Remove warnings (huggingface#111)

- Replace `use_pretrained_backbone` with `pretrained_backbone_weights`
- Bump diffusers' minimum version `0.26.3` -> `0.27.2`
- Add ignore flags in CI's pytest
- Change Box observation spaces in simulation environments
- Set `version_base="1.2"` in Hydra initializations
- Bump einops' minimum version `0.7.0` -> `0.8.0`

Author: aliberts

.github/workflows/test.yml

@@ -53,7 +53,14 @@ jobs:
           poetry install --all-extras
 
       - name: Test with pytest
-        run: pytest -v --cov=./lerobot --durations=0 tests
+        run: |
+          pytest tests -v --cov=./lerobot --durations=0 \
+            -W ignore::DeprecationWarning:imageio_ffmpeg._utils:7 \
+            -W ignore::UserWarning:torch.utils.data.dataloader:558 \
+            -W ignore::UserWarning:gymnasium.utils.env_checker:247 \
+            && rm -rf tests/outputs outputs
 
       - name: Test end-to-end
-        run: make test-end-to-end
+        run: |
+          make test-end-to-end \
+            && rm -rf outputs

lerobot/common/policies/act/configuration_act.py

@@ -33,8 +33,8 @@ class ActionChunkingTransformerConfig:
             deviation and "min_max" which rescale in a [-1, 1] range.
         unnormalize_output_modes: Similar dictionary as `normalize_input_modes`, but to unormalize in original scale.
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
-        use_pretrained_backbone: Whether the backbone should be initialized with pretrained weights from
-            torchvision.
+        pretrained_backbone_weights: Pretrained weights from torchvision to initalize the backbone.
+            `None` means no pretrained weights.
         replace_final_stride_with_dilation: Whether to replace the ResNet's final 2x2 stride with a dilated
             convolution.
         pre_norm: Whether to use "pre-norm" in the transformer blocks.
@@ -90,7 +90,7 @@ class ActionChunkingTransformerConfig:
     # Architecture.
     # Vision backbone.
     vision_backbone: str = "resnet18"
-    use_pretrained_backbone: bool = True
+    pretrained_backbone_weights: str | None = "ResNet18_Weights.IMAGENET1K_V1"
     replace_final_stride_with_dilation: int = False
     # Transformer layers.
     pre_norm: bool = False

lerobot/common/policies/act/modeling_act.py

@@ -104,7 +104,7 @@ def __init__(self, cfg: ActionChunkingTransformerConfig | None = None, dataset_s
         # Backbone for image feature extraction.
         backbone_model = getattr(torchvision.models, cfg.vision_backbone)(
             replace_stride_with_dilation=[False, False, cfg.replace_final_stride_with_dilation],
-            pretrained=cfg.use_pretrained_backbone,
+            weights=cfg.pretrained_backbone_weights,
             norm_layer=FrozenBatchNorm2d,
         )
         # Note: The assumption here is that we are using a ResNet model (and hence layer4 is the final feature

lerobot/common/policies/diffusion/configuration_diffusion.py

@@ -35,8 +35,8 @@ class DiffusionConfig:
             within the image size. If None, no cropping is done.
         crop_is_random: Whether the crop should be random at training time (it's always a center crop in eval
             mode).
-        use_pretrained_backbone: Whether the backbone should be initialized with pretrained weights from
-            torchvision.
+        pretrained_backbone_weights: Pretrained weights from torchvision to initalize the backbone.
+            `None` means no pretrained weights.
         use_group_norm: Whether to replace batch normalization with group normalization in the backbone.
             The group sizes are set to be about 16 (to be precise, feature_dim // 16).
         spatial_softmax_num_keypoints: Number of keypoints for SpatialSoftmax.
@@ -96,7 +96,7 @@ class DiffusionConfig:
     vision_backbone: str = "resnet18"
     crop_shape: tuple[int, int] | None = (84, 84)
     crop_is_random: bool = True
-    use_pretrained_backbone: bool = False
+    pretrained_backbone_weights: str | None = None
     use_group_norm: bool = True
     spatial_softmax_num_keypoints: int = 32
     # Unet.

