config: add canonical model configuration
Browse files- config.yaml +86 -0
config.yaml
ADDED
|
@@ -0,0 +1,86 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
_target_: fastwam.rift.create_fastwam_anticip
|
| 2 |
+
|
| 3 |
+
# Canonical RIFT: full anticipation grid, two-pass video co-training,
|
| 4 |
+
# flow-matching render objective, and a detached L2 diagnostic probe retained
|
| 5 |
+
# for checkpoint-schema compatibility.
|
| 6 |
+
anticip_future_tokens: 196
|
| 7 |
+
anticip_latent_t: 3
|
| 8 |
+
anticip_lambda: 1.0
|
| 9 |
+
anticip_action_lambda: 1.0
|
| 10 |
+
|
| 11 |
+
# Motion-aware target and late loss rebalance.
|
| 12 |
+
anticip_delta_mix: 0.8
|
| 13 |
+
anticip_lambda_final: 0.2
|
| 14 |
+
anticip_anneal_start_frac: 0.7
|
| 15 |
+
anticip_anneal_total_steps: 21700
|
| 16 |
+
|
| 17 |
+
# Late render-only conditioning-noise curriculum.
|
| 18 |
+
anticip_cond_noise_p: 0.3
|
| 19 |
+
anticip_cond_noise_sigma: 0.06
|
| 20 |
+
anticip_cond_noise_ramp_start_frac: 0.7
|
| 21 |
+
anticip_cond_noise_action_free: true
|
| 22 |
+
|
| 23 |
+
# The FM head is the released representation-shaping head. The L2 probe
|
| 24 |
+
# remains a detached diagnostic probe and does not shape anticipation tokens.
|
| 25 |
+
anticip_fm_lambda: 1.0
|
| 26 |
+
anticip_fm_width: 512
|
| 27 |
+
anticip_fm_blocks: 2
|
| 28 |
+
anticip_fm_lambda_final: 0.2
|
| 29 |
+
|
| 30 |
+
model_id: Wan-AI/Wan2.2-TI2V-5B
|
| 31 |
+
tokenizer_model_id: Wan-AI/Wan2.1-T2V-1.3B
|
| 32 |
+
tokenizer_max_len: 128
|
| 33 |
+
load_text_encoder: false
|
| 34 |
+
proprio_dim: ${data.train.processor.proprio_output_dim}
|
| 35 |
+
redirect_common_files: true
|
| 36 |
+
mot_checkpoint_mixed_attn: true
|
| 37 |
+
action_dit_pretrained_path: checkpoints/ActionDiT_linear_interp_Wan22_alphascale_1024hdim.pt
|
| 38 |
+
skip_dit_load_from_pretrain: false
|
| 39 |
+
|
| 40 |
+
video_dit_config:
|
| 41 |
+
has_image_input: false
|
| 42 |
+
patch_size: [1, 2, 2]
|
| 43 |
+
in_dim: 48
|
| 44 |
+
hidden_dim: 3072
|
| 45 |
+
ffn_dim: 14336
|
| 46 |
+
freq_dim: 256
|
| 47 |
+
text_dim: 4096
|
| 48 |
+
out_dim: 48
|
| 49 |
+
num_heads: 24
|
| 50 |
+
attn_head_dim: 128
|
| 51 |
+
num_layers: 30
|
| 52 |
+
eps: 1.0e-06
|
| 53 |
+
seperated_timestep: true
|
| 54 |
+
require_clip_embedding: false
|
| 55 |
+
require_vae_embedding: false
|
| 56 |
+
fuse_vae_embedding_in_latents: true
|
| 57 |
+
use_gradient_checkpointing: ${model.mot_checkpoint_mixed_attn}
|
| 58 |
+
video_attention_mask_mode: first_frame_causal
|
| 59 |
+
action_conditioned: false
|
| 60 |
+
action_dim: ${data.train.processor.action_output_dim}
|
| 61 |
+
action_group_causal_mask_mode: group_diagonal
|
| 62 |
+
|
| 63 |
+
action_dit_config:
|
| 64 |
+
action_dim: ${data.train.processor.action_output_dim}
|
| 65 |
+
hidden_dim: 1024
|
| 66 |
+
ffn_dim: 4096
|
| 67 |
+
num_heads: 24
|
| 68 |
+
attn_head_dim: 128
|
| 69 |
+
num_layers: 30
|
| 70 |
+
text_dim: 4096
|
| 71 |
+
freq_dim: 256
|
| 72 |
+
eps: 1.0e-06
|
| 73 |
+
use_gradient_checkpointing: ${model.mot_checkpoint_mixed_attn}
|
| 74 |
+
|
| 75 |
+
video_scheduler:
|
| 76 |
+
train_shift: 5.0
|
| 77 |
+
infer_shift: 5.0
|
| 78 |
+
num_train_timesteps: 1000
|
| 79 |
+
|
| 80 |
+
action_scheduler:
|
| 81 |
+
train_shift: 5.0
|
| 82 |
+
infer_shift: 5.0
|
| 83 |
+
num_train_timesteps: 1000
|
| 84 |
+
|
| 85 |
+
loss:
|
| 86 |
+
lambda_action: 1.0
|