Put in common gnn features extractor between ray.rllib and sb3 wrappers

The GNN features extractor used by rallib and sb3 are the same (except
the one for sb3 derives from a BaseFeaturesExtractor) and so is the
function conveting GraphInstance into torch_geometric.data.Data.

We put that in common in hub/solver/utils/gnn.

We also remove unused code in ray.rllib gnn code.

skdecide/hub/solver/ray_rllib/gnn/models/torch/gnn.py

@@ -8,11 +8,11 @@
 from ray.rllib.utils.typing import ModelConfigDict
 from torch import nn
 
-from skdecide.hub.solver.ray_rllib.gnn.torch_layers import GraphFeaturesExtractor
 from skdecide.hub.solver.ray_rllib.gnn.utils.spaces.space_utils import (
     convert_dict_space_to_graph_space,
     is_graph_dict_space,
 )
+from skdecide.hub.solver.utils.gnn.torch_layers import GraphFeaturesExtractor
 
 
 class GnnBasedModel(TorchModelV2, nn.Module):

skdecide/hub/solver/ray_rllib/gnn/utils/torch_utils.py

@@ -15,30 +15,7 @@
     is_graph_dict_multiinput,
     is_masked_obs,
 )
-
-
-def graph_obs_to_thg_data(
-    obs: gym.spaces.GraphInstance,
-    device: Optional[th.device] = None,
-    pin_memory: bool = False,
-) -> thg.data.Data:
-    # Node features
-    flatten_node_features = obs.nodes.reshape((len(obs.nodes), -1))
-    x = th.tensor(flatten_node_features).float()
-    # Edge features
-    if obs.edges is None:
-        edge_attr = None
-    else:
-        flatten_edge_features = obs.edges.reshape((len(obs.edges), -1))
-        edge_attr = th.tensor(flatten_edge_features).float()
-    edge_index = th.tensor(obs.edge_links, dtype=th.long).t().contiguous().view(2, -1)
-    # thg.Data
-    data = thg.data.Data(x=x, edge_index=edge_index, edge_attr=edge_attr)
-    # Pin the tensor's memory (for faster transfer to GPU later).
-    if pin_memory and th.cuda.is_available():
-        data.pin_memory()
-
-    return data if device is None else data.to(device)
+from skdecide.hub.solver.utils.gnn.torch_utils import graph_obs_to_thg_data
 
 
 def convert_to_torch_tensor(

skdecide/hub/solver/stable_baselines/gnn/common/torch_layers.py

@@ -9,6 +9,8 @@
 from torch import nn
 from torch_geometric.nn import global_max_pool
 
+from skdecide.hub.solver.utils.gnn import torch_layers
+
 
 class GraphFeaturesExtractor(BaseFeaturesExtractor):
     """Graph feature extractor for Graph observation spaces.
@@ -20,6 +22,10 @@ class GraphFeaturesExtractor(BaseFeaturesExtractor):
     - gnn: a 2-layers GCN
     - reduction layer: global_max_pool + linear layer + relu
 
+    This merely wraps `skdecide.hub.solver.utils.gnn.torch_layers.GraphFeaturesExtractor` to
+    makes it a `stable_baselines3.common.torch_layers.BaseFeaturesExtractor`. See the former documentation
+    for more precisions about its arguments.
+
     Args:
         observation_space:
         features_dim: Number of extracted features
@@ -45,75 +51,18 @@ def __init__(
         reduction_layer_class: Optional[type[nn.Module]] = None,
         reduction_layer_kwargs: Optional[dict[str, Any]] = None,
     ):
-
-        super().__init__(observation_space, features_dim=features_dim)
-
-        if gnn_out_dim is None:
-            if gnn_class is None:
-                gnn_out_dim = 2 * features_dim
-            else:
-                raise ValueError(
-                    "`gnn_out_dim` cannot be None if `gnn` is not None, "
-                    "and should match `gnn` output."
-                )
-
-        if gnn_class is None:
-            node_features_dim = int(np.prod(observation_space.node_space.shape))
-            self.gnn = thg.nn.models.GCN(
-                in_channels=node_features_dim,
-                hidden_channels=gnn_out_dim,
-                num_layers=2,
-                dropout=0.2,
-            )
-        else:
-            if gnn_kwargs is None:
-                gnn_kwargs = {}
-            self.gnn = gnn_class(**gnn_kwargs)
-
-        if reduction_layer_class is None:
-            self.reduction_layer = _DefaultReductionLayer(
-                gnn_out_dim=gnn_out_dim, features_dim=features_dim
-            )
-        else:
-            if reduction_layer_kwargs is None:
-                reduction_layer_kwargs = {}
-            self.reduction_layer = reduction_layer_class(**reduction_layer_kwargs)
+        self._extractor = torch_layers.GraphFeaturesExtractor(
+            observation_space=observation_space,
+            features_dim=features_dim,
+            gnn_out_dim=gnn_out_dim,
+            gnn_class=gnn_class,
+            gnn_kwargs=gnn_kwargs,
+            reduction_layer_class=reduction_layer_class,
+            reduction_layer_kwargs=reduction_layer_kwargs,
+        )
 
