new .yaml composable configs

pufferlib/config/ocean/drive/eval/behaviors.yaml

@@ -0,0 +1,104 @@
+# Driving-behaviour evaluation: nuPlan scenes labeled by scene type. Each
+# behavior is one [eval.behaviors_*] section inheriting the shared template.
+
+behaviors_defaults:
+  enabled: false
+  interval: 250
+  mode: inline
+  clean: true
+  render: true
+  render_views: ["sim_state", "bev"]
+  env:
+    simulation_mode: replay
+    control_mode: control_sdc_only
+    init_mode: create_all_valid
+    scenario_length: 201
+    obs_slots_partners_n: 32
+  eval:
+    num_scenarios: 50
+    render_num_scenarios: 2
+    render_max_steps: 200
+
+behaviors_full_dir:
+  inherits: behaviors_defaults
+  type: behavior_class
+  enabled: true
+  env:
+    map_dir: /scratch/ev2237/data/nuplan/nuplan_mini_train_bins
+
+behaviors_hard_stop:
+  inherits: behaviors_defaults
+  type: behavior_class
+  enabled: true
+  env:
+    map_dir: /scratch/ev2237/data/nuplan/categories_v021/hard_stop
+
+behaviors_highway_straight:
+  inherits: behaviors_defaults
+  type: behavior_class
+  enabled: true
+  env:
+    map_dir: /scratch/ev2237/data/nuplan/categories_v021/highway_straight
+
+behaviors_lane_change:
+  inherits: behaviors_defaults
+  type: behavior_class
+  enabled: true
+  env:
+    map_dir: /scratch/ev2237/data/nuplan/categories_v021/lane_change
+
+behaviors_merge:
+  inherits: behaviors_defaults
+  type: behavior_class
+  enabled: true
+  env:
+    map_dir: /scratch/ev2237/data/nuplan/categories_v021/merge
+
+behaviors_parked_cars:
+  inherits: behaviors_defaults
+  type: behavior_class
+  enabled: true
+  env:
+    map_dir: /scratch/ev2237/data/nuplan/categories_v021/parked_cars
+
+behaviors_roundabout:
+  inherits: behaviors_defaults
+  type: behavior_class
+  enabled: true
+  env:
+    map_dir: /scratch/ev2237/data/nuplan/categories_v021/roundabout
+
+behaviors_stopped_traffic:
+  inherits: behaviors_defaults
+  type: behavior_class
+  enabled: true
+  env:
+    map_dir: /scratch/ev2237/data/nuplan/categories_v021/stopped_traffic
+
+behaviors_traffic_light_green:
+  inherits: behaviors_defaults
+  type: behavior_class
+  enabled: true
+  env:
+    map_dir: /scratch/ev2237/data/nuplan/categories_v021/traffic_light_green
+
+behaviors_traffic_light_stop:
+  inherits: behaviors_defaults
+  type: behavior_class
+  enabled: true
+  env:
+    map_dir: /scratch/ev2237/data/nuplan/categories_v021/traffic_light_stop
+
+behaviors_unprotected_left:
+  inherits: behaviors_defaults
+  type: behavior_class
+  enabled: true
+  env:
+    map_dir: /scratch/ev2237/data/nuplan/categories_v021/unprotected_left
+
+behaviors_unprotected_right:
+  inherits: behaviors_defaults
+  type: behavior_class
+  enabled: true
+  env:
+    map_dir: /scratch/ev2237/data/nuplan/categories_v021/unprotected_right

pufferlib/config/ocean/drive/eval/validation.yaml

@@ -0,0 +1,97 @@
+# Multi-scenario validation suite. `validation_defaults` is a template (no
+# `type`); the replay/gigaflow/triage sections inherit it and add their
+# sim-specific map set + agent layout. Each lands under [eval.<name>].
+
+validation_defaults:
+  enabled: true
+  interval: 250
+  mode: inline
+  clean: true
+  env:
+    eval_mode: 1
+    collision_behavior: 1
+    offroad_behavior: 1
+    # Explicit 0 wins over the clean macro's red-light enforcement.
+    traffic_light_behavior: 0
+    reward_randomization: false
+    termination_mode: 0
+    num_agents: 1024
+    target_type: static
+    goal_speed: 3.0
+    reward_collision: 3.0
+    reward_offroad: 3.0
+    reward_stop_line: 1.0
+    reward_goal: 1.0
+    reward_overspeed: 0.05
+    reward_comfort: 0.05
+    reward_velocity: 0.0025
+    reward_lane_align: 0.025
+    reward_lane_center: 0.0038
+    reward_timestep: 0.000025
+    reward_reverse: 0.005
+    reward_ade: 0.0
+    obs_dropout_lane: 0.0
+    obs_dropout_boundary: 0.0
+    obs_slots_lane_n: 80
+    obs_slots_boundary_n: 80
+  eval:
+    num_scenarios: 250
+    export_episode_csv: true
+    verify_coverage: true
+
+validation_replay:
+  inherits: validation_defaults
+  type: multi_scenario
+  enabled: true
+  render: true
+  render_backend: triage_html
+  env:
+    simulation_mode: replay
+    control_mode: control_sdc_only
+    map_dir: /scratch/ev2237/data/nuplan/nuplan_mini_train_bins
+    num_maps: 250
+    max_agents_per_env: 64
+    scenario_length: 200
+    resample_frequency: 200
+  eval:
+    render_num_scenarios: 5
+    render_max_steps: 200
+
+validation_gigaflow:
+  inherits: validation_defaults
+  type: multi_scenario
+  enabled: true
+  render: true
+  render_backend: egl
+  render_views: ["sim_state", "bev"]
+  env:
+    simulation_mode: gigaflow
+    map_dir: pufferlib/resources/drive/binaries/carla
+    num_maps: 8
+    num_agents: 1024
+    min_agents_per_env: 40
+    max_agents_per_env: 40
+    scenario_length: 500
+    resample_frequency: 500
+  eval:
+    render_num_scenarios: 8
+    render_max_steps: 300
+
+dnf_triage:
+  inherits: validation_defaults
+  type: multi_scenario
+  enabled: false
+  render: true
+  render_backend: triage_html
+  env:
+    simulation_mode: gigaflow
+    map_dir: pufferlib/resources/drive/binaries/carla/opendrive__Town10HD.bin
+    num_maps: 1
+    min_agents_per_env: 1
+    max_agents_per_env: 1
+    scenario_length: 500
+    resample_frequency: 500
+  eval:
+    num_scenarios: 32
+    render_num_scenarios: 16
+    render_max_steps: 300

