Hoar012 commited on
Commit
d0528cd
·
verified ·
1 Parent(s): e0e5acc

Upload folder using huggingface_hub

Browse files
robomimic__can/IQL-critic-0331-2QHead_demos10/seed0/config.yaml ADDED
@@ -0,0 +1,132 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ act: SiLU
2
+ action_dim: -1
3
+ action_horizon: 1
4
+ action_repeat: 1
5
+ batch_length: 32
6
+ batch_size: 16
7
+ compile: true
8
+ cont_head: {layers: 2, loss_scale: 1.0, outscale: 1.0}
9
+ critic: {dist: normal_std_fixed, eps: 1e-05, grad_clip: 100.0, layers: 2, lr: 3e-05,
10
+ num_models: 5, num_subsample: 2, outscale: 0.0, slow_target: true, slow_target_fraction: 0.02,
11
+ slow_target_update: 1}
12
+ datadir: datasets/robomimic_datasets
13
+ debug: false
14
+ decoder: {act: SiLU, cnn_depth: 32, cnn_keys: .*image.*, cnn_sigmoid: false, image_dist: mse,
15
+ kernel_size: 4, minres: 4, mlp_keys: .*state.*, mlp_layers: 5, mlp_units: 1024,
16
+ norm: true, outscale: 1.0, vector_dist: symlog_mse}
17
+ deterministic_run: false
18
+ device: cuda:0
19
+ discount: 0.99
20
+ discount_lambda: 0.95
21
+ done_mode: 1
22
+ dp:
23
+ ac_chunk: 8
24
+ action_space: deltas
25
+ batch_size: 64
26
+ debug: false
27
+ device: cuda:0
28
+ devices: 1
29
+ eval_freq: 150000
30
+ img_chunk: 2
31
+ log_freq: 100
32
+ lr: 1e-05
33
+ max_iterations: 500000
34
+ num_cams: 2
35
+ pref_ckpt: ''
36
+ pretrained_ckpt:
37
+ /zfsauton2/home/haoranha/code/SAILOR/scratch_dir/logs/robomimic__can/train_dp-0111-p64_demos10/seed0/DP_Pretrain_base_policy_latest.pt
38
+ print_config: false
39
+ schedule_freq: 1
40
+ seed: 292285
41
+ shared_mlp: []
42
+ train_steps: 1000
43
+ train_transform: gpu_medium
44
+ use_if_ckpt_present: false
45
+ dyn_deter: 512
46
+ dyn_discrete: 32
47
+ dyn_hidden: 512
48
+ dyn_mean_act: none
49
+ dyn_min_std: 0.1
50
+ dyn_rec_depth: 1
51
+ dyn_scale: 0.5
52
+ dyn_std_act: sigmoid2
53
+ dyn_stoch: 32
54
+ encoder: {act: SiLU, cnn_depth: 32, cnn_keys: .*image.*, kernel_size: 4, minres: 4,
55
+ mlp_keys: .*state.*, mlp_layers: 5, mlp_units: 1024, norm: true, symlog_inputs: true}
56
+ env_max_steps: {can: 500000.0, lift: 100000.0, square: 500000.0}
57
+ env_time_limits: {can: 200, lift: 100, square: 400}
58
+ eval_epochs: 5
59
+ eval_num_runs: 100
60
+ generate_highres_eval: false
61
+ grad_clip: 1000
62
+ grad_heads: [decoder, cont]
63
+ high_res_render: false
64
+ highres_img_size: 1024
65
+ imag_horizon: 7
66
+ image_size: 64
67
+ initial: learned
68
+ iql_critic: {batch_size: 64, ckpt_path: null, epochs: 1, eps: 1e-05, expectile: 0.7,
69
+ gamma: 0.99, grad_clip: 100.0, lr: 3e-05, num_models: 5, num_subsample: 2, outscale: 0.0,
70
+ ranking_weight: 0.1, tau: 0.005, train_steps: 1000, warmup_steps: 15000}
71
+ kl_free: 1.0
72
+ log_every: 50
73
+ logdir: scratch_dir//logs/robomimic__can/IQL-critic-0331-2QHead_demos10/seed0
74
+ max_failure_buffer_size: 500
75
+ max_recovery_buffer_size: 500
76
+ max_retries: 5
77
+ max_success_buffer_size: 500
78
+ model_lr: 1e-05
79
+ mppi: {abs_residual: 0.2, action_l2_cost: 0, discount: 0.99, horizon: 8, init_std: 0.05,
80
+ iterations: 6, max_std: 0.1, min_std: 0.01, num_elites: 32, num_samples: 256, pretrained_ckpt: '',
81
+ temperature: 0.5, uncertainty_cost: 1.0}
82
+ norm: True'
83
+ normalize_state_actions: false
84
+ num_buffer_transitions: 100000
85
+ num_envs: 1
86
+ num_exp_trajs: 10
87
+ num_exp_val_trajs: 1
88
+ obs_horizon: 2
89
+ opt: adam
90
+ opt_eps: 1e-08
91
+ precision: 32
92
+ pred_horizon: 8
93
+ pref: {batch_size: 128, dpo_loss_scale: 1.0, exp_loss_scale: 1.0, lr: 0.0001, rec_loss_scale: 1.0,
94
+ train_steps: 200}
95
+ rep_scale: 0.1
96
+ reset_step: {can: 25, lift: 25, square: 30}
97
+ residual_actor: {absmax: null, dist: normal, entropy: 0.0003, eps: 1e-05, grad_clip: 100.0,
98
+ layers: 2, lr: 3e-05, max_std: 0.1, min_std: 0.01, outscale: 1.0, std: learned,
99
+ temp: 0.1, unimix_ratio: 0.01}
100
+ reward_EMA: true
101
+ reward_head: {dist: normal_std_fixed, layers: 2, outscale: 0.0}
102
+ scratch_dir: scratch_dir/
103
+ seed: 0
104
+ set:
105
+ - [image_size, '64']
106
+ - [dp.pretrained_ckpt,
107
+ /zfsauton2/home/haoranha/code/SAILOR/scratch_dir/logs/robomimic__can/train_dp-0111-p64_demos10/seed0/DP_Pretrain_base_policy_latest.pt]
