Source code for torchrl.envs.transforms.vc1
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import importlib
import os
import subprocess
from functools import partial
from typing import Union
import torch
from tensordict import TensorDictBase
from torch import nn
from torchrl._utils import logger as torchrl_logger
from torchrl.data.tensor_specs import Composite, DEVICE_TYPING, TensorSpec, Unbounded
from torchrl.envs.transforms.transforms import (
CenterCrop,
Compose,
ObservationNorm,
Resize,
ToTensorImage,
Transform,
)
from torchrl.envs.transforms.utils import _set_missing_tolerance
_has_vc = importlib.util.find_spec("vc_models") is not None
[docs]class VC1Transform(Transform):
"""VC1 Transform class.
VC1 provides pre-trained ResNet weights aimed at facilitating visual
embedding for robotic tasks. The models are trained using Ego4d.
See the paper:
VC1: A Universal Visual Representation for Robot Manipulation (Suraj Nair,
Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, Abhinav Gupta)
https://arxiv.org/abs/2203.12601
The VC1Transform is created in a lazy manner: the object will be initialized
only when an attribute (a spec or the forward method) will be queried.
The reason for this is that the :obj:`_init()` method requires some attributes of
the parent environment (if any) to be accessed: by making the class lazy we
can ensure that the following code snippet works as expected:
Examples:
>>> transform = VC1Transform("default", in_keys=["pixels"])
>>> env.append_transform(transform)
>>> # the forward method will first call _init which will look at env.observation_spec
>>> env.reset()
Args:
in_keys (list of NestedKeys): list of input keys. If left empty, the
"pixels" key is assumed.
out_keys (list of NestedKeys, optional): list of output keys. If left empty,
"VC1_vec" is assumed.
model_name (str): One of ``"large"``, ``"base"`` or any other compatible
model name (see the `github repo <https://github.com/facebookresearch/eai-vc>`_ for more info). Defaults to ``"default"``
which provides a small, untrained model for testing.
del_keys (bool, optional): If ``True`` (default), the input key will be
discarded from the returned tensordict.
"""
inplace = False
IMPORT_ERROR = (
"Could not load vc_models. You can install it via "
"VC1Transform.install_vc_models()."
)
def __init__(self, in_keys, out_keys, model_name, del_keys: bool = True):
if model_name == "default":
self.make_noload_model()
model_name = "vc1_vitb_noload"
self.model_name = model_name
self.del_keys = del_keys
super().__init__(in_keys=in_keys, out_keys=out_keys)
self._init()
def _init(self):
try:
from vc_models.models.vit import model_utils
except ModuleNotFoundError as err:
raise ModuleNotFoundError(self.IMPORT_ERROR) from err
if self.model_name == "base":
model_name = model_utils.VC1_BASE_NAME
elif self.model_name == "large":
model_name = model_utils.VC1_LARGE_NAME
else:
model_name = self.model_name
model, embd_size, model_transforms, model_info = model_utils.load_model(
model_name
)
self.model = model
self.embd_size = embd_size
self.model_transforms = self._map_tv_to_torchrl(model_transforms)
def _map_tv_to_torchrl(
self,
model_transforms,
in_keys=None,
):
if in_keys is None:
in_keys = self.in_keys
from torchvision import transforms
if isinstance(model_transforms, transforms.Resize):
size = model_transforms.size
if isinstance(size, int):
size = (size, size)
return Resize(
*size,
in_keys=in_keys,
)
elif isinstance(model_transforms, transforms.CenterCrop):
size = model_transforms.size
if isinstance(size, int):
size = (size,)
return CenterCrop(
*size,
in_keys=in_keys,
)
elif isinstance(model_transforms, transforms.Normalize):
return ObservationNorm(
in_keys=in_keys,
loc=torch.as_tensor(model_transforms.mean).reshape(3, 1, 1),
scale=torch.as_tensor(model_transforms.std).reshape(3, 1, 1),
standard_normal=True,
)
elif isinstance(model_transforms, transforms.ToTensor):
return ToTensorImage(
in_keys=in_keys,
)
elif isinstance(model_transforms, transforms.Compose):
transform_list = []
for t in model_transforms.transforms:
if isinstance(t, transforms.ToTensor):
transform_list.insert(0, t)
else:
transform_list.append(t)
if len(transform_list) == 0:
raise RuntimeError("Did not find any transform.")
