video_backbone剖析

unitree_g1_pack_camera/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_g1_pack_camera-vid50-fs6,unitree_g1_pack_camera,0,34.508949756622314,34.254820823669434,140.46258544921875,,,
+unitree_g1_pack_camera-vid50-fs6,unitree_g1_pack_camera,1,34.90244817733765,34.26566195487976,141.24445700645447,216.52996826171875,0.24411267042160034,
+unitree_g1_pack_camera-vid50-fs6,unitree_g1_pack_camera,2,34.450573444366455,33.74520945549011,140.44232392311096,210.6576385498047,0.211828351020813,
+unitree_g1_pack_camera-vid50-fs6,unitree_g1_pack_camera,3,34.227173805236816,34.19744801521301,139.57855439186096,231.77476501464844,0.10076236724853516,
+unitree_g1_pack_camera-vid50-fs6,unitree_g1_pack_camera,4,34.42516374588013,33.859012603759766,140.9917664527893,196.30520629882812,0.24305355548858643,
+unitree_g1_pack_camera-vid50-fs6,unitree_g1_pack_camera,5,34.275354862213135,34.270713329315186,143.11559915542603,270.99072265625,0.9365180283784866,
+unitree_g1_pack_camera-vid50-fs6,unitree_g1_pack_camera,6,34.44929575920105,33.572381019592285,141.9271743297577,227.52503967285156,0.14944738149642944,
+unitree_g1_pack_camera-vid50-fs6,unitree_g1_pack_camera,7,34.52991771697998,34.2509183883667,141.21305179595947,256.9947509765625,0.1445719599723816,
+unitree_g1_pack_camera-vid50-fs6,unitree_g1_pack_camera,8,34.30908131599426,33.91361689567566,140.56905364990234,263.05572509765625,0.0568811297416687,
+unitree_g1_pack_camera-vid50-fs6,unitree_g1_pack_camera,9,34.12396168708801,34.13224148750305,139.65425729751587,249.03759765625,0.11548089981079102,
+unitree_g1_pack_camera-vid50-fs6,unitree_g1_pack_camera,10,34.32340431213379,34.3162624835968,142.47088241577148,246.99588012695312,0.5883344411849976,

unitree_g1_pack_camera/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_g1_pack_camera-vid50-fs6,unitree_g1_pack_camera,policy,34.41139314391396,,,,45.90909090909091,47.81818181818182,47.0,47.09090909090909,,236.98672943115236,0.279099078476429,141.06088235161522,34.41139314391396,34.07075331427834,
+unitree_g1_pack_camera-vid50-fs6,unitree_g1_pack_camera,world_model,34.07075331427834,,,,46.27272727272727,47.77777777777778,47.0,47.09090909090909,,236.98672943115236,0.279099078476429,141.06088235161522,34.41139314391396,34.07075331427834,

unitree_g1_pack_camera/case2/run_world_model_interaction.sh

@@ -2,11 +2,11 @@ res_dir="unitree_g1_pack_camera/case2"
 dataset="unitree_g1_pack_camera"
 
 {
-    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 "${res_dir}/output/sparse_8" \
         --bs 1 --height 320 --width 512 \
         --unconditional_guidance_scale 1.0 \
         --ddim_steps 50 \
@@ -22,5 +22,8 @@ dataset="unitree_g1_pack_camera"
         --guidance_rescale 0.7 \
         --perframe_ae \
         --analysis_log_metrics \
-        --analysis_reference_steps 50
+        --analysis_reference_steps 50 \
+        --head_schedule_steps 0 7 14 21 28 35 42 49 \
+        --head_skip_mode reuse_prediction \
+        --head_log_steps 40 43 46 47 48 49
 } 2>&1 | tee "${res_dir}/output.log"