T.at[i_step].set(states.terminated)

        prior_policies = jax.nn.softmax(logits, axis=-1)
        STATS = STATS.at[i_step].set(
            jnp.array(
                [
                    jnp.sum(prior_policies * qvalue, axis=1),
                    jnp.sum(improved_policies * qvalue, axis=1),
                    jax.vmap(kl_divergence)(prior_policies, prior_policies),
                    jax.vmap(kl_divergence)(improved_policies, prior_policies),
                    jax.vmap(entropy)(prior_policies),
                    jax.vmap(entropy)(improved_policies),
                ]
            ).T
        )

        return states, C, A, R, T, P, V, STATS, rng_key

    code = encode(jnp.zeros(env.observation_shape))
    C = jnp.zeros((n_step, n_vmap, *code.shape), dtype=code.dtype)
    A = jnp.zeros((n_step, n_vmap), dtype=jnp.int32)
    R = jnp.zeros((n_step, n_vmap))
    T = jnp.zeros((n_step, n_vmap), dtype=jnp.bool_)
    P = jnp.zeros((n_step, n_vmap, env.num_actions))
    V = jnp.zeros((n_step, n_vmap))
    STATS = jnp.zeros((n_step, n_vmap, 6))
    key1, key2 = jax.random.split(rng_key)
    states = jax.vmap(env.init)(jax.random.split(key1, n_vmap))
    _, C, A, R, T, P, V, STATS, _ = jax.lax.fori_loop(0, n_step, body_fn, (states, C, A, R, T, P, V, STATS, key2))

    C, A, R, T, P, V = map(lambda X: jnp.swapaxes(X, 0, 1), (C, A, R, T, P, V))
    G = jax.vmap(calculate_targets)(R, V, T)
    C = C.reshape((n_vmap * n_step, *C.shape[2:]))
    A = A.reshape((n_vmap * n_step, 1))
    P = P.reshape((n_vmap * n_step, env.num_actions))
    G = G.reshape((n_vmap * n_step, 1))
    STATS = jnp.mean(STATS, axis=(0, 1))
    return C, A, P, G, STATS


@jax.jit
def evaluate(rng_key, params, opp_coef):
    our_player = 0
    rng_key, sub_key = jax.random.split(rng_key)
    n_vmap = config.evaluation_vmap
    states = jax.vmap(env.init)(jax.random.split(sub_key, n_vmap))

    def body_fn(loop_state):
        rng_key, states, rewards = loop_state
        rng_key, sub_key = jax.random.split(rng_key)
        logits, qvalue = network(states.observation, params)
        our_logits = 10000 * logits
        opp_logits = opp_coef * baseline(states.observation)[0]
        logits = jnp.where((states.current_player == our_player).reshape(-1, 1), our_logits, opp_logits)
        logits = logits + jnp.log(states.legal_action_mask)
        actions = jax.random.categorical(sub_key, logits, axis=-1)
        states = jax.vmap(env.step)(states, actions)
        rewards = rewards + states.rewards[jnp.arange(n_vmap), our_player]
        return rng_key, states, rewards

    _, _, reward = jax.lax.while_loop(
        lambda x: ~(x[1].terminated.all()),
        body_fn,
        (rng_key, states, jnp.zeros(n_vmap)),
    )
    W, D, L = jnp.mean(reward == 1), jnp.mean(reward == 0), jnp.mean(reward == -1)
    return W, D, L


def enrich_log(log):
    iteration = log["cost/iteration"]
    sim_plan = 0
    sim_play = config.selfplay_vmap * config.selfplay_step * iteration
    log["cost/simulator_evaluations/planning"] = sim_plan
    log["cost/simulator_evaluations/playing"] = sim_play
    log["cost/simulator_evaluations/total"] = sim_plan + sim_play
    log["cost/simulator_evaluations/total [million]"] = (sim_plan + sim_play) / (10 ** 6)
    log["cost/hours/total"] = log["cost/hours/selfplay"] + log["cost/hours/preprocess"] + log["cost/hours/fit"]
    for key in ["selfplay", "fit", "total"]:
        log[f"cost/gpu_hours/{key}"] = log[f"cost/hours/{key}"]
    for opp in list(
        map(
            lambda key: key.split("/")[1],
            filter(lambda key: "vs_baseline" in key and "win_rate" in key, log.keys()),
        )
    ):
        W, D, L = log[f"eval/{opp}/win_rate"], log[f"eval/{opp}/draw_rate"], log[f"eval/{opp}/lose_rate"]
        log[f"eval/{opp}/avg_R"] = 1 * W + 0 * D + (-1) * L
        log[f"eval/{opp}/score"] = 1 * W + 0.5 * D + 0 * L
        log[f"score/{opp}"] = 1 * W + 0.5 * D + 0 * L
    log["train/total_loss"] = log["train/policy_loss"] + log["train/qvalue_loss"]
    log["stats/sample_util_ratio"] = log["cost/frames/used"] / log["cost/frames/total"]
    log["stats/effective_actions"] = log["stats/policy_target_mean_exp_entropy"]
    log["stats/effective_actions_v2"] = float(np.exp(log["stats/policy_target_mean_entropy"]))
    for s in ["return", "kl", "ent"]:
        log[f"selfplay_stats/{s}_diff"] = log[f"selfplay_stats/{s}_1"] - log[f"selfplay_stats/{s}_0"]
    return dict(sorted(log.items()))


def get_params(model):
    return {
        "trainable_variables": tuple(jnp.array(var.numpy()) for var in model.trainable_variables),
        "non_trainable_variables": tuple(jnp.array(var.numpy()) for var in model.non_trainable_variables),
    }


