tangger/lerobot_piper
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

revert dp changes, make act and tdmpc batch friendly

Browse Source
This commit is contained in:
Alexander Soare
2024-03-18 19:18:21 +00:00
parent 09ddd9bf92
commit 88347965c2
8 changed files with 32 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
lerobot/common/policies/diffusion/model/multi_image_obs_encoder.py
View File

@@ -1,37 +1,15 @@
import copy
from typing import Dict, Optional, Tuple, Union
from typing import Dict, 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,
@@ -46,7 +24,7 @@ class MultiImageObsEncoder(ModuleAttrMixin):
share_rgb_model: bool = False,
# renormalize rgb input with imagenet normalization
# assuming input in [0,1]
norm_mean_std: Optional[tuple[float, float]] = None,
imagenet_norm: bool = False,
):
"""
Assumes rgb input: B,C,H,W
@@ -120,9 +98,10 @@ class MultiImageObsEncoder(ModuleAttrMixin):
this_normalizer = torchvision.transforms.CenterCrop(size=(h, w))
# configure normalizer
this_normalizer = nn.Identity()
if norm_mean_std is not None:
if imagenet_norm:
# TODO(rcadene): move normalizer to dataset and env
this_normalizer = torchvision.transforms.Normalize(
mean=norm_mean_std[0], std=norm_mean_std[1]
mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
)
this_transform = nn.Sequential(this_resizer, this_randomizer, this_normalizer)