Can not align FID with provided checkpoint

[LiCHH]

Hello, I wrote a script based on the demo_sample.ipynb to generate 50,000 samples and tested them using OpenAI's FID evaluation toolkit. However, I found that the metrics did not align. Could you help me identify the problem?
I got

by using d20 checkpoint and

by using d30 checkpoint.
The script is as below:

################## 1. Download checkpoints and build models
import os
import os.path as osp
import torch, torchvision
import random
from tqdm import tqdm
import numpy as np
import PIL.Image as PImage, PIL.ImageDraw as PImageDraw
setattr(torch.nn.Linear, 'reset_parameters', lambda self: None)     # disable default parameter init for faster speed
setattr(torch.nn.LayerNorm, 'reset_parameters', lambda self: None)  # disable default parameter init for faster speed
from models import VQVAE, build_vae_var

MODEL_DEPTH = 20    # TODO: =====> please specify MODEL_DEPTH <=====
assert MODEL_DEPTH in {16, 20, 24, 30}


# download checkpoint
hf_home = 'https://huggingface.co/FoundationVision/var/resolve/main'
vae_ckpt, var_ckpt = 'vae_ch160v4096z32.pth', f'var_d{MODEL_DEPTH}.pth'
if not osp.exists(vae_ckpt): os.system(f'wget {hf_home}/{vae_ckpt}')
if not osp.exists(var_ckpt): os.system(f'wget {hf_home}/{var_ckpt}')

# build vae, var
patch_nums = (1, 2, 3, 4, 5, 6, 8, 10, 13, 16)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
if 'vae' not in globals() or 'var' not in globals():
    vae, var = build_vae_var(
        V=4096, Cvae=32, ch=160, share_quant_resi=4,    # hard-coded VQVAE hyperparameters
        device=device, patch_nums=patch_nums,
        num_classes=1000, depth=MODEL_DEPTH, shared_aln=False,
    )

# load checkpoints
vae.load_state_dict(torch.load(vae_ckpt, map_location='cpu'), strict=True)
var.load_state_dict(torch.load(var_ckpt, map_location='cpu'), strict=True)
vae.eval(), var.eval()
for p in vae.parameters(): p.requires_grad_(False)
for p in var.parameters(): p.requires_grad_(False)
print(f'prepare finished.')

############################# 2. Sample with classifier-free guidance

# set args
seed = 1 #@param {type:"number"}
torch.manual_seed(seed)
num_sampling_steps = 250 #@param {type:"slider", min:0, max:1000, step:1}
cfg = 1.5 #@param {type:"slider", min:1, max:10, step:0.1}
more_smooth = False # True for more smooth output

# seed
torch.manual_seed(seed)
random.seed(seed)
np.random.seed(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False

# run faster
tf32 = True
torch.backends.cudnn.allow_tf32 = bool(tf32)
torch.backends.cuda.matmul.allow_tf32 = bool(tf32)
torch.set_float32_matmul_precision('high' if tf32 else 'highest')

# sample
B = 25
for img_cls in tqdm(range(1000)):
    for i in range(50 // B):
        label_B = torch.tensor([img_cls] * 25, device=device)
        # B = len(class_labels)
        # label_B: torch.LongTensor = torch.tensor(class_labels, device=device)
        with torch.inference_mode():
            with torch.autocast('cuda', enabled=True, dtype=torch.float16, cache_enabled=True):    # using bfloat16 can be faster
                recon_B3HW = var.autoregressive_infer_cfg(B=B, label_B=label_B, cfg=cfg, top_k=900, top_p=0.96, g_seed=seed, more_smooth=more_smooth)
            bchw = recon_B3HW.permute(0, 2, 3, 1).mul_(255).cpu().numpy()
        bchw = bchw.astype(np.uint8)
        for j in range(B):
            img = PImage.fromarray(bchw[j])
            img.save(f"./samples_d20/sample_{img_cls * 50 + i * B + j}.png")

[ma-xu]

@keyu-tian Thanks for your great work! could you please help with this issue? Any insights could be helpful.

[Kumbong]

+1 Thanks for the great work! wondering if you are willing to share your evaluation scripts ?

[ChenDRAG]

@LiCHH @ma-xu @Kumbong . It seems the generation seed is replicate for each class.
recon_B3HW = var.autoregressive_infer_cfg(B=B, label_B=label_B, cfg=cfg, top_k=900, top_p=0.96, g_seed=seed, more_smooth=more_smooth)

gseed should be different . I would use

    for i in range(50 // B):
        label_B = torch.tensor([img_cls] * 25, device=device)
        # B = len(class_labels)
        # label_B: torch.LongTensor = torch.tensor(class_labels, device=device)
        with torch.inference_mode():
            with torch.autocast('cuda', enabled=True, dtype=torch.float16, cache_enabled=True):    # using bfloat16 can be faster
                recon_B3HW = var.autoregressive_infer_cfg(B=B, label_B=label_B, cfg=cfg, top_k=900, top_p=0.96, g_seed=seed + i, more_smooth=more_smooth)

[Tom-zgt]

I tried your suggestion, but the results got worse. The randomness of c2i has already been introduced by rng, so shouldn't it be unnecessary to change the random seed? Or could you share your evaluation code with me? @ChenDRAG @keyu-tian

[sen-ye]

You can try to use B=50, since random seed is shared across different batches.

[adreamwu]

You can try to use B=50, since random seed is shared across different batches.

Have you been able to reproduce the results in Table 1 of the paper? Could you share the inference script?

We use B=50 for each class and var_d16 for evaluation.

• report

FID	IS	Pre	Rec
3.60	257.5	0.85	0.48

• reproduced

FID	IS	Pre	Rec
3.49	281.71	0.85	0.50

The main issue is the IS. The results on other metrics are similar to those in the paper. Thanks.