Merge pull request #256 from bokveizen/fanchen_250306

[Bug fixes for #255] Added environment config files for FLP and MCP

configs/env/flp.yaml

@@ -0,0 +1,13 @@
+_target_: rl4co.envs.FLPEnv
+name: flp
+
+generator_params:
+  num_loc: 100
+  min_loc: 0.0
+  max_loc: 1.0
+  loc_distribution: uniform
+  to_choose: 10
+
+# data_dir: ${paths.root_dir}/data/mcp
+# val_file: mcp${env.generator_params.num_loc}_val_seed4321.npz
+# test_file: mcp${env.generator_params.num_loc}_test_seed1234.npz

configs/env/mcp.yaml

@@ -0,0 +1,17 @@
+_target_: rl4co.envs.MCPEnv
+name: mcp
+
+generator_params:
+  num_items: 200
+  num_sets: 100
+  min_weight: 1
+  max_weight: 10
+  min_size: 5
+  max_size: 15
+  n_sets_to_choose: 10
+  size_distribution: uniform
+  weight_distribution: uniform
+
+# data_dir: ${paths.root_dir}/data/mcp
+# val_file: mcp${env.generator_params.num_loc}_val_seed4321.npz
+# test_file: mcp${env.generator_params.num_loc}_test_seed1234.npz

configs/experiment/graph/am.yaml

@@ -2,7 +2,7 @@
 
 defaults:
   - override /model: am.yaml
-  - override /env: mcp.yaml
+  - override /env: flp.yaml
   - override /callbacks: default.yaml
   - override /trainer: default.yaml
   - override /logger: wandb.yaml

rl4co/models/nn/env_embeddings/context.py

@@ -35,6 +35,7 @@ def env_context_embedding(env_name: str, config: dict) -> nn.Module:
         "mdcpdp": MDCPDPContext,
         "mtvrp": MTVRPContext,
         "shpp": TSPContext,
+        "flp": FLPContext,
     }
 
     if env_name not in embedding_registry:
@@ -372,3 +373,22 @@ def _state_embedding(self, embeddings, td):
             ],
             -1,
         )
+
+class FLPContext(EnvContext):
+    """Context embedding for the Facility Location Problem (FLP).    
+    """
+    def __init__(self, embed_dim: int):
+        super(FLPContext, self).__init__(embed_dim=embed_dim)
+        self.embed_dim = embed_dim
+        # self.mlp_context = MLP(embed_dim, [embed_dim, embed_dim])
+        self.projection = nn.Linear(embed_dim, embed_dim, bias=True)
+
+    def forward(self, embeddings, td):        
+        cur_dist = td["distances"].unsqueeze(-2)  # (batch_size, 1, n_points)
+        dist_improve = cur_dist - td["orig_distances"]  # (batch_size, n_points, n_points)
+        dist_improve = torch.clamp(dist_improve, min=0).sum(-1) # (batch_size, n_points)       
+
+        # softmax
+        loc_best_soft = torch.softmax(dist_improve, dim=-1) # (batch_size, n_points)        
+        embed_best = (embeddings * loc_best_soft[..., None]).sum(-2)
+        return embed_best

rl4co/models/nn/env_embeddings/init.py

@@ -40,6 +40,7 @@ def env_init_embedding(env_name: str, config: dict) -> nn.Module:
         "jssp": FJSPInitEmbedding,
         "mtvrp": MTVRPInitEmbedding,
         "shpp": TSPInitEmbedding,
+        "flp": FLPInitEmbedding,
     }
 
     if env_name not in embedding_registry:
@@ -562,3 +563,12 @@ def forward(self, td):
             )
         )
         return torch.cat((depot_embedding, node_embeddings), -2)
+
+class FLPInitEmbedding(nn.Module):
+    def __init__(self, embed_dim: int):
+        super().__init__()
+        self.projection = nn.Linear(2, embed_dim, bias=True)
+
+    def forward(self, td: TensorDict):
+        hdim = self.projection(td["locs"])
+        return hdim