Merge branch 'rt1'

.gitignore

@@ -12,3 +12,142 @@ __pycache__/
 !/models/main_models/alfred/data/vis_feats/.gitkeep
 /dataset/*
 !/dataset/.gitkeep
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# Extras
+models/main_models/checkpoints/*
+models/main_models/datasets/*
+models/main_models/wandb/
+models/main_models/rt1_dataset/*
+models/main_models/traj_rollouts/*
+models/main_models/run_rt1_main.sh
+models/main_models/rt1_ft_env

models/main_models/rt1/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 Phil Wang
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

models/main_models/rt1/README.md

@@ -0,0 +1,66 @@
+# NPM-Dataset
+A comprehensive robotics dataset that includes navigation, perception, and manipulation data per data point.
+
+# RT-1 (Robotic Transformer) PyTorch Implementation
+<img src="./figures/rt1.png" width="450px"></img>
+
+A forked implementation of <a href = "https://github.com/Rohan138/rt1-pytorch.git"> RT1 (Robotic Transformer) </a> originally inspired by the <a href="https://ai.googleblog.com/2022/12/rt-1-robotics-transformer-for-real.html"> Google Research </a> paper.
+
+This implemenetation of RT-1 was pretrained on the <a href="https://sites.google.com/view/bridgedata"> Bridge </a> dataset and further fine-tuned on our LaNMP dataset for evaluation. Please find details of the repository below
+
+## Setup Instructions
+
+```bash
+git clone https://github.com/h2r/NPM-Dataset.git
+git checkout -b rt1
+pip install -e .
+```
+
+## Overview of files
+
+This repository has 7 critical files/folders whose use cases are described below
+
+1) ```main.py```: used to pretrain RT-1 on the bridge dataset. Modifying this file to accomodate different datasets requires changing the ```observation_space``` and ```action_space``` according to the dataset being loaded, as well as changing the dataset keys in ```rt1_pytorch/tokenizers/action_tokenizer.py```. Running this file saves a series of checkpoints and logs losses using weights and biases
+2) ```main_ft.py```: used to finetune RT-1 on the LaNMP dataset. This file has the ```observation_space``` and ```action_space``` and PyTorch ```DataLoader``` already modified to accomodate for the LaNMP dataset finetuning (AI2Thor). Running this file saves a series of checkpoints and logs losses using weights and biases
+3) ```main_ft_eval.py```: used to run RT-1 in inference mode on the LaNMP dataset. This file has the ```observation_space``` and ```action_space``` and PyTorch ```DataLoader``` already modified to accomodate for the LaNMP dataset (AI2Thor). The file iterates/loads all saved checkpoints from finetuning and runs RT-1 on inference mode for the validation dataset on each checkpoint. The script logs the test losses using weights and biases
+4) ```ai2thor_env.py```: contains a Gym environment style class to load and take steps in AI2Thor enivironment. This file is used to generate real-time trajectories based on the action tokens generated by a finetuned RT-1 model (specific for AI2Thor). The main ```step()``` function takes/executes the generated action by RT-1 and returns a success message along with information about the environment state e.g. object or agent metadata, which can be saved to capture the trajectory taken by the agent for a given task
+5) ```rollout_ai2thor.py```: interfaces between the finetuned RT-1 model (from a loaded checkpoint after finetuning on LaNMP) and the ```ai2thor_env.py``` Gym environment, in order to send observations from the AI2Thor environment to RT-1 and execute proposed action tokens by RT-1 on AI2Thor. Note that this file should not be run on a headless machine since it requires/deploys AI2Thor simulator GUI
+6) ```rt1_pytorch/rt1_policy.py```: contains the RT-1 model implementation in PyTorch. The ```loss()``` function performs forward pass of RT-1 for training and ```act()``` function performs the forward pass during inference.
+7) ```lanmp_dataloader/rt1_dataloader.py```: contains the ```DatasetManager``` class that extracts trajectories from the LaNMP ```sim_data.hdf5``` dataset file. The script automatically separates train and validation subsets according to different splits e.g. k-fold by scene, task wise or for diversity ablation. The ```DatasetManager``` also handles tokenizing/detokenizing the raw trajectory data into 256 discrete buckets, whilst also chunking trajectories across non-overlapping window lengths of 6 steps
+
+## Details about file arguments
+
+Most relevant files in this repository accept the same set of arguments that are detailed below
+* ```dataset```: only for the ```main.py``` file, specifies the dataset on which the RT-1 model should be pretrained
+* ```train-split```: specifies what fraction of the loaded dataset should be used for training v.s. evaluation
+* ```eval-split```: specifies what fraction of the laoded dataset should be used for evaluation v.s. training
+* ```epochs```: total number of passes over the all batches of the training set
+* ```lr```: learning rate for cross-entropy loss of RT1
+* ```train-batch-size```: the number of trajectories from which to sample data for the current training batch
+* ```eval-batch-size```: the number of trajectories from which to sample data for the current evaluation batch
+* ```trajectory-length```: the window size (context history of ```trajecotry-length``` previous images) used for each trajectory when feeding data to RT-1 model; this is set to 6 based on the RT-1 implementation 
+* ```sentence-transformer```: the language embedding to apply on the language-specified task
+* ```device```: the device to load the model/data onto during training/inference
+* ```eval-freq```: the interval of batches at which to run evaluation/inference on the validation dataset (currently set to 0 in ```main_ft.py```)
+* ```checkpoint-freq```: the interval of batches at which to save a checkpoint during training
+* ```checkpoint-dir```: the directory path at which to save a checkpoint during training
+* ```load-checkpoint```: (optional) path of the pretrained checkpoint to load for further fine-tuning 
+* ```wandb```: boolean determining if logging to weights and biases should happen
+* ```eval-scene```: the AI2Thor scene number in the dataset that is held out of the training set for evaluation during k-fold cross validation across scenes
+* ```split-type```: determines the split type (i.e. k-fold by scene, task wise or diversity ablation) between train and evaluation used by the ```DatasetManager``` in ```rt1_dataloader.py```
+* ```num-diversity-scenes```: only if ```split-type``` is ```diversity-ablation```, this is used to determine the total number of scenes to perform diversity ablation over i.e. maximum of 4 for LaNMP simulation data
+*  ```max-diversity-trajectories```: only if ```split-type``` is ```diversity-ablation```, this is used to determine the total number of trajectories that are divided evenly across the number of ```num-diversity-scenes``` scenes
+* ```train-subbatch```: the batch size to use during training/finetuning
+* ```eval-subbatch```: the batch size to use during evaluation
+
+## Checkpoint samples
+
+Please find the follow checkpoints samples that can be loaded to the RT-1 model. These can be found on the supplementary Google Drive associated with this project
+* ```sample_checkpoints/pretrained_bridge```: the final checkpoint saved when pretraining the RT-1 model on the Bridge dataset
+* ```sample_checkpoints/task_gen```: the final checkpoint saved after finetuning RT-1 model on the task-wise split for the task generalization experiment
+
+## Additional notes
+
+When running any of the finetuning or pretraining scripts, please ensure the following modules are loaded
+```module load cuda/11.8.0-lpttyok```
+```module load cudnn/8.7.0.84-11.8-lg2dpd5```