jakegrigsby commited on
Commit
4c3e56c
·
1 Parent(s): 3abed59

add small-rl-gen9beta

Browse files
small-rl-gen9beta/ckpts/policy_weights/policy_epoch_0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a4c0e2750fb1d118c645a49f1c3dd3b6266f1d6e04e22fcbd56895a45d96b17e
3
+ size 56576177
small-rl-gen9beta/ckpts/policy_weights/policy_epoch_10.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cd9c1296b431039a8ccb5ad3ddd13bde8e55cccd3c84a93623f6920de53989fa
3
+ size 56576351
small-rl-gen9beta/ckpts/policy_weights/policy_epoch_12.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c466284dd15acc688efae73e7e2f66ca9fc77070daaa28b58786091519ce82f7
3
+ size 56576351
small-rl-gen9beta/ckpts/policy_weights/policy_epoch_14.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bb33f5610cf213451343f04b73527591c7e2774e85402c88069168958e0859af
3
+ size 56576351
small-rl-gen9beta/ckpts/policy_weights/policy_epoch_16.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a9d97e58a357d2c3401f13ec5a05a22eb803ed26c716813cc866c7928bb1f204
3
+ size 56576351
small-rl-gen9beta/ckpts/policy_weights/policy_epoch_18.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f4e9b3a42f208c3e4538db8871b1a0d2c05f05266e2aaceda919f33181d4926a
3
+ size 56576351
small-rl-gen9beta/ckpts/policy_weights/policy_epoch_2.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3e8891b6c6a134a1bf6ea7d64c5e31a0c6ccc4e6fc34c1f6ecc378df833ddb6c
3
+ size 56576177
small-rl-gen9beta/ckpts/policy_weights/policy_epoch_20.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:beaa0fa4a4c400defc0801d3b31e383e045a497d792f4f6a9f6b18650385f8e6
3
+ size 56576351
small-rl-gen9beta/ckpts/policy_weights/policy_epoch_22.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d93a4598039280403913ec35c41dd87a489f9f381c26854481a7a3afc973f991
3
+ size 56576351
small-rl-gen9beta/ckpts/policy_weights/policy_epoch_24.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:921d0c1bfcbc9a4319bccad8003ee4fbb22bee93c9bc588e0118e6deb3c6b144
3
+ size 56576351
small-rl-gen9beta/ckpts/policy_weights/policy_epoch_4.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:02904f0048c05b7008460aac2bfa07ed43e078669be6b1f3641aa4108eb13dec
3
+ size 56576177
small-rl-gen9beta/ckpts/policy_weights/policy_epoch_6.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d4fad78a293aff52a9ec310e24a464c37b0969c783efa464c60e59df40a152f3
3
+ size 56576177
small-rl-gen9beta/ckpts/policy_weights/policy_epoch_8.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9124f620013bdaa82afe502454c0e89503507f31c8720b77cab2bbfe20e6d1f5
3
+ size 56576177
small-rl-gen9beta/config.txt ADDED
@@ -0,0 +1,130 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import amago.agent
2
+ import amago.experiment as amago2
3
+ import amago.nets.actor_critic as amago3
4
+ import amago.nets.traj_encoders as amago4
5
+
6
+ # Parameters for Actor:
7
+ # ==============================================================================
8
+ # None.
9
+
10
+ # Parameters for Agent:
11
+ # ==============================================================================
12
+ # None.
13
+
14
+ # Parameters for Discrete:
15
+ # ==============================================================================
16
+ Discrete.clip_prob_high = 0.99
17
+ Discrete.clip_prob_low = 0.001
18
+
19
+ # Parameters for Experiment:
20
+ # ==============================================================================
21
+ Experiment.always_load_latest = False
22
+ Experiment.always_save_latest = True
23
+ Experiment.has_dset_edit_rights = True
24
+ Experiment.save_trajs_as = 'npz'
25
+ Experiment.stagger_traj_file_lengths = True
26
+ Experiment.traj_save_len = 10000000000.0
27
+ Experiment.wandb_group_name = None
28
+
29
+ # Parameters for FlashAttention:
30
+ # ==============================================================================
31
+ FlashAttention.window_size = (32, 0)
32
+
33
+ # Parameters for MetamonAMAGOExperiment:
34
+ # ==============================================================================
35
+ MetamonAMAGOExperiment.agent_type = @agent.MultiTaskAgent
36
+ MetamonAMAGOExperiment.critic_loss_weight = 10.0
37