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -378,13 +378,13 @@ def __init__(self, cfg: DiffusionConfig):
 
         # Set up backbone.
         backbone_model = getattr(torchvision.models, cfg.vision_backbone)(
-            pretrained=cfg.use_pretrained_backbone
+            weights=cfg.pretrained_backbone_weights
         )
         # Note: This assumes that the layer4 feature map is children()[-3]
         # TODO(alexander-soare): Use a safer alternative.
         self.backbone = nn.Sequential(*(list(backbone_model.children())[:-2]))
         if cfg.use_group_norm:
-            if cfg.use_pretrained_backbone:
+            if cfg.pretrained_backbone_weights:
                 raise ValueError(
                     "You can't replace BatchNorm in a pretrained model without ruining the weights!"
                 )

lerobot/common/utils/io_utils.py

@@ -0,0 +1,12 @@
+import warnings
+
+import imageio
+
+
+def write_video(video_path, stacked_frames, fps):
+    # Filter out DeprecationWarnings raised from pkg_resources
+    with warnings.catch_warnings():
+        warnings.filterwarnings(
+            "ignore", "pkg_resources is deprecated as an API", category=DeprecationWarning
+        )
+        imageio.mimsave(video_path, stacked_frames, fps=fps)

lerobot/common/utils/utils.py

@@ -92,7 +92,8 @@ def init_hydra_config(config_path: str, overrides: list[str] | None = None) -> D
     hydra.core.global_hydra.GlobalHydra.instance().clear()
     # Hydra needs a path relative to this file.
     hydra.initialize(
-        str(_relative_path_between(Path(config_path).absolute().parent, Path(__file__).absolute().parent))
+        str(_relative_path_between(Path(config_path).absolute().parent, Path(__file__).absolute().parent)),
+        version_base="1.2",
     )
     cfg = hydra.compose(Path(config_path).stem, overrides)
     return cfg

lerobot/configs/policy/act.yaml

@@ -45,7 +45,7 @@ policy:
   # Architecture.
   # Vision backbone.
   vision_backbone: resnet18
-  use_pretrained_backbone: true
+  pretrained_backbone_weights: ResNet18_Weights.IMAGENET1K_V1
   replace_final_stride_with_dilation: false
   # Transformer layers.
   pre_norm: false

lerobot/configs/policy/diffusion.yaml

@@ -61,7 +61,7 @@ policy:
   vision_backbone: resnet18
   crop_shape: [84, 84]
   crop_is_random: True
-  use_pretrained_backbone: false
+  pretrained_backbone_weights: null
   use_group_norm: True
   spatial_softmax_num_keypoints: 32
   # Unet.

lerobot/scripts/eval.py

@@ -38,7 +38,6 @@
 
 import einops
 import gymnasium as gym
-import imageio
 import numpy as np
 import torch
 from datasets import Dataset, Features, Image, Sequence, Value
@@ -51,13 +50,10 @@
 from lerobot.common.envs.utils import postprocess_action, preprocess_observation
 from lerobot.common.logger import log_output_dir
 from lerobot.common.policies.factory import make_policy
+from lerobot.common.utils.io_utils import write_video
 from lerobot.common.utils.utils import get_safe_torch_device, init_hydra_config, init_logging, set_global_seed
 
 
-def write_video(video_path, stacked_frames, fps):
-    imageio.mimsave(video_path, stacked_frames, fps=fps)
-
-
 def eval_policy(
     env: gym.vector.VectorEnv,
     policy: torch.nn.Module,

lerobot/scripts/train.py

@@ -22,7 +22,7 @@
 from lerobot.scripts.eval import eval_policy
 
 
-@hydra.main(version_base=None, config_name="default", config_path="../configs")
+@hydra.main(version_base="1.2", config_name="default", config_path="../configs")
 def train_cli(cfg: dict):
     train(
         cfg,

lerobot/scripts/visualize_dataset.py

@@ -16,7 +16,7 @@
 FIRST_FRAME = 0
 
 
-@hydra.main(version_base=None, config_name="default", config_path="../configs")
+@hydra.main(version_base="1.2", config_name="default", config_path="../configs")
 def visualize_dataset_cli(cfg: dict):
     visualize_dataset(cfg, out_dir=hydra.core.hydra_config.HydraConfig.get().runtime.output_dir)
 

poetry.lock

@@ -34,13 +34,13 @@ wandb = "^0.16.3"
 imageio = {extras = ["ffmpeg"], version = "^2.34.0"}
 gdown = "^5.1.0"
 hydra-core = "^1.3.2"
-einops = "^0.7.0"
+einops = "^0.8.0"
 pymunk = "^6.6.0"
 zarr = "^2.17.0"
 numba = "^0.59.0"
 torch = "^2.2.1"
 opencv-python = "^4.9.0.80"
-diffusers = "^0.26.3"
+diffusers = "^0.27.2"
 torchvision = "^0.18.0"
 h5py = "^3.10.0"
 huggingface-hub = "^0.21.4"