Merge remote-tracking branch 'origin/main' into polynet

configs/env/default.yaml

@@ -3,4 +3,4 @@ name: tsp
 
 generator_params:
   num_loc: 20
-  loc_distribution: uniform
+  loc_distribution: uniform

configs/env/fjsp/15j-10m.yaml

@@ -0,0 +1,13 @@
+_target_: rl4co.envs.FJSPEnv
+name: fjsp
+
+generator_params:
+  num_jobs: 15
+  num_machines: 10
+  min_ops_per_job: 8
+  max_ops_per_job: 12
+  min_processing_time: 1
+  max_processing_time: 20
+  min_eligible_ma_per_op: 1
+
+data_dir: ${paths.root_dir}/data/fjsp

configs/env/fjsp/20j-10m.yaml

@@ -0,0 +1,13 @@
+_target_: rl4co.envs.FJSPEnv
+name: fjsp
+
+generator_params:
+  num_jobs: 20
+  num_machines: 10
+  min_ops_per_job: 8
+  max_ops_per_job: 12
+  min_processing_time: 1
+  max_processing_time: 20
+  min_eligible_ma_per_op: 1
+
+data_dir: ${paths.root_dir}/data/fjsp

configs/env/fjsp/20j-5m.yaml

@@ -0,0 +1,13 @@
+_target_: rl4co.envs.FJSPEnv
+name: fjsp
+
+generator_params:
+  num_jobs: 20
+  num_machines: 5
+  min_ops_per_job: 4
+  max_ops_per_job: 6
+  min_processing_time: 1
+  max_processing_time: 20
+  min_eligible_ma_per_op: 1
+
+data_dir: ${paths.root_dir}/data/fjsp

configs/env/jssp/10j-10m.yaml

@@ -0,0 +1,11 @@
+_target_: rl4co.envs.JSSPEnv
+name: jssp
+
+generator_params:
+  num_jobs: 10
+  num_machines: 10
+  min_processing_time: 1
+  max_processing_time: 99
+
+data_dir: ${paths.root_dir}/data/jssp/taillard
+test_file: ${env.generator_params.num_jobs}j_${env.generator_params.num_machines}m

configs/env/jssp/15j-15m.yaml

@@ -0,0 +1,11 @@
+_target_: rl4co.envs.JSSPEnv
+name: jssp
+
+generator_params:
+  num_jobs: 15
+  num_machines: 15
+  min_processing_time: 1
+  max_processing_time: 99
+
+data_dir: ${paths.root_dir}/data/jssp/taillard
+test_file: ${env.generator_params.num_jobs}j_${env.generator_params.num_machines}m

configs/env/jssp/20j-20m.yaml

@@ -0,0 +1,11 @@
+_target_: rl4co.envs.JSSPEnv
+name: jssp
+
+generator_params:
+  num_jobs: 20
+  num_machines: 20
+  min_processing_time: 1
+  max_processing_time: 99
+
+data_dir: ${paths.root_dir}/data/jssp/taillard
+test_file: ${env.generator_params.num_jobs}j_${env.generator_params.num_machines}m

configs/env/jssp/6j-6m.yaml

@@ -0,0 +1,11 @@
+_target_: rl4co.envs.JSSPEnv
+name: jssp
+
+generator_params:
+  num_jobs: 6
+  num_machines: 6
+  min_processing_time: 1
+  max_processing_time: 99
+
+data_dir: ${paths.root_dir}/data/jssp/taillard
+test_file: ${env.generator_params.num_jobs}j_${env.generator_params.num_machines}m

