Merge pull request #188 from GFNOrg/conditional_gfn

conditional gfn

README.md

@@ -82,7 +82,7 @@ module_PF = NeuralNet(
 module_PB = NeuralNet(
     input_dim=env.preprocessor.output_dim,
     output_dim=env.n_actions - 1,
-    torso=module_PF.torso  # We share all the parameters of P_F and P_B, except for the last layer
+    trunk=module_PF.trunk  # We share all the parameters of P_F and P_B, except for the last layer
 )
 
 # 3 - We define the estimators.
@@ -136,7 +136,7 @@ module_PF = NeuralNet(
 module_PB = NeuralNet(
     input_dim=env.preprocessor.output_dim,
     output_dim=env.n_actions - 1,
-    torso=module_PF.torso  # We share all the parameters of P_F and P_B, except for the last layer
+    trunk=module_PF.trunk  # We share all the parameters of P_F and P_B, except for the last layer
 )
 module_logF = NeuralNet(
     input_dim=env.preprocessor.output_dim,

src/gfn/containers/trajectories.py

@@ -1,11 +1,12 @@
 from __future__ import annotations
 
-from typing import TYPE_CHECKING, Sequence
+from typing import TYPE_CHECKING, Sequence, Union, Tuple
+
 
 if TYPE_CHECKING:
     from gfn.actions import Actions
     from gfn.env import Env
-    from gfn.states import States
+    from gfn.states import States, DiscreteStates
 
 import numpy as np
 import torch
@@ -50,6 +51,7 @@ def __init__(
         self,
         env: Env,
         states: States | None = None,
+        conditioning: torch.Tensor | None = None,
         actions: Actions | None = None,
         when_is_done: TT["n_trajectories", torch.long] | None = None,
         is_backward: bool = False,
@@ -76,6 +78,7 @@ def __init__(
         is used to compute the rewards, at each call of self.log_rewards
         """
         self.env = env
+        self.conditioning = conditioning
         self.is_backward = is_backward
         self.states = (
             states if states is not None else env.states_from_batch_shape((0, 0))
@@ -315,6 +318,15 @@ def extend(self, other: Trajectories) -> None:
 
     def to_transitions(self) -> Transitions:
         """Returns a `Transitions` object from the trajectories."""
+        if self.conditioning is not None:
+            traj_len = self.actions.batch_shape[0]
+            expand_dims = (traj_len,) + tuple(self.conditioning.shape)
+            conditioning = self.conditioning.unsqueeze(0).expand(expand_dims)[
+                ~self.actions.is_dummy
+            ]
+        else:
+            conditioning = None
+
         states = self.states[:-1][~self.actions.is_dummy]
         next_states = self.states[1:][~self.actions.is_dummy]
         actions = self.actions[~self.actions.is_dummy]
@@ -348,6 +360,7 @@ def to_transitions(self) -> Transitions:
         return Transitions(
             env=self.env,
             states=states,
+            conditioning=conditioning,
             actions=actions,
             is_done=is_done,
             next_states=next_states,
@@ -363,7 +376,10 @@ def to_states(self) -> States:
 
     def to_non_initial_intermediary_and_terminating_states(
         self,
-    ) -> tuple[States, States]:
+    ) -> Union[
+        Tuple[States, States, torch.Tensor, torch.Tensor],
+        Tuple[States, States, None, None],
+    ]:
         """Returns all intermediate and terminating `States` from the trajectories.
 
         This is useful for the flow matching loss, that requires its inputs to be distinguished.
@@ -373,10 +389,27 @@ def to_non_initial_intermediary_and_terminating_states(
             are not s0.
         """
         states = self.states
+
+        if self.conditioning is not None:
+            traj_len = self.states.batch_shape[0]
+            expand_dims = (traj_len,) + tuple(self.conditioning.shape)
+            intermediary_conditioning = self.conditioning.unsqueeze(0).expand(
+                expand_dims
+            )[~states.is_sink_state & ~states.is_initial_state]
+            conditioning = self.conditioning  # n_final_states == n_trajectories.
+        else:
+            intermediary_conditioning = None
+            conditioning = None
+
         intermediary_states = states[~states.is_sink_state & ~states.is_initial_state]
         terminating_states = self.last_states
         terminating_states.log_rewards = self.log_rewards
-        return intermediary_states, terminating_states
+        return (
+            intermediary_states,
+            terminating_states,
+            intermediary_conditioning,
+            conditioning,
+        )
 
 
 def pad_dim0_to_target(a: torch.Tensor, target_dim0: int) -> torch.Tensor:

src/gfn/containers/transitions.py

@@ -34,6 +34,7 @@ def __init__(
         self,
         env: Env,
         states: States | None = None,
+        conditioning: torch.Tensor | None = None,
         actions: Actions | None = None,
         is_done: TT["n_transitions", torch.bool] | None = None,
         next_states: States | None = None,
@@ -65,6 +66,7 @@ def __init__(
                 `batch_shapes`.
         """
         self.env = env
+        self.conditioning = conditioning
         self.is_backward = is_backward
         self.states = (
             states

src/gfn/gflownet/base.py

@@ -1,10 +1,9 @@
 import math
 from abc import ABC, abstractmethod
-from typing import Generic, Tuple, TypeVar, Union
+from typing import Generic, Tuple, TypeVar, Union, Any
 
 import torch
 import torch.nn as nn
-from torch import Tensor
 from torchtyping import TensorType as TT
 
 from gfn.containers import Trajectories
@@ -14,6 +13,10 @@
 from gfn.samplers import Sampler
 from gfn.states import States
 from gfn.utils.common import has_log_probs
+from gfn.utils.handlers import (
+    has_conditioning_exception_handler,
+    no_conditioning_exception_handler,
+)
 
 TrainingSampleType = TypeVar(
     "TrainingSampleType", bound=Union[Container, tuple[States, ...]]
@@ -32,48 +35,48 @@ class GFlowNet(ABC, nn.Module, Generic[TrainingSampleType]):
     def sample_trajectories(
         self,
         env: Env,
-        n_samples: int,
+        n: int,
         save_logprobs: bool = True,
         save_estimator_outputs: bool = False,
     ) -> Trajectories:
         """Sample a specific number of complete trajectories.
 
