code/RAEv2/src/eval/generation.py

"""Distributed generation evaluation — FID, CLIPScore, VQAScore, GenEval, DPG-Bench."""

import os
import sys
from typing import Dict, List, Optional, Union

import numpy as np
import torch
import torch.distributed as dist
from torch.cuda.amp import autocast
from tqdm import tqdm

try:
    from .clipscore import CLIPScoreEvaluator
except Exception:
    CLIPScoreEvaluator = None
from .distributed import create_eval_dataloader, gather_and_cleanup_shards, setup_eval_tmpdir
from .distributional import compute_distributional_metrics, filter_distributional
try:
    from .dpgbench import DPGEvaluator
except Exception:
    DPGEvaluator = None
try:
    from .geneval import GenEvalEvaluator
except Exception:
    GenEvalEvaluator = None
from .lpips import LPIPSEvaluator
try:
    from .vqascore import VQAScoreEvaluator
except Exception:
    VQAScoreEvaluator = None


def evaluate_image_set(
    images: np.ndarray,
    *,
    metrics_to_compute: List[str],
    reference_npz_path: Optional[Union[str, List[str]]] = None,
    data_dir: Optional[str] = None,
    device: torch.device,
    metric_batch_size: int = 128,
) -> Dict[str, float]:
    """Run rank-0 generation metrics via fd_evaluator on a uint8 NHWC array.

    Called both from the distributed-generation path (after shard gather) and
    the offline_eval `--npz` short-circuit. `reference_npz_path` overrides the
    auto-resolved FID stats; pass None to use the catalogue default.
    """
    distributional = filter_distributional(metrics_to_compute)
    if not distributional:
        return {}
    ref = reference_npz_path[0] if isinstance(reference_npz_path, list) else reference_npz_path
    return compute_distributional_metrics(
        images,
        distributional,
        reference_npz=ref,
        data_dir=data_dir,
        device=device,
        batch_size=metric_batch_size,
    )


def _init_evaluators(metrics_to_compute: List[str], condition_type: str, device: torch.device):
    """Initialize metric evaluators and local score accumulators."""
    evaluators = {}
    local_scores = {}

    if 'clipscore' in metrics_to_compute and condition_type == 'text':
        evaluators['clipscore'] = CLIPScoreEvaluator(device=str(device))
        local_scores['clipscore'] = {'sum': 0.0, 'count': 0}

    if any(elem.startswith('vqascore') for elem in metrics_to_compute) and condition_type == 'text':
        vqascore_models = [elem for elem in metrics_to_compute if elem.startswith('vqascore')]
        vqascore_evaluators = {}
        for model_name in vqascore_models:
            model_name_ = model_name.split('_')[-1] if '_' in model_name else 'clip-flant5-xl'
            vqascore_evaluators[model_name] = VQAScoreEvaluator(model_name=model_name_, device=str(device))
            local_scores[model_name] = {'sum': 0.0, 'count': 0}
        evaluators['vqascore'] = vqascore_evaluators

    if 'geneval' in metrics_to_compute and condition_type == 'text':
        evaluators['geneval'] = GenEvalEvaluator(device=str(device))
        local_scores['geneval'] = {'sum': 0.0, 'count': 0}

    if 'dpgbench' in metrics_to_compute and condition_type == 'text':
        evaluators['dpgbench'] = DPGEvaluator(device=str(device))
        local_scores['dpgbench'] = {'sum': 0.0, 'count': 0}

    if 'lpips' in metrics_to_compute and condition_type == 'nwm':
        evaluators['lpips'] = LPIPSEvaluator(device=str(device))
        local_scores['lpips'] = {'sum': 0.0, 'count': 0}

    return evaluators, local_scores


def _aggregate_distributed_metrics(local_scores: dict, device: torch.device) -> Dict[str, float]:
    """All-reduce local score sums/counts across ranks and return averaged metrics."""
    metrics = {}
    for metric_name, scores in local_scores.items():
        sum_tensor = torch.tensor([scores['sum']], device=device, dtype=torch.float64)
        count_tensor = torch.tensor([scores['count']], device=device, dtype=torch.float64)
        dist.all_reduce(sum_tensor, op=dist.ReduceOp.SUM)
        dist.all_reduce(count_tensor, op=dist.ReduceOp.SUM)
        if count_tensor.item() > 0:
            metrics[metric_name] = sum_tensor.item() / count_tensor.item()
    return metrics


