fix stats computation

tests/test_datasets.py

@@ -1,5 +1,8 @@
+import einops
 import pytest
 import torch
+from torchrl.data.replay_buffers.replay_buffers import TensorDictReplayBuffer
+from torchrl.data.replay_buffers.samplers import SamplerWithoutReplacement
 
 from lerobot.common.datasets.factory import make_offline_buffer
 from lerobot.common.utils import init_hydra_config
@@ -30,3 +33,30 @@ def test_factory(env_name, dataset_id):
         # TODO(rcadene): we assume for now that image normalization takes place in the model
         assert img.max() <= 1.0
         assert img.min() >= 0.0
+
+
+def test_compute_stats():
+    """Check that the correct statistics are computed.
+
+    We compare with taking a straight min, mean, max, std of all the data in one pass (which we can do
+    because we are working with a small dataset).
+
+    This test does not check that the stats_patterns are correct (instead, it relies on them).
+    """
+    cfg = init_hydra_config(
+        DEFAULT_CONFIG_PATH, overrides=["env=aloha", "env.task=sim_transfer_cube_human"]
+    )
+    buffer = make_offline_buffer(cfg)
+    # Get all of the data.
+    all_data = TensorDictReplayBuffer(
+            storage=buffer._storage,
+            batch_size=len(buffer),
+            sampler=SamplerWithoutReplacement(),
+    ).sample().float()
+    computed_stats = buffer._compute_stats()
+    for k, pattern in buffer.stats_patterns.items():
+        expected_mean = einops.reduce(all_data[k], pattern, "mean")
+        assert torch.allclose(computed_stats[k]["mean"], expected_mean)
+        assert torch.allclose(computed_stats[k]["std"], torch.sqrt(einops.reduce((all_data[k] - expected_mean) ** 2, pattern, "mean")))
+        assert torch.allclose(computed_stats[k]["min"], einops.reduce(all_data[k], pattern, "min"))
+        assert torch.allclose(computed_stats[k]["max"], einops.reduce(all_data[k], pattern, "max"))