A tutorial on pin_memory and non_blocking usage

.ci/docker/requirements.txt

@@ -28,8 +28,9 @@ tensorboard
 jinja2==3.1.3
 pytorch-lightning
 torchx
-torchrl==0.3.0
-tensordict==0.3.0
+# TODO: use stable 0.5 when released
+-e git+https://github.com/pytorch/rl.git#egg=torchrl
+-e git+https://github.com/pytorch/tensordict.git#egg=tensordict
 ax-platform
 nbformat>==5.9.2
 datasets

advanced_source/coding_ddpg.py

@@ -182,7 +182,7 @@
 # Later, we will see how the target parameters should be updated in TorchRL.
 #
 
-from tensordict.nn import TensorDictModule
+from tensordict.nn import TensorDictModule, TensorDictSequential
 
 
 def _init(
@@ -290,12 +290,11 @@ def _loss_actor(
 ) -> torch.Tensor:
     td_copy = tensordict.select(*self.actor_in_keys)
     # Get an action from the actor network: since we made it functional, we need to pass the params
-    td_copy = self.actor_network(td_copy, params=self.actor_network_params)
+    with self.actor_network_params.to_module(self.actor_network):
+        td_copy = self.actor_network(td_copy)
     # get the value associated with that action
-    td_copy = self.value_network(
-        td_copy,
-        params=self.value_network_params.detach(),
-    )
+    with self.value_network_params.detach().to_module(self.value_network):
+        td_copy = self.value_network(td_copy)
     return -td_copy.get("state_action_value")
 
 
@@ -317,7 +316,8 @@ def _loss_value(
     td_copy = tensordict.clone()
 
     # V(s, a)
-    self.value_network(td_copy, params=self.value_network_params)
+    with self.value_network_params.to_module(self.value_network):
+        self.value_network(td_copy)
     pred_val = td_copy.get("state_action_value").squeeze(-1)
 
     # we manually reconstruct the parameters of the actor-critic, where the first
@@ -332,9 +332,8 @@ def _loss_value(
         batch_size=self.target_actor_network_params.batch_size,
         device=self.target_actor_network_params.device,
     )
-    target_value = self.value_estimator.value_estimate(
-        tensordict, target_params=target_params
-    ).squeeze(-1)
+    with target_params.to_module(self.actor_critic):
+        target_value = self.value_estimator.value_estimate(tensordict).squeeze(-1)
 
     # Computes the value loss: L2, L1 or smooth L1 depending on `self.loss_function`
     loss_value = distance_loss(pred_val, target_value, loss_function=self.loss_function)
@@ -717,7 +716,7 @@ def get_env_stats():
     ActorCriticWrapper,
     DdpgMlpActor,
     DdpgMlpQNet,
-    OrnsteinUhlenbeckProcessWrapper,
+    OrnsteinUhlenbeckProcessModule,
     ProbabilisticActor,
     TanhDelta,
     ValueOperator,
@@ -776,15 +775,18 @@ def make_ddpg_actor(
 # Exploration
 # ~~~~~~~~~~~
 #
-# The policy is wrapped in a :class:`~torchrl.modules.OrnsteinUhlenbeckProcessWrapper`
+# The policy is passed into a :class:`~torchrl.modules.OrnsteinUhlenbeckProcessModule`
 # exploration module, as suggested in the original paper.
 # Let's define the number of frames before OU noise reaches its minimum value
 annealing_frames = 1_000_000
 
-actor_model_explore = OrnsteinUhlenbeckProcessWrapper(
+actor_model_explore = TensorDictSequential(
     actor,
-    annealing_num_steps=annealing_frames,
-).to(device)
+    OrnsteinUhlenbeckProcessModule(
+        spec=actor.spec.clone(),
+        annealing_num_steps=annealing_frames,
+    ).to(device),
+)
 if device == torch.device("cpu"):
     actor_model_explore.share_memory()
 
@@ -1168,7 +1170,7 @@ def ceil_div(x, y):
         )
 
     # update the exploration strategy
-    actor_model_explore.step(current_frames)
+    actor_model_explore[1].step(current_frames)
 
 collector.shutdown()
 del collector

en-wordlist.txt

@@ -1,3 +1,4 @@
+
 ACL
 ADI
 AOT
@@ -50,6 +51,7 @@ DDP
 DDPG
 DDQN
 DLRM
+DMA
 DNN
 DQN
 DataLoaders
@@ -68,6 +70,8 @@ Ecker
 ExportDB
 FC
 FGSM
+tensordict
+DataLoader's
 FLAVA
 FSDP
 FX
@@ -139,6 +143,7 @@ MKLDNN
 MLP
 MLPs
 MNIST
+MPS
 MUC
 MacBook
 MacOS
@@ -219,6 +224,7 @@ STR
 SVE
 SciPy
 Sequentials
+Sharding
 Sigmoid
 SoTA
 Sohn
@@ -254,6 +260,7 @@ VLDB
 VQA
 VS Code
 ViT
+Volterra
 WMT
 WSI
 WSIs
@@ -336,11 +343,11 @@ dataset’s
 deallocation
 decompositions
 decorrelated
-devicemesh
 deserialize
 deserialized
 desynchronization
 deterministically
+devicemesh
 dimensionality
 dir
 discontiguous
@@ -384,6 +391,7 @@ hessian
 hessians
 histoencoder
 histologically
+homonymous
 hotspot
 hvp
 hyperparameter
@@ -459,6 +467,7 @@ optimizer's
 optimizers
 otsu
 overfitting
+pageable
 parallelizable
 parallelization
 parametrization
@@ -522,7 +531,6 @@ runtime
 runtimes
 scalable
 sharded
-Sharding
 softmax
 sparsified
 sparsifier
@@ -609,4 +617,4 @@ warmstarting
 warmup
 webp
 wsi
-wsis
+wsis

index.rst

@@ -93,6 +93,13 @@ Welcome to PyTorch Tutorials
    :link: intermediate/tensorboard_tutorial.html
    :tags: Interpretability,Getting-Started,TensorBoard
 
+.. customcarditem::
+   :header: Good usage of `non_blocking` and `pin_memory()` in PyTorch
+   :card_description: A guide on best practices to copy data from CPU to GPU.
+   :image: _static/img/pinmem.png
+   :link: intermediate/pinmem_nonblock.html
+   :tags: Getting-Started
+
 .. Image/Video
 
 .. customcarditem::
@@ -942,6 +949,7 @@ Additional Resources
    beginner/basics/autogradqs_tutorial
    beginner/basics/optimization_tutorial
    beginner/basics/saveloadrun_tutorial
+   intermediate/pinmem_nonblock
    advanced/custom_ops_landing_page
 
 .. toctree::

intermediate_source/dqn_with_rnn_tutorial.py

@@ -298,7 +298,7 @@
 # either by passing a string or an action-spec. This allows us to use
 # Categorical (sometimes called "sparse") encoding or the one-hot version of it.
 #
-qval = QValueModule(action_space=env.action_spec)
+qval = QValueModule(spec=env.action_spec)
 
 ######################################################################
 # .. note::