108
+ shape_rewards: true
109
+ state_dim: -1
110
+ state_only: false
111
+ task: robomimic__can
112
+ time_limit: 1000
113
+ train_dp_contrastive: true
114
+ train_dp_mppi: false
115
+ train_dp_mppi_params: {data_collect_noise_std: 0.1, discrim_state_only: true, dp_expert_buffer_ratio: 0.5,
116
+ eval_every_round: 10, force_warmup: true, min_env_steps_per_round: 3500, n_dp_train_itrs: 1000,
117
+ n_dp_traj_buffer_size: 64, n_env_steps: 500000.0, n_traj_to_relabel_per_round: 64,
118
+ rounds_train_ratio: 1.5, upate_discrim_every: 100, update_dp_every: 10, use_discrim: true,
119
+ warmstart_percentage_env_steps: 0.2, warmstart_train_ratio: 1.5}
120
+ train_estimator: false
121
+ training_epochs: 50
122
+ unimix_ratio: 0.01
123
+ units: 512
124
+ update_dp: false
125
+ use_wandb: true
126
+ visualize_eval: true
127
+ viz_expert_buffer: false
128
+ wandb_entity: njuai
129
+ wandb_exp_name: robomimic__can/IQL-critic-0331-2QHead_demos10
130
+ wandb_project: Recovery
131
+ warmup_epochs: 5
132
+ weight_decay: 0.0
robomimic__can/IQL-critic-0331-2QHead_demos10/seed0/preference_ckpt/Critic-epoch-0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:63d8f3fe3f790d1558892d9c221b28d4afa48f70c116b9191755f8e39d8517eb
3
+ size 189652670
robomimic__can/IQL-critic-0331-2QHead_demos10/seed36/config.yaml ADDED
@@ -0,0 +1,132 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ act: SiLU
2
+ action_dim: -1
3
+ action_horizon: 1
4
+ action_repeat: 1
5
+ batch_length: 32
6
+ batch_size: 16
7
+ compile: true
8
+ cont_head: {layers: 2, loss_scale: 1.0, outscale: 1.0}
9
+ critic: {dist: normal_std_fixed, eps: 1e-05, grad_clip: 100.0, layers: 2, lr: 3e-05,
10
+ num_models: 5, num_subsample: 2, outscale: 0.0, slow_target: true, slow_target_fraction: 0.02,
11
+ slow_target_update: 1}
12
+ datadir: datasets/robomimic_datasets
13
+ debug: false
14
+ decoder: {act: SiLU, cnn_depth: 32, cnn_keys: .*image.*, cnn_sigmoid: false, image_dist: mse,
15
+ kernel_size: 4, minres: 4, mlp_keys: .*state.*, mlp_layers: 5, mlp_units: 1024,
16
+ norm: true, outscale: 1.0, vector_dist: symlog_mse}
17
+ deterministic_run: false
18
+ device: cuda:0
19
+ discount: 0.99
20
+ discount_lambda: 0.95
21
+ done_mode: 1
22
+ dp:
23
+ ac_chunk: 8
24
+ action_space: deltas
25
+ batch_size: 64
26
+ debug: false
27
+ device: cuda:0
28
+ devices: 1
29
+ eval_freq: 150000
30
+ img_chunk: 2
31
+ log_freq: 100
32
+ lr: 1e-05
33
+ max_iterations: 500000
34
+ num_cams: 2
35
+ pref_ckpt: ''
36
+ pretrained_ckpt:
37
+ /zfsauton2/home/haoranha/code/SAILOR/scratch_dir/logs/robomimic__can/train_dp-0111-p64_demos10/seed36/DP_Pretrain_base_policy_latest.pt
38
+ print_config: false
39
+ schedule_freq: 1
40
+ seed: 292285
41
+ shared_mlp: []
42
+ train_steps: 1000
43
+ train_transform: gpu_medium
44
+ use_if_ckpt_present: false
45
+ dyn_deter: 512
46
+ dyn_discrete: 32
47
+ dyn_hidden: 512
48
+ dyn_mean_act: none
49
+ dyn_min_std: 0.1
50
+ dyn_rec_depth: 1
51
+ dyn_scale: 0.5
52
+ dyn_std_act: sigmoid2
53
+ dyn_stoch: 32
54
+ encoder: {act: SiLU, cnn_depth: 32, cnn_keys: .*image.*, kernel_size: 4, minres: 4,
55
+ mlp_keys: .*state.*, mlp_layers: 5, mlp_units: 1024, norm: true, symlog_inputs: true}
56
+ env_max_steps: {can: 500000.0, lift: 100000.0, square: 500000.0}
57
+ env_time_limits: {can: 200, lift: 100, square: 400}
58
+ eval_epochs: 5
59
+ eval_num_runs: 100
60
+ generate_highres_eval: false
61
+ grad_clip: 1000
62
+ grad_heads: [decoder, cont]
63
+ high_res_render: false
64
+ highres_img_size: 1024
65
+ imag_horizon: 7
66
+ image_size: 64
67
+ initial: learned
68
+ iql_critic: {batch_size: 64, ckpt_path: null, epochs: 1, eps: 1e-05, expectile: 0.7,
69
+ gamma: 0.99, grad_clip: 100.0, lr: 3e-05, num_models: 5, num_subsample: 2, outscale: 0.0,
70
+ ranking_weight: 0.1, tau: 0.005, train_steps: 1000, warmup_steps: 15000}
71
+ kl_free: 1.0
72
+ log_every: 50
73
+ logdir: scratch_dir//logs/robomimic__can/IQL-critic-0331-2QHead_demos10/seed36
74
+ max_failure_buffer_size: 500
75
+ max_recovery_buffer_size: 500
76
+ max_retries: 5
77
+ max_success_buffer_size: 500
78
+ model_lr: 1e-05
79
+ mppi: {abs_residual: 0.2, action_l2_cost: 0, discount: 0.99, horizon: 8, init_std: 0.05,
80