pufferlib/config/ocean/drive/eval/wosac.yaml

@@ -0,0 +1,18 @@
+# Optional WOSAC realism eval. Off by default; runs as a subprocess.
+wosac:
+  type: wosac
+  enabled: false
+  interval: 500
+  mode: subprocess
+  clean: true
+  render: false
+  env:
+    control_mode: control_wosac
+    init_mode: create_all_valid
+    init_step: 10
+    goal_radius: 2.0
+  eval:
+    wosac_num_rollouts: 32
+    wosac_num_agents: 256
+    wosac_sanity_check: false
+    wosac_aggregate_results: true

pufferlib/config/ocean/drive/model/gigaflow.yaml

@@ -0,0 +1,36 @@
+# Policy/network definition (internal sections: [base] net ids, [policy], [rnn]).
+base:
+  policy_name: Drive
+  # Options: Recurrent, None
+  rnn_name: null
+
+policy:
+  # Encoder layer (per-encoder embedding width)
+  ego_input_size: 64
+  partner_input_size: 64
+  lane_input_size: 64
+  boundary_input_size: 64
+  traffic_control_input_size: 64
+  context_input_size: 64
+  # Encoder activation - options: "relu", "tanh", "gelu"
+  encoder_activation: relu
+  encoder_layer_norm: true
+  mask_padded_features: false
+  # Shared backbone layer
+  backbone_hidden_size: 512
+  backbone_num_layers: 4
+  # Backbone activation - options: "relu", "tanh", "gelu"
+  backbone_activation: gelu
+  backbone_layer_norm: false
+  # Actor head layer
+  actor_hidden_size: 512
+  actor_num_layers: 0
+  # Critic head layer
+  critic_hidden_size: 512
+  critic_num_layers: 0
+  # Dual or shared actor-critic backbone
+  shared_network: true
+
+rnn:
+  input_size: 512
+  hidden_size: 512

pufferlib/config/ocean/drive/recipes/default.yaml

@@ -0,0 +1,35 @@
+# Default PufferDrive training recipe. Selects one component per group plus a
+# list of eval suites. Swap a group on the CLI, e.g.
+#   puffer train puffer_drive --config <this file>
+# or point --config at a sibling recipe to define a different experiment.
+
+package: ocean
+env_name: puffer_drive
+
+vecenv: default              # -> ../vecenv/default.yaml          ([vec])
+train: default               # -> ../train/default.yaml           ([train])
+simulator: gigaflow_carla    # -> ../simulator/gigaflow_carla.yaml ([env])
+model: gigaflow              # -> ../model/gigaflow.yaml          ([base]/[policy]/[rnn])
+
+eval:                        # each -> ../eval/<name>.yaml
+  - validation
+  - behaviors
+  - wosac
+
+# --- Inline sections (merged verbatim, not from component files) ---
+
+# Failure-mining workflow: `puffer mine_failures puffer_drive ...`
+mine:
+  output_dir: ""
+  num_episodes: 100
+  score_threshold: -inf
+  render: true
+
+# Controlled experiments: `puffer controlled_exp puffer_drive` runs the cross
+# product of these value lists.
+controlled_exp:
+  train:
+    learning_rate:
+      values: [0.001, 0.003, 0.01]
+    ent_coef:
+      values: [0.01, 0.005]

pufferlib/config/ocean/drive/recipes/replay_nuplan.yaml

@@ -0,0 +1,25 @@
+# nuPlan replay training: ego SDC learns against log-replayed traffic.
+# Same puffer_drive env as the default recipe, swapped to the replay_nuplan
+# simulator profile, with validation switched off (empty eval list).
+#
+#   puffer train puffer_drive --config pufferlib/config/ocean/drive/recipes/replay_nuplan.yaml
+
+package: ocean
+env_name: puffer_drive
+
+vecenv: default
+train: default
+model: gigaflow
+simulator: replay_nuplan     # replay mode, nuPlan maps, termination_mode 1, 200 steps
+
+eval: []                     # no validation suites
+
+# Patch a few values on top of the chosen components. Replay batches are
+# smaller than gigaflow, so a smaller minibatch and higher entropy bonus help,
+# as in the old replay_sdc config.
+overrides:
+  train:
+    seed: 555
+    ent_coef: 0.19
+    minibatch_size: 8192
+    checkpoint_interval: 1000