configs/experiment/routing/tsp-stepwise-ppo.yaml

@@ -0,0 +1,57 @@
+# @package _global_
+
+defaults:
+  - override /model: l2d.yaml
+  - override /callbacks: default.yaml
+  - override /trainer: default.yaml
+  - override /logger: wandb.yaml
+
+env:
+  _target_: rl4co.envs.TSPEnv4PPO
+  generator_params:
+    num_loc: 20
+
+logger:
+  wandb:
+    project: "rl4co"
+    tags: ["am-stepwise-ppo", "${env.name}"]
+    group: ${env.name}${env.generator_params.num_loc}
+    name: ppo-${env.name}${env.generator_params.num_loc}
+
+trainer:
+  max_epochs: 10
+  precision: 32-true
+
+embed_dim: 256
+num_heads: 8
+model:
+  _target_: rl4co.models.StepwisePPO
+  policy:
+    _target_: rl4co.models.L2DPolicy4PPO
+    decoder:
+      _target_: rl4co.models.zoo.l2d.decoder.L2DDecoder
+      env_name: ${env.name}
+      embed_dim: ${embed_dim}
+      feature_extractor:
+        _target_: rl4co.models.zoo.am.encoder.AttentionModelEncoder
+        embed_dim: ${embed_dim}
+        num_heads: ${num_heads}
+        num_layers: 4
+        normalization: "batch"
+        env_name: "tsp"
+      actor:
+        _target_: rl4co.models.zoo.l2d.decoder.AttnActor
+        embed_dim: ${embed_dim}
+        num_heads: ${num_heads}
+        env_name: ${env.name}
+    embed_dim: ${embed_dim}
+    env_name: ${env.name}
+    het_emb: False
+  batch_size: 512
+  mini_batch_size: 512
+  train_data_size: 20000
+  val_data_size: 1_000
+  test_data_size: 1_000
+  reward_scale: scale
+  optimizer_kwargs:
+    lr: 1e-4

configs/experiment/scheduling/am-pomo.yaml

@@ -0,0 +1,23 @@
+# @package _global_
+
+defaults:
+  - scheduling/base
+
+logger:
+  wandb:
+    tags: ["am-pomo", "${env.name}"]
+    name: "am-pomo-${env.name}-${env.generator_params.num_jobs}j-${env.generator_params.num_machines}m"
+
+model:
+  _target_: rl4co.models.POMO
+  policy:
+    _target_: rl4co.models.L2DAttnPolicy
+    env_name: ${env.name}
+    scaling_factor: ${scaling_factor}
+  batch_size: 64
+  num_starts: 10
+  num_augment: 0
+  baseline: "shared"
+  metrics:
+    val: ["reward", "max_reward"]
+    test: ${model.metrics.val}

configs/experiment/scheduling/am-ppo.yaml

@@ -0,0 +1,56 @@
+# @package _global_
+
+defaults:
+  - scheduling/base
+
+logger:
+  wandb:
+    tags: ["am-ppo", "${env.name}"]
+    name: "am-ppo-${env.name}-${env.generator_params.num_jobs}j-${env.generator_params.num_machines}m"
+
+embed_dim: 256
+num_heads: 8
+
+model:
+  _target_: rl4co.models.StepwisePPO
+  policy:
+    _target_: rl4co.models.L2DPolicy4PPO
+    decoder:
+      _target_: rl4co.models.zoo.l2d.decoder.L2DDecoder
+      env_name: ${env.name}
+      embed_dim: ${embed_dim}
+      feature_extractor:
+        _target_: rl4co.models.zoo.matnet.matnet_w_sa.Encoder
+        embed_dim: ${embed_dim}
+        num_heads: ${num_heads}
+        num_layers: 4
+        normalization: "batch"
+        init_embedding:
+          _target_: rl4co.models.nn.env_embeddings.init.FJSPMatNetInitEmbedding
+          embed_dim: ${embed_dim}
+          scaling_factor: ${scaling_factor}
+      actor:
+        _target_: rl4co.models.zoo.l2d.decoder.L2DAttnActor
+        embed_dim: ${embed_dim}
+        num_heads: ${num_heads}
+        env_name: ${env.name}
+        scaling_factor:  ${scaling_factor}
+        stepwise: True
+    env_name: ${env.name}
+    embed_dim: ${embed_dim}
+    scaling_factor: ${scaling_factor}
+    het_emb: True
+  batch_size: 128
+  val_batch_size: 512
+  test_batch_size: 64
+  # Song et al use 1000 iterations over batches of 20 = 20_000
+  # We train 10 epochs on a set of 2000 instance = 20_000
+  train_data_size: 2000
+  mini_batch_size: 512
+  reward_scale: scale
+  optimizer_kwargs:
+    lr: 1e-4
+
+env:
+  stepwise_reward: True
+  _torchrl_mode: True