@torch.no_grad()
def evaluate_generation_distributed(
    model_fn,
    sample_fn,
    latent_size,
    additional_model_kwargs,
    use_guidance: bool,
    rae,
    val_dataset,
    num_samples: int,
    batch_size: int,
    rank: int,
    world_size: int,
    device: torch.device,
    experiment_dir: str,
    global_step: int,
    autocast_kwargs: dict,
    metric_batch_size: int = 128,
    reference_npz_path: Optional[Union[str, List[str]]] = None,
    shared_tmpdir: Optional[str] = None,
    condition_type: str = "label",
    null_label: int = 1000,
    text_encoder=None,
    metrics_to_compute: Optional[List[str]] = None,
    data_dir: Optional[str] = None,
) -> Optional[Dict[str, float]]:
    """
    Evaluate generation metrics using all GPUs in a distributed manner.

    Args:
        model_fn: Model forward function
        sample_fn: Sampling function
        latent_size: Shape of latent noise
        additional_model_kwargs: Additional kwargs for model forward
        use_guidance: Whether to use classifier-free guidance
        rae: RAE model for decoding latents to images
        val_dataset: Validation dataset (returns (image, label) or (image, text))
        num_samples: Number of samples to generate
        batch_size: Batch size per GPU for generation
        rank: Current GPU rank
        world_size: Total number of GPUs
        device: Device to use
        experiment_dir: Experiment directory
        global_step: Current training step
        autocast_kwargs: Autocast configuration
        metric_batch_size: Batch size for metric computation (on rank 0)
        reference_npz_path: Optional path (or list of paths) to existing reference NPZ
            files. If a list, FID is computed once per reference and emitted as
            fid_<tag> (tag derived from filename: jit, adm, or stem); fid is set
            to the first reference's value for backwards compatibility.
        shared_tmpdir: Optional shared directory for multi-node eval
        condition_type: Type of conditioning - "label" or "text"
        null_label: Null label index for CFG (label conditioning only)
        text_encoder: Text encoder for text conditioning (required if condition_type="text")
        metrics_to_compute: List of metrics to compute (default: ['fid'])

    Returns:
        Dictionary of metrics (only on rank 0, None on other ranks)
    """
    temp_dir = setup_eval_tmpdir(experiment_dir, global_step, rank,
                                  shared_tmpdir=shared_tmpdir, eval_type="sampling")
    loader = create_eval_dataloader(val_dataset, rank, world_size, num_samples, batch_size)

    # Initialize evaluators
    if metrics_to_compute is None:
        metrics_to_compute = ['fid']
    evaluators, local_scores = _init_evaluators(metrics_to_compute, condition_type, device)

    # Generate images on this rank
    generations = []
    iterator = tqdm(loader, desc=f"[Rank {rank}] Sampling", file=sys.stdout) if rank == 0 else loader

    with torch.inference_mode():
        for gt_img, cond in iterator:
            # Handle conditioning based on type
            if condition_type == "text":
                n = len(cond)
                z = torch.randn(n, *latent_size, device=device)
                enc_out = text_encoder(list(cond))
                context = enc_out["tokens"]
                context_attn_mask = enc_out["attention_mask"]
                if use_guidance:
                    z = torch.cat([z, z], dim=0)
                    enc_null = text_encoder([""] * n)
                    context_null = enc_null["tokens"]
                    context_attn_mask_null = enc_null["attention_mask"]
                    context = torch.cat([context, context_null], dim=0)
                    context_attn_mask = torch.cat([context_attn_mask, context_attn_mask_null], dim=0)
            elif condition_type == "nwm":
                from stage2 import nwm_cond
                z, context, context_attn_mask, n = nwm_cond.encode_eval_context(
                    cond, rae, device, latent_size, use_guidance,
                )
            else:
                n = cond.size(0)
                z = torch.randn(n, *latent_size, device=device)
                context = cond.to(device)
                context_attn_mask = None
                if use_guidance:
                    z = torch.cat([z, z], dim=0)
                    context_null = torch.full((n,), null_label, device=device)
                    context = torch.cat([context, context_null], dim=0)

            model_kwargs = dict(context=context, attn_mask=context_attn_mask, **additional_model_kwargs)
            with autocast(**autocast_kwargs):
                samples = sample_fn(z, model_fn, **model_kwargs)[-1]
                if use_guidance:
                    samples = samples.chunk(2, dim=0)[0]
                samples = rae.decode(samples).clamp(0, 1)
            gen_np = samples.mul(255).permute(0, 2, 3, 1).to("cpu", dtype=torch.uint8).numpy()

