inference.py

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

import numpy as np
import jax
import jax.numpy as jnp

from loading_model import generator
from configuration import LATENT_DIM, RESOLUTIONS

BASE_DIR       = os.path.dirname(os.path.abspath(__file__))
WEIGHTS_H5     = os.path.join(BASE_DIR, "generator.weights.h5")
CHECKPOINT_PKL = os.path.join(BASE_DIR, "checkpoint.pkl")

_gen_state     = None
_use_stateless = False
_saved_alpha   = 1.0
TARGET_RES     = 256


def _postprocess(images: np.ndarray) -> list:
    images = (np.array(images) + 1.0) / 2.0
    images = np.clip(images, 0.0, 1.0)
    return [images[i] for i in range(images.shape[0])]


def _build_model():
    dummy_alpha = jnp.array(1.0)
    dummy_rng   = jax.random.PRNGKey(0)
    dummy_z     = jnp.zeros((1, LATENT_DIM))
    for res in RESOLUTIONS:
        generator.current_resolution = res
        _ = generator(dummy_z, alpha=dummy_alpha, rng_key=dummy_rng)
    generator.current_resolution = TARGET_RES
    print(f"✅ Model built — trainable: {len(generator.trainable_variables)}  non-trainable: {len(generator.non_trainable_variables)}")


def load_weights(resolution=256):
    global _gen_state, _use_stateless, _saved_alpha
    _build_model()

    if os.path.isfile(WEIGHTS_H5):
        print(f"📂 Loading Keras weights: {WEIGHTS_H5}")
        generator.load_weights(WEIGHTS_H5)
        _use_stateless = False
        print("✅ Keras weights loaded.")

    elif os.path.isfile(CHECKPOINT_PKL):
        print(f"📂 Loading checkpoint: {CHECKPOINT_PKL}")
        import joblib
        data = joblib.load(CHECKPOINT_PKL)
        print(f"   Keys: {list(data.keys())}")

        # Support both key naming conventions
        gen_trainable = (
            data.get("ema_trainable") or      # ← new key from your training code
            data.get("gen_trainable")          # ← old key
        )
        gen_non_trainable = (
            data.get("gen_non_trainable") or
            data.get("non_trainable", [])
        )
        _saved_alpha = float(data.get("alpha", 1.0))

        print(f"   trainable={len(gen_trainable)}  non_trainable={len(gen_non_trainable)}  alpha={_saved_alpha:.4f}")

        n_model = len(generator.trainable_variables)
        if len(gen_trainable) != n_model:
            raise ValueError(
                f"❌ Mismatch — checkpoint:{len(gen_trainable)} model:{n_model}\n"
                f"   Make sure _build_model() mirrors training exactly."
            )

        n_non = len(generator.non_trainable_variables)
        _gen_state = {
            "ema_trainable": [jnp.asarray(t) for t in gen_trainable],
            "non_trainable": (
                [jnp.asarray(t) for t in gen_non_trainable[:n_non]]
                if gen_non_trainable
                else [jnp.asarray(v.value) for v in generator.non_trainable_variables]
            ),
        }
        _use_stateless = True
        print("✅ Checkpoint loaded.")

    else:
        raise FileNotFoundError(f"❌ No weights found in: {BASE_DIR}")

    generator.current_resolution = TARGET_RES
    print(f"🖼️  Resolution locked to: {TARGET_RES}")
    return True


def generate_images(n_images, resolution=None, seed=None):
    if not _use_stateless and _gen_state is None:
        raise RuntimeError("❌ Call load_weights() first.")
    if seed is None:
        seed = int(np.random.randint(0, 2**31))
    rng = jax.random.PRNGKey(seed)
    if _use_stateless and _gen_state is not None:
        return _generate_stateless(n_images, rng)
    return _generate_keras(n_images, rng)


def _generate_keras(n_images, rng):
    generator.current_resolution = TARGET_RES
    rng, z_key, noise_key = jax.random.split(rng, 3)
    z   = jax.random.normal(z_key, (n_images, LATENT_DIM))
    out = generator(z, alpha=1.0, rng_key=noise_key)
    images = out[0] if isinstance(out, (list, tuple)) else out
    return _postprocess(np.array(images))


def _generate_stateless(n_images, rng):
    generator.current_resolution = TARGET_RES
    rng, z_key, noise_key = jax.random.split(rng, 3)
    z = jax.random.normal(z_key, (n_images, LATENT_DIM))
    images, _ = generator.stateless_call(
        _gen_state["ema_trainable"],
        _gen_state["non_trainable"],
        z,
        jnp.array(_saved_alpha),
        noise_key,
    )
    return _postprocess(np.array(images))