dreamer4 0.0.72__tar.gz → 0.0.74__tar.gz

{dreamer4-0.0.72 → dreamer4-0.0.74}/.github/workflows/test.yml

@@ -5,11 +5,11 @@ jobs:
   build:
 
     runs-on: ubuntu-latest
-    timeout-minutes: 20
+    timeout-minutes: 60
     strategy:
       fail-fast: false
       matrix:
-        group: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
+        group: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19]
 
     steps:
     - uses: actions/checkout@v4
@@ -24,4 +24,4 @@ jobs:
         python -m uv pip install -e .[test]
     - name: Test with pytest
       run: |
-        python -m pytest --num-shards 10 --shard-id ${{ matrix.group }} tests/
+        python -m pytest --num-shards 20 --shard-id ${{ matrix.group }} tests/

{dreamer4-0.0.72 → dreamer4-0.0.74}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dreamer4
-Version: 0.0.72
+Version: 0.0.74
 Summary: Dreamer 4
 Project-URL: Homepage, https://pypi.org/project/dreamer4/
 Project-URL: Repository, https://github.com/lucidrains/dreamer4

{dreamer4-0.0.72 → dreamer4-0.0.74}/dreamer4/dreamer4.py

@@ -82,6 +82,7 @@ class Experience:
     log_probs: tuple[Tensor, Tensor] | None = None
     values: Tensor | None = None
     step_size: int | None = None
+    lens: Tensor | None = None,
     agent_index: int = 0
     is_from_world_model: bool = True
 
@@ -90,6 +91,18 @@ def combine_experiences(
 ) -> Experience:
 
     assert len(exps) > 0
+
+    # set lens if not there
+
+    for exp in exps:
+        latents = exp.latents
+        batch, time, device = *latents.shape[:2], latents.device
+
+        if not exists(exp.lens):
+            exp.lens = torch.full((batch,), time, device = device)
+
+    # convert to dictionary
+
     exps_dict = [asdict(exp) for exp in exps]
 
     values, tree_specs = zip(*[tree_flatten(exp_dict) for exp_dict in exps_dict])
@@ -108,7 +121,11 @@ def combine_experiences(
     concatted = []
 
     for field_values in all_field_values:
+
         if is_tensor(first(field_values)):
+
+            field_values = pad_tensors_at_dim_to_max_len(field_values, dims = (1, 2))
+
             new_field_value = cat(field_values)
         else:
             new_field_value = first(list(set(field_values)))
@@ -222,6 +239,27 @@ def pad_at_dim(
     zeros = ((0, 0) * dims_from_right)
     return F.pad(t, (*zeros, *pad), value = value)
 
+def pad_to_len(t, target_len, *, dim):
+    curr_len = t.shape[dim]
+
+    if curr_len >= target_len:
+        return t
+
+    return pad_at_dim(t, (0, target_len - curr_len), dim = dim)
+
+def pad_tensors_at_dim_to_max_len(
+    tensors: list[Tensor],
+    dims: tuple[int, ...]
+):
+    for dim in dims:
+        if dim >= first(tensors).ndim:
+            continue
+
+        max_time = max([t.shape[dim] for t in tensors])
+        tensors = [pad_to_len(t, max_time, dim = dim) for t in tensors]
+
+    return tensors
+
 def align_dims_left(t, aligned_to):
     shape = t.shape
     num_right_dims = aligned_to.ndim - t.ndim
@@ -2560,12 +2598,16 @@ class DynamicsWorldModel(Module):
 
         # returning agent actions, rewards, and log probs + values for policy optimization
 
+        batch, device = latents.shape[0], latents.device
+        experience_lens = torch.full((batch,), time_steps, device = device)
+
         gen = Experience(
             latents = latents,
             video = video,
             proprio = proprio if has_proprio else None,
             step_size = step_size,
             agent_index = agent_index,
+            lens = experience_lens,
             is_from_world_model = True
         )
 

{dreamer4-0.0.72 → dreamer4-0.0.74}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "dreamer4"
-version = "0.0.72"
+version = "0.0.74"
 description = "Dreamer 4"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{dreamer4-0.0.72 → dreamer4-0.0.74}/tests/test_dreamer.py

@@ -653,7 +653,7 @@ def test_online_rl(
 
     # manually
 
-    one_experience = world_model_and_policy.interact_with_env(mock_env, max_timesteps = 16, env_is_vectorized = vectorized)
+    one_experience = world_model_and_policy.interact_with_env(mock_env, max_timesteps = 8, env_is_vectorized = vectorized)
     another_experience = world_model_and_policy.interact_with_env(mock_env, max_timesteps = 16, env_is_vectorized = vectorized)
 
     combined_experience = combine_experiences([one_experience, another_experience])