video_backbone剖析

unitree_z1_dual_arm_stackbox_v2/case2/output/sparse_8/inference/round_summary.csv

@@ -0,0 +1,12 @@
+sample_id,scene,round_id,policy_pass_total_time_s,world_model_pass_total_time_s,round_total_time_s,latent_init_dist_to_prev_round,action_drift_vs_prev_round,psnr_full50
+unitree_z1_dual_arm_stackbox_v2-vid15-fs4,unitree_z1_dual_arm_stackbox_v2,0,34.487425088882446,33.954912424087524,142.91271448135376,,,
+unitree_z1_dual_arm_stackbox_v2-vid15-fs4,unitree_z1_dual_arm_stackbox_v2,1,34.368257999420166,34.4118230342865,143.43725562095642,237.16329956054688,0.43214982748031616,
+unitree_z1_dual_arm_stackbox_v2-vid15-fs4,unitree_z1_dual_arm_stackbox_v2,2,34.422353744506836,34.37915921211243,142.58857226371765,261.67279052734375,1.1038902625441551,
+unitree_z1_dual_arm_stackbox_v2-vid15-fs4,unitree_z1_dual_arm_stackbox_v2,3,34.07564926147461,34.27230763435364,140.5536015033722,238.81314086914062,0.3953864574432373,
+unitree_z1_dual_arm_stackbox_v2-vid15-fs4,unitree_z1_dual_arm_stackbox_v2,4,34.21823239326477,34.477816343307495,141.6662359237671,246.5524444580078,0.6671305894851685,
+unitree_z1_dual_arm_stackbox_v2-vid15-fs4,unitree_z1_dual_arm_stackbox_v2,5,34.06297254562378,34.3808650970459,140.73531913757324,216.4503936767578,0.4289618134498596,
+unitree_z1_dual_arm_stackbox_v2-vid15-fs4,unitree_z1_dual_arm_stackbox_v2,6,34.32488775253296,34.4990758895874,141.98702192306519,235.13278198242188,0.9781416803598404,
+unitree_z1_dual_arm_stackbox_v2-vid15-fs4,unitree_z1_dual_arm_stackbox_v2,7,33.835288286209106,34.33413743972778,140.40734887123108,197.0900421142578,0.21572017669677734,
+unitree_z1_dual_arm_stackbox_v2-vid15-fs4,unitree_z1_dual_arm_stackbox_v2,8,34.21294689178467,34.610371828079224,141.6176176071167,198.307373046875,0.4027923345565796,
+unitree_z1_dual_arm_stackbox_v2-vid15-fs4,unitree_z1_dual_arm_stackbox_v2,9,34.39130401611328,34.23867988586426,140.93059992790222,218.06784057617188,0.0562475323677063,
+unitree_z1_dual_arm_stackbox_v2-vid15-fs4,unitree_z1_dual_arm_stackbox_v2,10,34.440922260284424,34.388654470443726,141.8792908191681,229.65902709960938,0.4281993508338928,

unitree_z1_dual_arm_stackbox_v2/case2/output/sparse_8/inference/sample_summary.csv

@@ -0,0 +1,3 @@
+sample_id,scene,pass_type,pass_total_time_s,action_first_stable_step,state_first_stable_step,latent_first_stable_step,action_vs_full50_90pct_step,action_vs_full50_95pct_step,oracle_budget_action,oracle_budget_state,oracle_budget_latent,latent_init_dist_to_prev_round,action_drift_vs_prev_round,round_total_time_s,policy_pass_total_time_s,world_model_pass_total_time_s,psnr_full50
+unitree_z1_dual_arm_stackbox_v2-vid15-fs4,unitree_z1_dual_arm_stackbox_v2,policy,34.25820365819064,,,,47.22222222222222,49.166666666666664,46.81818181818182,47.54545454545455,,227.8909133911133,0.5108620025217533,141.70141618902034,34.25820365819064,34.35889120535417,
+unitree_z1_dual_arm_stackbox_v2-vid15-fs4,unitree_z1_dual_arm_stackbox_v2,world_model,34.35889120535417,,,,47.2,49.25,46.90909090909091,47.54545454545455,,227.8909133911133,0.5108620025217533,141.70141618902034,34.25820365819064,34.35889120535417,

unitree_z1_dual_arm_stackbox_v2/case2/run_world_model_interaction.sh

@@ -2,11 +2,11 @@ res_dir="unitree_z1_dual_arm_stackbox_v2/case2"
 dataset="unitree_z1_dual_arm_stackbox_v2"
 
 {
-    time CUDA_VISIBLE_DEVICES=0 python3 scripts/evaluation/world_model_interaction.py \
+    time CUDA_VISIBLE_DEVICES=0 "${PYTHON_BIN:-python}" scripts/evaluation/world_model_interaction.py \
         --seed 123 \
         --ckpt_path ckpts/unifolm_wma_dual.ckpt \
         --config configs/inference/world_model_interaction.yaml \
-        --savedir "${res_dir}/output" \
+        --savedir "${res_dir}/output/sparse_8" \
         --bs 1 --height 320 --width 512 \
         --unconditional_guidance_scale 1.0 \
         --ddim_steps 50 \
@@ -22,5 +22,8 @@ dataset="unitree_z1_dual_arm_stackbox_v2"
         --guidance_rescale 0.7 \
         --perframe_ae \
         --analysis_log_metrics \
-        --analysis_reference_steps 50
+        --analysis_reference_steps 50 \
+        --head_schedule_steps 0 7 14 21 28 35 42 49 \
+        --head_skip_mode reuse_prediction \
+        --head_log_steps 40 43 46 47 48 49
 } 2>&1 | tee "${res_dir}/output.log"