+ iterations: 6, max_std: 0.1, min_std: 0.01, num_elites: 32, num_samples: 256, pretrained_ckpt: '',
81
+ temperature: 0.5, uncertainty_cost: 1.0}
82
+ norm: True'
83
+ normalize_state_actions: false
84
+ num_buffer_transitions: 100000
85
+ num_envs: 1
86
+ num_exp_trajs: 10
87
+ num_exp_val_trajs: 1
88
+ obs_horizon: 2
89
+ opt: adam
90
+ opt_eps: 1e-08
91
+ precision: 32
92
+ pred_horizon: 8
93
+ pref: {batch_size: 128, dpo_loss_scale: 1.0, exp_loss_scale: 1.0, lr: 0.0001, rec_loss_scale: 1.0,
94
+ train_steps: 200}
95
+ rep_scale: 0.1
96
+ reset_step: {can: 25, lift: 25, square: 30}
97
+ residual_actor: {absmax: null, dist: normal, entropy: 0.0003, eps: 1e-05, grad_clip: 100.0,
98
+ layers: 2, lr: 3e-05, max_std: 0.1, min_std: 0.01, outscale: 1.0, std: learned,
99
+ temp: 0.1, unimix_ratio: 0.01}
100
+ reward_EMA: true
101
+ reward_head: {dist: normal_std_fixed, layers: 2, outscale: 0.0}
102
+ scratch_dir: scratch_dir/
103
+ seed: 36
104
+ set:
105
+ - [image_size, '64']
106
+ - [dp.pretrained_ckpt,
107
+ /zfsauton2/home/haoranha/code/SAILOR/scratch_dir/logs/robomimic__can/train_dp-0111-p64_demos10/seed36/DP_Pretrain_base_policy_latest.pt]
108
+ shape_rewards: true
109
+ state_dim: -1
110
+ state_only: false
111
+ task: robomimic__can
112
+ time_limit: 1000
113
+ train_dp_contrastive: true
114
+ train_dp_mppi: false
115
+ train_dp_mppi_params: {data_collect_noise_std: 0.1, discrim_state_only: true, dp_expert_buffer_ratio: 0.5,
116
+ eval_every_round: 10, force_warmup: true, min_env_steps_per_round: 3500, n_dp_train_itrs: 1000,
117
+ n_dp_traj_buffer_size: 64, n_env_steps: 500000.0, n_traj_to_relabel_per_round: 64,
118
+ rounds_train_ratio: 1.5, upate_discrim_every: 100, update_dp_every: 10, use_discrim: true,
119
+ warmstart_percentage_env_steps: 0.2, warmstart_train_ratio: 1.5}
120
+ train_estimator: false
121
+ training_epochs: 50
122
+ unimix_ratio: 0.01
123
+ units: 512
124
+ update_dp: false
125
+ use_wandb: true
126
+ visualize_eval: true
127
+ viz_expert_buffer: false
128
+ wandb_entity: njuai
129
+ wandb_exp_name: robomimic__can/IQL-critic-0331-2QHead_demos10
130
+ wandb_project: Recovery
131
+ warmup_epochs: 5
132
+ weight_decay: 0.0
robomimic__can/IQL-critic-0331-2QHead_demos10/seed36/preference_ckpt/Critic-epoch-0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8475ec893cb32d9e3a9697203467f508a51cf49cd0a96423d5c7a66b40e35a04
3
+ size 189652670
robomimic__can/IQL-critic-0331-2QHead_demos10/seed72/config.yaml ADDED
@@ -0,0 +1,132 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ act: SiLU
2
+ action_dim: -1
3
+ action_horizon: 1
4
+ action_repeat: 1
5
+ batch_length: 32
6
+ batch_size: 16
7
+ compile: true
8
+ cont_head: {layers: 2, loss_scale: 1.0, outscale: 1.0}
9
+ critic: {dist: normal_std_fixed, eps: 1e-05, grad_clip: 100.0, layers: 2, lr: 3e-05,
10
+ num_models: 5, num_subsample: 2, outscale: 0.0, slow_target: true, slow_target_fraction: 0.02,
11
+ slow_target_update: 1}
12
+ datadir: datasets/robomimic_datasets
13
+ debug: false
14
+ decoder: {act: SiLU, cnn_depth: 32, cnn_keys: .*image.*, cnn_sigmoid: false, image_dist: mse,
15
+ kernel_size: 4, minres: 4, mlp_keys: .*state.*, mlp_layers: 5, mlp_units: 1024,
16
+ norm: true, outscale: 1.0, vector_dist: symlog_mse}
17
+ deterministic_run: false
18
+ device: cuda:0
19
+ discount: 0.99
20
+ discount_lambda: 0.95
21
+ done_mode: 1
22
+ dp:
23
+ ac_chunk: 8
24
+ action_space: deltas
25
+ batch_size: 64
26
+ debug: false
27
+ device: cuda:0
28
+ devices: 1
29
+ eval_freq: 150000
30
+ img_chunk: 2
31
+ log_freq: 100
32
+ lr: 1e-05
33
+ max_iterations: 500000
34
+ num_cams: 2
35
+ pref_ckpt: ''
36
+ pretrained_ckpt:
37
+ /zfsauton2/home/haoranha/code/SAILOR/scratch_dir/logs/robomimic__can/train_dp-0111-p64_demos10/seed72/DP_Pretrain_base_policy_latest.pt
38
+ print_config: false
39
+ schedule_freq: 1
40
+ seed: 292285
41
+ shared_mlp: []
42
+ train_steps: 1000
43
+ train_transform: gpu_medium
44
+ use_if_ckpt_present: false
45
+ dyn_deter: 512
46
+ dyn_discrete: 32
47
+ dyn_hidden: 512
48
+ dyn_mean_act: none
49
+ dyn_min_std: 0.1
50
+ dyn_rec_depth: 1
51
+ dyn_scale: 0.5
52
+ dyn_std_act: sigmoid2
53
+ dyn_stoch: 32
54