for i, t in enumerate(transform_list):
if i == 0:
transform_list[i] = self._map_tv_to_torchrl(t)
else:
transform_list[i] = self._map_tv_to_torchrl(t)
return Compose(*transform_list)
else:
raise NotImplementedError(type(model_transforms))
def _call(self, tensordict):
if not self.del_keys:
in_keys = [
in_key
for in_key, out_key in zip(self.in_keys, self.out_keys)
if in_key != out_key
]
saved_td = tensordict.select(*in_keys)
with tensordict.view(-1) as tensordict_view:
super()._call(self.model_transforms(tensordict_view))
if self.del_keys:
tensordict.exclude(*self.in_keys, inplace=True)
else:
# reset in_keys
tensordict.update(saved_td)
return tensordict
forward = _call
def _reset(
self, tensordict: TensorDictBase, tensordict_reset: TensorDictBase
) -> TensorDictBase:
# TODO: Check this makes sense
with _set_missing_tolerance(self, True):
tensordict_reset = self._call(tensordict_reset)
return tensordict_reset
@torch.no_grad()
def _apply_transform(self, obs: torch.Tensor) -> None:
shape = None
if obs.ndimension() > 4:
shape = obs.shape[:-3]
obs = obs.flatten(0, -4)
out = self.model(obs)
if shape is not None:
out = out.view(*shape, *out.shape[1:])
return out
[docs] def transform_observation_spec(self, observation_spec: TensorSpec) -> TensorSpec:
if not isinstance(observation_spec, Composite):
raise ValueError("VC1Transform can only infer Composite")
keys = [key for key in observation_spec.keys(True, True) if key in self.in_keys]
device = observation_spec[keys[0]].device
dim = observation_spec[keys[0]].shape[:-3]
observation_spec = observation_spec.clone()
if self.del_keys:
for in_key in keys:
del observation_spec[in_key]
for out_key in self.out_keys:
observation_spec[out_key] = Unbounded(
shape=torch.Size([*dim, self.embd_size]), device=device
)
return observation_spec
[docs] def to(self, dest: Union[DEVICE_TYPING, torch.dtype]):
if isinstance(dest, torch.dtype):
self._dtype = dest
else:
self._device = dest
return super().to(dest)
@property
def device(self):
return self._device
@property
def dtype(self):
return self._dtype
@classmethod
def install_vc_models(cls, auto_exit=False):
try:
from vc_models import models # noqa: F401
torchrl_logger.info("vc_models found, no need to install.")
except ModuleNotFoundError:
HOME = os.environ.get("HOME")
vcdir = HOME + "/.cache/torchrl/eai-vc"
parentdir = os.path.dirname(os.path.abspath(vcdir))
os.makedirs(parentdir, exist_ok=True)
try:
from git import Repo
except ModuleNotFoundError as err:
raise ModuleNotFoundError(
"Could not load git. Make sure that `git` has been installed "
"in your virtual environment."
) from err
Repo.clone_from("https://github.com/facebookresearch/eai-vc.git", vcdir)
os.chdir(vcdir + "/vc_models")
subprocess.call(["python", "setup.py", "develop"])
if not auto_exit:
input(
"VC1 has been successfully installed. Exit this python run and "
"relaunch it again. Press Enter to exit..."
)
exit()
[docs] @classmethod
def make_noload_model(cls):
"""Creates an naive model at a custom destination."""
import vc_models
models_filepath = os.path.dirname(os.path.abspath(vc_models.__file__))
cfg_path = os.path.join(
models_filepath, "conf", "model", "vc1_vitb_noload.yaml"
)
if os.path.exists(cfg_path):
return
config = """_target_: vc_models.models.load_model
model:
_target_: vc_models.models.vit.vit.load_mae_encoder
checkpoint_path:
model:
_target_: torchrl.envs.transforms.vc1._vit_base_patch16
img_size: 224
use_cls: True
drop_path_rate: 0.0
transform:
_target_: vc_models.transforms.vit_transforms
metadata:
algo: mae
model: vit_base_patch16
data:
- ego
- imagenet
- inav
comment: 182_epochs
"""
with open(cfg_path, "w") as file:
file.write(config)
def _vit_base_patch16(**kwargs):
from vc_models.models.vit.vit import VisionTransformer
model = VisionTransformer(
patch_size=16,
embed_dim=16,
depth=4,
num_heads=4,
mlp_ratio=4,
qkv_bias=True,
norm_layer=partial(nn.LayerNorm, eps=1e-6),
**kwargs,
)
return model