added paper configs (#28)

.gitignore

@@ -170,8 +170,8 @@ scripts/*.pdf
 scripts/*.jpg
 scripts/*.png
 src/xminigrid/envs/xland_tmp.py
-
+g
 # will remove later
 scripts/*testing*
-configs
+configs/wip
 scripts/eval_example.py

configs/debug/meta.yaml

@@ -0,0 +1,11 @@
+# params for quick debug on the CPU, should be 0.8+ score
+env_id: "XLand-MiniGrid-R1-9x9"
+benchmark_id: "trivial-1m"
+rnn_hidden_dim: 32
+head_hidden_dim: 16
+num_envs: 256
+num_steps_per_env: 736
+num_steps_per_update: 16
+total_timesteps: 2_000_000
+eval_num_envs: 64
+eval_num_episodes: 3

configs/debug/single.yaml

@@ -0,0 +1,6 @@
+# params for quick debug on the CPU, should be 0.8+ score
+env_id: "MiniGrid-Empty-6x6"
+rnn_hidden_dim: 16
+head_hidden_dim: 16
+num_envs: 64
+total_timesteps: 100000

configs/meta.yaml

@@ -0,0 +1,12 @@
+# params used in the paper for meta-RL experiments. Overwrite default benchmark_id if needed.
+env_id: "XLand-MiniGrid-R4-13x13"
+benchmark_id: "high-1m"
+total_timesteps: 10_000_000_000
+num_envs: 16384
+num_steps_per_env: 12800
+num_steps_per_update: 256
+update_epochs: 1
+num_minibatches: 32
+eval_num_envs: 16384
+eval_num_episodes: 25
+train_seed: 5

configs/meta_imgs.yaml

@@ -0,0 +1,14 @@
+# params used in the paper for meta-RL with RGB observations experiments. Overwrite default benchmark_id if needed.
+env_id: "XLand-MiniGrid-R1-13x13"
+benchmark_id: "high-1m"
+total_timesteps: 500_000_000
+num_envs: 1024
+num_steps_per_env: 12800
+num_steps_per_update: 128
+update_epochs: 1
+num_minibatches: 32
+eval_num_envs: 8192
+eval_num_episodes: 25
+img_obs: True
+lr: 0.0001
+train_seed: 5

configs/meta_trillion.yaml

@@ -0,0 +1,13 @@
+# params used in the paper for trillion transitions experiment.
+# params optimized for 8xA100 GPU node
+env_id: "XLand-MiniGrid-R4-13x13"
+benchmark_id: "high-1m"
+total_timesteps: 1_000_000_000_000
+num_envs: 131072
+num_steps_per_env: 12800
+num_steps_per_update: 256
+update_epochs: 1
+num_minibatches: 32
+eval_num_envs: 16384
+eval_num_episodes: 25
+train_seed: 5

configs/single.yaml

@@ -0,0 +1,11 @@
+# params used in the paper for single-RL experiments.
+env_id: "XLand-MiniGrid-R4-13x13"
+benchmark_id: "trivial-1m"
+ruleset_id: 0
+total_timesteps: 1_000_000_000
+num_envs: 16384
+num_steps: 256
+update_epochs: 1
+num_minibatches: 8
+eval_episodes: 512
+train_seed: 5