SAFE/scripts/batch_training/submit_pi0diff_libero.bash at main · vla-safe/SAFE · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
#!/bin/bash

# Run all experiments for Pi0 model on the LIBERO rollouts

GROUP_NAME=pi0diff_libero_v1

# LSTM and MLP
for REG in 1e-3 1e-2 1e-1; do
    python -m failure_prob.train \
        --multirun \
        train.wandb_group_name=${GROUP_NAME} \
        dataset=pizero \
        dataset.data_path_prefix=${SAFE_OPENPI_ROLLOUT_ROOT} \
        dataset.horizon_idx_rel=0.0,1.0,concat-2 \
        dataset.diff_idx_rel=0.0,1.0,concat-2 \
        model=lstm \
        model.lr=1e-5,3e-5,1e-4,3e-4,1e-3 \
        model.lambda_reg=${REG} \
        train.seed=0-1-2 \
        train.exp_suffix=lstm
    python -m failure_prob.train \
        --multirun \
        train.wandb_group_name=${GROUP_NAME} \
        dataset=pizero \
        dataset.data_path_prefix=${SAFE_OPENPI_ROLLOUT_ROOT} \
        dataset.horizon_idx_rel=0.0,1.0,concat-2 \
        dataset.diff_idx_rel=0.0,1.0,concat-2 \
        model=indep \
        model.lr=1e-5,3e-5,1e-4,3e-4,1e-3 \
        model.lambda_reg=${REG} \
        train.seed=0-1-2 \
        train.exp_suffix=mlp
done

# Embed model
for DIST in cosine euclid; do
    python -m failure_prob.train \
        --multirun \
        train.wandb_group_name=${GROUP_NAME} \
        dataset=pizero \
        dataset.data_path_prefix=${SAFE_OPENPI_ROLLOUT_ROOT} \
        dataset.horizon_idx_rel=0.0,1.0,concat-2 \
        dataset.diff_idx_rel=0.0,1.0,concat-2 \
        dataset.load_to_cuda=False \
        model=embed \
        model.n_epochs=1 \
        model.distance=${DIST} \
        model.use_success_only=False \
        model.topk=1,5,10 \
        model.cumsum=False,True \
        train.seed=0-1-2 \
        train.exp_suffix=embed
done

python -m failure_prob.train \
    --multirun \
    train.wandb_group_name=${GROUP_NAME} \
    dataset=pizero \
    dataset.data_path_prefix=${SAFE_OPENPI_ROLLOUT_ROOT} \
    dataset.horizon_idx_rel=0.0,1.0,concat-2 \
    dataset.diff_idx_rel=0.0,1.0,concat-2 \
    dataset.load_to_cuda=False \
    model=embed \
    model.n_epochs=1 \
    model.distance=mahala \
    model.use_success_only=False \
    model.cumsum=False,True \
    train.seed=0-1-2 \
    train.exp_suffix=embed

for PCA_DIM in 32 64 128; do
    python -m failure_prob.train \
        --multirun \
        train.wandb_group_name=${GROUP_NAME} \
        dataset=pizero \
        dataset.data_path_prefix=${SAFE_OPENPI_ROLLOUT_ROOT} \
        dataset.horizon_idx_rel=0.0,1.0,concat-2 \
        dataset.diff_idx_rel=0.0,1.0,concat-2 \
        dataset.load_to_cuda=False \
        model=embed \
        model.distance=pca_kmeans \
        model.pca_dim=${PCA_DIM} \
        model.n_clusters=16,32,64 \
        model.use_success_only=False \
        model.cumsum=False,True \
        train.seed=0-1-2 \
        train.exp_suffix=embed
done

# Chen's baselines
for HORIZON_IDX in 0.0 1.0 concat-2; do
    python -m failure_prob.train \
        --multirun \
        train.wandb_group_name=${GROUP_NAME} \
        dataset=pizero \
        dataset.data_path_prefix=${SAFE_OPENPI_ROLLOUT_ROOT} \
        dataset.horizon_idx_rel=${HORIZON_IDX} \
        dataset.diff_idx_rel=0.0,1.0,concat-2 \
        dataset.load_to_cuda=False \
        model=rnd \
        train.roc_every=50 \
        model.batch_size=32 \
        model.use_success_only=False \
        train.seed=0-1-2 \
        train.exp_suffix=chen
    python -m failure_prob.train \
        --multirun \
        train.wandb_group_name=${GROUP_NAME} \
        dataset=pizero \
        dataset.data_path_prefix=${SAFE_OPENPI_ROLLOUT_ROOT} \
        dataset.horizon_idx_rel=${HORIZON_IDX} \
        dataset.diff_idx_rel=0.0,1.0,concat-2 \
        dataset.load_to_cuda=False \
        model=logpzo \
        train.roc_every=50 \
        model.batch_size=32 \
        model.forward_chunk_size=512 \
        model.use_success_only=False \
        train.seed=0-1-2 \
        train.exp_suffix=chen
done

# Handcrafted metrics
python -m failure_prob.train \
    --multirun \
    train.wandb_group_name=${GROUP_NAME} \
    dataset=pizero \
    dataset.data_path_prefix=${SAFE_OPENPI_ROLLOUT_ROOT} \
    train.log_precomputed_only=True \
    train.seed=0-1-2 \
    train.exp_suffix=handcrafted


# Handcreafted baselines with multiple action samples
python -m failure_prob.train \
    --multirun \
    train.wandb_group_name=${GROUP_NAME} \
    dataset=pizero_libero_sample10 \
    dataset.data_path_prefix=${SAFE_OPENPI_ROLLOUT_ROOT} \
    train.log_precomputed_only=True \
    train.seed=0-1-2 \
    train.exp_suffix=handcrafted_multi