+ encoder: {act: SiLU, cnn_depth: 32, cnn_keys: .*image.*, kernel_size: 4, minres: 4,
55
+ mlp_keys: .*state.*, mlp_layers: 5, mlp_units: 1024, norm: true, symlog_inputs: true}
56
+ env_max_steps: {can: 500000.0, lift: 100000.0, square: 500000.0}
57
+ env_time_limits: {can: 200, lift: 100, square: 400}
58
+ eval_epochs: 5
59
+ eval_num_runs: 100
60
+ generate_highres_eval: false
61
+ grad_clip: 1000
62
+ grad_heads: [decoder, cont]
63
+ high_res_render: false
64
+ highres_img_size: 1024
65
+ imag_horizon: 7
66
+ image_size: 64
67
+ initial: learned
68
+ iql_critic: {batch_size: 64, ckpt_path: null, epochs: 1, eps: 1e-05, expectile: 0.7,
69
+ gamma: 0.99, grad_clip: 100.0, lr: 3e-05, num_models: 5, num_subsample: 2, outscale: 0.0,
70
+ ranking_weight: 0.1, tau: 0.005, train_steps: 1000, warmup_steps: 15000}
71
+ kl_free: 1.0
72
+ log_every: 50
73
+ logdir: scratch_dir//logs/robomimic__can/IQL-critic-0331-2QHead_demos10/seed72
74
+ max_failure_buffer_size: 500
75
+ max_recovery_buffer_size: 500
76
+ max_retries: 5
77
+ max_success_buffer_size: 500
78
+ model_lr: 1e-05
79
+ mppi: {abs_residual: 0.2, action_l2_cost: 0, discount: 0.99, horizon: 8, init_std: 0.05,
80
+ iterations: 6, max_std: 0.1, min_std: 0.01, num_elites: 32, num_samples: 256, pretrained_ckpt: '',
81
+ temperature: 0.5, uncertainty_cost: 1.0}
82
+ norm: True'
83
+ normalize_state_actions: false
84
+ num_buffer_transitions: 100000
85
+ num_envs: 1
86
+ num_exp_trajs: 10
87
+ num_exp_val_trajs: 1
88
+ obs_horizon: 2
89
+ opt: adam
90
+ opt_eps: 1e-08
91
+ precision: 32
92
+ pred_horizon: 8
93
+ pref: {batch_size: 128, dpo_loss_scale: 1.0, exp_loss_scale: 1.0, lr: 0.0001, rec_loss_scale: 1.0,
94
+ train_steps: 200}
95
+ rep_scale: 0.1
96
+ reset_step: {can: 25, lift: 25, square: 30}
97
+ residual_actor: {absmax: null, dist: normal, entropy: 0.0003, eps: 1e-05, grad_clip: 100.0,
98
+ layers: 2, lr: 3e-05, max_std: 0.1, min_std: 0.01, outscale: 1.0, std: learned,
99
+ temp: 0.1, unimix_ratio: 0.01}
100
+ reward_EMA: true
101
+ reward_head: {dist: normal_std_fixed, layers: 2, outscale: 0.0}
102
+ scratch_dir: scratch_dir/
103
+ seed: 72
104
+ set:
105
+ - [image_size, '64']
106
+ - [dp.pretrained_ckpt,
107
+ /zfsauton2/home/haoranha/code/SAILOR/scratch_dir/logs/robomimic__can/train_dp-0111-p64_demos10/seed72/DP_Pretrain_base_policy_latest.pt]
108
+ shape_rewards: true
109
+ state_dim: -1
110
+ state_only: false
111
+ task: robomimic__can
112
+ time_limit: 1000
113
+ train_dp_contrastive: true
114
+ train_dp_mppi: false
115
+ train_dp_mppi_params: {data_collect_noise_std: 0.1, discrim_state_only: true, dp_expert_buffer_ratio: 0.5,
116
+ eval_every_round: 10, force_warmup: true, min_env_steps_per_round: 3500, n_dp_train_itrs: 1000,
117
+ n_dp_traj_buffer_size: 64, n_env_steps: 500000.0, n_traj_to_relabel_per_round: 64,
118
+ rounds_train_ratio: 1.5, upate_discrim_every: 100, update_dp_every: 10, use_discrim: true,
119
+ warmstart_percentage_env_steps: 0.2, warmstart_train_ratio: 1.5}
120
+ train_estimator: false
121
+ training_epochs: 50
122
+ unimix_ratio: 0.01
123
+ units: 512
124
+ update_dp: false
125
+ use_wandb: true
126
+ visualize_eval: true
127
+ viz_expert_buffer: false
128
+ wandb_entity: njuai
129
+ wandb_exp_name: robomimic__can/IQL-critic-0331-2QHead_demos10
130
+ wandb_project: Recovery
131
+ warmup_epochs: 5
132
+ weight_decay: 0.0
robomimic__can/IQL-critic-0331-2QHead_demos10/seed72/preference_ckpt/Critic-epoch-0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:496ce440012eafe6a0c9e0e06bfc98ba0103742f3fadd708be243e9de817b7ac
3
+ size 189652670
robomimic__lift/IQL-critic-0331-2QHead_demos10/seed0/config.yaml ADDED
@@ -0,0 +1,143 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ act: SiLU
2
+ act_policy: {backbone: resnet18, batch_size: 128, dec_layers: 7, dim_feedforward: 3200,
3
+ enc_layers: 4, eval_freq: 10000, hidden_dim: 512, kl_weight: 10, log_freq: 100,
4
+ lr: 0.0001, lr_backbone: 1e-05, nheads: 8, pretrained_ckpt: '', schedule_freq: 1,
5
+ train_steps: 30000}
6
+ action_dim: -1
7
+ action_horizon: 1
8
+ action_repeat: 1
9
+ batch_length: 32
10
+ batch_size: 16
11
+ compile: true
12
+ cont_head: {layers: 2, loss_scale: 1.0, outscale: 1.0}
13
