Merge branch 'user/alexander-soare/train_pusht' into user/alexander-soare/multistep_policy_and_serial_env

lerobot/common/policies/diffusion/model/multi_image_obs_encoder.py

@@ -1,15 +1,37 @@
 import copy
-from typing import Dict, Tuple, Union
+from typing import Dict, Optional, Tuple, Union
 
+import timm
 import torch
 import torch.nn as nn
 import torchvision
+from robomimic.models.base_nets import SpatialSoftmax
 
 from lerobot.common.policies.diffusion.model.crop_randomizer import CropRandomizer
 from lerobot.common.policies.diffusion.model.module_attr_mixin import ModuleAttrMixin
 from lerobot.common.policies.diffusion.pytorch_utils import replace_submodules
 
 
+class RgbEncoder(nn.Module):
+    """Following `VisualCore` from Robomimic 0.2.0."""
+
+    def __init__(self, input_shape, model_name="resnet18", pretrained=False, num_keypoints=32):
+        """
+        resnet_name: a timm model name.
+        pretrained: whether to use timm pretrained weights.
+        num_keypoints: Number of keypoints for SpatialSoftmax (default value of 32 matches PushT Image).
+        """
+        super().__init__()
+        self.backbone = timm.create_model(model_name, pretrained, num_classes=0, global_pool="")
+        # Figure out the feature map shape.
+        with torch.inference_mode():
+            feat_map_shape = tuple(self.backbone(torch.zeros(size=(1, *input_shape))).shape[1:])
+        self.pool = SpatialSoftmax(feat_map_shape, num_kp=num_keypoints)
+
+    def forward(self, x):
+        return torch.flatten(self.pool(self.backbone(x)), start_dim=1)
+
+
 class MultiImageObsEncoder(ModuleAttrMixin):
     def __init__(
         self,
@@ -24,7 +46,7 @@ class MultiImageObsEncoder(ModuleAttrMixin):
         share_rgb_model: bool = False,
         # renormalize rgb input with imagenet normalization
         # assuming input in [0,1]
-        imagenet_norm: bool = False,
+        norm_mean_std: Optional[tuple[float, float]] = None,
     ):
         """
         Assumes rgb input: B,C,H,W
@@ -98,10 +120,9 @@ class MultiImageObsEncoder(ModuleAttrMixin):
                         this_normalizer = torchvision.transforms.CenterCrop(size=(h, w))
                 # configure normalizer
                 this_normalizer = nn.Identity()
-                if imagenet_norm:
-                    # TODO(rcadene): move normalizer to dataset and env
+                if norm_mean_std is not None:
                     this_normalizer = torchvision.transforms.Normalize(
-                        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
+                        mean=norm_mean_std[0], std=norm_mean_std[1]
                     )
 
                 this_transform = nn.Sequential(this_resizer, this_randomizer, this_normalizer)

lerobot/common/policies/diffusion/policy.py

@@ -7,7 +7,7 @@ import torch
 from lerobot.common.policies.abstract import AbstractPolicy
 from lerobot.common.policies.diffusion.diffusion_unet_image_policy import DiffusionUnetImagePolicy
 from lerobot.common.policies.diffusion.model.lr_scheduler import get_scheduler
-from lerobot.common.policies.diffusion.model.multi_image_obs_encoder import MultiImageObsEncoder
+from lerobot.common.policies.diffusion.model.multi_image_obs_encoder import MultiImageObsEncoder, RgbEncoder
 
 
 class DiffusionPolicy(AbstractPolicy):
@@ -38,7 +38,7 @@ class DiffusionPolicy(AbstractPolicy):
         self.cfg = cfg
 
         noise_scheduler = hydra.utils.instantiate(cfg_noise_scheduler)
-        rgb_model = hydra.utils.instantiate(cfg_rgb_model)
+        rgb_model = RgbEncoder(input_shape=shape_meta.obs.image.shape, **cfg_rgb_model)
         obs_encoder = MultiImageObsEncoder(
             rgb_model=rgb_model,
             **cfg_obs_encoder,

lerobot/configs/policy/diffusion.yaml

@@ -42,8 +42,8 @@ policy:
   num_inference_steps: 100
   obs_as_global_cond: ${obs_as_global_cond}
   # crop_shape: null
-  diffusion_step_embed_dim: 256  # before 128
-  down_dims: [256, 512, 1024]  # before [512, 1024, 2048]
+  diffusion_step_embed_dim: 128
+  down_dims: [512, 1024, 2048]
   kernel_size: 5
   n_groups: 8
   cond_predict_scale: True
@@ -81,12 +81,12 @@ obs_encoder:
   # random_crop: True
   use_group_norm: True
   share_rgb_model: False
-  imagenet_norm: True
+  norm_mean_std: [0.5, 0.5]  # for PushT the original impl normalizes to [-1, 1] (maybe not the case for robomimic envs)
 
 rgb_model:
-  _target_: lerobot.common.policies.diffusion.pytorch_utils.get_resnet
-  name: resnet18
-  weights: null
+  model_name: resnet18
+  pretrained: false
+  num_keypoints: 32
 
 ema:
   _target_: lerobot.common.policies.diffusion.model.ema_model.EMAModel
@@ -109,13 +109,13 @@ training:
   debug: False
   resume: True
   # optimization
-  # lr_scheduler: cosine
-  # lr_warmup_steps: 500
-  num_epochs: 8000
+  lr_scheduler: cosine
+  lr_warmup_steps: 500
+  num_epochs: 500
   # gradient_accumulate_every: 1
   # EMA destroys performance when used with BatchNorm
   # replace BatchNorm with GroupNorm.
-  # use_ema: True
+  use_ema: True
   freeze_encoder: False
   # training loop control
   # in epochs