slat_vae_dec_rf_swin8_B_64l8_fp16.json

{
    "models": {
        "decoder": {
            "name": "ElasticSLatRadianceFieldDecoder",
            "args": {
                "resolution": 64,
                "model_channels": 768,
                "latent_channels": 8,
                "num_blocks": 12,
                "num_heads": 12,
                "mlp_ratio": 4,
                "attn_mode": "swin",
                "window_size": 8,
                "use_fp16": true,
                "representation_config": {
                    "rank": 16,
                    "dim": 8
                }
            }
        }
    },
    "dataset": {
        "name": "SLat2Render",
        "args": {
            "image_size": 512,
            "latent_model": "dinov2_vitl14_reg_slat_enc_swin8_B_64l8_fp16",
            "min_aesthetic_score": 4.5,
            "max_num_voxels": 32768
        }
    },
    "trainer": {
        "name": "SLatVaeRadianceFieldDecoderTrainer",
        "args": {
            "max_steps": 1000000,
            "batch_size_per_gpu": 4,
            "batch_split": 2,
            "optimizer": {
                "name": "AdamW",
                "args": {
                    "lr": 1e-4,
                    "weight_decay": 0.0
                }
            },
            "ema_rate": [
                0.9999
            ],
            "fp16_mode": "inflat_all",
            "fp16_scale_growth": 0.001,
            "elastic": {
                "name": "LinearMemoryController",
                "args": {
                    "target_ratio": 0.75,
                    "max_mem_ratio_start": 0.5
                }
            },
            "grad_clip": {
                "name": "AdaptiveGradClipper",
                "args": {
                    "max_norm": 1.0,
                    "clip_percentile": 95
                }
            },
            "i_log": 500,
            "i_sample": 10000,
            "i_save": 10000,
            "loss_type": "l1",
            "lambda_ssim": 0.2,
            "lambda_lpips": 0.2
        }
    }
}