            # Compute distributed metrics during generation
            if 'clipscore' in evaluators:
                batch_scores = evaluators['clipscore'].compute_batch_scores(gen_np, list(cond))
                local_scores['clipscore']['sum'] += batch_scores.sum().item()
                local_scores['clipscore']['count'] += len(cond)
            if 'vqascore' in evaluators:
                for model_name, evaluator in evaluators['vqascore'].items():
                    batch_scores = evaluator.compute_batch_scores(gen_np, list(cond))
                    local_scores[model_name]['sum'] += batch_scores.sum().item()
                    local_scores[model_name]['count'] += len(cond)
            if 'geneval' in evaluators:
                batch_scores = evaluators['geneval'].compute_batch_scores(gen_np, list(cond))
                local_scores['geneval']['sum'] += batch_scores.sum().item()
                local_scores['geneval']['count'] += len(cond)
            if 'dpgbench' in evaluators:
                batch_scores = evaluators['dpgbench'].compute_batch_scores(gen_np, list(cond))
                local_scores['dpgbench']['sum'] += batch_scores.sum().item()
                local_scores['dpgbench']['count'] += len(cond)
            if 'lpips' in evaluators:
                gt_np = gt_img.clamp(0, 1).mul(255).permute(0, 2, 3, 1).to("cpu", dtype=torch.uint8).numpy()
                batch_scores = evaluators['lpips'].compute_batch_scores(gen_np, gt_np)
                local_scores['lpips']['sum'] += batch_scores.sum().item()
                local_scores['lpips']['count'] += gen_np.shape[0]

            for img in gen_np:
                generations.append(img)

    generations = np.stack(generations)
    shard_path = os.path.join(temp_dir, f"gen_{global_step:07d}_{rank:02d}.npz")
    np.savez(shard_path, arr_0=generations)

    if rank == 0:
        print(f"[Rank {rank}] Saved {len(generations)} generation to {shard_path}")

    # Wait for all ranks to finish generation
    dist.barrier()

    # Distributed metrics: all_reduce sum and count across all ranks
    metrics = _aggregate_distributed_metrics(local_scores, device)

    # Rank 0 computes FID (requires gathering all samples)
    save_gen_npz = os.environ.get("SAVE_GEN_NPZ")
    if rank == 0:
        need_combined = (
            'fid' in metrics_to_compute
            or 'inception_score' in metrics_to_compute
            or bool(filter_distributional(metrics_to_compute))
            or save_gen_npz
        )
        if need_combined:
            combined_recons = gather_and_cleanup_shards(temp_dir, "gen", global_step, world_size, num_samples)
            print(f"[Eval] Combined generation NPZ shape: {combined_recons.shape}")

            if save_gen_npz:
                os.makedirs(os.path.dirname(save_gen_npz), exist_ok=True)
                np.savez(save_gen_npz, arr_0=combined_recons)
                print(f"[Eval] Saved gen NPZ to {save_gen_npz}")

            metrics.update(
                evaluate_image_set(
                    combined_recons,
                    metrics_to_compute=metrics_to_compute,
                    reference_npz_path=reference_npz_path,
                    data_dir=data_dir,
                    device=device,
                    metric_batch_size=metric_batch_size,
                )
            )
        else:
            for r in range(world_size):
                shard_file = os.path.join(temp_dir, f"gen_{global_step:07d}_{r:02d}.npz")
                if os.path.exists(shard_file):
                    os.remove(shard_file)

        # Print results
        print(f"[Eval] Step {global_step} Metrics:")
        for key, value in metrics.items():
            print(f"  {key}: {value:.6f}")

    dist.barrier()
    return metrics if metrics else None