+ critic: {dist: normal_std_fixed, eps: 1e-05, grad_clip: 100.0, layers: 2, lr: 3e-05,
14
+ num_models: 5, num_subsample: 2, outscale: 0.0, slow_target: true, slow_target_fraction: 0.02,
15
+ slow_target_update: 1}
16
+ datadir: datasets/robomimic_datasets
17
+ debug: false
18
+ decoder: {act: SiLU, cnn_depth: 32, cnn_keys: .*image.*, cnn_sigmoid: false, image_dist: mse,
19
+ kernel_size: 4, minres: 4, mlp_keys: .*state.*, mlp_layers: 5, mlp_units: 1024,
20
+ norm: true, outscale: 1.0, vector_dist: symlog_mse}
21
+ deterministic_run: false
22
+ device: cuda:0
23
+ discount: 0.99
24
+ discount_lambda: 0.95
25
+ done_mode: 1
26
+ dp:
27
+ ac_chunk: 8
28
+ action_space: deltas
29
+ batch_size: 128
30
+ debug: false
31
+ device: cuda:0
32
+ devices: 1
33
+ eval_freq: 150000
34
+ img_chunk: 2
35
+ log_freq: 100
36
+ lr: 1e-05
37
+ max_iterations: 500000
38
+ num_cams: 2
39
+ pref_ckpt: ''
40
+ pretrained_ckpt:
41
+ /zfsauton2/home/haoranha/code/SAILOR/scratch_dir/logs/robomimic__lift/train_dp-0111-p64_demos10/seed0/DP_Pretrain_base_policy_latest.pt
42
+ print_config: false
43
+ schedule_freq: 1
44
+ seed: 292285
45
+ shared_mlp: []
46
+ train_steps: 1000
47
+ train_transform: gpu_medium
48
+ use_if_ckpt_present: false
49
+ dyn_deter: 512
50
+ dyn_discrete: 32
51
+ dyn_hidden: 512
52
+ dyn_mean_act: none
53
+ dyn_min_std: 0.1
54
+ dyn_rec_depth: 1
55
+ dyn_scale: 0.5
56
+ dyn_std_act: sigmoid2
57
+ dyn_stoch: 32
58
+ encoder: {act: SiLU, cnn_depth: 32, cnn_keys: .*image.*, kernel_size: 4, minres: 4,
59
+ mlp_keys: .*state.*, mlp_layers: 5, mlp_units: 1024, norm: true, symlog_inputs: true}
60
+ env_max_steps: {can: 500000.0, lift: 100000.0, square: 500000.0}
61
+ env_time_limits: {can: 200, lift: 100, square: 200}
62
+ eval_epochs: 10
63
+ eval_num_runs: 100
64
+ generate_highres_eval: false
65
+ grad_clip: 1000
66
+ grad_heads: [decoder, cont]
67
+ high_res_render: false
68
+ highres_img_size: 1024
69
+ imag_horizon: 7
70
+ image_size: 64
71
+ initial: learned
72
+ iql_critic: {batch_size: 128, ckpt_path: null, epochs: 1, eps: 1e-05, expectile: 0.7,
73
+ gamma: 0.99, grad_clip: 100.0, lr: 0.0001, ranking_weight: 0.1, tau: 0.005, train_steps: 2000,
74
+ warmup_steps: 15000}
75
+ kl_free: 1.0
76
+ log_every: 50
77
+ logdir: scratch_dir//logs/robomimic__lift/debug-critic-0331-2QHead_demos10/seed0
78
+ max_failure_buffer_size: 500
79
+ max_recovery_buffer_size: 500
80
+ max_retries: 5
81
+ max_retries_rollout: 5
82
+ max_success_buffer_size: 500
83
+ model_lr: 1e-05
84
+ mppi: {abs_residual: 0.2, action_l2_cost: 0, discount: 0.99, horizon: 8, init_std: 0.05,
85
+ iterations: 6, max_std: 0.1, min_std: 0.01, num_elites: 32, num_samples: 256, pretrained_ckpt: '',
86
+ temperature: 0.5, uncertainty_cost: 1.0}
87
+ noise_epsilon: 1.0
88
+ noise_std: 0.1
89
+ noise_std_decay: 1.0
90
+ norm: True'
91
+ normalize_state_actions: false
92
+ num_buffer_transitions: 100000
93
+ num_envs: 1
94
+ num_exp_trajs: 10
95
+ num_exp_val_trajs: 1
96
+ obs_horizon: 2
97
+ opt: adam
98
+ opt_eps: 1e-08
99
+ policy: dp
100
+ precision: 32
101
+ pred_horizon: 8
102
+ pref: {batch_size: 128, dpo_loss_scale: 1.0, exp_loss_scale: 1.0, lr: 0.0001, rec_loss_scale: 1.0,
103
+ train_steps: 200}
104
+ rep_scale: 0.1
105
+ reset_step: {can: 1, lift: 1, square: 1}
106
+ residual_actor: {absmax: null, dist: normal, entropy: 0.0003, eps: 1e-05, grad_clip: 100.0,
107
+ layers: 2, lr: 3e-05, max_std: 0.1, min_std: 0.01, outscale: 1.0, std: learned,
108
+ temp: 0.1, unimix_ratio: 0.01}
109
+ reward_EMA: true
110
+ reward_head: {dist: normal_std_fixed, layers: 2, outscale: 0.0}
111
+ scratch_dir: scratch_dir/
112
+ seed: 0
113
+ set:
114
+ - [image_size, '64']
115
+ - [dp.pretrained_ckpt,
116
+ /zfsauton2/home/haoranha/code/SAILOR/scratch_dir/logs/robomimic__lift/train_dp-0111-p64_demos10/seed0/DP_Pretrain_base_policy_latest.pt]
117
+ shape_rewards: true
118
+ state_dim: -1
119
+ state_only: false
120
+ task: robomimic__lift
121
+ time_limit: 500
122
+ train_dp_contrastive: true
123
+ train_dp_mppi: false
124
+ train_dp_mppi_params: {data_collect_noise_std: 0.1, discrim_state_only: true, dp_expert_buffer_ratio: 0.5,
125
+ eval_every_round: 10, force_warmup: true, min_env_steps_per_round: 3500, n_dp_train_itrs: 1000,