     def forward(self, observations: thg.data.Data) -> th.Tensor:
-        x, edge_index, edge_attr, batch = (
-            observations.x,
-            observations.edge_index,
-            observations.edge_attr,
-            observations.batch,
-        )
-        # construct edge weights, for GNNs needing it, as the first edge feature
-        edge_weight = edge_attr[:, 0]
-        h = self.gnn(
-            x=x, edge_index=edge_index, edge_weight=edge_weight, edge_attr=edge_attr
-        )
-        embedded_observations = thg.data.Data(
-            x=h, edge_index=edge_index, edge_attr=edge_attr, batch=batch
-        )
-        h = self.reduction_layer(embedded_observations=embedded_observations)
-        return h
-
-
-class _DefaultReductionLayer(nn.Module):
-    def __init__(self, gnn_out_dim: int, features_dim: int):
-        super().__init__()
-        self.gnn_out_dim = gnn_out_dim
-        self.features_dim = features_dim
-        self.linear_layer = nn.Linear(gnn_out_dim, features_dim)
-
-    def forward(self, embedded_observations: thg.data.Data) -> th.Tensor:
-        x, edge_index, batch = (
-            embedded_observations.x,
-            embedded_observations.edge_index,
-            embedded_observations.batch,
-        )
-        h = global_max_pool(x, batch)
-        h = self.linear_layer(h).relu()
-        return h
+        return self._extractor.forward(observations=observations)
 
 
 class CombinedFeaturesExtractor(BaseFeaturesExtractor):

skdecide/hub/solver/stable_baselines/gnn/common/utils.py

@@ -6,6 +6,8 @@
 import torch as th
 import torch_geometric as thg
 
+from skdecide.hub.solver.utils.gnn.torch_utils import graph_obs_to_thg_data
+
 SubObsType = Union[np.ndarray, gym.spaces.GraphInstance, list[gym.spaces.GraphInstance]]
 ObsType = Union[SubObsType, dict[str, SubObsType]]
 TorchSubObsType = Union[th.Tensor, thg.data.Data]
@@ -27,22 +29,6 @@ def copy_np_array_or_list_of_graph_instances(
         return np.copy(obs)
 
 
-def graph_obs_to_thg_data(
-    obs: gym.spaces.GraphInstance, device: th.device
-) -> thg.data.Data:
-    # Node features
-    flatten_node_features = obs.nodes.reshape((len(obs.nodes), -1))
-    x = th.tensor(flatten_node_features).float()
-    # Edge features
-    if obs.edges is None:
-        edge_attr = None
-    else:
-        flatten_edge_features = obs.edges.reshape((len(obs.edges), -1))
-        edge_attr = th.tensor(flatten_edge_features).float()
-    edge_index = th.tensor(obs.edge_links, dtype=th.long).t().contiguous().view(2, -1)
-    return thg.data.Data(x=x, edge_index=edge_index, edge_attr=edge_attr).to(device)
-
-
 def obs_as_tensor(
     obs: ObsType,
     device: th.device,

skdecide/hub/solver/ray_rllib/gnn/torch_layers.py → skdecide/hub/solver/utils/gnn/torch_layers.py

@@ -77,7 +77,7 @@ def __init__(
                 reduction_layer_kwargs = {}
             self.reduction_layer = reduction_layer_class(**reduction_layer_kwargs)
 
-    def forward(self, observations) -> th.Tensor:
+    def forward(self, observations: thg.data.Data) -> th.Tensor:
         x, edge_index, edge_attr, batch = (
             observations.x,
             observations.edge_index,

skdecide/hub/solver/utils/gnn/torch_utils.py

@@ -0,0 +1,29 @@
+from typing import Optional
+
+import gymnasium as gym
+import torch as th
+import torch_geometric as thg
+
+
+def graph_obs_to_thg_data(
+    obs: gym.spaces.GraphInstance,
+    device: Optional[th.device] = None,
+    pin_memory: bool = False,
+) -> thg.data.Data:
+    # Node features
+    flatten_node_features = obs.nodes.reshape((len(obs.nodes), -1))
+    x = th.tensor(flatten_node_features).float()
+    # Edge features
+    if obs.edges is None:
+        edge_attr = None
+    else:
+        flatten_edge_features = obs.edges.reshape((len(obs.edges), -1))
+        edge_attr = th.tensor(flatten_edge_features).float()
+    edge_index = th.tensor(obs.edge_links, dtype=th.long).t().contiguous().view(2, -1)
+    # thg.Data
+    data = thg.data.Data(x=x, edge_index=edge_index, edge_attr=edge_attr)
+    # Pin the tensor's memory (for faster transfer to GPU later).
+    if pin_memory and th.cuda.is_available():
+        data.pin_memory()
+
+    return data if device is None else data.to(device)