video_backbone剖析

unitree_z1_dual_arm_stackbox_v2/case4/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-vid35-fs4,unitree_z1_dual_arm_stackbox_v2,0,34.58336591720581,34.35174894332886,142.05735111236572,,,
+unitree_z1_dual_arm_stackbox_v2-vid35-fs4,unitree_z1_dual_arm_stackbox_v2,1,34.807106018066406,34.86327862739563,144.16768527030945,235.5131378173828,0.3439255952835083,
+unitree_z1_dual_arm_stackbox_v2-vid35-fs4,unitree_z1_dual_arm_stackbox_v2,2,34.52380657196045,34.07038402557373,141.07973194122314,259.55450439453125,1.1979307979345322,
+unitree_z1_dual_arm_stackbox_v2-vid35-fs4,unitree_z1_dual_arm_stackbox_v2,3,34.73806142807007,34.51879525184631,140.4986548423767,228.2439727783203,0.7015565633773804,
+unitree_z1_dual_arm_stackbox_v2-vid35-fs4,unitree_z1_dual_arm_stackbox_v2,4,34.56364035606384,34.5863573551178,141.2532799243927,233.71279907226562,0.5349503457546234,
+unitree_z1_dual_arm_stackbox_v2-vid35-fs4,unitree_z1_dual_arm_stackbox_v2,5,34.095723152160645,34.4039843082428,140.3526315689087,204.47723388671875,0.2700654864311218,
+unitree_z1_dual_arm_stackbox_v2-vid35-fs4,unitree_z1_dual_arm_stackbox_v2,6,34.46692442893982,34.41077995300293,141.16259741783142,253.46627807617188,0.8147276341915131,
+unitree_z1_dual_arm_stackbox_v2-vid35-fs4,unitree_z1_dual_arm_stackbox_v2,7,34.02199149131775,34.60529637336731,140.86116909980774,219.72276306152344,0.4421466588973999,
+unitree_z1_dual_arm_stackbox_v2-vid35-fs4,unitree_z1_dual_arm_stackbox_v2,8,34.57047963142395,34.605647802352905,144.16168022155762,214.1248016357422,0.05718517303466797,
+unitree_z1_dual_arm_stackbox_v2-vid35-fs4,unitree_z1_dual_arm_stackbox_v2,9,34.293241024017334,34.06853461265564,142.45936822891235,192.68353271484375,0.04669904708862305,
+unitree_z1_dual_arm_stackbox_v2-vid35-fs4,unitree_z1_dual_arm_stackbox_v2,10,34.26138091087341,34.6850950717926,142.63205909729004,197.61029052734375,0.18949097394943237,

unitree_z1_dual_arm_stackbox_v2/case4/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-vid35-fs4,unitree_z1_dual_arm_stackbox_v2,policy,34.44779281182723,,,,46.9,48.42857142857143,46.72727272727273,47.27272727272727,,223.9109313964844,0.45986782759428024,141.88056442954323,34.44779281182723,34.469991120425135,
+unitree_z1_dual_arm_stackbox_v2-vid35-fs4,unitree_z1_dual_arm_stackbox_v2,world_model,34.469991120425135,,,,47.0,48.875,46.63636363636363,47.09090909090909,,223.9109313964844,0.45986782759428024,141.88056442954323,34.44779281182723,34.469991120425135,

unitree_z1_dual_arm_stackbox_v2/case4/run_world_model_interaction.sh

@@ -2,11 +2,11 @@ res_dir="unitree_z1_dual_arm_stackbox_v2/case4"
 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"