126
+ n_dp_traj_buffer_size: 64, n_env_steps: 100000.0, n_traj_to_relabel_per_round: 64,
127
+ rounds_train_ratio: 1.5, upate_discrim_every: 100, update_dp_every: 10, use_discrim: true,
128
+ warmstart_percentage_env_steps: 0.2, warmstart_train_ratio: 1.5}
129
+ train_estimator: false
130
+ training_epochs: 100
131
+ tta: {epsilon: 0.5, lr: 1e-05, num_pos: 8, num_samples: 64, update_steps: 10}
132
+ tta_cl: true
133
+ unimix_ratio: 0.01
134
+ units: 512
135
+ update_dp: false
136
+ use_wandb: true
137
+ visualize_eval: true
138
+ viz_expert_buffer: false
139
+ wandb_entity: njuai
140
+ wandb_exp_name: robomimic__lift/debug-critic-0331-2QHead_demos10
141
+ wandb_project: Recovery
142
+ warmup_epochs: 5
143
+ weight_decay: 0.0
robomimic__lift/IQL-critic-0331-2QHead_demos10/seed0/preference_ckpt/Critic-epoch-0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ec416ac6614a4d0d42c8bf886491f198794b4ca26c9dbba1eb883cb425a4753a
3
+ size 189652670
robomimic__lift/IQL-critic-0331-2QHead_demos10/seed36/config.yaml ADDED
@@ -0,0 +1,143 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ act: SiLU
2
+ act_policy: {backbone: resnet18, batch_size: 128, dec_layers: 7, dim_feedforward: 3200,
3
+ enc_layers: 4, eval_freq: 10000, hidden_dim: 512, kl_weight: 10, log_freq: 100,
4
+ lr: 0.0001, lr_backbone: 1e-05, nheads: 8, pretrained_ckpt: '', schedule_freq: 1,
5
+ train_steps: 30000}
6
+ action_dim: -1
7
+ action_horizon: 1
8
+ action_repeat: 1
9
+ batch_length: 32
10
+ batch_size: 16
11
+ compile: true
12
+ cont_head: {layers: 2, loss_scale: 1.0, outscale: 1.0}
13
+ critic: {dist: normal_std_fixed, eps: 1e-05, grad_clip: 100.0, layers: 2, lr: 3e-05,
14
+ num_models: 5, num_subsample: 2, outscale: 0.0, slow_target: true, slow_target_fraction: 0.02,
15
+ slow_target_update: 1}
16
+ datadir: datasets/robomimic_datasets
17
+ debug: false
18
+ decoder: {act: SiLU, cnn_depth: 32, cnn_keys: .*image.*, cnn_sigmoid: false, image_dist: mse,
19
+ kernel_size: 4, minres: 4, mlp_keys: .*state.*, mlp_layers: 5, mlp_units: 1024,
20
+ norm: true, outscale: 1.0, vector_dist: symlog_mse}
21
+ deterministic_run: false
22
+ device: cuda:0
23
+ discount: 0.99
24
+ discount_lambda: 0.95
25
+ done_mode: 1
26
+ dp:
27
+ ac_chunk: 8
28
+ action_space: deltas
29
+ batch_size: 128
30
+ debug: false
31
+ device: cuda:0
32
+ devices: 1
33
+ eval_freq: 150000
34
+ img_chunk: 2
35
+ log_freq: 100
36
+ lr: 1e-05
37
+ max_iterations: 500000
38
+ num_cams: 2
39
+ pref_ckpt: ''
40
+ pretrained_ckpt:
41
+ /zfsauton2/home/haoranha/code/SAILOR/scratch_dir/logs/robomimic__lift/train_dp-0111-p64_demos10/seed36/DP_Pretrain_base_policy_latest.pt
42
+ print_config: false
43
+ schedule_freq: 1
44
+ seed: 292285
45
+ shared_mlp: []
46
+ train_steps: 1000
47
+ train_transform: gpu_medium
48
+ use_if_ckpt_present: false
49
+ dyn_deter: 512
50
+ dyn_discrete: 32
51
+ dyn_hidden: 512
52
+ dyn_mean_act: none
53
+ dyn_min_std: 0.1
54
+ dyn_rec_depth: 1
55
+ dyn_scale: 0.5
56
+ dyn_std_act: sigmoid2
57
+ dyn_stoch: 32
58
+ encoder: {act: SiLU, cnn_depth: 32, cnn_keys: .*image.*, kernel_size: 4, minres: 4,
59
+ mlp_keys: .*state.*, mlp_layers: 5, mlp_units: 1024, norm: true, symlog_inputs: true}
60
+ env_max_steps: {can: 500000.0, lift: 100000.0, square: 500000.0}
61
+ env_time_limits: {can: 200, lift: 100, square: 200}
62
+ eval_epochs: 10
63
+ eval_num_runs: 100
64
+ generate_highres_eval: false
65
+ grad_clip: 1000
66
+ grad_heads: [decoder, cont]
67
+ high_res_render: false
68
+ highres_img_size: 1024
69
+ imag_horizon: 7
70
+ image_size: 64
71
+ initial: learned
72
+ iql_critic: {batch_size: 128, ckpt_path: null, epochs: 1, eps: 1e-05, expectile: 0.7,
73
+ gamma: 0.99, grad_clip: 100.0, lr: 0.0001, ranking_weight: 0.1, tau: 0.005, train_steps: 2000,
74
+ warmup_steps: 15000}
75
+ kl_free: 1.0
76
+ log_every: 50
77
+ logdir: scratch_dir//logs/robomimic__lift/debug-critic-0331-2QHead_demos10/seed36
78
+ max_failure_buffer_size: 500
79
+ max_recovery_buffer_size: 500
80
+ max_retries: 5
81
+ max_retries_rollout: 5
82
+ max_success_buffer_size: 500
83
+ model_lr: 1e-05
84
+ mppi: {abs_residual: 0.2, action_l2_cost: 0, discount: 0.99, horizon: 8, init_std: 0.05,
85
