Fix diffusion (rm transpose), Add prefetch

lerobot/common/datasets/factory.py

@@ -1,9 +1,8 @@
 import torch
+from torchrl.data.replay_buffers import PrioritizedSliceSampler
 
 from lerobot.common.datasets.pusht import PushtExperienceReplay
 from lerobot.common.datasets.simxarm import SimxarmExperienceReplay
-# from rl.torchrl.data.replay_buffers.samplers import PrioritizedSliceSampler
-from torchrl.data.replay_buffers import PrioritizedSliceSampler
 
 # TODO(rcadene): implement
 
@@ -27,6 +26,17 @@ from torchrl.data.replay_buffers import PrioritizedSliceSampler
 
 
 def make_offline_buffer(cfg, sampler=None):
+    if cfg.policy.balanced_sampling:
+        assert cfg.online_steps > 0
+        batch_size = None
+        pin_memory = False
+        prefetch = None
+    else:
+        assert cfg.online_steps == 0
+        num_slices = cfg.policy.batch_size
+        batch_size = cfg.policy.horizon * num_slices
+        pin_memory = cfg.device == "cuda"
+        prefetch = cfg.prefetch
 
     overwrite_sampler = sampler is not None
 
@@ -52,6 +62,9 @@ def make_offline_buffer(cfg, sampler=None):
             streaming=False,
             root="data",
             sampler=sampler,
+            batch_size=batch_size,
+            pin_memory=pin_memory,
+            prefetch=prefetch,
         )
     elif cfg.env.name == "pusht":
         offline_buffer = PushtExperienceReplay(
@@ -61,6 +74,9 @@ def make_offline_buffer(cfg, sampler=None):
             streaming=False,
             root="data",
             sampler=sampler,
+            batch_size=batch_size,
+            pin_memory=pin_memory,
+            prefetch=prefetch,
         )
     else:
         raise ValueError(cfg.env.name)

lerobot/common/policies/diffusion.py

@@ -119,9 +119,9 @@ class DiffusionPolicy(nn.Module):
         assert batch_size % num_slices == 0
 
         def process_batch(batch, horizon, num_slices):
-            # trajectory t = 256, horizon h = 5
-            # (t h) ... -> h t ...
-            batch = batch.reshape(num_slices, horizon).transpose(1, 0).contiguous()
+            # trajectory t = 64, horizon h = 16
+            # (t h) ... -> t h ...
+            batch = batch.reshape(num_slices, horizon)  # .transpose(1, 0).contiguous()
 
             out = {
                 "obs": {
@@ -132,7 +132,10 @@ class DiffusionPolicy(nn.Module):
             }
             return out
 
-        batch = replay_buffer.sample(batch_size)
+        if self.cfg.balanced_sampling:
+            batch = replay_buffer.sample(batch_size)
+        else:
+            batch = replay_buffer.sample()
         batch = process_batch(batch, self.cfg.horizon, num_slices)
 
         loss = self.diffusion.compute_loss(batch)
@@ -149,4 +152,17 @@ class DiffusionPolicy(nn.Module):
             "total_loss": loss.item(),
             "lr": self.lr_scheduler.get_last_lr()[0],
         }
+
+        # TODO(rcadene): remove hardcoding
+        # in diffusion_policy, len(dataloader) is 168 for a batch_size of 64
+        if step % 168 == 0:
+            self.global_step += 1
+
         return metrics
+
+    def save(self, fp):
+        torch.save(self.state_dict(), fp)
+
+    def load(self, fp):
+        d = torch.load(fp)
+        self.load_state_dict(d)

lerobot/common/policies/tdmpc.py

@@ -359,7 +359,11 @@ class TDMPC(nn.Module):
             weights = batch["_weight"][FIRST_FRAME, :, None]
             return obs, action, next_obses, reward, mask, done, idxs, weights
 
-        batch = replay_buffer.sample(batch_size)
+        if self.cfg.balanced_sampling:
+            batch = replay_buffer.sample(batch_size)
+        else:
+            batch = replay_buffer.sample()
+
         obs, action, next_obses, reward, mask, done, idxs, weights = process_batch(
             batch, self.cfg.horizon, num_slices
         )

lerobot/configs/default.yaml

@@ -12,6 +12,7 @@ hydra:
 seed: 1337
 device: cuda
 buffer_device: cuda
+prefetch: 4
 eval_freq: ???
 save_freq: ???
 eval_episodes: ???

lerobot/configs/policy/diffusion.yaml

@@ -21,7 +21,12 @@ past_action_visible: False
 keypoint_visible_rate: 1.0
 obs_as_global_cond: True
 
-offline_steps: 50000
+eval_episodes: 50
+eval_freq: 10000
+save_freq: 100000
+log_freq: 250
+
+offline_steps: 1344000
 online_steps: 0
 
 policy:
@@ -48,8 +53,7 @@ policy:
   per_alpha: 0.6
   per_beta: 0.4
 
-  balanced_sampling: true
-
+  balanced_sampling: false
   utd: 1
   offline_steps: ${offline_steps}
   use_ema: true

sbatch.sh

@@ -6,7 +6,7 @@
 #SBATCH --time=2-00:00:00
 #SBATCH --output=/home/rcadene/slurm/%j.out
 #SBATCH --error=/home/rcadene/slurm/%j.err
-#SBATCH --qos=low
+#SBATCH --qos=medium
 #SBATCH --mail-user=re.cadene@gmail.com
 #SBATCH --mail-type=ALL