+ MetamonAMAGOExperiment.grad_clip = 1.5
38
+ MetamonAMAGOExperiment.l2_coeff = 0.0001
39
+ MetamonAMAGOExperiment.learning_rate = 0.00015
40
+ MetamonAMAGOExperiment.lr_warmup_steps = 1000
41
+ MetamonAMAGOExperiment.max_seq_len = 200
42
+ MetamonAMAGOExperiment.traj_encoder_type = @traj_encoders.TformerTrajEncoder
43
+ MetamonAMAGOExperiment.tstep_encoder_type = @MetamonTstepEncoder
44
+
45
+ # Parameters for MetamonMaskedActor:
46
+ # ==============================================================================
47
+ MetamonMaskedActor.activation = 'leaky_relu'
48
+ MetamonMaskedActor.continuous_dist_type = None
49
+ MetamonMaskedActor.d_hidden = 300
50
+ MetamonMaskedActor.dropout_p = 0.0
51
+ MetamonMaskedActor.mask_illegal_actions = True
52
+ MetamonMaskedActor.n_layers = 2
53
+
54
+ # Parameters for MetamonTstepEncoder:
55
+ # ==============================================================================
56
+ MetamonTstepEncoder.d_model = 100
57
+ MetamonTstepEncoder.dropout = 0.05
58
+ MetamonTstepEncoder.extra_emb_dim = 18
59
+ MetamonTstepEncoder.n_heads = 5
60
+ MetamonTstepEncoder.n_layers = 3
61
+ MetamonTstepEncoder.numerical_tokens = 6
62
+ MetamonTstepEncoder.scratch_tokens = 4
63
+ MetamonTstepEncoder.token_mask_aug = False
64
+
65
+ # Parameters for Multigammas:
66
+ # ==============================================================================
67
+ Multigammas.continuous = [0.1, 0.9, 0.95, 0.97, 0.99, 0.995]
68
+ Multigammas.discrete = [0.1, 0.9, 0.95, 0.97, 0.99, 0.995]
69
+
70
+ # Parameters for MultiTaskAgent:
71
+ # ==============================================================================
72
+ MultiTaskAgent.actor_type = @MetamonMaskedActor
73
+ MultiTaskAgent.critic_type = @actor_critic.NCriticsTwoHot
74
+ MultiTaskAgent.fake_filter = False
75
+ MultiTaskAgent.fbc_filter_func = @agent.exp_filter
76
+ MultiTaskAgent.gamma = 0.999
77
+ MultiTaskAgent.num_actions_for_value_in_actor_loss = 3
78
+ MultiTaskAgent.num_actions_for_value_in_critic_loss = 5
79
+ MultiTaskAgent.num_critics = 4
80
+ MultiTaskAgent.num_critics_td = 2
81
+ MultiTaskAgent.offline_coeff = 1.0
82
+ MultiTaskAgent.online_coeff = 0.0
83
+ MultiTaskAgent.pass_obs_keys_to_actor = ['illegal_actions']
84
+ MultiTaskAgent.popart = True
85
+ MultiTaskAgent.reward_multiplier = 10.0
86
+ MultiTaskAgent.tau = 0.004
87
+ MultiTaskAgent.use_multigamma = True
88
+ MultiTaskAgent.use_target_actor = True
89
+
90
+ # Parameters for NCriticsTwoHot:
91
+ # ==============================================================================
92
+ NCriticsTwoHot.activation = 'leaky_relu'
93
+ NCriticsTwoHot.d_hidden = 300
94
+ NCriticsTwoHot.dropout_p = 0.0
95
+ NCriticsTwoHot.max_return = 1100
96
+ NCriticsTwoHot.min_return = -1100
97
+ NCriticsTwoHot.n_layers = 2
98
+ NCriticsTwoHot.output_bins = 64
99
+ NCriticsTwoHot.use_symlog = False
100
+
101
+ # Parameters for PopArtLayer:
102
+ # ==============================================================================
103
+ PopArtLayer.beta = 0.0005
104
+ PopArtLayer.init_nu = 100.0
105
+
106
+ # Parameters for SigmaReparam:
107
+ # ==============================================================================
108
+ SigmaReparam.bias = True
109
+ SigmaReparam.fast_init = False
110
+
111
+ # Parameters for TformerTrajEncoder:
112
+ # ==============================================================================
113
+ TformerTrajEncoder.activation = 'leaky_relu'
114
+ TformerTrajEncoder.d_ff = 2048
115
+ TformerTrajEncoder.d_model = 512
116
+ TformerTrajEncoder.dropout_attn = 0.0
117
+ TformerTrajEncoder.dropout_emb = 0.05
118
+ TformerTrajEncoder.dropout_ff = 0.05
119
+ TformerTrajEncoder.dropout_qkv = 0.0
120
+ TformerTrajEncoder.head_scaling = True
121
+ TformerTrajEncoder.n_heads = 8
122
+ TformerTrajEncoder.n_layers = 3
123
+ TformerTrajEncoder.norm = 'layer'
124
+ TformerTrajEncoder.normformer_norms = True
125
+ TformerTrajEncoder.pos_emb = 'fixed'
126
+ TformerTrajEncoder.sigma_reparam = True
127
+
128
+ # Parameters for TransformerTurnEmbedding:
129
+ # ==============================================================================
130
+ # None.