         Args:
             env: the environment to sample trajectories from.
-            n_samples: number of trajectories to be sampled.
+            n: number of trajectories to be sampled.
             save_logprobs: whether to save the logprobs of the actions - useful for on-policy learning.
             save_estimator_outputs: whether to save the estimator outputs - useful for off-policy learning
                         with tempered policy
         Returns:
             Trajectories: sampled trajectories object.
         """
 
-    def sample_terminating_states(self, env: Env, n_samples: int) -> States:
+    def sample_terminating_states(self, env: Env, n: int) -> States:
         """Rolls out the parametrization's policy and returns the terminating states.
 
         Args:
             env: the environment to sample terminating states from.
-            n_samples: number of terminating states to be sampled.
+            n: number of terminating states to be sampled.
         Returns:
             States: sampled terminating states object.
         """
         trajectories = self.sample_trajectories(
-            env, n_samples, save_estimator_outputs=False, save_logprobs=False
+            env, n, save_estimator_outputs=False, save_logprobs=False
         )
         return trajectories.last_states
 
     def logz_named_parameters(self):
-        return {"logZ": dict(self.named_parameters())["logZ"]}
+        return {k: v for k, v in dict(self.named_parameters()).items() if "logZ" in k}
 
     def logz_parameters(self):
-        return [dict(self.named_parameters())["logZ"]]
+        return [v for k, v in dict(self.named_parameters()).items() if "logZ" in k]
 
     @abstractmethod
     def to_training_samples(self, trajectories: Trajectories) -> TrainingSampleType:
         """Converts trajectories to training samples. The type depends on the GFlowNet."""
 
     @abstractmethod
-    def loss(self, env: Env, training_objects):
+    def loss(self, env: Env, training_objects: Any):
         """Computes the loss given the training objects."""
 
 
@@ -93,18 +96,20 @@ def __init__(self, pf: GFNModule, pb: GFNModule):
     def sample_trajectories(
         self,
         env: Env,
-        n_samples: int,
+        n: int,
+        conditioning: torch.Tensor | None = None,
         save_logprobs: bool = True,
         save_estimator_outputs: bool = False,
-        **policy_kwargs,
+        **policy_kwargs: Any,
     ) -> Trajectories:
         """Samples trajectories, optionally with specified policy kwargs."""
         sampler = Sampler(estimator=self.pf)
         trajectories = sampler.sample_trajectories(
             env,
-            n_trajectories=n_samples,
-            save_estimator_outputs=save_estimator_outputs,
+            n=n,
+            conditioning=conditioning,
             save_logprobs=save_logprobs,
+            save_estimator_outputs=save_estimator_outputs,
             **policy_kwargs,
         )
 
@@ -176,7 +181,20 @@ def get_pfs_and_pbs(
                     ~trajectories.actions.is_dummy
                 ]
             else:
-                estimator_outputs = self.pf(valid_states)
+                if trajectories.conditioning is not None:
+                    cond_dim = (-1,) * len(trajectories.conditioning.shape)
+                    traj_len = trajectories.states.tensor.shape[0]
+                    masked_cond = trajectories.conditioning.unsqueeze(0).expand(
+                        (traj_len,) + cond_dim
+                    )[~trajectories.states.is_sink_state]
+
+                    # Here, we pass all valid states, i.e., non-sink states.
+                    with has_conditioning_exception_handler("pf", self.pf):
+                        estimator_outputs = self.pf(valid_states, masked_cond)
+                else:
+                    # Here, we pass all valid states, i.e., non-sink states.
+                    with no_conditioning_exception_handler("pf", self.pf):
+                        estimator_outputs = self.pf(valid_states)
 
             # Calculates the log PF of the actions sampled off policy.
             valid_log_pf_actions = self.pf.to_probability_distribution(
@@ -196,7 +214,23 @@ def get_pfs_and_pbs(
 
         # Using all non-initial states, calculate the backward policy, and the logprobs
         # of those actions.
-        estimator_outputs = self.pb(non_initial_valid_states)
+        if trajectories.conditioning is not None:
+
+            # We need to index the conditioning vector to broadcast over the states.
+            cond_dim = (-1,) * len(trajectories.conditioning.shape)
+            traj_len = trajectories.states.tensor.shape[0]
+            masked_cond = trajectories.conditioning.unsqueeze(0).expand(
+                (traj_len,) + cond_dim
+            )[~trajectories.states.is_sink_state][~valid_states.is_initial_state]
+
+            # Pass all valid states, i.e., non-sink states, except the initial state.
+            with has_conditioning_exception_handler("pb", self.pb):
+                estimator_outputs = self.pb(non_initial_valid_states, masked_cond)
+        else:
+            # Pass all valid states, i.e., non-sink states, except the initial state.
+            with no_conditioning_exception_handler("pb", self.pb):
+                estimator_outputs = self.pb(non_initial_valid_states)
+
         valid_log_pb_actions = self.pb.to_probability_distribution(
             non_initial_valid_states, estimator_outputs
         ).log_prob(non_exit_valid_actions.tensor)