+ iterations: 6, max_std: 0.1, min_std: 0.01, num_elites: 32, num_samples: 256, pretrained_ckpt: '',
86
+ temperature: 0.5, uncertainty_cost: 1.0}
87
+ noise_epsilon: 1.0
88
+ noise_std: 0.1
89
+ noise_std_decay: 1.0
90
+ norm: True'
91
+ normalize_state_actions: false
92
+ num_buffer_transitions: 100000
93
+ num_envs: 1
94
+ num_exp_trajs: 10
95
+ num_exp_val_trajs: 1
96
+ obs_horizon: 2
97
+ opt: adam
98
+ opt_eps: 1e-08
99
+ policy: dp
100
+ precision: 32
101
+ pred_horizon: 8
102
+ pref: {batch_size: 128, dpo_loss_scale: 1.0, exp_loss_scale: 1.0, lr: 0.0001, rec_loss_scale: 1.0,
103
+ train_steps: 200}
104
+ rep_scale: 0.1
105
+ reset_step: {can: 1, lift: 1, square: 1}
106
+ residual_actor: {absmax: null, dist: normal, entropy: 0.0003, eps: 1e-05, grad_clip: 100.0,
107
+ layers: 2, lr: 3e-05, max_std: 0.1, min_std: 0.01, outscale: 1.0, std: learned,
108
+ temp: 0.1, unimix_ratio: 0.01}
109
+ reward_EMA: true
110
+ reward_head: {dist: normal_std_fixed, layers: 2, outscale: 0.0}
111
+ scratch_dir: scratch_dir/
112
+ seed: 36
113
+ set:
114
+ - [image_size, '64']
115
+ - [dp.pretrained_ckpt,
116
+ /zfsauton2/home/haoranha/code/SAILOR/scratch_dir/logs/robomimic__lift/train_dp-0111-p64_demos10/seed36/DP_Pretrain_base_policy_latest.pt]
117
+ shape_rewards: true
118
+ state_dim: -1
119
+ state_only: false
120
+ task: robomimic__lift
121
+ time_limit: 500
122
+ train_dp_contrastive: true
123
+ train_dp_mppi: false
124
+ train_dp_mppi_params: {data_collect_noise_std: 0.1, discrim_state_only: true, dp_expert_buffer_ratio: 0.5,
125
+ eval_every_round: 10, force_warmup: true, min_env_steps_per_round: 3500, n_dp_train_itrs: 1000,
126
+ n_dp_traj_buffer_size: 64, n_env_steps: 100000.0, n_traj_to_relabel_per_round: 64,
127
+ rounds_train_ratio: 1.5, upate_discrim_every: 100, update_dp_every: 10, use_discrim: true,
128
+ warmstart_percentage_env_steps: 0.2, warmstart_train_ratio: 1.5}
129
+ train_estimator: false
130
+ training_epochs: 100
131
+ tta: {epsilon: 0.5, lr: 1e-05, num_pos: 8, num_samples: 64, update_steps: 10}
132
+ tta_cl: true
133
+ unimix_ratio: 0.01
134
+ units: 512
135
+ update_dp: false
136
+ use_wandb: true
137
+ visualize_eval: true
138
+ viz_expert_buffer: false
139
+ wandb_entity: njuai
140
+ wandb_exp_name: robomimic__lift/debug-critic-0331-2QHead_demos10
141
+ wandb_project: Recovery
142
+ warmup_epochs: 5
143
+ weight_decay: 0.0
robomimic__lift/IQL-critic-0331-2QHead_demos10/seed36/preference_ckpt/Critic-epoch-0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a62ff1161c7bcf08445a660a77d7d8fc8b37ea8b35ed0e77863a4f8d813bfa03
3
+ size 189652670
robomimic__lift/IQL-critic-0331-2QHead_demos10/seed72/config.yaml ADDED
@@ -0,0 +1,143 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ act: SiLU
2
+ act_policy: {backbone: resnet18, batch_size: 128, dec_layers: 7, dim_feedforward: 3200,
3
+ enc_layers: 4, eval_freq: 10000, hidden_dim: 512, kl_weight: 10, log_freq: 100,
4
+ lr: 0.0001, lr_backbone: 1e-05, nheads: 8, pretrained_ckpt: '', schedule_freq: 1,
5
+ train_steps: 30000}
6
+ action_dim: -1
7
+ action_horizon: 1
8
+ action_repeat: 1
9
+ batch_length: 32
10
+ batch_size: 16
11
+ compile: true
12
+ cont_head: {layers: 2, loss_scale: 1.0, outscale: 1.0}
13
+ critic: {dist: normal_std_fixed, eps: 1e-05, grad_clip: 100.0, layers: 2, lr: 3e-05,
14
+ num_models: 5, num_subsample: 2, outscale: 0.0, slow_target: true, slow_target_fraction: 0.02,
15
+ slow_target_update: 1}
16
+ datadir: datasets/robomimic_datasets
17
+ debug: false
18
+ decoder: {act: SiLU, cnn_depth: 32, cnn_keys: .*image.*, cnn_sigmoid: false, image_dist: mse,
19
+ kernel_size: 4, minres: 4, mlp_keys: .*state.*, mlp_layers: 5, mlp_units: 1024,
20
+ norm: true, outscale: 1.0, vector_dist: symlog_mse}
21
+ deterministic_run: false
22
+ device: cuda:0
23
+ discount: 0.99
24
+ discount_lambda: 0.95
25
+ done_mode: 1
26
+ dp:
27
+ ac_chunk: 8
28
+ action_space: deltas
29
+ batch_size: 128
30
+ debug: false
31
+ device: cuda:0
32
+ devices: 1
33
+ eval_freq: 150000
34
+ img_chunk: 2
35
+ log_freq: 100
36
+ lr: 1e-05
37
+ max_iterations: 500000
38
+ num_cams: 2
39
+ pref_ckpt: ''
40
+ pretrained_ckpt:
41
+ /zfsauton2/home/haoranha/code/SAILOR/scratch_dir/logs/robomimic__lift/train_dp-0111-p64_demos10/seed72/DP_Pretrain_base_policy_latest.pt
