Update

[ghstack-poisoned]

examples/agents/recurrent_actor.py

@@ -0,0 +1,205 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+"""
+This code exemplifies how an actor that uses a RNN backbone can be built.
+
+It is based on snippets from the DQN with RNN tutorial.
+
+There are two main APIs to be aware of when using RNNs, and dedicated notes regarding these can be found at the end
+of this example: the `set_recurrent_mode` context manager, and the `make_tensordict_primer` method.
+
+"""
+from collections import OrderedDict
+
+import torch
+from tensordict.nn import TensorDictModule as Mod, TensorDictSequential as Seq
+from torch import nn
+
+from torchrl.envs import (
+    Compose,
+    GrayScale,
+    GymEnv,
+    InitTracker,
+    ObservationNorm,
+    Resize,
+    RewardScaling,
+    StepCounter,
+    ToTensorImage,
+    TransformedEnv,
+)
+from torchrl.modules import ConvNet, LSTMModule, MLP, QValueModule, set_recurrent_mode
+
+# Define the device to use for computations (GPU if available, otherwise CPU)
+device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+
+# Create a transformed environment using the CartPole-v1 gym environment
+env = TransformedEnv(
+    GymEnv("CartPole-v1", from_pixels=True, device=device),
+    # Apply a series of transformations to the environment:
+    # 1. Convert observations to tensor images
+    # 2. Convert images to grayscale
+    # 3. Resize images to 84x84 pixels
+    # 4. Keep track of the step count
+    # 5. Initialize a tracker for the environment
+    # 6. Scale rewards by a factor of 0.1
+    # 7. Normalize observations to have zero mean and unit variance (we'll adapt that dynamically later)
+    Compose(
+        ToTensorImage(),
+        GrayScale(),
+        Resize(84, 84),
+        StepCounter(),
+        InitTracker(),
+        RewardScaling(loc=0.0, scale=0.1),
+        ObservationNorm(standard_normal=True, in_keys=["pixels"]),
+    ),
+)
+
+# Initialize the normalization statistics for the observation norm transform
+env.transform[-1].init_stats(1000, reduce_dim=[0, 1, 2], cat_dim=0, keep_dims=[0])
+
+# Reset the environment to get an initial observation
+td = env.reset()
+
+# Define a feature extractor module that takes pixel observations as input
+# and outputs an embedding vector
+feature = Mod(
+    ConvNet(
+        num_cells=[32, 32, 64],
+        squeeze_output=True,
+        aggregator_class=nn.AdaptiveAvgPool2d,
+        aggregator_kwargs={"output_size": (1, 1)},
+        device=device,
+    ),
+    in_keys=["pixels"],
+    out_keys=["embed"],
+)
+
+# Get the shape of the embedding vector output by the feature extractor
+with torch.no_grad():
+    n_cells = feature(env.reset())["embed"].shape[-1]
+
+# Define an LSTM module that takes the embedding vector as input and outputs
+# a new embedding vector
+lstm = LSTMModule(
+    input_size=n_cells,
+    hidden_size=128,
+    device=device,
+    in_key="embed",
+    out_key="embed",
+)
+
+# Define a multi-layer perceptron (MLP) module that takes the LSTM output as
+# input and outputs action values
+mlp = MLP(
+    out_features=2,
+    num_cells=[
+        64,
+    ],
+    device=device,
+)
+
+# Initialize the bias of the last layer of the MLP to zero
+mlp[-1].bias.data.fill_(0.0)
+
+# Wrap the MLP in a TensorDictModule to handle input/output keys
+mlp = Mod(mlp, in_keys=["embed"], out_keys=["action_value"])
+
+# Define a Q-value module that computes the Q-value of the current state
+qval = QValueModule(action_space=None, spec=env.action_spec)
+
+# Add a TensorDictPrimer to the environment to ensure that the policy is aware
+# of the supplementary inputs and outputs (recurrent states) during rollout execution
+# This is necessary when using batched environments or parallel data collection
+env.append_transform(lstm.make_tensordict_primer())
+
+# Create a sequential module that combines the feature extractor, LSTM, MLP, and Q-value modules
+policy = Seq(OrderedDict(feature=feature, lstm=lstm, mlp=mlp, qval=qval))
+
+# Roll out the policy in the environment for 100 steps
+rollout = env.rollout(100, policy)
+print(rollout)
+
+# Print result:
+#
+# TensorDict(
+#     fields={
+#         action: Tensor(shape=torch.Size([10, 2]), device=cpu, dtype=torch.int64, is_shared=False),
+#         action_value: Tensor(shape=torch.Size([10, 2]), device=cpu, dtype=torch.float32, is_shared=False),
+#         chosen_action_value: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.float32, is_shared=False),
+#         done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False),
+#         embed: Tensor(shape=torch.Size([10, 128]), device=cpu, dtype=torch.float32, is_shared=False),
+#         is_init: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False),
+#         next: TensorDict(
+#             fields={
+#                 done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False),
+#                 is_init: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False),
+#                 pixels: Tensor(shape=torch.Size([10, 1, 84, 84]), device=cpu, dtype=torch.float32, is_shared=False),