configs/experiment/scheduling/base.yaml

@@ -0,0 +1,38 @@
+# @package _global_
+
+defaults:
+  - override /model: l2d.yaml
+  - override /callbacks: default.yaml
+  - override /trainer: default.yaml
+  - override /logger: wandb.yaml
+
+logger:
+  wandb:
+    project: "rl4co"
+    log_model: "all"
+    group: "${env.name}-${env.generator_params.num_jobs}-${env.generator_params.num_machines}"
+    tags: ???
+    name: ???
+
+trainer:
+  max_epochs: 10
+  # NOTE for some reason l2d is extremely sensitive to precision
+  # ONLY USE 32-true for l2d!
+  precision: 32-true
+
+seed: 12345678
+
+scaling_factor: 20
+
+model:
+  _target_: ???
+  batch_size: ???
+  train_data_size: 2_000
+  val_data_size: 1_000
+  test_data_size: 1_000
+  optimizer_kwargs:
+    lr: 1e-4
+    weight_decay: 1e-6
+  lr_scheduler: "ExponentialLR"
+  lr_scheduler_kwargs:
+    gamma: 0.95

configs/experiment/scheduling/gnn-ppo.yaml

@@ -0,0 +1,35 @@
+# @package _global_
+
+defaults:
+  - scheduling/base
+
+logger:
+  wandb:
+    tags: ["gnn-ppo", "${env.name}"]
+    name: "gnn-ppo-${env.name}-${env.generator_params.num_jobs}j-${env.generator_params.num_machines}m"
+
+# params from Song et al.
+model:
+  _target_: rl4co.models.L2DPPOModel
+  policy_kwargs:
+    embed_dim: 128
+    num_encoder_layers: 3
+    scaling_factor: ${scaling_factor}
+    max_grad_norm: 1
+    ppo_epochs: 3
+    het_emb: False
+  batch_size: 128
+  val_batch_size: 512
+  test_batch_size: 64
+  mini_batch_size: 512
+  reward_scale: scale
+  optimizer_kwargs:
+    lr: 1e-4
+
+trainer:
+  max_epochs: 10
+
+
+env:
+  stepwise_reward: True
+  _torchrl_mode: True

configs/experiment/scheduling/hgnn-pomo.yaml

@@ -0,0 +1,27 @@
+# @package _global_
+
+defaults:
+  - scheduling/base
+
+logger:
+  wandb:
+    tags: ["hgnn-pomo", "${env.name}"]
+    name: "hgnn-pomo-${env.name}-${env.generator_params.num_jobs}j-${env.generator_params.num_machines}m"
+
+model:
+  _target_: rl4co.models.POMO
+  policy:
+    _target_: rl4co.models.L2DPolicy
+    env_name: ${env.name}
+    embed_dim: 256
+    num_encoder_layers: 3
+    stepwise_encoding: False
+    scaling_factor: ${scaling_factor}
+    het_emb: True
+  num_starts: 10
+  batch_size: 64
+  num_augment: 0
+  baseline: "shared"
+  metrics:
+    val: ["reward", "max_reward"]
+    test: ${model.metrics.val}

configs/experiment/scheduling/hgnn-ppo.yaml

@@ -0,0 +1,35 @@
+# @package _global_
+
+defaults:
+  - scheduling/base
+
+logger:
+  wandb:
+    tags: ["hgnn-ppo", "${env.name}"]
+    name: "hgnn-ppo-${env.name}-${env.generator_params.num_jobs}j-${env.generator_params.num_machines}m"
+
+# params from Song et al.
+model:
+  _target_: rl4co.models.L2DPPOModel
+  policy_kwargs:
+    embed_dim: 128
+    num_encoder_layers: 3
+    scaling_factor: ${scaling_factor}
+    max_grad_norm: 1
+    ppo_epochs: 3
+    het_emb: True
+  batch_size: 128
+  val_batch_size: 512
+  test_batch_size: 64
+  mini_batch_size: 512
+  reward_scale: scale
+  optimizer_kwargs:
+    lr: 1e-4
+
+trainer:
+  max_epochs: 10
+
+
+env:
+  stepwise_reward: True
+  _torchrl_mode: True