42
+ print_config: false
43
+ schedule_freq: 1
44
+ seed: 292285
45
+ shared_mlp: []
46
+ train_steps: 1000
47
+ train_transform: gpu_medium
48
+ use_if_ckpt_present: false
49
+ dyn_deter: 512
50
+ dyn_discrete: 32
51
+ dyn_hidden: 512
52
+ dyn_mean_act: none
53
+ dyn_min_std: 0.1
54
+ dyn_rec_depth: 1
55
+ dyn_scale: 0.5
56
+ dyn_std_act: sigmoid2
57
+ dyn_stoch: 32
58
+ encoder: {act: SiLU, cnn_depth: 32, cnn_keys: .*image.*, kernel_size: 4, minres: 4,
59
+ mlp_keys: .*state.*, mlp_layers: 5, mlp_units: 1024, norm: true, symlog_inputs: true}
60
+ env_max_steps: {can: 500000.0, lift: 100000.0, square: 500000.0}
61
+ env_time_limits: {can: 200, lift: 100, square: 200}
62
+ eval_epochs: 10
63
+ eval_num_runs: 100
64
+ generate_highres_eval: false
65
+ grad_clip: 1000
66
+ grad_heads: [decoder, cont]
67
+ high_res_render: false
68
+ highres_img_size: 1024
69
+ imag_horizon: 7
70
+ image_size: 64
71
+ initial: learned
72
+ iql_critic: {batch_size: 128, ckpt_path: null, epochs: 1, eps: 1e-05, expectile: 0.7,
73
+ gamma: 0.99, grad_clip: 100.0, lr: 0.0001, ranking_weight: 0.1, tau: 0.005, train_steps: 2000,
74
+ warmup_steps: 15000}
75
+ kl_free: 1.0
76
+ log_every: 50
77
+ logdir: scratch_dir//logs/robomimic__lift/debug-critic-0331-2QHead_demos10/seed72
78
+ max_failure_buffer_size: 500
79
+ max_recovery_buffer_size: 500
80
+ max_retries: 5
81
+ max_retries_rollout: 5
82
+ max_success_buffer_size: 500
83
+ model_lr: 1e-05
84
+ mppi: {abs_residual: 0.2, action_l2_cost: 0, discount: 0.99, horizon: 8, init_std: 0.05,
85
+ iterations: 6, max_std: 0.1, min_std: 0.01, num_elites: 32, num_samples: 256, pretrained_ckpt: '',
86
+ temperature: 0.5, uncertainty_cost: 1.0}
87
+ noise_epsilon: 1.0
88
+ noise_std: 0.1
89
+ noise_std_decay: 1.0
90
+ norm: True'
91
+ normalize_state_actions: false
92
+ num_buffer_transitions: 100000
93
+ num_envs: 1
94
+ num_exp_trajs: 10
95
+ num_exp_val_trajs: 1
96
+ obs_horizon: 2
97
+ opt: adam
98
+ opt_eps: 1e-08
99
+ policy: dp
100
+ precision: 32
101
+ pred_horizon: 8
102
+ pref: {batch_size: 128, dpo_loss_scale: 1.0, exp_loss_scale: 1.0, lr: 0.0001, rec_loss_scale: 1.0,
103
+ train_steps: 200}
104
+ rep_scale: 0.1
105
+ reset_step: {can: 1, lift: 1, square: 1}
106
+ residual_actor: {absmax: null, dist: normal, entropy: 0.0003, eps: 1e-05, grad_clip: 100.0,
107
+ layers: 2, lr: 3e-05, max_std: 0.1, min_std: 0.01, outscale: 1.0, std: learned,
108
+ temp: 0.1, unimix_ratio: 0.01}
109
+ reward_EMA: true
110
+ reward_head: {dist: normal_std_fixed, layers: 2, outscale: 0.0}
111
+ scratch_dir: scratch_dir/
112
+ seed: 72
113
+ set:
114
+ - [image_size, '64']
115
+ - [dp.pretrained_ckpt,
116
+ /zfsauton2/home/haoranha/code/SAILOR/scratch_dir/logs/robomimic__lift/train_dp-0111-p64_demos10/seed72/DP_Pretrain_base_policy_latest.pt]
117
+ shape_rewards: true
118
+ state_dim: -1
119
+ state_only: false
120
+ task: robomimic__lift
121
+ time_limit: 500
122
+ train_dp_contrastive: true
123
+ train_dp_mppi: false
124
+ train_dp_mppi_params: {data_collect_noise_std: 0.1, discrim_state_only: true, dp_expert_buffer_ratio: 0.5,
125
+ eval_every_round: 10, force_warmup: true, min_env_steps_per_round: 3500, n_dp_train_itrs: 1000,
126
+ n_dp_traj_buffer_size: 64, n_env_steps: 100000.0, n_traj_to_relabel_per_round: 64,
127
+ rounds_train_ratio: 1.5, upate_discrim_every: 100, update_dp_every: 10, use_discrim: true,
128
+ warmstart_percentage_env_steps: 0.2, warmstart_train_ratio: 1.5}
129
+ train_estimator: false
130
+ training_epochs: 100
131
+ tta: {epsilon: 0.5, lr: 1e-05, num_pos: 8, num_samples: 64, update_steps: 10}
132
+ tta_cl: true
133
+ unimix_ratio: 0.01
134
+ units: 512
135
+ update_dp: false
136
+ use_wandb: true
137
+ visualize_eval: true
138
+ viz_expert_buffer: false
139
+ wandb_entity: njuai
140
+ wandb_exp_name: robomimic__lift/debug-critic-0331-2QHead_demos10
141
+ wandb_project: Recovery
142
+ warmup_epochs: 5
143
+ weight_decay: 0.0
robomimic__lift/IQL-critic-0331-2QHead_demos10/seed72/preference_ckpt/Critic-epoch-0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a678fbc25174913bb51107db98365bd328880f061893c55b3dce060dc3a3ab19
3
+ size 189652670