把instant-ngp中编码方式改为hashgrid之后，torsor阶段的预测质量很低是为什么

[zckcola]

这是改成hash以后torsor阶段的 这是head的 这是config文件accumulate_grad_batches: 1 ambient_out_dim: 2 amp: true base_config:

• ./lm3d_radnerf_hash.yaml binary_data_dir: data/binary/videos bound: 1 camera_offset:
• 0
• 0
• 0 camera_scale: 4.0 clip_grad_norm: 0 clip_grad_value: 0 cond_out_dim: 64 cond_type: idexp_lm3d_normalized cond_win_size: 1 cuda_ray: true debug: false density_thresh: 10 density_thresh_torso: 0.01 desired_resolution: 2048 dt_gamma: 0.00390625 eval_max_batches: 100 far: 0.9 finetune_lips: true finetune_lips_start_iter: 200000 geo_feat_dim: 128 grid_interpolation_type: linear grid_size: 128 grid_type: hashgrid gui_fovy: 21.24 gui_h: 512 gui_max_spp: 1 gui_radius: 3.35 gui_w: 512 head_model_dir: checkpoints/May/lm3d_radnerf_hash hidden_dim_ambient: 128 hidden_dim_color: 128 hidden_dim_sigma: 128 individual_embedding_dim: 4 individual_embedding_num: 13000 infer_audio_source_name: '' infer_bg_img_fname: '' infer_c2w_name: '' infer_cond_name: '' infer_lm3d_clamp_std: 2.5 infer_lm3d_lle_percent: 0.0 infer_lm3d_smooth_sigma: 0.0 infer_out_video_name: '' infer_scale_factor: 1.0 infer_smo_std: 0.0 infer_smooth_camera_path: true infer_smooth_camera_path_kernel_size: 7 lambda_ambient: 0.1 lambda_lpips_loss: 0.01 lambda_weights_entropy: 0.0001 load_ckpt: '' load_imgs_to_memory: false log2_hashmap_size: 16 lr: 0.0005 max_ray_batch: 4096 max_steps: 16 max_updates: 250000 min_near: 0.05 n_rays: 65536 near: 0.3 num_ckpt_keep: 1 num_layers_ambient: 3 num_layers_color: 2 num_layers_sigma: 3 num_sanity_val_steps: 2 num_steps: 16 num_valid_plots: 5 optimizer_adam_beta1: 0.9 optimizer_adam_beta2: 0.999 print_nan_grads: false processed_data_dir: data/processed/videos raw_data_dir: data/raw/videos resume_from_checkpoint: 0 save_best: true save_codes:
• tasks
• modules
• egs save_gt: true scheduler: exponential seed: 9999 smo_win_size: 5 smooth_lips: false task_cls: tasks.radnerfs.radnerf_torso.RADNeRFTorsoTask tb_log_interval: 100 torso_head_aware: false torso_individual_embedding_dim: 8 torso_shrink: 0.8 torso_train_mode: 1 update_extra_interval: 16 upsample_steps: 0 use_window_cond: true val_check_interval: 2000 valid_infer_interval: 10000 valid_monitor_key: val_loss valid_monitor_mode: min video_id: May warmup_updates: 0 weight_decay: 0 with_att: true work_dir: checkpoints/May/lm3d_radnerf_torso_hash