revert dp changes, make act and tdmpc batch friendly

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: 128
-  down_dims: [512, 1024, 2048]
+  diffusion_step_embed_dim: 256  # before 128
+  down_dims: [256, 512, 1024]  # before [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
-  norm_mean_std: [0.5, 0.5]  # for PushT the original impl normalizes to [-1, 1] (maybe not the case for robomimic envs)
+  imagenet_norm: True
 
 rgb_model:
-  model_name: resnet18
-  pretrained: false
-  num_keypoints: 32
+  _target_: lerobot.common.policies.diffusion.pytorch_utils.get_resnet
+  name: resnet18
+  weights: null
 
 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: 500
+  # lr_scheduler: cosine
+  # lr_warmup_steps: 500
+  num_epochs: 8000
   # 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