[Feature] Plotting TensorDictSequential graphs

ghstack-source-id: ff93fb4
Pull Request resolved: #1144

Author: Vincent Moens

tensordict/nn/probabilistic.py

@@ -9,7 +9,7 @@
 import warnings
 
 from textwrap import indent
-from typing import Any, Dict, List, Optional, overload, OrderedDict
+from typing import Any, Dict, List, Optional, OrderedDict, overload
 
 import torch
 
@@ -800,8 +800,7 @@ def __init__(
         aggregate_probabilities: bool | None = None,
         include_sum: bool | None = None,
         inplace: bool | None = None,
-    ) -> None:
-        ...
+    ) -> None: ...
 
     @overload
     def __init__(
@@ -812,8 +811,7 @@ def __init__(
         aggregate_probabilities: bool | None = None,
         include_sum: bool | None = None,
         inplace: bool | None = None,
-    ) -> None:
-        ...
+    ) -> None: ...
 
     def __init__(
         self,

tensordict/nn/sequence.py

@@ -6,6 +6,7 @@
 from __future__ import annotations
 
 import collections
+import contextlib
 import logging
 from copy import deepcopy
 from typing import Any, Callable, Iterable, List, OrderedDict, overload
@@ -574,3 +575,64 @@ def __setitem__(
 
     def __delitem__(self, index: int | slice | str) -> None:
         self.module.__delitem__(idx=index)
+
+    def plot(self, example_input: TensorDictBase | None = None, **kwargs):
+        import pydot
+
+        graph = pydot.Dot(
+            "my_graph", graph_type="digraph", bgcolor="yellow", splines="curved"
+        )
+        graph.set_bgcolor("white")
+
+        if example_input is not None:
+            from torch._subclasses.fake_tensor import FakeTensorMode
+
+            fake_mode = FakeTensorMode()
+            converter = fake_mode.fake_tensor_converter
+            fake_td = example_input.apply(
+                lambda x: converter.from_real_tensor(fake_mode, x)
+            )
+        else:
+            fake_td = None
+            fake_mode = contextlib.nullcontext()
+
+        with fake_mode:
+            iterator = (
+                enumerate(self._module_iter())
+                if not isinstance(self.module, nn.ModuleDict)
+                else self.module.items()
+            )
+            for name, module in iterator:
+                graph.add_node(
+                    pydot.Node(str(name), shape="box")
+                )  # label=str(node.module)))
+
+                # Check if in_keys are there already
+                in_keys = module.in_keys
+                for in_key in in_keys:
+                    if in_key not in graph.obj_dict["nodes"]:
+                        in_key_node = pydot.Node(
+                            in_key, label=in_key, shape="plaintext"
+                        )
+                        graph.add_node(in_key_node)
+                    in_key_edge = pydot.Edge(
+                        in_key, str(name), color="blue", style="arrow"
+                    )
+                    graph.add_edge(in_key_edge)
+
+                if not isinstance(module, TensorDictModule):
+                    fake_td = self._run_module(module, fake_td, **kwargs)
+
+                out_keys = module.out_keys
+                for out_key in out_keys:
+                    if out_key not in graph.obj_dict["nodes"]:
+                        out_key_node = pydot.Node(
+                            out_key, label=out_key, shape="plaintext"
+                        )
+                        graph.add_node(out_key_node)
+                    out_key_edge = pydot.Edge(
+                        str(name), out_key, color="blue", style="arrow"
+                    )
+                    graph.add_edge(out_key_edge)
+
+        graph.write_png("/Users/vmoens/Downloads/my_graph.png")