def qvalue_loss_fn(y_true, y_pred):
    A = y_true[:, 0].astype(int)
    G = y_true[:, 1]
    q_pred = y_pred[jnp.arange(y_pred.shape[0]), A]
    squared_error = jnp.square(q_pred - G)
    return squared_error


def entropy(p):
    return jnp.sum(jnp.where(p == 0, 0, -p * jnp.log(p)))


def kl_divergence(p, q):
    return jnp.sum(jnp.where(p == 0, 0, p * (jnp.log(p) - jnp.log(q))))


conf_dict = OmegaConf.from_cli()
config = Config(**conf_dict)
env = pgx.make(config.env_id)
prod_observation_shape = int(jnp.prod(jnp.array(env.observation_shape)))

model = PQNet(
    input_shape=env.observation_shape,
    num_actions=env.num_actions,
    zero_init=config.zero_init,
    num_channels=config.num_channels,
    num_blocks=config.num_blocks,
)

losses = {
    "logits": keras.losses.CategoricalCrossentropy(from_logits=True),
    "qvalue": qvalue_loss_fn,
}
model.compile(
    optimizer=getattr(keras.optimizers, config.optimizer)(learning_rate=config.learning_rate),
    loss=losses,
    metrics=losses,
)
config.num_params = model.count_params()

baseline_id = config.env_id + "_v0"
if baseline_id in get_args(pgx.BaselineModelId):
    baseline = pgx.make_baseline_model(baseline_id)
else:

    def baseline(obs):
        return jnp.zeros((obs.shape[0], env.num_actions)), None


ckpt_dir = os.path.join(config.checkpoints_dir, f"{config.env_id}_{config.seed}")
os.makedirs(ckpt_dir, exist_ok=True)

jit_vmap_decoder = jax.jit(jax.vmap(decode))


def data_generator(C, A, P, G, rng_key):
    N = len(C)
    batch_size = config.fitting_batch_size
    while True:
        rng_key, sub_key = jax.random.split(rng_key)
        idx = jax.random.permutation(sub_key, jnp.arange(N))
        for start in range(0, N, batch_size):
            c, a, p, g = (x[idx[start : start + batch_size]] for x in (C, A, P, G))
            o = jit_vmap_decoder(c)
            target = {"logits": p, "qvalue": jnp.concatenate([a, g], axis=1)}
            yield o, target


def main():
    if config.wandb_on:
        import wandb
        wandb.init(project="klent-ablation", config=config.model_dump())
    key = KeyGenerator(config.seed)
    hours_selfplay, hours_preprocess, hours_fit = 0.0, 0.0, 0.0
    frames_total, frames_used = 0, 0

    for iteration in itertools.count():
        t0 = time()
        selfplay_output = selfplay(key(), get_params(model))
        t1 = time()
        C, A, P, G, STATS = jax.device_get(selfplay_output)
        del selfplay_output
        jax.clear_caches()
        mask = jnp.squeeze(jnp.isfinite(G))
        C, A, P, G = map(lambda x: x[mask], (C, A, P, G))
        N = len(C)
        t2 = time()
        history = model.fit(
            x=data_generator(C, A, P, G, key()),
            steps_per_epoch=N // config.fitting_batch_size,
            epochs=config.fitting_epochs,
        )
        t3 = time()

        eval_log = {}
        for opp_coef in [1.00]:
            W, D, L = evaluate(key(), get_params(model), opp_coef)
            eval_log[f"eval/vs_baseline_{int(100*opp_coef):03}/win_rate"] = float(W)
            eval_log[f"eval/vs_baseline_{int(100*opp_coef):03}/draw_rate"] = float(D)
            eval_log[f"eval/vs_baseline_{int(100*opp_coef):03}/lose_rate"] = float(L)

        hours_selfplay += (t1 - t0) / 3600
        hours_preprocess += (t2 - t1) / 3600
        hours_fit += (t3 - t2) / 3600
        frames_total += config.selfplay_vmap * config.selfplay_step
        frames_used += N

        ENT = np.array(jax.lax.map(entropy, P))
        log = enrich_log(
            eval_log
            | {
                "cost/iteration": iteration + 1,
                "cost/hours/selfplay": hours_selfplay,
                "cost/hours/preprocess": hours_preprocess,
                "cost/hours/fit": hours_fit,
                "cost/frames/total": frames_total,
                "cost/frames/used": frames_used,
                "train/policy_loss": float(np.mean(history.history["logits_categorical_crossentropy"])),
                "train/qvalue_loss": float(np.mean(history.history["qvalue_qvalue_loss_fn"])),
                "stats/policy_target_mean_entropy": float(np.mean(ENT)),
                "stats/policy_target_mean_exp_entropy": float(np.mean(np.exp(ENT))),
                "selfplay_stats/return_0": float(STATS[0]),
                "selfplay_stats/return_1": float(STATS[1]),
                "selfplay_stats/kl_0": float(STATS[2]),
                "selfplay_stats/kl_1": float(STATS[3]),
                "selfplay_stats/ent_0": float(STATS[4]),
                "selfplay_stats/ent_1": float(STATS[5]),
            }
        )
        log["mode"] = config.mode
        log["env_id"] = config.env_id
        log["seed"] = config.seed
        print(json.dumps(log))
        if config.wandb_on:
            import wandb
            wandb.log(log)

        del C, A, P, G, ENT, STATS, W, D, L, history
        jax.clear_caches()

        if (iteration + 1) % config.save_interval == 0:
            model.save(ckpt_dir + f"/{iteration+1:04}.keras")
        if log["cost/simulator_evaluations/total"] >= config.limit_simulator_evaluations:
            model.save(ckpt_dir + "/final.keras")
            break


result = main()