video_backbone剖析

unitree_g1_pack_camera/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_g1_pack_camera-vid200-fs6,unitree_g1_pack_camera,0,34.357739210128784,34.586522579193115,140.96898198127747,,,
+unitree_g1_pack_camera-vid200-fs6,unitree_g1_pack_camera,1,34.29071259498596,35.00397062301636,141.79955863952637,229.6180419921875,0.32235634326934814,
+unitree_g1_pack_camera-vid200-fs6,unitree_g1_pack_camera,2,34.918580293655396,34.66921854019165,145.18225049972534,231.34951782226562,0.08397442102432251,
+unitree_g1_pack_camera-vid200-fs6,unitree_g1_pack_camera,3,33.87128400802612,34.726802349090576,141.85736179351807,220.95947265625,0.16776376962661743,
+unitree_g1_pack_camera-vid200-fs6,unitree_g1_pack_camera,4,34.74736452102661,34.61718559265137,141.8124771118164,257.54010009765625,0.5236250162124634,
+unitree_g1_pack_camera-vid200-fs6,unitree_g1_pack_camera,5,33.95931363105774,34.89797496795654,140.93837070465088,257.5311279296875,0.28147268295288086,
+unitree_g1_pack_camera-vid200-fs6,unitree_g1_pack_camera,6,34.651822328567505,34.53969979286194,141.137797832489,236.1876983642578,0.1898796558380127,
+unitree_g1_pack_camera-vid200-fs6,unitree_g1_pack_camera,7,34.88510847091675,34.01325011253357,139.92974042892456,259.3033752441406,0.1549433469772339,
+unitree_g1_pack_camera-vid200-fs6,unitree_g1_pack_camera,8,34.31177830696106,34.80933094024658,142.9210181236267,282.3804016113281,0.5841841697692871,
+unitree_g1_pack_camera-vid200-fs6,unitree_g1_pack_camera,9,34.6734459400177,33.96956968307495,141.88233709335327,262.050048828125,0.22107374668121338,
+unitree_g1_pack_camera-vid200-fs6,unitree_g1_pack_camera,10,34.9462194442749,34.84579944610596,142.91395735740662,199.34121704101562,0.014238595962524414,

unitree_g1_pack_camera/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_g1_pack_camera-vid200-fs6,unitree_g1_pack_camera,policy,34.51030624996532,,,,45.81818181818182,47.6,46.63636363636363,46.72727272727273,,243.6261001586914,0.25435117483139036,141.94035014239225,34.51030624996532,34.607211329720236,
+unitree_g1_pack_camera-vid200-fs6,unitree_g1_pack_camera,world_model,34.607211329720236,,,,46.36363636363637,47.77777777777778,46.81818181818182,46.81818181818182,,243.6261001586914,0.25435117483139036,141.94035014239225,34.51030624996532,34.607211329720236,

unitree_g1_pack_camera/case4/run_world_model_interaction.sh

@@ -2,11 +2,11 @@ res_dir="unitree_g1_pack_camera/case4"
 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"