Upload RenderFormer raynerf4 config
Browse files
config_hybrid_cam_20l_raynerf4.json
ADDED
|
@@ -0,0 +1,97 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_comment": "Hybrid linear/softmax RenderFormer (20 layers). Alternates VDA linear attention (even layers) and full softmax attention (odd layers) across both radiance encoder (20 layers) and camera decoder (20 layers). ray_pos_mode is camera (RoPE over 3D camera origin, Swin windows for image space). ABLATION raynerf4: per-pixel ray direction gets NeRF positional encoding with L=4 frequencies (include_input=True), ray_map_encoder input becomes 1728-dim. Fresh from scratch (ray_map_encoder shape differs from all existing checkpoints).",
|
| 3 |
+
"_comment_attnres": "attn_res_block_size 4 gives 5 banked checkpoints per 20-layer stack.",
|
| 4 |
+
"_comment_linear": "Gate on the linear modules only, fla's low-rank g_a/g_b (rank = head_dim). delta_rule is vdn_solve: VDN's unscaled ridge S = B(I+A)^-1, the rule their released checkpoints train with, routed through RobustVdnDelta (same math, fp64 Cholesky + cholesky_ex with escalating jitter on the rare indefinite batch).",
|
| 5 |
+
"_comment_position": "ray_pos_mode camera, as the dense reference uses: every ray patch takes the camera origin, so RoPE contributes among 3D relative camera-to-voxel displacements while Swin windows carry the image-space structure. Voxels rotate in half-dims [20,50).",
|
| 6 |
+
"_comment_schedule": "0-based indices taking the linear operator. 10 linear layers [0, 2, 4, 6, 8, 10, 12, 14, 16, 18] and 10 softmax layers [1, 3, 5, 7, 9, 11, 13, 15, 17, 19]. Encoder finishes with softmax at layer 19 so the context write is full-rank. Decoder taps [4, 9, 14, 19] into DPT at quarter intervals.",
|
| 7 |
+
"activation": "situ",
|
| 8 |
+
"additive_pos_encoding": true,
|
| 9 |
+
"attn_res_block_size": 4,
|
| 10 |
+
"attn_variant": "hybrid",
|
| 11 |
+
"bias": false,
|
| 12 |
+
"decoder_linear_layers": [
|
| 13 |
+
0,
|
| 14 |
+
2,
|
| 15 |
+
4,
|
| 16 |
+
6,
|
| 17 |
+
8,
|
| 18 |
+
10,
|
| 19 |
+
12,
|
| 20 |
+
14,
|
| 21 |
+
16,
|
| 22 |
+
18
|
| 23 |
+
],
|
| 24 |
+
"dim_feedforward": 4032,
|
| 25 |
+
"dpt_features": 256,
|
| 26 |
+
"dpt_out_channels": [
|
| 27 |
+
256,
|
| 28 |
+
512,
|
| 29 |
+
1024,
|
| 30 |
+
1024
|
| 31 |
+
],
|
| 32 |
+
"dpt_out_layers": [
|
| 33 |
+
4,
|
| 34 |
+
9,
|
| 35 |
+
14,
|
| 36 |
+
19
|
| 37 |
+
],
|
| 38 |
+
"dropout": 0.0,
|
| 39 |
+
"encoder_linear_layers": [
|
| 40 |
+
0,
|
| 41 |
+
2,
|
| 42 |
+
4,
|
| 43 |
+
6,
|
| 44 |
+
8,
|
| 45 |
+
10,
|
| 46 |
+
12,
|
| 47 |
+
14,
|
| 48 |
+
16,
|
| 49 |
+
18
|
| 50 |
+
],
|
| 51 |
+
"include_alpha": false,
|
| 52 |
+
"latent_dim": 1024,
|
| 53 |
+
"latent_grid": 32,
|
| 54 |
+
"linear_conv_checkpoint": true,
|
| 55 |
+
"linear_conv_projections": [
|
| 56 |
+
"q",
|
| 57 |
+
"k",
|
| 58 |
+
"v"
|
| 59 |
+
],
|
| 60 |
+
"linear_conv_radius": 1,
|
| 61 |
+
"linear_delta_rule": "vdn_solve",
|
| 62 |
+
"linear_gate_rank": "low",
|
| 63 |
+
"linear_solve_dtype": "fp64",
|
| 64 |
+
"norm_first": true,
|
| 65 |
+
"norm_type": "rms_norm",
|
| 66 |
+
"num_heads": 8,
|
| 67 |
+
"num_layers": 20,
|
| 68 |
+
"num_register_tokens": 32,
|
| 69 |
+
"patch_size": 8,
|
| 70 |
+
"pe_type": "rope",
|
| 71 |
+
"qk_norm": true,
|
| 72 |
+
"ray_pos_mode": "camera",
|
| 73 |
+
"rope_double_max_freq": true,
|
| 74 |
+
"rope_ray_start_index": 0,
|
| 75 |
+
"rope_type": "triangle",
|
| 76 |
+
"rope_voxel_start_index": 20,
|
| 77 |
+
"texture_channels": 13,
|
| 78 |
+
"texture_encode_patch_size": 1,
|
| 79 |
+
"texture_encoder_norm_type": "rms_norm",
|
| 80 |
+
"turn_to_cam_coord": true,
|
| 81 |
+
"use_dpt_decoder": true,
|
| 82 |
+
"use_ldr": false,
|
| 83 |
+
"use_vn_encoder": true,
|
| 84 |
+
"vdir_num_freqs": 4,
|
| 85 |
+
"vdir_pe_type": "nerf",
|
| 86 |
+
"vertex_pe_num_freqs": 12,
|
| 87 |
+
"view_indep_qk_norm": true,
|
| 88 |
+
"view_transformer_ffn_hidden_dim": 4032,
|
| 89 |
+
"view_transformer_include_self_attn": true,
|
| 90 |
+
"view_transformer_latent_dim": 1024,
|
| 91 |
+
"view_transformer_n_heads": 8,
|
| 92 |
+
"view_transformer_n_layers": 20,
|
| 93 |
+
"view_transformer_use_swin_attn": true,
|
| 94 |
+
"view_transformer_window_size": 8,
|
| 95 |
+
"vn_encoder_norm_type": "rms_norm",
|
| 96 |
+
"vn_pe_num_freqs": 6
|
| 97 |
+
}
|