+#                 recurrent_state_c: Tensor(shape=torch.Size([10, 1, 128]), device=cpu, dtype=torch.float32, is_shared=False),
+#                 recurrent_state_h: Tensor(shape=torch.Size([10, 1, 128]), device=cpu, dtype=torch.float32, is_shared=False),
+#                 reward: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.float32, is_shared=False),
+#                 step_count: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.int64, is_shared=False),
+#                 terminated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False),
+#                 truncated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
+#             batch_size=torch.Size([10]),
+#             device=cpu,
+#             is_shared=False),
+#         pixels: Tensor(shape=torch.Size([10, 1, 84, 84]), device=cpu, dtype=torch.float32, is_shared=False),
+#         recurrent_state_c: Tensor(shape=torch.Size([10, 1, 128]), device=cpu, dtype=torch.float32, is_shared=False),
+#         recurrent_state_h: Tensor(shape=torch.Size([10, 1, 128]), device=cpu, dtype=torch.float32, is_shared=False),
+#         step_count: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.int64, is_shared=False),
+#         terminated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False),
+#         truncated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
+#     batch_size=torch.Size([10]),
+#     device=cpu,
+#     is_shared=False)
+#
+
+# Notes:
+# 1. make_tensordict_primer
+#
+# Regarding make_tensordict_primer, it creates a TensorDictPrimer object that ensures the policy is aware
+# of the supplementary inputs and outputs (recurrent states) during rollout execution.
+# This is necessary when using batched environments or parallel data collection, as the recurrent states
+# need to be shared across processes and dealt with properly.
+#
+# In other words, make_tensordict_primer adds the LSTM's hidden states to the environment's specs,
+# allowing the environment to properly handle the recurrent states during rollouts. Without it, the policy
+# would not be able to use the LSTM's memory buffers correctly, leading to poorly defined behaviors,
+# especially in parallel settings.
+#
+# By adding the TensorDictPrimer to the environment, you ensure that the policy can correctly use the
+# LSTM's recurrent states, even when running in parallel or batched environments. This is why
+# env.append_transform(lstm.make_tensordict_primer()) is called before creating the policy and rolling it
+# out in the environment.
+#
+# 2. Using the LSTM to process multiple steps at once.
+#
+# When set_recurrent_mode("recurrent") is used, the LSTM will process the entire input tensordict as a sequence, using
+# its recurrent connections to maintain state across time steps. This mode may utilize CuDNN to accelerate the processing
+# of the sequence on CUDA devices. The behavior in this mode is akin to torch.nn.LSTM, where the LSTM expects the input
+# data to be organized in batches of sequences.
+#
+# On the other hand, when set_recurrent_mode("sequential") is used, the
+# LSTM will process each step in the input tensordict independently, without maintaining any state across time steps. This
+# mode makes the LSTM behave similarly to torch.nn.LSTMCell, where each input is treated as a separate, independent
+# element.
+#
+# In the example code, set_recurrent_mode("recurrent") is used to process a tensordict of shape [T], where T
+# is the number of steps. This allows the LSTM to use its recurrent connections to maintain state across the entire
+# sequence.
+#
+# In contrast, set_recurrent_mode("sequential") is used to process a single step from the tensordict (i.e.,
+# rollout[0]). In this case, the LSTM does not use its recurrent connections, and simply processes the single step as if
+# it were an independent input.
+
+with set_recurrent_mode("recurrent"):
+    # Process a tensordict of shape [T] where T is a number of steps
+    print(policy(rollout))
+
+with set_recurrent_mode("sequential"):
+    # Process a tensordict of shape [T] where T is a number of steps
+    print(policy(rollout[0]))

torchrl/modules/tensordict_module/rnn.py

@@ -1652,8 +1652,8 @@ class set_recurrent_mode(_DecoratorContextManager):
     """Context manager for setting RNNs recurrent mode.
 
     Args:
-        mode (bool, "recurrent" or "stateful"): the recurrent mode to be used within the context manager.
-            `"recurrent"` leads to `mode=True` and `"stateful"` leads to `mode=False`.
+        mode (bool, "recurrent" or "sequential"): the recurrent mode to be used within the context manager.
+            `"recurrent"` leads to `mode=True` and `"sequential"` leads to `mode=False`.
             An RNN executed with recurrent_mode "on" assumes that the data comes in time batches, otherwise
             it is assumed that each data element in a tensordict is independent of the others.
             The default value of this context manager is ``True``.